diff --git a/BASELINE_CATALOG.json b/BASELINE_CATALOG.json new file mode 100644 index 0000000000000000000000000000000000000000..5866281e2ee398d93b589a79b355f3ab0a26fa37 --- /dev/null +++ b/BASELINE_CATALOG.json @@ -0,0 +1,53 @@ +{ + "delivery_status": "partial: OMG-XL updated acceptance checkpoint upload blocked by HF private storage quota", + "pending_file": { + "model": "omg-xl", + "path": "handoff-v1/acceptance/updated.ckpt", + "reason": "HF 403 Private storage limit reached; verified local full checkpoint retained" + }, + "scope": "All five baseline families, seven actually trained versions; original representations preserved", + "repositories": { + "genmo": { + "repo_id": "starstone3/HFD-GENMO-SMPL-Training", + "url": "https://huggingface.co/starstone3/HFD-GENMO-SMPL-Training", + "private": true + }, + "hymotion": { + "repo_id": "starstone3/HFD-HY-Motion-SMPL-Training", + "url": "https://huggingface.co/starstone3/HFD-HY-Motion-SMPL-Training", + "private": true + }, + "kimodo": { + "repo_id": "starstone3/HFD-Kimodo-SMPLX-Training", + "url": "https://huggingface.co/starstone3/HFD-Kimodo-SMPLX-Training", + "private": true + }, + "omg-b": { + "repo_id": "starstone3/HFD-OMG-B-Training", + "private": true, + "url": "https://huggingface.co/starstone3/HFD-OMG-B-Training" + }, + "omg-l": { + "repo_id": "starstone3/HFD-OMG-L-Training", + "private": true, + "url": "https://huggingface.co/starstone3/HFD-OMG-L-Training" + }, + "omg-xl": { + "repo_id": "starstone3/HFD-OMG-XL-Training", + "private": true, + "url": "https://huggingface.co/starstone3/HFD-OMG-XL-Training" + }, + "textop": { + "repo_id": "starstone3/HFD-TextOp-Training", + "private": true, + "url": "https://huggingface.co/starstone3/HFD-TextOp-Training" + } + }, + "heartbeat": "416e9f54 cancelled by user; not recreated", + "quality_protocol": { + "repo": "starstone3/HFD-Motion-Quality-Eval", + "repo_type": "dataset", + "revision": "b4f10ba13d414f819b3ee60d472ae7ec1197f220" + }, + "limitations": "Read per-model README and acceptance-report; full-state checkpoint availability is distinct from post-budget continuation and bitwise reproducibility." +} \ No newline at end of file diff --git a/README.md b/README.md new file mode 100644 index 0000000000000000000000000000000000000000..0fada6fdb11ef68135f1fa8e3bbc6749a39eec3b --- /dev/null +++ b/README.md @@ -0,0 +1,117 @@ +--- +license: other +--- +# omg-b 实际训练工程交付 + +本版本在本机实际训练过。原生输出为 G1 机器人动作,不是 SMPL/SMPL-X 人体模型。此私有仓库按同样完整训练工程标准交付。OMG-B/L/XL 分别使用官方50m/300m/500m配置;官方100m仅在外部资源清单中,不冒充已训练版本。 + +## 内容与恢复边界 + +latest 与 best 均为100000,含完整状态;现存较早正式断点已被原保留策略清理,原入口要求 step < max_steps,因此本版本没有宣称完成100K之后的原方案短跑。提供完整状态字段审计,预算扩展需要接手方明确新阶段方案。 + +`handoff-v1/packages/original-training-workspace.tar.gz` 是实际完整工作树(含 upstream、未提交修改、自定义实现和配置),非精简评测包。Git commit/diff/status 在 inventory;`.git` 内部、Python缓存、凭据排除。`*-all-run-history` 包含此家族原 run、评测、原始轨迹和日志;checkpoint单独存放。包内各文件大小和SHA在相邻manifest;仓库全部交付文件由根MANIFEST.json核验。完整模型最新和best见 checkpoints 与 inventory/models.json,未伪造缺失状态。 + +实际单GPU/rank0,无缺失分布式分片。OMG完整Lightning状态含 optimizer_states、lr_schedulers、loops、callbacks、epoch/global_step、hfd_resume 的Python/NumPy/torch/CUDA RNG和已消费 train_pool/val_pool;使用bf16-mixed,无FP16 scaler,未单独启用EMA。TextOp完整字段见审计,VAE和CLIP是冻结外部依赖。 + +每卡batch/累积/有效batch:512 × 累积1 = 512。种子42;OMG AdamW LR6e-5、2000 warmup、cosine至1e-6;TextOp LR1e-4、FP32、三个100K课程阶段。完整损失权重、所有最终解析配置在 original-run/resolved_config.yaml 和 native_config.yaml(TextOp另cfg.yaml),不是仅默认YAML。 + +## 源码与环境证据 + +实际环境 Python3.10.21 / PyTorch2.7.1+cu128 / CUDA12.8 / cuDNN90701,RTX5090,驱动580.76.05。environment-lock.txt为实际pip冻结清单;原requirements也随工作区保留。自定义算子/机器人运动学源码随 upstream 提供;未声称可选部署/仿真组件全部在新机器编译通过。没有经过重建验收的容器镜像。 + +OMG-L/XL 的训练源码哈希与当前完整快照全部一致。B/TextOp有后续修改:initial和latest-resume审计分别保留;historical-source保存能从本地Git按原SHA恢复的精确旧文件,不应盲目覆盖整个目录。最后恢复时的数据采样 ondemand.py 与当前一致;早期初始记录的不同版本和不相关HY辅助文件缺失项明确记录。短跑使用当前实际工作区,不能将经验上的配对一致扩大成所有历史源码均完全相同。历史完整shell环境/启动命令未全部保留;实际配置为原证据,本次验收 command.json 才是逐字执行记录。 + +## 数据与表示 + +训练126950 / validation302 英文会话;原始完整train/val/test定义及每个样本的排序/过滤/语言/窗口采样见数据索引包、train_native_manifest.json、val_native_manifest.json及采样日志。绝不以Test1024代替训练。基准release固定为 Tinghong-Ye22/data_vqa_all_1025@bdca2fe7581c263e9ceb3a96204b314e199faddf;raw源 Tinghong-Ye22/hfd_data@93ba4a7a9c66b3e1b7041e5f4ea038ef2418b948。 + +675个raw分片未重复上传;resources/RAW_REQUIRED_FILES.json列出大小、SHA256和上游清单来源(本次未重新读完所有原始分片)。对齐数据库、release/index/startup缓存和预处理代码已提供。参考文本为原有因果对齐的前景转录,保留0.1s guard,不使用未来文本。OMG额外使用过去2s音频;TextOp仅CLIP文本无直接音频。冻结编码器的临时内存缓存可重建,不作为采样状态。 + +输入HFD为50Hz、IsaacLab 29关节、root xyz + wxyz四元数。OMG重采样30Hz、native29关节、历史60/未来60帧,表示和坐标规范由 hfd_bridge/training.py、joints.py 与 upstream codec 精确定义。TextOp为50Hz、23关节57D特征、历史2/未来8帧、4个primitive;统计meanstd.pkl、robot XML和CLIP tokenizer均有文件哈希。无另行新增增强;原数据合成/筛选/窗口规则由配置和源码保留。外部官方初始化、冻结T5-base或CLIP/VAE及全部模型资产的固定revision/URL/大小/SHA见EXTERNAL_ASSETS.json;这些大依赖不重复上传。 + +## 下载、路径和资源恢复 + +```bash +hf download starstone3/HFD-OMG-B-Training --local-dir ./handoff-omg-b +cd handoff-omg-b +python verify_manifest.py MANIFEST.json +mkdir -p restored +for f in handoff-v1/packages/*.tar.gz; do tar -xzf "$f" -C restored; done +``` + +优先在隔离机器/容器把 restored/root/autodl-tmp 和 restored/root/autodl-fs 映射回原绝对路径。不要覆盖现有训练;原配置有绝对路径,不能删除合同校验规避路径问题。checkpoint另从 handoff-v1/checkpoints 复制到 original-run/resolved_config.yaml 所述原run路径:OMG的best.ckpt放best/,step-*.ckpt放checkpoints/;L/XL共享目录另需原hfd_run.json指向原日志run目录。TextOp ckpt_*.pth放原run根目录。共享与本地兼容链接按原配置建立。原日志包中的symlink未打包,需据本段恢复。验收使用原本地run路径查找配置,故不要只拷一个权重文件。 + +使用原 tools/training_assets.py --help,按EXTERNAL_ASSETS逐项固定revision下载并校验;hf download 的 repo_type 和 filename 必须按清单,不用浮动main替代。原始数据仅下载RAW_REQUIRED_FILES所列路径并执行: + +```bash +python handoff-v1/resources/verify_raw.py handoff-v1/resources/RAW_REQUIRED_FILES.json /root/autodl-fs/pangxinlei/hfd_data +``` + +在本机/AutoDL上安装和下载使用 /root/.local/bin/turbo;HF失败只用hfmirror,不自动消耗个人代理。接手机器按自己的网络配置。Token仅放自己的认证环境,不写包内。 + +## 续训和配对验收 + +验收资料包含 command.json、sample_ids.jsonl、指标、更新后完整权重和逐张量SHA。样本日志标记逻辑pool step;OMG预取可能记录下一批尚未消费样本,消费边界以输出checkpoint hfd_resume.train_pool为准。单GPU、相同代码/依赖/硬件/数据哈希/缓存输入/算法设置是配对前提;不保证跨GPU逐位一致。短跑结果和失败尝试以 acceptance-report.json 为准。只有训练loss一致不代表完整更新参数逐位一致,更不等于动作质量通过。 + +## 推理、验证、已知问题 + +原 evaluate_omg.py / evaluate_textop.py、infer_overfit.py、完整HFD_bench评测脚本及配置随原工作区和actual-recovery-and-evaluation包提供;直接使用其 --help,原实际参数和结果路径可查工作区 docs/CURRENT_RUNS.md 与各summary.json。模型/step/代码SHA/最终配置/评测产物对应关系见result-map.json;不把训练曲线当正式FID。 + +OMG存在自主历史接段突变、位移不足/过度;TextOp存在停止指令响应和长时自主历史恢复问题,GT历史oracle改善不是有效自主成绩。历史日志含失败、临时补丁、质量覆盖限制;本交付不宣称行为验收通过。未完成项:新环境全量容器重建、跨硬件逐位证明、原预算结束后的新阶段扩展、缺失历史源码/检查点恢复,均未虚报完成。 + +## 完整环境、依赖资产与最终历史源码补充 + +vendored OMG/TextOp 的官方原始提交见 inventory/vendor-provenance.json;upstream 子目录 git rev-parse 返回的是外围工作区commit,不能混淆。B/TextOp最后一次正式恢复涉及的本模型源码均已按原记录SHA找回,差异文件位于 resources/final-resume-historical-source;其中resume.py来自先前保留diff精确还原。初始更早快照及不相关HY文件的缺失仍如实保留。将此overlay应用到隔离工作区可还原最后正式源码;本次短跑为当前工作区,配对结果另列。 + +在新机器建立对应Python3.10环境后,原训练环境依赖的实际安装清单如下(此处是重建命令,未宣称干净容器重建已验收): + +```bash +/root/.local/bin/turbo python -m pip install --extra-index-url https://download.pytorch.org/whl/cu128 -r handoff-v1/resources/environment-lock.txt +``` + +固定资产获取和完整SHA校验: + +```bash +cd /root/autodl-tmp/hfd-baseline-portable +/root/.local/bin/turbo /root/autodl-tmp/envs/omg/bin/python tools/training_assets.py --models-root /root/autodl-tmp/models --raw-root /root/autodl-fs/pangxinlei/hfd_data --models omg --omg-sizes 50m --download +``` + +训练数据分片按资源清单下载(须有上游权限),随后用verify_raw.py校验;下载脚本保存为 resources/download_raw.py: + +```bash +/root/.local/bin/turbo python handoff-v1/resources/download_raw.py handoff-v1/resources/RAW_REQUIRED_FILES.json /root/autodl-fs/pangxinlei/hfd_data +``` + +原入口每次保存检查点使用原原子保存/Lightning插件;本次有界验收输出checkpoint已读取校验。正常训练默认保留数量见最终配置,勿改成逐step无限保留。学习曲线保留全部历史日志顺序(包括原恢复重跑的重复step),不偷偷删掉不利数据。 + +在隔离恢复机器解压原工作区/数据/历史后,从仓库根执行以下命令恢复 checkpoint 原路径与L/XL owner/兼容链接(存在不同权重时拒绝覆盖): + +```bash +python handoff-v1/resources/restore_checkpoint_layout.py --repo . +``` + +环境锁中的本地安装源码全部额外保存在 original-installed-workspaces.tar.gz(含实际CLIP、Isaac工具和原OMG/TextOp工作树、OMG路径别名)。先解压映射原路径,再安装environment-lock;不要把原本地file://依赖替换成随意最新版。 + +## Lightning 恢复边界 + +本次L/XL验收为保护原训练,写到新的输出目录。Lightning因此重置检查点排名/保存路径等ModelCheckpoint回调元数据(stdout保留原警告);模型、优化器、LR调度和HFD RNG/已消费采样池由原恢复机制加载。原完整checkpoint仍保留原callbacks状态;要求连best保留回调也一致时,应在隔离机器恢复原路径。不能把本次独立输出验收说成所有回调元数据逐项不变。 + +Lightning还有通用的DataLoader非stateful提示;此工程另用HFDResumeState保存和恢复已消费逻辑pool与RNG,相关源码/输出状态和配对数值全部提供。OMG停止回调触发后原trainer可能执行末尾validation,已保留其日志和最终状态;验收没有跳过原验证逻辑。 + +## 原生 validation 命令 + +以下调用原来实际使用的验证入口,使用validation划分和固定seed;输出完整轨迹、初始/训练模型对照和诊断。已有正式Test1024结果仅归档,不用于本次选参。 + +```bash +cd /root/autodl-tmp/hfd-baseline-portable +/root/autodl-tmp/envs/omg/bin/python evaluate_omg.py --trained-run /root/autodl-tmp/runs/hfd-omg-english-split20261026 --checkpoint /root/autodl-tmp/runs/hfd-omg-english-split20261026/best/best.ckpt --output /path/to/new-omg-b-validation --gpu 0 --split val --seed 12345 --per-mode 1 --third-party 3 --stage all +``` + +复验短跑更新后的模型时,将--checkpoint改为已下载的 acceptance/updated 文件;--trained-run仍提供相应原模型配置。小样本诊断不叫正式FID;正式固定窗评分入口、依赖和结果复核协议见catalog锁定的公共质量评测工程。 + +## 完整历史质量结果及本次验收 + +本仓库额外附本版本 complete-quality-results.tar.gz:1024完整预测轨迹、packed评分输入、官方scores、run_config权重SHA及cohort锁。数据均为已有正式评测,没有在本次重新选参。公共评分工程固定版本见 BASELINE_CATALOG.json;质量产物内部大小/SHA清单与归档验证记录已提供。 + +本次短跑/拒绝恢复的完整结果见 handoff-v1/acceptance/acceptance-report.json。OMG-B没有更新步验收,原因和原入口实际拒绝日志已提供;其他版本的原始数值配对及最终参数SHA在同目录,不保证逐位一致。 + diff --git a/handoff-v1/acceptance/acceptance-report.json b/handoff-v1/acceptance/acceptance-report.json new file mode 100644 index 0000000000000000000000000000000000000000..b26e0d28790c0bcfba4145c1470055c18f795105 --- /dev/null +++ b/handoff-v1/acceptance/acceptance-report.json @@ -0,0 +1,41 @@ +{ + "status": "complete_checkpoint_present_but_post_budget_training_not_accepted", + "global_step": 100000, + "keys": [ + "epoch", + "global_step", + "pytorch-lightning_version", + "state_dict", + "loops", + "callbacks", + "optimizer_states", + "lr_schedulers", + "hfd_resume", + "omg_model_architecture" + ], + "all_parameters_finite": true, + "optimizer_steps": [ + 100000 + ], + "rng_keys": [ + "python", + "numpy", + "torch", + "cuda" + ], + "train_pool_present": true, + "scheduler_present": true, + "attempt_argv": [ + "/root/autodl-tmp/envs/omg/bin/python", + "train.py", + "--config", + "/root/autodl-tmp/runs/hfd-omg-english-split20261026/resolved_config.yaml", + "--resume", + "/root/autodl-tmp/runs/hfd-omg-english-split20261026/last.ckpt", + "--dry-run", + "--set", + "paths.output_dir=/root/autodl-fs/pangxinlei/handoffs/all-baselines-20261111/acceptance/omg-b/unstarted-run" + ], + "returncode": 1, + "reason": "All retained formal B checkpoints are at original max_steps=100000. Original entry rejects continuation; earlier checkpoints removed by original retention. No optimizer state fabricated or schedule contract bypassed. No short-update acceptance claimed." +} \ No newline at end of file diff --git a/handoff-v1/acceptance/resume-at-budget-limit.log b/handoff-v1/acceptance/resume-at-budget-limit.log new file mode 100644 index 0000000000000000000000000000000000000000..fe7c38548a5dc1cfa10143c9f29c5e7a8bb89966 --- /dev/null +++ b/handoff-v1/acceptance/resume-at-budget-limit.log @@ -0,0 +1,8 @@ +Traceback (most recent call last): + File "/root/autodl-tmp/hfd-baseline-portable/train.py", line 409, in + main() + File "/root/autodl-tmp/hfd-baseline-portable/train.py", line 366, in main + resume=select_checkpoint(c,a.resume if a.resume is not None else c.training.get('resume'),path) + File "/root/autodl-tmp/hfd-baseline-portable/hfd_bridge/resume.py", line 100, in select_checkpoint + raise ValueError('Resume requires a full optimizer checkpoint with 0 < step < max_steps') +ValueError: Resume requires a full optimizer checkpoint with 0 < step < max_steps diff --git a/handoff-v1/inventory/archive-verification.json b/handoff-v1/inventory/archive-verification.json new file mode 100644 index 0000000000000000000000000000000000000000..c53fedb45e03f778b058e0ca61e5eda442c8e530 --- /dev/null +++ b/handoff-v1/inventory/archive-verification.json @@ -0,0 +1,32 @@ +[ + { + "archive": "omg-all-run-history.tar.gz", + "files": 287, + "errors": [] + }, + { + "archive": "omg-small-assets.tar.gz", + "files": 0, + "errors": [] + }, + { + "archive": "textop-all-run-history.tar.gz", + "files": 225, + "errors": [] + }, + { + "archive": "textop-small-assets.tar.gz", + "files": 81, + "errors": [] + }, + { + "archive": "original-training-workspace.tar.gz", + "files": 1138, + "errors": [] + }, + { + "archive": "omg-native-assets.tar.gz", + "files": 5, + "errors": [] + } +] \ No newline at end of file diff --git a/handoff-v1/inventory/compiler-runtime.json b/handoff-v1/inventory/compiler-runtime.json new file mode 100644 index 0000000000000000000000000000000000000000..429868a256ee579cc2afad77f865ee225cc59ae7 --- /dev/null +++ b/handoff-v1/inventory/compiler-runtime.json @@ -0,0 +1,15 @@ +{ + "nvcc": { + "status": "not_in_PATH" + }, + "cxx": { + "returncode": 0, + "stdout": "c++ (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0\nCopyright (C) 2021 Free Software Foundation, Inc.\nThis is free software; see the source for copying conditions. There is NO\nwarranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.\n\n", + "stderr": "" + }, + "ldd": { + "returncode": 0, + "stdout": "ldd (Ubuntu GLIBC 2.35-0ubuntu3.8) 2.35\nCopyright (C) 2022 Free Software Foundation, Inc.\nThis is free software; see the source for copying conditions. There is NO\nwarranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.\nWritten by Roland McGrath and Ulrich Drepper.\n", + "stderr": "" + } +} \ No newline at end of file diff --git a/handoff-v1/inventory/evaluation-core-git.json b/handoff-v1/inventory/evaluation-core-git.json new file mode 100644 index 0000000000000000000000000000000000000000..c4386cef2e73fdad2f43595e738314ec9fa800eb --- /dev/null +++ b/handoff-v1/inventory/evaluation-core-git.json @@ -0,0 +1,5 @@ +{ + "commit": "fatal: not a git repository (or any parent up to mount point /root)\nStopping at filesystem boundary (GIT_DISCOVERY_ACROSS_FILESYSTEM not set).\n", + "status": "fatal: not a git repository (or any parent up to mount point /root)\nStopping at filesystem boundary (GIT_DISCOVERY_ACROSS_FILESYSTEM not set).\n", + "source_diff": "warning: Not a git repository. Use --no-index to compare two paths outside a working tree\nusage: git diff --no-index [] \n\nDiff output format options\n -p, --patch generate patch\n -s, --no-patch suppress diff output\n -u generate patch\n -U, --unified[=] generate diffs with lines context\n -W, --function-context\n generate diffs with lines context\n --raw generate the diff in raw format\n --patch-with-raw synonym for '-p --raw'\n --patch-with-stat synonym for '-p --stat'\n --numstat machine friendly --stat\n --shortstat output only the last line of --stat\n -X, --dirstat[=...]\n output the distribution of relative amount of changes for each sub-directory\n --cumulative synonym for --dirstat=cumulative\n --dirstat-by-file[=...]\n synonym for --dirstat=files,param1,param2...\n --check warn if changes introduce conflict markers or whitespace errors\n --summary condensed summary such as creations, renames and mode changes\n --name-only show only names of changed files\n --name-status show only names and status of changed files\n --stat[=[,[,]]]\n generate diffstat\n --stat-width generate diffstat with a given width\n --stat-name-width \n generate diffstat with a given name width\n --stat-graph-width \n generate diffstat with a given graph width\n --stat-count generate diffstat with limited lines\n --compact-summary generate compact summary in diffstat\n --binary output a binary diff that can be applied\n --full-index show full pre- and post-image object names on the \"index\" lines\n --color[=] show colored diff\n --ws-error-highlight \n highlight whitespace errors in the 'context', 'old' or 'new' lines in the diff\n -z do not munge pathnames and use NULs as output field terminators in --raw or --numstat\n --abbrev[=] use digits to display object names\n --src-prefix \n show the given source prefix instead of \"a/\"\n --dst-prefix \n show the given destination prefix instead of \"b/\"\n --line-prefix \n prepend an additional prefix to every line of output\n --no-prefix do not show any source or destination prefix\n --inter-hunk-context \n show context between diff hunks up to the specified number of lines\n --output-indicator-new \n specify the character to indicate a new line instead of '+'\n --output-indicator-old \n specify the character to indicate an old line instead of '-'\n --output-indicator-context \n specify the character to indicate a context instead of ' '\n\nDiff rename options\n -B, --break-rewrites[=[/]]\n break complete rewrite changes into pairs of delete and create\n -M, --find-renames[=]\n detect renames\n -D, --irreversible-delete\n omit the preimage for deletes\n -C, --find-copies[=]\n detect copies\n --find-copies-harder use unmodified files as source to find copies\n --no-renames disable rename detection\n --rename-empty use empty blobs as rename source\n --follow continue listing the history of a file beyond renames\n -l prevent rename/copy detection if the number of rename/copy targets exceeds given limit\n\nDiff algorithm options\n --minimal produce the smallest possible diff\n -w, --ignore-all-space\n ignore whitespace when comparing lines\n -b, --ignore-space-change\n ignore changes in amount of whitespace\n --ignore-space-at-eol\n ignore changes in whitespace at EOL\n --ignore-cr-at-eol ignore carrier-return at the end of line\n --ignore-blank-lines ignore changes whose lines are all blank\n -I, --ignore-matching-lines \n ignore changes whose all lines match \n --indent-heuristic heuristic to shift diff hunk boundaries for easy reading\n --patience generate diff using the \"patience diff\" algorithm\n --histogram generate diff using the \"histogram diff\" algorithm\n --diff-algorithm \n choose a diff algorithm\n --anchored generate diff using the \"anchored diff\" algorithm\n --word-diff[=] show word diff, using to delimit changed words\n --word-diff-regex \n use to decide what a word is\n --color-words[=]\n equivalent to --word-diff=color --word-diff-regex=\n --color-moved[=]\n moved lines of code are colored differently\n --color-moved-ws \n how white spaces are ignored in --color-moved\n\nOther diff options\n --relative[=]\n when run from subdir, exclude changes outside and show relative paths\n -a, --text treat all files as text\n -R swap two inputs, reverse the diff\n --exit-code exit with 1 if there were differences, 0 otherwise\n --quiet disable all output of the program\n --ext-diff allow an external diff helper to be executed\n --textconv run external text conversion filters when comparing binary files\n --ignore-submodules[=]\n ignore changes to submodules in the diff generation\n --submodule[=]\n specify how differences in submodules are shown\n --ita-invisible-in-index\n hide 'git add -N' entries from the index\n --ita-visible-in-index\n treat 'git add -N' entries as real in the index\n -S look for differences that change the number of occurrences of the specified string\n -G look for differences that change the number of occurrences of the specified regex\n --pickaxe-all show all changes in the changeset with -S or -G\n --pickaxe-regex treat in -S as extended POSIX regular expression\n -O control the order in which files appear in the output\n --rotate-to show the change in the specified path first\n --skip-to skip the output to the specified path\n --find-object \n look for differences that change the number of occurrences of the specified object\n --diff-filter [(A|C|D|M|R|T|U|X|B)...[*]]\n select files by diff type\n --output Output to a specific file\n\n" +} \ No newline at end of file diff --git a/handoff-v1/inventory/evaluation-core-verification.json b/handoff-v1/inventory/evaluation-core-verification.json new file mode 100644 index 0000000000000000000000000000000000000000..14db683421a893456ba2216e0bd1526cc7714c75 --- /dev/null +++ b/handoff-v1/inventory/evaluation-core-verification.json @@ -0,0 +1,4 @@ +{ + "files": 1211, + "all_archive_hashes_verified": true +} \ No newline at end of file diff --git a/handoff-v1/inventory/gpu.txt b/handoff-v1/inventory/gpu.txt new file mode 100644 index 0000000000000000000000000000000000000000..4eba653f77e75ca1865cdb8409d6567be4fbbc9f --- /dev/null +++ b/handoff-v1/inventory/gpu.txt @@ -0,0 +1,3 @@ +index, name, uuid, driver_version, memory.total [MiB] +0, NVIDIA GeForce RTX 5090, GPU-8927e2b9-6dd9-6b9a-c243-b3b98a2c3b46, 580.76.05, 32607 MiB +1, NVIDIA GeForce RTX 5090, GPU-98671241-dcab-5613-9475-692e4a351157, 580.76.05, 32607 MiB diff --git a/handoff-v1/inventory/hymotion-final-pilot.json b/handoff-v1/inventory/hymotion-final-pilot.json new file mode 100644 index 0000000000000000000000000000000000000000..62f8b0b5def436c036748584b500438a305b1984 --- /dev/null +++ b/handoff-v1/inventory/hymotion-final-pilot.json @@ -0,0 +1,29 @@ +{ + "step": 250, + "keys": [ + "format", + "training_scheme", + "initialization", + "model", + "optimizer", + "scheduler", + "step", + "best", + "rng", + "feedback_rng", + "train_pool", + "val_pool", + "train_chain", + "signature", + "base_assets", + "config" + ], + "completed_receipt": { + "time": 1794387797.1160033, + "steps": 250 + }, + "status": "bounded pilot completed, not formal training; no new quality pass claim", + "supersedes": "prior handoff step200 snapshot; retained as historical evidence", + "archive_entries": 20, + "archive_hashes_verified": true +} \ No newline at end of file diff --git a/handoff-v1/inventory/installed-workspaces-git.json b/handoff-v1/inventory/installed-workspaces-git.json new file mode 100644 index 0000000000000000000000000000000000000000..1df55658ac1ba9123f71bce97bea4f0e06c0b1a2 --- /dev/null +++ b/handoff-v1/inventory/installed-workspaces-git.json @@ -0,0 +1,56 @@ +{ + "/root/autodl-tmp/datasets/eval_textop_omg/OMG_code_20260927/OMG": { + "commit": { + "returncode": 128, + "output": "fatal: not a git repository (or any parent up to mount point /root)\nStopping at filesystem boundary (GIT_DISCOVERY_ACROSS_FILESYSTEM not set).\n" + }, + "diff": { + "returncode": 129, + "output": "warning: Not a git repository. Use --no-index to compare two paths outside a working tree\nusage: git diff --no-index [] \n\nDiff output format options\n -p, --patch generate patch\n -s, --no-patch suppress diff output\n -u generate patch\n -U, --unified[=] generate diffs with lines context\n -W, --function-context\n generate diffs with lines context\n --raw generate the diff in raw format\n --patch-with-raw synonym for '-p --raw'\n --patch-with-stat synonym for '-p --stat'\n --numstat machine friendly --stat\n --shortstat output only the last line of --stat\n -X, --dirstat[=...]\n output the distribution of relative amount of changes for each sub-directory\n --cumulative synonym for --dirstat=cumulative\n --dirstat-by-file[=...]\n synonym for --dirstat=files,param1,param2...\n --check warn if changes introduce conflict markers or whitespace errors\n --summary condensed summary such as creations, renames and mode changes\n --name-only show only names of changed files\n --name-status show only names and status of changed files\n --stat[=[,[,]]]\n generate diffstat\n --stat-width generate diffstat with a given width\n --stat-name-width \n generate diffstat with a given name width\n --stat-graph-width \n generate diffstat with a given graph width\n --stat-count generate diffstat with limited lines\n --compact-summary generate compact summary in diffstat\n --binary output a binary diff that can be applied\n --full-index show full pre- and post-image object names on the \"index\" lines\n --color[=] show colored diff\n --ws-error-highlight \n highlight whitespace errors in the 'context', 'old' or 'new' lines in the diff\n -z do not munge pathnames and use NULs as output field terminators in --raw or --numstat\n --abbrev[=] use digits to display object names\n --src-prefix \n show the given source prefix instead of \"a/\"\n --dst-prefix \n show the given destination prefix instead of \"b/\"\n --line-prefix \n prepend an additional prefix to every line of output\n --no-prefix do not show any source or destination prefix\n --inter-hunk-context \n show context between diff hunks up to the specified number of lines\n --output-indicator-new \n specify the character to indicate a new line instead of '+'\n --output-indicator-old \n specify the character to indicate an old line instead of '-'\n --output-indicator-context \n specify the character to indicate a context instead of ' '\n\nDiff rename options\n -B, --break-rewrites[=[/]]\n break complete rewrite changes into pairs of delete and create\n -M, --find-renames[=]\n detect renames\n -D, --irreversible-delete\n omit the preimage for deletes\n -C, --find-copies[=]\n detect copies\n --find-copies-harder use unmodified files as source to find copies\n --no-renames disable rename detection\n --rename-empty use empty blobs as rename source\n --follow continue listing the history of a file beyond renames\n -l prevent rename/copy detection if the number of rename/copy targets exceeds given limit\n\nDiff algorithm options\n --minimal produce the smallest possible diff\n -w, --ignore-all-space\n ignore whitespace when comparing lines\n -b, --ignore-space-change\n ignore changes in amount of whitespace\n --ignore-space-at-eol\n ignore changes in whitespace at EOL\n --ignore-cr-at-eol ignore carrier-return at the end of line\n --ignore-blank-lines ignore changes whose lines are all blank\n -I, --ignore-matching-lines \n ignore changes whose all lines match \n --indent-heuristic heuristic to shift diff hunk boundaries for easy reading\n --patience generate diff using the \"patience diff\" algorithm\n --histogram generate diff using the \"histogram diff\" algorithm\n --diff-algorithm \n choose a diff algorithm\n --anchored generate diff using the \"anchored diff\" algorithm\n --word-diff[=] show word diff, using to delimit changed words\n --word-diff-regex \n use to decide what a word is\n --color-words[=]\n equivalent to --word-diff=color --word-diff-regex=\n --color-moved[=]\n moved lines of code are colored differently\n --color-moved-ws \n how white spaces are ignored in --color-moved\n\nOther diff options\n --relative[=]\n when run from subdir, exclude changes outside and show relative paths\n -a, --text treat all files as text\n -R swap two inputs, reverse the diff\n --exit-code exit with 1 if there were differences, 0 otherwise\n --quiet disable all output of the program\n --ext-diff allow an external diff helper to be executed\n --textconv run external text conversion filters when comparing binary files\n --ignore-submodules[=]\n ignore changes to submodules in the diff generation\n --submodule[=]\n specify how differences in submodules are shown\n --ita-invisible-in-index\n hide 'git add -N' entries from the index\n --ita-visible-in-index\n treat 'git add -N' entries as real in the index\n -S look for differences that change the number of occurrences of the specified string\n -G look for differences that change the number of occurrences of the specified regex\n --pickaxe-all show all changes in the changeset with -S or -G\n --pickaxe-regex treat in -S as extended POSIX regular expression\n -O control the order in which files appear in the output\n --rotate-to show the change in the specified path first\n --skip-to skip the output to the specified path\n --find-object \n look for differences that change the number of occurrences of the specified object\n --diff-filter [(A|C|D|M|R|T|U|X|B)...[*]]\n select files by diff type\n --output Output to a specific file\n\n" + }, + "status": { + "returncode": 128, + "output": "fatal: not a git repository (or any parent up to mount point /root)\nStopping at filesystem boundary (GIT_DISCOVERY_ACROSS_FILESYSTEM not set).\n" + }, + "submodules": { + "returncode": 128, + "output": "fatal: not a git repository (or any parent up to mount point /root)\nStopping at filesystem boundary (GIT_DISCOVERY_ACROSS_FILESYSTEM not set).\n" + } + }, + "/root/autodl-tmp/datasets/eval_textop_omg/TextOp_code_20260927/TextOp": { + "commit": { + "returncode": 128, + "output": "fatal: not a git repository (or any parent up to mount point /root)\nStopping at filesystem boundary (GIT_DISCOVERY_ACROSS_FILESYSTEM not set).\n" + }, + "diff": { + "returncode": 129, + "output": "warning: Not a git repository. Use --no-index to compare two paths outside a working tree\nusage: git diff --no-index [] \n\nDiff output format options\n -p, --patch generate patch\n -s, --no-patch suppress diff output\n -u generate patch\n -U, --unified[=] generate diffs with lines context\n -W, --function-context\n generate diffs with lines context\n --raw generate the diff in raw format\n --patch-with-raw synonym for '-p --raw'\n --patch-with-stat synonym for '-p --stat'\n --numstat machine friendly --stat\n --shortstat output only the last line of --stat\n -X, --dirstat[=...]\n output the distribution of relative amount of changes for each sub-directory\n --cumulative synonym for --dirstat=cumulative\n --dirstat-by-file[=...]\n synonym for --dirstat=files,param1,param2...\n --check warn if changes introduce conflict markers or whitespace errors\n --summary condensed summary such as creations, renames and mode changes\n --name-only show only names of changed files\n --name-status show only names and status of changed files\n --stat[=[,[,]]]\n generate diffstat\n --stat-width generate diffstat with a given width\n --stat-name-width \n generate diffstat with a given name width\n --stat-graph-width \n generate diffstat with a given graph width\n --stat-count generate diffstat with limited lines\n --compact-summary generate compact summary in diffstat\n --binary output a binary diff that can be applied\n --full-index show full pre- and post-image object names on the \"index\" lines\n --color[=] show colored diff\n --ws-error-highlight \n highlight whitespace errors in the 'context', 'old' or 'new' lines in the diff\n -z do not munge pathnames and use NULs as output field terminators in --raw or --numstat\n --abbrev[=] use digits to display object names\n --src-prefix \n show the given source prefix instead of \"a/\"\n --dst-prefix \n show the given destination prefix instead of \"b/\"\n --line-prefix \n prepend an additional prefix to every line of output\n --no-prefix do not show any source or destination prefix\n --inter-hunk-context \n show context between diff hunks up to the specified number of lines\n --output-indicator-new \n specify the character to indicate a new line instead of '+'\n --output-indicator-old \n specify the character to indicate an old line instead of '-'\n --output-indicator-context \n specify the character to indicate a context instead of ' '\n\nDiff rename options\n -B, --break-rewrites[=[/]]\n break complete rewrite changes into pairs of delete and create\n -M, --find-renames[=]\n detect renames\n -D, --irreversible-delete\n omit the preimage for deletes\n -C, --find-copies[=]\n detect copies\n --find-copies-harder use unmodified files as source to find copies\n --no-renames disable rename detection\n --rename-empty use empty blobs as rename source\n --follow continue listing the history of a file beyond renames\n -l prevent rename/copy detection if the number of rename/copy targets exceeds given limit\n\nDiff algorithm options\n --minimal produce the smallest possible diff\n -w, --ignore-all-space\n ignore whitespace when comparing lines\n -b, --ignore-space-change\n ignore changes in amount of whitespace\n --ignore-space-at-eol\n ignore changes in whitespace at EOL\n --ignore-cr-at-eol ignore carrier-return at the end of line\n --ignore-blank-lines ignore changes whose lines are all blank\n -I, --ignore-matching-lines \n ignore changes whose all lines match \n --indent-heuristic heuristic to shift diff hunk boundaries for easy reading\n --patience generate diff using the \"patience diff\" algorithm\n --histogram generate diff using the \"histogram diff\" algorithm\n --diff-algorithm \n choose a diff algorithm\n --anchored generate diff using the \"anchored diff\" algorithm\n --word-diff[=] show word diff, using to delimit changed words\n --word-diff-regex \n use to decide what a word is\n --color-words[=]\n equivalent to --word-diff=color --word-diff-regex=\n --color-moved[=]\n moved lines of code are colored differently\n --color-moved-ws \n how white spaces are ignored in --color-moved\n\nOther diff options\n --relative[=]\n when run from subdir, exclude changes outside and show relative paths\n -a, --text treat all files as text\n -R swap two inputs, reverse the diff\n --exit-code exit with 1 if there were differences, 0 otherwise\n --quiet disable all output of the program\n --ext-diff allow an external diff helper to be executed\n --textconv run external text conversion filters when comparing binary files\n --ignore-submodules[=]\n ignore changes to submodules in the diff generation\n --submodule[=]\n specify how differences in submodules are shown\n --ita-invisible-in-index\n hide 'git add -N' entries from the index\n --ita-visible-in-index\n treat 'git add -N' entries as real in the index\n -S look for differences that change the number of occurrences of the specified string\n -G look for differences that change the number of occurrences of the specified regex\n --pickaxe-all show all changes in the changeset with -S or -G\n --pickaxe-regex treat in -S as extended POSIX regular expression\n -O control the order in which files appear in the output\n --rotate-to show the change in the specified path first\n --skip-to skip the output to the specified path\n --find-object \n look for differences that change the number of occurrences of the specified object\n --diff-filter [(A|C|D|M|R|T|U|X|B)...[*]]\n select files by diff type\n --output Output to a specific file\n\n" + }, + "status": { + "returncode": 128, + "output": "fatal: not a git repository (or any parent up to mount point /root)\nStopping at filesystem boundary (GIT_DISCOVERY_ACROSS_FILESYSTEM not set).\n" + }, + "submodules": { + "returncode": 128, + "output": "fatal: not a git repository (or any parent up to mount point /root)\nStopping at filesystem boundary (GIT_DISCOVERY_ACROSS_FILESYSTEM not set).\n" + } + }, + "/root/autodl-tmp/setup/textop/CLIP": { + "commit": { + "returncode": 0, + "output": "d05afc436d78f1c48dc0dbf8e5980a9d471f35f6\n" + }, + "diff": { + "returncode": 0, + "output": "" + }, + "status": { + "returncode": 0, + "output": "?? build/\n" + }, + "submodules": { + "returncode": 0, + "output": "" + } + } +} \ No newline at end of file diff --git a/handoff-v1/inventory/models.json b/handoff-v1/inventory/models.json new file mode 100644 index 0000000000000000000000000000000000000000..063cfd4c92fc09b14b1bc79a76bbac90a9614363 --- /dev/null +++ b/handoff-v1/inventory/models.json @@ -0,0 +1,197 @@ +{ + "omg-b": { + "run": "/root/autodl-tmp/runs/hfd-omg-english-split20261026", + "checkpoints": [ + { + "path": "/root/autodl-tmp/runs/hfd-omg-english-split20261026/last.ckpt", + "bytes": 1059534816, + "step": 100000, + "keys": [ + "epoch", + "global_step", + "pytorch-lightning_version", + "state_dict", + "loops", + "callbacks", + "optimizer_states", + "lr_schedulers", + "hfd_resume", + "omg_model_architecture" + ] + }, + { + "path": "/root/autodl-tmp/runs/hfd-omg-english-split20261026/best/best.ckpt", + "bytes": 1059548064, + "step": 100000, + "keys": [ + "epoch", + "global_step", + "pytorch-lightning_version", + "state_dict", + "loops", + "callbacks", + "optimizer_states", + "lr_schedulers", + "hfd_resume", + "omg_model_architecture" + ] + } + ] + }, + "omg-l": { + "run": "/root/autodl-tmp/runs/hfd-omg-l-english-split20261026", + "checkpoints": [ + { + "path": "/autodl-fs/data/pangxinlei/models/OMG/hfd-l-english-split20261026/last.ckpt", + "bytes": 4745207121, + "step": 100000, + "keys": [ + "epoch", + "global_step", + "pytorch-lightning_version", + "state_dict", + "loops", + "callbacks", + "optimizer_states", + "lr_schedulers", + "hfd_resume", + "omg_model_architecture" + ] + }, + { + "path": "/autodl-fs/data/pangxinlei/models/OMG/hfd-l-english-split20261026/best/best.ckpt", + "bytes": 4745224529, + "step": 98000, + "keys": [ + "epoch", + "global_step", + "pytorch-lightning_version", + "state_dict", + "loops", + "callbacks", + "optimizer_states", + "lr_schedulers", + "hfd_resume", + "omg_model_architecture" + ] + }, + { + "path": "/autodl-fs/data/pangxinlei/models/OMG/hfd-l-english-split20261026/checkpoints/step-00098000.ckpt", + "bytes": 4745238365, + "step": 98000, + "keys": [ + "epoch", + "global_step", + "pytorch-lightning_version", + "state_dict", + "loops", + "callbacks", + "optimizer_states", + "lr_schedulers", + "hfd_resume", + "omg_model_architecture" + ] + }, + { + "path": "/autodl-fs/data/pangxinlei/models/OMG/hfd-l-english-split20261026/checkpoints/step-00100000.ckpt", + "bytes": 4745238365, + "step": 100000, + "keys": [ + "epoch", + "global_step", + "pytorch-lightning_version", + "state_dict", + "loops", + "callbacks", + "optimizer_states", + "lr_schedulers", + "hfd_resume", + "omg_model_architecture" + ] + } + ] + }, + "omg-xl": { + "run": "/root/autodl-tmp/runs/hfd-omg-xl-english-split20261026", + "checkpoints": [ + { + "path": "/autodl-fs/data/pangxinlei/models/OMG/hfd-xl-english-split20261026/best/best.ckpt", + "bytes": 7555617609, + "step": 88000, + "keys": [ + "epoch", + "global_step", + "pytorch-lightning_version", + "state_dict", + "loops", + "callbacks", + "optimizer_states", + "lr_schedulers", + "hfd_resume", + "omg_model_architecture" + ] + }, + { + "path": "/autodl-fs/data/pangxinlei/models/OMG/hfd-xl-english-split20261026/checkpoints/step-00088000.ckpt", + "bytes": 7555630617, + "step": 88000, + "keys": [ + "epoch", + "global_step", + "pytorch-lightning_version", + "state_dict", + "loops", + "callbacks", + "optimizer_states", + "lr_schedulers", + "hfd_resume", + "omg_model_architecture" + ] + }, + { + "path": "/autodl-fs/data/pangxinlei/models/OMG/hfd-xl-english-split20261026/checkpoints/step-00090000.ckpt", + "bytes": 7555630617, + "step": 90000, + "keys": [ + "epoch", + "global_step", + "pytorch-lightning_version", + "state_dict", + "loops", + "callbacks", + "optimizer_states", + "lr_schedulers", + "hfd_resume", + "omg_model_architecture" + ] + } + ] + }, + "textop": { + "run": "/root/autodl-tmp/runs/hfd-textop-english-split20261026", + "checkpoints": [ + { + "path": "/root/autodl-tmp/runs/hfd-textop-english-split20261026/ckpt_300000.pth", + "bytes": 223658977, + "step": 300000, + "keys": [ + "denoiser", + "optimizer", + "step", + "hfd_resume" + ] + }, + { + "path": "/root/autodl-tmp/runs/hfd-textop-english-split20261026/ckpt_299776.pth", + "bytes": 223658977, + "step": 299776, + "keys": [ + "denoiser", + "optimizer", + "step", + "hfd_resume" + ] + } + ] + } +} \ No newline at end of file diff --git a/handoff-v1/inventory/omg-EXTERNAL_ASSETS.json b/handoff-v1/inventory/omg-EXTERNAL_ASSETS.json new file mode 100644 index 0000000000000000000000000000000000000000..4c5e2d4d3cc488aceb4efdd00d4459abc37cd9fa --- /dev/null +++ b/handoff-v1/inventory/omg-EXTERNAL_ASSETS.json @@ -0,0 +1,198 @@ +{ + "version": 1, + "description": "External file identities only. Model groups use models-root; hfd-metadata uses raw-root. Llama and licensed body model require authorized manual provisioning.", + "files": [ + { + "group": "omg-100m", + "path": "OMG/official/checkpoints/paper/100m/sstep=100000.ckpt", + "bytes": 1933267854, + "sha256": "5e25dc7fd7251ec6df9dda8d3e008ff2540a5c8d1d691c576d4c64f9916eedaa", + "source": { + "url": "https://huggingface.co/THU-MARS/OMG/resolve/06a88419da9f9125d0b5480cc10966741a189358/checkpoints/paper/100m/sstep%3D100000.ckpt", + "repo": "THU-MARS/OMG", + "revision": "06a88419da9f9125d0b5480cc10966741a189358", + "repo_type": "model", + "filename": "checkpoints/paper/100m/sstep=100000.ckpt", + "local_prefix": "OMG/official" + } + }, + { + "group": "omg-300m", + "path": "OMG/official/checkpoints/paper/300m/sstep=055000.ckpt", + "bytes": 5065534010, + "sha256": "df4ce9aeced84fd96a39583c7f6e2c5773b76f942cb43012c6c51dd9eb806806", + "source": { + "url": "https://huggingface.co/THU-MARS/OMG/resolve/06a88419da9f9125d0b5480cc10966741a189358/checkpoints/paper/300m/sstep%3D055000.ckpt", + "repo": "THU-MARS/OMG", + "revision": "06a88419da9f9125d0b5480cc10966741a189358", + "repo_type": "model", + "filename": "checkpoints/paper/300m/sstep=055000.ckpt", + "local_prefix": "OMG/official" + } + }, + { + "group": "omg-500m", + "path": "OMG/official/checkpoints/paper/500m/sstep=050000.ckpt", + "bytes": 8332144642, + "sha256": "f13bc58f07f104fcff67e35e79650def0fec5a6eb62a193fa1ae9f979fb911cf", + "source": { + "url": "https://huggingface.co/THU-MARS/OMG/resolve/06a88419da9f9125d0b5480cc10966741a189358/checkpoints/paper/500m/sstep%3D050000.ckpt", + "repo": "THU-MARS/OMG", + "revision": "06a88419da9f9125d0b5480cc10966741a189358", + "repo_type": "model", + "filename": "checkpoints/paper/500m/sstep=050000.ckpt", + "local_prefix": "OMG/official" + } + }, + { + "group": "omg-50m", + "path": "OMG/official/checkpoints/paper/50m/sstep=090000.ckpt", + "bytes": 1106432846, + "sha256": "389fae1aa2ba31234e0ef621706362a19387ac485dce1f929408e06a3b012b2c", + "source": { + "url": "https://huggingface.co/THU-MARS/OMG/resolve/06a88419da9f9125d0b5480cc10966741a189358/checkpoints/paper/50m/sstep%3D090000.ckpt", + "repo": "THU-MARS/OMG", + "revision": "06a88419da9f9125d0b5480cc10966741a189358", + "repo_type": "model", + "filename": "checkpoints/paper/50m/sstep=090000.ckpt", + "local_prefix": "OMG/official" + } + }, + { + "group": "omg-shared", + "path": "OMG/t5-base-local/README.md", + "bytes": 8477, + "sha256": "c43c23f205b6839a6adb10f0757265f76c1c6f96b7793a4e53582d3a61cbfb23", + "source": { + "url": "https://huggingface.co/google-t5/t5-base/resolve/a9723ea7f1b39c1eae772870f3b547bf6ef7e6c1/README.md", + "repo": "google-t5/t5-base", + "revision": "a9723ea7f1b39c1eae772870f3b547bf6ef7e6c1", + "repo_type": "model", + "filename": "README.md", + "local_prefix": "OMG/t5-base-local" + } + }, + { + "group": "omg-shared", + "path": "OMG/t5-base-local/config.json", + "bytes": 1208, + "sha256": "46dd7cb62d29c81fb551e0ef1ea274c24a46ba441eeb948897706252933df033", + "source": { + "url": "https://huggingface.co/google-t5/t5-base/resolve/a9723ea7f1b39c1eae772870f3b547bf6ef7e6c1/config.json", + "repo": "google-t5/t5-base", + "revision": "a9723ea7f1b39c1eae772870f3b547bf6ef7e6c1", + "repo_type": "model", + "filename": "config.json", + "local_prefix": "OMG/t5-base-local" + } + }, + { + "group": "omg-shared", + "path": "OMG/t5-base-local/generation_config.json", + "bytes": 147, + "sha256": "f5a1c7e2be8092018d8835128987edf0111637dd98e90599cc80310fef75d95a", + "source": { + "url": "https://huggingface.co/google-t5/t5-base/resolve/a9723ea7f1b39c1eae772870f3b547bf6ef7e6c1/generation_config.json", + "repo": "google-t5/t5-base", + "revision": "a9723ea7f1b39c1eae772870f3b547bf6ef7e6c1", + "repo_type": "model", + "filename": "generation_config.json", + "local_prefix": "OMG/t5-base-local" + } + }, + { + "group": "omg-shared", + "path": "OMG/t5-base-local/model.safetensors", + "bytes": 891646390, + "sha256": "a90903540cc02cbeb7ff9f823f1a80eb778c7e22426a0e620b01c77a5ec8f5b4", + "source": { + "url": "https://huggingface.co/google-t5/t5-base/resolve/a9723ea7f1b39c1eae772870f3b547bf6ef7e6c1/model.safetensors", + "repo": "google-t5/t5-base", + "revision": "a9723ea7f1b39c1eae772870f3b547bf6ef7e6c1", + "repo_type": "model", + "filename": "model.safetensors", + "local_prefix": "OMG/t5-base-local" + } + }, + { + "group": "omg-shared", + "path": "OMG/t5-base-local/spiece.model", + "bytes": 791656, + "sha256": "d60acb128cf7b7f2536e8f38a5b18a05535c9e14c7a355904270e15b0945ea86", + "source": { + "url": "https://huggingface.co/google-t5/t5-base/resolve/a9723ea7f1b39c1eae772870f3b547bf6ef7e6c1/spiece.model", + "repo": "google-t5/t5-base", + "revision": "a9723ea7f1b39c1eae772870f3b547bf6ef7e6c1", + "repo_type": "model", + "filename": "spiece.model", + "local_prefix": "OMG/t5-base-local" + } + }, + { + "group": "omg-shared", + "path": "OMG/t5-base-local/tokenizer.json", + "bytes": 1389353, + "sha256": "d2acde0d8d71dd30a711834b07781b9c89feaac33fd332f60507699282740066", + "source": { + "url": "https://huggingface.co/google-t5/t5-base/resolve/a9723ea7f1b39c1eae772870f3b547bf6ef7e6c1/tokenizer.json", + "repo": "google-t5/t5-base", + "revision": "a9723ea7f1b39c1eae772870f3b547bf6ef7e6c1", + "repo_type": "model", + "filename": "tokenizer.json", + "local_prefix": "OMG/t5-base-local" + } + }, + { + "group": "hfd-metadata", + "path": "MANIFEST.json", + "bytes": 283191, + "sha256": "a7d91a52dabd142d73a8e93ed86403a5763385b3115ff2217777bf82d2bb492e", + "source": { + "repo": "Tinghong-Ye22/hfd_data", + "revision": "93ba4a7a9c66b3e1b7041e5f4ea038ef2418b948", + "repo_type": "dataset", + "filename": "MANIFEST.json", + "local_prefix": "" + } + }, + { + "group": "hfd-metadata", + "path": "PATH_MAP.json", + "bytes": 1151, + "sha256": "0e89066ad6758bfbd2811ba497e541aaf2d8f51aeeda3959379b765930f53098", + "source": { + "repo": "Tinghong-Ye22/hfd_data", + "revision": "93ba4a7a9c66b3e1b7041e5f4ea038ef2418b948", + "repo_type": "dataset", + "filename": "PATH_MAP.json", + "local_prefix": "" + } + }, + { + "group": "hfd-metadata", + "path": "index.sqlite", + "bytes": 454651904, + "sha256": "e41b7ba92d5ccd57092b7c3a190828cc0e8a545c79a8bce575633a68f2bb19bf", + "source": { + "repo": "Tinghong-Ye22/hfd_data", + "revision": "93ba4a7a9c66b3e1b7041e5f4ea038ef2418b948", + "repo_type": "dataset", + "filename": "index.sqlite", + "local_prefix": "" + } + }, + { + "group": "hfd-metadata", + "path": "metadata/seed_g1_smplx_pairs.parquet", + "bytes": 26580927, + "sha256": "63fcd6eb1deb1c270d93a7944f9f2fe95a4433c3bf557cb4ca8671bd18197b10", + "source": { + "repo": "Tinghong-Ye22/hfd_data", + "revision": "93ba4a7a9c66b3e1b7041e5f4ea038ef2418b948", + "repo_type": "dataset", + "filename": "metadata/seed_g1_smplx_pairs.parquet", + "local_prefix": "" + } + } + ] +} \ No newline at end of file diff --git a/handoff-v1/inventory/omg-b-best.ckpt.json b/handoff-v1/inventory/omg-b-best.ckpt.json new file mode 100644 index 0000000000000000000000000000000000000000..a352fac1a18d149fdea40f794cd5eeca7f68f10c --- /dev/null +++ b/handoff-v1/inventory/omg-b-best.ckpt.json @@ -0,0 +1,2941 @@ +{ + "epoch": 1, + "global_step": 100000, + "pytorch-lightning_version": "2.6.6", + "state_dict": { + "denoiser.time_embed.proj.0.weight": { + "shape": [ + 512, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.time_embed.proj.0.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.time_embed.proj.2.weight": { + "shape": [ + 512, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.time_embed.proj.2.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.input_proj.weight": { + "shape": [ + 512, + 637 + ], + "dtype": "torch.float32" + }, + "denoiser.input_proj.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.text_proj.weight": { + "shape": [ + 512, + 768 + ], + "dtype": "torch.float32" + }, + "denoiser.text_proj.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.extra_proj.0.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.extra_proj.0.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.extra_proj.1.weight": { + "shape": [ + 512, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.extra_proj.1.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.norm_cross.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.norm_cross.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.cross_attn.in_proj_weight": { + "shape": [ + 1536, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.cross_attn.in_proj_bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.cross_attn.out_proj.weight": { + "shape": [ + 512, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.cross_attn.out_proj.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.norm_self.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.norm_self.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.self_attn.qkv.weight": { + "shape": [ + 1536, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.self_attn.qkv.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.self_attn.out_proj.weight": { + "shape": [ + 512, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.self_attn.out_proj.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.norm_ff.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.norm_ff.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.ff.0.weight": { + "shape": [ + 2048, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.ff.0.bias": { + "shape": [ + 2048 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.ff.3.weight": { + "shape": [ + 512, + 2048 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.ff.3.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.audio_film.0.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.audio_film.0.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.audio_film.1.weight": { + "shape": [ + 1024, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.audio_film.1.bias": { + "shape": [ + 1024 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.human_motion_film.0.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.human_motion_film.0.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.human_motion_film.1.weight": { + "shape": [ + 1024, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.human_motion_film.1.bias": { + "shape": [ + 1024 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.norm_cross.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.norm_cross.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.cross_attn.in_proj_weight": { + "shape": [ + 1536, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.cross_attn.in_proj_bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.cross_attn.out_proj.weight": { + "shape": [ + 512, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.cross_attn.out_proj.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.norm_self.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.norm_self.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.self_attn.qkv.weight": { + "shape": [ + 1536, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.self_attn.qkv.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.self_attn.out_proj.weight": { + "shape": [ + 512, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.self_attn.out_proj.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.norm_ff.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.norm_ff.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.ff.0.weight": { + "shape": [ + 2048, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.ff.0.bias": { + "shape": [ + 2048 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.ff.3.weight": { + "shape": [ + 512, + 2048 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.ff.3.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.audio_film.0.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.audio_film.0.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.audio_film.1.weight": { + "shape": [ + 1024, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.audio_film.1.bias": { + "shape": [ + 1024 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.human_motion_film.0.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.human_motion_film.0.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.human_motion_film.1.weight": { + "shape": [ + 1024, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.human_motion_film.1.bias": { + "shape": [ + 1024 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.norm_cross.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.norm_cross.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.cross_attn.in_proj_weight": { + "shape": [ + 1536, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.cross_attn.in_proj_bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.cross_attn.out_proj.weight": { + "shape": [ + 512, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.cross_attn.out_proj.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.norm_self.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.norm_self.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.self_attn.qkv.weight": { + "shape": [ + 1536, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.self_attn.qkv.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.self_attn.out_proj.weight": { + "shape": [ + 512, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.self_attn.out_proj.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.norm_ff.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.norm_ff.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.ff.0.weight": { + "shape": [ + 2048, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.ff.0.bias": { + "shape": [ + 2048 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.ff.3.weight": { + "shape": [ + 512, + 2048 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.ff.3.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.audio_film.0.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.audio_film.0.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.audio_film.1.weight": { + "shape": [ + 1024, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.audio_film.1.bias": { + "shape": [ + 1024 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.human_motion_film.0.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.human_motion_film.0.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.human_motion_film.1.weight": { + "shape": [ + 1024, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.human_motion_film.1.bias": { + "shape": [ + 1024 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.norm_cross.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.norm_cross.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.cross_attn.in_proj_weight": { + "shape": [ + 1536, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.cross_attn.in_proj_bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.cross_attn.out_proj.weight": { + "shape": [ + 512, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.cross_attn.out_proj.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.norm_self.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.norm_self.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.self_attn.qkv.weight": { + "shape": [ + 1536, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.self_attn.qkv.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.self_attn.out_proj.weight": { + "shape": [ + 512, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.self_attn.out_proj.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.norm_ff.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.norm_ff.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.ff.0.weight": { + "shape": [ + 2048, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.ff.0.bias": { + "shape": [ + 2048 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.ff.3.weight": { + "shape": [ + 512, + 2048 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.ff.3.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.audio_film.0.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.audio_film.0.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.audio_film.1.weight": { + "shape": [ + 1024, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.audio_film.1.bias": { + "shape": [ + 1024 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.human_motion_film.0.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.human_motion_film.0.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.human_motion_film.1.weight": { + "shape": [ + 1024, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.human_motion_film.1.bias": { + "shape": [ + 1024 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.norm_cross.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.norm_cross.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.cross_attn.in_proj_weight": { + "shape": [ + 1536, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.cross_attn.in_proj_bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.cross_attn.out_proj.weight": { + "shape": [ + 512, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.cross_attn.out_proj.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.norm_self.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.norm_self.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.self_attn.qkv.weight": { + "shape": [ + 1536, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.self_attn.qkv.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.self_attn.out_proj.weight": { + "shape": [ + 512, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.self_attn.out_proj.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.norm_ff.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.norm_ff.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.ff.0.weight": { + "shape": [ + 2048, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.ff.0.bias": { + "shape": [ + 2048 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.ff.3.weight": { + "shape": [ + 512, + 2048 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.ff.3.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.audio_film.0.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.audio_film.0.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.audio_film.1.weight": { + "shape": [ + 1024, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.audio_film.1.bias": { + "shape": [ + 1024 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.human_motion_film.0.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.human_motion_film.0.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.human_motion_film.1.weight": { + "shape": [ + 1024, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.human_motion_film.1.bias": { + "shape": [ + 1024 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.norm_cross.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.norm_cross.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.cross_attn.in_proj_weight": { + "shape": [ + 1536, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.cross_attn.in_proj_bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.cross_attn.out_proj.weight": { + "shape": [ + 512, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.cross_attn.out_proj.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.norm_self.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.norm_self.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.self_attn.qkv.weight": { + "shape": [ + 1536, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.self_attn.qkv.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.self_attn.out_proj.weight": { + "shape": [ + 512, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.self_attn.out_proj.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.norm_ff.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.norm_ff.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.ff.0.weight": { + "shape": [ + 2048, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.ff.0.bias": { + "shape": [ + 2048 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.ff.3.weight": { + "shape": [ + 512, + 2048 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.ff.3.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.audio_film.0.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.audio_film.0.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.audio_film.1.weight": { + "shape": [ + 1024, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.audio_film.1.bias": { + "shape": [ + 1024 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.human_motion_film.0.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.human_motion_film.0.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.human_motion_film.1.weight": { + "shape": [ + 1024, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.human_motion_film.1.bias": { + "shape": [ + 1024 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.norm_cross.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.norm_cross.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.cross_attn.in_proj_weight": { + "shape": [ + 1536, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.cross_attn.in_proj_bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.cross_attn.out_proj.weight": { + "shape": [ + 512, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.cross_attn.out_proj.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.norm_self.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.norm_self.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.self_attn.qkv.weight": { + "shape": [ + 1536, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.self_attn.qkv.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.self_attn.out_proj.weight": { + "shape": [ + 512, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.self_attn.out_proj.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.norm_ff.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.norm_ff.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.ff.0.weight": { + "shape": [ + 2048, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.ff.0.bias": { + "shape": [ + 2048 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.ff.3.weight": { + "shape": [ + 512, + 2048 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.ff.3.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.audio_film.0.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.audio_film.0.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.audio_film.1.weight": { + "shape": [ + 1024, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.audio_film.1.bias": { + "shape": [ + 1024 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.human_motion_film.0.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.human_motion_film.0.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.human_motion_film.1.weight": { + "shape": [ + 1024, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.human_motion_film.1.bias": { + "shape": [ + 1024 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.norm_cross.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.norm_cross.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.cross_attn.in_proj_weight": { + "shape": [ + 1536, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.cross_attn.in_proj_bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.cross_attn.out_proj.weight": { + "shape": [ + 512, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.cross_attn.out_proj.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.norm_self.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.norm_self.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.self_attn.qkv.weight": { + "shape": [ + 1536, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.self_attn.qkv.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.self_attn.out_proj.weight": { + "shape": [ + 512, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.self_attn.out_proj.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.norm_ff.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.norm_ff.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.ff.0.weight": { + "shape": [ + 2048, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.ff.0.bias": { + "shape": [ + 2048 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.ff.3.weight": { + "shape": [ + 512, + 2048 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.ff.3.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.audio_film.0.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.audio_film.0.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.audio_film.1.weight": { + "shape": [ + 1024, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.audio_film.1.bias": { + "shape": [ + 1024 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.human_motion_film.0.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.human_motion_film.0.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.human_motion_film.1.weight": { + "shape": [ + 1024, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.human_motion_film.1.bias": { + "shape": [ + 1024 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.norm_cross.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.norm_cross.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.cross_attn.in_proj_weight": { + "shape": [ + 1536, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.cross_attn.in_proj_bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.cross_attn.out_proj.weight": { + "shape": [ + 512, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.cross_attn.out_proj.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.norm_self.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.norm_self.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.self_attn.qkv.weight": { + "shape": [ + 1536, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.self_attn.qkv.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.self_attn.out_proj.weight": { + "shape": [ + 512, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.self_attn.out_proj.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.norm_ff.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.norm_ff.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.ff.0.weight": { + "shape": [ + 2048, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.ff.0.bias": { + "shape": [ + 2048 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.ff.3.weight": { + "shape": [ + 512, + 2048 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.ff.3.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.audio_film.0.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.audio_film.0.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.audio_film.1.weight": { + "shape": [ + 1024, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.audio_film.1.bias": { + "shape": [ + 1024 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.human_motion_film.0.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.human_motion_film.0.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.human_motion_film.1.weight": { + "shape": [ + 1024, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.human_motion_film.1.bias": { + "shape": [ + 1024 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.norm_cross.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.norm_cross.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.cross_attn.in_proj_weight": { + "shape": [ + 1536, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.cross_attn.in_proj_bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.cross_attn.out_proj.weight": { + "shape": [ + 512, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.cross_attn.out_proj.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.norm_self.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.norm_self.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.self_attn.qkv.weight": { + "shape": [ + 1536, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.self_attn.qkv.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.self_attn.out_proj.weight": { + "shape": [ + 512, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.self_attn.out_proj.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.norm_ff.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.norm_ff.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.ff.0.weight": { + "shape": [ + 2048, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.ff.0.bias": { + "shape": [ + 2048 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.ff.3.weight": { + "shape": [ + 512, + 2048 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.ff.3.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.audio_film.0.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.audio_film.0.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.audio_film.1.weight": { + "shape": [ + 1024, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.audio_film.1.bias": { + "shape": [ + 1024 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.human_motion_film.0.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.human_motion_film.0.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.human_motion_film.1.weight": { + "shape": [ + 1024, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.human_motion_film.1.bias": { + "shape": [ + 1024 + ], + "dtype": "torch.float32" + }, + "denoiser.norm.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.norm.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.output.weight": { + "shape": [ + 125, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.output.bias": { + "shape": [ + 125 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.shared.weight": { + "shape": [ + 32128, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.embed_tokens.weight": { + "shape": [ + 32128, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.0.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.0.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.0.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.0.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.0.layer.0.SelfAttention.relative_attention_bias.weight": { + "shape": [ + 32, + 12 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.0.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.0.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.0.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.0.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.1.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.1.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.1.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.1.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.1.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.1.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.1.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.1.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.2.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.2.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.2.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.2.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.2.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.2.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.2.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.2.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.3.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.3.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.3.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.3.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.3.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.3.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.3.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.3.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.4.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.4.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.4.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.4.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.4.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.4.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.4.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.4.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.5.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.5.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.5.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.5.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.5.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.5.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.5.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.5.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.6.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.6.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.6.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.6.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.6.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.6.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.6.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.6.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.7.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.7.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.7.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.7.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.7.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.7.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.7.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.7.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.8.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.8.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.8.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.8.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.8.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.8.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.8.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.8.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.9.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.9.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.9.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.9.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.9.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.9.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.9.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.9.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.10.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.10.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.10.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.10.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.10.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.10.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.10.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.10.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.11.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.11.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.11.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.11.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.11.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.11.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.11.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.11.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.final_layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "history_projector.0.weight": { + "shape": [ + 125 + ], + "dtype": "torch.float32" + }, + "history_projector.0.bias": { + "shape": [ + 125 + ], + "dtype": "torch.float32" + }, + "history_projector.1.weight": { + "shape": [ + 512, + 125 + ], + "dtype": "torch.float32" + }, + "history_projector.1.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "history_projector.3.weight": { + "shape": [ + 512, + 512 + ], + "dtype": "torch.float32" + }, + "history_projector.3.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "audio_embedder.0.weight": { + "shape": [ + 35 + ], + "dtype": "torch.float32" + }, + "audio_embedder.0.bias": { + "shape": [ + 35 + ], + "dtype": "torch.float32" + }, + "audio_embedder.1.weight": { + "shape": [ + 1024, + 35 + ], + "dtype": "torch.float32" + }, + "audio_embedder.1.bias": { + "shape": [ + 1024 + ], + "dtype": "torch.float32" + }, + "audio_embedder.4.weight": { + "shape": [ + 512, + 1024 + ], + "dtype": "torch.float32" + }, + "audio_embedder.4.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "human_motion_embedder.0.weight": { + "shape": [ + 66 + ], + "dtype": "torch.float32" + }, + "human_motion_embedder.0.bias": { + "shape": [ + 66 + ], + "dtype": "torch.float32" + }, + "human_motion_embedder.1.weight": { + "shape": [ + 1024, + 66 + ], + "dtype": "torch.float32" + }, + "human_motion_embedder.1.bias": { + "shape": [ + 1024 + ], + "dtype": "torch.float32" + }, + "human_motion_embedder.4.weight": { + "shape": [ + 512, + 1024 + ], + "dtype": "torch.float32" + }, + "human_motion_embedder.4.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + } + }, + "loops": { + "fit_loop": { + "state_dict": { + "keys": [], + "count": 0 + }, + "epoch_loop.state_dict": { + "keys": [ + "_batches_that_stepped" + ], + "count": 1 + }, + "epoch_loop.batch_progress": { + "keys": [ + "total", + "current", + "is_last_batch" + ], + "count": 3 + }, + "epoch_loop.scheduler_progress": { + "keys": [ + "total", + "current" + ], + "count": 2 + }, + "epoch_loop.automatic_optimization.state_dict": { + "keys": [], + "count": 0 + }, + "epoch_loop.automatic_optimization.optim_progress": { + "keys": [ + "optimizer" + ], + "count": 1 + }, + "epoch_loop.manual_optimization.state_dict": { + "keys": [], + "count": 0 + }, + "epoch_loop.manual_optimization.optim_step_progress": { + "keys": [ + "total", + "current" + ], + "count": 2 + }, + "epoch_loop.val_loop.state_dict": { + "keys": [], + "count": 0 + }, + "epoch_loop.val_loop.batch_progress": { + "keys": [ + "total", + "current", + "is_last_batch" + ], + "count": 3 + }, + "epoch_progress": { + "keys": [ + "total", + "current" + ], + "count": 2 + } + }, + "validate_loop": { + "state_dict": { + "keys": [], + "count": 0 + }, + "batch_progress": { + "keys": [ + "total", + "current", + "is_last_batch" + ], + "count": 3 + } + }, + "test_loop": { + "state_dict": { + "keys": [], + "count": 0 + }, + "batch_progress": { + "keys": [ + "total", + "current", + "is_last_batch" + ], + "count": 3 + } + }, + "predict_loop": { + "state_dict": { + "keys": [], + "count": 0 + }, + "batch_progress": { + "keys": [ + "total", + "current" + ], + "count": 2 + } + } + }, + "callbacks": { + "ModelCheckpoint{'monitor': 'step', 'mode': 'max', 'every_n_train_steps': 256, 'every_n_epochs': 0, 'train_time_interval': None}": { + "monitor": "step", + "best_model_score": { + "shape": [], + "dtype": "torch.int64" + }, + "best_model_path": "/root/autodl-tmp/runs/hfd-omg-english-split20261026/checkpoints/step-00099840.ckpt", + "current_score": { + "shape": [], + "dtype": "torch.int64" + }, + "dirpath": "/root/autodl-tmp/runs/hfd-omg-english-split20261026/checkpoints", + "best_k_models": { + "keys": [ + "/root/autodl-tmp/runs/hfd-omg-english-split20261026/checkpoints/step-00099328.ckpt", + "/root/autodl-tmp/runs/hfd-omg-english-split20261026/checkpoints/step-00099584.ckpt", + "/root/autodl-tmp/runs/hfd-omg-english-split20261026/checkpoints/step-00099840.ckpt" + ], + "count": 3 + }, + "kth_best_model_path": "/root/autodl-tmp/runs/hfd-omg-english-split20261026/checkpoints/step-00099328.ckpt", + "kth_value": { + "shape": [], + "dtype": "torch.int64" + }, + "last_model_path": "/root/autodl-tmp/runs/hfd-omg-english-split20261026/checkpoints/last-v1.ckpt" + }, + "ModelCheckpoint{'monitor': 'val/loss', 'mode': 'min', 'every_n_train_steps': 0, 'every_n_epochs': 1, 'train_time_interval': None}": { + "monitor": "val/loss", + "best_model_score": { + "shape": [], + "dtype": "torch.float32" + }, + "best_model_path": "/root/autodl-tmp/runs/hfd-omg-english-split20261026/best/best.ckpt", + "current_score": { + "shape": [], + "dtype": "torch.float32" + }, + "dirpath": "/root/autodl-tmp/runs/hfd-omg-english-split20261026/best", + "best_k_models": { + "keys": [ + "/root/autodl-tmp/runs/hfd-omg-english-split20261026/best/best.ckpt" + ], + "count": 1 + }, + "kth_best_model_path": "/root/autodl-tmp/runs/hfd-omg-english-split20261026/best/best.ckpt", + "kth_value": { + "shape": [], + "dtype": "torch.float32" + }, + "last_model_path": "" + } + }, + "optimizer_states": { + "count": 1, + "sample": [ + { + "state": { + "keys": [ + "0", + "1", + "2", + "3", + "4", + "5", + "6", + "7", + "8", + "9", + "10", + "11", + "12", + "13", + "14", + "15", + "16", + "17", + "18", + "19", + "20", + "21", + "22", + "23", + "24", + "25", + "26", + "27", + "28", + "29" + ], + "count": 248 + }, + "param_groups": { + "count": 1, + "sample": [ + { + "keys": [ + "lr", + "betas", + "eps", + "weight_decay", + "amsgrad", + "maximize", + "foreach", + "capturable", + "differentiable", + "fused", + "decoupled_weight_decay", + "initial_lr", + "params" + ], + "count": 13 + } + ] + } + } + ] + }, + "lr_schedulers": { + "count": 1, + "sample": [ + { + "_milestones": { + "count": 1, + "sample": [ + 2000 + ] + }, + "last_epoch": 100000, + "_last_lr": { + "count": 1, + "sample": [ + 1e-06 + ] + }, + "_schedulers": { + "count": 2, + "sample": [ + { + "keys": [ + "start_factor", + "end_factor", + "total_iters", + "base_lrs", + "last_epoch", + "_step_count", + "_get_lr_called_within_step", + "_last_lr" + ], + "count": 8 + }, + { + "keys": [ + "T_max", + "eta_min", + "base_lrs", + "last_epoch", + "_step_count", + "_get_lr_called_within_step", + "_last_lr" + ], + "count": 7 + } + ] + } + } + ] + }, + "hfd_resume": { + "version": 1, + "rng": { + "python": { + "count": 3, + "sample": [ + 3, + { + "count": 625, + "sample": [ + 2147483648, + 3564348608 + ] + } + ] + }, + "numpy": { + "count": 5, + "sample": [ + "MT19937", + { + "count": 624, + "sample": [ + 42, + 3107752595 + ] + } + ] + }, + "torch": { + "shape": [ + 5056 + ], + "dtype": "torch.uint8" + }, + "cuda": { + "count": 1, + "sample": [ + { + "shape": [ + 16 + ], + "dtype": "torch.uint8" + } + ] + } + }, + "train_pool": { + "step": 100000, + "slots": { + "count": 512, + "sample": [ + 248743, + 347419 + ] + }, + "batch_size": 512, + "refresh": 256, + "rng": { + "keys": [ + "bit_generator", + "state", + "has_uint32", + "uinteger" + ], + "count": 4 + }, + "identity": { + "keys": [ + "sessions", + "native_source_sha256", + "asr", + "recipe_sha256", + "foreground_language", + "ordered_session_selection_sha256", + "final_release" + ], + "count": 7 + } + }, + "val_pool": { + "step": 16, + "slots": { + "count": 512, + "sample": [ + 6295, + 7193 + ] + }, + "batch_size": 512, + "refresh": 256, + "rng": { + "keys": [ + "bit_generator", + "state", + "has_uint32", + "uinteger" + ], + "count": 4 + }, + "identity": { + "keys": [ + "sessions", + "native_source_sha256", + "asr", + "recipe_sha256", + "foreground_language", + "ordered_session_selection_sha256", + "final_release" + ], + "count": 7 + } + } + }, + "omg_model_architecture": { + "format": "omg.model_architecture", + "version": 1, + "denoiser_type": "omg.generation.denoisers.transformer.MotionTransformerDenoiser", + "frame_cond_injection": "per_layer_film", + "attention": { + "rotary_self_attention_qk_norm": true, + "cross_attention_qk_norm": true + } + } +} \ No newline at end of file diff --git a/handoff-v1/inventory/omg-b-final-resume-historical-source.json b/handoff-v1/inventory/omg-b-final-resume-historical-source.json new file mode 100644 index 0000000000000000000000000000000000000000..3ffaa7c5967205b0f2004a814184fb55bc6196c5 --- /dev/null +++ b/handoff-v1/inventory/omg-b-final-resume-historical-source.json @@ -0,0 +1,63 @@ +{ + "hfd_bridge/omg_checks.py": { + "expected": "1c80361412892538400cc225ab28aa03d5357ccab7a53e3a6d46deb5da9d32e8", + "recovered_commit": "ea269e3716e289e64fa3cec17f1366b57e3fab9b", + "used_by_this_model": true + }, + "hfd_bridge/resume.py": { + "expected": "be1b665778f3c3bc137f4242b0ac4d1f587903357954725ed509e712dac0ab8f", + "recovered_commit": "exact bytes recovered from prior stored local diff; SHA matches recorded final resume", + "used_by_this_model": true, + "recovered_sha256": "be1b665778f3c3bc137f4242b0ac4d1f587903357954725ed509e712dac0ab8f" + }, + "hfd_bridge/genmo_training.py": { + "expected": "8276d4f266b44e2c112e14021eb29c442753f2de07abf29e9adb321226dc87d8", + "recovered_commit": "ea269e3716e289e64fa3cec17f1366b57e3fab9b", + "used_by_this_model": false + }, + "hfd_bridge/genmo_prefetch.py": { + "expected": "3375352f7668dfd5c76ff629853aa8e4d0bcd2e2294a6e96932280f2eafa68e4", + "recovered_commit": "ea269e3716e289e64fa3cec17f1366b57e3fab9b", + "used_by_this_model": false + }, + "hfd_bridge/hymotion_model.py": { + "expected": "db0968384b87929c6495b87647a37e92e13729a94a34f6a5347f4d23ed05be5c", + "recovered_commit": null, + "used_by_this_model": false + }, + "hfd_bridge/hymotion_data.py": { + "expected": "06170526a18b60fa486f084864e2e65f2608274ed967ef367b29de269f8c9147", + "recovered_commit": null, + "used_by_this_model": false + }, + "hfd_bridge/hymotion_rollout.py": { + "expected": "a7fefe8d72f9e3614103632d68d1d63ce3ee072650739683131f5315d0a41181", + "recovered_commit": null, + "used_by_this_model": false + }, + "hfd_bridge/hymotion_validation.py": { + "expected": "cf6fd4b14fdbafc65a8a65abaade620b43d5b1f95e7a7a6194a21247f2ad4127", + "recovered_commit": null, + "used_by_this_model": false + }, + "hfd_bridge/hymotion_training.py": { + "expected": "67f40e79cc43950f6876f8c0c2c67fe1f606f1b67e692a7dd239457fa4b96a2c", + "recovered_commit": null, + "used_by_this_model": false + }, + "hfd_bridge/hymotion_clock.py": { + "expected": "1eaf969e9198f6902759457488ff8c80fc20c973b72388443b3ef0af7c8282a2", + "recovered_commit": null, + "used_by_this_model": false + }, + "hfd_bridge/hymotion_recovery.py": { + "expected": "e47b841f85dfc19b37a646a6cf6e865dd2642718073a3419a97195f16024c379", + "recovered_commit": null, + "used_by_this_model": false + }, + "train.py": { + "expected": "0724b3db6d374259db9f6378e9b663ef8438e5b969093f3b4d1e46ee56456ddc", + "recovered_commit": "ea269e3716e289e64fa3cec17f1366b57e3fab9b", + "used_by_this_model": true + } +} \ No newline at end of file diff --git a/handoff-v1/inventory/omg-b-historical-source.json b/handoff-v1/inventory/omg-b-historical-source.json new file mode 100644 index 0000000000000000000000000000000000000000..388618b46ba331266973b367ea702d5bd5262095 --- /dev/null +++ b/handoff-v1/inventory/omg-b-historical-source.json @@ -0,0 +1,82 @@ +{ + "hfd_bridge/storage.py": { + "expected": "8b553d204c19ddad7cec223c96f42146865266bc578e7207ad80eb9c7c8f396d", + "recovered_commit": null, + "status": "historical_bytes_missing_current_snapshot_provided" + }, + "hfd_bridge/omg_checks.py": { + "expected": "1c80361412892538400cc225ab28aa03d5357ccab7a53e3a6d46deb5da9d32e8", + "recovered_commit": "ea269e3716e289e64fa3cec17f1366b57e3fab9b", + "status": "exact_historical_bytes" + }, + "hfd_bridge/ondemand.py": { + "expected": "ddf6c76a6b3d736de9dbf5e27c9d8d6e9384156cee68fc96bba0ce08237aa7c5", + "recovered_commit": null, + "status": "historical_bytes_missing_current_snapshot_provided" + }, + "hfd_bridge/human_sessions.py": { + "expected": "8b93bbe9450fef5eabd05c827b596a41dfa4748aeaaf93a63849877f5e8df38e", + "recovered_commit": "ea269e3716e289e64fa3cec17f1366b57e3fab9b", + "status": "exact_historical_bytes" + }, + "hfd_bridge/resume.py": { + "expected": "e8f9c06323415cbdf011272efa4504f12e0c44079d4648a92277f27a5e5d0cc3", + "recovered_commit": "ea269e3716e289e64fa3cec17f1366b57e3fab9b", + "status": "exact_historical_bytes" + }, + "hfd_bridge/kimodo.py": { + "expected": "196f8ff99caec99470b04adb5e24efac83b5f4239aa2fdc3083b4438fa4cf460", + "recovered_commit": "ea269e3716e289e64fa3cec17f1366b57e3fab9b", + "status": "exact_historical_bytes" + }, + "hfd_bridge/genmo_training.py": { + "expected": "8276d4f266b44e2c112e14021eb29c442753f2de07abf29e9adb321226dc87d8", + "recovered_commit": "ea269e3716e289e64fa3cec17f1366b57e3fab9b", + "status": "exact_historical_bytes" + }, + "hfd_bridge/genmo_prefetch.py": { + "expected": "3375352f7668dfd5c76ff629853aa8e4d0bcd2e2294a6e96932280f2eafa68e4", + "recovered_commit": "ea269e3716e289e64fa3cec17f1366b57e3fab9b", + "status": "exact_historical_bytes" + }, + "hfd_bridge/hymotion_model.py": { + "expected": "db0968384b87929c6495b87647a37e92e13729a94a34f6a5347f4d23ed05be5c", + "recovered_commit": null, + "status": "historical_bytes_missing_current_snapshot_provided" + }, + "hfd_bridge/hymotion_data.py": { + "expected": "06170526a18b60fa486f084864e2e65f2608274ed967ef367b29de269f8c9147", + "recovered_commit": null, + "status": "historical_bytes_missing_current_snapshot_provided" + }, + "hfd_bridge/hymotion_rollout.py": { + "expected": "a7fefe8d72f9e3614103632d68d1d63ce3ee072650739683131f5315d0a41181", + "recovered_commit": null, + "status": "historical_bytes_missing_current_snapshot_provided" + }, + "hfd_bridge/hymotion_validation.py": { + "expected": "cf6fd4b14fdbafc65a8a65abaade620b43d5b1f95e7a7a6194a21247f2ad4127", + "recovered_commit": null, + "status": "historical_bytes_missing_current_snapshot_provided" + }, + "hfd_bridge/hymotion_training.py": { + "expected": "67f40e79cc43950f6876f8c0c2c67fe1f606f1b67e692a7dd239457fa4b96a2c", + "recovered_commit": null, + "status": "historical_bytes_missing_current_snapshot_provided" + }, + "hfd_bridge/hymotion_clock.py": { + "expected": "1eaf969e9198f6902759457488ff8c80fc20c973b72388443b3ef0af7c8282a2", + "recovered_commit": null, + "status": "historical_bytes_missing_current_snapshot_provided" + }, + "hfd_bridge/hymotion_recovery.py": { + "expected": "e47b841f85dfc19b37a646a6cf6e865dd2642718073a3419a97195f16024c379", + "recovered_commit": null, + "status": "historical_bytes_missing_current_snapshot_provided" + }, + "train.py": { + "expected": "0724b3db6d374259db9f6378e9b663ef8438e5b969093f3b4d1e46ee56456ddc", + "recovered_commit": "ea269e3716e289e64fa3cec17f1366b57e3fab9b", + "status": "exact_historical_bytes" + } +} \ No newline at end of file diff --git a/handoff-v1/inventory/omg-b-last.ckpt.json b/handoff-v1/inventory/omg-b-last.ckpt.json new file mode 100644 index 0000000000000000000000000000000000000000..4672e619594a7bd6a358f1c10ca75f8b960350d6 --- /dev/null +++ b/handoff-v1/inventory/omg-b-last.ckpt.json @@ -0,0 +1,2941 @@ +{ + "epoch": 2, + "global_step": 100000, + "pytorch-lightning_version": "2.6.6", + "state_dict": { + "denoiser.time_embed.proj.0.weight": { + "shape": [ + 512, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.time_embed.proj.0.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.time_embed.proj.2.weight": { + "shape": [ + 512, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.time_embed.proj.2.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.input_proj.weight": { + "shape": [ + 512, + 637 + ], + "dtype": "torch.float32" + }, + "denoiser.input_proj.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.text_proj.weight": { + "shape": [ + 512, + 768 + ], + "dtype": "torch.float32" + }, + "denoiser.text_proj.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.extra_proj.0.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.extra_proj.0.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.extra_proj.1.weight": { + "shape": [ + 512, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.extra_proj.1.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.norm_cross.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.norm_cross.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.cross_attn.in_proj_weight": { + "shape": [ + 1536, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.cross_attn.in_proj_bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.cross_attn.out_proj.weight": { + "shape": [ + 512, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.cross_attn.out_proj.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.norm_self.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.norm_self.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.self_attn.qkv.weight": { + "shape": [ + 1536, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.self_attn.qkv.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.self_attn.out_proj.weight": { + "shape": [ + 512, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.self_attn.out_proj.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.norm_ff.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.norm_ff.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.ff.0.weight": { + "shape": [ + 2048, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.ff.0.bias": { + "shape": [ + 2048 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.ff.3.weight": { + "shape": [ + 512, + 2048 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.ff.3.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.audio_film.0.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.audio_film.0.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.audio_film.1.weight": { + "shape": [ + 1024, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.audio_film.1.bias": { + "shape": [ + 1024 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.human_motion_film.0.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.human_motion_film.0.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.human_motion_film.1.weight": { + "shape": [ + 1024, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.human_motion_film.1.bias": { + "shape": [ + 1024 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.norm_cross.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.norm_cross.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.cross_attn.in_proj_weight": { + "shape": [ + 1536, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.cross_attn.in_proj_bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.cross_attn.out_proj.weight": { + "shape": [ + 512, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.cross_attn.out_proj.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.norm_self.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.norm_self.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.self_attn.qkv.weight": { + "shape": [ + 1536, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.self_attn.qkv.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.self_attn.out_proj.weight": { + "shape": [ + 512, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.self_attn.out_proj.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.norm_ff.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.norm_ff.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.ff.0.weight": { + "shape": [ + 2048, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.ff.0.bias": { + "shape": [ + 2048 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.ff.3.weight": { + "shape": [ + 512, + 2048 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.ff.3.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.audio_film.0.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.audio_film.0.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.audio_film.1.weight": { + "shape": [ + 1024, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.audio_film.1.bias": { + "shape": [ + 1024 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.human_motion_film.0.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.human_motion_film.0.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.human_motion_film.1.weight": { + "shape": [ + 1024, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.human_motion_film.1.bias": { + "shape": [ + 1024 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.norm_cross.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.norm_cross.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.cross_attn.in_proj_weight": { + "shape": [ + 1536, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.cross_attn.in_proj_bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.cross_attn.out_proj.weight": { + "shape": [ + 512, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.cross_attn.out_proj.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.norm_self.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.norm_self.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.self_attn.qkv.weight": { + "shape": [ + 1536, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.self_attn.qkv.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.self_attn.out_proj.weight": { + "shape": [ + 512, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.self_attn.out_proj.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.norm_ff.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.norm_ff.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.ff.0.weight": { + "shape": [ + 2048, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.ff.0.bias": { + "shape": [ + 2048 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.ff.3.weight": { + "shape": [ + 512, + 2048 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.ff.3.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.audio_film.0.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.audio_film.0.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.audio_film.1.weight": { + "shape": [ + 1024, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.audio_film.1.bias": { + "shape": [ + 1024 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.human_motion_film.0.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.human_motion_film.0.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.human_motion_film.1.weight": { + "shape": [ + 1024, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.human_motion_film.1.bias": { + "shape": [ + 1024 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.norm_cross.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.norm_cross.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.cross_attn.in_proj_weight": { + "shape": [ + 1536, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.cross_attn.in_proj_bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.cross_attn.out_proj.weight": { + "shape": [ + 512, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.cross_attn.out_proj.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.norm_self.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.norm_self.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.self_attn.qkv.weight": { + "shape": [ + 1536, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.self_attn.qkv.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.self_attn.out_proj.weight": { + "shape": [ + 512, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.self_attn.out_proj.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.norm_ff.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.norm_ff.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.ff.0.weight": { + "shape": [ + 2048, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.ff.0.bias": { + "shape": [ + 2048 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.ff.3.weight": { + "shape": [ + 512, + 2048 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.ff.3.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.audio_film.0.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.audio_film.0.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.audio_film.1.weight": { + "shape": [ + 1024, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.audio_film.1.bias": { + "shape": [ + 1024 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.human_motion_film.0.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.human_motion_film.0.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.human_motion_film.1.weight": { + "shape": [ + 1024, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.human_motion_film.1.bias": { + "shape": [ + 1024 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.norm_cross.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.norm_cross.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.cross_attn.in_proj_weight": { + "shape": [ + 1536, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.cross_attn.in_proj_bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.cross_attn.out_proj.weight": { + "shape": [ + 512, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.cross_attn.out_proj.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.norm_self.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.norm_self.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.self_attn.qkv.weight": { + "shape": [ + 1536, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.self_attn.qkv.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.self_attn.out_proj.weight": { + "shape": [ + 512, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.self_attn.out_proj.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.norm_ff.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.norm_ff.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.ff.0.weight": { + "shape": [ + 2048, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.ff.0.bias": { + "shape": [ + 2048 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.ff.3.weight": { + "shape": [ + 512, + 2048 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.ff.3.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.audio_film.0.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.audio_film.0.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.audio_film.1.weight": { + "shape": [ + 1024, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.audio_film.1.bias": { + "shape": [ + 1024 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.human_motion_film.0.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.human_motion_film.0.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.human_motion_film.1.weight": { + "shape": [ + 1024, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.human_motion_film.1.bias": { + "shape": [ + 1024 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.norm_cross.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.norm_cross.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.cross_attn.in_proj_weight": { + "shape": [ + 1536, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.cross_attn.in_proj_bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.cross_attn.out_proj.weight": { + "shape": [ + 512, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.cross_attn.out_proj.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.norm_self.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.norm_self.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.self_attn.qkv.weight": { + "shape": [ + 1536, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.self_attn.qkv.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.self_attn.out_proj.weight": { + "shape": [ + 512, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.self_attn.out_proj.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.norm_ff.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.norm_ff.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.ff.0.weight": { + "shape": [ + 2048, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.ff.0.bias": { + "shape": [ + 2048 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.ff.3.weight": { + "shape": [ + 512, + 2048 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.ff.3.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.audio_film.0.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.audio_film.0.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.audio_film.1.weight": { + "shape": [ + 1024, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.audio_film.1.bias": { + "shape": [ + 1024 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.human_motion_film.0.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.human_motion_film.0.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.human_motion_film.1.weight": { + "shape": [ + 1024, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.human_motion_film.1.bias": { + "shape": [ + 1024 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.norm_cross.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.norm_cross.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.cross_attn.in_proj_weight": { + "shape": [ + 1536, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.cross_attn.in_proj_bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.cross_attn.out_proj.weight": { + "shape": [ + 512, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.cross_attn.out_proj.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.norm_self.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.norm_self.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.self_attn.qkv.weight": { + "shape": [ + 1536, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.self_attn.qkv.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.self_attn.out_proj.weight": { + "shape": [ + 512, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.self_attn.out_proj.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.norm_ff.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.norm_ff.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.ff.0.weight": { + "shape": [ + 2048, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.ff.0.bias": { + "shape": [ + 2048 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.ff.3.weight": { + "shape": [ + 512, + 2048 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.ff.3.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.audio_film.0.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.audio_film.0.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.audio_film.1.weight": { + "shape": [ + 1024, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.audio_film.1.bias": { + "shape": [ + 1024 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.human_motion_film.0.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.human_motion_film.0.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.human_motion_film.1.weight": { + "shape": [ + 1024, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.human_motion_film.1.bias": { + "shape": [ + 1024 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.norm_cross.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.norm_cross.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.cross_attn.in_proj_weight": { + "shape": [ + 1536, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.cross_attn.in_proj_bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.cross_attn.out_proj.weight": { + "shape": [ + 512, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.cross_attn.out_proj.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.norm_self.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.norm_self.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.self_attn.qkv.weight": { + "shape": [ + 1536, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.self_attn.qkv.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.self_attn.out_proj.weight": { + "shape": [ + 512, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.self_attn.out_proj.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.norm_ff.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.norm_ff.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.ff.0.weight": { + "shape": [ + 2048, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.ff.0.bias": { + "shape": [ + 2048 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.ff.3.weight": { + "shape": [ + 512, + 2048 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.ff.3.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.audio_film.0.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.audio_film.0.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.audio_film.1.weight": { + "shape": [ + 1024, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.audio_film.1.bias": { + "shape": [ + 1024 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.human_motion_film.0.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.human_motion_film.0.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.human_motion_film.1.weight": { + "shape": [ + 1024, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.human_motion_film.1.bias": { + "shape": [ + 1024 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.norm_cross.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.norm_cross.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.cross_attn.in_proj_weight": { + "shape": [ + 1536, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.cross_attn.in_proj_bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.cross_attn.out_proj.weight": { + "shape": [ + 512, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.cross_attn.out_proj.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.norm_self.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.norm_self.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.self_attn.qkv.weight": { + "shape": [ + 1536, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.self_attn.qkv.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.self_attn.out_proj.weight": { + "shape": [ + 512, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.self_attn.out_proj.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.norm_ff.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.norm_ff.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.ff.0.weight": { + "shape": [ + 2048, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.ff.0.bias": { + "shape": [ + 2048 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.ff.3.weight": { + "shape": [ + 512, + 2048 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.ff.3.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.audio_film.0.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.audio_film.0.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.audio_film.1.weight": { + "shape": [ + 1024, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.audio_film.1.bias": { + "shape": [ + 1024 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.human_motion_film.0.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.human_motion_film.0.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.human_motion_film.1.weight": { + "shape": [ + 1024, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.human_motion_film.1.bias": { + "shape": [ + 1024 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.norm_cross.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.norm_cross.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.cross_attn.in_proj_weight": { + "shape": [ + 1536, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.cross_attn.in_proj_bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.cross_attn.out_proj.weight": { + "shape": [ + 512, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.cross_attn.out_proj.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.norm_self.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.norm_self.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.self_attn.qkv.weight": { + "shape": [ + 1536, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.self_attn.qkv.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.self_attn.out_proj.weight": { + "shape": [ + 512, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.self_attn.out_proj.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.norm_ff.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.norm_ff.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.ff.0.weight": { + "shape": [ + 2048, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.ff.0.bias": { + "shape": [ + 2048 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.ff.3.weight": { + "shape": [ + 512, + 2048 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.ff.3.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.audio_film.0.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.audio_film.0.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.audio_film.1.weight": { + "shape": [ + 1024, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.audio_film.1.bias": { + "shape": [ + 1024 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.human_motion_film.0.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.human_motion_film.0.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.human_motion_film.1.weight": { + "shape": [ + 1024, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.human_motion_film.1.bias": { + "shape": [ + 1024 + ], + "dtype": "torch.float32" + }, + "denoiser.norm.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.norm.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.output.weight": { + "shape": [ + 125, + 512 + ], + "dtype": "torch.float32" + }, + "denoiser.output.bias": { + "shape": [ + 125 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.shared.weight": { + "shape": [ + 32128, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.embed_tokens.weight": { + "shape": [ + 32128, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.0.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.0.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.0.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.0.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.0.layer.0.SelfAttention.relative_attention_bias.weight": { + "shape": [ + 32, + 12 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.0.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.0.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.0.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.0.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.1.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.1.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.1.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.1.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.1.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.1.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.1.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.1.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.2.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.2.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.2.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.2.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.2.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.2.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.2.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.2.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.3.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.3.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.3.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.3.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.3.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.3.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.3.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.3.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.4.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.4.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.4.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.4.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.4.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.4.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.4.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.4.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.5.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.5.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.5.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.5.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.5.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.5.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.5.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.5.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.6.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.6.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.6.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.6.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.6.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.6.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.6.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.6.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.7.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.7.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.7.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.7.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.7.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.7.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.7.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.7.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.8.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.8.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.8.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.8.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.8.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.8.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.8.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.8.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.9.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.9.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.9.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.9.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.9.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.9.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.9.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.9.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.10.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.10.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.10.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.10.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.10.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.10.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.10.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.10.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.11.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.11.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.11.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.11.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.11.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.11.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.11.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.11.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.final_layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "history_projector.0.weight": { + "shape": [ + 125 + ], + "dtype": "torch.float32" + }, + "history_projector.0.bias": { + "shape": [ + 125 + ], + "dtype": "torch.float32" + }, + "history_projector.1.weight": { + "shape": [ + 512, + 125 + ], + "dtype": "torch.float32" + }, + "history_projector.1.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "history_projector.3.weight": { + "shape": [ + 512, + 512 + ], + "dtype": "torch.float32" + }, + "history_projector.3.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "audio_embedder.0.weight": { + "shape": [ + 35 + ], + "dtype": "torch.float32" + }, + "audio_embedder.0.bias": { + "shape": [ + 35 + ], + "dtype": "torch.float32" + }, + "audio_embedder.1.weight": { + "shape": [ + 1024, + 35 + ], + "dtype": "torch.float32" + }, + "audio_embedder.1.bias": { + "shape": [ + 1024 + ], + "dtype": "torch.float32" + }, + "audio_embedder.4.weight": { + "shape": [ + 512, + 1024 + ], + "dtype": "torch.float32" + }, + "audio_embedder.4.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "human_motion_embedder.0.weight": { + "shape": [ + 66 + ], + "dtype": "torch.float32" + }, + "human_motion_embedder.0.bias": { + "shape": [ + 66 + ], + "dtype": "torch.float32" + }, + "human_motion_embedder.1.weight": { + "shape": [ + 1024, + 66 + ], + "dtype": "torch.float32" + }, + "human_motion_embedder.1.bias": { + "shape": [ + 1024 + ], + "dtype": "torch.float32" + }, + "human_motion_embedder.4.weight": { + "shape": [ + 512, + 1024 + ], + "dtype": "torch.float32" + }, + "human_motion_embedder.4.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + } + }, + "loops": { + "fit_loop": { + "state_dict": { + "keys": [], + "count": 0 + }, + "epoch_loop.state_dict": { + "keys": [ + "_batches_that_stepped" + ], + "count": 1 + }, + "epoch_loop.batch_progress": { + "keys": [ + "total", + "current", + "is_last_batch" + ], + "count": 3 + }, + "epoch_loop.scheduler_progress": { + "keys": [ + "total", + "current" + ], + "count": 2 + }, + "epoch_loop.automatic_optimization.state_dict": { + "keys": [], + "count": 0 + }, + "epoch_loop.automatic_optimization.optim_progress": { + "keys": [ + "optimizer" + ], + "count": 1 + }, + "epoch_loop.manual_optimization.state_dict": { + "keys": [], + "count": 0 + }, + "epoch_loop.manual_optimization.optim_step_progress": { + "keys": [ + "total", + "current" + ], + "count": 2 + }, + "epoch_loop.val_loop.state_dict": { + "keys": [], + "count": 0 + }, + "epoch_loop.val_loop.batch_progress": { + "keys": [ + "total", + "current", + "is_last_batch" + ], + "count": 3 + }, + "epoch_progress": { + "keys": [ + "total", + "current" + ], + "count": 2 + } + }, + "validate_loop": { + "state_dict": { + "keys": [], + "count": 0 + }, + "batch_progress": { + "keys": [ + "total", + "current", + "is_last_batch" + ], + "count": 3 + } + }, + "test_loop": { + "state_dict": { + "keys": [], + "count": 0 + }, + "batch_progress": { + "keys": [ + "total", + "current", + "is_last_batch" + ], + "count": 3 + } + }, + "predict_loop": { + "state_dict": { + "keys": [], + "count": 0 + }, + "batch_progress": { + "keys": [ + "total", + "current" + ], + "count": 2 + } + } + }, + "callbacks": { + "ModelCheckpoint{'monitor': 'step', 'mode': 'max', 'every_n_train_steps': 256, 'every_n_epochs': 0, 'train_time_interval': None}": { + "monitor": "step", + "best_model_score": { + "shape": [], + "dtype": "torch.int64" + }, + "best_model_path": "/root/autodl-tmp/runs/hfd-omg-english-split20261026/checkpoints/step-00099840.ckpt", + "current_score": { + "shape": [], + "dtype": "torch.int64" + }, + "dirpath": "/root/autodl-tmp/runs/hfd-omg-english-split20261026/checkpoints", + "best_k_models": { + "keys": [ + "/root/autodl-tmp/runs/hfd-omg-english-split20261026/checkpoints/step-00099328.ckpt", + "/root/autodl-tmp/runs/hfd-omg-english-split20261026/checkpoints/step-00099584.ckpt", + "/root/autodl-tmp/runs/hfd-omg-english-split20261026/checkpoints/step-00099840.ckpt" + ], + "count": 3 + }, + "kth_best_model_path": "/root/autodl-tmp/runs/hfd-omg-english-split20261026/checkpoints/step-00099328.ckpt", + "kth_value": { + "shape": [], + "dtype": "torch.int64" + }, + "last_model_path": "/root/autodl-tmp/runs/hfd-omg-english-split20261026/checkpoints/last-v1.ckpt" + }, + "ModelCheckpoint{'monitor': 'val/loss', 'mode': 'min', 'every_n_train_steps': 0, 'every_n_epochs': 1, 'train_time_interval': None}": { + "monitor": "val/loss", + "best_model_score": { + "shape": [], + "dtype": "torch.float32" + }, + "best_model_path": "/root/autodl-tmp/runs/hfd-omg-english-split20261026/best/best.ckpt", + "current_score": { + "shape": [], + "dtype": "torch.float32" + }, + "dirpath": "/root/autodl-tmp/runs/hfd-omg-english-split20261026/best", + "best_k_models": { + "keys": [ + "/root/autodl-tmp/runs/hfd-omg-english-split20261026/best/best.ckpt" + ], + "count": 1 + }, + "kth_best_model_path": "/root/autodl-tmp/runs/hfd-omg-english-split20261026/best/best.ckpt", + "kth_value": { + "shape": [], + "dtype": "torch.float32" + }, + "last_model_path": "" + } + }, + "optimizer_states": { + "count": 1, + "sample": [ + { + "state": { + "keys": [ + "0", + "1", + "2", + "3", + "4", + "5", + "6", + "7", + "8", + "9", + "10", + "11", + "12", + "13", + "14", + "15", + "16", + "17", + "18", + "19", + "20", + "21", + "22", + "23", + "24", + "25", + "26", + "27", + "28", + "29" + ], + "count": 248 + }, + "param_groups": { + "count": 1, + "sample": [ + { + "keys": [ + "lr", + "betas", + "eps", + "weight_decay", + "amsgrad", + "maximize", + "foreach", + "capturable", + "differentiable", + "fused", + "decoupled_weight_decay", + "initial_lr", + "params" + ], + "count": 13 + } + ] + } + } + ] + }, + "lr_schedulers": { + "count": 1, + "sample": [ + { + "_milestones": { + "count": 1, + "sample": [ + 2000 + ] + }, + "last_epoch": 100000, + "_last_lr": { + "count": 1, + "sample": [ + 1e-06 + ] + }, + "_schedulers": { + "count": 2, + "sample": [ + { + "keys": [ + "start_factor", + "end_factor", + "total_iters", + "base_lrs", + "last_epoch", + "_step_count", + "_get_lr_called_within_step", + "_last_lr" + ], + "count": 8 + }, + { + "keys": [ + "T_max", + "eta_min", + "base_lrs", + "last_epoch", + "_step_count", + "_get_lr_called_within_step", + "_last_lr" + ], + "count": 7 + } + ] + } + } + ] + }, + "hfd_resume": { + "version": 1, + "rng": { + "python": { + "count": 3, + "sample": [ + 3, + { + "count": 625, + "sample": [ + 2147483648, + 3564348608 + ] + } + ] + }, + "numpy": { + "count": 5, + "sample": [ + "MT19937", + { + "count": 624, + "sample": [ + 42, + 3107752595 + ] + } + ] + }, + "torch": { + "shape": [ + 5056 + ], + "dtype": "torch.uint8" + }, + "cuda": { + "count": 1, + "sample": [ + { + "shape": [ + 16 + ], + "dtype": "torch.uint8" + } + ] + } + }, + "train_pool": { + "step": 100000, + "slots": { + "count": 512, + "sample": [ + 248743, + 347419 + ] + }, + "batch_size": 512, + "refresh": 256, + "rng": { + "keys": [ + "bit_generator", + "state", + "has_uint32", + "uinteger" + ], + "count": 4 + }, + "identity": { + "keys": [ + "sessions", + "native_source_sha256", + "asr", + "recipe_sha256", + "foreground_language", + "ordered_session_selection_sha256", + "final_release" + ], + "count": 7 + } + }, + "val_pool": { + "step": 16, + "slots": { + "count": 512, + "sample": [ + 6295, + 7193 + ] + }, + "batch_size": 512, + "refresh": 256, + "rng": { + "keys": [ + "bit_generator", + "state", + "has_uint32", + "uinteger" + ], + "count": 4 + }, + "identity": { + "keys": [ + "sessions", + "native_source_sha256", + "asr", + "recipe_sha256", + "foreground_language", + "ordered_session_selection_sha256", + "final_release" + ], + "count": 7 + } + } + }, + "omg_model_architecture": { + "format": "omg.model_architecture", + "version": 1, + "denoiser_type": "omg.generation.denoisers.transformer.MotionTransformerDenoiser", + "frame_cond_injection": "per_layer_film", + "attention": { + "rotary_self_attention_qk_norm": true, + "cross_attention_qk_norm": true + } + } +} \ No newline at end of file diff --git a/handoff-v1/inventory/omg-b-latest-resume-source-audit.json b/handoff-v1/inventory/omg-b-latest-resume-source-audit.json new file mode 100644 index 0000000000000000000000000000000000000000..8b983d58de3d86355124c1cdd828e3337b1334e1 --- /dev/null +++ b/handoff-v1/inventory/omg-b-latest-resume-source-audit.json @@ -0,0 +1,569 @@ +{ + "source": "/root/autodl-tmp/runs/hfd-omg-english-split20261026/resume_history/20261028T112852271305Z/code_sha256.json", + "files": { + "hfd_bridge/storage.py": { + "expected": "02c4fe4eed5f6d17bf140568af365b0abc8e5843df018d65553828890058a1fa", + "actual": "02c4fe4eed5f6d17bf140568af365b0abc8e5843df018d65553828890058a1fa" + }, + "hfd_bridge/__init__.py": { + "expected": "885528eb35015189a0bcd2dc62b60c61878efcf959caef68ff69e93501cafd5c", + "actual": "885528eb35015189a0bcd2dc62b60c61878efcf959caef68ff69e93501cafd5c" + }, + "hfd_bridge/loader.py": { + "expected": "13772137dc5cbb036bd5b5e4ee6ffe8cb45ba8d033ad580887b4dd63b253ac9b", + "actual": "13772137dc5cbb036bd5b5e4ee6ffe8cb45ba8d033ad580887b4dd63b253ac9b" + }, + "hfd_bridge/sessions.py": { + "expected": "c271aecf69e52acaee178b1381e465f22af82732f606d4e0bd5f891e2bb86076", + "actual": "c271aecf69e52acaee178b1381e465f22af82732f606d4e0bd5f891e2bb86076" + }, + "hfd_bridge/asr.py": { + "expected": "4baaf4faecfdce2054844a41326760220af301eb11d6fad892bed8c7380dceb0", + "actual": "4baaf4faecfdce2054844a41326760220af301eb11d6fad892bed8c7380dceb0" + }, + "hfd_bridge/training.py": { + "expected": "b513a632bcf010c83c23a35bcfd09e189149ff8c0d7d667d8966ec09e6e55970", + "actual": "b513a632bcf010c83c23a35bcfd09e189149ff8c0d7d667d8966ec09e6e55970" + }, + "hfd_bridge/native.py": { + "expected": "72ca9f675c2dfc7a2e304191da7b4d4016cb6bc226ccc22f9e7d0aeb2889c9f4", + "actual": "72ca9f675c2dfc7a2e304191da7b4d4016cb6bc226ccc22f9e7d0aeb2889c9f4" + }, + "hfd_bridge/paths.py": { + "expected": "2aaee82ecad571dcc1186c275e270af6c854c670db4ae354ee984d23a50b8bf3", + "actual": "2aaee82ecad571dcc1186c275e270af6c854c670db4ae354ee984d23a50b8bf3" + }, + "hfd_bridge/configured_data.py": { + "expected": "f916e0f9d4cdd2a4120fe94a9f1dfe59bbd24606d71ce1941023f04f97f181cd", + "actual": "f916e0f9d4cdd2a4120fe94a9f1dfe59bbd24606d71ce1941023f04f97f181cd" + }, + "hfd_bridge/textop_checks.py": { + "expected": "00662346781a638298cfa344b9ce8a58b6e8bba11261cd5783fdca956f4df7c8", + "actual": "00662346781a638298cfa344b9ce8a58b6e8bba11261cd5783fdca956f4df7c8" + }, + "hfd_bridge/omg_checks.py": { + "expected": "1c80361412892538400cc225ab28aa03d5357ccab7a53e3a6d46deb5da9d32e8", + "actual": "094d997043d4e0d868261b7f25b9452fca832dfa48572aeac4a3f2ba9f42d172" + }, + "hfd_bridge/textop_condition.py": { + "expected": "4fca0f4a02df3b843a908b28b64766877585aa3300977359d21fdb3adb5a3c28", + "actual": "4fca0f4a02df3b843a908b28b64766877585aa3300977359d21fdb3adb5a3c28" + }, + "hfd_bridge/joints.py": { + "expected": "7827c82ca3463a026db6007c41a66528e1294485b4acb081f27f09449da2d154", + "actual": "7827c82ca3463a026db6007c41a66528e1294485b4acb081f27f09449da2d154" + }, + "hfd_bridge/audio_condition.py": { + "expected": "cab6795bdddc36aa63d4f54c3c8e8aaee7958c86882abe35ae76f67a79b4b945", + "actual": "cab6795bdddc36aa63d4f54c3c8e8aaee7958c86882abe35ae76f67a79b4b945" + }, + "hfd_bridge/batch_asr.py": { + "expected": "9d4f0ba1d382f45c7a0adf562bc008485fce7cd00e86e29607528676f503f8f1", + "actual": "9d4f0ba1d382f45c7a0adf562bc008485fce7cd00e86e29607528676f503f8f1" + }, + "hfd_bridge/ondemand.py": { + "expected": "940339aa16a6fa25a27ace23a956017f7325813b0cd2a49f76e6f1fb1d014cc0", + "actual": "940339aa16a6fa25a27ace23a956017f7325813b0cd2a49f76e6f1fb1d014cc0" + }, + "hfd_bridge/ondemand_data.py": { + "expected": "24c61e36623eaca0bdc2f7bff2c8ce307946c4a5afef18f0427b96c06c9ed57e", + "actual": "24c61e36623eaca0bdc2f7bff2c8ce307946c4a5afef18f0427b96c06c9ed57e" + }, + "hfd_bridge/prefetch.py": { + "expected": "b227251cc6470f980231999bbecbeb1fc583c754eda56e70d4f3c84ed1845d64", + "actual": "b227251cc6470f980231999bbecbeb1fc583c754eda56e70d4f3c84ed1845d64" + }, + "hfd_bridge/human_sessions.py": { + "expected": "0461ea66b1c324ec38e9e833ee8eb19b95a24c3f478b094fb600dcdc3c53cdea", + "actual": "0461ea66b1c324ec38e9e833ee8eb19b95a24c3f478b094fb600dcdc3c53cdea" + }, + "hfd_bridge/genmo.py": { + "expected": "3c898d2828267f38f8ba9f41aae62c4a52cbae87cf63dd2cef2b6634f7160060", + "actual": "3c898d2828267f38f8ba9f41aae62c4a52cbae87cf63dd2cef2b6634f7160060" + }, + "hfd_bridge/genmo_audio.py": { + "expected": "1cdfa6ebc19a75af7540b2280f5dc9c3ecc584ed1c386c434de32e3777020c6f", + "actual": "1cdfa6ebc19a75af7540b2280f5dc9c3ecc584ed1c386c434de32e3777020c6f" + }, + "hfd_bridge/clocks.py": { + "expected": "27054c0b2d3e7d4087eb8fc3c59c6bcf5a581ca2e3022341e306732dcbe92637", + "actual": "27054c0b2d3e7d4087eb8fc3c59c6bcf5a581ca2e3022341e306732dcbe92637" + }, + "hfd_bridge/semantics.py": { + "expected": "42ed59d58609095b8053d0598a9c575d52af9b8508e699162743cc11599f03b4", + "actual": "42ed59d58609095b8053d0598a9c575d52af9b8508e699162743cc11599f03b4" + }, + "hfd_bridge/reference_text.py": { + "expected": "50da93fed3be2bad05035bf978c9be41ba7b48aaf3ce7c43a879519e7386a121", + "actual": "50da93fed3be2bad05035bf978c9be41ba7b48aaf3ce7c43a879519e7386a121" + }, + "hfd_bridge/reference_runtime.py": { + "expected": "2b7fe129152d0f7ea3a4a4e3f5fc9ae799e805138a0d76487e9f59065fe7d975", + "actual": "2b7fe129152d0f7ea3a4a4e3f5fc9ae799e805138a0d76487e9f59065fe7d975" + }, + "hfd_bridge/resume.py": { + "expected": "be1b665778f3c3bc137f4242b0ac4d1f587903357954725ed509e712dac0ab8f", + "actual": "e4e373802e0dd21e39fa2d899eff524080ce1735ca86eefbf4ef2080f170ffbf" + }, + "hfd_bridge/resume_lightning.py": { + "expected": "69e4c6bf11a199f3a979227eb551870c44e371fcae0af924ec64d2d58d3ce7a6", + "actual": "69e4c6bf11a199f3a979227eb551870c44e371fcae0af924ec64d2d58d3ce7a6" + }, + "hfd_bridge/index_cache.py": { + "expected": "d75cbcb0105448c61c841e525a1f17e63c55c83ad9201104e8118e22f7728b17", + "actual": "d75cbcb0105448c61c841e525a1f17e63c55c83ad9201104e8118e22f7728b17" + }, + "hfd_bridge/genmo_data.py": { + "expected": "4b102d1eecb0af9757a76d9600ede345c3f21f92166276a22cbf8e144fa9a38c", + "actual": "4b102d1eecb0af9757a76d9600ede345c3f21f92166276a22cbf8e144fa9a38c" + }, + "hfd_bridge/genmo_validation.py": { + "expected": "fa265cf9df51a4b36db977be945547dd78bdf105db9aed8ccba940b1a977f5e4", + "actual": "fa265cf9df51a4b36db977be945547dd78bdf105db9aed8ccba940b1a977f5e4" + }, + "hfd_bridge/kimodo_data.py": { + "expected": "35b3b7d8fc8f36fe65eb311e6e65e667fe1fdbcdfc23e58f37fb75c1c9e9feae", + "actual": "35b3b7d8fc8f36fe65eb311e6e65e667fe1fdbcdfc23e58f37fb75c1c9e9feae" + }, + "hfd_bridge/kimodo.py": { + "expected": "081a1047b853a2ee143aad2a5699f33cbecd0b4f93a07ad085760e16665f20ce", + "actual": "081a1047b853a2ee143aad2a5699f33cbecd0b4f93a07ad085760e16665f20ce" + }, + "hfd_bridge/kimodo_recovery.py": { + "expected": "b9177ff13b55b753430dd14d46778b723d2a03ce8927313c246c5d7ce780809b", + "actual": "b9177ff13b55b753430dd14d46778b723d2a03ce8927313c246c5d7ce780809b" + }, + "hfd_bridge/kimodo_closed_loop.py": { + "expected": "38d7e022b1450a715b35542cd23a70b1491a6dbd307dbb4be730b9f562100f91", + "actual": "38d7e022b1450a715b35542cd23a70b1491a6dbd307dbb4be730b9f562100f91" + }, + "hfd_bridge/kimodo_causal_rep.py": { + "expected": "c0c40003a3d36ec35e8490fa1165ac88723178b15a0e118d2d1d4c9f7ed7c537", + "actual": "c0c40003a3d36ec35e8490fa1165ac88723178b15a0e118d2d1d4c9f7ed7c537" + }, + "hfd_bridge/kimodo_diffusion_training.py": { + "expected": "078fe8cf1ee39d910a6ca5ef94d67f20c4398cfe505895a089acb80c05336539", + "actual": "078fe8cf1ee39d910a6ca5ef94d67f20c4398cfe505895a089acb80c05336539" + }, + "hfd_bridge/kimodo_text_worker.py": { + "expected": "a275c121a9364ae85b6233a67f3acb63ceddb2c9da490c2ea7d476f4cc710751", + "actual": "a275c121a9364ae85b6233a67f3acb63ceddb2c9da490c2ea7d476f4cc710751" + }, + "hfd_bridge/kimodo_full_data.py": { + "expected": "c62b4bb2673f1875792edb3403347f76bb52fa2360d92a9c1091eae10d43f905", + "actual": "c62b4bb2673f1875792edb3403347f76bb52fa2360d92a9c1091eae10d43f905" + }, + "hfd_bridge/kimodo_checkpoint.py": { + "expected": "b9cd24df225c91a6212c354983e3f51acb495373747f303c6096000af37f7d59", + "actual": "b9cd24df225c91a6212c354983e3f51acb495373747f303c6096000af37f7d59" + }, + "hfd_bridge/kimodo_finetuning.py": { + "expected": "a2c54f04af2b516c14d00025c8b695093ff09da6f04b0b73eb3269817bc0a544", + "actual": "a2c54f04af2b516c14d00025c8b695093ff09da6f04b0b73eb3269817bc0a544" + }, + "hfd_bridge/reference_assets.py": { + "expected": "e977a1d804807629fe925d511acdd25b488a11acc3767c0234020af8f699a43c", + "actual": "e977a1d804807629fe925d511acdd25b488a11acc3767c0234020af8f699a43c" + }, + "hfd_bridge/omg_model_loader.py": { + "expected": "f972080dd6140e1574124be615c9351001b51a73c5774b63bd1c063671bd56aa", + "actual": "f972080dd6140e1574124be615c9351001b51a73c5774b63bd1c063671bd56aa" + }, + "hfd_bridge/genmo_geometry.py": { + "expected": "4d3e912a858b8d3d4ddf73941c3120d5304ed70187c02fc969730816e155b2cb", + "actual": "4d3e912a858b8d3d4ddf73941c3120d5304ed70187c02fc969730816e155b2cb" + }, + "hfd_bridge/genmo_training.py": { + "expected": "8276d4f266b44e2c112e14021eb29c442753f2de07abf29e9adb321226dc87d8", + "actual": "556100608e466ca801abd623d24af30eebe7f9e3ac598c6709f813f07d5d45e9" + }, + "hfd_bridge/genmo_prefetch.py": { + "expected": "3375352f7668dfd5c76ff629853aa8e4d0bcd2e2294a6e96932280f2eafa68e4", + "actual": "0f00c9c5c951f53a0b55cad0c71d470c9312381e75ec31f756f606cf20309141" + }, + "hfd_bridge/hymotion_rep.py": { + "expected": "ee2f8282720d311b98d3ceb8cc72f63219622b5f61a138af6eb3221181a89f7a", + "actual": "ee2f8282720d311b98d3ceb8cc72f63219622b5f61a138af6eb3221181a89f7a" + }, + "hfd_bridge/hymotion_model.py": { + "expected": "db0968384b87929c6495b87647a37e92e13729a94a34f6a5347f4d23ed05be5c", + "actual": "ed29d2d03b247ed681a09dc230697885c5ab8fa8b30f82c4a827aa8fe0b41597" + }, + "hfd_bridge/hymotion_data.py": { + "expected": "06170526a18b60fa486f084864e2e65f2608274ed967ef367b29de269f8c9147", + "actual": "79503602a9a4de4f67931fa944a198d6f438b531a363ffbe545ecc4dd54f97e7" + }, + "hfd_bridge/hymotion_text.py": { + "expected": "f0238ffc3656e2a90a26360c5e9bc8c6de3d2d6be0af6e87787c83237914e98b", + "actual": "f0238ffc3656e2a90a26360c5e9bc8c6de3d2d6be0af6e87787c83237914e98b" + }, + "hfd_bridge/hymotion_rollout.py": { + "expected": "a7fefe8d72f9e3614103632d68d1d63ce3ee072650739683131f5315d0a41181", + "actual": "6e27c4e24bd42a211df7be2a0271337af4e658d66e5d303633912bbd40313517" + }, + "hfd_bridge/hymotion_validation.py": { + "expected": "cf6fd4b14fdbafc65a8a65abaade620b43d5b1f95e7a7a6194a21247f2ad4127", + "actual": "6043090f19315c9fde677168fed06dca1fb390b559a7b9668e0ba46799074cd2" + }, + "hfd_bridge/hymotion_training.py": { + "expected": "67f40e79cc43950f6876f8c0c2c67fe1f606f1b67e692a7dd239457fa4b96a2c", + "actual": "792b76cfc35f58b0c7c94b40a660f00c4971fff1d978ffd0c923b9f26a5fa733" + }, + "hfd_bridge/hymotion_clock.py": { + "expected": "1eaf969e9198f6902759457488ff8c80fc20c973b72388443b3ef0af7c8282a2", + "actual": null + }, + "hfd_bridge/hymotion_recovery.py": { + "expected": "e47b841f85dfc19b37a646a6cf6e865dd2642718073a3419a97195f16024c379", + "actual": null + }, + "hfd_bridge/human_eval.py": { + "expected": "c943c332315f16f41cb53beee8303f2d94a4b522e678cfd346e8a6c9ec203b71", + "actual": "c943c332315f16f41cb53beee8303f2d94a4b522e678cfd346e8a6c9ec203b71" + }, + "hfd_bridge/omg_rollout.py": { + "expected": "24ecdec4a2c698bd8cdc2b4ac6e1c0ad960c252b5a9f9a15d57a7716545fc4d7", + "actual": "24ecdec4a2c698bd8cdc2b4ac6e1c0ad960c252b5a9f9a15d57a7716545fc4d7" + }, + "hfd_bridge/final_split.py": { + "expected": "d0f0adef291d17d3b73477258b5de74bb8ed477ca1952c9b83ddb115efd34305", + "actual": "d0f0adef291d17d3b73477258b5de74bb8ed477ca1952c9b83ddb115efd34305" + }, + "hfd_bridge/final_third_party.py": { + "expected": "3926a0f72d419bb5006d8a62136a8f5ffd5ddca439c39dd7768f7cf03422d716", + "actual": "3926a0f72d419bb5006d8a62136a8f5ffd5ddca439c39dd7768f7cf03422d716" + }, + "hfd_bridge/replay_numerics.py": { + "expected": "827c70aa9b2308ac33b6fccee2c4dc88df58d6b89d902d8c215adfdf918ab231", + "actual": "827c70aa9b2308ac33b6fccee2c4dc88df58d6b89d902d8c215adfdf918ab231" + }, + "hfd_bridge/textop_text_cache.py": { + "expected": "3d5c8a187e461144aa97cc539661e6e90b475800379c0c096b243bcd337a0e38", + "actual": "3d5c8a187e461144aa97cc539661e6e90b475800379c0c096b243bcd337a0e38" + }, + "hfd_bridge/hot_assets.py": { + "expected": "43830324054a551ef4ed364ad45d9b2dfd8830b7dde2661f4404b1ffc8cabe10", + "actual": "43830324054a551ef4ed364ad45d9b2dfd8830b7dde2661f4404b1ffc8cabe10" + }, + "upstream/omg/download_data.py": { + "expected": "a247118ce874864affc5153a23d94788871d3b6d327bcc33ceeba45279d7732c", + "actual": "a247118ce874864affc5153a23d94788871d3b6d327bcc33ceeba45279d7732c" + }, + "upstream/omg/download_models.py": { + "expected": "811132811b35321174adc69a754cf64f9437eeba5d19f6ad02625ef9d9f73b57", + "actual": "811132811b35321174adc69a754cf64f9437eeba5d19f6ad02625ef9d9f73b57" + }, + "upstream/omg/a.py": { + "expected": "66e1c68c0a631d6fef0d7d13948299671baf72ec813fe025c1cac1e222fd4c63", + "actual": "66e1c68c0a631d6fef0d7d13948299671baf72ec813fe025c1cac1e222fd4c63" + }, + "upstream/omg/smoke.py": { + "expected": "8008c82e92250f148512e3127b939df3aded7025a89473738d024444d21edcfd", + "actual": "8008c82e92250f148512e3127b939df3aded7025a89473738d024444d21edcfd" + }, + "upstream/omg/validate_model.py": { + "expected": "36023f739cd8dfe24179f9247b5de04ca425d7a0d73c99b166bbeaea778a975f", + "actual": "36023f739cd8dfe24179f9247b5de04ca425d7a0d73c99b166bbeaea778a975f" + }, + "upstream/omg/validate_setup.py": { + "expected": "4e086d4615c621386cb54885aad0ede341c24a71a1668c3e5386f368c02f3741", + "actual": "4e086d4615c621386cb54885aad0ede341c24a71a1668c3e5386f368c02f3741" + }, + "upstream/omg/paper_table.py": { + "expected": "97c16c1c044639362f46cac439159384a1d79dc80a90c839c5e2ce471d8060d2", + "actual": "97c16c1c044639362f46cac439159384a1d79dc80a90c839c5e2ce471d8060d2" + }, + "upstream/omg/paper_main_table.py": { + "expected": "8f95f11b00882e4574df37e2f44d5843a15073004f0b9f78450895c65939196f", + "actual": "8f95f11b00882e4574df37e2f44d5843a15073004f0b9f78450895c65939196f" + }, + "upstream/omg/scripts/prepare_omg_inputs.py": { + "expected": "99bbc364b5f380ae7c2cc0fa53657722f896d3613d8b98937881d06197beccb4", + "actual": "99bbc364b5f380ae7c2cc0fa53657722f896d3613d8b98937881d06197beccb4" + }, + "upstream/omg/scripts/extract_omg_paper_row.py": { + "expected": "f64cb31b5a19efba6f3e1b1771d3d5175f6610392a588d118014e98790b82503", + "actual": "f64cb31b5a19efba6f3e1b1771d3d5175f6610392a588d118014e98790b82503" + }, + "upstream/omg/scripts/psi0_omg_generate.py": { + "expected": "38ccbd4c8f1d62b470193ae6e3598445f22408e2df6edc6b267ed46b4873f579", + "actual": "38ccbd4c8f1d62b470193ae6e3598445f22408e2df6edc6b267ed46b4873f579" + }, + "upstream/omg/scripts/verify_omg_artifact_match.py": { + "expected": "46815bd446b956583ae7feeb0ed03a58b6dde8cc78ecb4a1d4fd1a3a6b737ee7", + "actual": "46815bd446b956583ae7feeb0ed03a58b6dde8cc78ecb4a1d4fd1a3a6b737ee7" + }, + "upstream/omg/src/omg/__init__.py": { + "expected": "e06eff3bd6a146e0082518eb791164e26b80b548f9bdcd21c9f0aca1220bca34", + "actual": "e06eff3bd6a146e0082518eb791164e26b80b548f9bdcd21c9f0aca1220bca34" + }, + "upstream/omg/src/omg/callbacks/__init__.py": { + "expected": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", + "actual": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855" + }, + "upstream/omg/src/omg/callbacks/divergence_guard.py": { + "expected": "c1ec2773dddda3bd800247bca0abd3eca46aeae3c78ebae8050b98739292b61a", + "actual": "c1ec2773dddda3bd800247bca0abd3eca46aeae3c78ebae8050b98739292b61a" + }, + "upstream/omg/src/omg/callbacks/global_step_progress.py": { + "expected": "bd21180de0140a38fa2601f51b56c94aef346c757c140786e88b907720f3eb15", + "actual": "bd21180de0140a38fa2601f51b56c94aef346c757c140786e88b907720f3eb15" + }, + "upstream/omg/src/omg/callbacks/grad_norm.py": { + "expected": "c41e7e0bac01c984c85c38baf6de4b9e10ded95875a79ddc1c9d85ded5b9b742", + "actual": "c41e7e0bac01c984c85c38baf6de4b9e10ded95875a79ddc1c9d85ded5b9b742" + }, + "upstream/omg/src/omg/callbacks/speed_timer.py": { + "expected": "29ae0ec76a0545576bc31f240a4beab74f5c5e9ae50a0078c1ed46b8326fc360", + "actual": "29ae0ec76a0545576bc31f240a4beab74f5c5e9ae50a0078c1ed46b8326fc360" + }, + "upstream/omg/src/omg/callbacks/training_diagnostics.py": { + "expected": "6e348e5cca9e690379cdb7a3364678b666d7c53d0a7405d5269ca713574a6917", + "actual": "6e348e5cca9e690379cdb7a3364678b666d7c53d0a7405d5269ca713574a6917" + }, + "upstream/omg/src/omg/cli/__init__.py": { + "expected": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", + "actual": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855" + }, + "upstream/omg/src/omg/cli/data/__init__.py": { + "expected": "052be341b9c1b05585d3a27073a741d768974cfdb2a91659a687ed1d3d6291f9", + "actual": "052be341b9c1b05585d3a27073a741d768974cfdb2a91659a687ed1d3d6291f9" + }, + "upstream/omg/src/omg/cli/data/materialize_episode_cache.py": { + "expected": "c8e61cbcf17f55e17b47ad2a3e5c2aa0919d9f5bcc2a93edbf5a519e79f9793a", + "actual": "c8e61cbcf17f55e17b47ad2a3e5c2aa0919d9f5bcc2a93edbf5a519e79f9793a" + }, + "upstream/omg/src/omg/cli/data/validate_episode_cache.py": { + "expected": "6840e28c774964583d88c6402c201ab36c0150ec67139403a66782364173c878", + "actual": "6840e28c774964583d88c6402c201ab36c0150ec67139403a66782364173c878" + }, + "upstream/omg/src/omg/cli/generation/__init__.py": { + "expected": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", + "actual": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855" + }, + "upstream/omg/src/omg/cli/generation/compute_stats.py": { + "expected": "849ed1ac16612e23e09a887e5a030c5e0c496610e9afe882ce98e99d570bd6ca", + "actual": "849ed1ac16612e23e09a887e5a030c5e0c496610e9afe882ce98e99d570bd6ca" + }, + "upstream/omg/src/omg/cli/generation/train.py": { + "expected": "a39cc0c6cfccd7b6660bbd9ab4f88f90bc634bb95ce30e22492c8a6cb74a7c8c", + "actual": "a39cc0c6cfccd7b6660bbd9ab4f88f90bc634bb95ce30e22492c8a6cb74a7c8c" + }, + "upstream/omg/src/omg/cli/generation/benchmark.py": { + "expected": "ff813f344953da4b02388837f5a510230f8d4430393a88f205261faaf28c5dee", + "actual": "ff813f344953da4b02388837f5a510230f8d4430393a88f205261faaf28c5dee" + }, + "upstream/omg/src/omg/cli/generation/physical_benchmark.py": { + "expected": "f48b2b9d2d312c44d0e33eebf0c7fd95d8f2ce65a871fe253ef8ff9ceee30dd0", + "actual": "f48b2b9d2d312c44d0e33eebf0c7fd95d8f2ce65a871fe253ef8ff9ceee30dd0" + }, + "upstream/omg/src/omg/cli/generation/generate.py": { + "expected": "66eda839732e76a3b0fc39a767e6a238aa93bbf13ef2f3ddb6e1f9021948bc4a", + "actual": "66eda839732e76a3b0fc39a767e6a238aa93bbf13ef2f3ddb6e1f9021948bc4a" + }, + "upstream/omg/src/omg/cli/generation/export_onnx.py": { + "expected": "9cacee46fea5e18f8a9f9ffd52b25110f89e80be20d21acce85fe066392c00c1", + "actual": "9cacee46fea5e18f8a9f9ffd52b25110f89e80be20d21acce85fe066392c00c1" + }, + "upstream/omg/src/omg/core/__init__.py": { + "expected": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", + "actual": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855" + }, + "upstream/omg/src/omg/core/logging.py": { + "expected": "86c57564f7d8614d71d95e0ce377724642c8c3aa44501d9c317c33bf9ba7dce7", + "actual": "86c57564f7d8614d71d95e0ce377724642c8c3aa44501d9c317c33bf9ba7dce7" + }, + "upstream/omg/src/omg/core/paths.py": { + "expected": "6bb254b7219e70b58fc8581479857bfb7ee77a9f8937faf5dcec784dd1173a4d", + "actual": "6bb254b7219e70b58fc8581479857bfb7ee77a9f8937faf5dcec784dd1173a4d" + }, + "upstream/omg/src/omg/core/tensor.py": { + "expected": "998e6f24948983c15f3257f055a452e07b96cc0f13b357cb4881f4545bddb7d0", + "actual": "998e6f24948983c15f3257f055a452e07b96cc0f13b357cb4881f4545bddb7d0" + }, + "upstream/omg/src/omg/data/__init__.py": { + "expected": "4e8cd9d824b5b834a673011d1cd95cf0227a1deb60ee917ba85ec364d51db21f", + "actual": "4e8cd9d824b5b834a673011d1cd95cf0227a1deb60ee917ba85ec364d51db21f" + }, + "upstream/omg/src/omg/data/datamodule.py": { + "expected": "94cbaf84b019d559dd7810494694c156803d318f23a5ec95549315c59ea795d6", + "actual": "94cbaf84b019d559dd7810494694c156803d318f23a5ec95549315c59ea795d6" + }, + "upstream/omg/src/omg/data/episode_cache.py": { + "expected": "96639991cb75a3db602c608d929e4f742caef615166198e66731c406f068207e", + "actual": "96639991cb75a3db602c608d929e4f742caef615166198e66731c406f068207e" + }, + "upstream/omg/src/omg/data/episode_cache_inspect.py": { + "expected": "b7f43c2a5f7a1c63dcb9c243e79018ebff3ecce056fba22ac9b1948e754feb20", + "actual": "b7f43c2a5f7a1c63dcb9c243e79018ebff3ecce056fba22ac9b1948e754feb20" + }, + "upstream/omg/src/omg/data/lerobot_dataset.py": { + "expected": "78172eb344a37cab8457133a957eb25691ad78054f71db0c39171f241be12831", + "actual": "78172eb344a37cab8457133a957eb25691ad78054f71db0c39171f241be12831" + }, + "upstream/omg/src/omg/data/windowing.py": { + "expected": "88710f0c24d6c696670b702b9ce77609c59acd2106389fa769a00d39a98f2afb", + "actual": "88710f0c24d6c696670b702b9ce77609c59acd2106389fa769a00d39a98f2afb" + }, + "upstream/omg/src/omg/generation/__init__.py": { + "expected": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", + "actual": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855" + }, + "upstream/omg/src/omg/generation/architecture.py": { + "expected": "f77548e52394c78b813ad36dcd5fb8c8981ba393e5c723126491df6c171aeabf", + "actual": "f77548e52394c78b813ad36dcd5fb8c8981ba393e5c723126491df6c171aeabf" + }, + "upstream/omg/src/omg/generation/conditions/__init__.py": { + "expected": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", + "actual": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855" + }, + "upstream/omg/src/omg/generation/conditions/t5.py": { + "expected": "67c0f834106ff860e612a978bdba6ef503d2cf0dfb24e831af0bba723e95c890", + "actual": "67c0f834106ff860e612a978bdba6ef503d2cf0dfb24e831af0bba723e95c890" + }, + "upstream/omg/src/omg/generation/denoisers/__init__.py": { + "expected": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", + "actual": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855" + }, + "upstream/omg/src/omg/generation/denoisers/interface.py": { + "expected": "c1f8b5f2a93e6cf0a4b5df730f9f744c95d9b58bc5245e3cb15176a61065e1dd", + "actual": "c1f8b5f2a93e6cf0a4b5df730f9f744c95d9b58bc5245e3cb15176a61065e1dd" + }, + "upstream/omg/src/omg/generation/denoisers/transformer.py": { + "expected": "ae968a16ba6193607721f5c0c1910122818797407df3ee0f0fafa3728c0e398a", + "actual": "ae968a16ba6193607721f5c0c1910122818797407df3ee0f0fafa3728c0e398a" + }, + "upstream/omg/src/omg/generation/diffusion/__init__.py": { + "expected": "fab4af0ec5b7b38892ffb309d6073841de00f8f37e5b4c2da1657e1b30c92e1c", + "actual": "fab4af0ec5b7b38892ffb309d6073841de00f8f37e5b4c2da1657e1b30c92e1c" + }, + "upstream/omg/src/omg/generation/diffusion/forcing.py": { + "expected": "935b2f38db1d1d3c496df0b03cadaeb032ef1539522169f2731b87a8724362b5", + "actual": "935b2f38db1d1d3c496df0b03cadaeb032ef1539522169f2731b87a8724362b5" + }, + "upstream/omg/src/omg/generation/diffusion/guided.py": { + "expected": "3fd42c7ae9d01497223b08421ae3becc0a7be81846c1cf08c8bb944d8949b870", + "actual": "3fd42c7ae9d01497223b08421ae3becc0a7be81846c1cf08c8bb944d8949b870" + }, + "upstream/omg/src/omg/generation/diffusion/objectives.py": { + "expected": "db05bcbc32268e5d21298e19130e9e8f4b23eff7d1b679886306cd7c94195675", + "actual": "db05bcbc32268e5d21298e19130e9e8f4b23eff7d1b679886306cd7c94195675" + }, + "upstream/omg/src/omg/generation/diffusion/schedules.py": { + "expected": "a06c3f3ab1d81e07f680d44f8fb9ef56e8bd2a6e4c5e4ac911c316217f6e92e0", + "actual": "a06c3f3ab1d81e07f680d44f8fb9ef56e8bd2a6e4c5e4ac911c316217f6e92e0" + }, + "upstream/omg/src/omg/generation/export/__init__.py": { + "expected": "0ee7ed60aa43848dad0a4ba0e5dd29d815aab6dc37f4ae1c46d8e56a2ebc1770", + "actual": "0ee7ed60aa43848dad0a4ba0e5dd29d815aab6dc37f4ae1c46d8e56a2ebc1770" + }, + "upstream/omg/src/omg/generation/export/onnx.py": { + "expected": "168ad808867669b69cd69a419d5b8d2c3fd124c69c20daeca53b6fc59a0a21dd", + "actual": "168ad808867669b69cd69a419d5b8d2c3fd124c69c20daeca53b6fc59a0a21dd" + }, + "upstream/omg/src/omg/generation/losses/__init__.py": { + "expected": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", + "actual": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855" + }, + "upstream/omg/src/omg/generation/losses/contact.py": { + "expected": "75927b00fc5584988f2ecedbcfbba9af3ba80206966b0a91642ec133db5e4267", + "actual": "75927b00fc5584988f2ecedbcfbba9af3ba80206966b0a91642ec133db5e4267" + }, + "upstream/omg/src/omg/generation/losses/fk.py": { + "expected": "75927b00fc5584988f2ecedbcfbba9af3ba80206966b0a91642ec133db5e4267", + "actual": "75927b00fc5584988f2ecedbcfbba9af3ba80206966b0a91642ec133db5e4267" + }, + "upstream/omg/src/omg/generation/losses/motion.py": { + "expected": "00ee200046658dafcae2b9ec439ed0c3203ab533f4dd9d4a8475541df74c2fa5", + "actual": "00ee200046658dafcae2b9ec439ed0c3203ab533f4dd9d4a8475541df74c2fa5" + }, + "upstream/omg/src/omg/generation/losses/terrain.py": { + "expected": "75927b00fc5584988f2ecedbcfbba9af3ba80206966b0a91642ec133db5e4267", + "actual": "75927b00fc5584988f2ecedbcfbba9af3ba80206966b0a91642ec133db5e4267" + }, + "upstream/omg/src/omg/generation/metrics/__init__.py": { + "expected": "f686c86d57bab24aa1dd27cfae2bbab88757d7909e74cd7d4b9525a892ba025e", + "actual": "f686c86d57bab24aa1dd27cfae2bbab88757d7909e74cd7d4b9525a892ba025e" + }, + "upstream/omg/src/omg/generation/metrics/physical.py": { + "expected": "7aa2d8fc7d29b9d3ec836c3d23d0b5a7fc64a1eea5359c7300f6e6e191904f29", + "actual": "7aa2d8fc7d29b9d3ec836c3d23d0b5a7fc64a1eea5359c7300f6e6e191904f29" + }, + "upstream/omg/src/omg/generation/models/__init__.py": { + "expected": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", + "actual": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855" + }, + "upstream/omg/src/omg/generation/models/motion_generator.py": { + "expected": "05947c8e73646f634e9bbedf0a60f2becaf1ded21b9d1093f1084eaaefef2a3b", + "actual": "05947c8e73646f634e9bbedf0a60f2becaf1ded21b9d1093f1084eaaefef2a3b" + }, + "upstream/omg/src/omg/motion/__init__.py": { + "expected": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", + "actual": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855" + }, + "upstream/omg/src/omg/motion/feature_codec.py": { + "expected": "3d148816baf239dbc938210c959954c3c2fac4c6d9914ad48b851976ed7e2960", + "actual": "3d148816baf239dbc938210c959954c3c2fac4c6d9914ad48b851976ed7e2960" + }, + "upstream/omg/src/omg/motion/representation.py": { + "expected": "457328078b2a2652e06c39e5fa06c8a0d8ba9223dcf24e508a33cdd1806f45a7", + "actual": "457328078b2a2652e06c39e5fa06c8a0d8ba9223dcf24e508a33cdd1806f45a7" + }, + "upstream/omg/src/omg/pipeline/__init__.py": { + "expected": "3806048000a9bdc918ace57c68aabcd3c56bb503c4998b103d1593553a23a895", + "actual": "3806048000a9bdc918ace57c68aabcd3c56bb503c4998b103d1593553a23a895" + }, + "upstream/omg/src/omg/pipeline/audio.py": { + "expected": "d88a0b51fd72d80fdf7da1344082e7799f481fb114a246ed4e66c0b9745f3d09", + "actual": "d88a0b51fd72d80fdf7da1344082e7799f481fb114a246ed4e66c0b9745f3d09" + }, + "upstream/omg/src/omg/pipeline/condition_sequence.py": { + "expected": "e831393b0b843219520ee8a836b7cbc0b406295e60947cba1a443f1f78a81377", + "actual": "e831393b0b843219520ee8a836b7cbc0b406295e60947cba1a443f1f78a81377" + }, + "upstream/omg/src/omg/pipeline/human.py": { + "expected": "d5be4338bc69308450c18331da4cf9a93fc9e02bd5596dce92a3e58248ff14d0", + "actual": "d5be4338bc69308450c18331da4cf9a93fc9e02bd5596dce92a3e58248ff14d0" + }, + "upstream/omg/src/omg/pipeline/planner.py": { + "expected": "fa7aa07eebf9f56287301a6ab13fb7a42bf8b310e152b0916033019d7d55ce0c", + "actual": "fa7aa07eebf9f56287301a6ab13fb7a42bf8b310e152b0916033019d7d55ce0c" + }, + "upstream/omg/src/omg/robots/__init__.py": { + "expected": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", + "actual": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855" + }, + "upstream/omg/src/omg/robots/g1/__init__.py": { + "expected": "dc46237fccb843612112274586e9e00dee6a0e472f4ce6b93da267b64f8f4c19", + "actual": "dc46237fccb843612112274586e9e00dee6a0e472f4ce6b93da267b64f8f4c19" + }, + "upstream/omg/src/omg/robots/g1/constants.py": { + "expected": "1f2d0fd495de8765349ad24bdb8e8b327ffdf80c16b13c665fabfe8e9b1a39ea", + "actual": "1f2d0fd495de8765349ad24bdb8e8b327ffdf80c16b13c665fabfe8e9b1a39ea" + }, + "upstream/omg/src/omg/robots/g1/kinematics.py": { + "expected": "c8adbae0d96ef85ba948e27f1fb16bed35d8529c08a6d71d1943d31644585db4", + "actual": "c8adbae0d96ef85ba948e27f1fb16bed35d8529c08a6d71d1943d31644585db4" + }, + "upstream/omg/src/omg/runtime/__init__.py": { + "expected": "9728869f63d93ada09a0f77fc10e98f2add27f4ab84e999d228becca102315fe", + "actual": "9728869f63d93ada09a0f77fc10e98f2add27f4ab84e999d228becca102315fe" + }, + "upstream/omg/src/omg/runtime/checkpoint_io.py": { + "expected": "1b9ce178d3e5c1bba39c77d3098ea7e64788d8c81c48acdebaec09cc1be88e5e", + "actual": "1b9ce178d3e5c1bba39c77d3098ea7e64788d8c81c48acdebaec09cc1be88e5e" + }, + "upstream/omg/src/omg/runtime/onnx_providers.py": { + "expected": "5ed87e442ebf9c56b4065fc011c89d7745ac5bf163158fbf0732db160d0e5673", + "actual": "5ed87e442ebf9c56b4065fc011c89d7745ac5bf163158fbf0732db160d0e5673" + }, + "upstream/omg/src/omg/utils/rotation_conversions.py": { + "expected": "627ead38f2e15b0001eea501f6be09e7b6c309c21f1938ef419f65f8e1eb7ea5", + "actual": "627ead38f2e15b0001eea501f6be09e7b6c309c21f1938ef419f65f8e1eb7ea5" + }, + "train.py": { + "expected": "0724b3db6d374259db9f6378e9b663ef8438e5b969093f3b4d1e46ee56456ddc", + "actual": "61f49739e14275a671dc3343afcf2d6b225da547a952f963037d545afe680dd4" + } + } +} \ No newline at end of file diff --git a/handoff-v1/inventory/omg-b-source-audit.json b/handoff-v1/inventory/omg-b-source-audit.json new file mode 100644 index 0000000000000000000000000000000000000000..06777061d0f085f5cb91e5280d63473d68fa33d4 --- /dev/null +++ b/handoff-v1/inventory/omg-b-source-audit.json @@ -0,0 +1,562 @@ +{ + "hfd_bridge/storage.py": { + "expected": "8b553d204c19ddad7cec223c96f42146865266bc578e7207ad80eb9c7c8f396d", + "actual": "02c4fe4eed5f6d17bf140568af365b0abc8e5843df018d65553828890058a1fa" + }, + "hfd_bridge/__init__.py": { + "expected": "885528eb35015189a0bcd2dc62b60c61878efcf959caef68ff69e93501cafd5c", + "actual": "885528eb35015189a0bcd2dc62b60c61878efcf959caef68ff69e93501cafd5c" + }, + "hfd_bridge/loader.py": { + "expected": "13772137dc5cbb036bd5b5e4ee6ffe8cb45ba8d033ad580887b4dd63b253ac9b", + "actual": "13772137dc5cbb036bd5b5e4ee6ffe8cb45ba8d033ad580887b4dd63b253ac9b" + }, + "hfd_bridge/sessions.py": { + "expected": "c271aecf69e52acaee178b1381e465f22af82732f606d4e0bd5f891e2bb86076", + "actual": "c271aecf69e52acaee178b1381e465f22af82732f606d4e0bd5f891e2bb86076" + }, + "hfd_bridge/asr.py": { + "expected": "4baaf4faecfdce2054844a41326760220af301eb11d6fad892bed8c7380dceb0", + "actual": "4baaf4faecfdce2054844a41326760220af301eb11d6fad892bed8c7380dceb0" + }, + "hfd_bridge/training.py": { + "expected": "b513a632bcf010c83c23a35bcfd09e189149ff8c0d7d667d8966ec09e6e55970", + "actual": "b513a632bcf010c83c23a35bcfd09e189149ff8c0d7d667d8966ec09e6e55970" + }, + "hfd_bridge/native.py": { + "expected": "72ca9f675c2dfc7a2e304191da7b4d4016cb6bc226ccc22f9e7d0aeb2889c9f4", + "actual": "72ca9f675c2dfc7a2e304191da7b4d4016cb6bc226ccc22f9e7d0aeb2889c9f4" + }, + "hfd_bridge/paths.py": { + "expected": "2aaee82ecad571dcc1186c275e270af6c854c670db4ae354ee984d23a50b8bf3", + "actual": "2aaee82ecad571dcc1186c275e270af6c854c670db4ae354ee984d23a50b8bf3" + }, + "hfd_bridge/configured_data.py": { + "expected": "f916e0f9d4cdd2a4120fe94a9f1dfe59bbd24606d71ce1941023f04f97f181cd", + "actual": "f916e0f9d4cdd2a4120fe94a9f1dfe59bbd24606d71ce1941023f04f97f181cd" + }, + "hfd_bridge/textop_checks.py": { + "expected": "00662346781a638298cfa344b9ce8a58b6e8bba11261cd5783fdca956f4df7c8", + "actual": "00662346781a638298cfa344b9ce8a58b6e8bba11261cd5783fdca956f4df7c8" + }, + "hfd_bridge/omg_checks.py": { + "expected": "1c80361412892538400cc225ab28aa03d5357ccab7a53e3a6d46deb5da9d32e8", + "actual": "094d997043d4e0d868261b7f25b9452fca832dfa48572aeac4a3f2ba9f42d172" + }, + "hfd_bridge/textop_condition.py": { + "expected": "4fca0f4a02df3b843a908b28b64766877585aa3300977359d21fdb3adb5a3c28", + "actual": "4fca0f4a02df3b843a908b28b64766877585aa3300977359d21fdb3adb5a3c28" + }, + "hfd_bridge/joints.py": { + "expected": "7827c82ca3463a026db6007c41a66528e1294485b4acb081f27f09449da2d154", + "actual": "7827c82ca3463a026db6007c41a66528e1294485b4acb081f27f09449da2d154" + }, + "hfd_bridge/audio_condition.py": { + "expected": "cab6795bdddc36aa63d4f54c3c8e8aaee7958c86882abe35ae76f67a79b4b945", + "actual": "cab6795bdddc36aa63d4f54c3c8e8aaee7958c86882abe35ae76f67a79b4b945" + }, + "hfd_bridge/batch_asr.py": { + "expected": "9d4f0ba1d382f45c7a0adf562bc008485fce7cd00e86e29607528676f503f8f1", + "actual": "9d4f0ba1d382f45c7a0adf562bc008485fce7cd00e86e29607528676f503f8f1" + }, + "hfd_bridge/ondemand.py": { + "expected": "ddf6c76a6b3d736de9dbf5e27c9d8d6e9384156cee68fc96bba0ce08237aa7c5", + "actual": "940339aa16a6fa25a27ace23a956017f7325813b0cd2a49f76e6f1fb1d014cc0" + }, + "hfd_bridge/ondemand_data.py": { + "expected": "24c61e36623eaca0bdc2f7bff2c8ce307946c4a5afef18f0427b96c06c9ed57e", + "actual": "24c61e36623eaca0bdc2f7bff2c8ce307946c4a5afef18f0427b96c06c9ed57e" + }, + "hfd_bridge/prefetch.py": { + "expected": "b227251cc6470f980231999bbecbeb1fc583c754eda56e70d4f3c84ed1845d64", + "actual": "b227251cc6470f980231999bbecbeb1fc583c754eda56e70d4f3c84ed1845d64" + }, + "hfd_bridge/human_sessions.py": { + "expected": "8b93bbe9450fef5eabd05c827b596a41dfa4748aeaaf93a63849877f5e8df38e", + "actual": "0461ea66b1c324ec38e9e833ee8eb19b95a24c3f478b094fb600dcdc3c53cdea" + }, + "hfd_bridge/genmo.py": { + "expected": "3c898d2828267f38f8ba9f41aae62c4a52cbae87cf63dd2cef2b6634f7160060", + "actual": "3c898d2828267f38f8ba9f41aae62c4a52cbae87cf63dd2cef2b6634f7160060" + }, + "hfd_bridge/genmo_audio.py": { + "expected": "1cdfa6ebc19a75af7540b2280f5dc9c3ecc584ed1c386c434de32e3777020c6f", + "actual": "1cdfa6ebc19a75af7540b2280f5dc9c3ecc584ed1c386c434de32e3777020c6f" + }, + "hfd_bridge/clocks.py": { + "expected": "27054c0b2d3e7d4087eb8fc3c59c6bcf5a581ca2e3022341e306732dcbe92637", + "actual": "27054c0b2d3e7d4087eb8fc3c59c6bcf5a581ca2e3022341e306732dcbe92637" + }, + "hfd_bridge/semantics.py": { + "expected": "42ed59d58609095b8053d0598a9c575d52af9b8508e699162743cc11599f03b4", + "actual": "42ed59d58609095b8053d0598a9c575d52af9b8508e699162743cc11599f03b4" + }, + "hfd_bridge/reference_text.py": { + "expected": "50da93fed3be2bad05035bf978c9be41ba7b48aaf3ce7c43a879519e7386a121", + "actual": "50da93fed3be2bad05035bf978c9be41ba7b48aaf3ce7c43a879519e7386a121" + }, + "hfd_bridge/reference_runtime.py": { + "expected": "2b7fe129152d0f7ea3a4a4e3f5fc9ae799e805138a0d76487e9f59065fe7d975", + "actual": "2b7fe129152d0f7ea3a4a4e3f5fc9ae799e805138a0d76487e9f59065fe7d975" + }, + "hfd_bridge/resume.py": { + "expected": "e8f9c06323415cbdf011272efa4504f12e0c44079d4648a92277f27a5e5d0cc3", + "actual": "e4e373802e0dd21e39fa2d899eff524080ce1735ca86eefbf4ef2080f170ffbf" + }, + "hfd_bridge/resume_lightning.py": { + "expected": "69e4c6bf11a199f3a979227eb551870c44e371fcae0af924ec64d2d58d3ce7a6", + "actual": "69e4c6bf11a199f3a979227eb551870c44e371fcae0af924ec64d2d58d3ce7a6" + }, + "hfd_bridge/index_cache.py": { + "expected": "d75cbcb0105448c61c841e525a1f17e63c55c83ad9201104e8118e22f7728b17", + "actual": "d75cbcb0105448c61c841e525a1f17e63c55c83ad9201104e8118e22f7728b17" + }, + "hfd_bridge/genmo_data.py": { + "expected": "4b102d1eecb0af9757a76d9600ede345c3f21f92166276a22cbf8e144fa9a38c", + "actual": "4b102d1eecb0af9757a76d9600ede345c3f21f92166276a22cbf8e144fa9a38c" + }, + "hfd_bridge/genmo_validation.py": { + "expected": "fa265cf9df51a4b36db977be945547dd78bdf105db9aed8ccba940b1a977f5e4", + "actual": "fa265cf9df51a4b36db977be945547dd78bdf105db9aed8ccba940b1a977f5e4" + }, + "hfd_bridge/kimodo_data.py": { + "expected": "35b3b7d8fc8f36fe65eb311e6e65e667fe1fdbcdfc23e58f37fb75c1c9e9feae", + "actual": "35b3b7d8fc8f36fe65eb311e6e65e667fe1fdbcdfc23e58f37fb75c1c9e9feae" + }, + "hfd_bridge/kimodo.py": { + "expected": "196f8ff99caec99470b04adb5e24efac83b5f4239aa2fdc3083b4438fa4cf460", + "actual": "081a1047b853a2ee143aad2a5699f33cbecd0b4f93a07ad085760e16665f20ce" + }, + "hfd_bridge/kimodo_recovery.py": { + "expected": "b9177ff13b55b753430dd14d46778b723d2a03ce8927313c246c5d7ce780809b", + "actual": "b9177ff13b55b753430dd14d46778b723d2a03ce8927313c246c5d7ce780809b" + }, + "hfd_bridge/kimodo_closed_loop.py": { + "expected": "38d7e022b1450a715b35542cd23a70b1491a6dbd307dbb4be730b9f562100f91", + "actual": "38d7e022b1450a715b35542cd23a70b1491a6dbd307dbb4be730b9f562100f91" + }, + "hfd_bridge/kimodo_causal_rep.py": { + "expected": "c0c40003a3d36ec35e8490fa1165ac88723178b15a0e118d2d1d4c9f7ed7c537", + "actual": "c0c40003a3d36ec35e8490fa1165ac88723178b15a0e118d2d1d4c9f7ed7c537" + }, + "hfd_bridge/kimodo_diffusion_training.py": { + "expected": "078fe8cf1ee39d910a6ca5ef94d67f20c4398cfe505895a089acb80c05336539", + "actual": "078fe8cf1ee39d910a6ca5ef94d67f20c4398cfe505895a089acb80c05336539" + }, + "hfd_bridge/kimodo_text_worker.py": { + "expected": "a275c121a9364ae85b6233a67f3acb63ceddb2c9da490c2ea7d476f4cc710751", + "actual": "a275c121a9364ae85b6233a67f3acb63ceddb2c9da490c2ea7d476f4cc710751" + }, + "hfd_bridge/kimodo_full_data.py": { + "expected": "c62b4bb2673f1875792edb3403347f76bb52fa2360d92a9c1091eae10d43f905", + "actual": "c62b4bb2673f1875792edb3403347f76bb52fa2360d92a9c1091eae10d43f905" + }, + "hfd_bridge/kimodo_checkpoint.py": { + "expected": "b9cd24df225c91a6212c354983e3f51acb495373747f303c6096000af37f7d59", + "actual": "b9cd24df225c91a6212c354983e3f51acb495373747f303c6096000af37f7d59" + }, + "hfd_bridge/kimodo_finetuning.py": { + "expected": "a2c54f04af2b516c14d00025c8b695093ff09da6f04b0b73eb3269817bc0a544", + "actual": "a2c54f04af2b516c14d00025c8b695093ff09da6f04b0b73eb3269817bc0a544" + }, + "hfd_bridge/reference_assets.py": { + "expected": "e977a1d804807629fe925d511acdd25b488a11acc3767c0234020af8f699a43c", + "actual": "e977a1d804807629fe925d511acdd25b488a11acc3767c0234020af8f699a43c" + }, + "hfd_bridge/omg_model_loader.py": { + "expected": "f972080dd6140e1574124be615c9351001b51a73c5774b63bd1c063671bd56aa", + "actual": "f972080dd6140e1574124be615c9351001b51a73c5774b63bd1c063671bd56aa" + }, + "hfd_bridge/genmo_geometry.py": { + "expected": "4d3e912a858b8d3d4ddf73941c3120d5304ed70187c02fc969730816e155b2cb", + "actual": "4d3e912a858b8d3d4ddf73941c3120d5304ed70187c02fc969730816e155b2cb" + }, + "hfd_bridge/genmo_training.py": { + "expected": "8276d4f266b44e2c112e14021eb29c442753f2de07abf29e9adb321226dc87d8", + "actual": "556100608e466ca801abd623d24af30eebe7f9e3ac598c6709f813f07d5d45e9" + }, + "hfd_bridge/genmo_prefetch.py": { + "expected": "3375352f7668dfd5c76ff629853aa8e4d0bcd2e2294a6e96932280f2eafa68e4", + "actual": "0f00c9c5c951f53a0b55cad0c71d470c9312381e75ec31f756f606cf20309141" + }, + "hfd_bridge/hymotion_rep.py": { + "expected": "ee2f8282720d311b98d3ceb8cc72f63219622b5f61a138af6eb3221181a89f7a", + "actual": "ee2f8282720d311b98d3ceb8cc72f63219622b5f61a138af6eb3221181a89f7a" + }, + "hfd_bridge/hymotion_model.py": { + "expected": "db0968384b87929c6495b87647a37e92e13729a94a34f6a5347f4d23ed05be5c", + "actual": "ed29d2d03b247ed681a09dc230697885c5ab8fa8b30f82c4a827aa8fe0b41597" + }, + "hfd_bridge/hymotion_data.py": { + "expected": "06170526a18b60fa486f084864e2e65f2608274ed967ef367b29de269f8c9147", + "actual": "79503602a9a4de4f67931fa944a198d6f438b531a363ffbe545ecc4dd54f97e7" + }, + "hfd_bridge/hymotion_text.py": { + "expected": "f0238ffc3656e2a90a26360c5e9bc8c6de3d2d6be0af6e87787c83237914e98b", + "actual": "f0238ffc3656e2a90a26360c5e9bc8c6de3d2d6be0af6e87787c83237914e98b" + }, + "hfd_bridge/hymotion_rollout.py": { + "expected": "a7fefe8d72f9e3614103632d68d1d63ce3ee072650739683131f5315d0a41181", + "actual": "6e27c4e24bd42a211df7be2a0271337af4e658d66e5d303633912bbd40313517" + }, + "hfd_bridge/hymotion_validation.py": { + "expected": "cf6fd4b14fdbafc65a8a65abaade620b43d5b1f95e7a7a6194a21247f2ad4127", + "actual": "6043090f19315c9fde677168fed06dca1fb390b559a7b9668e0ba46799074cd2" + }, + "hfd_bridge/hymotion_training.py": { + "expected": "67f40e79cc43950f6876f8c0c2c67fe1f606f1b67e692a7dd239457fa4b96a2c", + "actual": "792b76cfc35f58b0c7c94b40a660f00c4971fff1d978ffd0c923b9f26a5fa733" + }, + "hfd_bridge/hymotion_clock.py": { + "expected": "1eaf969e9198f6902759457488ff8c80fc20c973b72388443b3ef0af7c8282a2", + "actual": null + }, + "hfd_bridge/hymotion_recovery.py": { + "expected": "e47b841f85dfc19b37a646a6cf6e865dd2642718073a3419a97195f16024c379", + "actual": null + }, + "hfd_bridge/human_eval.py": { + "expected": "c943c332315f16f41cb53beee8303f2d94a4b522e678cfd346e8a6c9ec203b71", + "actual": "c943c332315f16f41cb53beee8303f2d94a4b522e678cfd346e8a6c9ec203b71" + }, + "hfd_bridge/omg_rollout.py": { + "expected": "24ecdec4a2c698bd8cdc2b4ac6e1c0ad960c252b5a9f9a15d57a7716545fc4d7", + "actual": "24ecdec4a2c698bd8cdc2b4ac6e1c0ad960c252b5a9f9a15d57a7716545fc4d7" + }, + "hfd_bridge/final_split.py": { + "expected": "d0f0adef291d17d3b73477258b5de74bb8ed477ca1952c9b83ddb115efd34305", + "actual": "d0f0adef291d17d3b73477258b5de74bb8ed477ca1952c9b83ddb115efd34305" + }, + "hfd_bridge/final_third_party.py": { + "expected": "3926a0f72d419bb5006d8a62136a8f5ffd5ddca439c39dd7768f7cf03422d716", + "actual": "3926a0f72d419bb5006d8a62136a8f5ffd5ddca439c39dd7768f7cf03422d716" + }, + "hfd_bridge/replay_numerics.py": { + "expected": "827c70aa9b2308ac33b6fccee2c4dc88df58d6b89d902d8c215adfdf918ab231", + "actual": "827c70aa9b2308ac33b6fccee2c4dc88df58d6b89d902d8c215adfdf918ab231" + }, + "hfd_bridge/textop_text_cache.py": { + "expected": "3d5c8a187e461144aa97cc539661e6e90b475800379c0c096b243bcd337a0e38", + "actual": "3d5c8a187e461144aa97cc539661e6e90b475800379c0c096b243bcd337a0e38" + }, + "upstream/omg/download_data.py": { + "expected": "a247118ce874864affc5153a23d94788871d3b6d327bcc33ceeba45279d7732c", + "actual": "a247118ce874864affc5153a23d94788871d3b6d327bcc33ceeba45279d7732c" + }, + "upstream/omg/download_models.py": { + "expected": "811132811b35321174adc69a754cf64f9437eeba5d19f6ad02625ef9d9f73b57", + "actual": "811132811b35321174adc69a754cf64f9437eeba5d19f6ad02625ef9d9f73b57" + }, + "upstream/omg/a.py": { + "expected": "66e1c68c0a631d6fef0d7d13948299671baf72ec813fe025c1cac1e222fd4c63", + "actual": "66e1c68c0a631d6fef0d7d13948299671baf72ec813fe025c1cac1e222fd4c63" + }, + "upstream/omg/smoke.py": { + "expected": "8008c82e92250f148512e3127b939df3aded7025a89473738d024444d21edcfd", + "actual": "8008c82e92250f148512e3127b939df3aded7025a89473738d024444d21edcfd" + }, + "upstream/omg/validate_model.py": { + "expected": "36023f739cd8dfe24179f9247b5de04ca425d7a0d73c99b166bbeaea778a975f", + "actual": "36023f739cd8dfe24179f9247b5de04ca425d7a0d73c99b166bbeaea778a975f" + }, + "upstream/omg/validate_setup.py": { + "expected": "4e086d4615c621386cb54885aad0ede341c24a71a1668c3e5386f368c02f3741", + "actual": "4e086d4615c621386cb54885aad0ede341c24a71a1668c3e5386f368c02f3741" + }, + "upstream/omg/paper_table.py": { + "expected": "97c16c1c044639362f46cac439159384a1d79dc80a90c839c5e2ce471d8060d2", + "actual": "97c16c1c044639362f46cac439159384a1d79dc80a90c839c5e2ce471d8060d2" + }, + "upstream/omg/paper_main_table.py": { + "expected": "8f95f11b00882e4574df37e2f44d5843a15073004f0b9f78450895c65939196f", + "actual": "8f95f11b00882e4574df37e2f44d5843a15073004f0b9f78450895c65939196f" + }, + "upstream/omg/scripts/prepare_omg_inputs.py": { + "expected": "99bbc364b5f380ae7c2cc0fa53657722f896d3613d8b98937881d06197beccb4", + "actual": "99bbc364b5f380ae7c2cc0fa53657722f896d3613d8b98937881d06197beccb4" + }, + "upstream/omg/scripts/extract_omg_paper_row.py": { + "expected": "f64cb31b5a19efba6f3e1b1771d3d5175f6610392a588d118014e98790b82503", + "actual": "f64cb31b5a19efba6f3e1b1771d3d5175f6610392a588d118014e98790b82503" + }, + "upstream/omg/scripts/psi0_omg_generate.py": { + "expected": "38ccbd4c8f1d62b470193ae6e3598445f22408e2df6edc6b267ed46b4873f579", + "actual": "38ccbd4c8f1d62b470193ae6e3598445f22408e2df6edc6b267ed46b4873f579" + }, + "upstream/omg/scripts/verify_omg_artifact_match.py": { + "expected": "46815bd446b956583ae7feeb0ed03a58b6dde8cc78ecb4a1d4fd1a3a6b737ee7", + "actual": "46815bd446b956583ae7feeb0ed03a58b6dde8cc78ecb4a1d4fd1a3a6b737ee7" + }, + "upstream/omg/src/omg/__init__.py": { + "expected": "e06eff3bd6a146e0082518eb791164e26b80b548f9bdcd21c9f0aca1220bca34", + "actual": "e06eff3bd6a146e0082518eb791164e26b80b548f9bdcd21c9f0aca1220bca34" + }, + "upstream/omg/src/omg/callbacks/__init__.py": { + "expected": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", + "actual": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855" + }, + "upstream/omg/src/omg/callbacks/divergence_guard.py": { + "expected": "c1ec2773dddda3bd800247bca0abd3eca46aeae3c78ebae8050b98739292b61a", + "actual": "c1ec2773dddda3bd800247bca0abd3eca46aeae3c78ebae8050b98739292b61a" + }, + "upstream/omg/src/omg/callbacks/global_step_progress.py": { + "expected": "bd21180de0140a38fa2601f51b56c94aef346c757c140786e88b907720f3eb15", + "actual": "bd21180de0140a38fa2601f51b56c94aef346c757c140786e88b907720f3eb15" + }, + "upstream/omg/src/omg/callbacks/grad_norm.py": { + "expected": "c41e7e0bac01c984c85c38baf6de4b9e10ded95875a79ddc1c9d85ded5b9b742", + "actual": "c41e7e0bac01c984c85c38baf6de4b9e10ded95875a79ddc1c9d85ded5b9b742" + }, + "upstream/omg/src/omg/callbacks/speed_timer.py": { + "expected": "29ae0ec76a0545576bc31f240a4beab74f5c5e9ae50a0078c1ed46b8326fc360", + "actual": "29ae0ec76a0545576bc31f240a4beab74f5c5e9ae50a0078c1ed46b8326fc360" + }, + "upstream/omg/src/omg/callbacks/training_diagnostics.py": { + "expected": "6e348e5cca9e690379cdb7a3364678b666d7c53d0a7405d5269ca713574a6917", + "actual": "6e348e5cca9e690379cdb7a3364678b666d7c53d0a7405d5269ca713574a6917" + }, + "upstream/omg/src/omg/cli/__init__.py": { + "expected": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", + "actual": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855" + }, + "upstream/omg/src/omg/cli/data/__init__.py": { + "expected": "052be341b9c1b05585d3a27073a741d768974cfdb2a91659a687ed1d3d6291f9", + "actual": "052be341b9c1b05585d3a27073a741d768974cfdb2a91659a687ed1d3d6291f9" + }, + "upstream/omg/src/omg/cli/data/materialize_episode_cache.py": { + "expected": "c8e61cbcf17f55e17b47ad2a3e5c2aa0919d9f5bcc2a93edbf5a519e79f9793a", + "actual": "c8e61cbcf17f55e17b47ad2a3e5c2aa0919d9f5bcc2a93edbf5a519e79f9793a" + }, + "upstream/omg/src/omg/cli/data/validate_episode_cache.py": { + "expected": "6840e28c774964583d88c6402c201ab36c0150ec67139403a66782364173c878", + "actual": "6840e28c774964583d88c6402c201ab36c0150ec67139403a66782364173c878" + }, + "upstream/omg/src/omg/cli/generation/__init__.py": { + "expected": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", + "actual": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855" + }, + "upstream/omg/src/omg/cli/generation/compute_stats.py": { + "expected": "849ed1ac16612e23e09a887e5a030c5e0c496610e9afe882ce98e99d570bd6ca", + "actual": "849ed1ac16612e23e09a887e5a030c5e0c496610e9afe882ce98e99d570bd6ca" + }, + "upstream/omg/src/omg/cli/generation/train.py": { + "expected": "a39cc0c6cfccd7b6660bbd9ab4f88f90bc634bb95ce30e22492c8a6cb74a7c8c", + "actual": "a39cc0c6cfccd7b6660bbd9ab4f88f90bc634bb95ce30e22492c8a6cb74a7c8c" + }, + "upstream/omg/src/omg/cli/generation/benchmark.py": { + "expected": "ff813f344953da4b02388837f5a510230f8d4430393a88f205261faaf28c5dee", + "actual": "ff813f344953da4b02388837f5a510230f8d4430393a88f205261faaf28c5dee" + }, + "upstream/omg/src/omg/cli/generation/physical_benchmark.py": { + "expected": "f48b2b9d2d312c44d0e33eebf0c7fd95d8f2ce65a871fe253ef8ff9ceee30dd0", + "actual": "f48b2b9d2d312c44d0e33eebf0c7fd95d8f2ce65a871fe253ef8ff9ceee30dd0" + }, + "upstream/omg/src/omg/cli/generation/generate.py": { + "expected": "66eda839732e76a3b0fc39a767e6a238aa93bbf13ef2f3ddb6e1f9021948bc4a", + "actual": "66eda839732e76a3b0fc39a767e6a238aa93bbf13ef2f3ddb6e1f9021948bc4a" + }, + "upstream/omg/src/omg/cli/generation/export_onnx.py": { + "expected": "9cacee46fea5e18f8a9f9ffd52b25110f89e80be20d21acce85fe066392c00c1", + "actual": "9cacee46fea5e18f8a9f9ffd52b25110f89e80be20d21acce85fe066392c00c1" + }, + "upstream/omg/src/omg/core/__init__.py": { + "expected": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", + "actual": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855" + }, + "upstream/omg/src/omg/core/logging.py": { + "expected": "86c57564f7d8614d71d95e0ce377724642c8c3aa44501d9c317c33bf9ba7dce7", + "actual": "86c57564f7d8614d71d95e0ce377724642c8c3aa44501d9c317c33bf9ba7dce7" + }, + "upstream/omg/src/omg/core/paths.py": { + "expected": "6bb254b7219e70b58fc8581479857bfb7ee77a9f8937faf5dcec784dd1173a4d", + "actual": "6bb254b7219e70b58fc8581479857bfb7ee77a9f8937faf5dcec784dd1173a4d" + }, + "upstream/omg/src/omg/core/tensor.py": { + "expected": "998e6f24948983c15f3257f055a452e07b96cc0f13b357cb4881f4545bddb7d0", + "actual": "998e6f24948983c15f3257f055a452e07b96cc0f13b357cb4881f4545bddb7d0" + }, + "upstream/omg/src/omg/data/__init__.py": { + "expected": "4e8cd9d824b5b834a673011d1cd95cf0227a1deb60ee917ba85ec364d51db21f", + "actual": "4e8cd9d824b5b834a673011d1cd95cf0227a1deb60ee917ba85ec364d51db21f" + }, + "upstream/omg/src/omg/data/datamodule.py": { + "expected": "94cbaf84b019d559dd7810494694c156803d318f23a5ec95549315c59ea795d6", + "actual": "94cbaf84b019d559dd7810494694c156803d318f23a5ec95549315c59ea795d6" + }, + "upstream/omg/src/omg/data/episode_cache.py": { + "expected": "96639991cb75a3db602c608d929e4f742caef615166198e66731c406f068207e", + "actual": "96639991cb75a3db602c608d929e4f742caef615166198e66731c406f068207e" + }, + "upstream/omg/src/omg/data/episode_cache_inspect.py": { + "expected": "b7f43c2a5f7a1c63dcb9c243e79018ebff3ecce056fba22ac9b1948e754feb20", + "actual": "b7f43c2a5f7a1c63dcb9c243e79018ebff3ecce056fba22ac9b1948e754feb20" + }, + "upstream/omg/src/omg/data/lerobot_dataset.py": { + "expected": "78172eb344a37cab8457133a957eb25691ad78054f71db0c39171f241be12831", + "actual": "78172eb344a37cab8457133a957eb25691ad78054f71db0c39171f241be12831" + }, + "upstream/omg/src/omg/data/windowing.py": { + "expected": "88710f0c24d6c696670b702b9ce77609c59acd2106389fa769a00d39a98f2afb", + "actual": "88710f0c24d6c696670b702b9ce77609c59acd2106389fa769a00d39a98f2afb" + }, + "upstream/omg/src/omg/generation/__init__.py": { + "expected": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", + "actual": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855" + }, + "upstream/omg/src/omg/generation/architecture.py": { + "expected": "f77548e52394c78b813ad36dcd5fb8c8981ba393e5c723126491df6c171aeabf", + "actual": "f77548e52394c78b813ad36dcd5fb8c8981ba393e5c723126491df6c171aeabf" + }, + "upstream/omg/src/omg/generation/conditions/__init__.py": { + "expected": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", + "actual": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855" + }, + "upstream/omg/src/omg/generation/conditions/t5.py": { + "expected": "67c0f834106ff860e612a978bdba6ef503d2cf0dfb24e831af0bba723e95c890", + "actual": "67c0f834106ff860e612a978bdba6ef503d2cf0dfb24e831af0bba723e95c890" + }, + "upstream/omg/src/omg/generation/denoisers/__init__.py": { + "expected": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", + "actual": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855" + }, + "upstream/omg/src/omg/generation/denoisers/interface.py": { + "expected": "c1f8b5f2a93e6cf0a4b5df730f9f744c95d9b58bc5245e3cb15176a61065e1dd", + "actual": "c1f8b5f2a93e6cf0a4b5df730f9f744c95d9b58bc5245e3cb15176a61065e1dd" + }, + "upstream/omg/src/omg/generation/denoisers/transformer.py": { + "expected": "ae968a16ba6193607721f5c0c1910122818797407df3ee0f0fafa3728c0e398a", + "actual": "ae968a16ba6193607721f5c0c1910122818797407df3ee0f0fafa3728c0e398a" + }, + "upstream/omg/src/omg/generation/diffusion/__init__.py": { + "expected": "fab4af0ec5b7b38892ffb309d6073841de00f8f37e5b4c2da1657e1b30c92e1c", + "actual": "fab4af0ec5b7b38892ffb309d6073841de00f8f37e5b4c2da1657e1b30c92e1c" + }, + "upstream/omg/src/omg/generation/diffusion/forcing.py": { + "expected": "935b2f38db1d1d3c496df0b03cadaeb032ef1539522169f2731b87a8724362b5", + "actual": "935b2f38db1d1d3c496df0b03cadaeb032ef1539522169f2731b87a8724362b5" + }, + "upstream/omg/src/omg/generation/diffusion/guided.py": { + "expected": "3fd42c7ae9d01497223b08421ae3becc0a7be81846c1cf08c8bb944d8949b870", + "actual": "3fd42c7ae9d01497223b08421ae3becc0a7be81846c1cf08c8bb944d8949b870" + }, + "upstream/omg/src/omg/generation/diffusion/objectives.py": { + "expected": "db05bcbc32268e5d21298e19130e9e8f4b23eff7d1b679886306cd7c94195675", + "actual": "db05bcbc32268e5d21298e19130e9e8f4b23eff7d1b679886306cd7c94195675" + }, + "upstream/omg/src/omg/generation/diffusion/schedules.py": { + "expected": "a06c3f3ab1d81e07f680d44f8fb9ef56e8bd2a6e4c5e4ac911c316217f6e92e0", + "actual": "a06c3f3ab1d81e07f680d44f8fb9ef56e8bd2a6e4c5e4ac911c316217f6e92e0" + }, + "upstream/omg/src/omg/generation/export/__init__.py": { + "expected": "0ee7ed60aa43848dad0a4ba0e5dd29d815aab6dc37f4ae1c46d8e56a2ebc1770", + "actual": "0ee7ed60aa43848dad0a4ba0e5dd29d815aab6dc37f4ae1c46d8e56a2ebc1770" + }, + "upstream/omg/src/omg/generation/export/onnx.py": { + "expected": "168ad808867669b69cd69a419d5b8d2c3fd124c69c20daeca53b6fc59a0a21dd", + "actual": "168ad808867669b69cd69a419d5b8d2c3fd124c69c20daeca53b6fc59a0a21dd" + }, + "upstream/omg/src/omg/generation/losses/__init__.py": { + "expected": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", + "actual": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855" + }, + "upstream/omg/src/omg/generation/losses/contact.py": { + "expected": "75927b00fc5584988f2ecedbcfbba9af3ba80206966b0a91642ec133db5e4267", + "actual": "75927b00fc5584988f2ecedbcfbba9af3ba80206966b0a91642ec133db5e4267" + }, + "upstream/omg/src/omg/generation/losses/fk.py": { + "expected": "75927b00fc5584988f2ecedbcfbba9af3ba80206966b0a91642ec133db5e4267", + "actual": "75927b00fc5584988f2ecedbcfbba9af3ba80206966b0a91642ec133db5e4267" + }, + "upstream/omg/src/omg/generation/losses/motion.py": { + "expected": "00ee200046658dafcae2b9ec439ed0c3203ab533f4dd9d4a8475541df74c2fa5", + "actual": "00ee200046658dafcae2b9ec439ed0c3203ab533f4dd9d4a8475541df74c2fa5" + }, + "upstream/omg/src/omg/generation/losses/terrain.py": { + "expected": "75927b00fc5584988f2ecedbcfbba9af3ba80206966b0a91642ec133db5e4267", + "actual": "75927b00fc5584988f2ecedbcfbba9af3ba80206966b0a91642ec133db5e4267" + }, + "upstream/omg/src/omg/generation/metrics/__init__.py": { + "expected": "f686c86d57bab24aa1dd27cfae2bbab88757d7909e74cd7d4b9525a892ba025e", + "actual": "f686c86d57bab24aa1dd27cfae2bbab88757d7909e74cd7d4b9525a892ba025e" + }, + "upstream/omg/src/omg/generation/metrics/physical.py": { + "expected": "7aa2d8fc7d29b9d3ec836c3d23d0b5a7fc64a1eea5359c7300f6e6e191904f29", + "actual": "7aa2d8fc7d29b9d3ec836c3d23d0b5a7fc64a1eea5359c7300f6e6e191904f29" + }, + "upstream/omg/src/omg/generation/models/__init__.py": { + "expected": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", + "actual": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855" + }, + "upstream/omg/src/omg/generation/models/motion_generator.py": { + "expected": "05947c8e73646f634e9bbedf0a60f2becaf1ded21b9d1093f1084eaaefef2a3b", + "actual": "05947c8e73646f634e9bbedf0a60f2becaf1ded21b9d1093f1084eaaefef2a3b" + }, + "upstream/omg/src/omg/motion/__init__.py": { + "expected": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", + "actual": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855" + }, + "upstream/omg/src/omg/motion/feature_codec.py": { + "expected": "3d148816baf239dbc938210c959954c3c2fac4c6d9914ad48b851976ed7e2960", + "actual": "3d148816baf239dbc938210c959954c3c2fac4c6d9914ad48b851976ed7e2960" + }, + "upstream/omg/src/omg/motion/representation.py": { + "expected": "457328078b2a2652e06c39e5fa06c8a0d8ba9223dcf24e508a33cdd1806f45a7", + "actual": "457328078b2a2652e06c39e5fa06c8a0d8ba9223dcf24e508a33cdd1806f45a7" + }, + "upstream/omg/src/omg/pipeline/__init__.py": { + "expected": "3806048000a9bdc918ace57c68aabcd3c56bb503c4998b103d1593553a23a895", + "actual": "3806048000a9bdc918ace57c68aabcd3c56bb503c4998b103d1593553a23a895" + }, + "upstream/omg/src/omg/pipeline/audio.py": { + "expected": "d88a0b51fd72d80fdf7da1344082e7799f481fb114a246ed4e66c0b9745f3d09", + "actual": "d88a0b51fd72d80fdf7da1344082e7799f481fb114a246ed4e66c0b9745f3d09" + }, + "upstream/omg/src/omg/pipeline/condition_sequence.py": { + "expected": "e831393b0b843219520ee8a836b7cbc0b406295e60947cba1a443f1f78a81377", + "actual": "e831393b0b843219520ee8a836b7cbc0b406295e60947cba1a443f1f78a81377" + }, + "upstream/omg/src/omg/pipeline/human.py": { + "expected": "d5be4338bc69308450c18331da4cf9a93fc9e02bd5596dce92a3e58248ff14d0", + "actual": "d5be4338bc69308450c18331da4cf9a93fc9e02bd5596dce92a3e58248ff14d0" + }, + "upstream/omg/src/omg/pipeline/planner.py": { + "expected": "fa7aa07eebf9f56287301a6ab13fb7a42bf8b310e152b0916033019d7d55ce0c", + "actual": "fa7aa07eebf9f56287301a6ab13fb7a42bf8b310e152b0916033019d7d55ce0c" + }, + "upstream/omg/src/omg/robots/__init__.py": { + "expected": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", + "actual": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855" + }, + "upstream/omg/src/omg/robots/g1/__init__.py": { + "expected": "dc46237fccb843612112274586e9e00dee6a0e472f4ce6b93da267b64f8f4c19", + "actual": "dc46237fccb843612112274586e9e00dee6a0e472f4ce6b93da267b64f8f4c19" + }, + "upstream/omg/src/omg/robots/g1/constants.py": { + "expected": "1f2d0fd495de8765349ad24bdb8e8b327ffdf80c16b13c665fabfe8e9b1a39ea", + "actual": "1f2d0fd495de8765349ad24bdb8e8b327ffdf80c16b13c665fabfe8e9b1a39ea" + }, + "upstream/omg/src/omg/robots/g1/kinematics.py": { + "expected": "c8adbae0d96ef85ba948e27f1fb16bed35d8529c08a6d71d1943d31644585db4", + "actual": "c8adbae0d96ef85ba948e27f1fb16bed35d8529c08a6d71d1943d31644585db4" + }, + "upstream/omg/src/omg/runtime/__init__.py": { + "expected": "9728869f63d93ada09a0f77fc10e98f2add27f4ab84e999d228becca102315fe", + "actual": "9728869f63d93ada09a0f77fc10e98f2add27f4ab84e999d228becca102315fe" + }, + "upstream/omg/src/omg/runtime/checkpoint_io.py": { + "expected": "1b9ce178d3e5c1bba39c77d3098ea7e64788d8c81c48acdebaec09cc1be88e5e", + "actual": "1b9ce178d3e5c1bba39c77d3098ea7e64788d8c81c48acdebaec09cc1be88e5e" + }, + "upstream/omg/src/omg/runtime/onnx_providers.py": { + "expected": "5ed87e442ebf9c56b4065fc011c89d7745ac5bf163158fbf0732db160d0e5673", + "actual": "5ed87e442ebf9c56b4065fc011c89d7745ac5bf163158fbf0732db160d0e5673" + }, + "upstream/omg/src/omg/utils/rotation_conversions.py": { + "expected": "627ead38f2e15b0001eea501f6be09e7b6c309c21f1938ef419f65f8e1eb7ea5", + "actual": "627ead38f2e15b0001eea501f6be09e7b6c309c21f1938ef419f65f8e1eb7ea5" + }, + "train.py": { + "expected": "0724b3db6d374259db9f6378e9b663ef8438e5b969093f3b4d1e46ee56456ddc", + "actual": "61f49739e14275a671dc3343afcf2d6b225da547a952f963037d545afe680dd4" + } +} \ No newline at end of file diff --git a/handoff-v1/inventory/omg-environment-lock.txt b/handoff-v1/inventory/omg-environment-lock.txt new file mode 100644 index 0000000000000000000000000000000000000000..16a94c8f01340df124fbabfc1c7bb3e13524c38e --- /dev/null +++ b/handoff-v1/inventory/omg-environment-lock.txt @@ -0,0 +1,119 @@ +absl-py==2.5.0 +aiohappyeyeballs==2.7.1 +aiohttp==3.14.3 +aiosignal==1.4.0 +annotated-types==0.8.0 +antlr4-python3-runtime==4.9.3 +anyio==4.15.1 +async-timeout==5.0.1 +attrs==26.1.0 +certifi==2026.7.22 +charset-normalizer==3.5.2 +click==8.5.0 +cloudpickle==3.1.2 +coloredlogs==15.0.1 +contourpy==1.3.2 +cycler==0.12.1 +datasets==4.8.5 +dill==0.4.1 +etils==1.13.0 +exceptiongroup==1.3.1 +filelock==4.0.7 +flatbuffers==25.12.19 +fonttools==4.65.0 +frozenlist==1.8.0 +fsspec==2026.2.0 +gitdb==4.0.12 +GitPython==3.2.0 +glfw==2.10.2 +h11==0.16.0 +hf-xet==1.6.0 +httpcore==1.0.9 +httpx==0.28.1 +huggingface_hub==0.36.2 +humanfriendly==10.0 +hydra-core==1.3.0 +idna==3.20 +ImageIO==2.38.0 +imageio-ffmpeg==0.6.0 +importlib_resources==7.1.0 +iniconfig==2.3.0 +Jinja2==3.1.6 +joblib==1.6.0 +kiwisolver==1.5.1 +lightning-utilities==0.15.3 +MarkupSafe==3.0.3 +matplotlib==3.10.9 +ml_dtypes==0.5.4 +mpmath==1.3.0 +mujoco==3.7.0 +multidict==6.9.1 +multiprocess==0.70.19 +networkx==3.4.2 +numpy==1.26.4 +nvidia-cublas-cu12==12.8.3.14 +nvidia-cuda-cupti-cu12==12.8.57 +nvidia-cuda-nvrtc-cu12==12.8.61 +nvidia-cuda-runtime-cu12==12.8.57 +nvidia-cudnn-cu12==9.7.1.26 +nvidia-cufft-cu12==11.3.3.41 +nvidia-cufile-cu12==1.13.0.11 +nvidia-curand-cu12==10.3.9.55 +nvidia-cusolver-cu12==11.7.2.55 +nvidia-cusparse-cu12==12.5.7.53 +nvidia-cusparselt-cu12==0.6.3 +nvidia-nccl-cu12==2.26.2 +nvidia-nvjitlink-cu12==12.8.61 +nvidia-nvtx-cu12==12.8.55 +omegaconf==2.3.0 +# Editable install with no version control (omg==0.1.0) +-e /root/autodl-tmp/models/OMG/OMG +onnx==1.23.1 +onnx-ir==1.0.0 +onnxruntime-gpu==1.23.2 +onnxscript==0.7.2 +opencv-python==4.11.0.86 +packaging==26.3 +pandas==2.3.3 +pillow==12.3.0 +platformdirs==4.12.2 +pluggy==1.6.0 +propcache==0.5.4 +protobuf==7.36.2 +pyarrow==21.0.0 +pydantic==2.13.5 +pydantic_core==2.46.5 +Pygments==2.21.0 +PyOpenGL==3.1.10 +pyparsing==3.3.3 +pytest==9.1.1 +python-dateutil==2.9.0.post0 +pytorch-lightning==2.6.6 +pytz==2026.4 +PyYAML==6.0.3 +pyzmq==26.4.0 +regex==2026.9.29 +requests==2.34.2 +sacremoses==0.2.0 +safetensors==0.8.0 +scipy==1.15.3 +sentencepiece==0.2.2 +sentry-sdk==2.71.0 +six==1.17.0 +smmap==5.0.3 +sympy==1.14.0 +tokenizers==0.22.2 +tomli==2.4.1 +torch==2.7.1+cu128 +torchmetrics==1.9.0 +tqdm==4.70.1 +transformers==4.57.6 +triton==3.3.1 +typing-inspection==0.4.4 +typing_extensions==4.16.0 +tzdata==2026.4 +urllib3==2.8.0 +wandb==0.26.1 +xxhash==4.0.1 +yarl==1.25.1 +zipp==4.1.0 diff --git a/handoff-v1/inventory/omg-evaluation-checkpoint-index.json b/handoff-v1/inventory/omg-evaluation-checkpoint-index.json new file mode 100644 index 0000000000000000000000000000000000000000..c3e1f73869870e4a25f8e9e826d55d53ffe979ce --- /dev/null +++ b/handoff-v1/inventory/omg-evaluation-checkpoint-index.json @@ -0,0 +1,2321 @@ +[ + { + "original_file": "/root/autodl-tmp/runs/hfd-omg-english-split20261026/code_sha256.json", + "archive_file": "root/autodl-tmp/runs/hfd-omg-english-split20261026/code_sha256.json", + "sha256": "5bfff102fd74e5bc490fc9cec803675674e351f1bafe9a76bc7d39061f47e773", + "references": [ + { + "key": "/hfd_bridge/configured_data.py", + "value": "f916e0f9d4cdd2a4120fe94a9f1dfe59bbd24606d71ce1941023f04f97f181cd" + }, + { + "key": "/hfd_bridge/kimodo_checkpoint.py", + "value": "b9cd24df225c91a6212c354983e3f51acb495373747f303c6096000af37f7d59" + }, + { + "key": "/upstream/omg/src/omg/callbacks/global_step_progress.py", + "value": "bd21180de0140a38fa2601f51b56c94aef346c757c140786e88b907720f3eb15" + }, + { + "key": "/upstream/omg/src/omg/runtime/checkpoint_io.py", + "value": "1b9ce178d3e5c1bba39c77d3098ea7e64788d8c81c48acdebaec09cc1be88e5e" + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/hfd-omg-english-split20261026/data_manifest.json", + "archive_file": "root/autodl-tmp/runs/hfd-omg-english-split20261026/data_manifest.json", + "sha256": "7d2a9006fa5d2751b7406bc763957f70eabb6a5d4571bdbc457f2a1541574500", + "references": [ + { + "key": "/native_options/config", + "value": "/root/autodl-tmp/hfd-baseline-portable/configs/sessions.yaml" + }, + { + "key": "/native_options/seed", + "value": 42 + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/hfd-omg-english-split20261026/resume_history/20261028T104559302890Z/code_sha256.json", + "archive_file": "root/autodl-tmp/runs/hfd-omg-english-split20261026/resume_history/20261028T104559302890Z/code_sha256.json", + "sha256": "bcda10212c2aa6c3146ef16a0591bc724f42d4b054631355b9944beaaa66d6db", + "references": [ + { + "key": "/hfd_bridge/configured_data.py", + "value": "f916e0f9d4cdd2a4120fe94a9f1dfe59bbd24606d71ce1941023f04f97f181cd" + }, + { + "key": "/hfd_bridge/kimodo_checkpoint.py", + "value": "b9cd24df225c91a6212c354983e3f51acb495373747f303c6096000af37f7d59" + }, + { + "key": "/upstream/omg/src/omg/callbacks/global_step_progress.py", + "value": "bd21180de0140a38fa2601f51b56c94aef346c757c140786e88b907720f3eb15" + }, + { + "key": "/upstream/omg/src/omg/runtime/checkpoint_io.py", + "value": "1b9ce178d3e5c1bba39c77d3098ea7e64788d8c81c48acdebaec09cc1be88e5e" + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/hfd-omg-english-split20261026/resume_history/20261028T104559302890Z/data_manifest.json", + "archive_file": "root/autodl-tmp/runs/hfd-omg-english-split20261026/resume_history/20261028T104559302890Z/data_manifest.json", + "sha256": "7d2a9006fa5d2751b7406bc763957f70eabb6a5d4571bdbc457f2a1541574500", + "references": [ + { + "key": "/native_options/config", + "value": "/root/autodl-tmp/hfd-baseline-portable/configs/sessions.yaml" + }, + { + "key": "/native_options/seed", + "value": 42 + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/hfd-omg-english-split20261026/resume_history/20261028T104559302890Z/previous-training_summary.json", + "archive_file": "root/autodl-tmp/runs/hfd-omg-english-split20261026/resume_history/20261028T104559302890Z/previous-training_summary.json", + "sha256": "11a499adedb64d2d9ccb31563e09d5cb61c998d95b8720b21ee0f326b290e281", + "references": [ + { + "key": "/global_step", + "value": 258 + }, + { + "key": "/diagnostic_validation_only/step", + "value": 258 + }, + { + "key": "/diagnostic_validation_only/formal_first_validation_step", + "value": 2000 + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/hfd-omg-english-split20261026/resume_history/20261028T104559302890Z/resume.json", + "archive_file": "root/autodl-tmp/runs/hfd-omg-english-split20261026/resume_history/20261028T104559302890Z/resume.json", + "sha256": "f0c1731bdb0ef77163e11a3da10abfcd021ccb3d13f75bf53d4edfd0955fcf9a", + "references": [ + { + "key": "/checkpoint", + "value": "/root/autodl-tmp/runs/hfd-omg-english-split20261026/checkpoints/last.ckpt" + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/hfd-omg-english-split20261026/resume_history/20261028T112852271305Z/code_sha256.json", + "archive_file": "root/autodl-tmp/runs/hfd-omg-english-split20261026/resume_history/20261028T112852271305Z/code_sha256.json", + "sha256": "56508be82d73d1d174aa8af0e4781cd4ad86402fdc8b9fb65845eb20b6b103b0", + "references": [ + { + "key": "/hfd_bridge/configured_data.py", + "value": "f916e0f9d4cdd2a4120fe94a9f1dfe59bbd24606d71ce1941023f04f97f181cd" + }, + { + "key": "/hfd_bridge/kimodo_checkpoint.py", + "value": "b9cd24df225c91a6212c354983e3f51acb495373747f303c6096000af37f7d59" + }, + { + "key": "/upstream/omg/src/omg/callbacks/global_step_progress.py", + "value": "bd21180de0140a38fa2601f51b56c94aef346c757c140786e88b907720f3eb15" + }, + { + "key": "/upstream/omg/src/omg/runtime/checkpoint_io.py", + "value": "1b9ce178d3e5c1bba39c77d3098ea7e64788d8c81c48acdebaec09cc1be88e5e" + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/hfd-omg-english-split20261026/resume_history/20261028T112852271305Z/data_manifest.json", + "archive_file": "root/autodl-tmp/runs/hfd-omg-english-split20261026/resume_history/20261028T112852271305Z/data_manifest.json", + "sha256": "a2bb9afaaaf479e28b365d377e53a206581f4d120176aaf3254ae278de725f28", + "references": [ + { + "key": "/native_options/config", + "value": "/root/autodl-tmp/hfd-baseline-portable/configs/sessions.yaml" + }, + { + "key": "/native_options/seed", + "value": 42 + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/hfd-omg-english-split20261026/resume_history/20261028T112852271305Z/previous-training_summary.json", + "archive_file": "root/autodl-tmp/runs/hfd-omg-english-split20261026/resume_history/20261028T112852271305Z/previous-training_summary.json", + "sha256": "11a499adedb64d2d9ccb31563e09d5cb61c998d95b8720b21ee0f326b290e281", + "references": [ + { + "key": "/global_step", + "value": 258 + }, + { + "key": "/diagnostic_validation_only/step", + "value": 258 + }, + { + "key": "/diagnostic_validation_only/formal_first_validation_step", + "value": 2000 + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/hfd-omg-english-split20261026/resume_history/20261028T112852271305Z/resume.json", + "archive_file": "root/autodl-tmp/runs/hfd-omg-english-split20261026/resume_history/20261028T112852271305Z/resume.json", + "sha256": "e9a0bd4c15223a5a30c64b747b506046f6bcad44327763ac221d8b6a1bc10c7f", + "references": [ + { + "key": "/checkpoint", + "value": "/root/autodl-tmp/runs/hfd-omg-english-split20261026/checkpoints/step-00001024.ckpt" + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/hfd-omg-english-split20261026/train_native_manifest.json", + "archive_file": "root/autodl-tmp/runs/hfd-omg-english-split20261026/train_native_manifest.json", + "sha256": "882482e5f1e8246b46500866d296b91b0df5ae24f7ff4f214337c2424aa027a9", + "references": [ + { + "key": "/seed", + "value": 42 + }, + { + "key": "/sampling_seed", + "value": 9000043 + }, + { + "key": "/native_config/audio_sample_step_frames", + "value": 20 + }, + { + "key": "/native_config/action_sample_step_frames", + "value": 1 + }, + { + "key": "/native_config/max_motion_steps", + "value": 0 + }, + { + "key": "/native_config/initial_idle_max_steps", + "value": 7 + }, + { + "key": "/native_config/control_post_idle_min_steps", + "value": 3 + }, + { + "key": "/native_config/control_post_idle_max_steps", + "value": 20 + }, + { + "key": "/native_config/max_speech_onset_steps", + "value": 3 + }, + { + "key": "/native_config/listen_reaction_steps", + "value": 1 + }, + { + "key": "/native_config/max_stream_steps", + "value": 1919 + }, + { + "key": "/final_release/seed_policy", + "value": "recorded_per_session" + }, + { + "key": "/third_party/final_release/seed_policy", + "value": "recorded_per_session" + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/hfd-omg-english-split20261026/training_summary.json", + "archive_file": "root/autodl-tmp/runs/hfd-omg-english-split20261026/training_summary.json", + "sha256": "ffe8940d4c02f6da226be40683577be090ab894c2bdecd5b10ee42eb02d1ab61", + "references": [ + { + "key": "/global_step", + "value": 100000 + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/hfd-omg-english-split20261026/val_native_manifest.json", + "archive_file": "root/autodl-tmp/runs/hfd-omg-english-split20261026/val_native_manifest.json", + "sha256": "2e37c5039c519b05cb83c20731eedac8db4aba23b7dfd185ec4c291f8fa91ef9", + "references": [ + { + "key": "/seed", + "value": 42 + }, + { + "key": "/sampling_seed", + "value": 9000049 + }, + { + "key": "/native_config/audio_sample_step_frames", + "value": 20 + }, + { + "key": "/native_config/action_sample_step_frames", + "value": 1 + }, + { + "key": "/native_config/max_motion_steps", + "value": 0 + }, + { + "key": "/native_config/initial_idle_max_steps", + "value": 7 + }, + { + "key": "/native_config/control_post_idle_min_steps", + "value": 3 + }, + { + "key": "/native_config/control_post_idle_max_steps", + "value": 20 + }, + { + "key": "/native_config/max_speech_onset_steps", + "value": 3 + }, + { + "key": "/native_config/listen_reaction_steps", + "value": 1 + }, + { + "key": "/native_config/max_stream_steps", + "value": 1919 + }, + { + "key": "/final_release/seed_policy", + "value": "recorded_per_session" + }, + { + "key": "/third_party/final_release/seed_policy", + "value": "recorded_per_session" + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/hfd-omg-l-english-split20261026/code_sha256.json", + "archive_file": "root/autodl-tmp/runs/hfd-omg-l-english-split20261026/code_sha256.json", + "sha256": "8acf726745f82c38b84e00660983bcfbbf63efe22997957af3194414e61194e1", + "references": [ + { + "key": "/hfd_bridge/configured_data.py", + "value": "f916e0f9d4cdd2a4120fe94a9f1dfe59bbd24606d71ce1941023f04f97f181cd" + }, + { + "key": "/hfd_bridge/kimodo_checkpoint.py", + "value": "b9cd24df225c91a6212c354983e3f51acb495373747f303c6096000af37f7d59" + }, + { + "key": "/upstream/omg/src/omg/callbacks/global_step_progress.py", + "value": "bd21180de0140a38fa2601f51b56c94aef346c757c140786e88b907720f3eb15" + }, + { + "key": "/upstream/omg/src/omg/runtime/checkpoint_io.py", + "value": "1b9ce178d3e5c1bba39c77d3098ea7e64788d8c81c48acdebaec09cc1be88e5e" + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/hfd-omg-l-english-split20261026/data_manifest.json", + "archive_file": "root/autodl-tmp/runs/hfd-omg-l-english-split20261026/data_manifest.json", + "sha256": "a2bb9afaaaf479e28b365d377e53a206581f4d120176aaf3254ae278de725f28", + "references": [ + { + "key": "/native_options/config", + "value": "/root/autodl-tmp/hfd-baseline-portable/configs/sessions.yaml" + }, + { + "key": "/native_options/seed", + "value": 42 + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/hfd-omg-l-english-split20261026/train_native_manifest.json", + "archive_file": "root/autodl-tmp/runs/hfd-omg-l-english-split20261026/train_native_manifest.json", + "sha256": "882482e5f1e8246b46500866d296b91b0df5ae24f7ff4f214337c2424aa027a9", + "references": [ + { + "key": "/seed", + "value": 42 + }, + { + "key": "/sampling_seed", + "value": 9000043 + }, + { + "key": "/native_config/audio_sample_step_frames", + "value": 20 + }, + { + "key": "/native_config/action_sample_step_frames", + "value": 1 + }, + { + "key": "/native_config/max_motion_steps", + "value": 0 + }, + { + "key": "/native_config/initial_idle_max_steps", + "value": 7 + }, + { + "key": "/native_config/control_post_idle_min_steps", + "value": 3 + }, + { + "key": "/native_config/control_post_idle_max_steps", + "value": 20 + }, + { + "key": "/native_config/max_speech_onset_steps", + "value": 3 + }, + { + "key": "/native_config/listen_reaction_steps", + "value": 1 + }, + { + "key": "/native_config/max_stream_steps", + "value": 1919 + }, + { + "key": "/final_release/seed_policy", + "value": "recorded_per_session" + }, + { + "key": "/third_party/final_release/seed_policy", + "value": "recorded_per_session" + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/hfd-omg-l-english-split20261026/training_summary.json", + "archive_file": "root/autodl-tmp/runs/hfd-omg-l-english-split20261026/training_summary.json", + "sha256": "c20f1dc4b6db9063f7bc69a645724fc43b4c03bf2c4f11f8d1369e5a4b1f1e73", + "references": [ + { + "key": "/global_step", + "value": 100000 + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/hfd-omg-l-english-split20261026/val_native_manifest.json", + "archive_file": "root/autodl-tmp/runs/hfd-omg-l-english-split20261026/val_native_manifest.json", + "sha256": "2e37c5039c519b05cb83c20731eedac8db4aba23b7dfd185ec4c291f8fa91ef9", + "references": [ + { + "key": "/seed", + "value": 42 + }, + { + "key": "/sampling_seed", + "value": 9000049 + }, + { + "key": "/native_config/audio_sample_step_frames", + "value": 20 + }, + { + "key": "/native_config/action_sample_step_frames", + "value": 1 + }, + { + "key": "/native_config/max_motion_steps", + "value": 0 + }, + { + "key": "/native_config/initial_idle_max_steps", + "value": 7 + }, + { + "key": "/native_config/control_post_idle_min_steps", + "value": 3 + }, + { + "key": "/native_config/control_post_idle_max_steps", + "value": 20 + }, + { + "key": "/native_config/max_speech_onset_steps", + "value": 3 + }, + { + "key": "/native_config/listen_reaction_steps", + "value": 1 + }, + { + "key": "/native_config/max_stream_steps", + "value": 1919 + }, + { + "key": "/final_release/seed_policy", + "value": "recorded_per_session" + }, + { + "key": "/third_party/final_release/seed_policy", + "value": "recorded_per_session" + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/hfd-omg-l-scale-smoke-20261102/code_sha256.json", + "archive_file": "root/autodl-tmp/runs/hfd-omg-l-scale-smoke-20261102/code_sha256.json", + "sha256": "8acf726745f82c38b84e00660983bcfbbf63efe22997957af3194414e61194e1", + "references": [ + { + "key": "/hfd_bridge/configured_data.py", + "value": "f916e0f9d4cdd2a4120fe94a9f1dfe59bbd24606d71ce1941023f04f97f181cd" + }, + { + "key": "/hfd_bridge/kimodo_checkpoint.py", + "value": "b9cd24df225c91a6212c354983e3f51acb495373747f303c6096000af37f7d59" + }, + { + "key": "/upstream/omg/src/omg/callbacks/global_step_progress.py", + "value": "bd21180de0140a38fa2601f51b56c94aef346c757c140786e88b907720f3eb15" + }, + { + "key": "/upstream/omg/src/omg/runtime/checkpoint_io.py", + "value": "1b9ce178d3e5c1bba39c77d3098ea7e64788d8c81c48acdebaec09cc1be88e5e" + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/hfd-omg-l-scale-smoke-20261102/data_manifest.json", + "archive_file": "root/autodl-tmp/runs/hfd-omg-l-scale-smoke-20261102/data_manifest.json", + "sha256": "a2bb9afaaaf479e28b365d377e53a206581f4d120176aaf3254ae278de725f28", + "references": [ + { + "key": "/native_options/config", + "value": "/root/autodl-tmp/hfd-baseline-portable/configs/sessions.yaml" + }, + { + "key": "/native_options/seed", + "value": 42 + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/hfd-omg-l-scale-smoke-20261102/pipeline_check.json", + "archive_file": "root/autodl-tmp/runs/hfd-omg-l-scale-smoke-20261102/pipeline_check.json", + "sha256": "792a9bf193a2f9bd6d702770733ef10d9947819796511dd2d4628b4663faf1a9", + "references": [ + { + "key": "/resumed_from_step", + "value": 0 + }, + { + "key": "/native_checkpoint", + "value": "/autodl-fs/data/pangxinlei/models/OMG/hfd-l-scale-smoke-20261102/pipeline.ckpt" + }, + { + "key": "/strict_checkpoint_reload", + "value": true + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/hfd-omg-l-scale-smoke-20261102/resume_history/20261102T100222051409Z/code_sha256.json", + "archive_file": "root/autodl-tmp/runs/hfd-omg-l-scale-smoke-20261102/resume_history/20261102T100222051409Z/code_sha256.json", + "sha256": "8acf726745f82c38b84e00660983bcfbbf63efe22997957af3194414e61194e1", + "references": [ + { + "key": "/hfd_bridge/configured_data.py", + "value": "f916e0f9d4cdd2a4120fe94a9f1dfe59bbd24606d71ce1941023f04f97f181cd" + }, + { + "key": "/hfd_bridge/kimodo_checkpoint.py", + "value": "b9cd24df225c91a6212c354983e3f51acb495373747f303c6096000af37f7d59" + }, + { + "key": "/upstream/omg/src/omg/callbacks/global_step_progress.py", + "value": "bd21180de0140a38fa2601f51b56c94aef346c757c140786e88b907720f3eb15" + }, + { + "key": "/upstream/omg/src/omg/runtime/checkpoint_io.py", + "value": "1b9ce178d3e5c1bba39c77d3098ea7e64788d8c81c48acdebaec09cc1be88e5e" + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/hfd-omg-l-scale-smoke-20261102/resume_history/20261102T100222051409Z/data_manifest.json", + "archive_file": "root/autodl-tmp/runs/hfd-omg-l-scale-smoke-20261102/resume_history/20261102T100222051409Z/data_manifest.json", + "sha256": "a2bb9afaaaf479e28b365d377e53a206581f4d120176aaf3254ae278de725f28", + "references": [ + { + "key": "/native_options/config", + "value": "/root/autodl-tmp/hfd-baseline-portable/configs/sessions.yaml" + }, + { + "key": "/native_options/seed", + "value": 42 + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/hfd-omg-l-scale-smoke-20261102/resume_history/20261102T100222051409Z/previous-pipeline_check.json", + "archive_file": "root/autodl-tmp/runs/hfd-omg-l-scale-smoke-20261102/resume_history/20261102T100222051409Z/previous-pipeline_check.json", + "sha256": "792a9bf193a2f9bd6d702770733ef10d9947819796511dd2d4628b4663faf1a9", + "references": [ + { + "key": "/resumed_from_step", + "value": 0 + }, + { + "key": "/native_checkpoint", + "value": "/autodl-fs/data/pangxinlei/models/OMG/hfd-l-scale-smoke-20261102/pipeline.ckpt" + }, + { + "key": "/strict_checkpoint_reload", + "value": true + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/hfd-omg-l-scale-smoke-20261102/resume_history/20261102T100222051409Z/previous-training_summary.json", + "archive_file": "root/autodl-tmp/runs/hfd-omg-l-scale-smoke-20261102/resume_history/20261102T100222051409Z/previous-training_summary.json", + "sha256": "0aa904632c7ed30cc7269f9316159a255c7373c6c5a17146b816de0e3735853e", + "references": [ + { + "key": "/global_step", + "value": 2 + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/hfd-omg-l-scale-smoke-20261102/resume_history/20261102T100222051409Z/resume.json", + "archive_file": "root/autodl-tmp/runs/hfd-omg-l-scale-smoke-20261102/resume_history/20261102T100222051409Z/resume.json", + "sha256": "425e8c296f68cc885c0643e97135aee368f6ea985da50278528a357478026a4b", + "references": [ + { + "key": "/checkpoint", + "value": "/root/autodl-tmp/runs/hfd-omg-l-scale-smoke-20261102/last.ckpt" + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/hfd-omg-l-scale-smoke-20261102/train_native_manifest.json", + "archive_file": "root/autodl-tmp/runs/hfd-omg-l-scale-smoke-20261102/train_native_manifest.json", + "sha256": "882482e5f1e8246b46500866d296b91b0df5ae24f7ff4f214337c2424aa027a9", + "references": [ + { + "key": "/seed", + "value": 42 + }, + { + "key": "/sampling_seed", + "value": 9000043 + }, + { + "key": "/native_config/audio_sample_step_frames", + "value": 20 + }, + { + "key": "/native_config/action_sample_step_frames", + "value": 1 + }, + { + "key": "/native_config/max_motion_steps", + "value": 0 + }, + { + "key": "/native_config/initial_idle_max_steps", + "value": 7 + }, + { + "key": "/native_config/control_post_idle_min_steps", + "value": 3 + }, + { + "key": "/native_config/control_post_idle_max_steps", + "value": 20 + }, + { + "key": "/native_config/max_speech_onset_steps", + "value": 3 + }, + { + "key": "/native_config/listen_reaction_steps", + "value": 1 + }, + { + "key": "/native_config/max_stream_steps", + "value": 1919 + }, + { + "key": "/final_release/seed_policy", + "value": "recorded_per_session" + }, + { + "key": "/third_party/final_release/seed_policy", + "value": "recorded_per_session" + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/hfd-omg-l-scale-smoke-20261102/training_summary.json", + "archive_file": "root/autodl-tmp/runs/hfd-omg-l-scale-smoke-20261102/training_summary.json", + "sha256": "036994afa5bcc21b816f501277cb9e48f850b2eec18ca0b8efba963acd2c97b7", + "references": [ + { + "key": "/global_step", + "value": 4 + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/hfd-omg-l-scale-smoke-20261102/val_native_manifest.json", + "archive_file": "root/autodl-tmp/runs/hfd-omg-l-scale-smoke-20261102/val_native_manifest.json", + "sha256": "2e37c5039c519b05cb83c20731eedac8db4aba23b7dfd185ec4c291f8fa91ef9", + "references": [ + { + "key": "/seed", + "value": 42 + }, + { + "key": "/sampling_seed", + "value": 9000049 + }, + { + "key": "/native_config/audio_sample_step_frames", + "value": 20 + }, + { + "key": "/native_config/action_sample_step_frames", + "value": 1 + }, + { + "key": "/native_config/max_motion_steps", + "value": 0 + }, + { + "key": "/native_config/initial_idle_max_steps", + "value": 7 + }, + { + "key": "/native_config/control_post_idle_min_steps", + "value": 3 + }, + { + "key": "/native_config/control_post_idle_max_steps", + "value": 20 + }, + { + "key": "/native_config/max_speech_onset_steps", + "value": 3 + }, + { + "key": "/native_config/listen_reaction_steps", + "value": 1 + }, + { + "key": "/native_config/max_stream_steps", + "value": 1919 + }, + { + "key": "/final_release/seed_policy", + "value": "recorded_per_session" + }, + { + "key": "/third_party/final_release/seed_policy", + "value": "recorded_per_session" + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/hfd-omg-scale-launch-20261102/launch_state.json", + "archive_file": "root/autodl-tmp/runs/hfd-omg-scale-launch-20261102/launch_state.json", + "sha256": "3d2666f48b5c9fcca7aa0963fe8ac87cc85d07ed4eb23ed65f9609f6dfc976aa", + "references": [ + { + "key": "/models/l/config_sha256", + "value": "60f8cfad9443c7d608b3c20a17958749eb64aad757fd3e1546a252c5ef95ef4f" + }, + { + "key": "/models/xl/config_sha256", + "value": "ea268ae329767222bd28defdb9717da48c81d898c8e7b7f7d32833e03bec832b" + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/hfd-omg-scale-launch-20261102/preflight.json", + "archive_file": "root/autodl-tmp/runs/hfd-omg-scale-launch-20261102/preflight.json", + "sha256": "ccd2040b6bc8d5c49723173289a708dd019f95616cbe243951c7910832bb27a0", + "references": [ + { + "key": "/source_weights_sha_verified", + "value": true + }, + { + "key": "/models/l/initial_checkpoint", + "value": "/root/autodl-fs/pangxinlei/models/OMG/official/checkpoints/paper/300m/sstep=055000.ckpt" + }, + { + "key": "/models/xl/initial_checkpoint", + "value": "/root/autodl-fs/pangxinlei/models/OMG/official/checkpoints/paper/500m/sstep=050000.ckpt" + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/hfd-omg-scale-launch-20261102/recovery-l.json", + "archive_file": "root/autodl-tmp/runs/hfd-omg-scale-launch-20261102/recovery-l.json", + "sha256": "d14d8ab001f77c36daa772e7ffe8cc1ae21abdb281752ee3aac800da21af6765", + "references": [ + { + "key": "/from_step", + "value": 2 + }, + { + "key": "/to_step", + "value": 4 + }, + { + "key": "/logical_pool_step", + "value": 4 + }, + { + "key": "/initial_training_checks/resumed_from_step", + "value": 0 + }, + { + "key": "/initial_training_checks/native_checkpoint", + "value": "/autodl-fs/data/pangxinlei/models/OMG/hfd-l-scale-smoke-20261102/pipeline.ckpt" + }, + { + "key": "/initial_training_checks/strict_checkpoint_reload", + "value": true + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/hfd-omg-scale-launch-20261102/recovery-xl.json", + "archive_file": "root/autodl-tmp/runs/hfd-omg-scale-launch-20261102/recovery-xl.json", + "sha256": "010204e7aa929b258f7d146144504181739ed524e3ec07bfc360fb652fa8c6fa", + "references": [ + { + "key": "/from_step", + "value": 2 + }, + { + "key": "/to_step", + "value": 4 + }, + { + "key": "/logical_pool_step", + "value": 4 + }, + { + "key": "/initial_training_checks/resumed_from_step", + "value": 0 + }, + { + "key": "/initial_training_checks/native_checkpoint", + "value": "/autodl-fs/data/pangxinlei/models/OMG/hfd-xl-scale-smoke-20261102/pipeline.ckpt" + }, + { + "key": "/initial_training_checks/strict_checkpoint_reload", + "value": true + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/hfd-omg-scale-launch-20261102/startup-verification.json", + "archive_file": "root/autodl-tmp/runs/hfd-omg-scale-launch-20261102/startup-verification.json", + "sha256": "f1ec72c5e93909dc76b3500355eca61363ff1c3db96c19d6dfd1398ebf93f131", + "references": [ + { + "key": "/models/l/last_metric/step", + "value": 119 + }, + { + "key": "/models/l/checkpoint_root", + "value": "/root/autodl-fs/pangxinlei/models/OMG/hfd-l-english-split20261026" + }, + { + "key": "/models/l/configuration_sha256", + "value": "60f8cfad9443c7d608b3c20a17958749eb64aad757fd3e1546a252c5ef95ef4f" + }, + { + "key": "/models/xl/last_metric/step", + "value": 29 + }, + { + "key": "/models/xl/checkpoint_root", + "value": "/root/autodl-fs/pangxinlei/models/OMG/hfd-xl-english-split20261026" + }, + { + "key": "/models/xl/configuration_sha256", + "value": "ea268ae329767222bd28defdb9717da48c81d898c8e7b7f7d32833e03bec832b" + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/hfd-omg-xl-english-split20261026/code_sha256.json", + "archive_file": "root/autodl-tmp/runs/hfd-omg-xl-english-split20261026/code_sha256.json", + "sha256": "8acf726745f82c38b84e00660983bcfbbf63efe22997957af3194414e61194e1", + "references": [ + { + "key": "/hfd_bridge/configured_data.py", + "value": "f916e0f9d4cdd2a4120fe94a9f1dfe59bbd24606d71ce1941023f04f97f181cd" + }, + { + "key": "/hfd_bridge/kimodo_checkpoint.py", + "value": "b9cd24df225c91a6212c354983e3f51acb495373747f303c6096000af37f7d59" + }, + { + "key": "/upstream/omg/src/omg/callbacks/global_step_progress.py", + "value": "bd21180de0140a38fa2601f51b56c94aef346c757c140786e88b907720f3eb15" + }, + { + "key": "/upstream/omg/src/omg/runtime/checkpoint_io.py", + "value": "1b9ce178d3e5c1bba39c77d3098ea7e64788d8c81c48acdebaec09cc1be88e5e" + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/hfd-omg-xl-english-split20261026/data_manifest.json", + "archive_file": "root/autodl-tmp/runs/hfd-omg-xl-english-split20261026/data_manifest.json", + "sha256": "a2bb9afaaaf479e28b365d377e53a206581f4d120176aaf3254ae278de725f28", + "references": [ + { + "key": "/native_options/config", + "value": "/root/autodl-tmp/hfd-baseline-portable/configs/sessions.yaml" + }, + { + "key": "/native_options/seed", + "value": 42 + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/hfd-omg-xl-english-split20261026/train_native_manifest.json", + "archive_file": "root/autodl-tmp/runs/hfd-omg-xl-english-split20261026/train_native_manifest.json", + "sha256": "882482e5f1e8246b46500866d296b91b0df5ae24f7ff4f214337c2424aa027a9", + "references": [ + { + "key": "/seed", + "value": 42 + }, + { + "key": "/sampling_seed", + "value": 9000043 + }, + { + "key": "/native_config/audio_sample_step_frames", + "value": 20 + }, + { + "key": "/native_config/action_sample_step_frames", + "value": 1 + }, + { + "key": "/native_config/max_motion_steps", + "value": 0 + }, + { + "key": "/native_config/initial_idle_max_steps", + "value": 7 + }, + { + "key": "/native_config/control_post_idle_min_steps", + "value": 3 + }, + { + "key": "/native_config/control_post_idle_max_steps", + "value": 20 + }, + { + "key": "/native_config/max_speech_onset_steps", + "value": 3 + }, + { + "key": "/native_config/listen_reaction_steps", + "value": 1 + }, + { + "key": "/native_config/max_stream_steps", + "value": 1919 + }, + { + "key": "/final_release/seed_policy", + "value": "recorded_per_session" + }, + { + "key": "/third_party/final_release/seed_policy", + "value": "recorded_per_session" + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/hfd-omg-xl-english-split20261026/val_native_manifest.json", + "archive_file": "root/autodl-tmp/runs/hfd-omg-xl-english-split20261026/val_native_manifest.json", + "sha256": "2e37c5039c519b05cb83c20731eedac8db4aba23b7dfd185ec4c291f8fa91ef9", + "references": [ + { + "key": "/seed", + "value": 42 + }, + { + "key": "/sampling_seed", + "value": 9000049 + }, + { + "key": "/native_config/audio_sample_step_frames", + "value": 20 + }, + { + "key": "/native_config/action_sample_step_frames", + "value": 1 + }, + { + "key": "/native_config/max_motion_steps", + "value": 0 + }, + { + "key": "/native_config/initial_idle_max_steps", + "value": 7 + }, + { + "key": "/native_config/control_post_idle_min_steps", + "value": 3 + }, + { + "key": "/native_config/control_post_idle_max_steps", + "value": 20 + }, + { + "key": "/native_config/max_speech_onset_steps", + "value": 3 + }, + { + "key": "/native_config/listen_reaction_steps", + "value": 1 + }, + { + "key": "/native_config/max_stream_steps", + "value": 1919 + }, + { + "key": "/final_release/seed_policy", + "value": "recorded_per_session" + }, + { + "key": "/third_party/final_release/seed_policy", + "value": "recorded_per_session" + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/hfd-omg-xl-scale-smoke-20261102/code_sha256.json", + "archive_file": "root/autodl-tmp/runs/hfd-omg-xl-scale-smoke-20261102/code_sha256.json", + "sha256": "8acf726745f82c38b84e00660983bcfbbf63efe22997957af3194414e61194e1", + "references": [ + { + "key": "/hfd_bridge/configured_data.py", + "value": "f916e0f9d4cdd2a4120fe94a9f1dfe59bbd24606d71ce1941023f04f97f181cd" + }, + { + "key": "/hfd_bridge/kimodo_checkpoint.py", + "value": "b9cd24df225c91a6212c354983e3f51acb495373747f303c6096000af37f7d59" + }, + { + "key": "/upstream/omg/src/omg/callbacks/global_step_progress.py", + "value": "bd21180de0140a38fa2601f51b56c94aef346c757c140786e88b907720f3eb15" + }, + { + "key": "/upstream/omg/src/omg/runtime/checkpoint_io.py", + "value": "1b9ce178d3e5c1bba39c77d3098ea7e64788d8c81c48acdebaec09cc1be88e5e" + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/hfd-omg-xl-scale-smoke-20261102/data_manifest.json", + "archive_file": "root/autodl-tmp/runs/hfd-omg-xl-scale-smoke-20261102/data_manifest.json", + "sha256": "a2bb9afaaaf479e28b365d377e53a206581f4d120176aaf3254ae278de725f28", + "references": [ + { + "key": "/native_options/config", + "value": "/root/autodl-tmp/hfd-baseline-portable/configs/sessions.yaml" + }, + { + "key": "/native_options/seed", + "value": 42 + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/hfd-omg-xl-scale-smoke-20261102/pipeline_check.json", + "archive_file": "root/autodl-tmp/runs/hfd-omg-xl-scale-smoke-20261102/pipeline_check.json", + "sha256": "ca2766a19063379290009672ebcc55d09fc7383a6f1304eccd5c4e0a1ea9a14c", + "references": [ + { + "key": "/resumed_from_step", + "value": 0 + }, + { + "key": "/native_checkpoint", + "value": "/autodl-fs/data/pangxinlei/models/OMG/hfd-xl-scale-smoke-20261102/pipeline.ckpt" + }, + { + "key": "/strict_checkpoint_reload", + "value": true + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/hfd-omg-xl-scale-smoke-20261102/resume_history/20261102T100211467696Z/code_sha256.json", + "archive_file": "root/autodl-tmp/runs/hfd-omg-xl-scale-smoke-20261102/resume_history/20261102T100211467696Z/code_sha256.json", + "sha256": "8acf726745f82c38b84e00660983bcfbbf63efe22997957af3194414e61194e1", + "references": [ + { + "key": "/hfd_bridge/configured_data.py", + "value": "f916e0f9d4cdd2a4120fe94a9f1dfe59bbd24606d71ce1941023f04f97f181cd" + }, + { + "key": "/hfd_bridge/kimodo_checkpoint.py", + "value": "b9cd24df225c91a6212c354983e3f51acb495373747f303c6096000af37f7d59" + }, + { + "key": "/upstream/omg/src/omg/callbacks/global_step_progress.py", + "value": "bd21180de0140a38fa2601f51b56c94aef346c757c140786e88b907720f3eb15" + }, + { + "key": "/upstream/omg/src/omg/runtime/checkpoint_io.py", + "value": "1b9ce178d3e5c1bba39c77d3098ea7e64788d8c81c48acdebaec09cc1be88e5e" + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/hfd-omg-xl-scale-smoke-20261102/resume_history/20261102T100211467696Z/data_manifest.json", + "archive_file": "root/autodl-tmp/runs/hfd-omg-xl-scale-smoke-20261102/resume_history/20261102T100211467696Z/data_manifest.json", + "sha256": "a2bb9afaaaf479e28b365d377e53a206581f4d120176aaf3254ae278de725f28", + "references": [ + { + "key": "/native_options/config", + "value": "/root/autodl-tmp/hfd-baseline-portable/configs/sessions.yaml" + }, + { + "key": "/native_options/seed", + "value": 42 + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/hfd-omg-xl-scale-smoke-20261102/resume_history/20261102T100211467696Z/previous-pipeline_check.json", + "archive_file": "root/autodl-tmp/runs/hfd-omg-xl-scale-smoke-20261102/resume_history/20261102T100211467696Z/previous-pipeline_check.json", + "sha256": "ca2766a19063379290009672ebcc55d09fc7383a6f1304eccd5c4e0a1ea9a14c", + "references": [ + { + "key": "/resumed_from_step", + "value": 0 + }, + { + "key": "/native_checkpoint", + "value": "/autodl-fs/data/pangxinlei/models/OMG/hfd-xl-scale-smoke-20261102/pipeline.ckpt" + }, + { + "key": "/strict_checkpoint_reload", + "value": true + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/hfd-omg-xl-scale-smoke-20261102/resume_history/20261102T100211467696Z/previous-training_summary.json", + "archive_file": "root/autodl-tmp/runs/hfd-omg-xl-scale-smoke-20261102/resume_history/20261102T100211467696Z/previous-training_summary.json", + "sha256": "7d624be40ddf8c76b34ccc9a95c459afefba4f2dc42bebf04c14bb0b3d35b9bc", + "references": [ + { + "key": "/global_step", + "value": 2 + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/hfd-omg-xl-scale-smoke-20261102/resume_history/20261102T100211467696Z/resume.json", + "archive_file": "root/autodl-tmp/runs/hfd-omg-xl-scale-smoke-20261102/resume_history/20261102T100211467696Z/resume.json", + "sha256": "02ab58aa64a205667573f5301751b7e1ea025447393c3111e848fca4f6885694", + "references": [ + { + "key": "/checkpoint", + "value": "/root/autodl-tmp/runs/hfd-omg-xl-scale-smoke-20261102/last.ckpt" + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/hfd-omg-xl-scale-smoke-20261102/train_native_manifest.json", + "archive_file": "root/autodl-tmp/runs/hfd-omg-xl-scale-smoke-20261102/train_native_manifest.json", + "sha256": "882482e5f1e8246b46500866d296b91b0df5ae24f7ff4f214337c2424aa027a9", + "references": [ + { + "key": "/seed", + "value": 42 + }, + { + "key": "/sampling_seed", + "value": 9000043 + }, + { + "key": "/native_config/audio_sample_step_frames", + "value": 20 + }, + { + "key": "/native_config/action_sample_step_frames", + "value": 1 + }, + { + "key": "/native_config/max_motion_steps", + "value": 0 + }, + { + "key": "/native_config/initial_idle_max_steps", + "value": 7 + }, + { + "key": "/native_config/control_post_idle_min_steps", + "value": 3 + }, + { + "key": "/native_config/control_post_idle_max_steps", + "value": 20 + }, + { + "key": "/native_config/max_speech_onset_steps", + "value": 3 + }, + { + "key": "/native_config/listen_reaction_steps", + "value": 1 + }, + { + "key": "/native_config/max_stream_steps", + "value": 1919 + }, + { + "key": "/final_release/seed_policy", + "value": "recorded_per_session" + }, + { + "key": "/third_party/final_release/seed_policy", + "value": "recorded_per_session" + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/hfd-omg-xl-scale-smoke-20261102/training_summary.json", + "archive_file": "root/autodl-tmp/runs/hfd-omg-xl-scale-smoke-20261102/training_summary.json", + "sha256": "1eefd6c02d66a093a3e320e0f51ac791d0289a56671de43919cb3dfb7ba7e310", + "references": [ + { + "key": "/global_step", + "value": 4 + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/hfd-omg-xl-scale-smoke-20261102/val_native_manifest.json", + "archive_file": "root/autodl-tmp/runs/hfd-omg-xl-scale-smoke-20261102/val_native_manifest.json", + "sha256": "2e37c5039c519b05cb83c20731eedac8db4aba23b7dfd185ec4c291f8fa91ef9", + "references": [ + { + "key": "/seed", + "value": 42 + }, + { + "key": "/sampling_seed", + "value": 9000049 + }, + { + "key": "/native_config/audio_sample_step_frames", + "value": 20 + }, + { + "key": "/native_config/action_sample_step_frames", + "value": 1 + }, + { + "key": "/native_config/max_motion_steps", + "value": 0 + }, + { + "key": "/native_config/initial_idle_max_steps", + "value": 7 + }, + { + "key": "/native_config/control_post_idle_min_steps", + "value": 3 + }, + { + "key": "/native_config/control_post_idle_max_steps", + "value": 20 + }, + { + "key": "/native_config/max_speech_onset_steps", + "value": 3 + }, + { + "key": "/native_config/listen_reaction_steps", + "value": 1 + }, + { + "key": "/native_config/max_stream_steps", + "value": 1919 + }, + { + "key": "/final_release/seed_policy", + "value": "recorded_per_session" + }, + { + "key": "/third_party/final_release/seed_policy", + "value": "recorded_per_session" + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/omg-test-split20261026-20261029/ablations/no_audio/inference/report.json", + "archive_file": "root/autodl-tmp/runs/omg-test-split20261026-20261029/ablations/no_audio/inference/report.json", + "sha256": "c1a5ac6b2d56fe505dcb905d7b347e900d5833b6642da2a029d936968434a9ad", + "references": [ + { + "key": "/seed", + "value": 12345 + }, + { + "key": "/checkpoint", + "value": "/root/autodl-tmp/runs/hfd-omg-english-split20261026/best/best.ckpt" + }, + { + "key": "/omg_sampling_steps", + "value": 10 + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/omg-test-split20261026-20261029/ablations/no_text/inference/report.json", + "archive_file": "root/autodl-tmp/runs/omg-test-split20261026-20261029/ablations/no_text/inference/report.json", + "sha256": "254849c879d3c0328e349aa584e97ff0e587ed498719069014c176e1ec0803d3", + "references": [ + { + "key": "/seed", + "value": 12345 + }, + { + "key": "/checkpoint", + "value": "/root/autodl-tmp/runs/hfd-omg-english-split20261026/best/best.ckpt" + }, + { + "key": "/omg_sampling_steps", + "value": 10 + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/omg-test-split20261026-20261029/arguments.json", + "archive_file": "root/autodl-tmp/runs/omg-test-split20261026-20261029/arguments.json", + "sha256": "c87d567479cc5f8e15faceee7e2f07b28d8ea097a94a2d8956e61172679e4e9a", + "references": [ + { + "key": "/checkpoint", + "value": "/root/autodl-tmp/runs/hfd-omg-english-split20261026/best/best.ckpt" + }, + { + "key": "/seed", + "value": 12345 + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/omg-test-split20261026-20261029/audit.json", + "archive_file": "root/autodl-tmp/runs/omg-test-split20261026-20261029/audit.json", + "sha256": "bf2afb588370cb1871c85cf141dfbb4e09469c37ddcc13bcc0e85ef66944610f", + "references": [ + { + "key": "/checkpoint_step", + "value": 100000 + }, + { + "key": "/verification/gt_equal_across_checkpoints", + "value": true + }, + { + "key": "/stop_input_ablation/same_checkpoint_seed_gt_pcm_and_request_clocks", + "value": true + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/omg-test-split20261026-20261029/cases/test-000033-seed2314733246/initial/inference/report.json", + "archive_file": "root/autodl-tmp/runs/omg-test-split20261026-20261029/cases/test-000033-seed2314733246/initial/inference/report.json", + "sha256": "e731045d54a626a993b385be85fe33f1de575479a573ac545f9fba521c8ea199", + "references": [ + { + "key": "/seed", + "value": 12345 + }, + { + "key": "/checkpoint", + "value": "/root/autodl-tmp/models/OMG/official/checkpoints/paper/50m/sstep=090000.ckpt" + }, + { + "key": "/omg_sampling_steps", + "value": 10 + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/omg-test-split20261026-20261029/cases/test-000033-seed2314733246/trained/inference/report.json", + "archive_file": "root/autodl-tmp/runs/omg-test-split20261026-20261029/cases/test-000033-seed2314733246/trained/inference/report.json", + "sha256": "8ff8af049c06b0f7b82cf4ad3bc386265f944738f870f92182a7c4b594bc0bf4", + "references": [ + { + "key": "/seed", + "value": 12345 + }, + { + "key": "/checkpoint", + "value": "/root/autodl-tmp/runs/hfd-omg-english-split20261026/best/best.ckpt" + }, + { + "key": "/omg_sampling_steps", + "value": 10 + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/omg-test-split20261026-20261029/cases/test-000045-seed1597101952/initial/inference/report.json", + "archive_file": "root/autodl-tmp/runs/omg-test-split20261026-20261029/cases/test-000045-seed1597101952/initial/inference/report.json", + "sha256": "bf9d2dec0398331154dbb2213a897996b50d81c133e647683b71e470ef81f0b3", + "references": [ + { + "key": "/seed", + "value": 12345 + }, + { + "key": "/checkpoint", + "value": "/root/autodl-tmp/models/OMG/official/checkpoints/paper/50m/sstep=090000.ckpt" + }, + { + "key": "/omg_sampling_steps", + "value": 10 + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/omg-test-split20261026-20261029/cases/test-000045-seed1597101952/trained/inference/report.json", + "archive_file": "root/autodl-tmp/runs/omg-test-split20261026-20261029/cases/test-000045-seed1597101952/trained/inference/report.json", + "sha256": "c0d075e455b569fe3fedfcc69bbcc0cef0589d154e3a485e7b1f6a7b459c763d", + "references": [ + { + "key": "/seed", + "value": 12345 + }, + { + "key": "/checkpoint", + "value": "/root/autodl-tmp/runs/hfd-omg-english-split20261026/best/best.ckpt" + }, + { + "key": "/omg_sampling_steps", + "value": 10 + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/omg-test-split20261026-20261029/cases/test-000050-seed4281401434/initial/inference/report.json", + "archive_file": "root/autodl-tmp/runs/omg-test-split20261026-20261029/cases/test-000050-seed4281401434/initial/inference/report.json", + "sha256": "751c010a5c95c3066d374b9a1a7d24e45492e7f03c4a9f1ce93d0fe21df524e6", + "references": [ + { + "key": "/seed", + "value": 12345 + }, + { + "key": "/checkpoint", + "value": "/root/autodl-tmp/models/OMG/official/checkpoints/paper/50m/sstep=090000.ckpt" + }, + { + "key": "/omg_sampling_steps", + "value": 10 + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/omg-test-split20261026-20261029/cases/test-000050-seed4281401434/trained/inference/report.json", + "archive_file": "root/autodl-tmp/runs/omg-test-split20261026-20261029/cases/test-000050-seed4281401434/trained/inference/report.json", + "sha256": "2cade7b5ceff3f1586e45c74a584469b41fee1f4f9e4713015cc9724198d2350", + "references": [ + { + "key": "/seed", + "value": 12345 + }, + { + "key": "/checkpoint", + "value": "/root/autodl-tmp/runs/hfd-omg-english-split20261026/best/best.ckpt" + }, + { + "key": "/omg_sampling_steps", + "value": 10 + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/omg-test-split20261026-20261029/cases/test-000063-seed1287447995/initial/inference/report.json", + "archive_file": "root/autodl-tmp/runs/omg-test-split20261026-20261029/cases/test-000063-seed1287447995/initial/inference/report.json", + "sha256": "b6fa36f6fdcece5af4d176ba60195a844c0d02e1392b4b94ad27f476e6728bdc", + "references": [ + { + "key": "/seed", + "value": 12345 + }, + { + "key": "/checkpoint", + "value": "/root/autodl-tmp/models/OMG/official/checkpoints/paper/50m/sstep=090000.ckpt" + }, + { + "key": "/omg_sampling_steps", + "value": 10 + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/omg-test-split20261026-20261029/cases/test-000063-seed1287447995/trained/inference/report.json", + "archive_file": "root/autodl-tmp/runs/omg-test-split20261026-20261029/cases/test-000063-seed1287447995/trained/inference/report.json", + "sha256": "7e38f04a71634a13ee4c0f9146b2649393ec224c1be2f29d86b3db01b3510ec0", + "references": [ + { + "key": "/seed", + "value": 12345 + }, + { + "key": "/checkpoint", + "value": "/root/autodl-tmp/runs/hfd-omg-english-split20261026/best/best.ckpt" + }, + { + "key": "/omg_sampling_steps", + "value": 10 + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/omg-test-split20261026-20261029/cases/test-000111-seed3714435529/initial/inference/report.json", + "archive_file": "root/autodl-tmp/runs/omg-test-split20261026-20261029/cases/test-000111-seed3714435529/initial/inference/report.json", + "sha256": "d769ab5445003072a8fe3ffc6136f37a2f58350d73b0d7b672f02ba8d1995a2d", + "references": [ + { + "key": "/seed", + "value": 12345 + }, + { + "key": "/checkpoint", + "value": "/root/autodl-tmp/models/OMG/official/checkpoints/paper/50m/sstep=090000.ckpt" + }, + { + "key": "/omg_sampling_steps", + "value": 10 + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/omg-test-split20261026-20261029/cases/test-000111-seed3714435529/trained/inference/report.json", + "archive_file": "root/autodl-tmp/runs/omg-test-split20261026-20261029/cases/test-000111-seed3714435529/trained/inference/report.json", + "sha256": "146d480e6ae6a4a959bc38fb5f6659a0f2b6f755d983cae89cc60e4df7ff793a", + "references": [ + { + "key": "/seed", + "value": 12345 + }, + { + "key": "/checkpoint", + "value": "/root/autodl-tmp/runs/hfd-omg-english-split20261026/best/best.ckpt" + }, + { + "key": "/omg_sampling_steps", + "value": 10 + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/omg-test-split20261026-20261029/cases/test-000421-seed1978152656/initial/inference/report.json", + "archive_file": "root/autodl-tmp/runs/omg-test-split20261026-20261029/cases/test-000421-seed1978152656/initial/inference/report.json", + "sha256": "b8aff977824f8e416c416f5246bfe7f41ea8be39319ad4e4804d7ba327ad3c4b", + "references": [ + { + "key": "/seed", + "value": 12345 + }, + { + "key": "/checkpoint", + "value": "/root/autodl-tmp/models/OMG/official/checkpoints/paper/50m/sstep=090000.ckpt" + }, + { + "key": "/omg_sampling_steps", + "value": 10 + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/omg-test-split20261026-20261029/cases/test-000421-seed1978152656/trained/inference/report.json", + "archive_file": "root/autodl-tmp/runs/omg-test-split20261026-20261029/cases/test-000421-seed1978152656/trained/inference/report.json", + "sha256": "24151081882233f3da5245f580d55afef8b1f53f90760fc379fbbbf4aee5a669", + "references": [ + { + "key": "/seed", + "value": 12345 + }, + { + "key": "/checkpoint", + "value": "/root/autodl-tmp/runs/hfd-omg-english-split20261026/best/best.ckpt" + }, + { + "key": "/omg_sampling_steps", + "value": 10 + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/omg-test-split20261026-20261029/cases/third-test-35fb534d3cf361105fc557a2d0028b7da05d4a12828138ecbd30fb01c537c8cf/initial/inference/report.json", + "archive_file": "root/autodl-tmp/runs/omg-test-split20261026-20261029/cases/third-test-35fb534d3cf361105fc557a2d0028b7da05d4a12828138ecbd30fb01c537c8cf/initial/inference/report.json", + "sha256": "9dd867ef25ad1bda1fc7ac9c537860ff933425ae1a491c1d834ff2cc42367995", + "references": [ + { + "key": "/seed", + "value": 12345 + }, + { + "key": "/checkpoint", + "value": "/root/autodl-tmp/models/OMG/official/checkpoints/paper/50m/sstep=090000.ckpt" + }, + { + "key": "/omg_sampling_steps", + "value": 10 + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/omg-test-split20261026-20261029/cases/third-test-35fb534d3cf361105fc557a2d0028b7da05d4a12828138ecbd30fb01c537c8cf/trained/inference/report.json", + "archive_file": "root/autodl-tmp/runs/omg-test-split20261026-20261029/cases/third-test-35fb534d3cf361105fc557a2d0028b7da05d4a12828138ecbd30fb01c537c8cf/trained/inference/report.json", + "sha256": "096a0028c6c04f08916ce7a1bd6147139ce1f20b0cb4fc05eda41ba2bfd0e2a5", + "references": [ + { + "key": "/seed", + "value": 12345 + }, + { + "key": "/checkpoint", + "value": "/root/autodl-tmp/runs/hfd-omg-english-split20261026/best/best.ckpt" + }, + { + "key": "/omg_sampling_steps", + "value": 10 + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/omg-test-split20261026-20261029/cases/third-test-8933eb6de44a23953103048740c88aff70eac83163896dc0e9f83b6f88da8408/initial/inference/report.json", + "archive_file": "root/autodl-tmp/runs/omg-test-split20261026-20261029/cases/third-test-8933eb6de44a23953103048740c88aff70eac83163896dc0e9f83b6f88da8408/initial/inference/report.json", + "sha256": "6d2dd683c2d818941c9afbeeb254443d13cd59bf43b296ab65c2c0dbaccafaae", + "references": [ + { + "key": "/seed", + "value": 12345 + }, + { + "key": "/checkpoint", + "value": "/root/autodl-tmp/models/OMG/official/checkpoints/paper/50m/sstep=090000.ckpt" + }, + { + "key": "/omg_sampling_steps", + "value": 10 + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/omg-test-split20261026-20261029/cases/third-test-8933eb6de44a23953103048740c88aff70eac83163896dc0e9f83b6f88da8408/trained/inference/report.json", + "archive_file": "root/autodl-tmp/runs/omg-test-split20261026-20261029/cases/third-test-8933eb6de44a23953103048740c88aff70eac83163896dc0e9f83b6f88da8408/trained/inference/report.json", + "sha256": "bf0aef3af05e1612af0b7e119cc24c531dd39da1837c6c5dc9ba70aa69b34317", + "references": [ + { + "key": "/seed", + "value": 12345 + }, + { + "key": "/checkpoint", + "value": "/root/autodl-tmp/runs/hfd-omg-english-split20261026/best/best.ckpt" + }, + { + "key": "/omg_sampling_steps", + "value": 10 + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/omg-test-split20261026-20261029/cases/third-test-f40ec4be3f9aaa51418929e3d258c25eb32216445dc2d867e5563891a59c4770/initial/inference/report.json", + "archive_file": "root/autodl-tmp/runs/omg-test-split20261026-20261029/cases/third-test-f40ec4be3f9aaa51418929e3d258c25eb32216445dc2d867e5563891a59c4770/initial/inference/report.json", + "sha256": "d8db5828bcfc989591a35c776710113bfb507588b6897f1da395fdf17248bd73", + "references": [ + { + "key": "/seed", + "value": 12345 + }, + { + "key": "/checkpoint", + "value": "/root/autodl-tmp/models/OMG/official/checkpoints/paper/50m/sstep=090000.ckpt" + }, + { + "key": "/omg_sampling_steps", + "value": 10 + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/omg-test-split20261026-20261029/cases/third-test-f40ec4be3f9aaa51418929e3d258c25eb32216445dc2d867e5563891a59c4770/trained/inference/report.json", + "archive_file": "root/autodl-tmp/runs/omg-test-split20261026-20261029/cases/third-test-f40ec4be3f9aaa51418929e3d258c25eb32216445dc2d867e5563891a59c4770/trained/inference/report.json", + "sha256": "c55e0fcb6d6a941a32869659fb835b676c47aa38c53482dd44a0e60d12650465", + "references": [ + { + "key": "/seed", + "value": 12345 + }, + { + "key": "/checkpoint", + "value": "/root/autodl-tmp/runs/hfd-omg-english-split20261026/best/best.ckpt" + }, + { + "key": "/omg_sampling_steps", + "value": 10 + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/omg-test-split20261026-20261029/evaluation.json", + "archive_file": "root/autodl-tmp/runs/omg-test-split20261026-20261029/evaluation.json", + "sha256": "5f0eb678ce410760e10094c9a8489be55a99092a02fa74be2ebf1099faad4d60", + "references": [ + { + "key": "/seed", + "value": 12345 + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/omg-test-split20261026-20261029/selection.json", + "archive_file": "root/autodl-tmp/runs/omg-test-split20261026-20261029/selection.json", + "sha256": "a191dd26027c70f6ac255941d81ffd971c8470f453a512bfefb9166fbb93d9c0", + "references": [ + { + "key": "/provider_manifest/seed", + "value": 42 + }, + { + "key": "/provider_manifest/sampling_seed", + "value": 9000049 + }, + { + "key": "/provider_manifest/native_config/audio_sample_step_frames", + "value": 20 + }, + { + "key": "/provider_manifest/native_config/action_sample_step_frames", + "value": 1 + }, + { + "key": "/provider_manifest/native_config/max_motion_steps", + "value": 0 + }, + { + "key": "/provider_manifest/native_config/initial_idle_max_steps", + "value": 7 + }, + { + "key": "/provider_manifest/native_config/control_post_idle_min_steps", + "value": 3 + }, + { + "key": "/provider_manifest/native_config/control_post_idle_max_steps", + "value": 20 + }, + { + "key": "/provider_manifest/native_config/max_speech_onset_steps", + "value": 3 + }, + { + "key": "/provider_manifest/native_config/listen_reaction_steps", + "value": 1 + }, + { + "key": "/provider_manifest/native_config/max_stream_steps", + "value": 1919 + }, + { + "key": "/provider_manifest/final_release/seed_policy", + "value": "recorded_per_session" + }, + { + "key": "/provider_manifest/third_party/final_release/seed_policy", + "value": "recorded_per_session" + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/omg-test-split20261026-20261029/sessions/test-000033-seed2314733246.json", + "archive_file": "root/autodl-tmp/runs/omg-test-split20261026-20261029/sessions/test-000033-seed2314733246.json", + "sha256": "84417ec6730c196f8bbe62a47b2b55949f41fb30e2ccd01942a7936f8ff7140c", + "references": [ + { + "key": "/disfluency_events/0/gap_steps", + "value": 2 + }, + { + "key": "/terminal_idle_steps", + "value": 7 + }, + { + "key": "/idle_steps", + "value": 13 + }, + { + "key": "/listen_steps", + "value": 13 + }, + { + "key": "/motion_steps", + "value": 8 + }, + { + "key": "/semantic_steps", + "value": 34 + }, + { + "key": "/action_idle_steps", + "value": 249 + }, + { + "key": "/action_listen_steps", + "value": 260 + }, + { + "key": "/action_motion_steps", + "value": 160 + }, + { + "key": "/action_steps", + "value": 669 + }, + { + "key": "/sample_seed", + "value": 2314733246 + }, + { + "key": "/config/audio_sample_step_frames", + "value": 20 + }, + { + "key": "/config/action_sample_step_frames", + "value": 1 + }, + { + "key": "/config/max_motion_steps", + "value": 0 + }, + { + "key": "/config/initial_idle_max_steps", + "value": 7 + }, + { + "key": "/config/control_post_idle_min_steps", + "value": 3 + }, + { + "key": "/config/control_post_idle_max_steps", + "value": 20 + }, + { + "key": "/config/max_speech_onset_steps", + "value": 3 + }, + { + "key": "/config/listen_reaction_steps", + "value": 1 + }, + { + "key": "/config/max_stream_steps", + "value": 1919 + }, + { + "key": "/final_release/seed_policy", + "value": "recorded_per_session" + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/omg-test-split20261026-20261029/sessions/test-000045-seed1597101952.json", + "archive_file": "root/autodl-tmp/runs/omg-test-split20261026-20261029/sessions/test-000045-seed1597101952.json", + "sha256": "bbb575dcedd9f90f06a7b68c9ac4821043bbcfe319520856c643bbfae3fe8228", + "references": [ + { + "key": "/terminal_idle_steps", + "value": 5 + }, + { + "key": "/idle_steps", + "value": 10 + }, + { + "key": "/listen_steps", + "value": 11 + }, + { + "key": "/motion_steps", + "value": 25 + }, + { + "key": "/semantic_steps", + "value": 46 + }, + { + "key": "/action_idle_steps", + "value": 189 + }, + { + "key": "/action_listen_steps", + "value": 220 + }, + { + "key": "/action_motion_steps", + "value": 500 + }, + { + "key": "/action_steps", + "value": 909 + }, + { + "key": "/sample_seed", + "value": 1597101952 + }, + { + "key": "/config/audio_sample_step_frames", + "value": 20 + }, + { + "key": "/config/action_sample_step_frames", + "value": 1 + }, + { + "key": "/config/max_motion_steps", + "value": 0 + }, + { + "key": "/config/initial_idle_max_steps", + "value": 7 + }, + { + "key": "/config/control_post_idle_min_steps", + "value": 3 + }, + { + "key": "/config/control_post_idle_max_steps", + "value": 20 + }, + { + "key": "/config/max_speech_onset_steps", + "value": 3 + }, + { + "key": "/config/listen_reaction_steps", + "value": 1 + }, + { + "key": "/config/max_stream_steps", + "value": 1919 + }, + { + "key": "/final_release/seed_policy", + "value": "recorded_per_session" + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/omg-test-split20261026-20261029/sessions/test-000050-seed4281401434.json", + "archive_file": "root/autodl-tmp/runs/omg-test-split20261026-20261029/sessions/test-000050-seed4281401434.json", + "sha256": "e2ca769127679ed1679a2801ba8ec118bebbfa2396619b3c721d881c1a833ffd", + "references": [ + { + "key": "/disfluency_events/0/gap_steps", + "value": 4 + }, + { + "key": "/terminal_idle_steps", + "value": 4 + }, + { + "key": "/idle_steps", + "value": 8 + }, + { + "key": "/listen_steps", + "value": 12 + }, + { + "key": "/motion_steps", + "value": 19 + }, + { + "key": "/semantic_steps", + "value": 39 + }, + { + "key": "/action_idle_steps", + "value": 149 + }, + { + "key": "/action_listen_steps", + "value": 240 + }, + { + "key": "/action_motion_steps", + "value": 380 + }, + { + "key": "/action_steps", + "value": 769 + }, + { + "key": "/sample_seed", + "value": 4281401434 + }, + { + "key": "/config/audio_sample_step_frames", + "value": 20 + }, + { + "key": "/config/action_sample_step_frames", + "value": 1 + }, + { + "key": "/config/max_motion_steps", + "value": 0 + }, + { + "key": "/config/initial_idle_max_steps", + "value": 7 + }, + { + "key": "/config/control_post_idle_min_steps", + "value": 3 + }, + { + "key": "/config/control_post_idle_max_steps", + "value": 20 + }, + { + "key": "/config/max_speech_onset_steps", + "value": 3 + }, + { + "key": "/config/listen_reaction_steps", + "value": 1 + }, + { + "key": "/config/max_stream_steps", + "value": 1919 + }, + { + "key": "/final_release/seed_policy", + "value": "recorded_per_session" + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/omg-test-split20261026-20261029/sessions/test-000063-seed1287447995.json", + "archive_file": "root/autodl-tmp/runs/omg-test-split20261026-20261029/sessions/test-000063-seed1287447995.json", + "sha256": "c00cc1ab3dff8a4e7025a3b1b9a1a15cdbc6294679d6d639b62853ee8bb63dfb", + "references": [ + { + "key": "/terminal_idle_steps", + "value": 5 + }, + { + "key": "/idle_steps", + "value": 8 + }, + { + "key": "/listen_steps", + "value": 4 + }, + { + "key": "/motion_steps", + "value": 19 + }, + { + "key": "/semantic_steps", + "value": 31 + }, + { + "key": "/action_idle_steps", + "value": 149 + }, + { + "key": "/action_listen_steps", + "value": 80 + }, + { + "key": "/action_motion_steps", + "value": 380 + }, + { + "key": "/action_steps", + "value": 609 + }, + { + "key": "/sample_seed", + "value": 1287447995 + }, + { + "key": "/config/audio_sample_step_frames", + "value": 20 + }, + { + "key": "/config/action_sample_step_frames", + "value": 1 + }, + { + "key": "/config/max_motion_steps", + "value": 0 + }, + { + "key": "/config/initial_idle_max_steps", + "value": 7 + }, + { + "key": "/config/control_post_idle_min_steps", + "value": 3 + }, + { + "key": "/config/control_post_idle_max_steps", + "value": 20 + }, + { + "key": "/config/max_speech_onset_steps", + "value": 3 + }, + { + "key": "/config/listen_reaction_steps", + "value": 1 + }, + { + "key": "/config/max_stream_steps", + "value": 1919 + }, + { + "key": "/final_release/seed_policy", + "value": "recorded_per_session" + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/omg-test-split20261026-20261029/sessions/test-000111-seed3714435529.json", + "archive_file": "root/autodl-tmp/runs/omg-test-split20261026-20261029/sessions/test-000111-seed3714435529.json", + "sha256": "feac8344999acc57ff8208830b8bfff6226d7d70ad37dca570e0ed1f886a4ec7", + "references": [ + { + "key": "/terminal_idle_steps", + "value": 11 + }, + { + "key": "/idle_steps", + "value": 23 + }, + { + "key": "/listen_steps", + "value": 8 + }, + { + "key": "/motion_steps", + "value": 11 + }, + { + "key": "/semantic_steps", + "value": 42 + }, + { + "key": "/action_idle_steps", + "value": 449 + }, + { + "key": "/action_listen_steps", + "value": 160 + }, + { + "key": "/action_motion_steps", + "value": 220 + }, + { + "key": "/action_steps", + "value": 829 + }, + { + "key": "/sample_seed", + "value": 3714435529 + }, + { + "key": "/config/audio_sample_step_frames", + "value": 20 + }, + { + "key": "/config/action_sample_step_frames", + "value": 1 + }, + { + "key": "/config/max_motion_steps", + "value": 0 + }, + { + "key": "/config/initial_idle_max_steps", + "value": 7 + }, + { + "key": "/config/control_post_idle_min_steps", + "value": 3 + }, + { + "key": "/config/control_post_idle_max_steps", + "value": 20 + }, + { + "key": "/config/max_speech_onset_steps", + "value": 3 + }, + { + "key": "/config/listen_reaction_steps", + "value": 1 + }, + { + "key": "/config/max_stream_steps", + "value": 1919 + }, + { + "key": "/final_release/seed_policy", + "value": "recorded_per_session" + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/omg-test-split20261026-20261029/sessions/test-000421-seed1978152656.json", + "archive_file": "root/autodl-tmp/runs/omg-test-split20261026-20261029/sessions/test-000421-seed1978152656.json", + "sha256": "efef3818710808bc8bcaee37144d16eae71fe4c624119fda05468f26b2cc05fc", + "references": [ + { + "key": "/terminal_idle_steps", + "value": 3 + }, + { + "key": "/idle_steps", + "value": 12 + }, + { + "key": "/listen_steps", + "value": 19 + }, + { + "key": "/motion_steps", + "value": 31 + }, + { + "key": "/semantic_steps", + "value": 62 + }, + { + "key": "/action_idle_steps", + "value": 229 + }, + { + "key": "/action_listen_steps", + "value": 380 + }, + { + "key": "/action_motion_steps", + "value": 620 + }, + { + "key": "/action_steps", + "value": 1229 + }, + { + "key": "/sample_seed", + "value": 1978152656 + }, + { + "key": "/config/audio_sample_step_frames", + "value": 20 + }, + { + "key": "/config/action_sample_step_frames", + "value": 1 + }, + { + "key": "/config/max_motion_steps", + "value": 0 + }, + { + "key": "/config/initial_idle_max_steps", + "value": 7 + }, + { + "key": "/config/control_post_idle_min_steps", + "value": 3 + }, + { + "key": "/config/control_post_idle_max_steps", + "value": 20 + }, + { + "key": "/config/max_speech_onset_steps", + "value": 3 + }, + { + "key": "/config/listen_reaction_steps", + "value": 1 + }, + { + "key": "/config/max_stream_steps", + "value": 1919 + }, + { + "key": "/final_release/seed_policy", + "value": "recorded_per_session" + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/omg-test-split20261026-20261029/stop-conditioning-ablation.json", + "archive_file": "root/autodl-tmp/runs/omg-test-split20261026-20261029/stop-conditioning-ablation.json", + "sha256": "f09ee73861aea7827818acc48c241c754e4bb71f4a6e797dccde18869eac4e5c", + "references": [ + { + "key": "/same_checkpoint_seed_gt_pcm_and_request_clocks", + "value": true + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/omg-test-split20261026-20261029/verification.json", + "archive_file": "root/autodl-tmp/runs/omg-test-split20261026-20261029/verification.json", + "sha256": "c57bccc5175818a99ea030bb26629aae9c15f30f7749038fc726fcb4bcfc36cb", + "references": [ + { + "key": "/gt_equal_across_checkpoints", + "value": true + } + ] + } +] \ No newline at end of file diff --git a/handoff-v1/inventory/omg-git-commit.txt b/handoff-v1/inventory/omg-git-commit.txt new file mode 100644 index 0000000000000000000000000000000000000000..b497fa5066ab5a016e8f4eb188e2c2e2fbe8ce67 --- /dev/null +++ b/handoff-v1/inventory/omg-git-commit.txt @@ -0,0 +1 @@ +ea269e3716e289e64fa3cec17f1366b57e3fab9b diff --git a/handoff-v1/inventory/omg-git-diff.txt b/handoff-v1/inventory/omg-git-diff.txt new file mode 100644 index 0000000000000000000000000000000000000000..c9c545369cebbcf6d2a05f027ead2f705f40b6ea --- /dev/null +++ b/handoff-v1/inventory/omg-git-diff.txt @@ -0,0 +1,6411 @@ +diff --git a/FINAL_BASELINE_METHOD.md b/FINAL_BASELINE_METHOD.md +deleted file mode 100644 +index f388a06..0000000 +--- a/FINAL_BASELINE_METHOD.md ++++ /dev/null +@@ -1,16 +0,0 @@ +-# 当前保留的方案 +- +-本文为历史完整 session 方法记录。完整调试记录见 [视频问题调试](docs/ROLLOUT_DEBUG.md)。对应产物已归档到 `/root/autodl-tmp/runs/archive/legacy-20261020/hfd-final/`;该目录不随 portable 代码迁移。当前正式训练与评测见 [runs索引](/root/autodl-tmp/runs/README.md)。 +- +-| 模型 | 配置 | 当前连续 joint RMSE | 状态 | +-| --- | --- | ---: | --- | +-| OMG | `configs/omg_full_session.yaml` | 0.0619 rad | 监听与收尾恢复,重复弯腰消失,残余接缝仍在 | +-| TextOP | `configs/textop_full_session.yaml` | 0.2992 rad | 监听姿态出现,任务切换延迟,尚未完整拟合 | +- +-共同改为 `full_session` 关节监督,不向模型输入 listen/idle 等控制标签。训练仍使用原生框架和官方初始化。OMG 保持过去两秒音频+因果 ASR,采用原生 DDIM 10 步、CFG=1、60 帧规划和剩余 24 帧的异步预取;TextOP 保持原生 8 帧输出、CFG=5。 +- +-OMG 三个固定种子的误差为 0.0619、0.1538、0.0585 rad,当前视频沿用原 seed=12345。它仍未证明泛化或完美实时执行。TextOP 真值历史诊断为 0.0125 rad,但这不是连续推理性能;长训练反馈链对照尚未可靠修复闭环。 +- +-视频位于 `videos/omg_full_session_playable.mp4` 和 `videos/textop_full_session_playable.mp4`,均保留原始混合音频。完整报告在各模型的 `inference/`;动作曲线和同 checkpoint 的对照集中在 `debug/`。论文独立测试集评测仍待完成。 +- +-重训前修改配置的 `paths.output_dir`。完整迁移步骤见 [迁移说明](docs/MIGRATION.md);`configs/*_full_session.yaml` 中的 `paths.data_disk` 可通过 `HFD_DATA_DISK` 覆盖。 +diff --git a/JOINT_MAPPING.md b/JOINT_MAPPING.md +deleted file mode 100644 +index 8ebd645..0000000 +--- a/JOINT_MAPPING.md ++++ /dev/null +@@ -1,113 +0,0 @@ +-# HFD 到 TextOP / OMG 的动作映射与历史 bug +- +-## 原始数据和目标接口 +- +-session 的 `motion_q29_xyz_quat` 是 `[T, 36]` 数组,列布局为: +- +-| 列(从 0 开始) | 内容 | 顺序 | +-| --- | --- | --- | +-| `0:29` | G1 的 29 个关节角(rad) | HFD / IsaacLab 顺序 | +-| `29:32` | 根位置 | xyz | +-| `32:36` | 根四元数 | wxyz | +- +-IsaacLab 关节顺序交错排列左右侧和躯干。OMG 的 G1 qpos 使用机器人树顺序:左腿 6 个、右腿 6 个、腰部 3 个、左臂 7 个、右臂 7 个。两边都是 29 个角度,但同一列不一定对应同一个关节。这里不改变机器人关节定义,也不拟合角度变换;按相同关节名重新排列列即可。 +- +-session 保留原始顺序,不在导出阶段预先重排。新导出的 JSON 明确标记 `joint_order=isaaclab`、`quaternion_order=wxyz`;历史 v1 session 缺少这两个字段时按同一原始契约读取。适配器会拒绝显式声明其他顺序的数据,避免二次转换。 +- +-## 按名称转换 +- +-实现集中在 [`hfd_bridge/joints.py`](hfd_bridge/joints.py): +- +-```python +-HFD_TO_NATIVE = [HFD_JOINT_NAMES.index(name) for name in NATIVE_JOINT_NAMES] +-native_joints = raw[:, :29][:, HFD_TO_NATIVE] +-qpos = np.concatenate((raw[:, 29:36], native_joints), axis=1) +-``` +- +-`HFD_TO_NATIVE[i]` 表示“目标第 i 个关节应从原始哪一列取值”,所有下标从 0 开始。得到的 36D qpos 布局是 `[xyz3, quat_wxyz4, native_joints29]`。函数 `hfd_motion_to_qpos` 不改变根位置和四元数。 +- +-实际排列为: +- +-```text +-[0, 3, 6, 9, 13, 17, 1, 4, 7, 10, 14, 18, 2, 5, 8, 11, 15, 19, 21, 23, 25, 27, 12, 16, 20, 22, 24, 26, 28] +-``` +- +-| 目标关节下标 | 关节名 | HFD 原始列 | +-| --- | --- | --- | +-| 0 | `left_hip_pitch_joint` | 0 | +-| 1 | `left_hip_roll_joint` | 3 | +-| 2 | `left_hip_yaw_joint` | 6 | +-| 3 | `left_knee_joint` | 9 | +-| 4 | `left_ankle_pitch_joint` | 13 | +-| 5 | `left_ankle_roll_joint` | 17 | +-| 6 | `right_hip_pitch_joint` | 1 | +-| 7 | `right_hip_roll_joint` | 4 | +-| 8 | `right_hip_yaw_joint` | 7 | +-| 9 | `right_knee_joint` | 10 | +-| 10 | `right_ankle_pitch_joint` | 14 | +-| 11 | `right_ankle_roll_joint` | 18 | +-| 12 | `waist_yaw_joint` | 2 | +-| 13 | `waist_roll_joint` | 5 | +-| 14 | `waist_pitch_joint` | 8 | +-| 15 | `left_shoulder_pitch_joint` | 11 | +-| 16 | `left_shoulder_roll_joint` | 15 | +-| 17 | `left_shoulder_yaw_joint` | 19 | +-| 18 | `left_elbow_joint` | 21 | +-| 19 | `left_wrist_roll_joint` | 23 | +-| 20 | `left_wrist_pitch_joint` | 25 | +-| 21 | `left_wrist_yaw_joint` | 27 | +-| 22 | `right_shoulder_pitch_joint` | 12 | +-| 23 | `right_shoulder_roll_joint` | 16 | +-| 24 | `right_shoulder_yaw_joint` | 20 | +-| 25 | `right_elbow_joint` | 22 | +-| 26 | `right_wrist_roll_joint` | 24 | +-| 27 | `right_wrist_pitch_joint` | 26 | +-| 28 | `right_wrist_yaw_joint` | 28 | +- +-HFD 部署代码 [`upstream/hfd/scripts/deploy/run_sonic_sim_record_online.py`](upstream/hfd/scripts/deploy/run_sonic_sim_record_online.py) 使用相反方向的表,把 MuJoCo 关节值写成 IsaacLab 状态: +- +-```text +-[0, 6, 12, 1, 7, 13, 2, 8, 14, 3, 9, 15, 22, 4, 10, +- 16, 23, 5, 11, 17, 24, 18, 25, 19, 26, 20, 27, 21, 28] +-``` +- +-转换回原生顺序要取这个排列的逆排列,不能直接沿用同一张表。当前代码按关节名生成转换索引,并用部署表独立核对。 +- +-## 两个模型分别如何使用 +- +-**TextOP:** 原生接口只建模 23 个关节。适配器按模型的关节名称从 HFD 29 列中取值,保持 50 Hz,再交给原生 FeatureVersion=3 编码得到 57D 动作特征。选取的原生 29 关节下标为 `0:19` 和 `22:26`,即双腿、腰、双臂的肩和肘;六个腕关节不属于模型输出。实现可以直接使用 `TEXTOP_HFD_JOINTS` 从 raw 取 23 列,等价于先重排 29 列再选取 23 列。TextOP 根旋转接口使用 xyzw,所以输入 wxyz 会归一化并改为 `[x, y, z, w]`;根位置保留。视频中的六个腕关节使用初始化时的值,模型没有预测它们,也没有填入未来真值。 +- +-**OMG:** 原生接口使用全部 29 个关节。适配器将 HFD 的 50 Hz 动作重采样到 30 Hz:关节角和位置做线性插值,四元数做 SLERP;随后把关节列重排为原生顺序。qpos 根四元数保持 wxyz。再通过原生 G1 FK 和 codec 生成 125D 特征。按用户指定的离线测试口径,音频与目标动作窗口对齐。 +- +-渲染统一使用原生 29 关节的 qpos。真值也要经过同一份正确的 HFD 到原生转换,不能把 raw 的前 29 列直接写给渲染器。实现见 [`hfd_bridge/training.py`](hfd_bridge/training.py)、[`infer_overfit.py`](infer_overfit.py) 和 [`render_overfit.py`](render_overfit.py)。 +- +-## 之前的 bug 及影响 +- +-旧适配器误认为 HFD 的前 29 列已经是原生机器人树顺序,只调整了 root 字段的位置,没有重排关节列;TextOP 也按原生数字下标直接选择 HFD 列,因而取错关节。 +- +-例如原始第 1 列(从 0 开始)是 `right_hip_pitch_joint`,而原生第 1 个关节是 `left_hip_roll_joint`。错误代码把右髋 pitch 角写给左髋 roll。多个关节同时错位,会造成姿态明显异常。真值视频也采用了错误解释,因此即使完全没有模型预测,真值机器人仍会表现奇怪。 +- +-之前的往返测试用同一套错误排列做编码和解码,能够还原输入数字,只证明转换自洽,不能证明数字对应了正确的命名关节。错误动作空间内的训练损失下降,也不能证明真实机器人动作拟合正确。 +- +-以下结果不能用于论文中的动作质量或过拟合结论: +- +-- 修复前由旧适配器生成的动作特征、训练检查点、动作误差和预测视频。 +-- `hfd-video-overfit-textop-20261005`、`hfd-video-overfit-omg-20261005` 和 `hfd-overfit-videos-20261005` 中的旧结果,以及更早依赖同一错误排列的动作实验。 +- +-原始 HFD session、混合音频和 ASR 转写可以继续使用;错误动作特征必须重新生成。两个 baseline 已从官方初始化权重重新进行 800 步小样本训练,正确结果目录带 `jointfixed`,摘要见 [`reports/history/overfit_videos.json`](reports/history/overfit_videos.json)。旧视频、废弃适配脚本及旧实验检查点已清理;历史小型配置、日志和检查摘要集中保存在 `runs/diagnostics/historical/`,不能作为修复后模型的评测证据。 +- +-## 如何验证映射 +- +-```bash +-# 在 TextOP 环境运行已有回归测试 +-python -m unittest discover -s tests -t . +- +-# 在 OMG 环境核对独立 FK,并导出相同真值的错误/正确解释对照 +-# --output 必须是新的目录;外部 assets 需要包含 G1 URDF 和 mesh +-python check_joint_mapping.py \ +- --session /data/sessions/sample.npz \ +- --assets-root /data/OMG/assets \ +- --output /data/runs/joint-mapping-audit +-``` +- +-核验包括:部署表的逆排列与按名称转换一致;29 个关节逐个脉冲时仅到达同名电机;根位置和 wxyz 保持不变;50/30 Hz 公共时刻的姿态一致;OMG FK 与独立 MuJoCo 身体位置一致。已有样本核验的最大身体位置误差约 `8.0e-7 m`,见 [`reports/joint_mapping_audit.json`](reports/joint_mapping_audit.json)。这些检查确认数据与机器人接口对应正确;小样本训练确认训练链路可用,尚不构成完整全双工评测结论。 +diff --git a/MANIFEST.json b/MANIFEST.json +index 645a2a1..187efd9 100644 +--- a/MANIFEST.json ++++ b/MANIFEST.json +@@ -5,21 +5,9 @@ + "bytes": 4810, + "sha256": "df3bcbf6da74ff04850db5f301b6b9f84bfbea2ed68798787ad4fbb3277ec63f" + }, +- "FINAL_BASELINE_METHOD.md": { +- "bytes": 1728, +- "sha256": "014307754a7ae7d347ca5b64b6d30cb4b3831639e010ace4b6ca0f2f0b644200" +- }, +- "JOINT_MAPPING.md": { +- "bytes": 7338, +- "sha256": "a5047dc18dd36107ce385f6391460c7782fee6578fc565e7b627792c5096405f" +- }, + "README.md": { +- "bytes": 3805, +- "sha256": "8c55fba302e73f73e43ee5ad0096b852a775c23fd7ba01ac6fc068dedd4a3dd0" +- }, +- "SESSION_FORMAT.md": { +- "bytes": 5054, +- "sha256": "35c27ea0f7fe575cfbe236846ed540c3c95df432c0af8c6148091739a6c92869" ++ "bytes": 9402, ++ "sha256": "a85b9ee971e5a855b347a2b097ae5e4437b38369f1fb250faf6b135664d7e040" + }, + "TRAINING_ASSETS.json": { + "bytes": 87123, +@@ -81,181 +69,9 @@ + "bytes": 16378, + "sha256": "a4245020980dcbe6f8ba1527e96317549937ecaf128d43b8b2f24fef6405b365" + }, +- "docs/ASR_CONTEXT_OVERFIT.md": { +- "bytes": 6819, +- "sha256": "9838a5173dde9bfcffc3bab190282cd04e932b0933b52f1dbb6b7558078610ba" +- }, +- "docs/BASELINE_TEXT_CONDITION.md": { +- "bytes": 9685, +- "sha256": "1d291382f90539d20bc37e491d7c684ee888531fc2241fe53a688b7443363b40" +- }, +- "docs/ENGLISH_ONLY_PROBE.md": { +- "bytes": 9948, +- "sha256": "4ef5a6967a1e9edec4739e9882d97acc18d119a6e23342e599c0f9168d6077c2" +- }, +- "docs/FORMAL_TRAINING_READINESS.md": { +- "bytes": 14837, +- "sha256": "3676c5395bb74d0092327d2db1f8db8ba11ee4327785277b9769f8d196f88621" +- }, +- "docs/GENMO_FORMAL_TRAINING.md": { +- "bytes": 11336, +- "sha256": "eb3761334b649a5806fa5dc1e5029356e4b55e85524621b1e01642bf67a488f6" +- }, +- "docs/GENMO_HFD_OVERFIT.md": { +- "bytes": 12848, +- "sha256": "ca24702e3cfb4f359bfba677b24a013711cd0471f4255d2f8b619ab9ca327e93" +- }, +- "docs/GENMO_LISTENING_DIAGNOSIS.md": { +- "bytes": 3858, +- "sha256": "d54bdc5141d0bde62f1d76258f4ab353cbfe2d7fc713bffc6788326d322f58ee" +- }, +- "docs/GENMO_PAST_AUDIO_TEST.md": { +- "bytes": 9364, +- "sha256": "2a95909082c8ad998c542a205f1c2d3b4600bac04b9334753e5a16428356dbef" +- }, +- "docs/GENMO_V2_TRAINING.md": { +- "bytes": 6611, +- "sha256": "5c02cd209ccd1f9cf87f4f60b2d5fc54f00b412bde97ffa261856a3d96f4a397" +- }, +- "docs/GENMO_V2_VALIDATION.md": { +- "bytes": 5740, +- "sha256": "40d7cf46d333e194205dedf310954111d9fdf5ec014c4ca86dea4e3417824285" +- }, +- "docs/HFD_HUMAN_BASELINE_INPUT_PROTOCOL.md": { +- "bytes": 12323, +- "sha256": "dec2e8df1522e6740ef53ee1147e8e65cd318408c7835f65c507d56b5d28edd5" +- }, + "docs/HFD_HUMAN_BASELINE_SOURCES.json": { +- "bytes": 10598, +- "sha256": "14dae98d032442b7c86cd5f4375ff8aac2b8a99a8e8fa6840d6266ecf93f93a8" +- }, +- "docs/HUMAN_FORMAL_VALIDATION_20261024.md": { +- "bytes": 4630, +- "sha256": "286eec4ad74f26c93aa5ca5c91acbfb8b0fc0412c4ce743269a67bb074081ec1" +- }, +- "docs/HUMAN_RETARGET_EVALUATION.md": { +- "bytes": 7664, +- "sha256": "9efd14122d7f8461ccc9cf2db2d5843e02aac9b1b3b392b728b13517d3775a87" +- }, +- "docs/KIMODO_CLOSED_LOOP_STABILITY.md": { +- "bytes": 17001, +- "sha256": "7c29773be52bd16548208c4365cf4b31e4a9fa5b0b87efa9ee0eba6400d9f3f5" +- }, +- "docs/KIMODO_FORMAL_VALIDATION.md": { +- "bytes": 6733, +- "sha256": "d89622850760e9889549495e1b93dd30d4bc4f099ff9faf67f3384089527a2cf" +- }, +- "docs/KIMODO_HFD_OVERFIT_RESULTS.md": { +- "bytes": 6166, +- "sha256": "19dc1cc2dd3a0c6dba023b7b6ca979f0478b82c427e582b6386a9d94b512c091" +- }, +- "docs/KIMODO_HFD_TRAINING.md": { +- "bytes": 13291, +- "sha256": "4b7033db2727ded4d64a015c43d12b6fcd547a40a2e38662b2a33e2e3c9f0037" +- }, +- "docs/KIMODO_MOTION_RECOVERY.md": { +- "bytes": 7245, +- "sha256": "dc4f6192b533301946b218a51c0ab442572dd96996544e27b100f5594d7d2f22" +- }, +- "docs/KIMODO_ROOT_COLLAPSE_DIAGNOSIS.md": { +- "bytes": 4618, +- "sha256": "42a7a07344f523abb3bfb4d22214655141e4d4b38f7ca4d081f688ab57dc695d" +- }, +- "docs/KIMODO_ROOT_TRAINING_CAUSE_ANALYSIS.md": { +- "bytes": 8012, +- "sha256": "e2c119d0cc7e0cc65bf3c741142ad88c7844c41b5d33228189288848caac29a2" +- }, +- "docs/KIMODO_TRAINING_REPAIR.md": { +- "bytes": 5601, +- "sha256": "880ba0e4477735500471f7ac36fd075d63d874b4578b4ba8284045530670915e" +- }, +- "docs/LOCAL_INDEX_TRAINING.md": { +- "bytes": 6004, +- "sha256": "ac2dc3538b1214bf348bb672a79a6d6e1ea0b4d74cee907fa2416fe0f66405bd" +- }, +- "docs/MIGRATION.md": { +- "bytes": 426, +- "sha256": "1bbfc85ca07c183bdb6f25143d08db7dc1e526e9fb29993225a3232e74c689af" +- }, +- "docs/OMG_ENGLISH_VALIDATION.md": { +- "bytes": 3791, +- "sha256": "86a2a91471737d81fe9ef741aac5e224d3f727db9f6c08161c3bb98b78dee89b" +- }, +- "docs/OMG_TRAINING_REVIEW.md": { +- "bytes": 14254, +- "sha256": "d38b0b5049701c620935e0bf51bf0da77aa3cb2de675c33ba348b972039609d8" +- }, +- "docs/OMG_VALIDATION_DIAGNOSIS.md": { +- "bytes": 6133, +- "sha256": "5269c34a70f40977a9848106a59e05abbe9ee8486870ca4682e808cd0ac4259b" +- }, +- "docs/OMG_VALIDATION_ROLLOUT.md": { +- "bytes": 6132, +- "sha256": "ff8c7c8f249c3c89607f080ff281e15acc3d8fccc710c0f8563e417ecfa7d82e" +- }, +- "docs/PORTABLE_TRAINING.md": { +- "bytes": 11281, +- "sha256": "4b47034c1d5916537605e96153450ba5d8f3e89cb5e1fa4796e4a18b960ba973" +- }, +- "docs/REFERENCE_DYNAMIC_TRAINING.md": { +- "bytes": 9587, +- "sha256": "e55ead5d171134b3ef2a9238e4f28bd5315ace74733481f5033f6befe8b175b7" +- }, +- "docs/ROLLOUT_DEBUG.md": { +- "bytes": 8444, +- "sha256": "c45aefa40fe66dd1504465ec8254d4438aa4325b2f9687afda871e7c312e971d" +- }, +- "docs/SEMANTICS_CLOCK_FIX.md": { +- "bytes": 5914, +- "sha256": "66d1d57f18e69176b5ad91bf980a8f180350137b9c2b02d97ae1a71fa5d627cc" +- }, +- "docs/TEXTOP_ENGLISH_VALIDATION.md": { +- "bytes": 5350, +- "sha256": "c653a371cb97f690349d96c9750366cb570bd53cedb64695b283be9f67371374" +- }, +- "docs/TRAINING_REPRODUCTION_MATRIX.md": { +- "bytes": 4693, +- "sha256": "0c977387862ae82c48c2667e5d3f4520a9c522e6394a8d4e68cfa9db2a9919ad" +- }, +- "docs/TRAINING_RESUME.md": { +- "bytes": 4215, +- "sha256": "3caa601fdaef09d30ab79c195299fc5dfe52cd0b32be4daaf97e6bc7d7484253" +- }, +- "docs/WORKSPACE_CLEANUP.md": { +- "bytes": 4422, +- "sha256": "d10560c0c62a61f6cbee040c90c5db4fbe7167aecb286c2d8d637d161c72e90b" +- }, +- "docs/history/BASELINE_METHOD_BEFORE_SEMANTIC_FIX.md": { +- "bytes": 10085, +- "sha256": "60cfdeb3ee30d306e40af466723eaa98d73c2efa0a806b8b219a9b79f8013847" +- }, +- "docs/history/DUPLEX_TASK.md": { +- "bytes": 7084, +- "sha256": "75c02d717b23b000d95b35fbda0799326dece83f0e6ec893a368fe5e07ba420c" +- }, +- "docs/history/HFD_HUMAN_BASELINE_INPUT_PROTOCOL_RESEARCH_20261017.md": { +- "bytes": 40979, +- "sha256": "ad30933320fc00b863d50f7f102ab5c546c5a82ce4919ab4956af6c93997f702" +- }, +- "docs/history/HFD_WINDOW_PROTOCOL.md": { +- "bytes": 7986, +- "sha256": "4a5a25129aa0b5cd1173af7d2c9d365cdcb118a56856cb197f547f1e5e05c3d4" +- }, +- "docs/history/NORMAL_TASK.md": { +- "bytes": 6244, +- "sha256": "9a5600eeb00436d77603501666c910ca3ed536a93c696015ed706b5f5bff9c55" +- }, +- "docs/history/OMG_CONTEXT_DEBUG.md": { +- "bytes": 6920, +- "sha256": "34d50e904a8cfc721656e7939f6aff3c07021726d9a46b818ac853945d90134e" +- }, +- "docs/history/README.md": { +- "bytes": 670, +- "sha256": "09b277b367872ab7203e6ee1e6e48d31c75269e651982a1286f074b343b2ef43" +- }, +- "docs/history/VIDEO_SAMPLES.md": { +- "bytes": 7997, +- "sha256": "ff3489e07a6233176669aa5f2567284dbd880cb36e3f6089dbc60f3d9516295b" ++ "bytes": 10428, ++ "sha256": "2cc65718bc6bf9d0d089ab341db266f9bfc18fee1d6cdec4f80978ddf74f1493" + }, + "env.sh": { + "bytes": 631, +@@ -478,8 +294,8 @@ + "sha256": "1dc280ec6f2c008924b1a69079ee50d8e8281c0c2da2d291fc5c9f9c91a36769" + }, + "tools/configure_training.py": { +- "bytes": 6935, +- "sha256": "1519a086061cda3bf93aa1f7188bc0746c95efe2084cafb7feefc09405855308" ++ "bytes": 7506, ++ "sha256": "88ebaa442e230394b7929a85f26a696a309c5315fc4ce2f0b3307af730b7f7bc" + }, + "tools/download_hfd_tars.py": { + "bytes": 6701, +@@ -2348,6 +2164,42 @@ + "validate_sessions.py": { + "bytes": 1431, + "sha256": "8f7c045d7e9f75deeffc5777eefe31760c893c929fb25d2dffda8d8e520bd880" ++ }, ++ "docs/CURRENT_RUNS.md": { ++ "bytes": 15135, ++ "sha256": "a65e3eb72a8da86e705d134a9c43aa97be17aa51f53f4d3d1bf4c858faf7c2eb" ++ }, ++ "train_hymotion.py": { ++ "bytes": 3070, ++ "sha256": "c93a486c828963e568750da26ff81256804633fd34d28d9415d3535168c7c0b0" ++ }, ++ "train_hymotion_autoregressive.py": { ++ "bytes": 15136, ++ "sha256": "1438ed061573814dfa5fe9afe6914fb59df1e70d0215958dcbd9e1b114be839f" ++ }, ++ "hfd_bridge/hymotion_model.py": { ++ "bytes": 7571, ++ "sha256": "ed29d2d03b247ed681a09dc230697885c5ab8fa8b30f82c4a827aa8fe0b41597" ++ }, ++ "hfd_bridge/hymotion_data.py": { ++ "bytes": 2203, ++ "sha256": "79503602a9a4de4f67931fa944a198d6f438b531a363ffbe545ecc4dd54f97e7" ++ }, ++ "hfd_bridge/hymotion_training.py": { ++ "bytes": 3382, ++ "sha256": "792b76cfc35f58b0c7c94b40a660f00c4971fff1d978ffd0c923b9f26a5fa733" ++ }, ++ "hfd_bridge/hymotion_rollout.py": { ++ "bytes": 2186, ++ "sha256": "6e27c4e24bd42a211df7be2a0271337af4e658d66e5d303633912bbd40313517" ++ }, ++ "hfd_bridge/hymotion_evaluation.py": { ++ "bytes": 9348, ++ "sha256": "fe72667281f04a55bc8c88375b7ca3cee1f1b6de2c23fcd7d45f1e73c1530a99" ++ }, ++ "configs/hymotion_c24.yaml": { ++ "bytes": 1123, ++ "sha256": "258c03cd4ed31c41575c4aafe0e3ddd5758a0654a5f75a903833be6ff0359bca" + } + } + } +diff --git a/README.md b/README.md +index be11660..aac9867 100644 +--- a/README.md ++++ b/README.md +@@ -1,43 +1,861 @@ +-# HFD baseline training ++# HFD 五模型训练约定 + +-OMG、TextOp、GenMo和Kimodo在全双工HFD上的训练代码、配置及复现文档。仓库包含动态session组装、因果文本释放、原生训练、验证loss和完整断点恢复。HY-Motion尚无完整训练入口。 ++本文件维护当前训练约定;运行进度、有效权重、评测结论和待办见 [CURRENT_RUNS.md](docs/CURRENT_RUNS.md)。除非用户明确要求,不另写方案、诊断、历史报告等文档。 + +-| 模型 | 入口 | 正式配置 | 预算 | +-|---|---|---|---| +-| OMG | `train.py` | `configs/omg.yaml`,另支持50M/100M/300M/500M生成配置 | 100K,有效batch512 | +-| TextOp | `train.py` | `configs/textop.yaml` | 300K,batch512,三阶段各100K | +-| GenMo/GEM-SMPL v2 | `train_genmo_full.py` | `configs/genmo.yaml`,旧配方`configs/genmo_legacy.yaml` | 100K,batch32 | +-| Kimodo-SMPLX | `train_kimodo.py` | `configs/kimodo.yaml` | 100K,batch8 | ++## 数据与信息释放 + +-输入采用按词时间释放的英文前景GT文本、最近30秒上下文及各自的动作历史。OMG另输入已释放的过去2秒音频。session在训练时动态组装,不预导出全量session,不运行ASR或中文翻译。详见[输入协议](docs/HFD_HUMAN_BASELINE_INPUT_PROTOCOL.md)。 ++- 共用 [configs/sessions.yaml](configs/sessions.yaml) 的 final_release:profile=isolated_benchmark_1024_20261026;索引 Tinghong-Ye22/data_vqa_all_1025@bdca2fe7581c263e9ceb3a96204b314e199faddf,原料 Tinghong-Ye22/hfd_data@93ba4a7a9c66b3e1b7041e5f4ea038ef2418b948。 ++- 完整发布 train/val/test=258661/512/1024,含中文。当前仅英文:OMG/TextOp=126950/302/436(原生+第三方);GenMo/Kimodo/HY=117290/297/419(有配对SMPL-X的原生样本)。英文子集不等于完整1024 benchmark。 ++- 只回放动作session和原始配方,不下载或传入VQA问答、视频、问答配音;语音—动作与视频—音频完全解耦。保留严格session身份、动作长度和音频时钟检查。先按最终split/语言过滤,再按记录的upstream_split、dataset_index、seed回放,不重新随机划分。 ++- 英文文本用 aligned_reference_context、text_conditioning=none;读取本地预建GT词时间映射 /root/autodl-tmp/cache/hfd/reference-alignments.sqlite,不在线ASR、翻译或重做对齐。6608条严格对齐失败的源资产沿用已建立的整句结束后保守释放映射。 ++- 麦克风窗长0.6秒、步长0.4秒,词末加0.1秒guard后在下一个释放点可见;动作时刻t只访问麦克风时刻≤t+0.2的已释放文本,保留最近30秒,编码器另有token上限。session切换时清空。 ++- 不输入未来PCM、未释放文字、GT turn/listen/idle标签、未来时刻、根路径或姿态。带噪未来训练目标不是额外观测条件。人体标签从配对SMPL-X按原生事件配方构造,不以G1关节角代替。 + +-GenMo默认启用执行段监督、预测历史恢复与CPU预取,使用新的输出目录;旧正式权重保留。训练链路和短程检查不代表已完成新策略的全量质量验收,详见[GenMo v2](docs/GENMO_V2_TRAINING.md)。四模型及OMG多个初始化权重的覆盖范围见[复现矩阵](docs/TRAINING_REPRODUCTION_MATRIX.md)。 ++## 五模型默认训练 + +-本次速度、断点恢复及过拟合检查的实际结果见[GenMo v2验证记录](docs/GENMO_V2_VALIDATION.md),包括仍存在的动作不足问题。 ++分别使用 /root/autodl-tmp/envs/模型名/bin/python,共同seed=42;精确超参数以配置和代码为准。 + +-从[训练复现指南](docs/PORTABLE_TRAINING.md)开始,准备对应环境和外部资产后生成本机配置: ++| 模型/配置 | 条件与输出 | 正式训练预算 | ++|---|---|---| ++| OMG-B:[omg.yaml](configs/omg.yaml) | 冻结T5-Base,50 token;过去2秒已释放PCM;10帧历史@30Hz。原生125D、G1全29关节,计划60帧,异步缓冲剩24帧重规划 | 50M,100K,batch512,BF16;LR6e-5,2000 warmup后cosine至1e-6 | ++| OMG-L:[omg_l.yaml](configs/omg_l.yaml) | 与B相同的HFD数据、因果条件、125D表示和diffusion-only微调损失;官方300M/55K初始化 | GPU0,100K,microbatch64×累积8=512;BF16,LR6e-5,warmup2000+cosine | ++| OMG-XL:[omg_xl.yaml](configs/omg_xl.yaml) | 与B相同的HFD协议;官方500M/50K初始化;沿用严格加载器,仅排除官方旧版52个未使用visual_film张量 | GPU1,100K,microbatch32×累积16=512;BF16,LR6e-5,warmup2000+cosine | ++| TextOp:[textop.yaml](configs/textop.yaml) | 冻结CLIP ViT-B/32,77 token含边界;2帧57D历史@50Hz,无音频。128D latent经冻结VAE解码8帧,G1原生23关节,未建模腕关节固定 | 300K,三个100K阶段,batch512,FP32,LR1e-4;保留原生分阶段历史恢复 | ++| GenMo:[genmo.yaml](configs/genmo.yaml) | GEM-SMPL全量训练,冻结T5-3B,50 token;60帧人体历史,past_audio=false。151D,未来60/执行12帧@30Hz | 100K,batch32,BF16,LR2e-5→1e-6;regression+diffusion、执行段权重3、根速度1.0、旋转增量0.05;预测历史最高25%,1000 warmup+4000 ramp,DDIM10 | ++| Kimodo:[kimodo.yaml](configs/kimodo.yaml) | 冻结LLM2Vec,最近400正文token,独立文本worker;60帧人体历史,无音频。SMPLX22原生273D,因果root分解、root→body两阶段,未来60/执行12帧 | 100K,batch8,BF16;normalized_x0+解码连续性/根历史重建;根只训投影,冻结根注意力骨干;根/身体分开优化裁剪,LR上限2e-6/1e-5 | ++| HY-Motion:[hymotion_c24.yaml](configs/hymotion_c24.yaml) | 官方Lite+rank16/alpha32 LoRA;冻结Qwen3-8B 128正文token+CLIP-L,无音频。201D,历史60/未来60/执行12帧;10段预测反馈,Euler10/CFG1,反馈detach | C24完整训练100K,batch4,BF16,LR2e-5 cosine至1e-6;24帧连续恢复,根速度1.0/旋转增量0.05,前200更新渐启,接段权重3 | + +-```bash +-python tools/configure_training.py \ +- --work-root /data/hfd_work --assets-root /data/hfd_assets \ +- --raw-root /data/hfd_assets/hfd_data --gpu 0 +-source /data/hfd_work/training-configs/training.env ++TextOp每条训练短链有4个8帧primitive,共0.64秒;链首始终使用GT历史,后续按课程选择预测历史,每个新batch重新初始化短链。第三阶段不等于整条session完全自主展开。 + +-/data/hfd_work/envs/omg/bin/python train.py --config /data/hfd_work/training-configs/omg.yaml +-/data/hfd_work/envs/textop/bin/python train.py --config /data/hfd_work/training-configs/textop.yaml +-/data/hfd_work/envs/genmo/bin/python train_genmo_full.py --config /data/hfd_work/training-configs/genmo.yaml +-/data/hfd_work/envs/kimodo/bin/python train_kimodo.py --config /data/hfd_work/training-configs/kimodo.yaml +-``` ++关节顺序及TextOp关节选择以 [joints.py](hfd_bridge/joints.py) 为准;沿用原生统计、冻结编码器/VAE、体型和表示变换。依赖与资产由 [requirements](requirements)、[UPSTREAM_LOCK.json](UPSTREAM_LOCK.json)、[TRAINING_ASSETS.json](TRAINING_ASSETS.json)、[HYMOTION_ASSETS.json](docs/HYMOTION_ASSETS.json) 记录。 + +-命令分别运行;同一GPU顺序执行。中断后原命令加`--resume auto`,Kimodo也支持`--resume`。恢复要求配置、代码、数据身份一致,预算是总步数。 ++## 缓存、存储与恢复 + +-- `hfd_bridge/`:训练数据、动作表示、适配、文本worker、验证与断点管理。 +-- `upstream/`:按训练依赖裁剪的HFD、OMG、TextOp原生代码,保留许可证;独立评测与部署代码归入本地`test/`。 +-- `tools/`:数据下载/对齐、配置生成、外部资源连接与构建。 +-- `requirements/`、`UPSTREAM_LOCK.json`、`TRAINING_ASSETS.json`:冻结环境、外部源码revision及必要外部文件的大小、SHA256和来源。 +-- `TRAINING_SNAPSHOT.json`:迁移前当前训练源码hash;Git历史保留对应实现。 +-- `docs/`:训练协议、复现操作和诊断说明,历史实验页不作为当前默认命令。 ++- 原始tar在 /root/autodl-fs/pangxinlei/hfd_data,保留分片、不批量解压;换索引无需删除原料。新下载默认共享盘models/datasets/downloads,下载/装依赖先用 /root/.local/bin/turbo,回退严格遵守 /root/AGENTS.md。 ++- 热索引、词映射、文本embedding、派生session放tmp。热资产 /root/autodl-tmp/cache/hfd/hot-assets 已约16.6 GiB有效内容,不用默认4 GiB准备命令缩小;长尾人体原料仍按需读共享tar。 ++- OMG/TextOp原生session内存缓存8 GiB,当前分别每256/64 batch刷新。GenMo train/val缓存4/8 GiB、文本512 MiB、预取16、训练每16 batch刷新、CPU4/BLAS1。GPU瞬时空闲不代表适合加batch或并发任务。 ++- 日志用 /root/autodl-tmp/runs;OMG/TextOp权重同目录,人体模型正式权重用 /root/autodl-fs/pangxinlei/models/模型名/。tmp给checkpoint和输出留空间;共享df不代表账号配额。 ++- 完整续训须恢复模型、优化器、scheduler、已消费采样池和原总预算,并校验代码/数据/资产合同;仅加载权重须明确重置优化器与步数。训练运行期间不改其合同约束源码/配置。 ++- OMG-B/TextOp每256步完整断点、每2000步验证;GenMo/Kimodo每2000步保存并验证。GenMo另在128/256/512/1024步滚动保存一份完整早期断点,正常2000步保存成功后清理早期断点,后续保留最近2份及best。 ++- OMG-L/XL先采样完整512窗口逻辑批,再切成GPU小批;保留B的512个session槽、每256逻辑批刷新及相同随机采样顺序。每2000次优化器更新验证8192窗口(16×512)并保存完整断点,保留最近2份及验证best;日志loss按完整更新聚合。`val_check_interval`已换算为小批次数,避免累积梯度时过频验证。 ++- OMG-L/XL日志与运行合同分别在`/root/autodl-tmp/runs/hfd-omg-{l,xl}-english-split20261026`,完整权重/优化器在共享盘`/root/autodl-fs/pangxinlei/models/OMG/hfd-{l,xl}-english-split20261026`,本地checkpoints/best/last路径为兼容链接。使用原生DirectTorchCheckpointIO流式临时文件+原子替换,恢复同时检查配置、原运行归属、RNG和已消费的逻辑批状态。 ++- GenMo环境固定NumPy1.26.4、SciPy1.15.3、Pandas2.2.3,主进程/CPU生产进程检查版本并写入合同及权重;保留无梯度AMP采样cache_enabled=False修复、主干梯度检查及严格回放。CPU审计失败时保存最近完成更新并停止。 + +-本地单元测试、过拟合/探针、评测/渲染代码和原始报告集中到`test/`,由`.gitignore`忽略。旧入口的兼容链接也被忽略。GitHub不包含这些代码、模型权重、数据、日志、环境、视频或缓存;CLIP自带的压缩分词词表是必要训练资源。原生validation及断点完整性检查属于训练实现。 ++以下仅用于任务停止后续训;不要对运行中的目录重复启动: + +-进一步阅读:[迁移操作](docs/PORTABLE_TRAINING.md)、[GenMo训练](docs/GENMO_FORMAL_TRAINING.md)、[Kimodo训练](docs/KIMODO_HFD_TRAINING.md)、[文本时间轴](docs/REFERENCE_DYNAMIC_TRAINING.md)、[resume](docs/TRAINING_RESUME.md)、[GenMo倾听姿态诊断](docs/GENMO_LISTENING_DIAGNOSIS.md)。 ++~~~bash ++cd /root/autodl-tmp/hfd-baseline-portable ++/root/autodl-tmp/envs/omg/bin/python -u train.py --config configs/omg.yaml --resume auto --set 'runtime.gpu_ids=[1]' ++/root/autodl-tmp/envs/omg/bin/python -u train.py --config configs/omg_l.yaml --resume auto ++/root/autodl-tmp/envs/omg/bin/python -u train.py --config configs/omg_xl.yaml --resume auto ++/root/autodl-tmp/envs/textop/bin/python -u train.py --config configs/textop.yaml --resume auto --set 'runtime.gpu_ids=[0]' ++/root/autodl-tmp/envs/genmo/bin/python -u train_genmo_full.py --config configs/genmo.yaml --resume auto ++/root/autodl-tmp/envs/kimodo/bin/python -u train_kimodo.py --config configs/kimodo.yaml --resume auto ++CUDA_VISIBLE_DEVICES=1 /root/autodl-tmp/envs/hymotion/bin/python -u train_hymotion_autoregressive.py --config configs/hymotion_c24.yaml --set gpu=0 --set text.device=cuda:0 --resume auto ++~~~ ++ ++HY训练入口为 train_hymotion_autoregressive.py,当前唯一正式配方为configs/hymotion_c24.yaml。train_hymotion.py只提供模型/合同公共函数,不再包含旧训练循环。 ++ ++## HY C24完整训练约定 ++ ++- 用户于2026-10-30授权完整训练及HY专项清理。原始训练数据、split、英文筛选、样本选择、采样顺序和时间映射不变;不加入静止历史增强。恢复监督在训练目标构造中计算,不回写GT。 ++- 从C24验证集选出的best仅加载权重。该best与上一轮C的2000步模型完全一致,已独立复制到共享models/hymotion/hfd-c24-full-split20261026/initialization.pt;初始化来源、SHA256和选择理由保存在新运行目录initialization.json。优化器/步数重置,新增100K预算,不声称完整续接旧优化器。 ++- 保留原XY/yaw对齐;姿态SO(3)差和高度差在24帧内按w=1−3s²+2s³消除。保留flow,根速度权重1.0、旋转增量0.05,前200更新渐启;执行段接段第一帧权重3,FP32几何、无效尾帧屏蔽。 ++- 预测历史策略仍以每链GT历史启动,之后反馈预测;每链10段共4秒,不能描述成跨整个session完全自主训练。训练数据和实际采样约定不因代码清理改变。 ++- 初始和每2000步用6条未参与旧38K选优的validation、两个固定噪声seed、最长20秒自主闭环。best按含运动误差的score选择,并要求平均接段比≤3;best只是当前最优,不等于行为验收合格。 ++- last每256步、验证结束及最终步原子保存;保留initial/best/last,不累计中间断点。启动先完成17步完整断点,再自动恢复原100K预算,检查恢复链路。 ++- 旧split完整训练runs/hfd-hymotion-english及共享models/hymotion/hfd-english保留;近期HY实验目录和中间权重清理后,正式训练不依赖它们。结果摘要见CURRENT_RUNS,必需初始化和基线指标保存在当前运行目录。 ++- 保留数据/连续性/恢复/模型回归检查;清除旧时钟输入、静止/单窗恢复实验分支和一次性对照脚本。当前测试评测入口为test/evaluation/evaluate_hymotion.py,可通过--run/--weights指定运行;仅在用户要求评测时启动GPU评测。 ++ ++## 评测约定 ++ ++自主闭环只开头给GT姿态,之后用预测历史,不逐窗重置GT;不做视频平滑、rootfit或DTW。人体模型先生成完整原始序列,再做整段GMR。训练loss或短smoke通过不等于行为、连续性和根运动验收;验证集用于选权重/调参,测试集只报告固定模型。 ++ ++## 动作评测脚本入口 ++ ++- 移植来源为 Tinghong-Ye22/HFD-motion-evaluation-handoff-20261101@7901e042c9cf06e0081d7303a4b2c229136ec70b;下载原件在 /root/autodl-fs/pangxinlei/downloads/HFD-motion-evaluation-handoff-20261101,工作目录在 /root/autodl-tmp/hfd-motion-evaluation/HFD_bench。 ++- 命令入口 /root/autodl-tmp/hfd-motion-evaluation/bench,使用 configs/local_motion.json;只读看板入口为同目录 review --port 8770。增量环境 /root/autodl-tmp/envs/hfd-motion-eval 复用 OMG 基础环境,不修改原训练环境;实际依赖见工作目录 requirements-local.lock。 ++- 冻结识别器5c04cec、阈值、GT+0.7秒计分参照和原始结果均保留;模型预测不平移。交互衡量发布的motion reference,非实机执行。UNKNOWN、未适用和未响应删失分别保留。 ++- 质量入口 motion-score 调用指定官方源码与本机已有 evaluator/T5-3B/HoloMotion,包含原生29DOF到MuJoCo的置换。旧113587的14项质量分数仍因 invalid_input_joint_order 失效;测试样例通过不代表新模型完整质量评测完成。大型参考缓存、视频和971502训练权重未全部下载,不直接启动旧调度/恢复队列。 ++ ++## 五模型全英文1024条动作测试集 ++ ++- 2026-11-01用户指定用于OMG、TextOp、GenMo、Kimodo、HY-Motion。独立派生集在 `/root/autodl-tmp/datasets/hfd-test-english1024-v1`:保留原1024个test场景及1937个动作轮,共2229个motion事件;原436条英文前景保留,588条中文前景重建为对应英文。动作场景、原始session身份和训练/验证划分不变,没有从其他划分补动作样本。 ++- 文件入口为 `manifest.jsonl`、`events.jsonl`、`membership.jsonl`、`evaluation_contract.json`;每例 `sessions/test-en-NNNN.npz` 包含50Hz IsaacLab G1与配对159D SMPL-X、16维betas、16kHz麦克风和评分标签。原文件保留,新版本不覆盖原始test合同或旧分数。 ++- 生成器与读取接口在 `/root/autodl-tmp/hfd-motion-evaluation/english1024`。使用OMG环境Python;`dataset_loader.English1024.input_at(index, motion_frame, model)` 只给初始已知姿态及已释放文本,后续历史由模型预测;`scoring_targets` 单独返回GT,`paired_human_30hz` 复用原人体几何变换。五模型各用相同1024例,不沿用旧人体provider的419条筛选。 ++- 英文均复用已有审计语音。更换语言后重算麦克风、控制事件和动作时间轴,沿用已发布的30帧终端时钟。2例改用其他合格英文音色,37例因英文提示语更长而最小延长停顿;20例中文第三方干扰换为上游held-out英文干扰,其中7例原前景已为英文。保留干扰类型、事件数及同音色aside关系;重叠背景按不相交聚合片段移动一次,aside避开前景指令。 ++- 文本协议仍为离线前景reference alignment,30秒上下文、0.1秒guard、microphone clock;它是oracle参考转写,不是新ASR结果,guard不代表经认证的时间误差上限。原训练/validation配置和已运行模型均未修改。 ++- 全量核验见数据集 `verification.json`:2002个实际英文语音资产SHA匹配;全部1024例G1/人体/时钟/语言/事件通过;419条原生英文的音频、动作和标签经旧加载器重放逐项一致;436条原英文动作轮标注一致,429例完整motion标注一致(另7例替换了中文背景,浮点时钟按1ns容差比较);35组任务×模型输入检查及2项混音/aside回归通过。 ++- 用户随后确认保留流式输入与listen/idle动作,已启动下述完整流式质量评测。交互识别/评分尚未重跑;本次质量使用预先固定的GT执行窗口,不需要推断预测首尾,也不复用旧971502分数。 ++ ++## 完整流式动作质量评测 ++ ++- 入口 `/root/autodl-tmp/hfd-motion-evaluation/stream-quality`;`status` 查看进度,`run --gpus 0 1` 启动或恢复(运行中会拒绝重复启动),`collect` 汇总已有评分。2026-11-01已启动,脚本在 `/root/autodl-tmp/hfd-motion-evaluation/stream_quality`;模型和训练代码未改。 ++- 协议 `full_stream_fixed_gt_windows_v1`:五模型逐包/逐次读取现有因果输入,单个已知姿态启动,之后只有执行过的预测历史。完整保留listen/idle/执行输出,不按GT任务边界重启,不把完整目标文本提前交给生成器。OMG保留过去2秒PCM及native async缓冲;其余模型保留既有因果reference文本协议,非在线ASR。 ++- 每模型生成全部1024例。1937个GT动作窗口在模型生成前按event ID排序,官方batch32取1920个,其余17个明确列入excluded;五模型成员一致。按固定绝对时间切出预测,响应慢、未响应或仍为listen的输出照常保留;不寻找最像GT的片段,不平移、不DTW、不平滑。模型没有收到这些评分窗口。 ++- 使用现有验证集best:OMG100000、GenMo72000、Kimodo100000、HY14000;TextOp使用最终300000。精确路径/SHA/配置见评测输出run_config.json、checkpoint_metadata.json、preflight.json。人体完整流先整段GMR,再切评分窗口,GMR只在session开头重置。 ++- 大文件在 `/root/autodl-fs/pangxinlei/evaluations/hfd-english1024-stream-quality-v1`;每模型 `models//sessions/` 保存完整预测及因果审计,`phase_diagnostics.json` 单独保存listen/motion/idle的关节、根误差和根速度,`packed/` 保存官方输入,`scores/motion_quality.json` 保存14项结果。汇总自动写summary.json/csv,不编造一个总分;阶段诊断不冒充官方14项。 ++- 日志与进度在 `/root/autodl-tmp/runs/hfd-five-stream-quality-english1024-v1`;两张GPU动态排队,CPU并行整段GMR,模型失败会保留错误并阻止该模型缺样本评分。GPU评分调用指定原版OmniHM/HoloMotion、tracker-mode=run与原有分块FK执行器。 ++- 恢复时校验并复用已完成的1024条生成结果与packed输入。评分经独立`stream_quality/score.py`调用交接包同一`OmniHM.evaluate`接口,将输出写入用户指定共享盘;仍执行冻结配置、官方源码及输入SHA校验。GMR读取进度文件遇到短暂缺失时等待下次读取,其他错误照常报告;summary同步显示运行/失败阶段。 ++- 启动前已通过五模型8秒流式生成、三模型整段GMR、4项因果/固定时窗回归,29个互异关节值经过baseline及指定官方转换后逐项一致。短检查不是完整1024评分,最终状态及分数以实际产物为准。 ++- 2026-11-02已完成五模型全部评分:各1024条完整生成、各1920固定窗口、14项均有值,完整结果见上述summary.json/csv与docs/CURRENT_RUNS.md。跟踪MPJPE等比较HoloMotion执行与生成目标;Diversity的GT参考约1.17965。显示FID已×100、C-Slide已×10,引用时保留该缩放。 ++ ++ ++## 隔离恢复训练 pilot(2026-11-09) ++ ++首轮 HY 与 Kimodo 的独立 500 更新恢复试验已结束,均未通过闭环验收;首轮入口 `/root/autodl-tmp/hfd-motion-evaluation/HFD_bench/scripts/recovery_training/supervise_pilot.py --config /config.yaml`,分别使用对应原训练环境。运行目录 `` 为共享盘 `models/hfd-recovery-pilot-20261109/{hymotion,kimodo}`。必须先有实际容量预留收据;supervisor 先做 2 更新保存与完整恢复检查,再继续固定预算。新试验加入单姿态启动的预测历史链和真实历史回放,HY 去掉 C24 混合标签,Kimodo 保留独立 expert 速度标签;从旧 best 显式权重初始化,保持原推理 steps/CFG,旧正式训练及已有成绩不替换。 ++ ++权重、缓存、临时文件和日志都有独立预算与持续磁盘/inode 检查。验证新增动作幅度、进度和误差非退化门槛;只有合格闭环 checkpoint 可成为新 best。500 更新结束不自动延长,不把 loss 降低等同于 FID 提升。具体预算、选优门槛、实时状态路径见 docs/CURRENT_RUNS.md。若受控停训,先排查对应错误收据并重新验证可写空间,再严格恢复,不能用删除合同检查的方式续跑。 ++ ++### 当前第二轮恢复试验 ++ ++第二轮已启动,入口改为 `HFD_bench/scripts/recovery_training_v2/supervise_pilot.py --config /config.yaml`,运行目录为共享盘 `models/hfd-recovery-pilot-v2-20261109/{hymotion,kimodo}`。从各自原正式 best 重新初始化,各 250 更新,0/125/250 固定验证;首轮输出保留,不采用其退化 last 替换正式模型。历史采样、推理 steps/CFG 和验收条件保持。 ++ ++监督改为 XY/yaw 对齐、六帧 SO(3)/高度过渡,第六帧起回到对齐 GT;位置与导数 loss 使用同一目标。HY 连续性辅助项恢复,warmup 200 更新,root velocity 权重 1、rotation velocity 0.05、边界权重 3。仅修改训练标签,推理不加入 GT 或平滑。增加段内突变、根 jerk 诊断及三次完整验证轨迹;125 更新额外保留一份权重快照(Kimodo 为 model-only,完整恢复使用 last)。只有原定门槛全部通过才写 best,尚无 FID 改善结论。 ++ ++实际新增容量预留 HY 2 GiB、Kimodo 8 GiB 已通过;checkpoint 单次写入上限分别为 256/2560 MiB,其他资源阈值和断点保护沿用首轮。第二轮进度命令为对应 Python 执行 `HFD_bench/scripts/status_recovery_training.py --root /root/autodl-fs/pangxinlei/models/hfd-recovery-pilot-v2-20261109`。证据与当前状态见 docs/CURRENT_RUNS.md。 ++ ++迭代资源维护:每轮核查已失效临时文件和可再生的重复缓存,确认进程结束、无打开句柄和保留副本有效后及时清理;当前运行的容量预留、有效断点和诊断证据保留。日志轮转、缓存限额、受控结束释放剩余预留继续执行。 ++ ++用户已授权每小时自主检查 HY/Kimodo 恢复训练,完成后分析并持续迭代,及时清理无用中间产物;验证效果有明确希望且空间预算通过时,可直接展开分阶段正式训练,无需例行再次确认。当前会话心跳为每小时整点(Etc/UTC),复用原 auto-review 权限;具体任务、状态与操作范围见 HFD_bench/artifacts/recovery_hourly_task.json 及 recovery_hourly_task_request.json。只用 train/validation 选参,保留既定验收,更多 validation 复核后再升级正式训练,最终测试协议保持。 ++ ++ ++### 第三轮恢复试验(2026-11-09 11:14 UTC) ++ ++第二轮两模型均已完成 250 更新,均未通过门槛。HY 最终活跃动作幅度比 0.358、姿态误差 0.443 rad(初始 0.372,恶化约 19%),行走案例 961 仍未响应;Kimodo 幅度比 0.842、接段比 2.355,但根位置误差 0.527 m(初始 0.442,恶化约 19%)且行走结果不稳定。不能升级正式训练,也没有新的正式 FID。 ++ ++当前最新目录为 `/root/autodl-fs/pangxinlei/models/hfd-recovery-pilot-v3-20261109/{hymotion,kimodo}`,入口 `HFD_bench/scripts/recovery_training_v3/supervise_pilot.py`。各从原正式 best 初始化,仍为 250 更新及固定 0/125/250 验证。每模型只改变一个训练变量:HY 预测历史链概率 0.75→0.25,增加真实历史回放;Kimodo 根网络从仅投影训练改为完整解冻,学习率不变。保留 C6 标签、因果输入、两随机种子与全部验收门槛。监督器已启动,先执行初始验证及两更新保存/严格恢复检查,未通过不会继续剩余预算。 ++ ++实际写入/fsync 预算 HY 2 GiB、Kimodo 12 GiB 已通过,Kimodo 完整断点预计 3.2 GiB、单次原子写入上限 4 GiB。第二轮重复启动索引已清理 656.24 MiB,正式权重、恢复断点、125 快照和完整轨迹保留。逐例结果、配置差异、启动与清理清单见 HFD_bench/artifacts/recovery_hourly_20261109T11_analysis.json、recovery_v3_config_diff.json、recovery_v3_launch.json、recovery_cleanup_hourly_20261109T11.json。每小时任务的 latest_run_root 已更新为 v3。 ++ ++ ++用户最新要求(2026-11-09):任一模型正式训练成功启动并确认进程、日志、容量及恢复检查正常后,立即删除每小时心跳任务 `416e9f54`,不等待训练或最终评测完成,不自动重建。当前有界恢复 pilot 仍继续每小时检查;取消心跳不停止训练,也不关闭训练监督器的资源与断点保护。此要求覆盖此前“最终评测完成后取消”的旧条件。 ++ ++ ++### 每小时检查与第四轮 Kimodo(2026-11-09 12:04 UTC) ++ ++当前两模型轮次不同:HY 仍在 `/root/autodl-fs/pangxinlei/models/hfd-recovery-pilot-v3-20261109/hymotion`,Kimodo 最新为 `/root/autodl-fs/pangxinlei/models/hfd-recovery-pilot-v4-20261109/kimodo`,以 `artifacts/recovery_hourly_task.json` 的 latest_models 分别定位。尚未启动正式训练,小时任务保持;任一正式训练健康启动后立即取消任务的要求不变。 ++ ++HY v3 在本次检查期间推进至 168/250,loss/梯度有限、无错误,完整两更新恢复检查通过(断点 136,897,993 bytes)。125 验证幅度比 0.261、接段比 1.672、姿态误差 0.423 rad、根位置误差 0.487 m,未通过;维持原有 250 更新上限,不延长或重启。 ++ ++Kimodo v3 已完成 250,两更新完整断点 3,394,252,981 bytes、恢复校验通过,无数值或资源故障。最终幅度比 0.942、接段比 2.299、姿态误差 0.422 rad,但根位置误差从 0.442 增至 0.675 m(约 +53%),根 jerk RMS 从 357.30 增至 700.46 m/s³。行走案例 961 两种子进度比 -0.466/-0.055;案例 2327 最大根速度仍有 10.98/18.81 m/s,尖峰仍落在每段第二帧。125 同样不合格。已读取 0/125/250 全部逐例完整轨迹,不能将动作幅度达标当成恢复成功,也无新的正式 FID。 ++ ++Kimodo v4 从相同原正式 best 重新开始,相对 v3 只新增 stage1_root_velocity 权重 0.5:对反归一化第一阶段根位置求物理速度,与相同 C6 因果目标比较,包含历史/未来边界并屏蔽无效帧。目标是阻止第一阶段错误被身体骨盆偏移短暂抵消。保持完整根网络解冻、学习率、250 更新、固定 0/125/250 验证、两个种子和原验收门槛;不改变推理或加入 GT 条件。新损失通过真实表征检查:精确目标为零、补偿式跳变仍被识别、梯度有限且仅作用于根分量、无效尾帧不计分,见 artifacts/recovery_v4_objective_checks.json。源码入口 scripts/recovery_training_v4/supervise_pilot.py;准备与运行独立于 HY v3。 ++ ++第四轮 12 GiB 预算已实际写入/fsync,通过物理分配校验;4 GiB 原子写入上限、8 小时墙钟及原存储阈值不变。监督器已启动,先进行初始验证和两更新完整恢复检查,该检查尚待本轮执行完成。当前本地约 1.54 GiB 空闲、共享约 34,900 空闲 inode,小文件写入/fsync 通过;没有把共享 df 总空闲当账号配额。 ++ ++清理 v3 Kimodo 的 16 个重复启动索引及已分析且不再使用的 125 步落败快照,共 1,407.06 MiB。已核对进程结束、无打开句柄、无配置引用、缓存保留副本 SHA 相同;保留最终完整断点、原正式 best、必要旧对照及全部验证原始轨迹。清单 artifacts/recovery_cleanup_hourly_20261109T12.json。分析、配置差异与启动收据分别为 recovery_hourly_20261109T12_analysis.json、recovery_v4_config_diff.json、recovery_v4_launch.json。下一检查点为 HY 最终验证及 Kimodo v4 恢复/125 验证,不因平均指标改善放宽逐例行走门槛。 ++ ++ ++### 第五轮会话覆盖对照(2026-11-09 13:05 UTC) ++ ++最新两模型统一位于 `/root/autodl-fs/pangxinlei/models/hfd-recovery-pilot-v5-20261109/{hymotion,kimodo}`,入口 `HFD_bench/scripts/recovery_training_v5/supervise_pilot.py`。HY v3 和 Kimodo v4 均已完成 250 更新,完整恢复检查通过、训练数值有限、无资源错误,但未通过闭环门槛。尚未正式开训,小时检查继续;任一正式训练健康启动后立即取消心跳。 ++ ++HY v3 最终幅度比 0.263、接段比 1.686、姿态误差 0.427 rad、根位置误差 0.467 m、根 jerk 279.24 m/s³;两行走案例四种子组合进度比 0.336/-0.037/0.352/-0.010,均低于下限。Kimodo v4 最终幅度比 0.856、接段比 2.336、姿态误差 0.408 rad、根位置误差 0.564 m、根 jerk 376.07 m/s³。相对 v3 根 jerk 改善,但行走仍不可靠:案例 2327 两种子进度比 0.046/0.790,961 为 -0.168/0.653。v4 的 125 步姿态/位置非退化门槛通过,但两例进度未通过,故保留该快照诊断,不升级正式训练。已读取各轮初始/125/250 全部逐例完整轨迹;这些不是正式 FID。 ++ ++采样链路发现:WindowPool.refresh 按链起点批次计数,不按优化器更新计数。Kimodo v4 最终完整断点的 pool.step=6、refresh=16、slots=[32781,251404,112021,179604],证明 250 更新一直使用最初四个会话;HY v3 pool.step=7、refresh=2,日志实际覆盖 16 个不同会话。长链重复同一池显著限制本轮样本覆盖,但不能据此认定它是质量差的唯一原因。第五轮每模型只将 data.refresh_batches 改为 1(Kimodo 16→1,HY 2→1),每批新链重新抽取会话;仍保留最多 50 段闭环链、原概率、各自上一轮损失和学习率、同一原正式 best 初始化及全部验收门槛,各 250 更新,0/125/250 验证。没有读取 Test1024 选参。Kimodo 日志新增实际采样位置和池计数,便于后续核查覆盖。 ++ ++真实 WindowPool 的受控测试验证六批起点在 refresh16/2/1 时分别覆盖 4/12/24 个模拟会话,保存并恢复后下一批完全一致;实际覆盖仍由运行日志判断。检查和差异收据为 artifacts/recovery_v5_sampling_checks.json、recovery_v5_config_diff.json,启动源码 SHA 与进程在 recovery_v5_launch.json。原因果标签及推理实现未改;两更新完整断点/严格恢复仍由监督器先验证,第五轮目前处于该启动阶段,尚无训练后改善结论。 ++ ++HY 2 GiB、Kimodo 12 GiB 预算已重新实际写入/fsync 并确认物理分配;本地约 1.54 GiB 空闲,共享约 34,800 空闲 inode。保持原子替换预算、8 小时墙钟、目录项/缓存/日志上限及资源受控停止,不将共享 df 视为账号配额。清理结束轮次的重复启动索引及 HY v3 125 落败快照 786.77 MiB,已检查无进程占用、无必要配置引用、缓存 SHA 一致,保留当前完整断点、Kimodo v4 125 候选及全部轨迹。清单 recovery_cleanup_hourly_20261109T13.json,完整分析 recovery_hourly_20261109T13_analysis.json。状态脚本已修复缺少某模型目录时的报错,返回 present=false;没有修改任何旧训练合同源码。 ++ ++下一检查点为第五轮两模型恢复收据、实际会话覆盖以及 125/250 固定验证。失败时继续依据证据调整,不能仅延长当前失败配方或降低验收条件。 ++ ++ ++### 小时检查与 Kimodo 第六轮(2026-11-09 14:07 UTC) ++ ++当前 HY 仍为 `models/hfd-recovery-pilot-v5-20261109/hymotion`,Kimodo 最新为 `models/hfd-recovery-pilot-v6-20261109/kimodo`,根目录均在 `/root/autodl-fs/pangxinlei/`,以任务 JSON 的 latest_models 为准。没有正式训练启动;达到正式启动并核实健康条件后,仍应立即删除心跳 416e9f54。 ++ ++HY 本次检查从 138 推进到约 154/250;恢复检查通过(136,897,929 bytes)、训练数值有限、无错误,实际已记录 17 个训练会话。125 验证幅度比 0.252、接段比 1.661、姿态误差 0.424 rad、根位置误差 0.478 m,未通过。保持当前有界训练与原定最终验证,不抢占 GPU0。 ++ ++Kimodo v5 完成 250 更新,严格恢复通过(3,394,253,109 bytes),实际训练覆盖 25 个会话。最终幅度比 0.919、接段比 2.365、姿态误差 0.409 rad、根位置误差 0.371 m;相比原初始 0.468 rad/0.442 m,两种误差均改善。但行走案例 2327 两种子进度 0.496/0.519、961 为 0.547/0.227,仍有低于 0.5 的案例,不能四舍五入放宽门槛。根 jerk RMS 411.10 m/s³,高于初始 357.30,段首尖峰仍存在。125 的根误差 0.592 m、jerk 606.01,亦未通过。已读取 0/125/250 全部逐例原始轨迹;平均数改善不等于正式训练资格,也不是 FID 改善。 ++ ++完成同一 v5 final 权重下四个 validation 历史窗口、两种子、预测过去与真实过去的 16 次配对诊断,输出 `/root/autodl-fs/pangxinlei/models/hfd-recovery-history-diagnostics-v5-20261109/`。输入仅为截至 decision 的历史,无未来 GT;真实历史是诊断对照,不能作为正式验证或替代闭环成绩。案例 2327 decision396 的第二帧最大根速度从预测历史约 5.96/5.97 m/s 降至真实过去 0.12/0.32,但 decision504 真实过去仍约 2.85/2.63;961 decision276 真实过去仍有 2.32/4.12 m/s。说明历史累积参与部分异常,段首问题也能在真实过去下出现。低速度窗口可能只是静止,不能认作正确完成行走。 ++ ++第六轮从同一原正式 best 重新开始,保持第五轮采样刷新1、原链长和概率、完整根解冻、学习率、stage1 速度权重0.5、250 更新及两种子固定验证,只新增 stage1_boundary_weight=3:第一阶段根速度的历史末帧→首未来帧及首→第二未来帧两条差分加权3倍。推理无平滑、无 GT,验收不变。真实表征检查确认边界/内部误差比3、精确目标零、无效尾帧忽略、有限且仅作用于根分量的梯度;收据 recovery_v6_boundary_checks.json,配置差异 recovery_v6_config_diff.json,入口 scripts/recovery_training_v6/supervise_pilot.py。监督器已启动,初始验证和两更新完整恢复检查尚待完成,不声称本轮已改善。 ++ ++新增 12 GiB 预算已实际写入/fsync 并验证物理分配,保留4 GiB原子 checkpoint 上限、8小时墙钟及原资源门槛。本地约1.54 GiB空闲、共享约34,750空闲inode,小文件写入/fsync通过。已清理 v5 Kimodo 125 落败快照及结束诊断后的重复索引共1,407.06 MiB;无进程句柄和必要配置引用,缓存保留副本 SHA 一致。保留 v5 最终完整断点、v4 125候选、所有轨迹与活跃 HY 预留。诊断预留已释放。清单 recovery_cleanup_hourly_20261109T14.json;完整分析 recovery_hourly_20261109T14_analysis.json;启动收据 recovery_v6_launch.json。下一检查点是 HY 最终结果、Kimodo v6恢复收据与固定验证,而非直接延长未通过配方。 ++ ++ ++### 第七轮监督权重对照(2026-11-09 15:06 UTC) ++ ++当前两模型均进入 `/root/autodl-fs/pangxinlei/models/hfd-recovery-pilot-v7-20261109/{hymotion,kimodo}`,入口 `HFD_bench/scripts/recovery_training_v7/supervise_pilot.py`。上一轮 HY v5、Kimodo v6 均正常完成250更新,恢复检查通过、训练数值有限,未通过闭环验收;无正式训练,继续小时调度。任一正式训练成功启动并验证健康后立即取消心跳的要求保持。 ++ ++HY v5 最终幅度比0.264、接段比1.666、姿态误差0.429 rad、根位置误差0.487 m、根jerk285.94 m/s³。行走案例2327两种子进度0.279/0.289、961为-0.008/0.004,明显不合格。Kimodo v6 最终幅度比1.018、接段比2.283、姿态误差0.412 rad、根位置误差0.362 m、根jerk330.86 m/s³;虽然jerk低于原初始357.30,进度仍不可靠:2327为0.114/0.510,961为0.885/0.285。两个模型都已读取初始/125/250全部逐例完整轨迹,保存位置/姿态/速度误差、幅度、接段/段内突变、jerk与路径;中途候选也未通过,不用平均改善掩盖逐例失败,未计算新的正式FID。 ++ ++第七轮每模型仅改一个参数,保持原正式best初始化、固定训练种子、refresh1会话采样、原链长/概率、学习率、C6标签、推理及原验收不变,仍250更新、0/125/250固定验证。HY根速度辅助权重1→0.1,旋转速度权重0.05和warmup200保持;依据是v5后期根速度项数值均值约0.102,大于flow约0.071,同时幅度和姿态退化,这只是权重失衡假设,不能将损失数值比较当成梯度主导的证明。Kimodo closed_loop.tail_weight由0.25→1,完整监督原有60帧预测,不再将后48帧降权;检验局部接段改善却仍走不够是否与短时域拟合有关。执行仍只用前12帧,stage1_boundary_weight=3保留,不加推理平滑或未来GT。 ++ ++真实函数检查确认尾段权重1时原有效性掩码、历史条件和未定义末帧导数掩码完全保留、输入不被修改;HY调度只改变根速度系数。检查、配置差异、源码SHA与启动收据为 artifacts/recovery_v7_weighting_checks.json、recovery_v7_config_diff.json、recovery_v7_launch.json。两监督器真实进程已启动,先初始验证和两更新完整保存/严格恢复,本轮恢复检查尚未结束,不能声称训练效果已改善。本轮选参只用train/validation。 ++ ++已重新实际写入/fsync HY2 GiB与Kimodo12 GiB预算并验证物理分配,原子替换上限256 MiB/4 GiB、8小时墙钟、目录项/缓存/日志及空间inode护栏均保持。本地约1.54 GiB空闲,共享约34,690空闲inode,小文件实写/fsync通过,未把共享挂载总空闲视为账号配额。 ++ ++清理约4.98 GiB:已结束轮次的重复启动索引、HY v5及Kimodo v6的125步落败快照,以及不再需要的Kimodo v3最终完整断点。删除前核对运行结束、无打开句柄和配置输入依赖,缓存保留副本哈希一致。v3的配置/源码/完整原始轨迹/指标仍保留;保留当前HY v5、Kimodo v6完整断点、较好Kimodo v5完整候选、v4的125候选和监督审计所需旧断点,不无限累积落败权重。清单 recovery_cleanup_hourly_20261109T15.json,完整分析 recovery_hourly_20261109T15_analysis.json。下次先检查v7恢复与资源健康,任一完成即分析逐例,不机械延长失败配方。 ++ ++ ++### 小时检查与 Kimodo 原生初始化对照(2026-11-09 16 UTC) ++ ++HY v7仍运行,检查时194/250更新,数值有限、两步完整恢复检查通过、无错误。125验证幅度比0.269、姿态误差0.443 rad、根位置误差0.464 m,不合格,继续原定有界运行。 ++ ++Kimodo v7正常完成250更新、完整恢复通过,已读取初始/125/250全部逐例轨迹。最终幅度比0.770、接段比2.394、姿态误差0.422 rad、根位置误差0.376 m;原正式初始分别为幅度0.405、姿态0.468 rad、根位置0.442 m。行走案例2327两种子进度0.513/0.896、961为0.911/0.346;最后一个未达到0.5。根jerk511.39 m/s³高于初始357.30和v6最终330.86;段内姿态变化p95均值0.07275 rad,GT为0.05927。125根误差0.496 m、jerk643.45亦不合格。不能因平均误差改善晋升正式训练,也未计算新FID。 ++ ++第八轮仅Kimodo:`/root/autodl-fs/pangxinlei/models/hfd-recovery-pilot-v8-20261110/kimodo`,入口`HFD_bench/scripts/recovery_training_v8/supervise_pilot.py`。目录日期采用当前会话日期,检查收据时间来自机器UTC。只把initial_checkpoint从旧HFD正式best改为None,由上游配置加载原生Kimodo-SMPLX-RP-v1/model.safetensors。CPU逐张量精确核验408个张量、282790161个元素与原文件完全一致,非随机初始化。其余学习率、完整根训练、250更新、采样、C6标签、因果输入、推理及0/125/250两种子验证不变。复制原HFD初始baseline-validation.json且逐字节核对一致,误差门槛仍以原正式初始为准,不能因原生初始化改变验收。 ++ ++新监督器558793、子进程558794已核实存活,处于初始验证/两更新恢复检查阶段;本轮恢复尚未通过,效果未知。收据artifacts/recovery_v8_initialization_check.json、recovery_v8_config_diff.json、recovery_v8_launch.json。新预算12 GiB已实写/fsync并验证物理分配,4 GiB原子保存上限及原空间/inode/目录项/日志/8小时保护保持。本地约1.54 GiB可用、共享约34600空闲inode,本地及共享小文件实写通过;不将共享df当账号配额。 ++ ++已清理v7 Kimodo重复索引和125落败快照1407.06 MiB,核实结束进程、无打开句柄、快照无配置输入引用,索引与v8保留副本SHA一致。保留v7最终完整恢复断点、必要旧候选、配置与全部轨迹;活跃HY预留不动。清单artifacts/recovery_cleanup_hourly_20261109T16.json;逐例分析recovery_hourly_20261109T16_analysis.json。 ++ ++无正式训练启动,小时心跳416e9f54继续;一旦正式训练成功启动并验证健康,立即删除心跳,不停止训练监督器。下次先检查HY最终结果和Kimodo v8完整恢复、初始及训练后闭环验证。不把本次原生初始化假设当成已证实改善。 ++ ++ ++### 第九轮有界对照(2026-11-09 17 UTC) ++ ++HY v7、Kimodo v8均完成250更新,完整两更新保存/恢复通过,训练及已读取0/125/250全部逐例轨迹数值有限,无错误或资源受控停止。没有正式训练;小时心跳416e9f54保持,正式训练成功启动并核实健康后立即删除。 ++ ++HY v7降低根速度辅助权重后,最终幅度比0.272、接段比2.044、身体姿态误差0.450 rad、根位置误差0.478 m、jerk291.97 m/s³。原初始姿态0.372 rad,最终退化超过10%,且有1个幅度坍缩案例;行走2327两种子进度0.350/0.315,961约0/0.005。中途125亦未过关,不能将较低jerk解释为动作恢复。 ++ ++Kimodo v8原生初始化在当前因果适配接口下,初始幅度1.258、姿态0.603 rad、根位置0.608 m、jerk723.84;最终幅度1.545、接段比2.948、姿态0.601 rad、根位置0.569 m、jerk758.54。段内姿态变化p95均值0.10262 rad(GT0.05927);行走2327进度0.014/-0.019,961为-0.029/-0.036。125也不合格。原生权重增加运动幅度但未恢复正确进度,不支持“换回原生权重即可解决”,也不能据此否定原生模型在其原协议下的能力;当前结果包含因果表征、历史输入及任务域的适配影响。未计算新FID。 ++ ++下一轮根目录`/root/autodl-fs/pangxinlei/models/hfd-recovery-pilot-v9-20261110/{hymotion,kimodo}`,入口`HFD_bench/scripts/recovery_training_v9/supervise_pilot.py`。HY以v7为对照,仅initial_checkpoint改为None,build_model仍加载官方latest.ckpt。CPU核对302个backbone张量及null文本参数与官方文件完全一致,新增观察列为0、LoRA B为0,非随机网络;其余损失和配置不变。Kimodo以更好的v7为对照(不是从v8续训),保持原正式best初始化,仅新增stage1_root_displacement权重0.5:对有效未来帧的第一阶段物理根位置相对最后历史帧的位移,用beta0.1的SmoothL1监督并按score掩码归一。检验短时速度监督下仍走不够的问题;不改变推理、标签、历史构造或执行长度。测试确认精确目标损失0、无效尾段与空掩码忽略、整体平移不变、有限且仅根通道梯度。此项是假设检验,不预设改善。 ++ ++两模型仍250更新、0/125/250固定validation与两个种子,train/validation选参,Test1024未使用。原HFD初始baseline-validation.json逐字节保留,误差门槛不随新初始化放宽。配置差异仅指定变量与路径/元数据,收据artifacts/recovery_v9_config_diff.json;测试recovery_v9_displacement_check.json、recovery_v9_hy_initialization_check.json;启动源码SHA和进程分别recovery_v9_hymotion_launch.json、recovery_v9_kimodo_launch.json。两监督器及子进程已启动,目前初始验证/两更新恢复检查尚未结束,不声称本轮恢复或效果已通过。 ++ ++HY2 GiB、Kimodo12 GiB预算重新实写/fsync并验证物理分配,256 MiB/4 GiB原子checkpoint上限、8小时墙钟、空间/inode/日志/文件数限制保持。本地约1.54 GiB可用、共享约34500空闲inode,另做本地/共享小文件实写;共享df不代表账号配额。清理HY v7与Kimodo v8重复启动索引及125落败快照1865.72 MiB,核实进程结束、无打开句柄、缓存与v9副本SHA一致、快照无配置输入引用。保留最终完整断点、原正式权重、少量既有对照、全部轨迹。清单artifacts/recovery_cleanup_hourly_20261109T17.json,完整分析recovery_hourly_20261109T17_analysis.json。下次先检查v9恢复收据和资源健康,完成后按原门槛分析逐例。 ++ ++ ++### 小时检查与 Kimodo 第十轮(2026-11-09 18 UTC) ++ ++HY v9仍运行,检查时195→198/250更新,真实进程560419存活,GPU0约20.6 GiB,loss和已记录梯度有限,完整两步恢复通过,无错误或资源停止。125验证幅度1.952、接段比5.406、姿态误差0.925 rad、根位置0.717 m,明显未通过。保留当前有界实验等待最终验证,不重复启动或占用其GPU。 ++ ++Kimodo v9完成250,完整恢复检查通过,无非有限值,实际训练覆盖25会话,记录中预测历史槽位占比0.835(设定抽样概率0.75;长链会影响实际槽位占比)。已读取初始/125/250全部逐例完整轨迹及指标。最终幅度0.730、接段2.450、姿态0.404 rad、根位置0.396 m、jerk353.50 m/s³;与v7的0.422 rad、0.376 m、511.39相比,姿态与jerk改善,根位置略差但仍优于原初始0.442 m。段内姿态p95均值0.06497 rad,GT0.05927。没有幅度坍缩,但2327两种子进度0.332/1.090、961为0.517/0.261,仍有两个低于0.5。125根误差0.497 m、jerk642.08、961种子2026进度0.272,亦不合格。不能据此正式开训,不将诊断指标当FID。 ++ ++第十轮仅Kimodo,目录`/root/autodl-fs/pangxinlei/models/hfd-recovery-pilot-v10-20261110/kimodo`,入口`HFD_bench/scripts/recovery_training_v10/supervise_pilot.py`。以v9为对照,仅pilot.predicted_probability由0.75降至0.25;假设长预测历史反馈误差限制正确运动监督,增加真实过去训练占比进行检验。保持原正式best初始化、位移权重0.5、所有损失、学习率、采样刷新、链长、C6标签和250更新。运行源码除准备器外逐字节一致;推理完全闭环、无未来GT,不把真实过去诊断当正式成绩。原初始baseline-validation.json逐字节保留,验收与0/125/250两种子协议不变,Test1024未使用。配置核验artifacts/recovery_v10_config_diff.json,启动进程和源码SHA在recovery_v10_launch.json。 ++ ++12 GiB预算已重新实写/fsync且物理分配通过,保留4 GiB原子checkpoint上限、8小时墙钟、空间/inode/文件数/日志保护。监督器561515及子进程已核实启动,处于初始验证和两步恢复检查,尚无本轮改善结论。本地约1.54 GiB可用、共享约34480空闲inode,本地与共享小文件实写通过;不把共享df当账号配额。活跃HY预留不动。 ++ ++清理4.535 GiB:v9 Kimodo重复启动索引、125落败快照,以及已分析且无后续用途的v8 Kimodo原生初始化最终权重。删除前核实完成收据、进程结束、无打开句柄、无配置输入引用,缓存与v10副本SHA一致。保留v9当前完整断点、较优既有对照、原正式权重,以及v8配置/全部轨迹/指标/原生权重核验;不再无限积累落败完整权重。清单artifacts/recovery_cleanup_hourly_20261109T18.json,完整分析recovery_hourly_20261109T18_analysis.json。 ++ ++当前无正式训练,心跳416e9f54继续。下次先检查HY最终结果与Kimodo v10恢复收据、资源和逐例验证;任何正式训练健康启动后立即取消心跳,不终止监督器。 ++ ++ ++### 小时检查与固定权重推理诊断 v11(2026-11-09 19 UTC) ++ ++HY v9及Kimodo v10均正常完成250更新,完整两步恢复通过,训练loss/已记录梯度与0/125/250全部读取的完整轨迹有限,无错误、无资源停止、预留已释放。均未通过固定门槛,没有正式训练,小时心跳416e9f54保持。 ++ ++HY v9原生初始化最终幅度1.888、接段5.423、姿态0.882 rad、根位置0.710 m、jerk2306.62 m/s³;原生初始分别为幅度1.694、接段5.710、姿态1.061 rad、根位置0.811 m、jerk2498.46。训练有局部改善但仍远差于原HFD正式初始姿态0.372/根位置0.449,验收不随初始化放宽。行走2327进度0.024/0.116、961为-0.173/-0.014;无幅度坍缩不代表正确运动。段内姿态p95均值0.11329 rad,GT0.05927。125也不合格。 ++ ++Kimodo v10降低预测历史抽样后,最终幅度0.723、接段2.020、姿态0.460 rad、根位置0.429 m、jerk408.32;125分别0.616、2.144、0.466 rad、0.431 m、329.05。最终行走2327进度1.097/0.882,但961为-0.015/0.285;较v9的961进度0.517/0.261,第一种子退化。因此不接受“增加真实历史一定改善”的假设,也不凭更平滑中途结果正式开训。没有新的正式FID。 ++ ++连续单变量训练配方未带来稳定逐例通过,本轮先做有界推理诊断,暂不再追加新的损失权重或延长失败训练。目录`/root/autodl-fs/pangxinlei/models/hfd-recovery-inference-v11-20261110/{hymotion,kimodo}`,入口`HFD_bench/scripts/recovery_inference_v11/supervise_pilot.py`。HY固定原HFD正式best,比较10/50采样步;Kimodo固定v9最终完整权重,比较20/100步。各自CFG、相同6个validation案例、两个种子、因果历史、坐标转换、执行长度和原验收不变;不使用Test1024。先跑原步数对照用于复现已有结果,然后跑更细采样。先前Kimodo单窗口20/100步诊断不能替代这次完整闭环比较。 ++ ++诊断分支在优化器创建前返回,零训练更新、不保存训练断点,不计作正式训练或恢复续训;无需伪造两步恢复检查。复用原监督器空间/inode/目录项保护、日志轮转和受控终止,单次墙钟上限4小时,完成后释放预留。静态检查artifacts/recovery_v11_static_check.json,固定输入路径和SHA及基线一致性recovery_v11_inputs.json,启动源码SHA/进程recovery_v11_launch.json。输出为logs/diagnostic.jsonl及validation-traces/step-000010或000050等(此处编号是采样步数,不是训练更新);完成收据标记training_updates=0。状态脚本last_validation可能为空,应读取diagnostic.jsonl。 ++ ++两监督器562275/562278及子进程562279/562280已核实存活,无启动错误;结果尚待完成。每模型2 GiB预算已实写/fsync且验证物理分配,因无新训练checkpoint只预算有界缓存、日志和两套轨迹。本地约1.54 GiB可用、共享约34400空闲inode,本地/共享小文件实写通过,不将共享df当配额。 ++ ++清理1.822 GiB:HY v9和Kimodo v10已结束轮次的重复启动索引、125不合格快照;删除前核验进程结束、无打开句柄、快照无配置引用、缓存与诊断副本SHA一致。保留最终完整断点、原正式权重、诊断所用Kimodo v9完整权重、其他必要对照及全部原始轨迹。清单artifacts/recovery_cleanup_hourly_20261109T19.json;完整分析recovery_hourly_20261109T19_analysis.json。下次检查诊断完成/错误、原步数复现及增加步数的逐例影响,再据此决定训练改动;不将诊断直接当正式训练资格。任一正式训练健康启动后立即取消心跳的条件保持。 ++ ++ ++### 小时检查、HY精度修复与Kimodo v12(2026-11-09 20 UTC) ++ ++v11两模型诊断均正常完成,零训练更新,无错误/资源停止,预留释放。读取全部两套完整逐例轨迹,数值有限,输入权重SHA与启动前完全一致。没有正式训练,心跳416e9f54继续。 ++ ++Kimodo20步与v9最终全部预测及参考轨迹逐元素相同,确认复现。100步幅度1.004、接段2.255、姿态0.451 rad、根位置0.434 m、jerk542.36 m/s³;20步分别0.730、2.450、0.404 rad、0.396 m、353.50。100步行走2327进度0.670/0.863,961为-0.003/0.487,未达标且jerk升高,不能用更多采样步替代训练修复,不采用100步为新默认。 ++ ++HY v11发现诊断入口缺少原训练验证的BF16 autocast上下文。10步原协议复现失败,预测轨迹最大元素绝对差0.5620(混合位置/旋转通道,不作为单一物理量解释);其jerk255.71与原基线335.19不同。v11的10/50步结果仍保留作FP32诊断证据,但不能用于判断原BF16协议增加步数的收益,也不能与既有BF16指标混合统计。原因是新增诊断分支只包torch.no_grad,原validate同时包torch.autocast;旧训练结果未受该新增分支影响。 ++ ++HY v12复制独立源码到scripts/recovery_inference_v12,添加与原validate一致的BF16 autocast和validation seed;目录`/root/autodl-fs/pangxinlei/models/hfd-recovery-inference-v12-20261110/hymotion`。仍固定原正式best、10/50步、同一6案例与两个种子,零更新,4小时监督上限,原误差基线不变。监督器563273、子进程563276已核实存活。下一轮必须先验证10步复现;若仍不一致,继续排查文本缓存/随机上下文,不先宣布改善。诊断输出logs/diagnostic.jsonl、validation-traces,编号是采样步数。 ++ ++Kimodo v12以v9为基线,仅预测历史概率0.75→1.0。v10增加真实历史后961第一种子进度从v9的0.517退到-0.015,因此测试增加闭环暴露,而非延长失败预算。目录`/root/autodl-fs/pangxinlei/models/hfd-recovery-pilot-v12-20261110/kimodo`,入口scripts/recovery_training_v12/supervise_pilot.py。保持原正式best初始化、位移监督0.5、所有其他损失/学习率/采样刷新/链长/推理、250更新和0/125/250固定两种子验证;不把v10结果作为初始化,不改验收,不用Test1024。除准备器外训练与推理源码和v9逐字节一致,原baseline-validation.json逐字节一致。监督器563316及训练子进程已启动,目前初始验证/两更新完整恢复检查阶段,恢复检查尚未完成、效果未知。 ++ ++HY2 GiB、Kimodo12 GiB预算重新实写/fsync并确认物理分配,保留原资源保护;本地约1.54 GiB可用、共享约34340空闲inode,本地与共享写入探针通过,不把df总量当账号配额。清理v11已完成诊断的重复索引656.24 MiB,核对进程结束、无打开句柄和v12保留副本SHA一致,保留所有诊断轨迹/指标/配置与正式及必要候选权重。清单artifacts/recovery_cleanup_hourly_20261109T20.json;完整分析recovery_hourly_20261109T20_analysis.json;配置/静态核验recovery_v12_config_diff.json、recovery_v12_static_check.json;两启动收据recovery_v12_hymotion_launch.json、recovery_v12_kimodo_launch.json。任一正式训练健康启动后立即删除心跳的要求保持。 ++ ++ ++### 小时检查与文本条件诊断 v13(2026-11-09 21 UTC) ++ ++HY v12固定权重诊断和Kimodo v12训练均正常完成,预留已释放,无错误/资源停止。已读取所有完整轨迹且有限;Kimodo训练loss/记录梯度有限、完整两更新恢复通过。无正式训练,心跳416e9f54继续。 ++ ++HY修复BF16上下文后,10步所有预测/参考轨迹与原正式初始逐元素完全一致,确认复现问题解决。50步幅度0.489、接段1.324、姿态0.404 rad、根位置0.498 m、jerk304.67;10步分别0.291、1.290、0.372 rad、0.449 m、335.19。50步幅度仍低于0.5,根位置超过原初始10%上限,行走2327进度1.188/1.869、961为-0.222/-0.283,仍失败。不放宽门槛、不采用50步作为已通过方案。 ++ ++Kimodo v12全预测历史250更新最终幅度1.100、接段2.386、姿态0.439 rad、根位置0.441 m、jerk405.63。行走2327进度0.827/0.678,961为0.107/0.510,仍有一个失败;125时961为1.108/0.390,jerk890.09。中途和最终对同一案例不同种子不稳定,不能以均值改善晋升正式训练,也不机械延长250预算。所有诊断都不是新FID。 ++ ++读取既有validation历史诊断的实际因果文本:961在decision276/432是“turns left Hmm... walks forward”;2327在decision396/504是“interchanging an item between hands stop jog forward left diagonally”。Source.condition按已发布时钟取完整30秒参考转录上下文,未用GT turn或控制标签选新指令。文本中旧/新动作混合是可检验假设,尚不能断言根因或擅自丢弃旧指令。 ++ ++v13固定权重原文本/空文本对照已启动,目录`/root/autodl-fs/pangxinlei/models/hfd-recovery-text-v13-20261110/{hymotion,kimodo}`,入口scripts/recovery_text_v13/supervise_pilot.py。HY用原正式best、10步、BF16;Kimodo用v12最终完整权重、20步;同一6个validation案例、两种子、CFG、因果历史、原始验收和其余推理保持。两臂只在text.encode入口把已发布文本保留或替换为空字符串;空文本是编码空转录,并非假定Kimodo得到严格CFG无条件向量。没有未来文本、GT动作/标签选择,也不改变训练或评测门槛。原文本先跑以检验复现。该诊断测量文本敏感性,空文本即使分数变好也不能当任务成功或正式训练资格。 ++ ++零训练更新、无优化器更新、无新训练checkpoint,4小时监督上限;输出logs/diagnostic.jsonl的arm字段为original/empty,validation-traces/step-000000与step-000001对应两臂,不是训练更新。静态检查artifacts/recovery_v13_static_check.json,输入权重SHA/步数recovery_v13_inputs.json,源码SHA/进程recovery_v13_launch.json。监督器564572/564576及子进程已核实存活,尚无结果结论。 ++ ++每模型2 GiB预算实写/fsync验证通过,保持磁盘/inode/目录项/日志保护;本地约1.54 GiB可用、共享约34240空闲inode,写入探针通过,不将df当账号配额。清理4.856 GiB:HY v12诊断与Kimodo v12训练的重复索引、Kimodo v12不合格125快照,以及无后续用途的Kimodo v10最终权重。核对结束收据、进程已退出、无打开句柄、无配置输入引用,缓存与新副本哈希一致。保留当前完整断点、必要对照/正式权重及全部轨迹/配置/指标。清单artifacts/recovery_cleanup_hourly_20261109T21.json,完整分析recovery_hourly_20261109T21_analysis.json。 ++ ++下一轮先检验原文本组复现,再读取两臂完整逐例轨迹判断文本敏感性,结合行走进度/幅度/接段/jerk而非单一平均误差确定后续训练。未达到正式条件,不提前取消心跳;任一正式训练健康启动后立即取消的要求保持。 ++ ++ ++### 小时检查与引导强度诊断 v14(2026-11-09 22 UTC) ++ ++v13两模型文本对照正常完成,零更新,无错误/资源停止,预留释放。已读取全部两臂完整逐例轨迹并确认有限;original组与既有对应权重结果逐元素完全一致。原文本对照可复现,空文本效果不是入口精度差异导致。无正式训练,心跳416e9f54保持。 ++ ++HY原文本幅度0.291、姿态0.372 rad、根位置0.449 m、jerk335.19;空文本为0.269、0.387 rad、0.530 m、334.10。行走2327从0.483/0.680降至0.173/0.056;961仍近零。Kimodo原文本幅度1.100、姿态0.439 rad、根位置0.441 m、jerk405.63;空文本为1.176、0.439 rad、0.526 m、56.93。空文本行走2327仅0.067/0.035,961为-0.022/-0.030,原文本分别0.827/0.678和0.107/0.510。空文本低jerk来自缺少正确行走,不能作为改善或替代输入。两臂都未通过门槛;不能据此断言历史文本冲突是唯一根因,也不能宣称模型忽略文本。未测新FID。 ++ ++下一有界诊断v14只改变引导强度,目录`/root/autodl-fs/pangxinlei/models/hfd-recovery-guidance-v14-20261110/{hymotion,kimodo}`,入口scripts/recovery_guidance_v14/supervise_pilot.py。保持v13相同固定权重、完整因果文本、同一6个validation案例/两个种子、HY10步BF16与Kimodo20步、历史/坐标/执行长度及原门槛。HY比较guidance1→2,检验增强条件响应;Kimodo比较2→1,检验文本引导是否放大根运动抖动。假设未证实,不以分数改善预设正确;仍须检查逐例进度与幅度。无未来文本/GT选指令,Test1024未使用。 ++ ++两臂original/adjusted_guidance写入logs/diagnostic.jsonl,guidance字段记实际数值;validation-traces/step-000000和000001分别对应两臂,非训练步。零优化更新、不保存训练断点,4小时监督上限与资源受控停止保留。输入权重SHA与步数在artifacts/recovery_v14_inputs.json;源码SHA与进程recovery_v14_launch.json;静态检查recovery_v14_static_check.json。监督器565596/565599及子进程已核实存活,尚无本轮结果。 ++ ++各2 GiB预算重新实写/fsync并验证物理分配,本地约1.54 GiB空闲、共享约34170空闲inode,探针写入通过;不将df当账号配额。清理已完成v13的重复索引656.24 MiB,确认旧进程结束、无打开句柄、与v14副本SHA一致,保留全部轨迹/指标/源码/配置及正式与必要候选权重。清单artifacts/recovery_cleanup_hourly_20261109T22.json;完整分析recovery_hourly_20261109T22_analysis.json。下次先检验original组复现,再分析调整引导强度的逐例效果,为后续恢复训练选有证据的设置;没有通过全部门槛不正式开训。正式训练健康启动后立即取消小时任务的条件保持。 ++ ++ ++### 第十五轮因果文本适配恢复(2026-11-09 23 UTC) ++ ++v14两模型引导诊断正常完成,无错误/资源停止,预留释放;全部两臂完整轨迹已读取且有限,original组逐元素复现既有结果。HY guidance2幅度0.341、接段1.530、姿态0.379 rad、根位置0.575 m、jerk529.91;行走2327进度2.108/2.568过冲,961仍为-0.024/-0.008。Kimodo guidance1幅度1.095、接段2.412、姿态0.429 rad、根位置0.475 m、jerk143.44,但2327进度0.120/0.366、961为0.075/0.206。两者均未通过,低jerk不能代替任务运动。不采用这些推理设置,无正式训练。 ++ ++恢复实验v15目录`/root/autodl-fs/pangxinlei/models/hfd-recovery-pilot-v15-20261110/{hymotion,kimodo}`,入口scripts/recovery_training_v15/supervise_pilot.py。HY以v7、Kimodo以v12为对照,均从原正式best重新初始化,不从诊断输出续训;保持各自损失/学习率/预测历史概率/250更新/0、125、250两种子验证/原推理设置,仅加入训练和推理共享的因果文本适配。原baseline-validation.json逐字节保持,原误差门槛不因输入变换放宽。 ++ ++适配器causal_text.py只读取截至当前发布时刻的文本前缀和发布时间。新文本为旧文本词前缀的延续时,距上次新增词不足1.2秒则继续当前片段,否则从新增加的词开始;不变文本保持当前片段,空上下文清空,修订或截断无法可靠分辨时保守保留当前已发布全文。无动作、turn ID、控制标签、GT边界或未来词。是新的因果输入预处理实验,不能把结果标为“原完整转录不变”的协议,也不能据其直接晋升正式训练。 ++ ++单元验证前缀不变性、追加未来不影响过去、空上下文过期、修订/截断保守处理、非法时间拒绝及确定性恢复。使用禁止未来索引的序列测试真实Source.condition绑定,确认只访问已发布前缀且发布时间不变。真实输入核验先发现按需provider用SimpleNamespace构造Source、绕过__init__,已在启动前改为共享Source.condition接口;未启动无效训练。修正后961 decision276/432从“turns left Hmm... walks forward”变为“walks forward”,120为“Hmm...”;2327仍可能保留“stop jog forward left diagonally”,此规则不保证解决全部歧义,不预设改善。收据artifacts/recovery_v15_causal_text_checks.json、recovery_v15_actual_text_audit.json、recovery_v15_config_diff.json。只用train/validation,Test1024留出。 ++ ++HY2 GiB、Kimodo12 GiB预算重新实写/fsync并验证物理分配;原子checkpoint上限256 MiB/4 GiB、8小时墙钟、磁盘/inode/目录项/日志和受控停止保护保持。本地约1.54 GiB空闲、共享约34130空闲inode,探针写入通过,不把df当配额。两监督器567139/567143与子进程已核实启动,先初始验证和完整两更新保存/严格恢复;当前恢复检查未完成、效果未知。源码SHA和启动收据artifacts/recovery_v15_launch.json。 ++ ++清理已完成v14重复索引656.24 MiB,确认旧进程退出、无打开句柄和新副本SHA一致,正式/必要候选/完整恢复断点、全部轨迹和配置指标保留。清单artifacts/recovery_cleanup_hourly_20261109T23.json;完整分析recovery_hourly_20261109T23_analysis.json。下一检查点为两模型恢复收据与实际训练文本、0/125/250闭环逐例改善,并同时关注旧指令被过早丢弃的失败案例。无正式训练,心跳416e9f54继续;任一正式训练健康启动后立即删除心跳的条件保持。 ++ ++ ++### 小时检查与固定权重文本复核 v16(2026-11-10 00 UTC) ++ ++HY v15仍运行,00:05核验194/250步,进程567475正常推进,已记录loss/梯度有限、完整两更新恢复通过,无错误或资源停止。125步幅度0.258、接段2.068、姿态0.447 rad、根位置0.512 m,行走2327进度0.111/0.032、961为-0.052/-0.015,仍不合格;等待有界250步最终验证,不重复启动。 ++ ++Kimodo v15已完成250步、完整恢复通过、无错误,预留已释放。读取初始/125/250全部完整轨迹并确认有限。最终幅度1.193、接段2.259、姿态0.465 rad、根位置0.685 m、jerk501.95;2327进度0.679/0.537,961为-0.938/-1.307,明显反向行走。相比v12最终根位置0.441 m、961进度0.107/0.510退化。v15第0步961已为-1.171/-1.180(根位置均值0.739 m),125步进一步到-2.071/-2.205(根位置0.996 m);因此退化并非全由训练更新造成,截短上下文可能丢失方向信息,但尚不能确认唯一根因。不延长此配方、不进入正式训练;这些结果不是FID。 ++ ++Kimodo v16仅固定v15最终权重,比较adapted与full_context两臂,原验证6案例/两种子、20步CFG2、全部其他输入/推理/门槛保持。目录`/root/autodl-fs/pangxinlei/models/hfd-recovery-text-v16-20261110/kimodo`,入口scripts/recovery_text_v16/supervise_pilot.py。两臂依次写logs/diagnostic.jsonl,轨迹step-000000/000001是诊断臂而非训练步;0优化更新,进入优化器构造前返回。先复现v15最终适配文本,再恢复完整已发布因果文本,区分输入变换和训练后权重效应;不使用未来词或GT选指令,Test1024未使用。通过共享模块开关切换已绑定Source.condition,验证切换/还原与发布时间保持;避免运行入口与模型入口各自加载配置导致开关无效。源码/静态核验artifacts/recovery_v16_static_check.json,启动收据recovery_v16_launch.json。监督器568479/子进程568481已确认存活,诊断结果尚未完成。 ++ ++v16重新实写/fsync预留2GiB,4小时监督上限,磁盘/inode/日志/目录项保护保留。本地约1.539GiB可用,共享34052空闲inode,本地与共享1MiB写入/fsync探针通过;不把共享df总量当账号配额。清理已结束Kimodo v15重复索引和失败125快照,共1407.06MiB:确认训练/监督器已退出、无打开句柄、缓存与v16副本SHA一致、中间权重无配置输入引用。保留v15最终完整断点、全部配置/指标/轨迹及必要历史/正式权重,不触碰仍运行HY资源。清单artifacts/recovery_cleanup_hourly_20261110T00.json;完整轨迹/健康分析recovery_hourly_20261110T00_analysis.json。 ++ ++下一检查点为HY v15最终结果与Kimodo v16原适配组复现、完整文本恢复后的逐例方向/进度/幅度/接段/jerk。无正式训练,小时任务416e9f54继续;任一正式训练健康启动后立即删除心跳,不等待正式训练结束。 ++ ++ ++### 第十七轮检查与恢复实验(2026-11-10 01 UTC) ++ ++HY v15完成250更新,完整两更新恢复通过、无错误/资源停止,预留释放。初始/125/最终全部完整轨迹已读取且有限,记录loss/梯度有限。最终幅度0.247、接段2.003、姿态0.452 rad、根位置0.524 m、jerk279.65;行走2327进度0.073/-0.018、961为-0.038/0.005。相对原正式初始幅度0.291、姿态0.372、根位置0.449退化,低jerk不代表恢复。原门槛不变,不正式开训。 ++ ++Kimodo v16两臂正常完成,零训练更新、预留释放。已读取全部轨迹,适配文本组预测/参考逐元素复现v15最终。完整文本组幅度1.212、接段2.311、姿态0.454 rad、根位置0.523 m、jerk473.87;2327进度1.052/0.669,961为-0.227/0.253。相比适配组根位置0.685和961 -0.938/-1.307改善,但仍明显未通过原门槛,且不如v12根位置0.441/961 0.107/0.510。说明截短因果上下文确有影响,恢复全文不足以解决全部问题;不将该诊断称为FID。 ++ ++HY v17固定v15最终权重进行adapted/full_context两臂验证,目录`/root/autodl-fs/pangxinlei/models/hfd-recovery-text-v17-20261110/hymotion`,入口scripts/recovery_text_v17/supervise_pilot.py。同一6案例/两种子、10步BF16、CFG1与原门槛保持;沿用v16已验证共享模块开关切换Source.condition,原发布时间不变,无未来词或GT选指令。0更新、优化器构造前返回、4小时上限;先检验适配组复现再分析全文效果,不默认可恢复。 ++ ++Kimodo v17回到v12完整文本配方及原正式best初始化,唯一算法变量stage1_root_displacement权重0.5→1.0,检验加强相对最后观测帧的未来根位移监督能否减少v12的行走进度不足。不能据单一失败案例宣称必然有效;监督只用训练目标,推理不输入未来GT。其余源码与v12逐字节一致、训练/推理设置和原baseline-validation.json保持。目录`/root/autodl-fs/pangxinlei/models/hfd-recovery-pilot-v17-20261110/kimodo`,入口scripts/recovery_training_v17/supervise_pilot.py。250更新、0/125/250固定两种子验证、先两更新完整保存/严格恢复;不延长或续训v15失败配方,不使用Test1024。当前仅启动及数据准备,恢复检查尚未完成。 ++ ++两监督器569655/569696及真实子进程均已核实存活,未发现错误;源码/配置核验artifacts/recovery_v17_checks.json,启动及源码SHA见recovery_v17_hymotion_launch.json、recovery_v17_kimodo_launch.json。HY2GiB/Kimodo12GiB实写fsync预留通过,Kimodo保留4GiB原子checkpoint上限、8小时墙钟及磁盘/inode/目录项/日志保护。本地约1.539GiB、共享约34000空闲inode,本地和共享1MiB写入/fsync探针通过;共享df不是账号配额。 ++ ++清理已结束HY v15及Kimodo v16重复索引、HY v15已分析失败125快照,共786.78MiB;核验旧训练/监督器退出、无打开句柄、重复索引与新副本SHA一致及权重无配置引用。保留最终完整断点、正式权重、必要对照和所有轨迹/指标/配置。清单artifacts/recovery_cleanup_hourly_20261110T01.json;完整审计recovery_hourly_20261110T01_analysis.json。下一检查点为HY两臂结果与Kimodo恢复收据/训练推进及逐例效果。无正式训练,心跳416e9f54继续;任一正式训练健康启动后立即取消的条件保持。 ++ ++ ++### 第十八轮扩大验证(2026-11-10 02 UTC) ++ ++HY v17固定权重文本诊断正常完成,无错误,预留释放;全部轨迹有限,adapted组逐元素复现v15最终。full_context幅度0.272、接段2.003、姿态0.444 rad、根位置0.459 m、jerk290.00,行走2327为0.454/0.449、961为-0.018/-0.007;无坍缩但幅度仍不足,姿态超过原初始10%上限。截短组仍有1197/2026坍缩。恢复全文有改善却未解决恢复问题,不再重复截短配方;本轮先集中验证Kimodo的新增积极证据。 ++ ++Kimodo v17正常完成250更新,完整2步恢复通过,所有记录loss/梯度有限、全部完整轨迹已读取有限,无资源停止,预留释放。125步幅度1.182、接段2.229、姿态0.463 rad、根位置0.386 m;2327进度0.924/1.035、961为1.464/0.541,无坍缩,首次此轮通过原小验证全部门槛。250步幅度1.187、接段2.317、姿态0.456、根位置0.369,但961第二种子进度0.325,最终不合格;不机械延长,不丢掉中途候选。125步jerk787.14、250步562.89,相比原正式初始357.30偏高,必须排除抖动收益,不能以门槛通过直接正式开训。原正式基线姿态0.468/根位置0.442,各阶段对照和逐例证据见artifacts/recovery_hourly_20261110T02_analysis.json;这些不是FID。 ++ ++v18是冻结权重的扩大validation,零更新。目录`/root/autodl-fs/pangxinlei/models/hfd-recovery-broad-v18-20261110/{baseline,candidate125}`,入口scripts/recovery_broad_v18/supervise_pilot.py。baseline用原正式best,candidate125用v17 weights/step-000125.pt。每交互模式确定性选择4个6–20秒validation案例,共24个、包含原6个、固定种子12345/2026;仅依据模式/长度/可用性选样,不能依据预测分数选样,无Test1024。两组均保持完整因果文本、20步CFG2、同一初始状态及其他推理协议。选择器验证确定性、各模式4例、原6例保留;源码/检查recovery_v18_checks.json,源码SHA/启动recovery_v18_launch.json。两组监督器570911/570912、子进程570913/570914已核实存活,当前数据准备,结果未知。 ++ ++每组logs/diagnostic.jsonl及validation-traces/step-000000.npz为单个冻结权重验证,不是训练步。日志内门槛暂以原六例baseline计算,不能据此自动晋升;最终必须用同一扩展案例原正式baseline重新计算非退化门槛,检查逐例进度、幅度、接段/段内突变、速度误差、根jerk及失败轨迹。即使小验证125合格也不绕过更广验证和正式容量/恢复配置要求。 ++ ++两组各2GiB预算实写/fsync通过,4小时监督上限、日志轮转/目录项/资源停止保持,扩大轨迹单文件上限128MiB以覆盖48条轨迹且纳入预算。本地约1.539GiB、共享33893空闲inode,本地和共享1MiB写入/fsync通过,共享df不视为账号配额。清理已结束Kimodo v17重复索引328.12MiB,核验旧训练/监督器退出、无打开句柄、与新副本SHA一致。保留125候选、best、最终完整恢复断点及全部证据;HY诊断缓存暂留待后续。清单recovery_cleanup_hourly_20261110T02.json。无正式训练,心跳416e9f54保持,下一检查03 UTC;任一正式训练健康启动后立即取消,不等待训练结束。 ++ ++ ++### 扩大验证故障修复与候选重试(2026-11-10 03 UTC) ++ ++v18 baseline已完成24案例×2种子,零训练更新;读取48条完整轨迹均有限。幅度0.446、接段1.995、姿态0.453 rad、根位置0.632 m、jerk587.06;24个GT明显位移的案例×种子组合中17个进度未达0.5–2。扩大集不能直接与原六例均值比较,候选必须和这份同集基线比较。 ++ ++candidate125没有完成,因CUDA OOM失败,不能将缺失结果解释成模型效果差。根因是v18并行主进程使用不同pilot.physical_gpu,但文本编码器FrozenTextWorker通过config.gpu独立指定物理GPU,两份配置均沿用gpu=1,导致两个约14.75GiB文本工作进程挤占GPU1。已保留phase-error/supervisor-error/日志和原源码配置;两个旧主进程与监督器均已退出,GPU任务检查为空,预留已释放。baseline已经正常完成,保留不重复计算。 ++ ++仅重跑候选,目录`/root/autodl-fs/pangxinlei/models/hfd-recovery-broad-v18-20261110/candidate125-retry`,入口scripts/recovery_broad_v18r/supervise_pilot.py。新冻结源码在run_pilot启动前强制config.gpu==pilot.physical_gpu,明确主进程和文本工作进程都用GPU1;本轮只运行这一组。权重仍v17 step-000125.pt、24案例选择规则/两个种子/全文/20步CFG2及全部模型推理设置不变,未改训练或放宽门槛,无Test1024。原运行与源码不覆盖。监督器571978、子进程571982已存活并正常准备数据;当前尚未完成文本加载和全部验证,不虚称效果或恢复成功。 ++ ++新目录2GiB容量重新实写/fsync通过,4小时监督/磁盘/inode/日志轮转/目录项和轨迹128MiB上限保持;保留失败证据及原正式/候选/完整恢复断点。修复及预留收据artifacts/recovery_v18_retry.json,启动与源码SHA recovery_v18_retry_launch.json。清理两旧组重复索引656.24MiB,确认旧监督器/主进程退出、无打开句柄、与重试副本SHA一致;错误日志和完成基线全部轨迹保留。清单recovery_cleanup_hourly_20261110T03.json。本地约1.539GiB、共享33869空闲inode,本地和共享1MiB写入/fsync通过;不以共享df推断账号配额。 ++ ++HY无新训练/验证结果,保留v17失败诊断,不重复失败配方。完整本轮分析recovery_hourly_20261110T03_analysis.json。下一检查先核验候选重试完成和实际GPU分配,再用同集baseline重算非退化门槛及逐例进度/抖动;原六例门槛标记不能替代扩大集判断。无正式训练,小时心跳416e9f54继续,下一次04 UTC;任一正式训练健康启动后立即取消的条件不变。 ++ ++ ++### 第十九轮恢复实验(2026-11-10 04 UTC) ++ ++v18候选单独重试正常完成,零更新,无错误/资源停止、预留释放。原正式与候选24案例×2种子的完整轨迹均有限,case_keys及全部参考轨迹完全匹配,已按同集重新计算门槛。候选幅度1.021(基线0.446)、接段2.209(1.995)、姿态0.429 rad(0.453)、根位置0.598 m(0.632),无坍缩(基线1例)。明显位移24个案例×种子中,失败从17降至10;两个种子各5个失败,不能用平均改善掩盖失败。根jerk956.56 vs587.06,增约63%;1863两种子进度0.917/1.100虽合格,jerk7422/6808,相比基线4476/3907明显恶化。幅度/接段/姿态位置非退化门槛通过,逐例行走门槛不通过,且抖动增加,不批准正式训练。指标为SMPL-X诊断,不是FID。 ++ ++下一有界实验v19以v17为对照,原正式best初始化,不从125候选或最终断点延长;唯一训练变量decoded_root_acceleration监督权重0.001→0.01,保持根位移1.0、全部其他损失/数据/学习率/因果全文/推理参数不变。已有该损失约束解码实际根位置的二阶差分与相同C6目标,并非独立辅助速度通道;源码不改、加强权重检验抖动是否可控,不预设有效,须防止以静止换低jerk。仍250更新、0/125/250固定原六例两种子验证及原正式初始门槛,通过后还必须复查扩展集。无Test1024。HY无新证据,本轮不重复失败配方,保留诊断。 ++ ++目录`/root/autodl-fs/pangxinlei/models/hfd-recovery-pilot-v19-20261110/kimodo`,入口scripts/recovery_training_v19/supervise_pilot.py。增加启动前config.gpu==pilot.physical_gpu校验,两者均1,GPU空闲检查后单独启动,避免v18文本工作进程冲突。训练源码除运行器GPU断言/准备入口外与v17逐字节一致,原baseline-validation.json不变。收据artifacts/recovery_v19_checks.json,源码SHA/启动recovery_v19_launch.json。监督器572624及真实子进程已确认存活,无错误;当前初始验证/两步保存与严格恢复阶段,恢复检查尚未完成,不宣称训练改善。 ++ ++12GiB预算重新实写/fsync通过(实际预留约12.265GB),保留4GiB原子checkpoint上限、8小时墙钟、日志轮转、目录项/磁盘/inode保护。目录项32,本地约1.538GiB,共享33844空闲inode,本地/共享1MiB写入fsync探针通过;共享df不作账号配额。清理已完成候选重试的重复索引328.12MiB,核验旧监督器/子进程退出、无打开句柄、与v19新副本SHA一致;125候选、必要历史对照、正式权重、最终完整断点和全部轨迹/配置/指标保留。清单recovery_cleanup_hourly_20261110T04.json;配对验收与失败明细recovery_hourly_20261110T04_analysis.json。 ++ ++下一检查为v19恢复收据、训练推进、0/125/250逐例进度/幅度及jerk,不因平均数改善放宽门槛。尚无正式训练,心跳416e9f54继续,下次05 UTC;任一正式训练健康启动后立即取消的条件保持。 ++ ++ ++### 第二十轮候选扩大验证(2026-11-10 05 UTC) ++ ++Kimodo v19完成250更新,完整两步保存/optimizer/scheduler恢复通过,已记录loss和嵌套梯度均有限;全部0/125/250完整轨迹读取且有限,无错误/资源停止,预留释放。125步幅度1.121、接段2.187、姿态0.449 rad、根位置0.308 m、jerk795.43;行走2327进度1.044/0.910、961为1.329/0.904,无坍缩,原小验证全部门槛通过。对照v17的125步根位置0.386 m有所改善,但jerk787.14→795.43,并未证明加强加速度监督减少抖动。250步幅度0.974、接段2.436、姿态0.416、根位置0.407、jerk530.99,961进度0.338/0.349不合格,2327为0.817/0.635;无坍缩。继续保留中途候选,不机械延长最终失败配方,不以低jerk代替正确行走。上述诊断不是FID。 ++ ++v20冻结v19 weights/step-000125.pt,扩大到与v18完全相同的24个validation案例/两个固定种子,复用已完成原正式基线及v17候选对照,不重跑基线。目录`/root/autodl-fs/pangxinlei/models/hfd-recovery-broad-v20-20261110/candidate125`,入口scripts/recovery_broad_v20/supervise_pilot.py。除准备入口外所有源码与修复GPU绑定的v18r逐字节相同,选择规则、原六例包含、完整因果文本、20步CFG2及推理不变;仅输入权重换为v19的125步。0训练更新,无Test1024。输出logs/diagnostic.jsonl与validation-traces/step-000000.npz;日志原六例门槛仅为临时标记,必须基于同一扩大集baseline重新计算非退化、逐例进度和抖动后判断,不能自动晋升。 ++ ++监督器573160及真实子进程已核验存活、未发现错误,目前数据准备。启动/源码SHA artifacts/recovery_v20_launch.json,源码一致性recovery_v20_checks.json。2GiB容量重新实写/fsync通过,单独运行GPU1,config.gpu与pilot.physical_gpu一致断言保持;4小时墙钟、128MiB轨迹上限、目录项/日志轮转/磁盘inode保护保持。完整本轮分析recovery_hourly_20261110T05_analysis.json。 ++ ++清理已结束v19训练的重复索引328.12MiB,核验旧监督器/子进程退出、无打开句柄且与v20副本SHA一致。保留125候选、best、最终完整恢复断点、正式和必要历史对照及全部原始轨迹/指标/配置。清单recovery_cleanup_hourly_20261110T05.json。本地约1.538GiB可用、共享约33780空闲inode,本地和共享1MiB写入fsync通过;共享df不视为账号配额。HY暂无新结果,不重复失败实验。 ++ ++下一检查将v20候选与v18基线/上一候选逐例匹配,核实范围更广的行走改善及jerk而非只看均值。尚未正式训练,心跳416e9f54继续,下次06 UTC;任一正式训练健康启动后立即取消的条件保持。 ++ ++ ++### 第二十一轮接段与执行节奏诊断(2026-11-10 06 UTC) ++ ++v20候选扩大验证正常完成,零训练更新,无错误/资源停止,预留释放,进程退出。读取原正式基线、v17候选、v19候选三组全部完整轨迹,数值有限、case_keys及参考逐元素匹配。v19候选幅度0.991、接段2.213、姿态0.416 rad、根位置0.528 m、jerk906.78;上一候选对应1.021/2.209/0.429/0.598/956.56,原正式基线0.446/1.995/0.453/0.632/587.06。明显位移失败17→10→8(总24案例×种子),最新无坍缩、其他均值门槛通过,但8个行走失败与较基线偏高jerk仍不满足正式条件,不能仅因持续改善就晋升。不是正式FID。 ++ ++进一步读取逐帧三阶差分,将跨过每12帧重规划首个新预测帧的差分窗口计为接段附近:约24.72%样本占基线88.18%、v17候选86.32%、v19候选86.48%的根jerk平方能量。表明高频接段附近值得定位,不等于已证明唯一根因;仅继续加大损失权重缺乏充分依据。 ++ ++v21固定v19 step-000125权重,同一24validation案例、两种子,执行12帧与24帧两臂对照。目录`/root/autodl-fs/pangxinlei/models/hfd-recovery-cadence-v21-20261110/kimodo`,入口scripts/recovery_cadence_v21/supervise_pilot.py。原12帧组先跑,要求复现v20完整预测;24帧组只改重规划/执行节奏,每次使用截至该决策时刻已发布文本和模型历史,不输入未来GT、不改变权重/采样步数/CFG。通过共享runtime CONFIG切换,避免配置副本导致开关无效;逐帧覆盖/无重复遗漏静态验证通过。0优化更新、无Test1024。 ++ ++24帧组改变响应延迟、文本更新时机与随机抽样次数,是定位频繁重规划接缝的诊断,不能当作原协议训练改善或正式晋升依据。logs/diagnostic.jsonl arms为execute_12/execute_24,完整轨迹step-000000/000001为两诊断臂,非训练步。旧12帧边界指标与资格标记对24帧组不能直接使用,下一轮需按各自执行节奏重算接段指标,并同时查看行走/停步延迟,避免平滑却不响应任务。 ++ ++监督器573629及真实子进程已存活、无错误,当前数据准备。配置/覆盖检查artifacts/recovery_v21_checks.json,源码SHA/启动recovery_v21_launch.json。2GiB预算重新实写fsync验证,4小时墙钟、单独GPU1、主/文本GPU一致断言、日志轮转/目录项/磁盘inode与128MiB轨迹上限保持。本地约1.538GiB、共享约33750空闲inode,本地/共享1MiB写入fsync通过,共享df不视为账号配额。 ++ ++清理已结束v20重复索引328.12MiB,核对旧进程/监督器退出、无打开句柄及与新副本SHA一致;正式模型、完整恢复断点、必要候选及全部证据保留。清单recovery_cleanup_hourly_20261110T06.json;三组配对验收和接段能量分析recovery_hourly_20261110T06_analysis.json。HY暂无新结果,不重复失败训练。无正式训练,心跳416e9f54保持,下次07 UTC;正式训练健康启动后立即取消的规则不变。 ++ ++ ++### 第二十二轮根轨迹组件诊断(2026-11-10 07 UTC) ++ ++v21两臂正常完成,零训练更新,无错误/资源停止,预留释放、旧进程退出。全部完整轨迹已读取且有限,12帧组与v20预测/参考逐元素一致。24帧组jerk1431.62,比12帧906.78更差;根位置0.593 vs0.528 m,幅度0.847 vs0.991,行走失败均8/24,不能据旧12帧接段均值下降宣称改善,也不采用24帧协议。 ++ ++按真实执行节奏重算:以新计划首帧为接段、SO3逐关节角增量求均值后取接段/段内比,12帧均值2.329、24帧2.924(此定义与已有指标实现细节不同,单列诊断不覆盖旧指标);根jerk跨接段窗口占样本24.72%/12.31%,对应能量86.48%/88.05%。减少重规划次数并未消除接缝,可能每次接缝更大;响应延迟同时改变,尚未单独量化停止延迟,不能以此代替任务协议。不是FID。 ++ ++v22固定相同v19 step-000125权重,恢复原12帧执行及24validation案例/两种子,仅增加只读根轨迹组件采集。目录`/root/autodl-fs/pangxinlei/models/hfd-recovery-components-v22-20261110/kimodo`,入口scripts/recovery_components_v22/supervise_pilot.py。采集每次最终inverse的unnormalized原生channels0:8(含stage1根0:3、骨盆偏移5:8)与decoded root xyz,原生帧59:73;首行最后已知历史,随后12行执行,末行仅诊断的未执行预测。保存为同一validation-traces/step-000000.npz中的root_components_i,形状[decisions,14,11],不新增逐步文件。钩子返回原解码对象,不修改输入/输出、不新增随机采样,并在finally恢复inverse;完整轨迹必须先逐元素复现v20,之后组件证据才能用于判断根/骨盆抵消或突跳。组件是模型预测,未向条件加入未来GT,无Test1024。 ++ ++该实验零优化更新、原20步CFG2/完整因果文本/坐标及其他推理保持。通过语法、钩子还原、12帧保持检查,收据artifacts/recovery_v22_checks.json;源码SHA/启动recovery_v22_launch.json。监督器574134及真实子进程已核验存活、无错误,当前数据准备,尚无组件诊断结论。4小时监督上限、2GiB实写fsync预留、GPU1主/文本一致断言、目录项/磁盘inode/日志轮转与128MiB轨迹上限保留。 ++ ++清理已结束v21重复索引328.12MiB,核验旧进程/监督器退出、无打开句柄,与新副本SHA一致;全部诊断轨迹/配置/指标、正式/必要候选及完整恢复权重保留。清单recovery_cleanup_hourly_20261110T07.json,完整分析recovery_hourly_20261110T07_analysis.json。本地约1.537GiB、共享33711空闲inode,本地和共享1MiB写入fsync通过,共享df不作为账号配额。HY无新证据,不重复失败训练。 ++ ++下一检查先确认v22输出复现,再比较失败/正常案例的stage1根、骨盆偏移和解码根在接段前后变化,以证据决定训练或表征修复。无正式训练,心跳416e9f54继续,下次08 UTC;正式训练健康启动后立即取消规则保持。 ++ ++ ++### 用户要求双模型同步推进;HY v23启动(2026-11-10) ++ ++用户明确要求:两个模型必须一起推进,不能因为Kimodo出现改善就搁置HY;只有某模型确认改进、通过验证并正式训练成功启动后,才可聚焦另一个。该要求已写入artifacts/recovery_hourly_task.json的parallel_progress_requirement,后续每轮必须分别检查并推进两者。此前任一正式训练健康启动后立即取消小时任务的条件仍保留,不把pilot当正式训练。 ++ ++HY新一轮v23已在空闲GPU0启动,目录`/root/autodl-fs/pangxinlei/models/hfd-recovery-pilot-v23-20261110/hymotion`,入口scripts/recovery_training_v23/supervise_pilot.py。以完整因果文本的v7为对照,原正式best初始化、同一数据/种子/250更新/0、125、250验证/根运动监督与全部原门槛保持;唯一算法变化rotation_velocity_weight 0.05→0,检验额外旋转速度辅助目标是否限制动作幅度。只是待验证假设,不声称根因已确定;原flow损失和根导数目标保留,不靠降低验收标准。无Test1024,训练源码除准备入口与v7逐字节相同。检查recovery_v23_checks.json,启动与源码SHA recovery_v23_launch.json。 ++ ++HY监督器574500和真实子进程已确认存活,目前初始验证/两更新完整保存恢复阶段,尚不能宣称恢复检查通过或效果改善。Kimodo v22在GPU1的根组件诊断真实进程仍存活且无错误,继续推进,不中断源码/配置。两模型健康检查recovery_v23_health.json。 ++ ++HY2GiB容量重新实写/fsync预留通过,256MiB原子checkpoint上限、8小时墙钟、日志轮转/目录项/磁盘inode保护保持。本地与共享1MiB写入/fsync检查通过,本地约1.537GiB可用、共享约33660空闲inode。保留当前有效权重与证据,重复缓存继续按每轮已结束/无占用/哈希一致规则清理。下一小时检查须同时分析HY v23恢复与训练进度、Kimodo v22组件结果,再分别确定后续动作;不能再次默认只推进Kimodo。 ++ ++ ++### 双模型同步检查与Kimodo v24恢复(2026-11-10 08 UTC) ++ ++HY v23仍在GPU0正常训练,检查时151步,结束核验163/250步,真实进程存活、完整2步保存/optimizer/scheduler恢复通过、无错误/资源停止。已读取初始/125完整轨迹且有限,记录loss/梯度有限;125幅度0.270、接段2.156、姿态0.442 rad、根位置0.463 m,仍不合格。继续既定有界250步及最终验证,不重启、不覆盖正在运行源码配置。 ++ ++Kimodo v22组件诊断完成,全部完整预测/参考与v20逐元素完全相同,组件数值有限。高抖动1863两种子每次第一帧stage1根平均跳变0.230/0.207 m,骨盆偏移变化0.262/0.236 m,方向余弦约-0.987/-0.989;解码第一帧仅0.043/0.038 m,第二帧达0.186/0.160 m。810与277也出现类似抵消/后续突跳。这支持针对根轨迹第一、第二帧直接约束,尚不声称唯一根因。已查原生inverse:平滑根仅加到水平x/z,根高度来自骨盆位置;水平组合的最大残差2.38e-7,三维直接相加的残差不是解码bug。完整证据artifacts/recovery_hourly_20261110T08_analysis.json。 ++ ++Kimodo v24以v19为基线、原正式best初始化,唯一算法变量stage1_boundary_weight 3→10,加强已有stage1_root_velocity中前两次差分权重;其余根位移1.0、解码加速度0.01、训练数据/学习率/种子/完整因果文本/推理及原门槛不变。并非延长失败最终断点,仍250更新、0/125/250固定两种子验证、先2步完整保存严格恢复;不以静止或抑制幅度换低抖动,无Test1024。目录`/root/autodl-fs/pangxinlei/models/hfd-recovery-pilot-v24-20261110/kimodo`,入口scripts/recovery_training_v24/supervise_pilot.py。配置差异与基线核验recovery_v24_checks.json,源码SHA/启动recovery_v24_launch.json。 ++ ++Kimodo监督器575150及子进程已核验存活,GPU1启动前空闲,GPU0的HY持续推进。目前Kimodo初始验证/两步恢复阶段,恢复测试尚未完成、效果未知。12GiB预算重新实写fsync通过,保留4GiB原子checkpoint上限、8小时墙钟、主/文本GPU绑定校验、日志轮转、目录项/磁盘inode保护。HY当前预留不清理。本地约1.537GiB、共享33638空闲inode,本地/共享1MiB写入fsync通过;共享df不当账号配额。 ++ ++清理已完成Kimodo v22与HY v17诊断重复索引共656.24MiB,确认旧训练/监督器退出、无打开句柄,与新v24/v23保留副本SHA一致。保留正式权重、完整恢复断点、必要候选和全部原始证据。清单recovery_cleanup_hourly_20261110T08.json。下次同时分析HY最终结果和Kimodo恢复/推进及逐例结果;双模型必须持续推进的要求保持。无正式训练,心跳416e9f54继续,下次09 UTC;任一模型正式训练健康启动后立即取消规则不变。 ++ ++ ++### 第二十五轮双模型推进(2026-11-10 09 UTC) ++ ++HY v23与Kimodo v24均正常完成250更新、完整2步optimizer/scheduler恢复通过,无错误/资源停止、预留释放。已读取两模型初始/125/最终所有完整轨迹且有限,loss及嵌套梯度记录有限。HY最终幅度0.275、接段2.098、姿态0.451 rad、根位置0.474 m、jerk293.03;2327进度0.359/0.334,961为0.000/0.004,仍不合格。移除旋转速度辅助并未恢复幅度,不继续延长此配方。 ++ ++Kimodo v24的125步通过小验证门槛,jerk775.87,2327进度1.540/0.822、961为1.327/0.609;但250步961第二种子0.209失败(第一0.608),2327为0.793/0.553,最终幅度1.008、接段2.431、姿态0.421 rad、根位置0.390 m、jerk561.37。小验证通过不等于更广交互恢复,不直接正式训练,也不把这些诊断称为FID。 ++ ++HY v25目录`/root/autodl-fs/pangxinlei/models/hfd-recovery-pilot-v25-20261110/hymotion`,入口scripts/recovery_training_v25/supervise_pilot.py。在v23基础上唯一算法变量root_velocity_weight 0.1→0(旋转速度权重已为0),保留原flow损失、C6目标/导数一致构造、完整因果文本、原正式best初始化、数据/种子/学习率和原门槛,做flow-only对照检验辅助目标是否冲突,而非预设根因。250更新、0/125/250两种子验证、完整2步保存恢复检查;当前恢复检查尚未完成。GPU0监督器575697与真实子进程已存活,无错误。 ++ ++Kimodo v25目录`/root/autodl-fs/pangxinlei/models/hfd-recovery-broad-v25-20261110/candidate125`,入口scripts/recovery_broad_v25/supervise_pilot.py。冻结v24 step-000125.pt,重复同一24validation案例/两种子/12帧执行/20步CFG2及完整因果文本,复用v18原正式基线及v20候选对照。零优化更新,只有权重与输出目录变化;要求用匹配扩大集原基线重算门槛并检查抖动,不用日志六例临时标记晋升。GPU1监督器575698与真实子进程已存活,无错误。两任务都不使用Test1024;双模型同步推进要求保持。 ++ ++每组2GiB预算重新实写fsync通过;HY8小时/256MiB原子checkpoint上限,Kimodo4小时/128MiB轨迹上限,原磁盘/inode/目录项/日志轮转保护保持。主模型与文本GPU分配保持各自0/1,不共享Kimodo文本工作进程。配置差异artifacts/recovery_v25_checks.json,源码SHA/启动recovery_v25_launch.json。原始结果、恢复与实际写入检查recovery_hourly_20261110T09_analysis.json。 ++ ++清理两旧训练重复索引及HY已分析失败125快照,共786.78MiB;确认旧训练/监督器退出、无打开句柄、缓存SHA与保留副本一致、中间快照无配置输入引用。保留两模型最终完整断点、Kimodo125候选与best、正式/必要历史对照及全部轨迹/配置/指标。清单recovery_cleanup_hourly_20261110T09.json。本地约1.536GiB、共享约33560空闲inode,本地/共享1MiB写入fsync通过,不把共享df当配额。下次同时检查HY恢复/训练推进、Kimodo扩大验证结果。无正式训练,小时心跳416e9f54继续,下次10 UTC;任一正式训练健康启动后立即取消规则保持。 ++ ++ ++### 第二十六轮同步检查与历史诊断(2026-11-10 10 UTC) ++ ++HY v25持续GPU0训练,开始检查161步,结束169/250步;完整2步恢复通过,记录loss/梯度及初始/125完整轨迹有限,无错误/资源停止。125幅度0.272、接段2.299、姿态0.450 rad、根位置0.469 m,仍不合格。继续本轮有界预算,不重启,不碰正在使用的预留或断点。 ++ ++Kimodo v25扩大验证正常完成,无错误,预留释放。三组同24案例/两种子完整轨迹均有限、参考与case_keys匹配。v24的125步候选幅度0.966、接段2.234、姿态0.412 rad、根位置0.585 m、jerk971.45;明显位移失败11/24,相比v19候选8/24、根位置0.528、jerk906.78退化。加强stage1前两帧速度权重没有带来扩大集改善,不采用、不机械延长。不能根据小验证通过正式开训。 ++ ++Kimodo v26固定较好的v19 step-000125权重,比较predicted_history与expert_past两臂,目录`/root/autodl-fs/pangxinlei/models/hfd-recovery-history-v26-20261110/kimodo`,入口scripts/recovery_history_v26/supervise_pilot.py。保持原24validation案例/两种子、12帧执行、20步CFG2、因果全文与其余推理设置。第一臂必须逐元素复现v20;第二臂每次用截至当前decision帧的真实过去60帧替换生成历史,索引绝不超过decision、不读取未来GT条件。目标是区分反馈误差累积与单次条件预测不足,非训练配方或部署结果;真实过去在部署不可获得,重置也会制造拼接突变,不能把该组较低误差/jerk或资格标记当恢复成功。零更新、无Test1024,保留完整两臂轨迹。 ++ ++静态验证过去索引上界、共享runtime CONFIG切换和原组复现要求,收据artifacts/recovery_v26_checks.json;启动及源码SHA recovery_v26_launch.json。监督器576671及真实子进程已核验存活,无错误,GPU1启动前空闲;HY继续GPU0。2GiB实写fsync预留、4小时墙钟、主/文本GPU绑定校验、日志轮转/目录项/磁盘inode与128MiB轨迹上限保留。当前尚无诊断结论。 ++ ++清理已完成Kimodo扩大验证重复索引328.12MiB,核验旧训练/监督器退出、无打开句柄及与新副本SHA一致;正式/必要候选、完整断点和全部证据保留。清单recovery_cleanup_hourly_20261110T10.json;三组配对轨迹与HY健康分析recovery_hourly_20261110T10_analysis.json。本地约1.536GiB、共享33503空闲inode,本地/共享1MiB写入fsync通过,共享df不是账号配额。下次同时分析HY最终结果和Kimodo历史对照,不搁置任一模型。无正式训练,心跳416e9f54继续,下次11 UTC;正式训练健康启动后立即取消规则不变。 ++ ++### 第二十七轮双模型历史诊断(2026-11-10 11 UTC) ++ ++HY v25完成250步,完整恢复检查通过,loss/梯度记录和0/125/250全部轨迹有限;最终幅度0.283、接段2.224、姿态0.457 rad、根位置0.461 m,仍不合格。去掉旋转及根速度辅助损失没有恢复动作幅度,不机械延长。Kimodo v26两臂完成;生成历史组逐元素复现v20。真实过去组姿态0.176 rad、根位置0.093 m,但接段6.186、jerk5296.08,且每段重置使行走进度看似通过,不能作为恢复成功。按当前段真实起点计算,真实历史组段内根位移RMSE仍0.139 m;生成历史组按自己当前起点计算为0.078 m,二者锚点定义不同,不能据此单独下因果结论。完整证据见HFD_bench/artifacts/recovery_hourly_20261110T11_analysis.json;其中mean_root_jerk_rms是坐标分量RMS,和既有向量模jerk相差sqrt(3),不替换旧指标。 ++ ++两个模型同步启动v27,根目录/root/autodl-fs/pangxinlei/models/hfd-recovery-history-v27-20261110。HY固定v25最终权重,原6validation×2种子对照生成历史与截至decision的真实过去;首臂应复现v25最终轨迹。Kimodo固定v19 step125,在原24validation×2种子中分别只用真实历史根平移、只用真实历史旋转替换,分解条件误差来源。混合历史可能存在平移/姿态不一致,因此仅作诊断,不能视为可部署配方或正式资格。索引不超过decision,未来占位和原采样设置保持;零更新、无Test1024。入口scripts/recovery_history_v27_{hymotion,kimodo}/supervise_pilot.py,源码/启动及健康收据recovery_v27_launch_*.json、recovery_v27_health.json。下次检查必须核对HY原臂复现,并结合Kimodo两组的段内误差/接缝而非全局重置进度选择训练改动。 ++ ++GPU0/1分别确认真实子进程存活并占用显存,当前为加载及诊断阶段,不宣称结果已改善。每组2GiB预算实写fsync预留通过,4小时监督器、磁盘/inode/目录项/日志保护保留。本地约1.536GiB,磁盘本地/共享1MiB实写fsync通过,共享空闲inode超过33000;共享df不作为账号配额。确认旧进程已退出、无打开句柄及新副本SHA相同后,清理HYv25和Kimv26重复启动索引共656.24MiB;所有正式权重、完整断点和轨迹保留,清单recovery_cleanup_hourly_20261110T11.json。 ++ ++尚无正式训练,心跳416e9f54继续;下一次12 UTC。任一模型正式训练通过验收且真实健康启动后立即取消心跳,不关闭训练监督器。两模型继续同步推进。 ++ ++### 第二十八轮双模型恢复试验(2026-11-10 12 UTC) ++ ++v27两模型正常完成,无错误/资源停止,预留已释放;两臂所有完整轨迹有限。HY生成历史组逐元素复现v25最终结果,幅度0.283、接段2.224、姿态0.457 rad、根位置0.461 m、jerk334.62,2个幅度坍缩案例;4个明显位移组合均未严格达到0.5(其中一个0.499574,不能四舍五入算通过)。真实过去组幅度0.644、姿态0.146、根位置0.034,但接段12.068、jerk1878.39;重置产生的行走改善不计资格。Kimodo只替换历史平移:幅度1.235、接段2.397、姿态0.394、根位置0.100、jerk5448.41;只替换历史旋转:幅度1.217、接段6.417、姿态0.183、根位置0.383、jerk1256.41,行走仍失败7/24。真实历史及混合条件均非部署方法,不能用其较低误差正式开训。完整逐案例/轨迹核验在artifacts/recovery_hourly_20261110T12_analysis.json,不是FID。 ++ ++两边同步启动v28各250步有界恢复试验,根目录/root/autodl-fs/pangxinlei/models/hfd-recovery-pilot-v28-20261110,入口scripts/recovery_training_v28/supervise_pilot.py。HY以v25配方为对照,仅将生成历史训练比例0.25提高到1,检验真实历史和自主历史的分布差异;Kimodo以v19为对照,仅将监督标签过渡6帧延长至12帧,检验更缓的姿态/高度纠正能否减小接段冲击。后者不是直接改变XY轨迹或推理后处理,也不能预设有效。均从原正式best初始化,保留0/125/250同6案例×2种子及原始验收门槛;未达幅度/行走条件的平滑不算改善,若小验证通过仍须扩大验证。Test1024未使用。 ++ ++配置差异记录recovery_v28_config_diff.json;12帧过渡测试验证第12帧起与对齐标签完全一致、历史不被改写、数值有限,推理条件/部署实现不改。HY与Kimodo关键运行源码分别匹配v25/v19;启动先做2步optimizer/scheduler断点回读与恢复,随后监督器继续至250步,当前恢复检查尚未完成,不宣称已通过。两个真实子进程已核验,GPU分别0/1,不存在正式训练。HY2GiB/Kimodo12GiB实写fsync预留均通过;8小时墙钟、日志轮转、目录项/容量/inode检查、原子断点失败保留旧文件机制沿用。启动/源码SHA和健康证据recovery_v28_launch_*.json、recovery_v28_health.json。 ++ ++本轮清理已结束v27两模型重复启动索引656.24MiB,删除前确认旧进程/监督器退出、无打开句柄并与新副本SHA一致;保留全部权重、断点和原始轨迹,清单recovery_cleanup_hourly_20261110T12.json。本地约1.536GiB,本地/共享1MiB实写fsync通过,共享inode超过33000,共享df不作账号配额。小时任务416e9f54保持,下一次13 UTC重点核对两边恢复收据、loss/梯度和初始/中途轨迹;任一正式训练健康启动立即取消,不终止监督器。 ++ ++### 双模型检查及Kimodo扩大验证(2026-11-10 13 UTC) ++ ++HY v28仍正常训练,检查时153步、结束记录158/250步;完整2步optimizer/scheduler断点回读与恢复通过,无错误或资源停止。125步幅度0.480(未达0.5)、接段1.695、姿态0.418 rad(仍超过原初始1.1倍)、根位置0.430 m、jerk351.95。2327进度0.965/1.173改善,961仍-0.007/0.001,尚不合格;继续当前有界训练,不重启、不动其预留。full_predicted_history_fraction表示当前窗口已有60帧自身历史的比例,训练初始段不足60帧,因此不应把该日志低于1误读为配置没有启用100%预测链。 ++ ++Kimodo v28正常完成250步,恢复收据通过,预留已释放。125步通过原6案例两种子小验证:幅度0.859、接段2.352、姿态0.425 rad、根位置0.344 m、jerk652.17;2327进度0.770/0.947、961为1.388/0.929。250步退化:幅度0.734、根位置0.506 m、jerk277.79,但2327一个种子-0.006、961一个种子0.031,不能因jerk降低采用末步。保留125候选和完整最终断点,不续训退化配方。两模型已保存全部轨迹有限,case_keys/参考与对应配方对照一致,初始逐元素复现;训练loss/梯度记录有限,证据artifacts/recovery_hourly_20261110T13_analysis.json。这些诊断不是正式FID。 ++ ++立即启动Kimodo v29扩大验证:/root/autodl-fs/pangxinlei/models/hfd-recovery-broad-v29-20261110/candidate125,入口scripts/recovery_broad_v29/supervise_pilot.py,固定v28 step-000125.pt,沿用原24validation案例(各模式覆盖)、两种子、12帧执行与20步CFG2;零训练更新,无Test1024。必须与v18 expanded baseline、v20候选比较全部配对轨迹、行走/坍缩和jerk;日志中的原六案例门槛标记仍仅为临时诊断,不能据此正式开训。候选jerk仍高于原初始,扩大验证不能只看平均误差。GPU1启动前空闲,真实子进程580064和监督器580063已核验存活;HY继续GPU0。2GiB预算实写fsync预留通过,4小时墙钟/日志轮转/容量inode监督保留;源码SHA/启动/健康证据recovery_v29_launch_kimodo.json、recovery_v29_health.json。 ++ ++清理已结束Kimodo v28的重复启动索引328.12MiB,确认旧进程/监督器退出、无打开句柄、新副本SHA一致;所有权重和轨迹保留,HY运行资源未动。清单recovery_cleanup_hourly_20261110T13.json。本地约1.536GiB、共享inode超过37000,本地/共享实写fsync通过;共享df不作账号配额。尚无正式训练,小时任务416e9f54继续,下一次14 UTC同步分析HY完成结果与Kimodo扩大集;正式训练健康启动后立即取消规则保持。 ++ ++### 第三十轮双模型恢复试验(2026-11-10 14 UTC) ++ ++HY v28正常完成250步,2步恢复收据通过,全部训练loss/梯度及0/125/250轨迹有限,初始与原对照精确复现。最终幅度0.561、接段1.772、无低于0.1坍缩,姿态0.421 rad、根位置0.441 m、jerk359.06;相对原始姿态0.372仍超过10%恶化限,961两种子行走进度-0.057/-0.006,未达资格。100%生成历史使幅度改善,但未解决行走,不能将整体平均改善当正式开训依据。 ++ ++Kimodo v29扩大验证正常完成,无错误/资源停止、预留释放;原始扩大基线、v19step125和v28step125三组48条完整轨迹均有限、case_keys/GT完全匹配。v28候选幅度0.794、接段2.239、姿态0.408 rad、根位置0.568 m、jerk961.69,无坍缩、明显位移失败8/24;v19候选根位置0.528、jerk906.78、同样8/24失败,因此不采用延长标签过渡。完整配对指标及案例保存在artifacts/recovery_hourly_20261110T14_analysis.json。这些是原始SMPL-X诊断,不是正式FID。 ++ ++同步启动v30各250步,根目录/root/autodl-fs/pangxinlei/models/hfd-recovery-pilot-v30-20261110,入口scripts/recovery_training_v30/supervise_pilot.py。HY以v28为对照,仅将根速度辅助损失0改为0.1,保留全生成历史与旋转辅助0,检验在动作幅度恢复后明确根速度目标是否改善行走。Kimodo回到扩大集较好的v19配方(6帧标签过渡),仅统一把root/body学习率峰值和下限乘0.5,warmup/衰减时点及总250步保持;目的是检验后期退化与更新幅度的关系,不预设成功、不无限延长。两者仍从原正式best初始化,保持原始验收基线、0/125/250同案例两种子,Test1024未使用。 ++ ++配置差异和源码匹配检查recovery_v30_checks.json通过;没有新增监督表示/推理路径代码,恢复合同按独立源码/配置生成。HY2GiB/Kimodo12GiB预算实写fsync预留完成;GPU0/1启动前空闲,双模型真实子进程和监督器核验存活,目前执行两步恢复检查,尚未宣称该检查通过。启动SHA收据recovery_v30_launch_*.json,资源和PID见recovery_v30_health.json。保留8小时监督器、磁盘/inode/目录项检查、日志轮转和原子断点失败保留旧文件机制。 ++ ++清理已结束HYv28及Kimodo v29重复启动索引656.24MiB,确认旧进程/监督器退出、无打开句柄及新副本SHA一致,清单recovery_cleanup_hourly_20261110T14.json;正式权重、完整恢复断点、候选和全部轨迹保留。本地约1.535GiB,共享inode超过37000,本地/共享写入fsync通过;共享df不当账号配额。尚未正式开训,小时任务416e9f54继续,下次15 UTC检查两边恢复收据和训练/验证;任一正式训练健康启动立即取消规则保持。 ++ ++调度更新(2026-11-10):用户要求将检查频率改为每半小时。已通过paseo heartbeat update将现有任务416e9f54改为 `*/30 * * * *`(Etc/UTC),持久化记录核验通过;每小时整点及30分执行,下次2026-11-10 15:00 UTC。任务旧名称仍含“每小时”,实际频率以cron为准。继续同步推进两模型;任一正式训练健康启动立即取消的条件不变,不创建重复任务。 ++ ++### 半小时检查及Kimodo v31扩大验证(2026-11-10 15:00 UTC) ++ ++HY v30持续GPU0训练,检查157步,结束记录160/250;两步optimizer/scheduler断点回读及恢复通过,无错误/资源停止。125步幅度0.446、接段1.732、姿态0.427 rad、根位置0.414 m,仍不合格,继续当前有界预算。训练loss/梯度记录及已保存0/125完整轨迹有限,不重启不清理正在使用的资源。 ++ ++Kimodo v30完成250步,恢复检查通过,无错误/资源停止、预留释放。125步通过小验证:幅度0.730、接段2.124、姿态0.425 rad、根位置0.363 m、jerk628.92;2327进度0.751/0.761,961为1.129/0.507(后者接近下限,须关注扩大集稳定性)。250步幅度0.718、接段2.346、姿态0.414、根位置0.403、jerk384.19,但2327一个种子0.461、961一个种子0.365不合格。减半学习率没有消除后期退化,不续训失败末步。所有0/125/250完整轨迹有限、case_keys/参考与v19一致,初始精确复现;对照证据artifacts/recovery_hourly_20261110T1500_analysis.json,非正式FID。 ++ ++GPU1已启动v31:/root/autodl-fs/pangxinlei/models/hfd-recovery-broad-v31-20261110/candidate125,入口scripts/recovery_broad_v31/supervise_pilot.py,固定v30 step-000125.pt。同24validation案例、两个种子、原采样和12帧执行;零更新无Test1024。原6案例资格标记不能替代扩大集验收,须与v18原始扩大基线、v20及v29完整轨迹比较,检查行走失败、接段、幅度、jerk与姿态/位置误差。实际子进程582073及监督器582072已核验存活;2GiB实写fsync预留通过,4小时及容量/inode/日志保护沿用。启动和健康证据recovery_v31_launch_kimodo.json、recovery_v31_health.json。 ++ ++清理已结束Kimodo v30重复索引328.12MiB,确认旧进程/监督器退出、无打开句柄且新副本SHA一致;权重/完整断点/轨迹保留,HY预留和运行缓存未动。清单recovery_cleanup_hourly_20261110T1500.json。本地约1.535GiB,共享inode超过37000,本地/共享实写fsync检查通过;不将共享df作为账号配额。无正式训练,保持半小时cron */30,下一次15:30 UTC,优先分析任一已完成模型;正式训练健康启动后立即取消任务416e9f54的条件不变。 ++ ++### 半小时检查及v32采样覆盖对照(2026-11-10 15:30 UTC) ++ ++HY v30正常完成250步,两步恢复检查通过,无错误/资源停止;最终幅度0.512、接段1.707、姿态0.419 rad、根位置0.436 m、jerk302.88。961行走进度-0.048/-0.006,姿态仍超原初始10%限,未合格;增加根速度辅助降低jerk但没有恢复该行走案例。训练loss/梯度和0/125/250完整轨迹有限,参考和初始复现核验通过。 ++ ++Kimodo v31扩大验证正常完成,预留释放、无错误。v30半学习率候选幅度0.708、接段2.061、姿态0.407 rad、根位置0.575 m、jerk732.01,无坍缩但行走失败13/24;v19候选为8/24、根位置0.528、jerk906.78。降低学习率有平滑收益但伴随运动不足,不能正式晋级。与原扩大基线/v19全部48条轨迹有限且case_keys/GT匹配,原始及逐案例证据artifacts/recovery_hourly_20261110T1530_analysis.json;非FID。 ++ ++转查采样覆盖:代码确认预测历史链都从decision0开始,原链长50个12帧窗口最多覆盖开头20秒。v32两模型各只把链长50改25(10秒),在固定250更新内增加会话切换机会;HY对照v30,Kimodo回到扩大集较好的v19对照。它减少连续窗口相关性,但也缩短后半段覆盖,属于待证假设,不能预设修复;下一次需从训练locations检查独立会话数和实际时域覆盖并结合运动/行走/姿态。仍从原正式best初始化、全预测历史、初始单姿态启动,不引入GT历史替换。验证不缩短,保持原完整时长、原案例和两个种子;原门槛不变,无Test1024。 ++ ++新根目录/root/autodl-fs/pangxinlei/models/hfd-recovery-pilot-v32-20261110,入口scripts/recovery_training_v32/supervise_pilot.py。配置差异及生效字段检查recovery_v32_checks.json;HY的curriculum.chain_segments和pilot.chain_segments均25,Kimodo读取pilot.chain_segments。未改监督表示/推理代码。HY2GiB、Kimodo12GiB实写fsync预留通过,启动前GPU0/1空闲,真实进程/监督器已核验存活,当前执行两步恢复检查,尚不宣称完成。8小时监督/有限日志/目录项与磁盘inode检查/原子断点保护保留,启动及健康收据recovery_v32_launch_*.json、recovery_v32_health.json。 ++ ++清理已结束HYv30及Kimodo v31重复索引656.24MiB,旧进程/监督器退出、无打开句柄及新副本SHA一致后删除,清单recovery_cleanup_hourly_20261110T1530.json。所有权重、完整恢复断点及轨迹保留。本地约1.534GiB、共享inode超过36000,本地/共享实写fsync通过,共享df不作账号配额。尚无正式训练,半小时任务416e9f54继续,下一次16:00 UTC;任一正式训练健康启动立即取消规则不变。 ++ ++### v32运行中半小时检查(服务器时间2026-11-10 16:00 UTC) ++ ++本次核查以服务器date和运行收据为时间依据;对话环境日期标为11月11日,但实际服务器为11月10日16:00 UTC。最新仍是v32,没有发现正式训练启动。两模型真实进程和监督器存活,HY结束检查79/250步,Kimodo 202/250步;两步optimizer/scheduler恢复收据均通过,未出现错误或资源停止,训练loss/梯度记录和所有已保存完整轨迹有限,初始复现及参考配对通过。不重启、不抢占GPU。 ++ ++HY尚只有初始验证;Kimodo125幅度1.035、接段2.372、姿态0.455 rad、根位置0.454 m,未合格:2327种子12345进度0.430、961种子2026进度0.364低于0.5,其余两个1.013/0.956。等待预算内中途/最终轨迹,不仅看loss或幅度。链长改动实际增加了覆盖:相同79步HY训练独立会话8→16;相同200步Kimodo为18→33,所有比较更新的locations均有记录。最远起点588→288帧,验证了减少后段覆盖的代价;这些变化不等于质量改善,仍按原完整验证门槛判定。 ++ ++证据artifacts/recovery_hourly_20261110T1600_analysis.json。本地约1.534GiB、共享空闲inode36968,本地/共享1MiB实写fsync通过,当前每运行目录项低于500。两模型正在使用的预留、缓存、断点保持;本轮无新增已核验可删除项,清理检查写recovery_cleanup_hourly_20261110T1600.json(0字节删除)。不重复创建实验或扩大预算。文档和任务状态已更新,半小时调度下一次16:30 UTC,任一正式训练健康启动立即取消规则保持。 ++ ++### 半小时检查及Kimodo模块诊断(服务器2026-11-10 16:30 UTC) ++ ++HY v32继续GPU0,检查158步,结束记录177/250。恢复检查通过,无错误/资源停止,记录loss/梯度和已保存初始/125完整轨迹有限。125幅度0.419、接段1.962、姿态0.451 rad、根位置0.419 m,仍不合格,继续既定有界预算。 ++ ++Kimodo v32正常完成250步,恢复检查通过,预留释放,无错误。125和250均未合格;250幅度0.865、接段2.551、姿态0.426 rad、根位置0.477 m、jerk325.49;2327进度0.312/0.537、961为0.186/0.438,4组合3失败。增加会话覆盖未解决行走,不能因jerk降低晋级。完整0/125/250轨迹有限、参考匹配及初始复现通过,证据artifacts/recovery_hourly_20261110T1630_analysis.json。保持原正式基线与门槛,无Test1024,不是FID。 ++ ++下一步停止连续标量调参,启动冻结v33模块交换诊断:/root/autodl-fs/pangxinlei/models/hfd-recovery-modules-v33-20261110/kimodo,入口scripts/recovery_modules_v33/supervise_pilot.py。以v19 step125为候选,三组分别原候选、只换回原正式best根模块、只换回原正式best身体模块;原正式best不是官方原生权重。各组先严格重新加载完整候选,再覆盖一个模块,24validation案例两种子同协议,零训练更新、无未来GT条件。权重键/形状匹配通过,root/body各204键,检查recovery_v33_checks.json。第一组必须逐元素复现v20;模块交互可能导致分布不匹配,不能将混合权重改善直接当因果损坏证明或正式资格。结果用于定位下一训练改动。 ++ ++GPU1启动前空闲,新子进程583485和监督器583482存活,HY原进程持续。2GiB实写fsync预留、4小时墙钟、日志轮转及容量/inode检查保留;启动/源码SHA和健康收据recovery_v33_launch_kimodo.json、recovery_v33_health.json。清理已结束Kimodo v32重复索引328.12MiB,旧进程退出、无打开句柄、新副本SHA一致后删除,清单recovery_cleanup_hourly_20261110T1630.json;全部模型断点和轨迹保留,HY在用资源不动。本地约1.534GiB、共享inode36935,本地/共享实写fsync通过,目录项低于500。无正式训练,半小时任务继续,下次17:00 UTC;正式训练健康启动即取消条件不变。 ++ ++### 双模型运行中检查(服务器2026-11-10 17:00 UTC) ++ ++HY v32检查时249步,结束记录249/250;真实训练进程和监督器存活,尚无completed收据及最终验证结论,不将步数达到预算当成已完成。两步恢复检查通过,loss/梯度记录及已保存轨迹有限,初始/参考复现通过。保留当前预算,等待后台最终验证与保存,不重复启动。 ++ ++Kimodo v33仍运行,原候选组已完成:全部48条轨迹逐元素复现v20,幅度0.991、接段2.213、姿态0.416 rad、根位置0.528 m、jerk906.78,说明对照基线一致。模块交换组尚无完整结果,不预判根或身体模块的作用。GPU1真实主进程583485及文本进程运行,GPU0为HY;无错误/监督器停止。证据artifacts/recovery_hourly_20261110T1700_analysis.json。诊断不当正式FID或正式训练资格。 ++ ++本地约1.534GiB、共享inode36931,本地/共享1MiB实写fsync通过,两个运行目录项均低于500。当前预留/缓存/断点正在使用,本轮未发现新增已核验可清理项,0字节删除,清理检查recovery_cleanup_hourly_20261110T1700.json。两份文档与任务状态更新,未正式开训,半小时任务保持,下一次17:30 UTC;任一正式训练健康启动立即取消任务416e9f54。 ++ ++### 双模型采样诊断v34(服务器2026-11-10 17:30 UTC) ++ ++HY v32完成250步,无错误/资源停止,恢复检查通过;全部训练记录及初始/中途/最终轨迹有限。最终幅度0.476、接段2.025、姿态0.454 rad、根位置0.443 m、jerk318.87,2327进度1.015/1.698,961为-0.055/-0.005。缩短链后仍不合格,不延长该配方。Kimodo v33三臂完成,无错误,预留释放;48案例种子轨迹有限、参考配对一致,候选逐元素复现v20。候选行走失败8/24、根位置0.528、jerk906.78;原根模块替换失败10/24、根位置0.550、jerk864.14;原身体模块替换幅度0.453、失败17/24、根位置0.633、jerk590.09,接近原正式模型。说明恢复后的身体模块对幅度/行走改善很重要,简单回滚根或身体都不解决问题;模块交互影响仍在,不能断言单一模块损坏。证据artifacts/recovery_hourly_20261110T1730_analysis.json,非FID。 ++ ++同步启动v34固定权重采样步数诊断,目录/root/autodl-fs/pangxinlei/models/hfd-recovery-sampling-v34-20261110/{hymotion,kimodo},入口scripts/recovery_sampling_v34_{hymotion,kimodo}/supervise_pilot.py。HY固定较好幅度的v28最终权重,6validation×2种子比较10与50步;Kimodo固定v19 step125,24validation×2种子比较20与50步。先要求原采样组逐元素复现HYv28最终/Kimv20,再分析50步是否改善运动、行走和jerk及延迟代价。之前50步HY对照针对原正式权重,本轮针对恢复候选,非重复同一配方。未更改初始状态/执行窗口/因果文本/GT条件,不训练更新、不读取Test1024,验收基线不重置。更高步数即使改善也须记录计算成本,不能混同原等预算评测。 ++ ++两组各2GiB实写fsync预留完成,GPU0/1启动前空闲,真实子进程/监督器核验存活,启动/源码SHA及健康见recovery_v34_launch_*.json、recovery_v34_health.json。4小时墙钟、磁盘/inode/目录项限制、有限日志/缓存保留。清理已结束HYv32和Kimv33重复索引656.24MiB,删除前确认旧进程退出、无句柄占用、新副本SHA一致,清单recovery_cleanup_hourly_20261110T1730.json;所有权重、完整恢复断点及轨迹保留。本地约1.533GiB、共享inode超过36800,本地/共享实写fsync通过,共享df不是账号配额。无正式训练,半小时调度继续,下次18:00 UTC;任一正式训练健康启动立即取消任务416e9f54。 ++ ++### HY文本条件诊断及Kimodo运行检查(服务器2026-11-10 18:00 UTC) ++ ++HY v34两组正常完成,无错误/资源停止、预留释放。10步组逐元素复现v28最终;50步幅度0.887、接段1.671、姿态0.454 rad、根位置0.687 m、jerk368.75,较10步根位置0.441明显退化;2327进度2.055/3.613过冲,961为-0.283/-0.573反向,4组全不合格。采样更精细不等于恢复,不采用50步。两组完整轨迹有限、case_keys/GT相同。 ++ ++Kimodo v34仍在运行50步组;已完成20步组全部48条轨迹逐元素复现v20。尚无50步完整结果,不推断改善。真实进程584260及监督器584259存活,无错误/资源停止。完整核验和可写性证据artifacts/recovery_hourly_20261110T1800_analysis.json。 ++ ++HY继续推进冻结v35:/root/autodl-fs/pangxinlei/models/hfd-recovery-text-v35-20261110/hymotion,入口scripts/recovery_text_v35/supervise_pilot.py,固定v28最终权重与原10步采样,比较full_text和empty_text两臂、原6案例两种子。检查已恢复幅度候选的持续行走失败是否依赖文本条件;此前空文本实验使用别的权重,本轮不冒充相同对照。空文本经原编码器编码,不是零向量;只改变推理文本,轨迹/未来GT不输入条件。空文本偏离任务条件,仅用于机制诊断,不能作为正式评测或放宽资格。要求full_text精确复现v28最终。零权重更新,无Test1024。 ++ ++GPU0启动前空闲,HY真实子进程584643和监督器584642存活,GPU1继续Kimodo。2GiB实写fsync预留通过,4小时墙钟、日志轮转、容量/inode/目录项保护保留;启动与健康收据recovery_v35_launch_hymotion.json、recovery_v35_health.json。清理已结束HYv34重复启动索引328.12MiB,核验旧进程退出、无打开句柄、新副本SHA相同后删除;Kimodo在用预留不动,所有权重和轨迹保留,清单recovery_cleanup_hourly_20261110T1800.json。本地约1.533GiB、共享inode36828,本地/共享1MiB实写fsync通过。尚无正式训练,半小时任务继续,下次18:30 UTC,正式训练健康启动即取消规则保持。 ++ ++### HY v36执行窗口权重试验(服务器2026-11-10 18:30 UTC) ++ ++HY v35正常完成,预留释放,无错误。全文组完整轨迹逐元素复现v28最终;空文本组幅度0.544、接段1.874、姿态0.413 rad、根位置0.514 m、jerk339.82,2327进度0.082/0.011,961为-0.051/-0.010。去掉文本未恢复961且破坏2327行走,不能采用;仅为机制诊断。两臂轨迹有限、案例/参考相同,证据artifacts/recovery_hourly_20261110T1830_analysis.json。Kimodo v34仍正常运行50步组,原20步组精确复现已核验,目前没有50步完整结果,不重启或抢占GPU1。 ++ ++HY恢复训练v36,/root/autodl-fs/pangxinlei/models/hfd-recovery-pilot-v36-20261110/hymotion,入口scripts/recovery_training_v36/supervise_pilot.py。对照v28配方,仅把training.execution_weight从3改6,保持全生成历史、50窗口链、原正式best初始化及全部其他训练设置。代码核实该权重作用于未来60帧中实际执行的前12帧,并以加权总和归一化,因此并非简单放大总体loss。目的是检验近端执行目标的相对权重,不预设解决行走。原10步推理/正常因果全文/固定validation两种子/原验收基线不改,无Test1024,各250步,0/125/250比较。配置差异、入口源码一致核验recovery_v36_checks.json。 ++ ++GPU0启动前空闲,真实子进程585015及监督器585014存活,当前执行2步恢复检查,尚不宣称检查已完成;2GiB实写fsync预留通过,8小时墙钟、日志轮转/目录项/磁盘inode/原子断点保护保留。启动SHA和健康收据recovery_v36_launch_hymotion.json、recovery_v36_health.json。Kimodo主进程584260持续,监督器无错误或停止。 ++ ++清理已结束HYv35重复索引328.12MiB,确认旧进程退出、无句柄占用及新副本SHA相同;清单recovery_cleanup_hourly_20261110T1830.json,全部权重/断点/轨迹保留,Kimodo在用资源不动。本地约1.533GiB,共享inode36795,本地/共享实写fsync通过,目录项低于500。未正式开训,半小时任务保持,下次19:00 UTC;任一正式训练健康启动即取消规则不变。 ++ ++ ++### 半小时检查与Kimodo引导强度诊断(服务器2026-11-10 19:00 UTC) ++ ++HY v36仍在GPU0,真实进程585133、监督器585014,记录124/250步,正等待中途验证结果;控制台有数据读取进展,容量监督收据更新,无错误/资源停止。两步optimizer/scheduler恢复和读回通过,全部训练记录有限、已保存完整轨迹有限且初始/参考复现通过,不重复启动。当前尚不能判断本轮改进。 ++ ++Kimodo v34两组已完成且预留释放。全部48条/组完整轨迹有限、case_keys及参考一致,20步组精确复现v20。50步相比20步:幅度0.991→1.293、接段2.213→2.063、姿态0.416→0.447rad、根位置0.528→0.563m、根jerk906.78→1103.45,行走失败仍8/24;不能凭接段变小和幅度增加采用更高采样步数。证据artifacts/recovery_hourly_20261110T1900_analysis.json,非正式FID,无Test1024。 ++ ++Kimodo继续冻结v37引导强度单变量诊断:/root/autodl-fs/pangxinlei/models/hfd-recovery-guidance-v37-20261110/kimodo,入口scripts/recovery_guidance_v37/supervise_pilot.py。固定v19step125、20步采样、原24validation两种子、12帧执行及正常因果全文,只比较guidance2与1,排查引导放大是否贡献抖动及行走失败。guidance2必须精确复现v20;不预设低引导能恢复,不改变验收门槛。参数通过closed_loop的config.guidance实际传入采样器;零训练更新,不是正式训练。 ++ ++GPU1启动前空闲,真实子进程585526与监督器585525存活、初始化日志推进。2GiB实写fsync预留通过,4小时墙钟/容量inode/目录项/日志轮转保护保留,源码SHA与健康证据recovery_v37_launch_kimodo.json、recovery_v37_health.json。清理已结束Kimodo v34重复索引328.12MiB,旧进程退出、无打开句柄及新副本SHA一致后删除;所有权重和轨迹保留,HY在用资源未动,清单recovery_cleanup_hourly_20261110T1900.json。本地约1.532GiB,共享空闲inode36761,本地/共享1MiB实写fsync通过;共享df不作账号配额。半小时任务416e9f54继续,下次19:30 UTC;尚无正式训练,正式训练健康启动即取消规则保持。 ++ ++ ++### 双模型半小时运行检查(服务器2026-11-10 19:30 UTC) ++ ++HY v36实际推进到224/250,进程585133/监督器585014存活,两步恢复读回通过,全部loss/梯度记录及初始/125完整轨迹有限,参考/案例匹配、初始精确复现。125步幅度0.533、接段1.733、姿态0.420rad、根位置0.438m、jerk364.83;2327进度1.150/1.325,961为-0.030/-0.004。幅度达线但行走仍失败,姿态较原初始0.3723增约12.8%,超过10%门槛,尚不能正式开训。继续既定250步预算等待最终结果,不以平均数或loss作恢复结论。 ++ ++Kimodo v37运行中,真实进程585526/监督器585525存活,GPU1约17GiB。guidance2对照48条完整轨迹精确复现v20;guidance1组尚无完整结果,不预判改善或重复启动。HY继续GPU0。无错误/资源停止,证据artifacts/recovery_hourly_20261110T1930_analysis.json。当前两模型均为pilot或冻结诊断,没有正式训练。 ++ ++本地约1.532GiB,共享空闲inode36748,本地/共享1MiB实写fsync通过;共享df不是账号配额。两运行目录项低于500,现有预留与缓存正在使用,本轮无新增核验可删资源,0字节清理,记录recovery_cleanup_hourly_20261110T1930.json。全部断点和完整轨迹保留。半小时任务416e9f54保持,下次20:00 UTC;任一正式训练健康启动后立即取消规则不变。 ++ ++ ++### 双模型v38诊断(服务器2026-11-10 20:00 UTC) ++ ++HY v36完成250步,恢复读回通过,无错误,预留释放。全部训练记录及0/125/250完整轨迹有限,初始精确复现,参考/案例一致。最终幅度0.609、接段1.780、姿态0.41234rad(较原初始0.37227约+10.77%)、根位置0.452m、jerk355.50;2327进度1.489/1.855,961为-0.074/-0.071。仍不合格,不延长失败配方。Kimodo v37完成两臂,48条/组轨迹有限且参考一致,原引导2精确复现v20;引导1幅度0.980、接段2.258、姿态0.409、根位置0.614、jerk529.29(原906.78),行走失败8/24→16/24,不能以更平滑替代任务能力。证据artifacts/recovery_hourly_20261110T2000_analysis.json,无Test1024,非FID。 ++ ++两模型同步进入冻结v38,目录/root/autodl-fs/pangxinlei/models/hfd-recovery-guidance-v38-20261110/{hymotion,kimodo},入口scripts/recovery_guidance_v38_模型名/supervise_pilot.py。HY固定v36最终权重及10步,原6validation两种子比较guidance1与2;此前更强引导针对其他权重,本轮检查已恢复幅度候选。Kimodo固定v19step125、20步及24validation两种子,对照constant2与前10次去噪引导2/后10次引导1,检验强引导行走收益与抖动是否可按去噪阶段分开;不是根/身体模块分离。仅改变引导,无未来GT输入,原验收门槛不变,零训练更新。原臂必须精确复现HYv36最终/Kimv20。Kimodo本地采样副本撤销该切换后逐字节还原原实现,反向去噪迭代顺序验证通过,检查recovery_v38_checks.json;不修改上游或历史运行源码。 ++ ++两GPU启动前空闲,各2GiB实写fsync预留完成;HY真实进程586140/监督器586137、Kimodo586144/586142存活,初始化中,无错误。启动SHA及健康证据recovery_v38_launch_*.json、recovery_v38_health.json。4小时墙钟、容量/inode/目录项/日志保护保留。清理已结束HYv36和Kimv37重复启动索引656.24MiB,旧进程退出、无打开句柄及新副本SHA一致后删除;所有权重、恢复断点及轨迹保留,清单recovery_cleanup_hourly_20261110T2000.json。本地约1.532GiB,共享inode36697,本地/共享实写fsync通过,共享df不作账号配额。没有正式训练,半小时调度保持,下次20:30 UTC;任一正式训练健康启动即取消任务416e9f54。 ++ ++ ++### v38分析、故障修复及双模型v39(服务器2026-11-10 20:30 UTC) ++ ++HY v38完成、预留释放。两臂完整轨迹有限且参考/案例相同,引导1逐元素复现v36最终。引导2幅度0.645、接段1.893、姿态0.423rad、根位置0.729m、jerk560.34,2327进度2.581/1.443、961为0.079/0.011,3/4行走组合失败;原引导1根位置0.452、jerk355.50,仍不采用更强引导。证据artifacts/recovery_hourly_20261110T2030_analysis.json,非FID。 ++ ++Kimodo v38失败退出,原因是新增独立recovery_generate.py保留包内相对导入,make_window调用from .kimodo_causal_rep报ImportError。此前AST检查未覆盖实际导入路径,是本轮入口错误,未产生完整诊断结果,不归因于模型。监督器已释放预留,失败日志保留。v39独立复制源码只把两处包内相对导入改为hfd_bridge绝对导入;模块及依赖实际导入检查通过、AST确认无相对导入,recovery_v39_checks.json。保持v38的冻结权重、constant2/late1两臂和全部验证协议;原臂轨迹复现仍待完成,不把导入通过当效果验证。 ++ ++HY同步v39有界250步,从原正式best初始化,以v36全部配置为对照,仅根速度辅助权重0→0.1,保留execution_weight6、全生成历史50窗口链及10步推理。依据是v36改善幅度,而v30在execution_weight3下的小速度辅助曾减少jerk;本轮检验该作用能否在v36保留,不能预设改善行走,不机械延长或改门槛。固定原validation两种子,无Test1024。 ++ ++目录/root/autodl-fs/pangxinlei/models/hfd-recovery-v39-20261110/{hymotion,kimodo},入口scripts/recovery_v39_模型名/supervise_pilot.py。HY真实进程586697/监督器586694处于两步恢复检查初始化;Kimodo586701/586699处于诊断初始化,无新错误,未宣称恢复/诊断已通过。两GPU启动前空闲,各2GiB实写fsync预留通过;HY8小时/Kim4小时墙钟、日志/缓存/容量inode/断点保护保留,启动与健康见recovery_v39_launch_*.json、recovery_v39_health.json。 ++ ++清理已结束HYv38和已失败Kimv38重复索引656.24MiB,确认旧进程退出、无打开句柄、新副本SHA相同后删除;全部权重、恢复断点、轨迹及错误证据保留,清单recovery_cleanup_hourly_20261110T2030.json。本地约1.531GiB、共享inode36632,本地/共享实写fsync通过,共享df不作账号配额,目录项低于500。尚无正式训练,半小时任务继续,下次21:00 UTC;任一正式训练健康启动即取消规则保持。 ++ ++ ++### v39半小时运行检查(服务器2026-11-10 21:00 UTC) ++ ++HY v39推进108/250,真实训练进程586859/监督器586694存活;两步optimizer/scheduler恢复及断点读回通过,已记录loss/梯度全部有限,初始完整轨迹精确复现及参考配对通过。尚未到125步中途验证,不能判断新辅助项效果,不重复启动或扩大预算。 ++ ++Kimodo v39修复重试正常运行,进程586701/监督器586699存活,GPU1约17GiB。constant2对照完成,全部48条完整轨迹逐元素精确复现v20,有限性及案例/参考匹配通过,确认本次导入修复与采样副本没有改变对照输出。late1诊断仍运行,无完整结果,不预判改善;没有再次出现导入错误或资源停止。HY继续GPU0。证据artifacts/recovery_hourly_20261110T2100_analysis.json。 ++ ++本地约1.531GiB,共享空闲inode36605,本地/共享1MiB实写fsync通过,共享df不作账号配额。两个运行目录项低于500,预留/缓存/恢复断点正在使用,本轮没有新增已核验可清理项,0字节删除,清单recovery_cleanup_hourly_20261110T2100.json。尚无正式训练,保留半小时任务416e9f54,下一检查21:30 UTC;任一正式训练健康启动即取消规则不变。 ++ ++ ++### Kimodo v40与HY运行检查(服务器2026-11-10 21:30 UTC) ++ ++HY v39检查209步、结束更新210/250,真实进程586859及监督器586694存活。恢复检查通过、全部loss/梯度及初始/125完整轨迹有限,参考配对/初始复现通过。125幅度0.499、接段1.747、姿态0.423rad、根位置0.417m、jerk302.33;2327进度1.017/1.255,961为-0.047/-0.019,仍未达门槛,继续原250步预算。预留实写期间出现一次114.8秒训练步耗时,进程仍推进,无资源停止,不将此解释为模型性能变化。 ++ ++Kimodo v39两组完成,原臂精确复现v20、完整48条/组轨迹有限且参考一致。late1幅度0.978、接段2.351、姿态0.417、根位置0.542m、jerk892.22(对照906.78,仅降约1.6%),行走失败仍8/24,根位置误差增大,不采用该推理改动。证据artifacts/recovery_hourly_20261110T2130_analysis.json,无Test1024,非FID。 ++ ++Kimodo启动有界训练v40:/root/autodl-fs/pangxinlei/models/hfd-recovery-pilot-v40-20261110/kimodo,入口scripts/recovery_training_v40/supervise_pilot.py。从原正式best初始化,以v19为基线只把recovery.temporal_weights.decoded_root_acceleration从0.01提高到0.1,其余训练/验证/恢复配方保持,250步并保存0/125/250对照。该项跟踪解码后根位置二阶差分与参考加速度(包含fps及评分掩码),不是把所有加速度压零。此前引导改动无法兼顾行走与抖动,检验加强监督是否有帮助,不预设改善或放宽门槛。原入口/runtime源码一致、单变量配置核验recovery_v40_checks.json。 ++ ++GPU1启动前空闲,12GiB预算实写fsync预留通过(实际reserve约11.42GiB,非shared df配额推断);监督器587407与真实子进程存活,当前恢复检查初始化,尚未宣称通过。8小时墙钟、原子断点/日志/缓存/容量inode保护保留,启动SHA/健康recovery_v40_launch_kimodo.json、recovery_v40_health.json。HY GPU0继续运行。清理已结束Kimodo v39重复索引328.12MiB,确认旧进程退出、无打开句柄及新副本SHA相同;权重/恢复断点/完整轨迹保留,清单recovery_cleanup_hourly_20261110T2130.json。本地约1.531GiB,共享inode36585,本地/共享实写fsync通过,目录项低于500。未正式开训,半小时调度继续,下次22:00 UTC;正式训练健康启动即取消规则保持。 ++ ++ ++### HY扩展验证v41及Kimodo v40(服务器2026-11-10 22:00 UTC) ++ ++HY v39正常完成250步、恢复读回通过,预留释放;所有训练记录及0/125/250完整轨迹有限,原初始精确复现、参考/案例配对通过。最终幅度0.563、接段1.724、姿态0.418rad、根位置0.439m、jerk302.30;2327进度1.069/1.454,961为-0.069/-0.007,姿态及行走仍不达门槛。对比v36降低jerk但未恢复行走,不延长此配方。Kimodo v40检查196步、结束200/250,真实587571/监督器587407存活;完整3.394GB断点两步恢复通过、loss/梯度和初始/125轨迹有限,初始/参考匹配通过。125小验证幅度0.886、接段2.343、姿态0.442rad、根位置0.367m,现有小验证门槛通过;尚需最终结果及更大validation,不能据此正式开训。证据artifacts/recovery_hourly_20261110T2200_analysis.json,非FID。 ++ ++HY继续冻结扩展诊断v41:/root/autodl-fs/pangxinlei/models/hfd-recovery-broad-v41-20261110/hymotion,入口scripts/recovery_broad_v41_hymotion/supervise_pilot.py。同24validation案例、两种子比较原正式best、v36最终、v39最终三臂;6种交互模式各4例,选择不依赖模型结果并包含原6例,复用Kimodo扩展选样规则。先产生相同24例原正式基线,不沿用6例误差作比较门槛;逐臂使用已有asset/adapter键/shape校验完整重载权重,保持10步/引导1/12帧执行及因果文本,无训练更新/无Test1024。原6例子集需精确复现,各臂参考需一致。此轮为失败分布与覆盖诊断,不是已通过小验证后的正式晋级复核,不宣称候选合格。代码检查recovery_v41_checks.json。 ++ ++GPU0启动前空闲,监督器587842和子进程存活,仍在初始化;2GiB实写fsync预留通过,4小时墙钟、日志/缓存/容量inode保护保留,启动SHA和健康recovery_v41_launch_hymotion.json、recovery_v41_health.json。Kimodo GPU1继续训练,当前预留与恢复断点不动。清理已结束HYv39重复索引328.12MiB,旧进程退出、无打开句柄和新副本SHA一致后删除,所有权重/完整轨迹保留,清单recovery_cleanup_hourly_20261110T2200.json。本地约1.531GiB,共享inode36532,本地/共享实写fsync通过,目录项低于500,共享df不作账号配额。没有正式训练,半小时任务继续,下次22:30 UTC;任一正式训练健康启动即取消规则保持。 ++ ++ ++### Kimodo最终分析及扩大验证v42(服务器2026-11-10 22:30 UTC) ++ ++Kimodo v40正常完成250步,恢复读回通过、预留释放。全部训练loss/梯度以及0/125/250完整轨迹有限,初始精确复现、参考与案例匹配。125步幅度0.886、接段2.343、姿态0.442rad、根位置0.367m、jerk536.63;2327进度0.887/1.015、961为1.272/0.888,四组均达标。250步幅度0.665、接段2.671、姿态0.418、根位置0.493、jerk206.09,但2327进度0.004/0.788、961为0.045/0.162,三组行走失败且根位置超原基线10%。更低jerk伴随行走退化,不能晋级最终权重,不继续延长。证据artifacts/recovery_hourly_20261110T2230_analysis.json,非FID,无Test1024。 ++ ++保留125步候选,启动冻结扩大验证v42:/root/autodl-fs/pangxinlei/models/hfd-recovery-broad-v42-20261110/candidate125,入口scripts/recovery_broad_v42/supervise_pilot.py。同24validation×2种子、各模式4例,原20步guidance2/12帧执行不变。启动前baseline-validation.json替换为v18/baseline的48条原正式模型结果,避免旧6案例误差门槛与扩展样本错配;来源及复现操作写experiment.json和recovery_v42_checks.json,门槛未放宽。需核验原6子集精确复现v40step125、全量参考/案例一致和逐失败案例。小验证通过不能直接启动正式训练。 ++ ++GPU1启动前空闲,新监督器588139与子进程存活,初始化推进;2GiB实写fsync预留通过,4小时墙钟/日志/缓存/容量inode保护保留,recovery_v42_launch_kimodo.json、recovery_v42_health.json。HY v41进程587845/监督器587842存活,GPU0约18GiB,容量心跳更新,第一臂尚无完整输出,不声称已复现,不重启。原来较小HY验证耗时不能直接套到此次24例×三权重上。 ++ ++清理已完成Kimodo v40重复索引328.12MiB,确认旧进程退出、无打开句柄及新副本SHA相同后删除,所有权重/完整恢复断点/轨迹保留;recovery_cleanup_hourly_20261110T2230.json。HY在用资源不动。本地约1.531GiB,共享inode约36500,本地/共享实写fsync通过,共享df不作账号配额,目录项低于500。尚无正式训练,半小时任务保持,下次23:00 UTC;任一正式训练健康启动即取消规则保持。 ++ ++ ++### 扩展结果及Kimodo中点诊断v43(服务器2026-11-10 23:00 UTC) ++ ++HY v41原正式模型扩展基线已完成:幅度0.271、接段1.250、姿态0.373rad、根位置0.768m、jerk331.81,明显位移组合失败23/24。48条轨迹有限、原6例子集逐元素复现;恢复候选尚未完成,进程587845/监督器587842和容量心跳正常,不能只根据基线宣布改善。Kimodo v42完成,预留释放、无错误;完整48条轨迹有限、与扩展原基线参考/案例相同,原6例精确复现v40step125。扩展幅度0.741、接段2.356、姿态0.403rad、根位置0.558m、jerk725.25,行走失败9/24,未过原门槛。对比v19候选失败8/24、jerk906.78、姿态0.416,平滑/姿态有所改善但行走没有整体改善,不能正式开训。证据artifacts/recovery_hourly_20261110T2300_analysis.json,无Test1024,非FID。 ++ ++启动冻结v43诊断:/root/autodl-fs/pangxinlei/models/hfd-recovery-midpoint-v43-20261110/kimodo,入口scripts/recovery_midpoint_v43/supervise_pilot.py。两臂是原v40step125对照及v40/v19两个step125完整denoiser浮点张量等权中点,非浮点缓冲保持一致;两者同初始化/架构,仅加速度训练权重不同。固定0.5而非扫描参数,检验行走与平滑性折中是否存在,不假设权重平均必然改善。实际检查键/shape/dtype一致、中点数值有限、非浮点缓冲相同,recovery_v43_checks.json。使用相同24validation两种子/20步guidance2与原48例基线,原臂需精确复现v42,零训练更新,不导出新大权重;原门槛不变。 ++ ++GPU1启动前空闲,监督器588526和真实子进程存活、初始化中;2GiB实写fsync预留完成,4小时墙钟/日志/缓存/容量inode保护继续,启动及健康recovery_v43_launch_kimodo.json、recovery_v43_health.json。HY GPU0继续。清理已结束v42重复索引328.12MiB,确认旧进程退出、无打开句柄、新副本SHA一致,所有权重/恢复断点/轨迹保留;recovery_cleanup_hourly_20261110T2300.json。本地约1.530GiB,共享inode约36450,本地/共享实写fsync通过,共享df不是账号配额,目录项低于500。尚无正式训练,半小时任务保持,下次23:30 UTC;任一正式训练健康启动即取消规则不变。 ++ ++ ++### 双模型扩展诊断运行检查(服务器2026-11-10 23:30 UTC) ++ ++HY v41进程587845及监督器587842存活,累计约88分钟,GPU0约18GiB且有实际计算活动;原正式模型基线已核验,恢复候选仍未完成,没有新完整结果,不重复启动。Kimodo v43进程588528及监督器588526存活,GPU1约17GiB且计算活跃;v40对照臂48条完整轨迹逐元素精确复现v42,数值有限、原6例和全量参考匹配。中点臂仍运行,不预判权重平均收益。两组无错误或资源停止,证据artifacts/recovery_hourly_20261110T2330_analysis.json。均为冻结诊断,零训练更新,不是正式训练/FID。 ++ ++本地约1.530GiB、共享空闲inode36448,本地/共享1MiB实写fsync通过,共享df不作账号配额。目录项低于500,运行中预留与缓存保持,本轮无新增已核验可清理项,0字节删除,清单recovery_cleanup_hourly_20261110T2330.json。两份文档及任务状态已更新;半小时调度保持,下一次11月11日00:00 UTC,任一正式训练健康启动立即取消416e9f54规则不变。 ++ ++ ++### Kimodo根训练范围v44及HY扩展进展(服务器2026-11-11 00:00 UTC) ++ ++HY v41完成v36臂:幅度0.544、接段1.776、姿态0.402rad、根位置0.738m、jerk357.01,行走失败22/24(原基线23/24);幅度恢复不等于行走恢复,未过门槛。完整48轨迹有限、全量参考/案例匹配、原6例子集逐元素复现v36最终。v39仍运行,真实587845/监督器587842存活,无错误,不抢占GPU0。Kimodo v43两臂完成,预留释放、无错误,中点幅度0.798、接段2.325、姿态0.408、根位置0.564m、jerk732.78,行走失败10/24,较v40的9/24和v19的8/24更差,不采用。完整轨迹有限且参考匹配。证据artifacts/recovery_hourly_20261111T0000_analysis.json,非FID,无Test1024。 ++ ++Kimodo启动v44有界250步:/root/autodl-fs/pangxinlei/models/hfd-recovery-pilot-v44-20261111/kimodo,入口scripts/recovery_training_v44/supervise_pilot.py。以v19为对照、原正式best初始化,仅training.root_trainable由full改projections;根模块input_linear/output_linear训练,内部参数冻结,身体模块全量训练,全部学习率/损失(加速度仍0.01)/链/验证配置保持。基于身体适配必要且多次根全量训练后期行走退化,检验限制根更新自由度是否有帮助,不预设根更新就是原因。使用已有受支持模式,不修改上游。单变量配置及优化器排除冻结参数验证通过,recovery_v44_checks.json;真实模型训练计数/恢复检查待启动完成,不宣称已通过。 ++ ++GPU1启动前空闲,监督器589285和真实子进程存活,处于恢复检查初始化;12GiB预算实写fsync预留通过(实际reserve约11.42GiB),8小时墙钟、日志/缓存/目录项/容量inode/原子断点保护保留,recovery_v44_launch_kimodo.json、recovery_v44_health.json。HY继续扩展诊断。清理已结束v43重复索引328.12MiB,确认旧进程退出、无打开句柄和新副本SHA相同;所有正式权重/恢复断点/候选/轨迹保留,清单recovery_cleanup_hourly_20261111T0000.json。本地约1.530GiB,共享inode36432,本地/共享实写fsync通过,共享df不作账号配额,目录项低于500。无正式训练,半小时调度保持,下次00:30 UTC;正式训练健康启动立即取消规则不变。 ++ ++ ++### 双模型运行检查(服务器2026-11-11 00:30 UTC) ++ ++HY v41监督器587842、真实子进程587845存活,扩展v39臂继续;原基线及v36臂已完成,v36行走失败22/24,尚无新完整候选结果。Kimodo v44监督器589285、训练子进程589455存活,审计182/250步,loss/梯度及已保存完整轨迹有限,初始复现、参考匹配通过。两步恢复读回通过,optimizer/scheduler均为2,完整断点2268515405字节。实际根可训练565253、身体141529356参数,符合根projections模式。125步小验证幅度1.144、接段2.255、姿态0.462rad、根位置0.317m,既定小验证门槛通过;jerk773.16仍需关注。最终及扩展验证未完成,不据此宣布正式训练资格。证据artifacts/recovery_hourly_20261111T0030_analysis.json;非FID,无Test1024。 ++ ++两组无错误收据、容量心跳继续,HY约2小时28分钟,仍在4小时诊断上限内;Kimodo受8小时训练上限保护。本地约1.530GiB、共享空闲inode36397,本地/共享1MiB实写fsync通过,共享df不作账号配额。目录项均低于500;保留运行中预留、缓存、完整断点及轨迹,无新增可安全清理项,删除0字节,清单recovery_cleanup_hourly_20261111T0030.json。未重复启动。尚无正式训练,半小时调度继续,下一检查点01:00 UTC,检查HY扩展结果及Kimodo最终验证;任一正式训练健康启动即取消416e9f54规则保持。 ++ ++ ++### 完成结果及双模型v45(服务器2026-11-11 01:00 UTC) ++ ++HY v41三臂完成、无错误且预留已释放,所有完整轨迹有限、48例参考匹配,原6例子集精确复现各自候选。v39扩展幅度0.495、接段1.718、姿态0.400rad、根位置0.739m、jerk303.50,行走失败22/24;对比v36幅度0.544、jerk357.01、同样22/24失败,平滑收益没有解决广泛行走不足,未过原门槛。Kimodo v44正常完成250步、恢复读回通过、预留释放;0/125/250完整轨迹和训练数值有限,初始及参考匹配。最终幅度1.043、接段2.447、姿态0.439rad、根位置0.454m、jerk394.23,但961两种子进度0.352/0.298,不合格。125步通过小验证(幅度1.144、根位置0.317m、jerk773.16),保留该候选进行扩展复核,不延长退化终点。证据artifacts/recovery_hourly_20261111T0100_analysis.json,非FID,无Test1024。 ++ ++已启动共享根/root/autodl-fs/pangxinlei/models/hfd-recovery-v45-20261111:HY入口scripts/recovery_v45_hymotion/supervise_pilot.py,从原正式best重置进行250步有界实验,相对v36仅training.text_dropout由0.1改0;其余训练参数、执行权重6、因果预测历史及原门槛不变。依据扩展行走普遍不足、此前空文本消融损害行走,检验训练时完整保留文本条件的假设,尚不能归因或宣布改善。Kimodo入口scripts/recovery_v45_kimodo/supervise_pilot.py,冻结v44step125,固定24validation×2种子,沿用原正式48例基线,验证原6例子集精确复现及扩大覆盖后的行走/抖动表现。配置单变量、48例基线和语法检查通过,recovery_v45_checks.json。 ++ ++启动前GPU0/1均空闲;HY监督器590055/子进程590059、Kimodo监督器590058/子进程590060存活,无错误,处于初始化(HY恢复检查尚未完成,不宣称已通过)。各2GiB预算实写fsync预留通过,HY8小时/Kimodo4小时上限及原容量/inode/日志/原子断点保护保持,健康及源码SHA见recovery_v45_health.json和launch收据。清理已结束HYv41、Kimodov44重复启动索引共656.24MiB,确认进程退出、无打开句柄、新副本SHA相同;全部权重/恢复断点/诊断轨迹保留,清单recovery_cleanup_hourly_20261111T0100.json。本地约1.529GiB,共享空闲inode36348,实写fsync检查通过,目录项低于500,共享df不作账号配额。尚无正式训练,保留半小时调度,下一检查点01:30 UTC;任一正式训练健康启动立即取消416e9f54条件不变。 ++ ++ ++### HY v45继续、Kimodo位移监督v46(服务器2026-11-11 01:30 UTC) ++ ++HY v45真实训练590274/监督器590055存活,检查55步、轨迹审计时57/250,loss/梯度有限;原初始轨迹逐元素复现v36初始,2步完整恢复读回通过(optimizer/scheduler为2,136884553字节断点)。尚未到125验证,保留正在运行配方,不重复启动。Kimodo v45扩展验证完成,无错误、预留释放;48条完整轨迹有限、参考/case_keys与原扩展基线一致,原6例精确复现v44step125。幅度0.938、接段2.249、姿态0.426rad、根位置0.561m、jerk900.12,行走失败9/24,较v19的8/24及根位置0.528m未改善。限制根训练范围未解决泛化不足,不晋级正式训练。证据artifacts/recovery_hourly_20261111T0130_analysis.json,非FID,无Test1024。 ++ ++Kimodo实际启动v46:/root/autodl-fs/pangxinlei/models/hfd-recovery-pilot-v46-20261111/kimodo,入口scripts/recovery_training_v46/supervise_pilot.py。以v19原正式best初始化配方为对照,仅recovery.temporal_weights.stage1_root_displacement由1增至2,其余训练/损失/学习率/根全量训练/因果输入/推理不变。检验更强已有位移目标能否减少扩展行走不足,不预设有效,仍需防止抖动与误差退化。既有损失实现不变,只使用训练目标和有效帧掩码;单变量与原基线保持检查recovery_v46_checks.json。250步有界、固定0/125/250两种子验证、原门槛不变,不从退化断点延长。 ++ ++GPU1空闲后启动监督器590548和真实子进程,健康收据recovery_v46_health.json、源码SHA见launch收据;处于恢复检查初始化,不宣称新轮恢复或效果通过。12GiB预算实际写入fsync预留通过,8小时墙钟、容量/inode/目录项/日志/原子断点保护保持;HY GPU0继续。清理已结束Kimodo v45重复索引328.12MiB,确认旧进程退出、无打开句柄和新副本SHA一致;所有模型权重、完整断点及轨迹保留,清单recovery_cleanup_hourly_20261111T0130.json。本地约1.529GiB、共享空闲inode超过36000,本地/共享实写fsync通过,目录项低于500,共享df不作为账号配额。无正式训练,半小时调度继续,下次02:00 UTC;正式训练健康启动即取消416e9f54规则不变。 ++ ++ ++### 双模型中期健康检查(服务器2026-11-11 02:00 UTC) ++ ++HY v45真实进程590274/监督器590055存活,检查131步、审计135/250;Kimodo v46真实进程590704/监督器590548存活,检查196步、审计200/250。双方训练loss/梯度有限,初始/125完整轨迹有限、参考及案例匹配,原初始逐元素复现;两步恢复读回通过,optimizer/scheduler均为2,HY完整断点136884553字节、Kimodo3394249205字节。无错误或watchdog停止收据,GPU0/1分别约20.1/20.3GiB,监督器继续保护。 ++ ++HY125幅度0.644、接段1.729、姿态0.409rad、根位置0.504m(比原初始约恶化12.3%)、jerk383.25;2327两种子进度1.644/2.514,961为-0.267/-0.129,3/4行走组合失败,尚未合格。Kimodo125幅度1.194、接段2.224、姿态0.452rad、根位置0.424m、jerk766.11;2327进度0.750/0.876,961为1.434/0.524,现有小验证门槛通过,最终及扩展验证待完成,不能据此正式开训。证据artifacts/recovery_hourly_20261111T0200_analysis.json,非FID,无Test1024。保持原250步有界预算,不重复启动或修改运行源码。 ++ ++本地约1.529GiB、共享空闲inode36269,本地/共享1MiB实写fsync通过,共享df不作账号配额;两运行目录项低于500,预留及旧断点保护正常。没有新核验可删项,删除0字节,保留活动预留、缓存、权重及轨迹,清单recovery_cleanup_hourly_20261111T0200.json。无正式训练,半小时调度保持,下次02:30 UTC检查最终结果;正式训练健康启动即取消416e9f54规则不变。 ++ ++ ++### Kimodo v46完成及扩展v47(服务器2026-11-11 02:30 UTC) ++ ++HY v45进程590274/监督器590055存活,最新日志249/250,尚无完成收据或最终验证结果;训练数值及已保存初始/125完整轨迹有限,原初始复现和参考匹配通过,恢复检查通过。保留运行中预留及源码,未重复启动。Kimodo v46正常完成250步,预留释放、无错误;初始/125/250完整轨迹有限、参考/案例匹配、原初始精确复现,两步完整恢复通过。最终幅度1.219、接段2.329、姿态0.451rad、根位置0.437m、jerk478.29;2327进度0.864/0.585,961进度0.242/0.237,最终不合格。125小验证通过,但这不等于扩大覆盖合格。证据artifacts/recovery_hourly_20261111T0230_analysis.json,非FID,无Test1024。 ++ ++已在空闲GPU1启动冻结扩展v47:/root/autodl-fs/pangxinlei/models/hfd-recovery-broad-v47-20261111/candidate125,入口scripts/recovery_broad_v47/supervise_pilot.py。使用v46的step125候选,同24validation案例×两种子及原正式48例基线,20步/guidance2,零训练更新,保持原门槛。需核对完整参考及原6例精确复现后比较v19等扩展对照,不延长失败终点。2GiB预算实写fsync预留通过;监督器591195和真实子进程存活,初始化中,无错误,4小时墙钟/容量inode/日志缓存/原子保存保护保持,recovery_v47_checks.json、recovery_v47_health.json和launch收据。 ++ ++清理已结束Kimodo v46重复启动索引328.12MiB,旧进程退出、无打开句柄、新副本SHA相同;保留所有模型权重、完整恢复断点及轨迹,清单recovery_cleanup_hourly_20261111T0230.json。本地约1.529GiB、共享空闲inode超过36000,本地/共享实写fsync通过,目录项低于500,共享df不作账号配额。仍无正式训练,半小时调度保持,下次03:00 UTC检查HY最终结果和Kimodo扩展结果;正式训练健康启动即取消416e9f54规则不变。 ++ ++ ++### 完成分析及历史通道诊断v48(服务器2026-11-11 03:00 UTC) ++ ++HY v45正常完成250步,恢复检查通过、预留释放;训练及0/125/250完整轨迹有限、原初始精确复现和参考配对通过。最终幅度0.700、接段1.758、姿态0.406rad、根位置0.522m(较原初始恶化约16.2%)、jerk366.42;2327进度1.740/2.607,961为-0.326/-0.119,3/4行走组合失败,取消文本dropout未解决行为问题,不延长。Kimodo v47扩展完成、预留释放,无错误;48完整轨迹有限、原扩展参考匹配、原6例精确复现v46step125。幅度0.971、接段2.228、姿态0.421rad、根位置0.581m、jerk961.83,无低于0.1坍缩,但行走失败10/24,比v19的8/24、根位置0.528m及jerk906.78更差,不采用。证据artifacts/recovery_hourly_20261111T0300_analysis.json,非FID,无Test1024。 ++ ++两边实际启动冻结诊断v48,共享根/root/autodl-fs/pangxinlei/models/hfd-recovery-history-v48-20261111/{hymotion,kimodo},入口scripts/recovery_history_v48_{hymotion,kimodo}/supervise_pilot.py。HY固定v36最终候选、原6validation×两种子;Kimodo固定v19step125候选、扩展24validation×两种子。各含自主预测历史对照、只替换过去根平移、只替换过去旋转三臂。根平移是原始轨迹0:3通道,其余旋转3:;历史索引仅0..decision,未来占位仍由末历史帧构造。只作反馈误差定位,真实过去臂不可部署,位置重置造成的进度收益不算自主行走,所有oracle臂明确禁止用于正式晋级;自主对照需精确复现既有结果后才解释差异。不更新训练权重、不导出新大模型,保留原门槛。检查recovery_v48_checks.json,源码SHA见launch收据。 ++ ++GPU0/1启动前空闲;HY监督器591617/子进程591619、Kimodo监督器591621/子进程591622存活,初始化中,无错误,健康recovery_v48_health.json。各2GiB预算实写fsync预留通过,4小时墙钟/容量inode/目录项/日志缓存/原子保存保护保持。清理已结束HYv45、Kimodov47重复索引共656.24MiB,确认旧进程退出、无打开句柄、新副本SHA相同;全部模型权重、完整断点和轨迹保留,清单recovery_cleanup_hourly_20261111T0300.json。本地约1.528GiB、共享inode36185,本地/共享实写fsync通过,目录项低于500,共享df不作账号配额。尚无正式训练,半小时调度保持,下次03:30 UTC检查诊断对照复现与进度;正式训练健康启动即取消416e9f54条件不变。 ++ ++ ++### HY历史诊断细分v49、Kimodo v48继续(服务器2026-11-11 03:30 UTC) ++ ++HY v48三臂完成,无错误且预留释放,完整轨迹有限、同参考匹配,自主预测历史逐元素精确复现v36最终。只替换过去平移后根误差0.055m、行走进度接近1,但jerk2521.70,属于位置重置收益,不是自主行走改善;身体幅度0.607基本不变。只替换过去全部旋转后幅度0.832、姿态误差0.183rad、根位置0.279m,961两种子进度0.507/0.717;接段比10.516、jerk476.27,不可部署。表明旋转历史可能参与行走失败,但尚未区分根朝向和身体姿态,不能把oracle收益作为资格。Kimodo v48监督器591621/子进程591622存活,GPU1约17GiB、计算活动正常,尚在后续诊断;自主对照48条完整轨迹有限并逐元素精确复现v20,参考匹配,原有8/24行走失败不变。证据artifacts/recovery_hourly_20261111T0330_analysis.json,非FID,无Test1024。 ++ ++HY启动冻结v49:/root/autodl-fs/pangxinlei/models/hfd-recovery-history-v49-20261111/hymotion,入口scripts/recovery_history_v49_hymotion/supervise_pilot.py。相同v36最终候选、6validation×两种子,三臂为自主对照、仅替换过去根朝向3:6、仅替换过去身体关节旋转6:69;已核对原始3:69为22个轴角关节,保持根平移预测不重置,历史索引仅0..decision,不读未来GT。零训练更新,oracle臂禁止正式晋级;先要求自主对照精确复现,再据通道差异设计恢复实验。检查recovery_v49_checks.json。 ++ ++GPU0启动前空闲,监督器591942及真实子进程存活,无错误、初始化中;2GiB预算实写fsync预留通过,4小时墙钟/容量inode/日志缓存/原子保存保护保持,健康recovery_v49_health.json和启动SHA收据。Kimodo原运行继续、不重复启动。清理已结束HYv48重复索引328.12MiB,进程退出、无打开句柄、新副本SHA相同;所有权重/断点/轨迹保留,清单recovery_cleanup_hourly_20261111T0330.json。本地约1.528GiB、共享inode超过36000,实写fsync通过、目录项低于500,共享df不作账号配额。尚无正式训练,半小时调度继续,下次04:00 UTC;正式训练健康启动即取消416e9f54条件保持。 ++ ++ ++### 历史诊断完成及双模型混合历史v50(服务器2026-11-11 04:00 UTC) ++ ++HY v49和Kimodo v48均正常完成、预留释放;三臂完整轨迹有限、参考配对通过,自主对照逐元素精确复现HYv36最终及Kimodov20扩展。HY仅替换过去根朝向时961进度0.209/0.146仍不足,2327一seed超2;仅替换身体旋转时961仍反向-0.256/-0.108、接段10.516。与此前全部旋转替换相比,单通道修正不足,不能简单归因为某一独立通道。Kimodo替换平移使进度接近1但jerk5448.41,属于位置重置;替换旋转后姿态0.183rad、根位置0.383m,但接段6.417、jerk1256.41、行走失败7/24(自主对照8/24),仍不可部署。oracle臂不作为正式晋级依据。证据artifacts/recovery_hourly_20261111T0400_analysis.json,非FID,无Test1024。 ++ ++已启动双模型v50:/root/autodl-fs/pangxinlei/models/hfd-recovery-pilot-v50-20261111/{hymotion,kimodo},入口scripts/recovery_training_v50_{hymotion,kimodo}/supervise_pilot.py。分别以HYv36、Kimodov19为对照,从原正式best重置,仅pilot.predicted_probability从1降至0.75,使训练含25%真实历史窗口;其余训练参数、损失、学习率、链长、文本和0/125/250固定两种子验证均不变。诊断提示历史与姿态误差耦合,但并未证明混合历史训练有效,此轮检验是否有助于改善反馈误差。推理仍完全自主、无真实过去替换,无未来GT,无Test1024,原门槛保持;不延长失败断点。配置检查recovery_v50_checks.json。 ++ ++启动前GPU0/1空闲;两监督器及真实子进程存活、恢复检查初始化中,无错误,不宣称本轮恢复或效果已通过。HY2GiB、Kimodo12GiB预算实写fsync预留完成,8小时墙钟及容量/inode/日志缓存/原子断点保护保持,进程及源码SHA见recovery_v50_health.json和launch收据。清理已结束HYv49、Kimodov48重复索引共656.24MiB,旧进程退出、无打开句柄、新副本SHA一致后删除;正式权重、候选、完整断点及轨迹全部保留,清单recovery_cleanup_hourly_20261111T0400.json。本地约1.528GiB、共享inode超过36000,实写fsync检查通过、目录项低于500,共享df不作账号配额。无正式训练,半小时调度保持,下次04:30 UTC检查恢复和训练进度;正式训练健康启动即取消416e9f54条件不变。 ++ ++ ++### 双模型v50运行检查(服务器2026-11-11 04:30 UTC) ++ ++HY真实训练592427/监督器592244存活,审计98/250,尚未完成125验证;Kimodo真实训练592540/监督器592293存活,检查200步、审计205/250。双方两步完整断点恢复读回通过,optimizer/scheduler均为2,HY断点136888521字节、Kimodo3394253045字节;训练loss/梯度和已保存完整轨迹有限,初始逐元素复现原对照、参考/案例匹配。Kimodo日志已包含预测与真实历史链,混合训练路径生效;推理保持完全自主。无错误或watchdog停止,GPU0/1分别约20.1/20.3GiB,监督器正常。 ++ ++Kimodo125幅度0.764、接段2.262、姿态0.413rad、根位置0.520m(比原初始恶化约17.8%)、jerk590.96;2327两种子进度0.907/0.720,961为0.806/0.143,一组明显位移失败,未过门槛,不能把jerk下降视为已恢复质量。等待原250步有界预算的最终验证,不重复启动、不改动运行源码。证据artifacts/recovery_hourly_20261111T0430_analysis.json,非FID,无Test1024。 ++ ++本地约1.527GiB、共享空闲inode36041,本地/共享1MiB实写fsync通过,目录项低于500,共享df不作账号配额。活动预留、完整断点、缓存及轨迹保留,本轮无新核验可清理项,删除0字节,清单recovery_cleanup_hourly_20261111T0430.json。尚无正式训练,半小时调度继续,下次05:00 UTC检查中期/最终结果;正式训练健康启动即取消416e9f54规则不变。 ++ ++ ++### Kimodo v50完成与细分历史诊断v51(服务器2026-11-11 05:00 UTC) ++ ++HY v50进程592427/监督器592244存活,最新日志249/250、无错误、最终验证待返回。125幅度0.447低于0.5,姿态误差0.448rad较原初始恶化约20.3%,961进度-0.045/-0.001,尚不合格,不重复启动或覆盖运行源码。Kimodo v50正常完成250步、恢复通过、预留释放;双方训练及已保存初始/中期/最终完整轨迹有限,原初始精确复现、参考/案例匹配。Kimodo最终幅度0.658、接段2.562、姿态0.427rad、根位置0.409m、jerk372.26;2327进度0.373/1.001,961为0.771/0.221,2/4行走组合失败,混合历史未过门槛,不能以低jerk当恢复成功。证据artifacts/recovery_hourly_20261111T0500_analysis.json,非FID,无Test1024。 ++ ++Kimodo已启动冻结v51:/root/autodl-fs/pangxinlei/models/hfd-recovery-history-v51-20261111/kimodo,入口scripts/recovery_history_v51_kimodo/supervise_pilot.py。沿用v19step125、相同24validation×两种子及原正式48例基线,将v48全部旋转历史诊断细分为自主对照、仅替换过去根朝向3:6、仅替换过去身体旋转6:69;根平移保持预测,历史索引0..decision,未来占位不取GT。用于区分反馈误差来源,零训练更新,oracle臂不可部署且禁止正式晋级,自主对照必须精确复现。检查recovery_v51_checks.json。 ++ ++GPU1空闲后启动监督器593012及真实子进程,存活、初始化中、无错误;2GiB预算实写fsync预留通过,4小时墙钟/容量inode/日志缓存/原子保存保护保持,健康recovery_v51_health.json及launch源码SHA。HY GPU0继续。清理已结束Kimodo v50重复索引328.12MiB,旧进程退出、无打开句柄、新副本SHA一致后删除,所有模型权重、完整断点及轨迹保留,清单recovery_cleanup_hourly_20261111T0500.json。本地约1.527GiB、共享inode超过36000,本地/共享实写fsync通过、目录项低于500,共享df不作账号配额。尚无正式训练,半小时调度继续,下次05:30 UTC;正式训练健康启动即取消416e9f54规则不变。 ++ ++ ++### HY v50完成及标签过渡v52(服务器2026-11-11 05:30 UTC) ++ ++HY v50正常完成250步、无错误、预留释放、恢复通过;完整初始/125/250轨迹有限、参考配对、原初始精确复现。最终幅度0.557、接段1.887、姿态0.429rad(原初始+15.2%)、根位置0.450m、jerk358.13;2327进度1.316/1.838,961为-0.062/-0.006,姿态及行走失败,不延长混合历史配方。Kimodo v51真实593013/监督器593012存活,GPU1约17GiB并有计算活动,无错误;自主对照48条完整轨迹精确复现v20,参考一致且有限,后续诊断尚未完成,不预判oracle收益。证据artifacts/recovery_hourly_20261111T0530_analysis.json,非FID,无Test1024。 ++ ++HY启动v52有界250步:/root/autodl-fs/pangxinlei/models/hfd-recovery-pilot-v52-20261111/hymotion,入口scripts/recovery_training_v52/supervise_pilot.py。回到v36原正式best初始化对照,唯一语义变量标签SO3/高度过渡6帧缩至3帧,pilot和curriculum字段同步;其他训练参数、全预测历史、文本、损失、执行12帧及原验证保持。依据旋转历史误差耦合及混合历史无效,检验更快纠正监督姿态是否有效;存在接段恶化风险,原门槛不变,不预设改善。已有过渡实现不改,测试第三个未来帧起精确对齐目标、XY保持、有效掩码保持、输入不修改、真实历史恒等及数值有限通过,recovery_v52_checks.json。只修改训练标签,不增加未来GT输入,不改变自主推理,无Test1024。 ++ ++GPU0启动前空闲,监督器593503及真实子进程存活、恢复检查初始化中、无错误;2GiB预算实写fsync预留通过,8小时墙钟及容量/inode/日志缓存/原子断点保护保持,健康recovery_v52_health.json及launch源码SHA。不宣称新轮恢复通过。Kimodo原诊断继续。清理已结束HYv50重复索引328.12MiB,旧进程退出、无打开句柄和新副本SHA一致后删除,权重/断点/轨迹保留,清单recovery_cleanup_hourly_20261111T0530.json。本地约1.527GiB、共享inode超过35000,实写fsync通过、目录项低于500,共享df不作账号配额。无正式训练,半小时调度继续,下次06:00 UTC;正式训练健康启动立即取消416e9f54规则不变。 ++ ++ ++### HY v52继续及Kimodo标签过渡v53(服务器2026-11-11 06:00 UTC) ++ ++HY v52真实训练593626/监督器593503存活,124/250步、中期验证待返回;两步完整恢复读回通过,optimizer/scheduler均为2,断点136884553字节。训练数值及已存完整轨迹有限,原初始精确复现、参考配对,无错误。不重复启动或覆盖活动源码。Kimodo v51完成三臂、预留释放;48例完整轨迹有限、参考匹配、自主对照精确复现v20。自主行走失败8/24;仅替换过去根朝向失败10/24,幅度1.115、接段2.343、姿态0.387rad、根位置0.617m、jerk1000.12;仅替换身体旋转失败11/24,幅度1.251、接段6.582、姿态0.205rad、根位置0.549m、jerk1635.70。单通道oracle修正没有恢复自主行走,oracle不可晋级。证据artifacts/recovery_hourly_20261111T0600_analysis.json,非FID,无Test1024。 ++ ++Kimodo已启动v53:/root/autodl-fs/pangxinlei/models/hfd-recovery-pilot-v53-20261111/kimodo,入口scripts/recovery_training_v53/supervise_pilot.py。从原正式best初始化,以v19为对照,仅pilot.blend_frames从6改3,检验更快的标签姿态/高度纠正;此前12帧过渡和混合历史均未解决行走失败,因此不重复旧配方。250步有界、0/125/250固定两种子验证;训练、损失、学习率、推理、原门槛均不变。只改变标签不增加未来GT输入,存在接段恶化风险,不预设改善。准备时已移除误带的HY专属curriculum字段,未启动任何错误配置训练。三帧目标数值测试及单变量合同检查通过,recovery_v53_checks.json。 ++ ++GPU1空闲后监督器启动,真实子进程初始化中;12GiB预算实写fsync预留通过,8小时墙钟、磁盘/inode、日志缓存及原子断点保护保持。新轮恢复结果尚待返回,不宣称通过。健康及源码SHA见recovery_v53_health.json和launch收据。HY GPU0继续运行。仅清理已结束Kimodo v51重复索引328.12MiB,确认旧进程退出、无打开句柄、新副本SHA一致,权重/断点/完整轨迹保留,清单recovery_cleanup_hourly_20261111T0600.json。本地约1.526GiB,共享inode超过35000,实写fsync通过,目录项均低于500;共享df不作账号配额。无正式训练,半小时调度继续,下次06:30 UTC;任一正式训练健康启动即取消416e9f54的规则保持。 ++ ++ ++### 双模型标签过渡实验运行检查(服务器2026-11-11 06:30 UTC) ++ ++HY v52真实训练593626/监督器593503存活,审计190/250;Kimodo v53真实训练594424/监督器594263存活,125/250且中期验证尚待返回。双方两步完整断点恢复读回通过,optimizer/scheduler均为2,HY断点136884553字节、Kimodo3394249141字节。训练日志数值与已保存完整轨迹有限、参考和案例匹配、原初始精确复现,运行源码SHA与启动收据一致,无错误/监督器停止。GPU0/1占用约20.1/20.3GiB。HY日志full_predicted_history_fraction统计历史60帧是否全由模型生成,链启动阶段可低于1,不代表配置概率变为0.75;pilot.predicted_probability仍为1。 ++ ++HY125幅度0.538、接段2.124、姿态0.419675rad(原初始+12.7%)、根位置0.440253m、jerk368.13;2327两种子进度1.159/1.348,961为-0.031/0.001,姿态与行走不合格。Kimodo尚无中期结果,不能以初始结果判断本轮改善。双方继续已授权250步有界预算,不重复启动、不延长失败配方、不覆盖运行源码;下一次分析各自完成结果。证据artifacts/recovery_hourly_20261111T0630_analysis.json,非FID,无Test1024。 ++ ++本地约1.526GiB、共享空闲inode35910,本地及共享1MiB实写fsync检查通过,目录项均低于500,共享df不作账号配额。活动预留、缓存、完整断点及轨迹保留,无新增核验可删除项,本轮清理0字节,清单recovery_cleanup_hourly_20261111T0630.json。尚无正式训练,半小时调度保持,下次07:00 UTC;正式训练健康启动立即取消416e9f54的条件不变。 ++ ++ ++### 标签过渡实验完成及双模型v54(服务器2026-11-11 07:00 UTC) ++ ++HYv52与Kimodov53正常完成250步、旧进程退出、预留释放;恢复检查均通过,初始/125/250完整轨迹有限、参考配对、原初始精确复现。HY最终幅度0.609、接段2.452、姿态0.413428rad(原初始+11.1%)、根位置0.460978m、jerk363.53;2327进度1.552/1.887,961为-0.076/-0.070,仍不合格;较v36接段1.780恶化而姿态/行走未改善。Kimodo125幅度1.151、接段2.387、姿态0.44648rad、根位置0.43974m、jerk816.99,961一seed进度0.471低于0.5,不能放宽门槛;最终幅度1.210、接段2.652、姿态0.44417rad、根位置0.39969m、jerk514.15,2327进度0.814/0.455,961为0.794/0.195,两组失败。证据recovery_hourly_20261111T0700_analysis.json包括每例段内突变、根速度/jerk、坍缩及完整轨迹核对,非FID,无Test1024。 ++ ++启动双模型v54有界250步,根目录/root/autodl-fs/pangxinlei/models/hfd-recovery-pilot-v54-20261111,入口scripts/recovery_training_v54_{hymotion,kimodo}/supervise_pilot.py。HY回到v36原正式best初始化,仅旋转速度辅助权重0改0.05,原暖启动200步保持,检验姿态动态监督;不同于早期带根速度辅助/不同历史配方,不宣称该损失首次尝试。Kimodo回到v19原正式best,仅身体学习率曲线减半(1e-6/2e-7变5e-7/1e-7),根学习率2e-7/5e-8保持,检验身体更新过快与中后期退化的关系;区别于v30同时减半双模块。两者均非旧断点续训。其余配置、标签6帧、因果输入、推理、固定两种子验证及原验收保持;假设尚待验证。逐字段单变量比较及HY辅助损失有限/非零梯度检查通过,recovery_v54_checks.json。 ++ ++HY2GiB/Kimodo12GiB预算实写fsync预留通过后,在空闲GPU0/1分别启动监督器与真实子进程;新轮恢复检查初始化中,不宣称已通过。8小时墙钟、容量/inode、日志缓存、原子断点保护保持,健康和源码SHA见recovery_v54_health.json及launch收据。清理已结束HYv52/Kimodov53重复索引共656.24MiB,确认旧进程退出、无打开句柄、新副本SHA一致,权重/完整恢复断点/轨迹全部保留,清单recovery_cleanup_hourly_20261111T0700.json。本地约1.526GiB、共享inode超过35000,实写fsync检查通过,目录项低于500,共享df不作为账号配额。无正式训练,半小时调度继续,下次07:30 UTC;任一正式训练健康启动立即取消416e9f54规则不变。 ++ ++ ++### 双模型v54运行及Kimodo中期候选(服务器2026-11-11 07:30 UTC) ++ ++HY真实训练595667/监督器595433存活,审计90/250;Kimodo真实训练595769/监督器595479存活,检查184步、审计188/250。双方两步完整恢复读回通过,optimizer/scheduler均为2,HY断点136884553字节、Kimodo3394249205字节;训练数值有限,已保存完整轨迹有限、参考/案例匹配、原初始精确复现。源码SHA与启动收据一致,无错误/监督停止,GPU0/1占用约20.1/20.3GiB,日志持续更新。HY尚无125中期结果,不据初始结果推断本轮效果。 ++ ++Kimodo125在当前6案例×两种子上通过原门槛:幅度0.791598、接段2.058991、姿态0.423648rad、根位置0.314195m、jerk639.63;2327进度0.696/0.787,961为1.148/0.685,明显位移四组均在0.5–2。jerk仍高于原初始357.30,不能把小样本门槛通过等同于动作质量完全恢复。保留125候选,先完成当前250步有界预算;GPU1释放后优先对125进行相同24案例×两种子的扩展验证,与原正式48例基线及v19候选比较。未完成扩展验证前不启动正式训练,不重复占用GPU,不延长预算。证据recovery_hourly_20261111T0730_analysis.json包含完整轨迹核对及每例幅度、段内突变、速度、jerk、进度;非FID,无Test1024。 ++ ++本地约1.526GiB、共享空闲inode35821,本地及共享1MiB实写fsync检查通过、目录项低于500,共享df不作账号配额。活动预留、缓存、完整断点及候选保留,无新增核验可删除项,本轮清理0字节,清单recovery_cleanup_hourly_20261111T0730.json。尚无正式训练,半小时调度保持,下次08:00 UTC;正式训练健康启动立即取消416e9f54条件不变。 ++ ++ ++### 用户结束定时检查(2026-11-11T07:52:28.163857+00:00) ++ ++按用户明确指令,已执行 paseo heartbeat delete 416e9f54,CLI返回deleted,并核实调度记录不存在。任务状态已更新为cancelled_by_user,不再自动检查,不自动重建。此次操作仅取消心跳调度,不终止已授权训练,也不关闭监督器的容量/inode、断点保护及资源清理。 ++ ++ ++## 2026-11-11 三模型训练交付 ++ ++原训练工作区、代码差异、正式 latest/best、恢复候选、数据索引和特征缓存、环境锁、历史日志及学习曲线已上传并核验到以下私有仓库: ++ ++- [genmo](https://huggingface.co/starstone3/HFD-GENMO-SMPL-Training),commit `979e5130d6d2289582e8171830700d49dd0cdaf2`,清单 94 个文件。 ++- [hymotion](https://huggingface.co/starstone3/HFD-HY-Motion-SMPL-Training),commit `f858d43583e5897f850394fef886e7860f3a84f9`,清单 97 个文件。 ++- [kimodo](https://huggingface.co/starstone3/HFD-Kimodo-SMPLX-Training),commit `c9a71c47e7e16e9700779e986a607da6d920669e`,清单 107 个文件。 ++ ++三个模型均已实际短程续训并保存完整更新断点:HY 14000→14002(loss/LR/样本与原记录一致);Kimodo 98000→98002(loss/LR一致);GENMO 98000→98002(LR一致,loss略有差异,不能称为逐步精确一致)。GENMO 保存时触发容量受控停止,Kimodo 验收监督器异常和 HY 路径身份修复均已保留证据。原始数据和部分外部基础/文本/许可资产以固定来源与 SHA 清单交付,未声称已做全新环境重建。三个正式最新断点均到100K原预算,后续新阶段调度迁移尚未验证。 ++ ++Kimodo 原 best 仅权重;另交付与 best 参数逐位相同的同一步完整 last 对应文件。每仓库 README/acceptance-report/MANIFEST 记录具体命令、缺口和验证范围。心跳416e9f54已删除,不重建;既有训练未被终止。 ++ ++本地交付证据:`/root/autodl-fs/pangxinlei/handoffs/three-human-models-20261111/FINAL_DELIVERY.json`。核验时间:2026-11-11T08:35:41.057469+00:00。 ++ ++HY pilot54 已补交 step200 的最新完整断点与明确截止时间的日志快照;活动训练后续变化不包含在快照中。最终 MANIFEST 已从固定远端提交下载回读核验。最终核验时间:2026-11-11T08:39:04.787213+00:00。 ++ +diff --git a/SESSION_FORMAT.md b/SESSION_FORMAT.md +deleted file mode 100644 +index 03f5053..0000000 +--- a/SESSION_FORMAT.md ++++ /dev/null +@@ -1,82 +0,0 @@ +-# HFD session 格式与当前输入边界 +- +-更新:2026-10-21。正式默认为英文前景、`aligned_reference_context`、 +-`clock_policy: microphone`、`text_conditioning: none`:离线原文词对齐、最近30秒 +-已释放文本,不运行ASR或翻译。五模型的具体输入/输出、帧率与实现状态见 +-[当前协议](docs/HFD_HUMAN_BASELINE_INPUT_PROTOCOL.md),不再沿用旧中文翻译/ASR默认说明。 +- +-portable不携带原始tar、模型权重、缓存、视频或训练产物。HFD原始数据在 +-`/root/autodl-fs/pangxinlei/hfd_data`,按索引读取tar,保留分片,不全量解压。 +-原生训练session按配方动态组装并缓存在进程内;不是必须提前导出所有NPZ。 +-第三方读取已经组装的生产session。人体源动作也需要按计划组装,不能直接当完整session。 +- +-## 导出格式 +- +-`prepare_sessions.py`调用原生FullDuplexAudioCommandDataset;导出用于固定评测和短测, +-不改变任务、背景、监听/保持/过渡或事件时间。命令示例: +- +-```bash +-python prepare_sessions.py --config configs/sessions.yaml +-python prepare_sessions.py --config configs/sessions.yaml --set split=val +-python validate_sessions.py /data/hfd-sessions/train +-``` +- +-每个G1 session为NPZ及同名JSON: +- +-| 字段 | 内容 | +-| --- | --- | +-| `motion_q29_xyz_quat` | `[T,36]`:IsaacLab顺序29关节+根xyz+wxyz四元数;50Hz | +-| `audio` | 16kHz单声道混合PCM,保留原前景语音与背景 | +-| `control_labels` | 0=listen、1=motion、2=idle;监督/评测元数据,不作为输入 | +-| `turn_index_at_decision` | turn元数据,仅审计/展示,不作为输入 | +-| `valid_start_count` | 原生有效决策起点数;各模型再检查自己的窗口覆盖 | +-| `.json` | 源动作/音频、实际事件位置、关节/四元数约定、split、哈希 | +- +-`full_session`包含listen、motion、idle、过渡及事件前后的合法训练窗口。 +-`motion_only`仅保留作历史消融。源片段在session内部按配方组合,但训练窗口不跨session。 +-候选session数不是独立原始动作数,也不是已经缓存到磁盘的NPZ数量。 +- +-## 当前文本sidecar +- +-`data.asr_dir`是历史字段名;在当前协议下实际读取 +-`.reference.json`,不是`.asr.json`。动态训练则从同一离线对齐库和 +-当次session音频组件偏移直接生成相同记录,不需要在线ASR。 +- +-格式 `hfd-aligned-reference-v1` 包含NPZ/元数据哈希、源TTS片段、估计词末时间、 +-可用时间、首次释放时间、30秒上下文记录及异常fallback。已释放的完整句子可以输入; +-尚未释放的句尾、最终修正指令或整条GT全文不能提前提供。 +- +-动作时刻t对应麦克风t+0.2秒。词末加0.1秒guard,向上取整到麦克风 +-0.6/1.0/1.4…秒释放,模型只选择 `r_k <= t+0.2` 的记录。 +-0.6秒音频窗每0.4秒更新,是重叠窗口,不是每步新增0.6秒。 +-文本按释放点最近30秒滚动,过期可以为空;空文本不无限沿用旧命令。 +-无效词跨度在整个对应音频组件结束后释放。对齐误差和guard未完成逐词人工认证。 +- +-这是正确前景说话人的oracle转写实验:背景PCM保留,背景文字不输入。 +-旧 `causal_asr_context`、`stable_latest_sentence`、Whisper与中文→英文翻译仅用于 +-显式历史复现,不与当前结果混用。现有原始双语数据和历史产物不改写。 +- +-## 动作与坐标 +- +-原始NPZ不是MuJoCo/native关节顺序,不能直接切片当qpos。 +-`hfd_bridge/joints.py`按名称转换:TextOp原生23关节/57D,OMG原生29关节/125D。 +-当前TextOp未建模的腕关节固定在bootstrap姿态;指标只对其原生23关节计分。 +-根四元数在原始HFD文件为wxyz,进入各框架时显式重排。 +- +-GenMo读取配对人体trans、root_orient、pose_body、betas和sample_times, +-训练与推理不使用G1关节作为网络输入或目标。当前人体组装仍使用原生G1配方/ +-配对映射追踪源与时序;不能说整个数据构造程序与G1文件完全无关。 +-人体输入与retarget契约见[人体协议](docs/HFD_HUMAN_BASELINE_INPUT_PROTOCOL.md) +-及[评测转换说明](docs/HUMAN_RETARGET_EVALUATION.md)。 +- +-## 训练与连续评测 +- +-各模型使用自己的原生表示、归一化、损失和框架。OMG与GenMo训练提供GT历史, +-TextOp沿用原生分阶段历史课程。连续评测只初始化一次GT状态,之后反馈预测; +-真值历史诊断必须明确标注。GT事件仅用于选择/标注片段,不作为模型条件。 +- +-评测保留原混合音轨,将共同0.2秒microphone前置帧和bootstrap从动作评分中排除, +-不做DTW或事后根对齐。OMG/TextOp原生G1输出不做人体retarget;人体模型对完整轨迹 +-做一次retarget,并同时报告人体本身误差及人体GT重定向对照。 +-当前独立test未建立;val用于训练监测,不能称完全未见测试集。视频为运动学回放, +-不声称物理跟踪、端到端实时语音系统或语义任务成功率。 +diff --git a/UPSTREAM_LOCK.json b/UPSTREAM_LOCK.json +index 6c9873d..f30a85f 100644 +--- a/UPSTREAM_LOCK.json ++++ b/UPSTREAM_LOCK.json +@@ -2,11 +2,17 @@ + "external_repositories": { + "genmo": { + "url": "https://github.com/NVlabs/GENMO.git", +- "commit": "16bebf402d8893184249ee206d957b8248cd8310" ++ "commit": "16bebf402d8893184249ee206d957b8248cd8310", ++ "local_packaging_patch": "patches/genmo-numpy1264.patch", ++ "replay_numpy": "1.26.4" + }, + "kimodo": { + "url": "https://github.com/nv-tlabs/kimodo.git", + "commit": "e1c219c99b4b94ba48be1fb6a1cbc64ecee4c954" ++ }, ++ "hymotion": { ++ "url": "https://github.com/Tencent-Hunyuan/HY-Motion-1.0.git", ++ "commit": "4e426f5a1021cbcf7f375458c37b840ee7225229" + } + }, + "genmo_helper_assets": { +@@ -118,5 +124,16 @@ + "revision": "93ba4a7a9c66b3e1b7041e5f4ea038ef2418b948", + "metadata_manifest": "TRAINING_ASSETS.json", + "note": "Includes paired SMPL-X and all prior shards; do not bootstrap using the older pre-SMPL-X revision." ++ }, ++ "hfd_final_release": { ++ "data_repo": "Tinghong-Ye22/data_vqa_all_1025", ++ "data_revision": "bdca2fe7581c263e9ceb3a96204b314e199faddf", ++ "pipeline_repo": "Tinghong-Ye22/data_vqa_pipeline_1025", ++ "pipeline_revision": "9a2407adeed36a2645110390f1521304e36a2215", ++ "profile": "isolated_benchmark_1024_20261026", ++ "replay_arithmetic": "numpy2_float32_scalars", ++ "vendored_files": { ++ "upstream/hfd/src/psi/data/seed/third_party_dataset.py": "212fd08f50a6211caf3442d19772c8437b4de36ca0fb6de7f6865bfd621e84da" ++ } + } + } +diff --git a/configs/genmo.yaml b/configs/genmo.yaml +index 68fd000..a442cf6 100644 +--- a/configs/genmo.yaml ++++ b/configs/genmo.yaml +@@ -2,13 +2,15 @@ + model: genmo + genmo_root: /root/autodl-tmp/hfd-human-models/genmo + python: /root/autodl-tmp/envs/genmo/bin/python +-gpu: 1 +-output: /root/autodl-tmp/runs/hfd-genmo-english-v2 +-weights_output: /root/autodl-fs/pangxinlei/models/genmo/hfd-english-v2 ++gpu: 0 ++output: /root/autodl-tmp/runs/hfd-genmo-english-v2-split20261026-replay1264 ++weights_output: /root/autodl-fs/pangxinlei/models/genmo/hfd-english-v2-split20261026-replay1264 + init_checkpoint: /root/autodl-fs/pangxinlei/models/genmo/gem_smpl.ckpt + text_encoder: /root/autodl-fs/pangxinlei/models/genmo/t5-3b-local + body_models: /root/autodl-fs/pangxinlei/models/body_models + data: ++ replay_audit_status: /root/autodl-tmp/cache/hfd/genmo-replay1264/audit.json ++ require_replay_risk_gate: true + config: /root/autodl-tmp/hfd-baseline-portable/configs/sessions.yaml + reference_cache: /root/autodl-tmp/cache/hfd/reference-alignments.sqlite + language: en +@@ -43,6 +45,7 @@ strategy: + max_frames: 900 + seed: 12345 + training: ++ early_checkpoint_steps: [128, 256, 512, 1024] + # Total budget; review validation and closed-loop rollouts at 20K/40K/60K. + # Cosine T_max follows max_steps; resume must retain this original budget. + max_steps: 100000 +diff --git a/configs/genmo_legacy.yaml b/configs/genmo_legacy.yaml +index 57a7ea1..fa7f7aa 100644 +--- a/configs/genmo_legacy.yaml ++++ b/configs/genmo_legacy.yaml +@@ -3,8 +3,8 @@ model: genmo + genmo_root: /root/autodl-tmp/hfd-human-models/genmo + python: /root/autodl-tmp/envs/genmo/bin/python + gpu: 1 +-output: /root/autodl-tmp/runs/hfd-genmo-english-train +-weights_output: /root/autodl-fs/pangxinlei/models/genmo/hfd-english-train ++output: /root/autodl-tmp/runs/hfd-genmo-english-legacy-split20261026 ++weights_output: /root/autodl-fs/pangxinlei/models/genmo/hfd-english-legacy-split20261026 + init_checkpoint: /root/autodl-fs/pangxinlei/models/genmo/gem_smpl.ckpt + text_encoder: /root/autodl-fs/pangxinlei/models/genmo/t5-3b-local + body_models: /root/autodl-fs/pangxinlei/models/body_models +diff --git a/configs/kimodo.yaml b/configs/kimodo.yaml +index 5c27a61..4537176 100644 +--- a/configs/kimodo.yaml ++++ b/configs/kimodo.yaml +@@ -1,11 +1,11 @@ + # Full native English paired SMPL-X; text-only causal 30s context. + model: kimodo +-gpu: 0 ++gpu: 1 + model_dir: /root/autodl-fs/pangxinlei/models/kimodo/Kimodo-SMPLX-RP-v1 + session_config: /root/autodl-tmp/hfd-baseline-portable/configs/sessions.yaml + reference_cache: /root/autodl-tmp/cache/hfd/reference-alignments.sqlite +-output: /root/autodl-tmp/runs/hfd-kimodo-english-train +-weights_output: /root/autodl-tmp/models/kimodo/hfd-english-train ++output: /root/autodl-tmp/runs/hfd-kimodo-english-split20261026 ++weights_output: /root/autodl-fs/pangxinlei/models/kimodo/hfd-english-split20261026 + representation: + causal_root: true + future_frames: 60 +diff --git a/configs/omg.yaml b/configs/omg.yaml +index 9d973ce..c94cf3e 100644 +--- a/configs/omg.yaml ++++ b/configs/omg.yaml +@@ -1,4 +1,4 @@ +-# English foreground only; original indices/seed preserved. No ASR or translation. ++# English foreground only; isolated final split and recorded per-session seeds. + # Full-data English subset; long-run step budget still requires calibration. + schema_version: 1 + mode: train +@@ -7,9 +7,9 @@ framework: omg_lightning + paths: + data_disk: ${oc.env:HFD_DATA_DISK,/root/autodl-tmp} + upstream_root: upstream/omg +- session_root: ${paths.data_disk}/datasets/hfd-sessions ++ session_root: ${paths.data_disk}/datasets/hfd-sessions-split20261026 + cache_root: ${paths.data_disk}/cache +- output_dir: ${paths.data_disk}/runs/hfd-omg-english-train ++ output_dir: ${paths.data_disk}/runs/hfd-omg-english-split20261026 + runtime: + python: ${paths.data_disk}/envs/omg/bin/python + num_gpus: 1 +@@ -21,6 +21,8 @@ runtime: + distributed_timeout_seconds: 1800 + environment: + HFD_LOCAL_INDEX_CACHE: '1' ++ HFD_NATIVE_STARTUP_CACHE: '1' ++ HFD_STARTUP_INDEX_CACHE_DIR: ${paths.cache_root}/hfd/startup-indexes + data: + train_sessions: + - ${paths.session_root}/train/*.npz +@@ -28,7 +30,7 @@ data: + - ${paths.session_root}/val/*.npz + window_policy: full_session + max_windows: null +- cache_sessions: 512 ++ cache_sessions: 2048 + asr_dir: ${paths.session_root}/reference + asr_latency_seconds: 0.0 + audio_policy: hfd_past_2s +@@ -41,7 +43,7 @@ data: + prefetch_batches: 1 + native: + config: configs/sessions.yaml +- memory_mib: 2048 ++ memory_mib: 8192 + reference_cache: ${paths.cache_root}/hfd/reference-alignments.sqlite + reference_guard_seconds: 0.1 + refresh_batches: 256 +@@ -56,7 +58,7 @@ training: + precision: bf16-mixed + accumulate_grad_batches: 1 + gradient_clip_val: 1.0 +- save_every: 2000 ++ save_every: 256 + eval_every: 2000 + experiment: 50m + scheduler: +diff --git a/configs/omg_variants.yaml b/configs/omg_variants.yaml +index 87e9636..9d98296 100644 +--- a/configs/omg_variants.yaml ++++ b/configs/omg_variants.yaml +@@ -16,11 +16,11 @@ + experiment: 300m + checkpoint: checkpoints/paper/300m/sstep=055000.ckpt + sha256: df4ce9aeced84fd96a39583c7f6e2c5773b76f942cb43012c6c51dd9eb806806 +- batch_size: 32 +- accumulate_grad_batches: 16 ++ batch_size: 64 ++ accumulate_grad_batches: 8 + 500m: + experiment: 500m + checkpoint: checkpoints/paper/500m/sstep=050000.ckpt + sha256: f13bc58f07f104fcff67e35e79650def0fec5a6eb62a193fa1ae9f979fb911cf +- batch_size: 16 +- accumulate_grad_batches: 32 ++ batch_size: 32 ++ accumulate_grad_batches: 16 +diff --git a/configs/sessions.yaml b/configs/sessions.yaml +index c73045e..fa5652a 100644 +--- a/configs/sessions.yaml ++++ b/configs/sessions.yaml +@@ -2,7 +2,7 @@ + code_root: upstream/hfd + raw_root: /root/autodl-fs/pangxinlei/hfd_data + cache_root: ${oc.env:HFD_DATA_DISK,/root/autodl-tmp}/cache/hfd +-output_dir: ${oc.env:HFD_DATA_DISK,/root/autodl-tmp}/datasets/hfd-sessions/${split} ++output_dir: ${oc.env:HFD_DATA_DISK,/root/autodl-tmp}/datasets/hfd-sessions-split20261026/${split} + split: train + limit_sources: 0 + count: null +@@ -14,3 +14,11 @@ candidate_pool: false + recipe_overrides: + command_interruption_session_variants: 2 + include_third_party: true ++ ++# Final outer split selects sessions; upstream split is used only for replay. ++final_release: ++ profile: isolated_benchmark_1024_20261026 ++ revision: bdca2fe7581c263e9ceb3a96204b314e199faddf ++ directory: ${oc.env:HFD_DATA_DISK,/root/autodl-tmp}/cache/hfd/releases/bdca2fe7581c263e9ceb3a96204b314e199faddf/final_session_delivery ++ selection_index: ${oc.env:HFD_DATA_DISK,/root/autodl-tmp}/cache/hfd/releases/bdca2fe7581c263e9ceb3a96204b314e199faddf/motion-selection.sqlite ++ replay_support_root: ${oc.env:HFD_DATA_DISK,/root/autodl-tmp}/cache/hfd/releases/bdca2fe7581c263e9ceb3a96204b314e199faddf/runtime +diff --git a/configs/textop.yaml b/configs/textop.yaml +index 2235dab..49be7df 100644 +--- a/configs/textop.yaml ++++ b/configs/textop.yaml +@@ -1,4 +1,4 @@ +-# English foreground only; original indices/seed preserved. No ASR or translation. ++# English foreground only; isolated final split and recorded per-session seeds. + # Full-data English subset; long-run step budget still requires calibration. + schema_version: 1 + mode: train +@@ -7,9 +7,9 @@ framework: textop_native + paths: + data_disk: ${oc.env:HFD_DATA_DISK,/root/autodl-tmp} + upstream_root: upstream/textop +- session_root: ${paths.data_disk}/datasets/hfd-sessions ++ session_root: ${paths.data_disk}/datasets/hfd-sessions-split20261026 + cache_root: ${paths.data_disk}/cache +- output_dir: ${paths.data_disk}/runs/hfd-textop-english-train ++ output_dir: ${paths.data_disk}/runs/hfd-textop-english-split20261026 + runtime: + python: ${paths.data_disk}/envs/textop/bin/python + num_gpus: 1 +@@ -19,6 +19,8 @@ runtime: + offline: true + environment: + HFD_LOCAL_INDEX_CACHE: '1' ++ HFD_NATIVE_STARTUP_CACHE: '1' ++ HFD_STARTUP_INDEX_CACHE_DIR: ${paths.cache_root}/hfd/startup-indexes + data: + train_sessions: + - ${paths.session_root}/train/*.npz +@@ -26,7 +28,7 @@ data: + - ${paths.session_root}/val/*.npz + window_policy: full_session + max_windows: null +- cache_sessions: 512 ++ cache_sessions: 2048 + asr_dir: ${paths.session_root}/reference + asr_latency_seconds: 0.0 + audio_policy: released_asr +@@ -36,7 +38,7 @@ data: + backend: native_ondemand + native: + config: configs/sessions.yaml +- memory_mib: 2048 ++ memory_mib: 8192 + reference_cache: ${paths.cache_root}/hfd/reference-alignments.sqlite + reference_guard_seconds: 0.1 + refresh_batches: 64 +@@ -51,7 +53,7 @@ training: + precision: 32-true + accumulate_grad_batches: 1 + gradient_clip_val: 1.0 +- save_every: 2000 ++ save_every: 256 + eval_every: 2000 + num_primitive: 4 + stages: +diff --git a/docs/ASR_CONTEXT_OVERFIT.md b/docs/ASR_CONTEXT_OVERFIT.md +deleted file mode 100644 +index ce25532..0000000 +--- a/docs/ASR_CONTEXT_OVERFIT.md ++++ /dev/null +@@ -1,73 +0,0 @@ +-# OMG / TextOp 完整历史 ASR 过拟合测试 +- +-2026-10-18。结果目录:`/root/autodl-tmp/runs/archive/legacy-20261020/hfd-asr-context-overfit`。 +- +-本实验复用 `hfd-final/videos/omg_full_session_playable.mp4` 对应的原生双任务 session、混合音频和因果 Whisper sidecar。从相同官方初始权重重新训练,完整监督监听、执行、收尾动作;不把 GT 指令、任务阶段或控制标签作为输入。原来的 `hfd-final` 文件没有被覆盖。 +- +-## 输入变化 +- +-新增 `data.caption_policy: causal_asr_context`:每个释放点使用本次 ASR 对最近最多 30 秒已释放音频的**完整识别结果**;非空立即更新,空结果沿用上一份非空文本。不拼接历次重识别结果,不等两次一致,不用未来识别结果回填过去。这里保留的是 ASR 当前假设,最后一句可能尚未识别完整。 +- +-旧 `stable_latest_sentence` 只保留最近一句,并要求连续两次识别一致。本实验同时改变了文本覆盖范围和确认等待规则,因此不能把结果变化单独归因于历史句子的数量。本轮最初通过独立实验配置选择新策略;随后用户确认将其设为五个 baseline 的默认文本策略,见 [共同文本约定](BASELINE_TEXT_CONDITION.md)。旧 run 配置保留用于历史复现。 +- +-两个模型使用同一字符串,各自经过原生编码器。TextOp 保留 CLIP 的 77-token 限制和已有的超长文本最近后缀策略;OMG 保留 T5 的 50-token 限制。**本 session 的全部 37 个释放点,两个模型都没有截断。** 第二条指令出现后,输入可同时包含 `Taking an item from someone. Picking an item up from the floor.`,没有使用 metadata 修正 ASR。 +- +-`input_audit.json` 保存逐释放点的完整文本、旧稳定句、实际编码文本/令牌、截断标志;还核对 session 哈希、每个 ASR 音频切片哈希、释放边界与前缀不变性。ASR 仍是预计算的因果 sidecar,不计入实际识别计算延迟。 +- +-## 保持的训练与推理设置 +- +-| 项目 | OMG | TextOp | +-| --- | --- | --- | +-| 训练初始化 | 原官方 50M checkpoint | 原官方预训练 checkpoint | +-| 训练 seed | 42 | 42 | +-| 更新次数 | 1600 | 3200,阶段 `[800,1600,800]` | +-| batch / learning rate | 8 / 6e-5 | 4 / 1e-4 | +-| 额外条件 | 过去 2 秒已释放音频,10 帧动作历史 | 2 帧动作历史;不输入 PCM | +-| 动作预测 | 60 帧,30Hz | 8 帧,50Hz;训练链 4 个 primitive | +-| 推理 | DDIM 10 步、joint_one CFG、原生 async、24 帧余量、测得规划延迟 | CFG 5,每轮执行 8 帧 | +- +-保持原生网络、损失、优化器框架和动作表示。训练数据窗口与旧 run 相同:OMG 366 个窗口、TextOp 88 条连续链。只有开场使用 GT 初始化,随后完全反馈已执行的预测历史。TextOp 旧 run 保存的 `audio_policy: offline_target_aligned` 是未使用的历史字段,本轮按当前校验器写为 `released_asr`;两次都只使用因果 ASR。完整配置差异见 `experiment.json`。 +- +-OMG 的异步交接使用本次测量的规划耗时,旧新耗时可能不同,故误差差异包含实际调度影响,不是固定交接时刻的纯文本消融。主种子平均规划耗时旧版约 0.157 秒、新版约 0.175 秒。 +- +-## 结果 +- +-主视频沿用 seed=12345,没有挑选最优种子。关节 RMSE 排除开场初始化;OMG 统计 29 关节、TextOp 统计原生 23 关节,跨模型绝对值不能直接排名。 +- +-| 模型 | 旧稳定句关节 RMSE | 完整历史 ASR 关节 RMSE | 旧 / 新根坐标 RMSE | +-| --- | ---: | ---: | ---: | +-| OMG | 0.06191 rad | 0.07455 rad | 0.00844 / 0.01052 m | +-| TextOp | 0.29922 rad | 0.17019 rad | 0.04596 / 0.01603 m | +- +-| 模型 / 文本策略 | seed 12345 | seed 12346 | seed 12347 | 三种子均值 | +-| --- | ---: | ---: | ---: | ---: | +-| OMG 旧稳定句 | 0.06191 | 0.15385 | 0.05854 | 0.09143 | +-| OMG 完整历史 | 0.07455 | 0.13126 | 0.08651 | 0.09744 | +-| TextOp 旧稳定句 | 0.29922 | 0.36330 | 0.28990 | 0.31747 | +-| TextOp 完整历史 | 0.17019 | 0.30671 | 0.32361 | 0.26684 | +- +-TextOp 主种子的深弯腰峰值从旧版约 11.66 秒提前至 9.74 秒,接近 GT 的约 9.58 秒;但第一段动作仍有时序偏差,三个种子中仍有一个变差。其固定 GT 历史短窗口 RMSE 为 0.00976 rad,不能代替闭环结果。**完整历史文本对本条 session 的 TextOp 有帮助迹象,尚未实现稳定过拟合;OMG 没有整体改善。** +- +-三种子仅改变推理采样,每种文本策略只有一次训练、同一条训练 session;这不是独立训练重复或泛化评测。也不能据此认定模型已理解“旧指令被新指令替代”的对话语义。 +- +-## 产物与核验 +- +-- `videos/omg_full_session_playable.mp4`、`videos/textop_full_session_playable.mp4`:主种子视频,预测与原始 G1 GT 对照,保留混合音轨;均为运动学回放,没有物理 tracker。 +-- `comparison.png`、`video_contact_sheet.jpg`:旧新轨迹及视频帧对照。 +-- `summary.json`:全部采样种子、根误差、关节误差、跳变和规划耗时。 +-- `input_audit.json`、`experiment.json`、`verification.json`、`code_sha256.json`:输入、配置、检查与代码记录。 +-- 两个 `pipeline_check.json` 均通过有限 loss/梯度、参数更新、预期优化步数、原生 checkpoint 严格重载检查。28 项相关单元测试和两个数据合同检查通过。两个主视频完整解码及音频检查通过。 +-- OMG 冗余 checkpoint 先逐 tensor 比对最终权重与 global_step,确认一致后删除,保留最终 `last.ckpt` 与全部日志;记录在 `omg/checkpoint_cleanup.json`。 +- +-## 复现 +- +-以下训练输出目录应尚不存在;重跑时先覆盖 `paths.output_dir` 为新目录。 +- +-```bash +-/root/autodl-tmp/envs/omg/bin/python train.py --config configs/history/omg_asr_context_overfit.yaml +-/root/autodl-tmp/envs/textop/bin/python train.py --config configs/history/textop_asr_context_overfit.yaml +-/root/autodl-tmp/envs/omg/bin/python infer_overfit.py --run /root/autodl-tmp/runs/archive/legacy-20261020/hfd-asr-context-overfit/omg --continuous-only --skip-window-eval --task-seconds 0 14.8 +-/root/autodl-tmp/envs/textop/bin/python infer_overfit.py --run /root/autodl-tmp/runs/archive/legacy-20261020/hfd-asr-context-overfit/textop --continuous-only --task-seconds 0 14.8 +-/root/autodl-tmp/envs/omg/bin/python render_overfit.py --inference /root/autodl-tmp/runs/archive/legacy-20261020/hfd-asr-context-overfit/omg/inference --assets-root /root/autodl-tmp/models/OMG/OMG/assets --gpu 1 +-/root/autodl-tmp/envs/omg/bin/python render_overfit.py --inference /root/autodl-tmp/runs/archive/legacy-20261020/hfd-asr-context-overfit/textop/inference --assets-root /root/autodl-tmp/models/OMG/OMG/assets --gpu 1 +-``` +diff --git a/docs/BASELINE_TEXT_CONDITION.md b/docs/BASELINE_TEXT_CONDITION.md +deleted file mode 100644 +index 8d819f7..0000000 +--- a/docs/BASELINE_TEXT_CONDITION.md ++++ /dev/null +@@ -1,63 +0,0 @@ +-# 五个 baseline 的共同文本条件 +- +-2026-10-21 更新:按用户决定,当前训练仅选英文前景 session,使用 `aligned_reference_context`、`clock_policy: microphone`、`text_conditioning: none`。停止对中文样本做训练和在线翻译。原始中文数据及历史报告保留,中文文本缓存已清理。完整输入/输出与实现状态见[当前五模型协议](HFD_HUMAN_BASELINE_INPUT_PROTOCOL.md)。 +- +-模型接收截至当前释放点、最近 30 秒的英文前景文字。文字来自原始 TTS 文本的离线词级对齐,不再经过 ASR。词结束后加实验性的 0.1 秒 guard,再向上取整到 `0.6+0.4k` 秒释放;该 guard 不是已证明的对齐误差上界。无效词跨度整段结束后释放,禁止提前展示完整指令。背景 PCM 原样保留,背景文字不作为输入。这是 oracle 前景转写实验。 +- +-原生动作时刻 `t` 对应麦克风 `t+0.2s`;只允许使用已经到达的文字和音频。文字不跨 session 累积,不输入控制标签。编码器 token 预算仍有效:OMG T5 50 token、TextOp CLIP 77 token(75正文+两个边界)、GenMo T5-3B 50 token,超长时保留最近后缀。前端 30 秒不等于全部文字都能进入编码器。文本过期后允许为空,不沿用旧ASR“空结果保留最后非空”的行为。 +- +-| Baseline | 文本与额外输入 | 当前接入范围 | +-| --- | --- | --- | +-| OMG | 原生 T5-Base;过去2秒已释放混合音频+10帧动作历史@30Hz | 动态英文训练、连续验证与视频 | +-| TextOp | 原生 CLIP ViT-B/32;2帧57D动作历史@50Hz,不输入音频 | 动态英文300K训练、连续验证与视频 | +-| GenMo / GEM-SMPL | 原生 T5-3B;60帧/2秒人体历史@30Hz,当前音频关闭 | 动态原生英文人体训练、验证、完整resume、retarget视频;100K配置 | +-| HY-Motion | 原生 Qwen3 与 CLIP-L | 共同契约;完整 HFD 训练入口尚未实现 | +-| Kimodo | 冻结 LLM2Vec / Llama-3-8B,最多400正文token,mean pooling为1×4096 | 共同契约;固定英文配对子集训练/闭环/GMR入口已接通,见[Kimodo训练](KIMODO_HFD_TRAINING.md) | +- +-OMG/TextOp 的正式默认配置与 reference 配置均设置 `data.native.language: en`。原生索引、session ID 和 seed 不重排,仅将中文候选的采样权重置零;第三方已组装 session 同样按前景语言过滤。剩余 train **168,583**、val **5,109** 个候选 session。共享声明见 `configs/english_text_protocol.yaml`,详细时间戳与动态组装见 [动态 reference 训练说明](REFERENCE_DYNAMIC_TRAINING.md)。 +- +-GenMo只使用配对人体的原生英文子集:train **155,117**、val **4,719**,排除尚无人体真值的第三方生产session。全量session没有提前落盘;离线缓存的是源语音词对齐,原生session在训练时动态组装并做有上限的内存缓存。三模型的数据范围与音频可见性不同,不能仅按step或误差绝对值直接比较架构优劣。 +- +-GenMo 新缓存示例: +- +-```bash +-python cache_genmo_text.py --asr SESSION.reference.json --output NEW_TEXT_CACHE.pt --caption-policy aligned_reference_context --text-conditioning none +-``` +- +-旧 ASR 对照配置归档在 `configs/history/omg_bilingual_asr.yaml` 和 `configs/history/textop_bilingual_asr.yaml`。历史策略 `causal_asr_context` 使用混合音频的因果 ASR 及冻结 OPUS 中文→英文转换,曾有动作语义误译。旧配置、权重和日志不改写,但重跑中文历史实验需重建已删除的中文缓存。历史依据见 [ASR 过拟合报告](ASR_CONTEXT_OVERFIT.md) 和 [语义时钟修复](SEMANTICS_CLOCK_FIX.md)。 +- +-实际误差、吞吐和清理记录见 [纯英文短测报告](ENGLISH_ONLY_PROBE.md)。 +- +-## 历史双语盘点(清理前) +- +-### 本地中英文文本配对盘点(2026-10-19) +- +-原始 `audio/seed_tts/metadata.jsonl` 共 142,220 行,包含 5,220 个原始音频目录及 5,220 对不同的中英文完整指令;按字符串去重后为中文 5,060 条、英文 5,220 条。中英字段缺失均为 0。同一指令会对应多份动作和多个声音版本,142,220 行不能当作独立句子数。 +- +-加入停止、附和、犹豫、修正及拆分的指令片段后,已对齐的前景资产包含 9,313 种不同中文字符串和 9,140 种不同英文字符串,中文音频 61,860 个、英文音频 61,860 个。这不是动态 session 组合文本的数量,也不包含对背景说话者做完整中英配对的承诺。停止短语有 12 种中文及 12 种英文,附和短语各 15 种。 +- +-完整动作指令都有英文对应字段,但不能据此保证全部等义或一一对应: +- +-- 141 个不同中文字符串对应多个英文字符串,有些仅为措辞变体,也有“单手/双手”“轻/重”等实质区别。例如“左手持小发光物向左转身”同时对应英文的 one hand 和 two hands 描述。该计数不是已确认错误条数。 +-- 已知 `retro`(复古)被译为“后退”的问题涉及原始 594 行、30 个音频目录。原生规则会排除对应的 120 个中文声线版本;实际选定第三方 recipe 中也未发现这些错误中文资产。完整资产对齐库仍保留这些原始资源,不能将其存在误认为都进入训练。 +-- 中英指令独立拆分后,prefix/suffix 不一定语义对齐。例如“检查全身”对应 checking whole body,但中文 prefix 是“检查”,英文 prefix 是 checking whole;不能直接逐片替换。 +- +-现有完整配对英文可以作为复用与质量审查的起点,但在中文只说出一部分时不能提前释放完整英文。前面“转写正确”指文字与合成语音所用原文一致,不代表已完成全部中英等义人工校验。数据报告:`/root/autodl-tmp/runs/archive/legacy-20261020/reference-full-preparation-20261019/bilingual-inventory-audit.json`。 +- +-## 官方语言支持核查(2026-10-19) +- +-核查对象为 OMG(Omni-Modal Motion Generation)、TextOp、NVlabs GENMO/GEM、HY-Motion-1.0 和 Kimodo。区分动作模型的语言验证、文本编码器的语言说明、以及前端翻译/改写能力。“未找到中文声明”不等于证明中文绝对不可用,也不等于已证明只支持英文。 +- +-| 模型 | 官方可核查的说明 | 对中文输入的结论 | +-| --- | --- | --- | +-| OMG | [论文附录 B.3](https://arxiv.org/html/2606.10340v1) 明确要求 action 和 style 标注写成英文;C.3 使用冻结 T5-Base、最多 50 token。 | 查阅的论文和官方 README 未列中文支持或中文动作评测。英文标注规范与本地中文 token 坍缩相符。 | +-| TextOp | [论文 III-B/III-C](https://arxiv.org/html/2602.07439v1) 使用 BABEL 动作文本标注和 CLIP;[CLIP 官方模型卡](https://github.com/openai/CLIP/blob/main/model-card.md#out-of-scope-use-cases) 说明未专门训练或评估英文以外语言,限定英文用途。 | TextOp 未给完整语言清单,未找到中文动作评测或中文支持承诺。不能把中文可分词视为已支持中文语义。 | +-| GenMo / GEM-SMPL | [论文附录](https://arxiv.org/html/2505.01425v1) 使用 T5 文本编码;[官方配置](https://github.com/NVlabs/GENMO/blob/main/configs/exp/gem_smpl.yaml) 指定 `t5-3b`。 | 未找到动作模型语言清单、中文支持声明或中文动作评测。英文示例和 T5 选型是使用证据,不是所有其他语言被排除的证明。 | +-| HY-Motion-1.0 | [官方模型卡](https://huggingface.co/tencent/HY-Motion-1.0) 标注 English/Chinese,但其 Prompting Guide 要求使用英文,其他语言由 Text2MotionPrompter 改写。[论文第 3.3 节](https://arxiv.org/html/2512.23464v1) 的改写模块训练输入包含中英混合请求。 | 官方支持链路包含中文用户输入,但通过前置改写生成英文动作描述。未找到绕过改写后,核心动作模型直接读中文的明确质量保证或专门评测。 | +-| Kimodo | [技术报告第 4.2 节](https://research.nvidia.com/labs/sil/projects/kimodo/assets/kimodo_tech_report.pdf) 使用 LLM2Vec;[官方编码器配置](https://github.com/nv-tlabs/kimodo/blob/main/kimodo/scripts/run_text_encoder_server.py) 指向 Llama-3-8B-Instruct 的 LLM2Vec 适配器,其[模型卡](https://huggingface.co/McGill-NLP/LLM2Vec-Meta-Llama-3-8B-Instruct-mntp-supervised) 标注 English。 | Kimodo 技术报告和 SMPLX-RP-v1 模型卡未提供动作提示词语言清单或中文动作评测。Llama 能处理某些中文,不构成动作模型中文能力的验证。 | +- +-HY-Motion 的官方实现直接要求改写输出为英文,见 [prompt engineering 的系统提示](https://github.com/Tencent-Hunyuan/HY-Motion-1.0/blob/master/hymotion/prompt_engineering/model_constants.py)。因此模型卡的中英文标签与“动作生成推荐英文”应一起阅读,不能仅依据标签宣称核心网络原生中英双语,也不能把整套系统说成完全不能接收中文。 +- +-另需澄清:[T5-Base](https://huggingface.co/google-t5/t5-base) 和 [T5-3B](https://huggingface.co/google-t5/t5-3b) 的官方 NLP 语言列表为英语、法语、罗马尼亚语、德语,没有中文。这也不表示采用 T5 的动作模型已验证了法语、德语或罗马尼亚语动作生成;编码器能力与下游训练覆盖应分别报告。 +- +-对当前计划的含义:使用共享英文文本前端符合已查到的主要官方使用路径,尤其是 HY-Motion 已明确建议其他语言先改写。共享前端仍须验证翻译质量;直接中文训练或更换多语编码器应单列实验,不能以“官方已支持中文”跳过检查。本轮没有修改训练或推理配置。 +diff --git a/docs/ENGLISH_ONLY_PROBE.md b/docs/ENGLISH_ONLY_PROBE.md +deleted file mode 100644 +index 62771a8..0000000 +--- a/docs/ENGLISH_ONLY_PROBE.md ++++ /dev/null +@@ -1,90 +0,0 @@ +-# 纯英文训练切换与短测(2026-10-19) +- +-后续性能更新:已启用本地只读索引镜像并通过原生训练对照;旧正式训练已停止并清理,详见 [最新性能报告](LOCAL_INDEX_TRAINING.md)。下文耗时属于优化前记录。 +- +-清理说明:本文记录的一次性实验脚本已归档,复现前请按[脚本恢复说明](WORKSPACE_CLEANUP.md)取回;历史指标与日志保留。 +- +-已将当前 OMG/TextOp 正式默认、reference、smoke 和完整 session 配置切换为英文前景、离线词级对齐、最近 30 秒上下文、不翻译。GenMo 单 session 配置和新建文本缓存默认策略同步;HY-Motion/Kimodo 记录共同约定,尚无完整 HFD 训练适配。本轮实际训练及误差、吞吐验证仅覆盖 OMG/TextOp。没有启动 100K/300K 正式长训练。 +- +-英文路线可完成训练和严格重载,但本轮没有证据证明它一致改善动作质量。OMG 相对旧双语短测的关节误差下降 4.4%,TextOp 上升 1.8%;OMG 仍未优于官方初始模型。数据准备仍是主要效率瓶颈。 +- +-## 数据与清理 +- +-原生 train 155,117、val 4,719;选定第三方 train 13,466、val 390,合计 **168,583 / 5,109**。原始全量索引与 session ID、seed 不重排,中文候选权重置零;六种交互类型均保留。native train 覆盖 4,245 种英文指令和 99,099 个动作文件 ID,这两个计数不包含第三方的额外去重统计。 +- +-删除中文资产对齐条目 61,860 个、中文孤立对齐记录 27 个、非英文全文 recipe 13,695 个、翻译缓存 10,692 条、含中文文本的旧 ASR 缓存 109,323 条,以及 63 份中文 reference/ASR sidecar,约回收 0.8 GB。英文资产内容摘要删除前后相同,当前对齐库只含 61,860 个英文资产,完整性检查通过。旧英文 ASR 条目 100,509 个作为历史缓存保留,当前训练不使用。 +- +-原始 tar、原始双语 metadata、日志、历史报告、模型及英文缓存保留。recipe 的 ASCII 语言标签保留以过滤中文 ID,中文全文 recipe 已删除。背景 PCM 不变,过滤的是前景目标语言。清理记录:[cache-cleanup.json](/root/autodl-tmp/runs/archive/legacy-20261020/english-only-20261019/cache-cleanup.json)。 +- +-文字是原始 TTS 文本的 oracle 前景转写,按实际动态 session 拼接位置平移词时间戳;不需要在线 ASR。释放时钟为 `0.6+0.4k` 秒、词结束后加实验性 0.1 秒 guard,保留最近 30 秒。英文中 6,608/61,860(10.68%)个资产词边界无效,整段结束后释放;该 guard 未证明能覆盖所有对齐误差。OMG 输入过去 2 秒音频、已释放文本及动作历史;TextOp 输入已释放文本及动作历史。两者均未加载 ASR 或翻译模型。 +- +-## 误差对照 +- +-两次新训练均从官方权重初始化。OMG batch64、512 步、lr6e-5、bf16;TextOp batch64、768 步、lr1e-4、FP32,三阶段各 256 步。对照旧双语试验采用相同训练步数、batch、学习率和 seed42;采样范围按实验目的改变为英文。不是在旧双语 checkpoint 上续训。 +- +-在同一组三个英文 val session(普通指令、打断、急停)上,对官方初始、旧双语、纯英文三个 checkpoint 做连续闭环回放。各 checkpoint 使用完全相同的原始音频、oracle 文本、释放时钟、一次 GT bootstrap、推理参数和 seed12345。OMG 固定规划延迟 4 帧、原生异步规划;TextOp 原生 CFG5、预测历史。计分排除 bootstrap;OMG 计 29 关节,TextOp 按其原生覆盖关节计分,不能横向比较两模型绝对值。 +- +-以下为三个 session 各自 RMSE 的等权平均,单位 rad: +- +-| 模型 | 官方初始 | 旧双语 | 纯英文 | 纯英文相对双语 | +-| --- | ---: | ---: | ---: | ---: | +-| OMG | 0.4043 | 0.4257 | 0.4070 | -4.4% | +-| TEXTOP | 0.4848 | 0.4117 | 0.4193 | +1.8% | +- +-根节点 xyz RMSE 的相同等权平均,单位 m: +- +-| 模型 | 官方初始 | 旧双语 | 纯英文 | +-| --- | ---: | ---: | ---: | +-| OMG | 0.3002 | 1.5859 | 0.7177 | +-| TEXTOP | 0.7033 | 0.5317 | 0.6196 | +- +-逐 session 的关节 RMSE: +- +-| 模型 | 交互 | 官方初始 | 旧双语 | 纯英文 | +-| --- | --- | ---: | ---: | ---: | +-| OMG | turn-taking | 0.3616 | 0.3794 | 0.3654 | +-| OMG | interruption | 0.3404 | 0.3319 | 0.3097 | +-| OMG | emergency-stop | 0.5109 | 0.5657 | 0.5459 | +-| TEXTOP | turn-taking | 0.3750 | 0.3913 | 0.3873 | +-| TEXTOP | interruption | 0.3582 | 0.3401 | 0.3461 | +-| TEXTOP | emergency-stop | 0.7211 | 0.5037 | 0.5244 | +- +-OMG 新版比旧双语更接近 GT,但官方初始关节误差仍略低,根节点误差也更低。TextOp 新版比双语略差,虽然优于官方初始。这只是三个样本、单一随机种子的短测,没有统计显著性保证;单一 GT 的运动 RMSE 也不是指令执行成功率,不能据此宣称全双工能力改善。 +- +-OMG 英文固定验证损失为 0.05414 → 0.05118 → 0.04272 → 0.04288(128/256/384/512 步)。TextOp 为 0.13086、0.13235、0.13669、0.14283、0.13936、0.12966(每 128 步);阶段历史分布在 256/512 步改变,不能把整段曲线当作同一目标的收敛曲线。旧双语 native validation 含中文,新版仅英文,二者验证损失不能直接比较优劣。 +- +-## 训练效率 +- +-OMG 固定 GPU1,TextOp 固定 GPU0。另从官方初始权重做 batch512 测速:OMG64步、TextOp96步,分别与 batch64 对照处理相同的 32,768 / 49,152 个训练实例;优化器更新次数不同,batch512 权重不用于上面的质量结论。TextOp 的“实例”按每次更新的 primitive 样本计,OMG 为动作窗,不能拿两模型的样本/秒直接比架构效率。 +- +-“池内”是排除超过 2 秒的冷池/验证间隔后的连续步间隔中位数,包含热数据准备和训练。“训练阶段”包含组池、验证和周期保存;“总墙钟”还包括索引/模型初始化和结束检查。batch512 复用了前一阶段的 CPU 原始索引,因此它的总墙钟不能与含冷索引构建的 batch64 直接当作正式加速比。 +- +-| 模型 | Batch | 步数 | 池内秒/步 | 池内实例/秒 | 训练阶段秒 | 训练阶段实例/秒 | 总墙钟秒 | +-| --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: | +-| OMG | 64 | 512 | 0.118 | 543 | 547.4 | 59.9 | 729.1 | +-| OMG | 512 | 64 | 0.535 | 957 | 385.6 | 85.0 | 397.9 | +-| TEXTOP | 64 | 768 | 0.083 | 775 | 300.9 | 163.3 | 478.8 | +-| TEXTOP | 512 | 96 | 0.149 | 3430 | 399.6 | 123.0 | 410.3 | +- +-OMG batch512 的首池准备 341.6 秒,TextOp 为 374.2 秒。训练阶段分别约 88.6% / 93.6% 时间用于该首池准备。batch64 OMG 8 池共 431.5 秒,TextOp 3 池共 169.4 秒。实际观察到共享盘 tar 读取等待,GPU 在组池时基本空闲。两秒采样、排除各卡训练后的推理后,包含组池/验证的平均 GPU 利用率仅约 GPU1 3.1%、GPU0 3.9%;采样从索引初始化后开始,不能当成全程精确 GPU 工时。 +- +-batch512 峰值 PyTorch 分配显存 OMG **16.21 GiB**、TextOp **2.62 GiB**,均可在当前卡运行。TextOp 的 batch512 短测没有覆盖完整 256 更新的数据池刷新周期,只用一个池跑了96步;batch64 跑了三个64-session池,故新鲜 session 数不同。大 batch 的池内吞吐高,但首池成本和数据复用程度必须一起看。 +- +-OMG 两阶段启用逐步训练健康检查;TextOp matched 为匹配旧试验关闭额外健康 hook,batch512 开启,因此不能把全部吞吐差异归因于 batch。历史双语 OMG 还多准备了一个未使用的尾池,当前已修复;其 1063.8 秒历史 fit 时间与本轮 547.4 秒的差异包含此独立修复及缓存状态差异,不是单独“去中文”的严格加速比。 +- +-当前瓶颈仍在数据准备。后续效率工作应优先测后台组池、原始资产/动作特征缓存及共享盘读取,再决定正式步数和刷新频率;仅增大 batch 或把旧 100K/300K 步原样启动,不能保证合理训练时长或动作质量。 +- +-## 验证与产物 +- +-60 项相关单元测试、两套正式英文配置 dry-run 通过。OMG matched/batch512 分别记录到 512/64 次真实优化器更新,有限损失/梯度、参数变化和严格重载通过;TextOp batch512 记录96次真实更新并通过同样检查。TextOp matched 记录768步有限损失/梯度范数,108个 denoiser 状态张量变化且有限,实际回放完成严格加载;该阶段没有单独 hook 计数优化器。没有训练结束后的多余冷池。 +- +-- [完整指标与逐样本误差](/root/autodl-tmp/runs/archive/legacy-20261020/english-only-20261019/summary.json) +-- [英文验证曲线](/root/autodl-tmp/runs/archive/legacy-20261020/english-only-20261019/english-validation.png) +-- [GPU 利用率](/root/autodl-tmp/runs/archive/legacy-20261020/english-only-20261019/gpu-utilization.png) +-- [最终数据和前端检查](/root/autodl-tmp/runs/archive/legacy-20261020/english-only-20261019/final-audit.json) +-- [OMG 急停视频](/root/autodl-tmp/runs/archive/legacy-20261020/english-only-20261019/evaluation/omg/emergency-stop/english/inference/complete_task.mp4) +-- [TextOp 急停视频](/root/autodl-tmp/runs/archive/legacy-20261020/english-only-20261019/evaluation/textop/emergency-stop/english/inference/complete_task.mp4) +- +-视频保留原始混合音频,显示预测与 GT、麦克风时钟和实际释放的文本;两份均已完整解码检查。OMG/TextOp 直接输出原生 G1,无需 SMPL-X retarget。GenMo 等人体模型仍须遵循现有 SMPL-X→G1 retarget 与视频评测链路,本轮没有宣称已完成它们的新训练。 +- +-运行配置、日志及 checkpoint 位于 `/root/autodl-tmp/runs/archive/legacy-20261020/english-only-20261019/`。脚本为 `tools/run_english_training_probe.py`、`tools/compare_english_checkpoints.py`、`tools/summarize_english_probe.py`。已更新正式默认配置,但正式长训练未启动;100K/300K 仍属于待校准预算。 +diff --git a/docs/FORMAL_TRAINING_READINESS.md b/docs/FORMAL_TRAINING_READINESS.md +deleted file mode 100644 +index 426a887..0000000 +--- a/docs/FORMAL_TRAINING_READINESS.md ++++ /dev/null +@@ -1,117 +0,0 @@ +-# 正式训练准备检查 +- +-后续性能更新:已启用本地只读索引镜像并通过原生训练对照;旧正式训练已停止并清理,详见 [最新性能报告](LOCAL_INDEX_TRAINING.md)。下文耗时属于优化前记录。 +- +-2026-10-19 英文路线更新:正式默认与 reference 配置均仅采样英文前景 session,离线对齐、无 ASR/翻译;train 168,583、val 5,109。中文缓存已清理。当前有步数上限的双卡误差与吞吐短测见 [纯英文短测报告](ENGLISH_ONLY_PROBE.md),100K/300K 正式训练未启动,旧预算仍需校准。 +- +-2026-10-19 较早更新:离线前景转写对齐和动态 session 读取已接入独立 reference 配置,详见 [离线对齐与动态训练](REFERENCE_DYNAMIC_TRAINING.md)。当前是一卡 OMG、一卡 TextOp;下文的双卡 OMG、TextOp batch64 和 ASR 预算属于历史记录。中文翻译质量、TextOp 连续回放和新版训练预算仍需验证,不能直接沿用旧文中的启动结论。 +- +-2026-10-17:OMG 已完成一轮 100k 更新,训练循环 108.55 小时。实际结果、数据范围、性能瓶颈和新配置见 [正式训练复查](OMG_TRAINING_REVIEW.md)。当前增加 `data.batched_motion=true`、`training.fixed_validation=true` 和 `training.save_best=true`;早停默认关闭,可用 `training.early_stopping.enabled=true` 启用。新增最佳 checkpoint 约多占 1 GiB。下面 10 月 10 日的短测和资源规划为历史准备记录,不应当作此次完整训练的实测速度。 +- +-同日后续:OMG 默认 `data.prefetch_batches=1`。双卡 64 步同步/异步分段对照,排除边界两步后,池内平均每步由 1.07 秒降到 0.62 秒。该数值不包含下一池刷新和 ASR,也不是整轮加速承诺。主机内存预算需另留每卡一个 CPU batch;`num_workers=0` 保持不变。 +- +-2026-10-10 更新:419 个原始 tar 已全部下载且 SHA256 校验通过;正式配置改为 OMG 双卡、TextOP 单卡,每次只运行一个模型。全量原生 session 配方已通过索引构造,原有两模型 batch=2 的短测已通过。OMG 每卡 batch 512 的显存探针已通过;低吞吐正式运行已按用户要求终止并删除,当前仅进行有明确步数上限的性能短测。 +- +-## 已确认的配方修正 +- +-用户确认将 `configs/sessions.yaml` 中的 `command_interruption_session_variants` 从 1 调到 2。原值在当前数据版本只能产生 36,836 个打断候选,低于原生配比要求的 39,743;修改后有 73,672 个候选,原生生成、筛选和最长 session 剔除流程均通过。 +- +-不复制 session,不跳过最终配比,不改变原生任务类型目标。增加 variant 可能改变后续随机组装结果,应把这个 override 同步记录到论文 HFD 和 baseline 数据配方中。原始 recipe 源码保持不变,覆盖参数与源码哈希保存在检查报告中。 +- +-| 类型 | train | val | +-| --- | ---: | ---: | +-| turn-taking | 158,812 | 4,539 | +-| interruption | 39,703 | 1,134 | +-| emergency-stop | 26,468 | 756 | +-| backchannel | 39,703 | 1,134 | +-| hesitation | 34,071 | 1,016 | +-| pause | 23,486 | 693 | +-| 合计 | 322,243 | 9,272 | +- +-这些是原生筛选及每类最长 0.1% 剔除后的计数。interrupt/control/backchannel 的目标按基础任务 session 数定义;hesitation/pause 按窗口数加入,因此不能直接拿最终六类 session 百分比验证全部配比。train/val 沿用原生动作来源分组划分。原生长度估计对应约 3,304.3/94.6 小时,实际导出长度会随 idle、打断时刻等而变化。 +- +-## 全量数据准备的实际限制 +- +-- 检查开始时,两张 RTX 5090 32 GB 空闲;TextOP、OMG、CLIP、T5、Whisper 权重均已在本机。 +-- 原始分片共 419 个,完整 tar 约 49.8 GiB,另有索引和 metadata。2026-10-09 已有 419 个完整分片,大小和 SHA256 均通过清单校验,无需继续下载。原始目录保持在共享盘。后续使用 `tools/download_hfd_tars.py` 校验 SHA256、跳过完整文件并续传 `.partial`;命令及空间要求见 [迁移说明](MIGRATION.md)。实际 hesitation 样本仍读取 v1 分片中的资源,不能仅凭配方名为 v2 就忽略这些文件。 +-- 按原生估计长度,在每个交互类型中选一个中位附近 session,六个均完成导出、`validate_sessions.py` 检查和完整因果 ASR;这些仅用于容量与吞吐测量,没有进入训练。 +-- 六个样本的压缩 NPZ 约为每音频秒 37,765 字节。乘以原生总估计时长,train+val 约 **430 GiB**,还不含 JSON、ASR 和 checkpoint。此为旧完整导出方案的六样本推算,当前按需方案无需分配这部分空间;收尾时数据盘剩余约 44 GiB。 +-- 逐窗口 ASR 在五个热身后样本上约 0.048–0.091 秒/窗口。六个样本共 483 个因果窗口,逐条总耗时 57.9 秒,batch=32 总耗时 7.45 秒(约 7.8 倍);两个测量的热身条件并不完全一致,不能作为严格性能基准。16/483 个窗口的文本有差异,PCM 和截止时刻一致。保持混合音频,不用 GT 替代 ASR。 +-- 旧 NPZ 加载器只有 4-session 缓存且 OMG 保留所有周期 checkpoint。当前正式配置改用按需读取、OMG 512-session 特征缓存、每个 split 2 GiB 数组预算和最近 3 个周期 checkpoint,旧方案只保留用于已有小样本复现。 +- +-已选择按需组装方案:`configs/textop.yaml` 和 `configs/omg.yaml` 使用 `data.backend: native_ondemand`。每个 batch 槽位按合法动作窗口的长度加权抽 session,再从 session 内均匀取窗口;OMG 当前每卡 batch 为 512,正式值每 256 个 batch 刷新一次 session 池。相邻 batch 有相关性,但保持合法窗口的边际采样概率;每个 batch 仍会在已保留 session 内重新抽动作点。较长刷新间隔用于摊薄原生组装和 ASR 成本,若需更强 session 随机性可改回 64。OMG 缓存覆盖一个完整槽位池;批量音频特征计算保留逐窗口 35 维特征定义,独立对照容差为 rtol=atol=1e-6(存在浮点运算顺序差异)。原生配比定义的是候选 session 集合,不保证每个 batch 都满足该比例。增强使用固定 seed=42,缓存失效后重组同一 session 不改变其内容,也不每个 epoch 重新随机增强。ASR 结果以压缩 SQLite 缓存,Whisper 使用 batch=32 的因果窗口;批处理参数和实现哈希记录在 run manifest,实际识别结果位于缓存。它与逐条识别不是字节级等价,做论文复现时应一并保存匹配的缓存。 +- +-OMG 和 TextOP 的按需 smoke 都已通过(各 4 次更新、有限梯度/参数变化、验证读取、严格 checkpoint 加载)。六个样本的动作/PCM 与旧导出逐元素一致。少数窗口文字因 batch/半精度数值产生差异,这是已记录的预处理口径变化,不是未来音频或 GT 输入。 +- +-按需训练不生成数百 GiB 的派生 NPZ;当前数据盘只存 raw-range cache、ASR SQLite、日志和模型输出。建议先补齐原始 tar,再长时间训练;当前代码支持缺失文件的 range 回退,但受网络可用性与重试影响,不能按本地数据速度估算。`runtime.offline: true` 控制模型权重加载,不会关闭这个 raw tar 网络回退。 +- +-## 正式训练命令和资源预算 +- +-2026-10-09 配置改为**每次只训练一个模型**:OMG 使用 GPU 0、1 的 DDP;TextOP 沿用原生单卡训练器。OMG 的按需加载器在每个 rank 使用独立窗口采样随机流、对应 GPU 的 Whisper,以及分别命名的采样记录;session 构造种子保持一致。两个 rank 从完整原生集合独立有放回抽样,允许偶然抽到同一个 session,不是两卡复制同一个 batch。每个 rank 保留完整索引,模型梯度由 DDP 同步。 +- +-先在仓库目录检查配置(不启动训练): +- +-```bash +-cd /root/autodl-tmp/hfd-baseline-portable +-export HFD_DATA_DISK=/root/autodl-tmp +-/root/autodl-tmp/envs/omg/bin/python train.py --config configs/omg.yaml --dry-run +-/root/autodl-tmp/envs/textop/bin/python train.py --config configs/textop.yaml --dry-run +-``` +- +-在 screen 中运行 OMG: +- +-```bash +-screen -S hfd-omg +-cd /root/autodl-tmp/hfd-baseline-portable +-/root/autodl-tmp/envs/omg/bin/python -u train.py --config configs/omg.yaml +-``` +- +-OMG 结束后,另开一个 screen 运行 TextOP: +- +-```bash +-screen -S hfd-textop +-cd /root/autodl-tmp/hfd-baseline-portable +-/root/autodl-tmp/envs/textop/bin/python -u train.py --config configs/textop.yaml +-``` +- +-已经在 screen 内时只执行 `cd` 和 Python 命令,不嵌套创建 screen。`Ctrl+A` 然后 `D` 脱离,`screen -r hfd-omg` 或 `screen -r hfd-textop` 恢复。不要同时执行这两个训练命令。终端直接显示数据准备和训练进度,无需 `nohup`、`tee` 或输出重定向;run 中仍保留配置、checkpoint 和小型结构化指标,供复查实验。 +- +-| 项目 | TextOP | OMG | +-| --- | --- | --- | +-| GPU | 1 张 RTX 5090 32 GB,GPU 0 | 2 张 RTX 5090 32 GB,GPU 0、1 | +-| batch | 64 | 每卡 512,全局 1024 | +-| 精度 | FP32 | bf16 mixed | +-| 更新步数 | 300,000;三阶段各 100,000 步 | 100,000 | +-| 学习率 | 1e-4 | 6e-5 | +-| 数据读取 | 按需原生组装,完整 train/val | 同左,DDP 独立采样 | +-| ASR | Whisper tiny,batch 32,共用训练卡 | 每个 rank 在对应训练卡使用 Whisper tiny,batch 32 | +-| CPU | 4 线程,num_workers=0 | 每 rank 4 线程,num_workers=0;音频特征按 batch 向量化 | +-| 输出目录 | `/root/autodl-tmp/runs/hfd-textop-train` | `/root/autodl-tmp/runs/archive/legacy-20261020/hfd-omg-train` | +-| 保存策略 | 每 20,000 步,保留最新 3 份 | 每 2,000 步,保留最新 3 份 | +- +-输出目录必须尚不存在;启动新实验可加 `--set paths.output_dir=/root/autodl-tmp/runs/新的名称`。OMG 入口自动调用 torchrun,无需手工加启动器。默认沿用本机可用的 Gloo,DDP 超时 1,800 秒,给按需组装和 ASR 留余量。OMG 的 `training.batch_size` 是每卡 batch;当前每卡 512、总 batch 1024。真实 forward/backward 探针峰值约为 21.9 GiB/卡(另有 Whisper 常驻);没有把 640 作为正式值,因为 640 已达到约 27.3 GiB,给优化器和 DDP 临时张量的余量较小。改回 OMG 单卡时加 `--set runtime.num_gpus=1 'runtime.gpu_ids=[0]' training.batch_size=512`;保持学习率 6e-5 以便与 baseline 口径一致。 +- +-两模型都对完整 session 的动作窗口进行监督,保留 listen/idle 对应关节动作,不输入控制标签。OMG 使用过去 2 秒音频+因果 ASR;TextOP 使用因果 ASR。此次双卡改动不改变上述输入或数据配方。 +- +-训练是按优化更新步数停止的随机窗口采样,并不是逐个遍历全部 session 的 epoch:TextOP 每次取一个 batch 后依次对 4 个 primitive 做 4 次更新,共取约 75,000 组 batch;OMG 共取 100,000 组双卡 batch。OMG 槽位每 256 个 batch 刷新一次,TextOP 每 64 个 batch 刷新;全量集合可被抽样,但有限步数不保证每个 session 都被访问。论文应记录这一训练预算。 +- +-资源规划:OMG 两进程建议预留 8 CPU 核、32 GiB 主机 RAM;TextOP 建议至少 4 核、16 GiB RAM。OMG 每个 rank 的 train/val 各有 2 GiB 数组 LRU,另有 512-session 特征缓存和完整索引;2 GiB 不是整个进程内存上限。仅数据回放进程的 RSS 约 3.9 GiB;这不包含训练模型、Whisper、验证集与优化器,不能把它当作完整训练的 RAM 占用。ASR SQLite 共用 2 GiB 有效记录预算;原始 tar 齐全时无需网络 range 回退。无需生成约 430 GiB 的全量派生 NPZ。 +- +-数据盘在已有权重和环境之外建议预留 15–20 GiB。TextOP 输出约 1.2 GiB;OMG 最新 3 个周期 checkpoint 加最终 last.ckpt 约 4 GiB,多卡不会各保存一套 checkpoint。其余用于缓存、SQLite 页/WAL 和保存瞬间空间。当前约 44 GiB 空闲足够这份规划。原始 tar 约 49.8 GiB+索引保存在指定共享盘,已下载完成;共享盘 df 的整体空闲仍不能视作账号配额。 +- +-本轮已终止的 OMG batch=512 实跑暴露出旧配置的问题:64-session 缓存无法覆盖一个 512-session 槽位池,普通 batch 也会重复构造 session 及运动特征;每 64 个 batch 重新抽 session 时还需要处理 ASR,导致运行近 10.5 小时只完成约 500 步。配置已改为 512-session 缓存、每 256 个 batch 刷新一次。此前 session 回放基准(ASR 命中磁盘缓存)显示,缓存 512、数组预算 2 GiB 时首次准备约 50 秒;另一轮文件读取较冷的回放约 188 秒。后者包含不同缓存状态,不能把首次准备时间差全归因于代码优化。热缓存下 batch 准备约 1.7 秒;同样缓存容量、旧逐窗口音频实现的整个 batch 准备约 7.5 秒,改为 batch 向量化后降至约 1.7 秒。双卡真实 train mode 短测中,模型计算约 0.43 秒/步,缓存命中时完整 optimizer step 约 2.0 秒,首次装载一个 512-session 池约 72 秒;这些测试不包含完整 refresh 周期。按 100,000 步估算,纯缓存命中下约 56 小时,实际时间还取决于每 256 步的刷新、ASR 命中率和共享盘读取速度。正式重启前仍应做 3–5 个 optimizer step 的端到端短测,再估计总耗时。 +- +-OMG 每 2,000 步每 rank 抽取 16 个验证 batch(双卡共 32 个),TextOP 每 2,000 步运行原生 eval_steps=12(3 组四 primitive batch)。这是训练监测,不是全量独立测试集评测。统一入口目前仍未提供恢复优化器及步数的 resume;`weights.init_checkpoint` 仅表示初始化。断点恢复和按需 run 的论文评测入口是后续工程项,不能把重新加载初始化权重当作续训。 +- +-## 可复查的证据 +- +-小型 JSON 结果在 [formal_training_readiness.json](../reports/formal_training_readiness.json)。本机测量样本和日志位于 `/root/autodl-tmp/cache/hfd/readiness/`,不进入 Git,也不混入 `runs/hfd-final`。 +- +-重建索引检查(不导出 session、不启动训练): +- +-```bash +-/root/autodl-tmp/envs/omg/bin/python prepare_sessions.py --config configs/sessions.yaml --audit-only --report /root/autodl-tmp/cache/hfd/train_audit.json +-/root/autodl-tmp/envs/omg/bin/python prepare_sessions.py --config configs/sessions.yaml --audit-only --report /root/autodl-tmp/cache/hfd/val_audit.json --set split=val +-``` +- +-迁移时替换 Python、数据盘和原始 tar 路径;详见 [迁移说明](MIGRATION.md)。`--audit-only` 通过只说明索引构造与配比可行,不保证全部 tar、ASR 或训练就绪。全量 tar 已校验、独立 train/val 索引已构造;本轮正式 batch 的短测结果见本文开头及报告。长时吞吐和论文独立测试集评测须另行衡量。 +diff --git a/docs/GENMO_FORMAL_TRAINING.md b/docs/GENMO_FORMAL_TRAINING.md +deleted file mode 100644 +index d2b273f..0000000 +--- a/docs/GENMO_FORMAL_TRAINING.md ++++ /dev/null +@@ -1,150 +0,0 @@ +-> 本文保留旧正式配方的记录。当前默认训练入口采用 [GenMo v2](GENMO_V2_TRAINING.md),包括执行段损失、预测历史恢复、CPU预取及闭环选权重;旧配置见 `configs/genmo_legacy.yaml`。 +- +-# GenMo 原生英文 HFD 训练 +- +-2026-10-24:本轮100K正式训练已完成,耗时约85.8小时。best为92K,已对官方初始化、92K和最终100K按既有六类验证规则完成闭环、整段GMR及有声视频评测。当前六段结果以92K略好,后退/移动任务仍明显不足;完整数值与视频位置见[正式评测报告](HUMAN_FORMAL_VALIDATION_20261024.md)。以下命令和训练约定保留;新正式权重未修改或删除。 +- +-正式入口为 `train_genmo_full.py --config configs/genmo.yaml`。旧 `train_genmo.py` +-和 `run_genmo_overfit.py` 仍是单session检查入口,不用于全数据训练。 +-沿用锁定版本的原生 GEM `training_step`、regression→diffusion 顺序、AdamW、151D表示、 +-冻结预训练统计量及原生3D损失;适配层保留此前验证的坐标/位移逆变换、历史锁定和未来评分mask。 +- +-## 数据范围与输入 +- +-只使用原生英文 HFD:155,117条train候选session、4,719条val候选session。 +-第三方生产包目前只有G1时间线,没有经过验证的配对SMPL-X session,明确排除。 +-因此不能把这次训练与包含第三方数据的OMG/TextOp结果直接作为同数据量架构对比。 +- +-每次按相同索引、原生seed组装音频和G1轨迹,同时观察同次执行的源选择、裁剪、 +-持姿、对齐、过渡和拼接,依据权威配对表构造人体轨迹。源SMPL-X SHA256逐条核验, +-所有原生记录及stationary源的配对ID在首次使用时检查;未支持的操作或缺失源直接失败。 +-全双工人体session不是从G1反解出来的,也不要求预先把整套session解压存盘。 +- +-每条已组装人体session的30Hz重采样只计算一次并受CPU缓存容量限制。决策点同时位于 +-50Hz和30Hz网格;历史切片的插值支撑均不晚于决策时刻,末端无效目标仍保持原50Hz终帧。 +-优化前后实际batch32的前4步loss逐值一致;未来动作扰动和末端mask回归检查通过。 +- +-输入为已知/已执行的2秒人体动作历史,以及当前已释放的最近30秒英文GT前景文本。 +-词时间取现有离线对齐缓存,guard=.1秒,0.6+.4k秒释放,motion到microphone偏移.2秒。 +-冻结T5-3B以BF16编码,原生50token预算超长时保留最近token;空文本特征为零。 +-编码器常驻当前GPU,CPU特征缓存有256MiB上限。无需ASR、翻译或重新跑强制对齐。 +-本版本保留已采用的文本+动作历史条件,不混入独立的past-audio实验。 +- +-输出为未来2秒人体动作,30Hz,每次执行12帧(.4秒)再规划。不输入GT阶段、 +-listen/motion/idle标签或未来音频。训练窗口可有初始重复姿态和末端无效目标, +-已知历史不计入未来几何损失;未知边界位移保留监督。 +- +-## 采样、验证、续训 +- +-在原生session的合法.4秒决策窗口上带放回抽样,使用保守上界拒绝采样维持窗口边际。 +-一个槽位绑定一条session,默认16批后刷新槽位,每批重新抽该session中的决策点。 +-训练按step预算运行,epoch不表示无重复遍历全部session。 +- +-每2K步进行固定验证:原生GEM双模式目标、固定窗口和噪声,无梯度;网络保持eval, +-只切换diffusion调度器的训练目标标志,防止原生图像验证入口不兼容或BN统计更新。 +-该loss不是连续推理或任务成功率,应同时使用下面的retarget视频评测。 +-默认每次16批×16窗口,验证每批刷新session,使用固定的256个窗口;这仍是验证子集, +-不是遍历4,719条候选session。训练池的16批复用策略不用于验证。 +-GEM的时间步采样使用NumPy,故验证同时保存、固定并恢复Python/NumPy/Torch随机数, +-不能仅固定Torch高斯噪声。同一真实checkpoint重复验证两次loss均为7.5329627991, +-仅用于验证复现性,不代表正式训练效果。 +- +-checkpoint保存模型、AdamW、按step的cosine调度器、RNG和已消费的采样池状态。 +-先写临时文件再原子替换,保留最近两份和最佳验证权重。完整checkpoint单份约5.1GiB, +-默认放共享盘 `models/genmo/hfd-english-train`,日志放本地 `runs/hfd-genmo-english-train`。 +-常态约需3份权重空间,保存时还需要临时文件空间;共享盘写入错误会中止并保留原checkpoint。 +- +-续训核对输入配置、模型适配实现、原生配方、离线文本缓存、配对表和body model身份。 +-只允许在原总步数预算内续训;max_steps不是“再训练多少步”,不能直接修改总预算或学习率后冒充恢复。 +-可更换GPU,允许把同一checkpoint分支到新的空输出目录。 +- +-```bash +-cd /root/autodl-tmp/hfd-baseline-portable +-/root/autodl-tmp/envs/genmo/bin/python -u train_genmo_full.py --config configs/genmo.yaml +- +-# 中断后恢复完整状态 +-/root/autodl-tmp/envs/genmo/bin/python -u train_genmo_full.py --config configs/genmo.yaml --resume auto +-``` +- +-当前总预算100K步,20K/40K/60K为阶段检查点,不表示已验证必须训练到100K才收敛。 +-学习率2e-5起步,按全部100K步做cosine衰减,在100K降到1e-6;`T_max`直接读取 +-`training.max_steps`,20K不再是学习率衰减终点。BF16、GPU1、batch32,合计320万次窗口采样。 +-采样仍有放回、槽位每16批刷新,不保证每个session都出现,不把100K解释为固定epoch数。 +-每2K步固定验证和保存,保留最近两份及最佳验证checkpoint;阶段检查应及时运行连续评测, +-20K/40K/60K权重不会永久保留。评测需单独使用空闲GPU或在训练暂停后进行,避免抢占显存。 +-验证或rollout停止改善时可提前停止,并从最佳checkpoint评测;当前不自动按val/loss早停, +-因为它不是语义成功率。100K是总上限,不是达到20K后再额外训练100K。 +-若中断,用同一100K配置恢复;旧20K协议checkpoint不能通过修改YAML直接兼容续训。 +-本次预算调整时正式输出目录尚未创建、权重目录为空,无旧正式训练需要迁移。 +- +-实测含T5-3B的GPU峰值约23.7GiB。热批次数据准备约4.9秒降到0.05–0.07秒, +-计算约0.4秒;换session池仍有组装开销,不能把热批次加速当作整轮训练的加速倍数。 +-不会自动开始长训练。 +- +-## 连续推理、retarget与视频 +- +-```bash +-/root/autodl-tmp/envs/genmo/bin/python -u evaluate_genmo.py \ +- --run /root/autodl-tmp/runs/hfd-genmo-english-train \ +- --output /root/autodl-tmp/runs/genmo-english-validation \ +- --gpu 1 +-``` +- +-默认读取最佳完整checkpoint,六类原生英文val各选首条6–30秒样本,可用 `--checkpoint PATH` +-和 `--modes turn-taking interruption ...` 显式指定。样本选择在推理前,使用与训练相同的 +-配对provider与文本对齐;只初始化一次GT历史,随后反馈预测,不按窗口重置GT。 +-推理中的指令显示时间读取实际session元数据;GT文本出处传递到retarget与视频,避免误标ASR。 +- +-每条完整预测和完整人体GT各retarget一次,保留三种比较:预测→G1/原G1、人体GT→G1/原G1、 +-预测→G1/人体GT→G1。retarget误差是独立诊断,不能从模型误差直接相减。 +-视频保留原混合音轨与共同microphone前缀。默认生成三条视频,属于运动学回放,不是物理跟踪。 +-原生task源按train/val隔离;val用于训练监测,不是独立test,预训练数据重合未审计。 +- +-## 2026-10-20 训练链路复核 +- +-可复用检查入口: +- +-```bash +-/root/autodl-tmp/envs/genmo/bin/python -u tools/audit_genmo_training.py \ +- --config configs/genmo.yaml --gpu 1 \ +- --output /root/autodl-tmp/setup/genmo-training-audit-20261020/audit.json +-``` +- +-此次检查对六类交互各取首条合格train session,共核对483个决策点,其中465个非空文本。 +-逐点重建文本并核对词末时间、guard和离散释放时间,没有违反标注释放规则; +-相对强制对齐词末的披露延迟为0.110875–0.48秒,一处不合格对齐采用整段结束后释放。 +-强制对齐是事后估计,0.1秒guard不是人工验证的误差上界;这仍是oracle文本实验。 +-六条样本最长29个T5 token,未发生截断,不代表全量语料截断率为零。 +- +-实测改变未来GT位置和姿态,regression预测逐值不变;清空文本后未来预测发生变化 +-(标准化151D特征的平均绝对变化0.05681275;不是语义成功率)。 +-regression和diffusion分别反向传播,文本投影、动作输入投影、主干和相关输出头均有 +-有限非零梯度;固定历史的原始姿态输出梯度为零,未知边界位移保留梯度。 +-执行一次用于检查的diffusion优化器更新后,文本/动作输入投影权重及bias均变化; +-T5参数保持冻结且无梯度。检查不保存模型权重,不代表收敛或闭环质量。 +-配套44项回归检查通过;报告位于上述output路径,日志在同级目录的 +-`genmo-training-audit-20261020.log`。 +- +-此次修复:位移目标i表示i→i+1,原生编码会复制窗口末段位移。当前屏蔽不存在下一有效帧 +-的位移和静止标签,包括窗口末帧及session尾部填充边界,保留有效姿态和历史→未来边界 +-位移监督。该修改改变训练目标mask和实现身份,旧实验loss不可直接逐值对比,旧完整状态 +-不能伪装为新协议resume。此次修改前未启动GenMo正式训练。 +- +-评测新增retarget之前的人体根朝向、根速度误差,与已有身体关节旋转及根位置误差一起报告; +-`evaluate_human.py`的`human_metrics`及`evaluate_genmo.py`摘要都会保留这些结果。 +-它们不依赖原G1轨迹;retarget后三种比较仍用于单独诊断。误差均不做根对齐或DTW, +-排除GT bootstrap。速度RMSE按三个坐标分量求均方根,单位m/s。 +- +-本设计适合“已释放文本+动作历史”的条件生成对照,存在以下能力/衡量限制: +- +-- 文本释放前没有语音起点信息,不能要求模型准确判断倾听开始;语调、韵律和背景声也不可见。 +-- 最近30秒是文本候选范围,实际最多50个T5 token;没有显式词时间或当前说话状态输入。 +-- 训练历史为GT,闭环历史为预测;未来2秒内的新指令尚不可见,目标天然存在不确定性。 +-- 输出预测2秒,每0.4秒重新规划执行;原生位置损失使用GT朝向积分,实际推理使用预测朝向, +- 因而必须同时看根朝向和连续rollout,不能只依据val/loss判定部署误差。 +-- 原生loss沿用全窗口均值乘mask,尾部有效目标少时loss会偏小;固定验证的窗口/噪声可比, +- 该loss不能直接当成“每个有效未来帧”的误差或跨模型通用分数。 +-- 现有指标衡量运动学轨迹一致性,尚不衡量独立定义的指令成功率、停止延迟或物理跟踪稳定性。 +- +-当前结论是结构与梯度检查通过,可开展该文本条件baseline训练;输入信息不足造成的反应差异 +-不能简单归为训练未收敛。按既定设置保留文本条件,音频条件仍作为另行定义的实验。 +diff --git a/docs/GENMO_HFD_OVERFIT.md b/docs/GENMO_HFD_OVERFIT.md +deleted file mode 100644 +index 805e6b3..0000000 +--- a/docs/GENMO_HFD_OVERFIT.md ++++ /dev/null +@@ -1,105 +0,0 @@ +-# GenMo / GEM-SMPL 的 HFD 数据适配与过拟合实验 +- +-清理说明:本文记录的一次性实验脚本已归档,复现前请按[脚本恢复说明](WORKSPACE_CLEANUP.md)取回;历史指标与日志保留。 +- +-新增原生过去音频分支的匹配实验见 [GENMO_PAST_AUDIO_TEST.md](GENMO_PAST_AUDIO_TEST.md)。音频输入已接通,当前小样本表现未优于 ASR-only,未替换本文工作版本。 +- +-本轮使用官方 `NVlabs/GENMO` commit `16bebf402d8893184249ee206d957b8248cd8310`、`nvidia/GEM-X` 的 **`gem_smpl.ckpt`**。模型源码没有修改;适配代码位于 portable 仓库。官方训练入口是 `scripts/train.py exp=gem_smpl`,本地桥接入口 `train_genmo.py` 直接继承官方 `GEM.training_step`、`train_step`、AdamW 和 regression→diffusion 执行顺序,使用原生 Lightning Trainer、噪声过程、151D 编解码与 3D 损失。 +- +-## 当前验证范围 +- +-只验证一条 train split 的双指令 interruption session:接别人递来的物品,随后切换到捡地上物品。输入、音频与此前 OMG/TextOp 的 `train-002203-seed42` 相同。human session 为 740 帧 @50Hz;连续导出覆盖 444 帧 @30Hz,视频 14.8 秒。首帧为已知初始化,其余帧闭环生成。此实验不是全数据训练、未见样本评估或物理跟踪。 +- +-原 human 包仍是源动作。`prepare_genmo_session.py` 使用 `hfd_bridge/human_sessions.py` 观察同一次原生 HFD recipe 执行,记录 source ID、精确采样时间、随机裁剪、listen 持姿帧、过渡 envelope、截断/恢复和 append 计划。人体端从 SQLite/tar 索引读取 SMPL-X,不批量解压;根据权威配对表和 SHA256 对应源动作。平移按秒插值,旋转按 SO(3)/SLERP 处理,使用人体自身的根高度和 pelvis offset 重新对齐。idle 采用局部旋转残差,listen 采用所选人体持姿与连续过渡,整个 session 保持同一个 16D shape。 +- +-原生 G1 复现与旧结果的 PCM、控制标签和决策索引逐值一致。受环境浮点实现影响,G1 数组最大差异为 `2.3841858e-7`;输出保留旧 G1 数组作为参考。该误差已记录,不宣称 G1 数组逐字节一致。50Hz/源时间的末端舍入最多允许一个 50Hz 帧的边界保持,实际保持时间写入计划,超过即报错。其他任务的几何/事件组合尚未逐一验收;观察器遇到无法追踪的姿态会拒绝导出。 +- +-## 输入与因果边界 +- +-- 原生 GEM-SMPL 151D:21×6D body rotations、10D shape、camera/gravity 两组根旋转、3D 根局部逐帧位移。沿用 checkpoint 的统计量,不另拟合 HFD 归一化。 +-- 使用相同 Whisper 因果 sidecar 的 `stable_latest_sentence`,经原生 T5-3B 编码,最多 50 tokens。冻结编码器,以 BF16 计算并缓存 FP32 特征;空文本缓存为零。缓存核对原 session 哈希。 +-- ASR 释放仍为 `0.6+0.4*k` 秒,动作决策为 `0.4*k` 秒;保留既有 HFD 三时钟约定。每次固定预测未来 60 帧,执行 12 帧,再重规划。 +-- 对照输入历史为 12 帧(0.4s)或 60 帧(2s)。历史包含当前状态。闭环仅使用实际已生成的历史,不在每次规划读取真值;首帧不足的历史重复该已知姿态。 +-- 视觉、2D、相机、music、speech 分支全部标缺失。没有未来音频、GT 指令、listen/motion/idle 标签、未来 Stop、任务进度、GT turn ID 或 GT 总时长作为条件。 +-- 固定虚拟坐标映射为 AMASS Z-up→GEM Y-up,camera 和 gravity 使用同一固定坐标;平移转换包含 shape 对应的 SMPL-X pelvis offset,导出时返回 Z-up。 +-- 原生根“速度”使用前向帧差。历史单独重采样/编码,最后一个已知姿态的下一步位移保持未知;其 mask=0,作为未来边界监督。不能使用第一帧未来 GT 计算该历史特征。 +- +-沿用原生 observed-motion overwrite,不添加新的 mask embedding 或网络层。已知前缀在网络输入及输出均固定;原生根位移 loss 会积分整段窗口,因此仅锁输入、却让任意历史输出补偿未来位移是不一致的。输出固定通过无参数 forward hook 实现,并覆盖训练和 DDIM 采样。未知边界位移保留梯度。 +- +-## 训练损失与执行 +- +-保留原生 simple MSE、root-relative 3D joints/437 vertices、world translation、static-contact BCE;缺少真实图像时关闭相机平移和 2D/投影损失。simple loss 的 mask 排除复制的已知特征;几何/contact loss 只对未来有效帧评分。注意原生 masked mean 仍除以固定窗口尺寸,因此**不同历史长度的原始 loss 数值不宜直接比较**。比较应使用同样长度的闭环输出指标。 +- +-full model 共约 518M 参数,原始权重 709 个 state tensors 全部严格匹配。motion 模型全部微调,T5 冻结;batch=2,BF16,AdamW,gradient clip=0.5。本轮先做 3 步训练烟雾检查,再探索 300 步及低学习率继续训练。排查完 prefix 输出一致性后,两种历史长度都从同一官方权重、相同 seed=42、lr=2e-5 开始,分别训练 600 步。`locked-h12-fit600` / `locked-h60-fit600` 是用于比较的修正版本;此前 `pilot`、`fit300`、`fit1300`、`history60-fit600` 是排查记录。 +- +-## 位移表示往返检查 +- +-固定版本的 `get_local_transl_vel` 文档写 R^T,但实际代码为 R×world_delta;`rollout_local_transl_vel` 也乘 R,两者不互逆。真值完整 session 的原生往返误差约 0.01168m,正确逆变换 R^T 的误差约 6.4e-10m。桥接保留原生编码/预训练特征定义,修正解码与 world-translation loss 的积分方向;不修改上游 checkout。随机旋转的往返测试防止仅在近似单位旋转上“偶然正确”。 +- +-`final-h12` / `final-h60` 是输出前缀锁定后的匹配 600 步对照,保留了原生积分方向,用于记录历史长度实验。发现积分问题后,以 h60 的 600 步权重为初值,lr=5e-6 再训练 300 步(AdamW 状态重新初始化),保存为 `corrected-h60-fit900`;最终复现入口默认采用正确逆变换。`initial-corrected-h60` 是相同解码规则下的官方权重基线。修正这项确定性错误不等于消除了自由闭环的速度偏置/误差累积。 +- +-## 重要发现:短历史存在观测歧义 +- +-`input-ambiguity-h12.json` 找到两对完全相同的文本和历史姿态: +- +-| 动作时刻 | 当前文本 | 两份未来 body rotation 标签的 RMSE 差异 | +-| --- | --- | --- | +-| 2.8s / 3.2s | Taking an item from someone | 0.230 rad | +-| 7.6s / 8.0s | Picking an item up from the floor | 0.264 rad | +- +-同样的输入不能唯一确定两份不同的未来真值,不能把无法精确过拟合都归为优化问题。2s 历史在这条样例中消除了上述精确重复,但这不证明所有任务已经可观测,也不保证模型能记住更长的语义状态。若仍有事件响应问题,后续应单列原生 past-audio 或因果状态记忆实验,不能注入 GT 阶段计时器。 +- +-## Retarget 与视频 +- +-`infer_genmo.py` 导出标准 AMASS SMPL-X。GEM 不预测手指,预测端使用零残差手姿态;known shape 使用 session bootstrap 的 16D betas,GEM 原生特征取前 10 维。human GT 保留实际手残差。评测使用 `flat_hand_mean=False`、`use_pca=False`、`ext=npz`,与 human 包约定一致。 +- +-`evaluate_human.py` 将**完整连续轨迹**各 retarget 一次到 G1,每条轨迹使用独立、跨帧连续的 GMR IK 状态。GMR commit `bb1bbe40774794fceb2a7c579a3464a28e68c844`,solver=daqp,`offset_to_ground=False`。仅稀疏检出 GMR package 与 Unitree G1 资产,约 79MB;AutoDL 安装依赖失败后通过清空子进程代理的阿里云 HTTPS PyPI 国内镜像直连完成依赖安装。 +- +-同时保留三组比较和视频:prediction→G1 vs 原 G1、human GT→G1 vs 原 G1、prediction→G1 vs human GT→G1。人体真值转换与原 G1 的关节 RMSE 约 0.202 rad,是几何转换/表示的诊断,不能从模型误差中直接相减。视频沿用 `render_overfit.py` / OMG MuJoCo renderer,含原始混音声音,展示真实输入的因果 ASR 和初始化/预测历史标记。 +- +-## 文件和复现 +- +-实验根目录:`/root/autodl-tmp/runs/archive/legacy-20261020/genmo-overfit`。模型训练权重放 `/root/autodl-fs/pangxinlei/models/genmo/hfd-overfit`,数据盘只保留软链接。源数据仍保留 tar;原生 GenMo 和 GMR checkout 没有修改。 +- +-配置入口为 `configs/genmo_overfit.yaml`,一键运行: +- +-```bash +-cd /root/autodl-tmp/hfd-baseline-portable +-/root/autodl-tmp/envs/genmo/bin/python run_genmo_overfit.py \ +- --config configs/genmo_overfit.yaml \ +- --output /root/autodl-tmp/runs/genmo-new-overfit \ +- --prepared-session /root/autodl-tmp/runs/archive/legacy-20261020/genmo-overfit/session \ +- --text-cache /root/autodl-tmp/runs/archive/legacy-20261020/genmo-overfit/text.pt +-``` +- +-不提供后两个参数时,会重新准备同一条 paired session 和文本缓存。输出目录必须是新的。此 runner 调用的只是本轮已逐步验证的准备、训练、推理、retarget、渲染入口;组合 runner 自身不额外启动一轮重复训练。 +- +-契约检查:`tests/test_genmo_contract.py`、`tests/test_human_eval.py`。真实输入/梯度审计为 `tools/audit_genmo_inputs.py`:扰动全部未来 GT 后,observed motion、mask、f_cond/uncond/empty 和 T5 条件逐值不变;原生 regression/diffusion 两模式实际反向,532 个 denoiser 梯度 tensors 有限且非零。另检查 SO(3) 跨 ±π 插值、坐标/pelvis roundtrip、前缀损失掩码与未知边界位移梯度。 +- +-## 本轮结果(2026-10-18) +- +-正式比较使用同样的 2s 历史、固定坐标和修正后的位移逆变换,单次已知初始化后闭环生成,不使用真值重置: +- +-| 设置 | 对原始 G1 的 29 关节 RMSE (rad) | 根位置每坐标 RMSE (m) | +-| --- | ---: | ---: | +-| 官方权重 | 0.4786 | 0.3804 | +-| 微调 600+300 步,自身历史 | 0.1998 | 0.1285 | +-| 同一权重,真值历史诊断 | 0.1857 | 0.0125 | +-| human GT 经同一 retarget | 0.2016 | 0.0093 | +- +-最终模型与 **同样 retarget 的 human GT** 比较为关节 0.1384 rad、根位置 0.1236m;这更直接衡量相同转换后的生成误差。人体原始空间的闭环根位置每坐标 RMSE 为 0.1426m,真值历史诊断为 0.0070m。曲线显示根位置存在随时间积累的偏差,不能将本轮称作完整成功过拟合。动作阶段也仍有偏差。 +- +-前缀输出锁定、原生积分方向下的匹配 600 步对照:0.4s 历史的 G1 关节 RMSE=0.3420、2s 历史=0.1975 rad;两者根位置误差均约 0.34–0.35m。这组仅用于记录输入长度实验,与修正后 900 步结果分开。 +- +-结论:**数据适配和完整推理评测链路通过;当前小样本闭环过拟合未完全通过。** 后续建议从 2s 历史开始,分别验证闭环预测历史训练、原生 contact 后处理或显式因果状态记忆;每一项单独标注适配和对照,不加入 GT 任务阶段或手工真值位置校正。增加其他任务前,先冻结并验收对应人体/G1 session。 +- +-结果入口: +- +-- `/root/autodl-tmp/runs/archive/legacy-20261020/genmo-overfit/README.md`、`summary.json`。 +-- 主视频:`final/evaluation/rollout.mp4`;相同 retarget 比较:`final/evaluation/retargeted_human_comparison.mp4`;转换诊断:`final/evaluation/retarget_reference.mp4`。 +-- 最终三条视频均为 444 帧、30fps、14.8 秒,原始混音声音已解码核验;`videos.json` 保存详细检查。人工查看了 listen、第一动作、打断后的 listen、第二动作的代表帧。 +-- `trajectory_diagnostics.png` 展示关节误差和根漂移;视频是参考姿态播放,不代表物理控制成功。 +-- 最终权重:`/root/autodl-fs/pangxinlei/models/genmo/hfd-overfit/corrected-h60-fit900/fitted.ckpt`。 +-- 14 个相关契约测试通过;完整官方权重严格加载、真实 regression/diffusion 反向、未来 GT 扰动后的采样输出逐值不变、真实 GMR 和音视频集成已验证。GPU 训练峰值约 9.17GB。ASR/T5 为预计算缓存,采样耗时不是端到端在线延迟。 +- +-复现最终继续训练阶段:在 portable 目录执行 `GENMO_HISTORY_FRAMES=60 /root/autodl-tmp/envs/genmo/bin/python train_genmo.py --session /root/autodl-tmp/runs/archive/legacy-20261020/genmo-overfit/session --text-cache /root/autodl-tmp/runs/archive/legacy-20261020/genmo-overfit/text.pt --checkpoint /root/autodl-tmp/runs/archive/legacy-20261020/genmo-overfit/locked-h60-fit600/fitted.ckpt --output /root/autodl-tmp/runs/genmo-reproduce900 --steps 300 --batch-size 2 --lr 5e-6`。当前 `run_genmo_overfit.py` 默认是从官方权重直接运行修正后的 600 步实验配方,与本轮分阶段 900 步训练的优化器历史不同,应在报告中区分。 +diff --git a/docs/GENMO_LISTENING_DIAGNOSIS.md b/docs/GENMO_LISTENING_DIAGNOSIS.md +deleted file mode 100644 +index a7ab5de..0000000 +--- a/docs/GENMO_LISTENING_DIAGNOSIS.md ++++ /dev/null +@@ -1,25 +0,0 @@ +-2026-10-24:针对GenMo多个episode长期保持倾听姿态,使用当前92K权重做离线对照。正式权重未修改,训练损失、采样配方和推理规则均未改动。 +- +-已经确认:失败发生在人体生成阶段,早于retarget。按词释放的英文文本确实进入模型,模型也能产生变化的未来特征;当前证据不支持“没有读到文本”或“整个网络已经输出常数”。主要暴露的问题是:GT历史训练下能延续运动,反馈自身预测时却容易停留在倾听状态,动作启动不足。 +- +-| 后退样本val-backchannel-000057 | 人体姿态RMSE(rad) | 根位置RMSE(m) | 整段根路径(m) | +-|---|---:|---:|---:| +-| 正常预测历史闭环 | 0.536 | 1.252 | 0.136 | +-| 每0.4秒使用真实历史,仅作诊断 | 0.155 | 0.030 | 4.702 | +-| 人体GT | — | — | 3.652 | +- +-GT历史对照会不断重置条件及空间位置,因此位置误差天然受保护,其速度误差反而较大;它不能作为正式评测或有效修复。该对照说明运动历史对结果影响很大,并不证明模型能独立完成后退。 +- +-同一8秒决策点、相同英文caption和噪声,单窗口把真实运动历史改为固定倾听历史:CFG1时未来两秒的根路径由0.266m降至0.011m;CFG2.5时由0.206m降至0.093m。改变历史会明显影响运动启动。直接清空文本也会改变预测:t=999时未来151D特征平均绝对差约0.00963,真实文本特征范数约17.05;此时未来特征的帧间RMS约0.0271。因此不能把倾听姿态卡住等同于编码器完全失效。 +- +-曾怀疑原生CFG的无文本分支移除历史条件特征。实际检查当前`in_attr`不含该独立历史特征,历史主要由`motion_mask_3d/observed_motion_3d`约束;强制让两分支共用`f_cond`后,单窗口和完整闭环输出均保持相同。该候选原因已经排除,没有据此修改生产推理。 +- +-训练数据抽查来自真实`train_sampling`记录的最早、约三分之一、三分之二和末尾四个池,共128个窗口。按根速度>0.1m/s或平均身体角速度>0.3rad/s、活动帧占比至少25%分类:80个窗口的两秒未来有活动,63个窗口的实际12帧执行段有活动,49个窗口的历史较安静,其中13个执行段转入活动,122个窗口有非空文字。此抽查不是全语料统计,但排除了“训练全部读成静止样本”的解释。起动转换仅占该抽查约10.2%,不能据此推断全库比例。 +- +-当前训练完整地监督未来60帧,闭环只执行最前12帧;全部历史来自GT,没有预测历史混入。模型一旦在起动处继续预测倾听,下一次重规划又看到自己的倾听历史;caption不变时容易形成持续保持姿态的闭环。近真实未来加噪的去噪损失下降,不能保证从安静历史自主启动动作。这是目前最直接的失效机制,具体损失权重、历史课程、数据分布的独立贡献尚未完成受控消融。 +- +-后续训练策略应优先验证:提高已执行12帧及起动窗口的监督占比,加入真实解码运动的变化监督,并用train-only的预测历史训练恢复能力。输入仍保持原协议,不能用GT阶段、未来文本或评测时真实历史替代这些训练工作。本轮没有直接启用这些未验收配方,也没有重新训练100K。 +- +-本地诊断代码在被忽略的`test/diagnostics/`,数值与中间轨迹在本机cache;GitHub仅保留本文结论和复现当前正式训练所需的代码。当前训练源快照见`TRAINING_SNAPSHOT.json`及对应Git历史;本次迁移改动仅处理路径、资源定位和训练依赖拆分,没有改变训练算法。 +- +-后续已实现的训练适配改动见 [GenMo v2](GENMO_V2_TRAINING.md)。本文诊断数值仍对应旧正式权重,不能当作新策略全量训练后的效果。 +diff --git a/docs/GENMO_PAST_AUDIO_TEST.md b/docs/GENMO_PAST_AUDIO_TEST.md +deleted file mode 100644 +index 4d9c162..0000000 +--- a/docs/GENMO_PAST_AUDIO_TEST.md ++++ /dev/null +@@ -1,93 +0,0 @@ +-# GEM-ASR-past-audio:原生音频分支测试 +- +-2026-10-18。在已有 GenMo/HFD bridge 中启用官方 speech encoder,使用同一条冻结 interruption session、同一因果 ASR、同一人体真值和 GMR 链路。不是新增语音语义模型;ASR 文本仍经 T5-3B,音频由 checkpoint 自带的卷积前端编码。 +- +-## 输入构造 +- +-第 k 次决策的动作时刻是 `t_k=0.4*k`,按既有 HFD 协议,PCM 释放点是 `r_k=0.6+0.4*k`。取**释放点之前**的 `[r_k-2,r_k)` 两秒混合 PCM;不能把动作轴、PCM 轴和墙钟直接视为同一个零点。这里保留 OMG/TextOp/HFD 的释放规则。 +- +-1. 从原 16kHz session 混音按采样点截取,右端严格为 `9600+6400*k`。启动时不足两秒的左侧补静音,只有真实覆盖的历史位置 availability=1;首个决策实际覆盖 18 个音频 bin。 +-2. **先截断,再重采样和卷积。** SciPy polyphase FIR 按 9/8 重采样到 18kHz,保持波形幅度。未来样本没有进入 FIR 或卷积。 +-3. 原生六个 BasicBlock 将 `[B,1,36000]` 编码为 `[B,128,60]`;stride=600 samples,恰为 30Hz。卷积感受野为 18,635 samples(约 1.035s)。所有卷积支持都限制在本次已释放窗口及边界补零内;不是在完整 session 上先提特征再切片。 +-4. **编码后**补 60 帧未来特征,再用原生 `audio_embedder`、`cond_exists_embedder` 和 CFG 逻辑融合。120 帧条件中只有历史音频位置可用,未来 audio availability 恒为 0。未来的 f_cond 数值可以有“缺失条件”的学习偏置,但与波形内容逐值无关。 +-5. 特征按时间顺序放入 60 个动作历史槽位。卷积的中心时间为 `r_k-2+j/30`,最后一个中心比释放点早 1/30 秒;保留原生卷积的 phase,没有为对齐而读取释放点之后的样本。原生训练 BatchNorm 会在已截断的 batch/time 内统计;推理使用 eval/running statistics。 +- +-其它输入保持:60 帧已执行动作历史+T5 当前稳定 ASR;预测 60 帧,执行 12 帧,30Hz。训练使用真值历史,闭环使用模型自身历史;仅首帧初始化已知。不输入未来音频、GT 指令、状态/turn 标签或任务进度。music、视觉、2D 和相机条件仍缺失。 +- +-## 实现边界 +- +-- 新增 `hfd_bridge/genmo_audio.py` 负责释放窗口、重采样和卷积后 padding。 +-- `GENMO_PAST_AUDIO=1` 且 `GENMO_HISTORY_FRAMES=60` 启用音频;默认关闭,因此已有 ASR-only 入口保留。 +-- 在原生 `GEM.create_condition_mask` 外接入已截断的波形;encoder 的无参数 forward hook 仅在卷积之后补齐 future features。没有添加新的学习层、mask embedding 或改变 709 个 state tensor 的结构。上游 checkout 保持 clean。 +-- 保留原生音频条件 dropout=0.1、BatchNorm、卷积和音频投影的训练方式。模型主体及音频前端一起微调,T5 保持冻结。 +-- 之前修正的 prefix 输出锁定、未知边界位移和实际位移编码逆变换在两组实验中一致。 +- +-## 对照设计 +- +-两组均从 `/root/autodl-tmp/runs/archive/legacy-20261020/genmo-overfit/corrected-h60-fit900/fitted.ckpt` 初始化,再各训练 600 个 optimizer steps,batch=2、lr=5e-6、seed=42。motion 模型累计微调 1500 步;音频前端在本轮训练 600 步。优化器均重新初始化。两组的窗口、文本、监督、损失权重、DDIM 10 步、CFG 和采样种子相同;音频前端 dropout 会消耗额外随机数,因此不宣称两组训练噪声逐位一致。 +- +-| 训练目录 | 输入 | +-| --- | --- | +-| `asr-control-fit600` | ASR+2s 动作历史 | +-| `past-audio-fit600` | 同上+2s 已释放混合 PCM | +- +-另对音频模型使用 `infer_genmo.py --drop-audio` 做消融:清零波形并把 audio availability 全设 0,ASR、动作历史和权重不变。视频仍播放原 session 的真实声音,标题明确显示 `AUDIO OFF`;播放声音不等于模型收到了音频。 +- +-## 验证 +- +-- 完整模型严格加载仍为 709 个 state tensors。 +-- 扰动释放点之后的原始 PCM,重采样输入逐值相同;同时扰动未来 GT,原生条件及 DDIM 输出逐值相同。 +-- 过去 PCM 改变实际条件和未来采样结果,未来 audio 条件槽与波形内容无关;CFG unconditioned branch 对波形不变。 +-- 原生 regression/diffusion 两模式实际反向,音频 encoder 与 projection 的梯度有限且非零。 +-- 训练完成后的 checkpoint 检查:ASR-only 的音频 encoder/projection state 与初值完全相同;音频组有实际更新,BN batch counter 增加 1200,符合每步 regression/diffusion 各一次前端调用。 +-- 四个音频契约测试覆盖未来 PCM 隔离、启动补零与可用帧、16→18kHz 的频率保持、拒绝未覆盖的释放点。原有六个 GenMo 契约测试在音频开/关两种进程配置下通过。 +- +-测试根目录:`/root/autodl-tmp/runs/archive/legacy-20261020/genmo-audio-overfit`,真实前端审计为 `audio-audit.json`,权重更新审计为 `trained-audio-state-audit.json`。 +- +-## 配置与运行 +- +-配置:`configs/genmo_audio_overfit.yaml`;组合 runner 为 `run_genmo_overfit.py`,设置 `past_audio: true` 会在训练和推理子进程中显式启用对应环境变量。要做匹配的 ASR-only 对照,复制配置只把 `past_audio` 改为 false,保持初值、步数和学习率一致。 +- +-本轮直接调用的训练方式(两组用不同的新输出目录): +- +-```bash +-GENMO_HISTORY_FRAMES=60 GENMO_PAST_AUDIO=1 \ +-/root/autodl-tmp/envs/genmo/bin/python train_genmo.py \ +- --session /root/autodl-tmp/runs/archive/legacy-20261020/genmo-overfit/session \ +- --text-cache /root/autodl-tmp/runs/archive/legacy-20261020/genmo-overfit/text.pt \ +- --checkpoint /root/autodl-tmp/runs/archive/legacy-20261020/genmo-overfit/corrected-h60-fit900/fitted.ckpt \ +- --output /root/autodl-tmp/runs/genmo-new-audio-test \ +- --steps 600 --batch-size 2 --lr 5e-6 +-``` +- +-权重默认写 fs 的 `models/genmo/hfd-overfit//fitted.ckpt`,数据盘只保存软链接。推理时使用同样的两个环境变量。评测和视频继续调用 `evaluate_human.py`、`render_overfit.py`,对完整连续轨迹 retarget;不在每个规划窗口重置 IK。ASR、T5 为预计算缓存,报告的采样耗时不是端到端在线耗时。 +- +-## 实测结果 +- +-以原始 G1 为参考,全部排除首帧初始化;根位置为每坐标 RMSE,未做根对齐或 DTW: +- +-| 设置 | 采样 seed | 29 关节 RMSE (rad) | 根位置 RMSE (m) | +-| --- | ---: | ---: | ---: | +-| 匹配 ASR-only 对照 | 42 | 0.2061 | 0.0366 | +-| 原生过去音频,eval/running BN | 42 | 0.2893 | 0.1030 | +-| 匹配 ASR-only 对照 | 43 | 0.1992 | 0.0261 | +-| 原生过去音频,eval/running BN | 43 | 0.2773 | 0.1096 | +-| 音频模型,关闭 speech 条件 | 42 | 0.1992 | 0.1076 | +-| 音频模型,当前窗口 BN 诊断 | 42 | 0.2380 | 0.1085 | +- +-与同样 retarget 的 human GT 比较,seed 42 的 ASR-only 为关节 0.1395 rad、根位置 0.0350m,音频版为 0.2480 rad、0.0989m,方向一致。因此,退化不能仅归因于原始 G1 与 human GT 的转换差异。`summary.json` 另存 listen/motion/idle 三类 GT 阶段的评分;这些标签仅用于评估,未输入模型。 +- +-诊断中的 `--audio-normalization window` 只改变推理时音频 BatchNorm:其它模块保持 eval、Dropout 关闭,BN 使用**当前已截断的两秒窗口**统计,关闭 running-stat 更新,不使用未来 PCM 或整段 session 的统计。这是单独的归一化适配诊断,不是默认原生推理。姿态有所改善,但根漂移没有解决,尚不足以将归一化差异认定为唯一原因。未来 PCM 扰动的逐值不变性检查针对原生 eval 推理;训练保留原生跨 batch/time 的 BN 统计语义。 +- +-**结论:当前版本的因果性、维度、权重加载、梯度和完整评测链路通过,但小样本性能没有优于 ASR-only,音频版尚未成功过拟合。** 两个采样种子都显示相同方向;它们不是两次独立训练,更不是未见数据泛化验证。暂保留 ASR-only 为工作版本,音频模型作为实验结果保存。 +- +-可能的后续检查包括:固定音频前端/BN 的训练对照、原生逐帧音频预训练与仅历史条件之间的分布差异、闭环预测历史训练。当前结果不证明原生音频能力无效,也不能证明上述某一项已经解释了退化。 +- +-## 产物 +- +-- 汇总与曲线:`/root/autodl-tmp/runs/archive/legacy-20261020/genmo-audio-overfit/README.md`、`summary.json`、`comparison.png`。 +-- 音频版视频:`past-audio/evaluation/rollout.mp4`;ASR-only 对照:`asr-only/evaluation/rollout.mp4`。 +-- 两组均另有 `retarget_reference.mp4`(human GT→G1 vs 原 G1)和 `retargeted_human_comparison.mp4`(相同转换后的预测比较)。保留原始混音声音;视频为参考姿态播放,不是物理控制跟踪。 +-- 消融和归一化诊断保存在 `audio-disabled`、`audio-window-bn`,明确标注模型收到的输入;seed 43 复核保存推理数组与完整 GMR 数值,不重复渲染视频。 +-- 训练权重:fs 的 `models/genmo/hfd-overfit/asr-control-fit600/fitted.ckpt`、`past-audio-fit600/fitted.ckpt`。数据盘训练目录内为软链接。 +-- 未安装新依赖、下载新模型或复制主体资产;沿用已准备的 GenMo、T5、SMPL-X 和 GMR。 +diff --git a/docs/GENMO_V2_TRAINING.md b/docs/GENMO_V2_TRAINING.md +deleted file mode 100644 +index 0cc76f7..0000000 +--- a/docs/GENMO_V2_TRAINING.md ++++ /dev/null +@@ -1,52 +0,0 @@ +-# GenMo v2:执行段监督、预测历史恢复与 CPU 预取 +- +-本版本针对长期保持倾听姿态的失效机制改造训练适配。模型骨干、英文前景 GT 词时间释放、最近 30 秒上下文、2 秒动作历史、2 秒预测及每次执行 0.4 秒的输入输出协议不变。没有新增音频、未来文本或 GT 阶段条件。正式旧权重保留;短程链路检查不能替代新策略的全量质量评估。 +- +-## 训练目标 +- +-保留原生 regression + diffusion 两个训练分支及姿态几何损失;全局根平移监督改为按预测朝向解码的实际轨迹,避免用GT朝向掩盖推理时的方向误差。151D 特征误差中,实际执行的前 12 帧及历史末帧到第一未来帧的位移权重设为 3。逐样本、逐特征归一化权重,保持有效监督总权重与原配方一致,避免仅仅放大总 loss。无效尾帧及未知末帧位移仍不参与监督。 +- +-另加前 12 帧的解码根速度与身体关节旋转增量 SmoothL1 监督,默认权重分别为 1 和 0.05。根位移按原生编码器的实际逆变换 `Rᵀ @ local_delta` 解码,旋转来自预测本身;不以未来 GT 朝向代替预测朝向。旋转增量在 SO(3) 上计算,避免轴角跨分支的假速度。速度项不是鼓励一直运动,正常倾听/停止的 GT 速度仍是低速或零。 +- +-## 预测历史恢复 +- +-每个训练 batch 最多选 25% 的可用窗口,保持原始窗口采样分布。前 1,000 步不混入预测历史,接下来 4,000 步逐步增加,5,000 步起达到该上限。源窗口不足 0.4 秒历史时不参与恢复。 +- +-对时刻 `t` 的恢复样本,在 `t−0.4s` 决策点用当时已释放的文字及过去动作生成一次原生 10 步 DDIM 预测,只取实际执行的 12 帧,替换 `t` 输入历史的末 12 帧。生成所需的未知未来位置填成末个已知姿态,不读 GT 未来;预测过程关闭梯度及 dropout/BN 更新。未来GT关节姿态、根旋转、位移增量及时间不变;将未来根轨迹整体平移到实际历史的末位置,保持相同的局部运动。GEM的151D输入含根速度而不含绝对根位置,因此不能要求它纠正输入中不可观测的全局位置偏移。目标编码及速度/旋转增量的边界以这份一致的历史/未来为起点,避免监督矛盾。目标编码、解码和新增时序损失使用FP32几何计算。恢复样本仍需从偏离的自身预测回到正确轨迹。 +- +-这是一段预测历史的恢复课程,不等同于对无限长闭环的保证;必要时可另做多段恢复消融,不能把评测时使用 GT 历史当作修复。 +- +-## 验证与权重选择 +- +-继续记录固定噪声 `val/loss`。默认另选两个不同 interaction mode 的固定英文验证 session,各最多 900 帧。只用一个初始真实姿态启动,之后全部反馈自身预测,文字沿原时间轴释放。训练/验证 RNG 在检查后恢复。 +- +-默认 `best` 按 `val/closed_loop_score` 选择,定义为两个 session 的 `body geodesic RMSE(rad) + root coordinate RMSE(m)` 的平均值。另保存根速度误差与预测路径长度。这是训练选权重的代理指标,不是全测试集分数,也不替代人体到 G1 retarget 后的正式评测。`last` 仍按训练步保存完整恢复状态。 +- +-## 吞吐改动 +- +-- 单独的 CPU producer 负责 session 组装和窗口构造,最多预取 16 个 batch(一个样本池周期,深度上限 32);GPU 主进程负责冻结 T5 与训练。进程间只传配置、CPU batch 和采样状态,不共享 SQLite 连接或 CUDA 对象。 +-- checkpoint 记录已消费 batch 的池/RNG 状态,忽略预取但尚未消费的数据。producer 及主进程训练/验证 DataLoader 均使用独立随机生成器,不挤占模型的扩散随机流。 +-- GenMo 的 SO(3) 插值/混合批量化;人体几何 observer 只挂接实际需要的函数返回事件,避免全进程 `sys.setprofile` 的逐函数开销。原 HFD 文件及其校验哈希不修改。此优化需通过与旧 observer 的成对 session 对照,不能只看速度。 +-- 训练 session 内存预算为 4 GiB,验证预算为 8 GiB,文本特征 LRU 为 512 MiB。另有进程及索引开销;内存较少的机器可降低这些预算,采样规则不变。 +-- 启用内容校验的本地原生 session 索引缓存,缓存的是源组合索引,不是 GT session 全量导出或模型输入。原始 tar 继续留在共享盘,离线词时间缓存、索引缓存、代码和日志位于工作盘。 +-- 保持 `refresh_batches=16`,没有通过延长同一批 session 的重复使用来虚增吞吐。GT 几何在两个训练分支间复用一次计算。NumPy/OpenBLAS默认限制为1线程,避免小矩阵使用64线程的调度开销;正式配置关闭逐 batch 进度条,保留结构化指标输出。 +- +-耗时记录新增 `cpu_batch_seconds`、`text_encoding_seconds` 和样本池刷新标记。`data_wait_seconds` 是训练批次间的主进程等待,不能直接等同于磁盘 IO;验证耗时不再计入该字段。评估速度时应跨越多个样本池刷新,并单独报告初始化。 +- +-本机稳定阶段实测约快 1.4–3.1 倍,较保守复测为 1.44 倍;冷启动和共享存储波动仍明显。完整测量口径、正确性与小规模质量限制见 [GENMO_V2_VALIDATION.md](GENMO_V2_VALIDATION.md)。 +- +-## 命令与旧实验复现 +- +-现有服务器从仓库根目录正常前台运行: +- +-```bash +-/root/autodl-tmp/envs/genmo/bin/python -u train_genmo_full.py --config configs/genmo.yaml +-``` +- +-新输出为 `runs/hfd-genmo-english-v2`,权重为共享盘 `models/genmo/hfd-english-v2`;不覆盖旧的 `hfd-english-train`。跨机器先按 [PORTABLE_TRAINING.md](PORTABLE_TRAINING.md) 生成本机配置。 +- +-新 run 的完整断点使用同配置加 `--resume auto`。如要从旧正式权重初始化新策略,使用 `--init-checkpoint /path/to/old.ckpt`;这是权重 warm start,优化器、调度与步数重新开始,不能伪装成旧 run 的续训。默认仍从官方 `gem_smpl.ckpt` 开始。 +- +-`configs/genmo_legacy.yaml` 保留旧配方;配置生成器 `--genmo-recipe legacy` 可以复现该配方。若要审计旧训练的原实现,Git `0bcc49e` 保留了与旧 GenMo/Kimodo 正式训练源码记录匹配的快照,`fa5c1a5` 是本次策略修改前的便携版。恢复旧完整断点应使用对应原配置、代码和资产。跨版本/路径改动不承诺逐位相同。 +- +-测试、基准和过拟合脚本均在忽略的 `test/`,不随训练代码发布;实验日志、权重和视频不进入 Git。 +diff --git a/docs/GENMO_V2_VALIDATION.md b/docs/GENMO_V2_VALIDATION.md +deleted file mode 100644 +index ab834f4..0000000 +--- a/docs/GENMO_V2_VALIDATION.md ++++ /dev/null +@@ -1,45 +0,0 @@ +-# GenMo v2 本机验证记录(2026-10-24) +- +-本页记录训练实现和短程实测,不能视为新策略已经完成全量训练或测试集验收。正式配置与设计见 [GENMO_V2_TRAINING.md](GENMO_V2_TRAINING.md),跨机器准备见 [PORTABLE_TRAINING.md](PORTABLE_TRAINING.md)。诊断脚本、原始日志、测试权重和视频均不发布到 Git。 +- +-## 原耗时来自哪里 +- +-旧 100K run 耗时约 85.8 小时,记录的平均计算约 0.367 秒/步,平均批次等待约 2.718 秒/步;旧等待字段还混入验证间隔,不能直接当作 IO 时间。一次冷启动的 8-session CPU profile 中,实际读取约 625.6 MB 的累计耗时为 1.027 秒,而旋转插值累计约 6.644 秒。session 组装、旋转处理及主进程串行等待也是瓶颈,单纯搬动 tar 不能解释全部耗时。 +- +-本次保持原始 tar 在共享盘、缓存和派生索引在本地盘;采用向量化几何、定点 observer hook、本地索引缓存、单 CPU producer 和最多 16-batch 预取。保持每 16 batch 刷新样本池,未通过延长同一批 session 的重复使用换取吞吐。 +- +-## 吞吐实测 +- +-同机 RTX 5090 32 GB,batch 32,各执行 64 次真实优化更新,跨越 4 个样本池。基线与优化版的 session ID、决策帧和采样顺序一致;新策略开启恢复训练,GPU 计算负担比旧策略更高。吞吐检查关闭验证,不含正式闭环验证开销;总耗时栏是训练指标计时,另有进程初始化和退出开销。 +- +-| 测量 | 前 16 步(秒) | 64 步(秒) | 第 17–64 步平均(秒/步) | 同区间计算 / 等待(秒/步) | +-|---|---:|---:|---:|---:| +-| 旧策略、串行数据 | 51.61 | 284.00 | 4.84 | 0.35 / 4.49 | +-| 新策略、预取 16,较热缓存的一次 | 91.60 | 167.05 | 1.57 | 0.68 / 0.89 | +-| 修正监督后的新策略复测 | 111.50 | 272.33 | 3.35 | 0.68 / 2.67 | +- +-稳定阶段观察到约 **1.4–3.1 倍**吞吐提升;更保守的复测是约 **1.44 倍**。包含前期等待的 64 步复测只缩短约 4%,不能承诺全量训练固定快 3 倍。共享存储/主机负载与缓存状态使样本池生成耗时明显波动,预取仍未完全遮蔽 CPU 等待。新策略两次峰值显存约 20.7 GB(十进制),基线约 25.4 GB;这不是最大 batch 搜索结果。根轨迹监督最后增加了平移重锚,最终源码另通过参数更新与恢复检查,未再重复整套吞吐基准。 +- +-## 正确性检查 +- +-- 14 个成对 session 覆盖 6 种交互模式:音频、G1、文字、时间及参考元数据完全相同;人体数组最大绝对差约 `7.45e-9`。原始 HFD 文件及哈希不变,旧 profiler 状态保持。 +-- 7 项单元检查通过,覆盖 SO(3) 插值、执行段权重总量/有效 mask/梯度、姿态键、显式与隐式返回 hook、恢复样本条件、配置迁移与资源链接。 +-- 原生编码器检查通过:扰动历史后,保持未来局部运动并平移重锚的完美预测,其新增速度、根轨迹损失接近零;错误预测有有限且非零的梯度。未来 GT 关节姿态、根旋转和位移增量保持,源 GT batch 未改写。 +-- 训练恢复以预取队列中**已消费**的样本状态为准。最终源码连续 4 步与 2 步后中断恢复至 4 步的对照通过:709 个模型张量、Adam、调度器、采样状态及 Python/NumPy/Torch/CUDA RNG 均逐位一致。对照开启预取、预测历史恢复、固定噪声验证和闭环验证;这是短程恢复检查,不是长期质量保证。 +-- 外部资产清单中 141 个文件(约 53.6 GB)的 SHA256 全部核验通过。OMG 50M/100M/300M/500M 均严格加载并以配置中的微批次执行一次丢弃的真实参数更新;这不代表各尺寸已完成 HFD 正式训练。 +-- 从 Git 暂存内容导出的干净副本检查训练入口和模块依赖,不依赖忽略的 `test/` 或本机旧适配仓库。OMG/TextOp dry-run、GenMo 配置及新增模块导入检查通过;Kimodo 原生离线检查通过,真实参数更新成功,文本敏感性为 0.0632,根/身体梯度均有限且非零。 +- +-## 小规模闭环质量复核 +- +-使用旧 GenMo 正式 best 初始化,在两个**训练集** session 的全部合法窗口上训练 600 步,batch 8;恢复课程压缩为 warmup 40、ramp 160。此实验只检查新框架能否学习与闭环是否改善,不是默认 100K 配方的验证集比较,也不是与旧策略同预算的完整消融。 +- +-评估各自前 450 帧(15 秒):一个 GT 初始姿态,之后仅反馈预测历史,执行段为 12 帧;文字按原时间释放。无根对齐或时间扭曲。误差从第二帧开始统计,路径长度包含整个轨迹。 +- +-| 训练 session | 根坐标 RMSE,前→后(m) | 身体旋转 RMSE,前→后(rad) | 根路径,预测前→后 / GT(m) | +-|---|---:|---:|---:| +-| 53390 | 2.794 → 0.922 | 0.462 → 0.317 | 2.973 → 4.328 / 6.075 | +-| 28035 | 0.125 → 0.142 | 0.630 → 0.619 | 0.549 → 0.468 / 1.588 | +- +-首 20 步平均训练 loss 从 2.623 降至末 20 步的 1.499,身体、边界及未来速度输出均有非零梯度。一段恢复明显,另一段仍动作不足;此前其他短程窗口实验也有改善和退化并存的情况,不能声称倾听停滞已经彻底解决。 +- +-两段预测和人体 GT 分别对完整连续序列 retarget 一次到 G1,保存原 G1、人体 GT retarget 与预测的三种对照视频。视频本机目录为 `runs/genmo-v2-debug-20261024/videos/{53390,28035}/`;它们是训练集诊断,不用于报告测试集成绩。新正式 run 应在 20K/40K/60K 查看自主闭环与 retarget 视频后判断效果;默认 best 已改按固定自主闭环分数选择,旧正式权重保留。 +diff --git a/docs/HFD_HUMAN_BASELINE_INPUT_PROTOCOL.md b/docs/HFD_HUMAN_BASELINE_INPUT_PROTOCOL.md +deleted file mode 100644 +index 6579d20..0000000 +--- a/docs/HFD_HUMAN_BASELINE_INPUT_PROTOCOL.md ++++ /dev/null +@@ -1,149 +0,0 @@ +-# HFD 五个 baseline:当前输入、输出与训练约定 +- +-更新:2026-10-21(Asia/Taipei)。本文记录当前本地实现与约定;源码/运行配置优先。 +-原2026-10-17调研、上游commit、论文链接和待实现方案已完整保存在 +-[历史调研](history/HFD_HUMAN_BASELINE_INPUT_PROTOCOL_RESEARCH_20261017.md)。 +-历史ASR/中文翻译、12帧GenMo历史等配方不作为当前默认。 +- +-## 共同文本协议 +- +-五个baseline统一约定如下;OMG、TextOp、GenMo已实现,Kimodo已增加全量原生英文动态人体训练与闭环评测入口,HY-Motion仍待适配。Kimodo的小规模验收不代表全库正式训练已经完成。 +- +-```yaml +-foreground_language: en +-caption_policy: aligned_reference_context +-clock_policy: microphone +-text_conditioning: none +-context_seconds: 30.0 +-alignment_guard_seconds: 0.1 +-``` +- +-输入是已经释放的英文前景GT文本。原始TTS原文与对应源音频由 +-Qwen3-ForcedAligner-0.6B离线估计词时间,再按动态session内的真实音频拼接偏移平移。 +-原生指令、停止、附和、犹豫、修正片段均按各自的实际播放位置处理,不提前给出最终完整指令。 +-不运行ASR、不翻译、不用动作标签选择“当前任务”,也不做摘要或prompt rewrite。 +-背景混合PCM原样保留,但背景文字不输入。这是oracle foreground transcript实验, +-不代表已经解决部署时的ASR或目标说话人识别。 +- +-当前词级对齐库含61,860条英文前景资产;这是音频资产数,不是不同句子数或session数。 +-对齐跨度无效、越界、非单调或文本覆盖不完整时,在对应完整音频片段结束后释放整段文字。 +-0.1秒guard是实验缓冲,不是经过人工验证的对齐误差上界;事后强制对齐也不等同在线识别。 +-实现见 `hfd_bridge/reference_text.py`、`reference_runtime.py`,共享声明见 +-`configs/english_text_protocol.yaml`。 +- +-## 时钟与文本释放 +- +-动作源为50Hz,混合音频16kHz。动作时刻t对应麦克风时刻 `t+0.2s`。 +-文本释放点采用麦克风时间 `r_k=0.6+0.4k`,k从0开始。 +-词末估计为e时,先等到 `e+0.1s`,再向上取整到下一个r_k: +- +-```text +-r(word) = 0.6 + 0.4 * max(0, ceil((e + 0.1 - 0.6) / 0.4)) +-当前可见文本 = 最近已到达释放点r_k上的文本记录,要求r_k <= t + 0.2 +-``` +- +-因此麦克风0.6/1.0/1.4秒对应动作0.4/0.8/1.2秒;动作起点没有预先获得0.6秒文本。 +-这是当前microphone协议;旧 `legacy_hfd` 按动作组提前取记录的行为仅供历史复现。 +-旧材料“同一动作组直接读取0.6秒结果”不能继续套入当前正式实验。 +- +-每条记录只保留该释放点前30秒内、已经完成并获准释放的词;不等待句子完整, +-不拼接历次窗口字符串,过期后可以为空,不沿用任意久远的最后一句。session切换时清空。 +-例如词末0.85秒,加guard到0.95秒,在麦克风1.0秒释放,动作0.8秒决策才能看见。 +-0.6秒音频窗与0.4秒步长是重叠窗口:相邻只新增0.4秒,不能说每步新增0.6秒。 +- +-模型不接收未来PCM、未释放文字、GT turn ID、listen/motion/idle标签、未来事件时刻、 +-真实剩余动作时长、未来根路径或未来姿态约束。标签仅用于监督、样本选择、评测与视频标注。 +-扩散训练的带噪未来目标是学习变量;不能将干净未来GT作为条件。 +- +-## 实际输入与输出 +- +-共同的是信息释放规则,不要求模型拥有相同张量或动作历史长度。 +- +-| 模型 | 当前输入 | 当前输出及执行 | 本地接入状态 | +-| --- | --- | --- | --- | +-| OMG | 已释放英文文本→冻结T5-Base,最多50 token、保留最近后缀;过去2秒已释放混合PCM;10帧动作历史@30Hz | 原生125D动作特征,解码G1全29关节及根状态;计划60帧/2秒;当前原生异步缓冲剩24帧触发规划 | 正式英文训练及验证评测已接通 | +-| TextOp | 已释放英文文本→冻结CLIP ViT-B/32,77 token含两个边界token;2帧57D原生动作历史特征@50Hz;不输入音频 | 128D primitive latent,经冻结VAE解码8帧57D/0.16秒动作;G1原生23关节及根状态,未建模腕关节固定 | 正式300K训练完成;连续验证和视频见TextOp评测入口 | +-| GenMo/GEM-SMPL | 已释放英文文本→冻结T5-3B,最多50 token、保留最近后缀;60帧/2秒人体历史@30Hz;固定已知体型;当前past_audio=false | 原生151D表示,解码身体21关节旋转、根朝向及根位移;未来60帧/2秒,每次执行12帧/0.4秒后重规划 | 100K已完成,best92K与最终100K的闭环/GMR/视频见[正式评测](HUMAN_FORMAL_VALIDATION_20261024.md) | +-| HY-Motion | 共同英文文本计划接原生Qwen3 token与CLIP-L全局特征;流式历史条件需显式扩展 | 原生201D人体运动表示;连续输出、历史帧数与mask尚未验收 | 已有模型环境/调研;本地完整HFD训练尚未接通,不声称已按此协议训练 | +-| Kimodo-SMPLX | 已释放英文文本→冻结LLM2Vec,最近400正文token,mean pooling为1×4096;60帧/2秒人体历史@30Hz;203维历史姿态约束,无音频;当前配置采用因果root分解 | root→body两阶段,SMPLX22的273D表示;默认未来60帧/2秒,执行12帧/0.4秒;中性手部,保留源体型;12帧预测诊断未优于默认 | 修复版100K已完成,best即100K;冻结根骨干仅训练根投影/身体,见[训练修复](KIMODO_TRAINING_REPAIR.md)及[新评测](HUMAN_FORMAL_VALIDATION_20261024.md);未宣称闭环质量已达标 | +- +-30秒是候选文本范围,不等于30秒所有文字都进入编码器。TextOp的75个正文token +-加边界共77;T5为原生50 token。当前各自编码和缓存embedding,不共享不同模型的向量。 +-Kimodo的400正文token进入原生512总预算及模板;独立文本worker验证双向注意力和两份适配器权重。HY-Motion的token预算和前缀长度仍待实际adapter固定。 +- +-Kimodo使用归一化x0主损失,历史70维未观察辅助量在每个DDIM步保持训练噪声分布,避免未监督历史输出继续回灌。2026-10-23修复版增加原始根历史重建及实际解码运动的连续性监督;根/身体分别优化和裁剪,冻结根注意力骨干,未增加未来观测或推理修正。早期2K+1K、在线历史课程属于历史诊断。全量配置从完整英文索引按窗口均匀边缘抽样,不采用诊断中的70%活动窗口筛选;文本worker与动作模型同卡、独立进程,已编码特征在本地SQLite复用。共同文本释放规则及其余四模型输入未更改。 +- +-TextOp每8帧重新查询当前可见文字,文本本身只在0.4秒释放点改变;不同模型的执行周期 +-不必等于文本释放周期。其原生历史特征含相邻帧差分,2帧特征需读到当前决策状态, +-不能简单理解成两个完全独立的旧姿态。TextOp原生57D包含根姿态/位移、23个关节、 +-关节差分和接触;沿用官方统计与VAE,不直接拿G1原始36D训练。 +- +-OMG的60×35声学位置来自已经截取的过去PCM,不是未来2秒音频;文本和音频释放时钟一致。 +-初始不足两秒的音频左补零。TextOp和GenMo主实验没有声学分支,因此在文字出现前 +-没有显式语音起点、韵律或背景说话信息,不能要求它们仅靠文字精确判断倾听开始。 +- +-## 人体数据与session构造 +- +-源人体资产包含 `trans[T,3]`、`root_orient[T,3]`、`pose_body[T,63]`、 +-`pose_hand[T,90]`、`betas[16]`、`sample_times`。它们是源动作,不是完整全双工人体session。 +-以 `metadata/seed_g1_smplx_pairs.parquet` 的motion_id连接源,按原生配方的真实选取、 +-裁剪、持姿、朝向/位置衔接、过渡和事件时间构造人体时间线。 +-当前实现仍借用原G1 session组合程序及配对表追踪计划;不是从G1反解人体。 +-G1关节不是GenMo或Kimodo网络输入或训练损失目标。 +- +-GenMo训练实际使用trans、root_orient、21个身体关节及前10维betas; +-16维体型还用于SMPL-X pelvis/坐标转换。手指与面部不作为生成目标,推理导出中手部为中性残差。 +-151D由21×6D身体旋转、10D体型、两组6D根朝向和3D根局部相邻帧位移构成。 +-位移编码没有乘FPS,不能直接称为m/s;坐标转换和推理积分必须互为逆变换。 +- +-原生源动作约50Hz对齐成完整session后,按时间与SO(3)插值重采样到30Hz。 +-输入历史不能通过插值使用决策之后的GT;GenMo历史单独编码,最后一个历史帧指向 +-第一未来帧的位移未知,保留为生成目标。已知历史锁定,无下一有效帧的位移/静止标签不计损失。 +- +-## 数据范围与缓存 +- +-| 模型范围 | 英文train候选session | 英文val候选session | +-| --- | ---: | ---: | +-| OMG/TextOp:原生+已选第三方 | 168,583 | 5,109 | +-| GenMo:仅原生配对人体 | 155,117 | 4,719 | +- +-第三方生产包有G1时间线,但尚无经验证的配对人体session;GenMo排除该部分。 +-因此不能直接把OMG/TextOp与GenMo的结果差异完全归因于模型结构。 +-原生索引/seed保持,英文过滤不重排session ID;原生任务源train/val隔离。 +-当前没有独立test集,代码中test别名val不能当作未触碰测试集;预训练/共享stationary源重叠未排除。 +- +-离线完成的是源英文语音词对齐及索引,不是提前导出全部人体session。 +-OMG/TextOp/GenMo正式训练均按需组装原生session,组装结果放有上限的进程内缓存,退出即失效; +-第三方则读取已组装生产session。GenMo额外缓存30Hz重采样,train/val各512MiB, +-冻结文本特征共享256MiB CPU LRU;这不是全量持久session缓存。 +-原始tar在共享盘保留,本地只读索引镜像减轻共享盘随机读取。 +-采样有放回、槽位复用,step/epoch不等于逐条无重复遍历整个候选集。 +- +-## 训练与评测 +- +-正式配置:`configs/omg.yaml`、`configs/textop.yaml`、`configs/genmo.yaml`。 +-OMG预算100K、TextOp三阶段合计300K、GenMo100K;不按步数直接比较训练数据量。 +-OMG/GenMo训练使用GT历史,TextOp沿用原生分阶段GT/rollout/static-history课程。 +-正式连续评测只做一次初始GT bootstrap,之后使用预测历史;重复喂GT的结果必须标为oracle诊断。 +-TextOp当前根位置与朝向锚点取已知初始决策状态,不能每个primitive重新用GT根状态对齐。 +- +-评测入口:`evaluate_omg.py`、`evaluate_textop.py`、`evaluate_genmo.py`。 +-OMG/TextOp直接输出G1,不需要人体retarget。 +-人体模型先比较人体关节旋转、根位置、根朝向和根速度,再对整条预测及整条人体GT各做一次 +-SMPL-X→G1 retarget,不能在每个规划窗口重置IK。三种比较分别是预测→G1/原G1、 +-人体GT→G1/原G1、预测→G1/人体GT→G1;重定向误差不能直接从模型误差中相减。 +-统一保存原混合音轨、microphone前置帧、文本释放审计和bootstrap排除范围。 +- +-验证loss用于监测或选权重,但不等于连续生成质量或语义成功率。特别是GenMo原生位移loss +-使用GT朝向积分,闭环使用预测朝向,必须同时看朝向和长时漂移。 +-当前视频是运动学回放,不代表物理控制跟踪、独立任务成功率或真实端到端停止延迟。 +-GenMo的双分支梯度/因果性与44项检查见[正式训练说明](GENMO_FORMAL_TRAINING.md)。 +- +-更多入口:[共同文本条件](BASELINE_TEXT_CONDITION.md)、[session格式](../SESSION_FORMAT.md)、 +-[动态reference组装](REFERENCE_DYNAMIC_TRAINING.md)、[GenMo正式训练](GENMO_FORMAL_TRAINING.md)、 +-[人体retarget](HUMAN_RETARGET_EVALUATION.md)。历史研究及社区训练代码核查仍见归档, +-不把当年的待实现建议、旧ASR输入或旧缓存描述当作当前已经验收的实现。 +- +-## GenMo v2训练适配补充(2026-10-24) +- +-输入/输出与本协议相同。训练额外使用一段模型预测历史作为因果恢复条件,GT未来仅作监督,前0.4秒增加损失权重;这不增加模型推理时可见的信息。默认最佳权重用自主闭环验证选择,执行细节见 [GENMO_V2_TRAINING.md](GENMO_V2_TRAINING.md)。旧配方保留在 `configs/genmo_legacy.yaml`。 +diff --git a/docs/HFD_HUMAN_BASELINE_SOURCES.json b/docs/HFD_HUMAN_BASELINE_SOURCES.json +index 8752b9f..5cf7f2f 100644 +--- a/docs/HFD_HUMAN_BASELINE_SOURCES.json ++++ b/docs/HFD_HUMAN_BASELINE_SOURCES.json +@@ -175,10 +175,6 @@ + } + ], + "local_evidence": [ +- { +- "path": "/root/autodl-tmp/hfd-baseline-portable/SESSION_FORMAT.md", +- "sha256": "b478da021db41576b023de4ee057350e93539b89a9427fa680f16004abfb226b" +- }, + { + "path": "/root/autodl-tmp/hfd-baseline-portable/README.md", + "sha256": "f2b9adf6cf12bcbe7ebd3af72e74355d7a1910f0238c2744571a03a6193c7f40" +diff --git a/docs/HUMAN_FORMAL_VALIDATION_20261024.md b/docs/HUMAN_FORMAL_VALIDATION_20261024.md +deleted file mode 100644 +index 531d697..0000000 +--- a/docs/HUMAN_FORMAL_VALIDATION_20261024.md ++++ /dev/null +@@ -1,38 +0,0 @@ +-2026-10-24:GenMo、修复版 Kimodo 均已完成 100K 步;本轮按既有六类 native English validation 规则执行闭环生成、整段 GMR 和有声视频。没有训练、修改或删除正式权重。 +- +-输入为按既定词时间释放的最近30秒英文GT文本与60帧已执行人体历史,无音频输入。只使用第一帧GT初始化,随后完全反馈预测;规划60帧、执行12帧、30Hz。Kimodo采用DDIM20/CFG2,GenMo采用DDIM10/CFG2.5,均seed12345。人体GT和完整预测分别连续retarget一次,无轨迹对齐、DTW或位置修正。 +- +-GenMo依原入口每类选首个预计6–30秒的样本;Kimodo依原入口选6–20秒。打断、急停的样本ID不同,其余四类相同。所有样本均在生成前确定。下表用于比较各模型微调前后,不能据此直接排名架构;也不是全HFD验证集、独立test或物理跟踪/任务成功率评测。 +- +-| 模型/权重 | 人体姿态RMSE(rad) | 人体根位置RMSE(m) | G1关节RMSE(rad) | G1根位置RMSE(m) | G1根朝向RMSE(rad) | +-|---|---:|---:|---:|---:|---:| +-| GenMo official | 0.651 | 0.995 | 0.688 | 0.864 | 2.342 | +-| GenMo best92k | 0.407 | 0.701 | 0.373 | 0.604 | 0.373 | +-| GenMo last100k | 0.427 | 0.758 | 0.376 | 0.655 | 0.390 | +-| Kimodo official | 0.627 | 0.502 | 0.464 | 0.433 | 0.575 | +-| Kimodo last100k | 0.422 | 0.306 | 0.373 | 0.264 | 0.487 | +- +-六段分别评分后等权平均,排除GT初始化帧。人体姿态是21个身体关节的SO(3)角误差RMSE;根位置为三坐标逐元素RMSE。G1指标参考同流程retarget的human GT,原始G1仅作另一个参考;retarget误差不从模型误差中相减。 +- +-GenMo:优先保留92K的best作为当前评测候选。固定去噪验证loss在92K最低(2.409),100K为2.464;本轮六段连续生成的人体姿态、根位置及根速度误差也均为92K略低。训练耗时约85.8小时。批量评测复用模型但恢复原始初始化后的采样RNG;官方和92K各抽一个固定样本与未修改的infer_genmo.py命令复核,人体位置/根旋转/身体旋转均逐位相同。 +- +-Kimodo:best与last均为100K,训练耗时约18.7小时,固定验证loss为1.179→0.0669。根骨干200个参数张量仍与官方初始化逐位一致;同一后退诊断窗口t=0的末层时序RMS为0.0360、输入梯度范数0.00586,旧失效100K分别约1.08e-7和4.76e-12。本轮未重现此前根网络恒定输出失效。 +- +-与上一轮失败的Kimodo 100K历史报告相比,G1根位置误差从0.544降至0.264m,人体窗口首帧根速度误差从约0.839降至0.109m/s;但G1根朝向误差从0.266升至0.487rad。修复包含可训练参数范围、学习率、损失和注意力后端的组合修改,不能把这些收益归因于某个单独超参数。 +- +-两者相对官方初始化均有平均误差收益,但动作质量仍有明显限制:后退、长距离移动与部分指令跟随不足。backchannel样本的人体GT净后退位移约3.50m,Kimodo仅沿GT后退方向移动约0.37m;GenMo该样本的整段根路径只有约0.14m,而GT约3.65m,接近保持位置。低平均误差不能当成动作任务已经完成。 +- +-保持初始姿态的对照也保存在JSON中。Kimodo新100K的人体姿态平均误差0.422略低于静止对照0.431,但G1关节0.373仍略高于静止对照0.366。GenMo92K的人体姿态0.407低于静止对照0.429。静止对照无法执行指令,只用于识别静态区间对平均分的影响。 +- +-本轮导出42段有声MP4:Kimodo 18段、GenMo 24段。均核对帧数、30Hz、音轨和完整音视频解码;抽帧检查六类预测/参考、文本与时钟显示。runs只保留视频,工作数据和日志在cache。 +- +-- `/root/autodl-tmp/runs/kimodo-formal-validation-20261024/last100k/`:六类预测→G1 / human-GT→G1及六段retarget参考;`official/`含六段官方初始化对照。 +-- `/root/autodl-tmp/runs/genmo-formal-validation-20261024/best92k/`、`last100k/`、`official/`:各六段预测→G1 / human-GT→G1;last100k另含六段GT retarget参考。 +-- `retargeted_human_comparison.mp4`左侧预测,右侧同流程retarget的human GT;`retarget_reference.mp4`比较human-GT→G1与原始G1。 +- +-完整报告与曲线: +- +-- [Kimodo数值、逐例结果、根骨干检查和视频校验](../reports/kimodo_formal_validation_20261024.json) +-- [GenMo数值、逐例结果、命令一致性和视频校验](../reports/genmo_formal_validation_20261024.json) +-- [Kimodo验证曲线与后退轨迹](../reports/kimodo_formal_validation_20261024.png) +-- [GenMo验证曲线与后退轨迹](../reports/genmo_formal_validation_20261024.png) +diff --git a/docs/HUMAN_RETARGET_EVALUATION.md b/docs/HUMAN_RETARGET_EVALUATION.md +deleted file mode 100644 +index 3176e0b..0000000 +--- a/docs/HUMAN_RETARGET_EVALUATION.md ++++ /dev/null +@@ -1,98 +0,0 @@ +-# 人体模型的 retarget 评测入口 +- +-`evaluate_human.py` 共用于 Kimodo、GenMo、HY-Motion。输入为模型导出的 **完整连续 rollout**,先转换为标准 SMPL-X / AMASS,再通过 GMR 转到 G1。该入口不加载生成模型、不下载权重,也不改变现有 TextOP / OMG 推理逻辑。 +- +-2026-10-21更新:GenMo已接通动态原生英文人体训练及`evaluate_genmo.py`连续推理、完整轨迹GMR与带音轨视频;Kimodo已增加固定原生英文子集训练与`evaluate_kimodo.py`完整轨迹导出、GMR和视频。两者当前均输入已释放英文GT文本+60帧人体历史,不输入音频。Kimodo范围见[训练说明](KIMODO_HFD_TRAINING.md),HY仍待接入。共同输入协议见[当前约定](HFD_HUMAN_BASELINE_INPUT_PROTOCOL.md)。GMR位于 `/root/autodl-tmp/hfd-human-models/gmr-runtime`,主体资产在fs的`models/body_models`。 +- +-## 同时保存三组结果 +- +-| report.json 指标 | 左侧 | 右侧 | 用途 | +-| --- | --- | --- | --- | +-| `prediction_vs_original_g1` | 预测 SMPL-X→G1 | 原始 G1 真值 | 与机器人 baseline 相同目标上的轨迹误差 | +-| `human_gt_vs_original_g1` | human 真值→G1 | 原始 G1 真值 | 检查体型、session 几何渲染及 retarget 自身的偏差 | +-| `prediction_vs_retargeted_human_gt` | 预测 SMPL-X→G1 | human 真值→G1 | 相同转换后的预测误差 | +- +-第二项是转换链路的诊断,不能从第一项中直接相减,也不是数学意义的误差下界。所有指标包含 29 关节 RMSE(rad)、根位置每坐标 RMSE(m)、根朝向 SO(3) 测地角 RMSE(rad)。不做 DTW、逐帧根对齐、截短或补齐。`bootstrap_frames` 在三组评分中一致排除。teacher-forced 必须标为 oracle,不能作为闭环结果。 +- +-此外,`report.json.human_metrics`在retarget前直接比较人体身体旋转、根位置、根朝向及根速度,排除同一bootstrap。根速度按实际FPS计算,单位m/s;人体关节角采用SO(3)测地角。优先用人体自身指标判断生成误差,G1比较辅助诊断转换与机器人覆盖差异。 +- +-一个连续序列仅创建一次 GMR retargeter,内部状态跨帧保留;预测和 human GT 分别新建,不能共享 IK 状态,也不能每个 0.4s 规划块重置。保持旧 benchmark 的 `offset_to_ground=False` 和身高估计 `1.66 + 0.1 * beta[0]`;体型使用 16 个 betas、`use_pca=False`、`flat_hand_mean=False`。不根据未来 GT 调整预测的朝向或地面。该身高估计是否适用于当前体型应在真实集成测试中核对。 +- +-## 输入契约 +- +-路径可相对 manifest 所在目录。manifest 的三份动作文件必须携带同一个已冻结的 `session_plan_id`。**这个 ID 只用于发现误配,脚本不能证明 human/G1 已按同一个 plan 正确渲染**;生成端必须先完成该保证。不要仅凭相同随机种子或相同动作名配对。 +- +-预测和 human 真值两个 NPZ 都必须包含: +- +-| 字段 | 要求 | +-| --- | --- | +-| `trans` / `root_orient` / `pose_body` | `[T,3]` / `[T,3]` / `[T,63]`,轴角,有限值 | +-| `betas` | 恒定 `[10]` 或 `[16]`,10 维补零至 16;不从 GT 偷换预测体型 | +-| `pose_hand` | 可选 `[T,90]` 左右手轴角;相对 SMPL-X 手均值的残差;缺省零表示模型默认手姿态 | +-| `gender` | 可选 `neutral` / `male` / `female`,缺省 neutral | +-| `mocap_frame_rate` | 正标量;预测和 human GT 必须相同 | +-| `coordinate_system` | 字符串 `amass_z_up`;`trans` 是模型平移,不是 pelvis 世界坐标 | +-| `sample_times` | 可选;若存在必须等于 `arange(T)/fps`,时间不均匀时先在导出端显式重采样 | +-| `session_plan_id` | 冻结 plan 的标识,推荐使用内容哈希 | +-| `start_seconds` | 该 rollout 第 0 帧在 session 中的时间,与 manifest 一致 | +- +-Kimodo 的 Y-up joint positions / rotations、HY 的 201D 特征、GenMo 的 151D 特征均不能直接送入此入口,须通过各模型原生解码与坐标逆变换导出 AMASS。`trans` 不做隐式 pelvis 偏移修正。非零 `pose_jaw` / `pose_eye` 暂不支持。预测与 human GT 的体型/性别是否相同会记录在 provenance,差异需要在实验说明中解释。 +- +-原始 G1 session NPZ 必须包含: +- +-- `layout="xyz_wxyz_native29"` + `qpos_36[T,36]`,或 `layout="hfd_isaac29_xyz_wxyz"` + `motion[T,36]`。后者走现有按关节名排列的转换,不能直接切片当 MuJoCo qpos。 +-- `fps`、`session_plan_id`、`start_seconds=0`。 +-- `audio`(一维PCM)、`audio_sample_rate=16000`、`audio_start_seconds`。当前`microphone`协议下原PCM起点填0,动作时刻t显示在麦克风t+0.2秒;从动作0开始的回放给两侧共同增加0.2秒前置姿态。不能把当前PCM起点机械填成-0.2或再次偏移。历史协议需按原manifest显式复现。 +- +-G1 真值按真实时间线重采样到 human rollout 帧率:位置/关节线性插值,根四元数 SLERP。覆盖不足立即报错;音频也必须覆盖完整的 `T/fps` 播放区间。当前入口面向均匀 rollout,不接受拼接成连续序列的 disconnected windows。 +- +-manifest 示例(示意路径,不是现有数据): +- +-```json +-{ +- "schema_version": 1, +- "model": "genmo", +- "session_plan_id": "sha256-of-frozen-plan", +- "prediction": "prediction_amass.npz", +- "human_reference": "human_gt_amass.npz", +- "g1_reference": "g1_session.npz", +- "start_seconds": 0.0, +- "bootstrap_frames": 1, +- "history": "predicted", +- "clock_policy": "microphone", +- "caption_policy": "aligned_reference_context", +- "text_conditioning": "none", +- "text_source": "oracle_foreground_transcript", +- "inputs": {"text": "released English reference text <=30s", "history": "60-frame history; initial pose held at startup, then predictions"}, +- "audio_policy": "Aligned reference text <=30s; no raw audio conditioning", +- "use_asr_text": true, +- "command_seconds": [0.4, 5.2] +-} +-``` +- +-`command_seconds` 仅供叠字展示,不用于生成或 retarget。输入策略字段是生成端声明,评测脚本不能据此证明模型没有使用未来信息。 +- +-## 执行与渲染 +- +-在具有 NumPy / SciPy 的模型环境中,先验证导出数据: +- +-```bash +-python evaluate_human.py --manifest /path/to/manifest.json \ +- --output-dir /root/autodl-tmp/runs/kimodo-eval \ +- --gmr-root /path/to/GMR --body-models /path/to/body_models --preflight +-``` +- +-preflight 检查文件、布局、时间覆盖和 plan ID,只报告 GMR/资产目录是否存在;不会导入 GMR 或验证资产许可证、内容、依赖是否完整,也不写评测目录。具备真实依赖后去掉 `--preflight` 执行转换。输出目录已存在时拒绝覆盖。 +- +-输出 `report.json`、三组配对 NPZ,以及汇总 `retargeted.npz`(预测、human GT、原始 G1 GT 和绝对 timestamps),包含输入 SHA256、GMR commit/dirty 状态、路径、转换配置和体型一致性。真实资产存在时记录主体 NPZ 的 SHA256,并记录相关包版本;GMR 工作树仍需保持固定。 +- +-现有渲染环境中执行: +- +-```bash +-python render_overfit.py --inference /root/autodl-tmp/runs/kimodo-eval \ +- --assets-root /path/to/OMG/assets --gpu 0 +-``` +- +-三组视频均使用原 session 音频;标题区分预测、human GT retarget 和原始 G1 真值。视频为参考姿态播放,不是物理跟踪结果。human GT 转换对照不显示模型 ASR 输入或预测历史提示。 +- +-开发验证:`python -m unittest tests.test_human_eval tests.test_joint_mapping -v`。合成测试覆盖三组指标、bootstrap 排除、序列级后端调用、拒绝误配/不均匀时间/覆盖不足、四元数双覆盖和 SLERP,以及 HFD→G1 关节名映射。 +diff --git a/docs/KIMODO_CLOSED_LOOP_STABILITY.md b/docs/KIMODO_CLOSED_LOOP_STABILITY.md +deleted file mode 100644 +index ca51599..0000000 +--- a/docs/KIMODO_CLOSED_LOOP_STABILITY.md ++++ /dev/null +@@ -1,144 +0,0 @@ +-# Kimodo 训练策略与闭环稳定性实验 +- +-本轮从官方权重重新构建实验。此前及本轮的测试权重均已清理,复现需重训;报告、代码和视频保留,正式GenMo训练不受影响。 +- +-## 当前保留的诊断路线与最终状态 +- +-**训练与推理的关键一致性问题已修正,但六种模式混合闭环过拟合、少样本验证均未通过;不启动全量正式Kimodo训练。** +- +-目前较好的诊断路线是:官方初始化→归一化x0目标训练2,000步→只加载模型权重、重置Adam,再用GT历史训练1,000步。输入协议不变,不增加GT阶段、语音结束状态或动作时钟。两个公开配置为`configs/kimodo_overfit.yaml`与`configs/kimodo_refine_probe.yaml`,准备/复用输入与续训命令见[训练说明](KIMODO_HFD_TRAINING.md)。首阶段配置与实测`normalized-x0`除输出路径外完全相同。 +- +-| 修正采样后的实验 | 累计步数 | 训练身体RMSE(rad) | 训练根RMSE(m) | 训练根速度RMSE(m/s) | +-| --- | ---: | ---: | ---: | ---: | +-| 物理Huber,同一模型旧历史反馈规则 | 2,000 | 0.3703 | 0.2580 | 0.4307 | +-| 物理Huber,同一模型修正历史反馈规则 | 2,000 | 0.3747 | 0.2137 | 0.3638 | +-| 归一化x0,50%高噪声/70%运动窗口 | 2,000 | 0.3822 | 0.2055 | 0.4102 | +-| 上述模型再用GT历史训练1,000步 | 3,000 | **0.3105** | **0.0915** | **0.3183** | +-| 同一初始化,再用在线历史训练1,000步 | 3,000 | 0.4355 | 0.2826 | 0.5754 | +-| 归一化x0,均匀扩散步/均匀窗口 | 2,000 | 0.3898 | 0.5479 | 0.3337 | +- +-第一、二行已核对checkpoint SHA相同,唯一改动是推理历史辅助量规则:根位置/根速度误差分别降低约17%/16%,身体误差没有改善。第四、五行已核对初始化权重相同、各追加1,000步;在线课程更差,**不采用为默认**。归一化x0在2,000步时并未全面优于物理Huber,不能把第4行的改善全部归因于损失更换,训练预算和优化器重启也改变了。 +- +-GT历史3K模型的六条结果如下,均为完整预测历史rollout: +- +-| 模式 | 身体RMSE(rad) | 根RMSE(m) | 预测/GT身体角速度比 | +-| --- | ---: | ---: | ---: | +-| turn-taking | 0.4387 | 0.0398 | 1.747 | +-| interruption | 0.3192 | 0.0686 | 1.165 | +-| emergency-stop | 0.0564 | 0.0396 | 1.149 | +-| backchannel | 0.3220 | 0.0236 | 1.107 | +-| hesitation | 0.3703 | 0.1470 | 1.103 | +-| pause | 0.3564 | 0.2307 | 1.148 | +- +-逐条精确值以[机器报告](../reports/kimodo_closed_loop_stability_20261021.json)为准。行走/停止已能较准确复现,其他模式虽运动幅度不再明显不足,动作相位和多余运动仍存在;幅度正确不等于动作正确。六条验证session宏平均为身体0.5220rad、根0.8850m、根速度0.3673m/s,身体只有1/6好于冻结初始姿态,泛化没有通过。当前训练入口仍是固定小规模子集,不是全库动态训练器。 +- +-优先查看本轮较好模型的三组预测↔人体GT视频(双方各完整retarget到G1,均含音轨): +- +-- [行走/停止](/root/autodl-tmp/runs/kimodo-closed-loop-20261021/score-control/train-emergency-stop-000036/retargeted_human_comparison.mp4) +-- [动作切换,仍有相位偏差](/root/autodl-tmp/runs/kimodo-closed-loop-20261021/score-control/train-turn-taking-000013/retargeted_human_comparison.mp4) +-- [犹豫,仍未准确复现](/root/autodl-tmp/runs/kimodo-closed-loop-20261021/score-control/train-hesitation-000087/retargeted_human_comparison.mp4) +- +-## 修正历史辅助量的推理反馈 +- +-训练时,历史未观察的70维速度/接触辅助量始终是`epsilon * sigma_t`;它们不参与历史损失。旧推理却把模型对这些未监督历史字段的预测继续经DDIM反馈,低噪声阶段也保留其信号,与训练不一致。 +- +-当前`inference.prefix_aux_policy: training_noise`在每次去噪调用前恢复相应扩散步的历史纯噪声边缘分布,然后重新锁定已知203维姿态。噪声来自本轮初始epsilon,没有额外GT或随机数;未来采样变量不被这一步改写,也不对输出做平滑。训练与20步稀疏DDIM的sigma映射有独立一致性检查,旧行为仅通过`legacy`供对照复现。在线历史刷新也使用同一规则。 +- +-新增`normalized_x0`主损失直接在原生归一化特征上做Gaussian x0最小二乘;root/body两阶段分别平均,已知/填充帧和未定义末端速度排除。它是训练目标对照,不宣称复现未公开的官方训练配方。`physical_huber`保留供历史结果对照。 +- +-当前代码41项测试通过;最新完整模型续训重放含修正后的采样与在线缓存,在确定性模式下逐位一致。该完整重放使用physical_huber分支;normalized_x0共用持久化路径,并完成了实际全模型训练与梯度检查。本轮共16条训练路径(含共享初始化和后续阶段),不是16次独立随机实验。以下为历史实验明细;旧表均使用修正前的`legacy`历史辅助量反馈,重新运行需显式指定该规则。 +- +-## 视频审计修正 +- +-快速probe最初没有导出每轮`executed_frames`,导致显示层误报“文本未释放”。模型实际使用的caption、运动轨迹和误差均未受影响。现已补齐字段,核对执行帧总数等于预测总长减去bootstrap,并重渲染全部72个视频。现有retarget结果未重新计算;报告记录了仅显示元数据修正及前后manifest哈希。 +- +-## 改动与对照原则 +- +-输入协议不变:按时释放的最近30秒英文GT文本、60帧已知人体历史、预测60帧、执行12帧、无音频。推理仅用一帧GT初始化,后续全部接自己的输出,不加入GT阶段、未来指令、动作时钟或目标姿态,也不做平滑后处理。 +- +-`hfd_bridge/kimodo_closed_loop.py` 实现两项可独立开关的训练策略: +- +-1. `closed_loop.tail_weight`:实际执行的前12个未来帧权重为1,余下48帧为较低的正权重。无效填充帧依然排除,分母使用有效权重之和。这个实验检验远期监督是否稀释当前决策,不代表已经证明未来文字是退化的唯一原因。 +-2. `closed_loop.online_history`:按 `refresh_every` 用当前模型重新闭环生成训练session,混入原GT历史窗口。`max_probability` 在 `ramp_steps` 内逐渐增加。刷新使用与推理相同的DDIM步数、CFG、文字释放时钟和单帧初始化;只刷新训练集,不对验证集训练。预测历史只作为条件,GT未来保留原世界坐标,不再平移根位置或混合前6帧目标姿态。 +- +-历史刷新不反传穿过整个采样轨迹;随后训练步对网络两阶段正常反传。独立的历史选择随机数不会改变原窗口抽样顺序,刷新会恢复训练的全局随机数和train/eval模式。完整checkpoint包含历史缓存、缓存对应的policy step和历史选择随机数;恢复训练不重新生成另一份历史替代原缓存。 +- +-原 `prepare_kimodo_predicted_history.py` 为历史实验保留,不作为本轮在线策略入口。 +- +-可选 `reset_schedule` 只控制训练历史课程:例如 `[2,5,0]` 在前两阶段每2轮、5轮从GT过去重新启动短rollout,最后阶段从一个GT初始姿态生成完整session。每个窗口保存模型当时实际看到的60帧历史,不能从包含后续GT重启的拼接轨迹倒推历史。正式闭环评测入口没有这个开关,始终不重启、不补GT历史。`curriculum_stage_steps` 决定阶段长度,实际历史在下一次刷新时切换。 +- +-## 固定实验 +- +-数据保持六种模式各一条训练session与验证session,分别205与215个窗口。所有数据来自原生英文配对人体session。81种合法释放文本使用同一份冻结LLM2Vec缓存。 +- +-- 先按原方案训练2,000步,作为公共初始化。 +-- 三组分别从同一初始化再训练1,000步:原方案对照;`tail_weight=0.1`;`tail_weight=0.1`加在线历史(最大概率0.5,500步渐增,每250步刷新)。 +-- 三组保持batch8、LR1e-5→1e-6、随机种子、窗口抽样顺序、原七项损失、文本dropout0.1、DDIM20、CFG2一致。本轮不同时叠加高噪声抽样、活跃采样或额外速度损失。 +-- GT历史去噪loss只用于诊断,不能跨不同权重目标直接比较。主要比较六条训练session的完整闭环误差、静止基线、角速度幅度、根速度误差和重规划加速度误差;验证集另列,不作为训练过拟合的替代。 +- +-前三组之后增加一个同预算的课程恢复候选:短历史→长历史→完整历史,50%样本t=999,并在root输出进入body阶段前锁定已知root历史。它是组合配方,不把其结果归因于单一因素;仍不增加活跃抽样或额外速度损失。 +- +-快速多session评测入口为 `tools/probe_kimodo_closed_loop.py`,导出的manifest兼容 `evaluate_human.py` 和 `render_overfit.py`。`--teacher-forced`只能生成明确标记的诊断结果,不计入闭环成绩。每次评测必须使用新目录,避免误用其他权重的旧输出。 +- +-工作数据、临时日志与权重分别在 `cache/kimodo-stability-20261021`、`setup/kimodo-stability-*` 和 `/dev/shm/kimodo-stability-*`,不重新往 `runs` 堆积中间数据。`runs`只保存最终对照视频。 +- +-## 结果 +- +-**结论:混合闭环过拟合仍未通过,不推荐启动正式Kimodo训练。** 已落地表示一致性、导数标签与续训状态修正;在线历史、执行帧加权、动态损失和短预测时长没有形成可靠的混合训练配方。 +- +-下表为原始六条训练session完整预测历史rollout的等权宏平均,全部仅第一帧GT初始化、固定seed12345、DDIM20、CFG2,无后处理。累计步数包含公共初始化,表中不同累计预算不能解释为单因素同预算对照。 +- +-| 配方 | 累计更新 | 身体RMSE(rad) | 根RMSE(m) | 根速度RMSE(m/s) | +-| --- | ---: | ---: | ---: | ---: | +-| 原方案公共初始化 | 2000 | 0.4785 | 0.5938 | 0.2945 | +-| 原方案继续训练 | 3000 | 0.4186 | 0.5964 | 0.2944 | +-| 执行帧加权 | 3000 | 0.4353 | 0.5961 | 0.2946 | +-| 加权+整段在线历史 | 3000 | 0.3540 | 0.5789 | 1.2121 | +-| 高噪声+历史课程 | 3000 | 0.3673 | 0.2337 | 0.4767 | +-| 强动态监督/GT历史 | 5000 | 0.3583 | 0.2913 | 1.1430 | +-| 强动态监督/课程历史 | 5000 | 0.3470 | 0.3921 | 0.4033 | +-| 因果编码+强动态监督/GT历史 | 5000 | 0.4352 | 0.4061 | 0.9534 | +-| 因果编码+较轻动态监督/课程历史 | 6000 | 0.3231 | 0.4638 | 1.0686 | +-| 同预算追加/12帧预测 | 7000 | 0.3854 | 0.4401 | 1.3143 | +-| 同预算追加/60帧预测 | 7000 | 0.3387 | 0.3743 | 1.0379 | +- +-较轻课程组身体误差比3K原方案低约23%,但根速度误差从0.2944升至1.0686m/s,不能以身体误差下降宣称修复。其行走/停止样本根路径21.83m,GT5.29m,根速度误差4.7618m/s;backchannel运动速度仍只有GT约15%。强动态监督的在线版本相对稳定一些,但hesitation/pause仍偏静止、行走仍有额外速度。因果root编码解决了实际表示不一致,却没有独立带来混合动作质量提升。 +- +-12帧与60帧预测从同一个较轻课程组最终权重开始,均追加1,000步GT历史训练,LR1e-5→1e-6、batch8、运动抽样70%,仅预测长度不同。12帧组的三项宏平均均未优于60帧组,故默认保留60帧。该对照不是把不同预算或不同初始化混为预测时长的效果。 +- +-较轻课程组的六条验证session宏平均为身体0.4430rad、根0.4699m、根速度0.1859m/s,仅作小规模诊断,不是全库泛化或独立测试成绩。CFG1训练宏平均为0.3747rad/0.4654m/1.2224m/s,也没有改善本轮问题;验证结果与逐条指标见机器报告。 +- +-此处为旧采样规则下的历史对照;当前诊断配置和修正后的采样结果见文首。在线历史、执行帧加权及12帧预测未采纳为默认;其他四个baseline配置未改动。 +- +-[完整配置、逐条指标与检查记录](../reports/kimodo_closed_loop_stability_20261021.json)。旧采样规则阶段共11条训练路径(包含公共初始化),有共享初始化关系,不是11次独立随机实验。 +- +-视频均含原始音轨,分别先对完整预测和人体GT做一次GMR,再比较G1轨迹;不是逐窗口重置retarget。示例: +- +-- [原方案:行走/停止](/root/autodl-tmp/runs/kimodo-closed-loop-20261021/control/train-emergency-stop-000036/retargeted_human_comparison.mp4) +-- [较轻课程:行走/停止,仍有抖动](/root/autodl-tmp/runs/kimodo-closed-loop-20261021/causal-online-lite/train-emergency-stop-000036/retargeted_human_comparison.mp4) +-- [较轻课程:犹豫样本](/root/autodl-tmp/runs/kimodo-closed-loop-20261021/causal-online-lite/train-hesitation-000087/retargeted_human_comparison.mp4) +-- [12帧预测诊断:犹豫样本](/root/autodl-tmp/runs/kimodo-closed-loop-20261021/horizon12/train-hesitation-000087/retargeted_human_comparison.mp4) +- +-保留范围:报告和代码在仓库,工作缓存位于cache;runs中本轮只新增MP4。所有本轮过拟合权重已清理,正式OMG/TextOp权重及GenMo活跃输出保留。旧中间初始化权重不再可直接复用,复现需按报告中各阶段重建。 +- +-## 表示与速度监督的额外修正 +- +-复查原205个训练窗口,独立编码的历史末帧与整段监督中的同一帧,其虚拟平滑root平均相差0.0213m、最大0.1492m。这是特征分解差异,不是人体GT发生了15cm跳变;但把两份表示在边界拼接后,会影响root→body阶段计算的局部root速度。原先只检查“未来不能改变observed”,没有检查“known部分与target使用同一分解”。 +- +-新增 `representation.causal_root: true`:在原273D表示中使用单向、半径0.06m受限的根轨迹滤波(alpha0.2),并同步重算hips相对位移。FK、全局旋转、人体root物理位置及预训练统计保持一致。此滤波只用于特征分解,不对预测轨迹做事后平滑。前203维已知姿态在prefix与整段编码中一致,改变未来不改变输入,解码后保持原人体运动;已对全部420个train/val窗口检查。 +- +-`hfd_bridge/kimodo_causal_rep.py` 是可选编码实现;准备数据、在线历史、单session推理、多sessionprobe均读取相同配置。启用或切换编码后必须重新准备windows,不能直接把旧编码缓存标成新编码。严格训练合同会拒绝代码或数据已经变化的旧resume。 +- +-对于预测历史,位置目标仍为原绝对GT未来;另外的 `temporal_target` 只提供真实GT速度/加速度标签。构造这些标签时使用最后3个GT过去姿态,但它们不进入observed/mask或网络条件。这样不会把“从错误预测位置瞬间跳回GT”的位移当成期望速度。原GT历史样本的temporal target就是原target。新增边界root速度及实际解码root加速度监督,未配置的权重为0。 +- +-两组动态监督对照均从课程组第1,000步初始化,再训练2,000步、LR3e-5→3e-6:GT历史组与在线课程组使用相同强度的五项导数损失(1、0.5、0.1、2、0.05)。因果编码另做相同预算的GT历史对照。最后的较轻组合改为3,000步,权重0.2、0.2、0.05、0.2、0.005,70%抽样为运动窗口,75%历史混入上限,750步刷新,短2轮→短5轮→完整session;不降低远期监督权重。该组合改变了预算和多项训练策略,不能将其改善唯一归因于因果编码。 +- +-## 续训与确定性 +- +-包含因果编码、预测历史缓存和分开的导数标签的完整282,790,161参数模型,已在确定性CUDA/math attention模式下验证:第3步checkpoint两次恢复到第6步,模型、Adam、scheduler、全局随机状态、窗口采样、历史选择随机数、缓存及刷新元数据逐项完全相同。默认CUDA attention的重放出现过微小数值差异,但缓存和随机状态一致,不能把默认运行称为逐位可复现。确定性检查使用`CUBLAS_WORKSPACE_CONFIG=:4096:8`、`torch.use_deterministic_algorithms(True)`并关闭flash/memory-efficient SDPA;正常实验保留原性能配置。 +- +-## 输入时机的限制 +- +-训练中发现一个具体例子:`train-hesitation-000087` 的首段“performing ride the pony”在mic2.2s释放,最后一个词在mic5.4s释放;GT持续根运动起点约motion6.67s(按>0.15m/s持续0.2s检测)。一些候选在motion2.5s就开始运动。释放时钟没有错位,但文本模型看到可执行的前半句时,并没有显式的讲话/结束状态。这个例子提示输入信息可能限制时机复现,没有证明所有提前/静止问题均由它引起。 +- +-本轮没有加入GT阶段、语音结束标签、动作时钟或额外音频条件,也没有改动其他四个baseline的输入协议。后续若测试语音状态/文字稳定时长,需明确区分可实时计算的因果信号与GT阶段标签,并单独报告协议变化。 +diff --git a/docs/KIMODO_FORMAL_VALIDATION.md b/docs/KIMODO_FORMAL_VALIDATION.md +deleted file mode 100644 +index 0cf05ca..0000000 +--- a/docs/KIMODO_FORMAL_VALIDATION.md ++++ /dev/null +@@ -1,72 +0,0 @@ +-2026-10-24:修复后新100K的正式评测已完成,见[最新报告](HUMAN_FORMAL_VALIDATION_20261024.md)。本文记录旧失败训练的历史评测,不能把下文100K结果当成新权重的结果。 +- +-2026-10-23 更新:用户随后要求修复训练并删除旧失败目录;原run及其全部权重已删除,本文数值、分析缓存和视频保留。训练配置及修改见[训练修复](KIMODO_TRAINING_REPAIR.md)。当前同名run/权重目录已被新训练重新创建。 +- +-2026-10-23:`322190.textop` screen 实际运行的是 Kimodo,已正常完成 100,000 步,耗时约 19.6 小时。 +-本轮对六类未参与训练的 validation session 做了完整闭环生成、整段 GMR retarget 和有声视频评测。 +-100K 权重已有明确的人体姿态与朝向收益,但根位置、窗口衔接和部分交互仍较弱,暂不能认定全双工指标已经稳定可用。 +- +-同日进一步检查确认 100K 权重的根网络已退化为近似常数;因此撤回单凭平均姿态误差推荐 100K 整模的建议。诊断结论见[根网络退化检查](KIMODO_ROOT_COLLAPSE_DIAGNOSIS.md)。 +-`best.pt` 实际为 2K 权重:它的去噪验证 loss 最低;初轮六段连续生成的平均姿态误差逊于 100K,但根网络仍有动态响应。 +-初轮评测没有修改训练配置或重新训练;失败权重已在后续修复阶段按用户要求删除。 +- +-评测约定: +- +-- 每类取 native validation index 中第一个符合条件的英文、具有人体配对、预计 6–20 秒的 session。样本在生成前选定,没有根据效果挑选。 +-- 六类:turn-taking、interruption、emergency-stop、backchannel、hesitation、pause。源动作共 88.4 秒;排除每段第一帧后评分 2,646 帧。 +-- 输入为按既定时间因果释放的英文 GT 词、最近 30 秒上下文和已执行的人体历史。无音频输入、无未来文本或阶段标签。 +-- 仅以第一帧 GT 姿态初始化,之后完全使用预测历史。历史 60 帧,规划 60 帧,每次执行 12 帧,30 Hz。 +-- DDIM 20 步、CFG 2.0、seed 12345。官方、2K、100K 使用相同的适配、文本、随机种子和闭环流程。 +-- 每段完整预测及 human GT 各连续 retarget 一次到 G1;不在窗口之间重置 IK,不作位置对齐、DTW 或轨迹修正。 +-- 下表是六段分别计算后等权平均,单位为 rad 或 m。位置误差为三个坐标的逐元素 RMSE。 +- +-| 权重/对照 | 人体关节旋转 RMSE (rad) | G1 关节 RMSE (rad) | G1 根位置 RMSE (m) | G1 根朝向 RMSE (rad) | +-|---|---:|---:|---:|---:| +-| 官方初始化,采用相同 HFD 适配 | 0.628 | 0.464 | 0.433 | 0.600 | +-| best.pt,2K | 0.643 | 0.467 | 0.463 | 0.825 | +-| last.pt,100K | **0.454** | **0.390** | 0.544 | **0.266** | +-| 保持每段初始 GT 姿态 | 0.431 | 0.366 | 0.365 | — | +- +-G1 主指标的参考是同流程 retarget 的 human GT。100K 相比官方初始化,人体姿态误差降低 27.7%,G1 关节误差降低 15.9%,根朝向误差降低 55.7%,但根位置误差增加 25.7%。 +-人体原生空间的位置 RMSE 同样从 0.502 m 升至 0.629 m,所以位置问题并非只由 retarget 造成。 +-GT 活动帧的人体旋转误差从 0.633 降至 0.478 rad;保持初始姿态的相应误差约 0.447 rad。 +-动作幅度和视频均表明 100K 并非完全静止,但在本轮静止/倾听时间较多的样本上,平均误差仍未超过保持初始姿态的简单对照。 +-该对照不能完成动作任务;也不能把低姿态误差直接当成语义任务成功。 +- +-| 验证样本 | 100K 人体旋转 RMSE (rad) | 100K G1 关节 RMSE (rad) | 100K G1 根位置 RMSE (m) | +-|---|---:|---:|---:| +-| turn-taking:向后跳 | 0.321 | 0.247 | 0.385 | +-| interruption:摸头后切换姿态 | 0.533 | 0.478 | 0.226 | +-| emergency-stop:向左行走并收到停止文本 | 0.350 | 0.286 | 0.899 | +-| backchannel:拄拐后退并收到附和文本 | 0.541 | 0.469 | 1.311 | +-| hesitation:拿物体指令中插入犹豫 | 0.572 | 0.529 | 0.242 | +-| pause:拿物体指令中暂停 | 0.406 | 0.332 | 0.201 | +- +-主要闭环问题集中在窗口接缝。将每个 12 帧执行块的第一帧标为重规划边界,100K 的边界根速度误差为 0.839 m/s,块内部为 0.118 m/s。 +-各段边界帧占比平均 8.35%,却贡献平均 81.5% 的根速度平方误差。 +-官方初始化对应为边界 0.724、内部 0.169 m/s;100K 改善了块内运动,但边界仍有明显跳变。 +-曲线还可见部分样本约 2 秒一次的位置脉冲。这些现象在人体输出中已存在;目前证据只支持“生成窗口衔接未解决”,尚不能断定是某一行代码或单一训练因素。 +- +-训练去噪验证 loss:初始 0.371、2K 最低 0.198、100K 为 0.340。 +-这种 GT 历史下的短窗口去噪指标与执行预测历史的整段质量不一致,不能据此简单判断 2K 最好或 100K 只是过拟合。 +-后续应将固定闭环样本的姿态、位置和边界速度一起用于 checkpoint 选择,并优先排查历史/未来交界处的根表示与解码,再考虑连续性监督。 +-当前没有证据支持直接追加长训练即可解决这些问题。 +- +-视频保存在 `/root/autodl-tmp/runs/kimodo-formal-validation-20261023/`: +- +-- `last100k/`:六段预测→G1 与 human-GT→G1 对比,以及六段 human-GT→G1 与原始 G1 对比,共 12 段。 +-- `best2k/`、`official/`:各六段预测→G1 与 human-GT→G1 对比。 +-- `retargeted_human_comparison.mp4` 左侧为预测,右侧为 human GT 经相同 retarget 得到的参考。 +-- `retarget_reference.mp4` 用于查看 retarget 本身的差异。human-GT→G1 对原始 G1 的平均关节误差 0.226 rad、根位置误差 0.045 m;这不是可从模型误差中减去的常数。 +-- 共 24 个有声 MP4,帧数、帧率、音轨和完整音视频解码均通过检查。runs 中仅导出视频,工作数据在 cache 中。 +- +-这是一轮六段、单随机种子的离线轨迹检查,未运行物理跟踪、任务成功率或完整 HFD 验证集评测。 +-可以确认身体姿态训练有收益;进一步诊断已确认 100K 根阶段退化,需修复后再选择正式比较权重,当前不足以保证正式全双工指标“不难看”。 +- +-完整数据与曲线: +- +-- [数值和样本溯源](../reports/kimodo_formal_validation_20261023.json) +-- [训练验证曲线](../reports/kimodo_formal_validation_curve_20261023.png) +-- [六段闭环运动速度曲线](../reports/kimodo_formal_validation_motion_20261023.png) +-- [24 段视频的校验结果](../reports/kimodo_formal_validation_videos_20261023.json) +-- 本轮输入、原生预测、G1 retarget、完整文本释放记录在 `/root/autodl-tmp/cache/kimodo-formal-validation-20261023/`。 +diff --git a/docs/KIMODO_HFD_OVERFIT_RESULTS.md b/docs/KIMODO_HFD_OVERFIT_RESULTS.md +deleted file mode 100644 +index f9830e2..0000000 +--- a/docs/KIMODO_HFD_OVERFIT_RESULTS.md ++++ /dev/null +@@ -1,66 +0,0 @@ +-# Kimodo 英文人体过拟合验收(2026-10-21) +- +-复核结论:**本轮仅通过训练链路检查,闭环过拟合未通过。** 用户指出视频动作偏静止后,运动幅度审计确认三条训练样本的关节运动速度仅约GT的三分之一。训练急停样本保持初始姿态的关节/根误差为0.3607rad/2.0503m,微调闭环为0.3624rad/2.0289m,基本未超过静止基线。下文误差表作为原始记录保留,不应据其宏平均下降宣称动作复现成功。详见评测目录的 `motion_amplitude_audit.json` 和该样本 `motion_audit.json`。 +- +-训练框架、输入和复现命令见 [Kimodo训练说明](KIMODO_HFD_TRAINING.md)。运行目录为 `/root/autodl-tmp/runs/kimodo-english-overfit-20261021`;视频与逐样本指标入口为 [/root/autodl-tmp/runs/kimodo-english-evaluation-20261021/README.md](/root/autodl-tmp/runs/kimodo-english-evaluation-20261021/README.md)。 +- +-## 已完成的检查 +- +-- train/val 各六个原生英文配对人体 session,分别为 84.4/88.4 秒、205/215 个窗口,覆盖六种交互模式。选择规则在训练前固定;不是全库正式训练。 +-- 27 项单元/数据契约测试通过。实际 session 的坐标转换回环误差:平移不超过 6e-8m、旋转不超过 2.8e-6rad;扰动未来不改变输入历史或 heading。 +-- 文本基础模型无缺失权重;MNTP/supervised 两份适配器各 448 个张量与文件逐项一致。冻结编码得到 81 个文本特征,包含空文本的零向量;双向注意力探针通过。动作环境 `pip check` 通过,旧版文本依赖单独隔离。 +-- 完整 282,790,161 参数模型的 root/body 梯度分别非零且有限,一次 Adam 更新实际改变参数;移除文本会改变预测。 +-- 从第 4 步独立恢复两次至第 6 步,全部模型参数、采样器和两步 loss 完全一致;Adam step、scheduler step 均正确恢复。 +- +-## 1,000 步训练 +- +-使用 batch8、AdamW 1e-5→1e-6、bf16,7 项结构损失和 10% 文本 dropout。固定噪声评估每次抽取 64 个窗口,训练/验证独立抽样,不能理解为遍历全数据集。 +- +-| 指标 | 初始 | 1,000步 | +-| --- | ---: | ---: | +-| 固定训练窗口去噪损失 | 1.61369 | 0.24366 | +-| 固定验证窗口去噪损失 | 1.28607 | 0.60110 | +- +-最佳固定验证损失为第 800 步的 0.55017。主闭环对照明确使用最后第 1,000 步,避免把窗口损失最优当成闭环最优。每步中位耗时 0.0896s,峰值显存 5.33GiB;第 6→1,000 步续训进程共约 107 秒,含验证和保存。此吞吐只代表小型预制窗口与缓存文本,不代表全量动态数据加载和 8B 文本编码成本。 +- +-## 闭环对照 +- +-训练和验证各选 turn-taking、interruption、emergency-stop 三条 session;官方初始化/微调模型均使用相同 seed、DDIM20、CFG2、单帧 GT 初始化和后续预测历史。下表为各三条的逐 clip 宏平均,无根对齐或 DTW。 +- +-| 范围 | 关节旋转RMSE,初始→微调(rad) | 根位置RMSE,初始→微调(m) | +-| --- | ---: | ---: | +-| 训练三条 | 0.5711 → 0.3492 | 1.5246 → 0.7362 | +-| 验证三条 | 0.6580 → 0.4471 | 0.9332 → 0.4737 | +- +-验证宏平均关节误差下降约32%,根位置误差下降约49%。这是小规模探索性验证,三条样本不足以证明泛化性能。 +- +-| 样本 | 关节初始→微调(rad) | 根位置初始→微调(m) | +-| --- | ---: | ---: | +-| train turn-taking 13 | 0.5262 → 0.3260 | 0.8745 → 0.0323 | +-| train interruption 73 | 0.6752 → 0.3591 | 1.0668 → 0.1476 | +-| train emergency-stop 36 | 0.5118 → 0.3624 | 2.6323 → 2.0289 | +-| val turn-taking 3 | 0.5458 → 0.3664 | 1.0103 → 0.3108 | +-| val interruption 75 | 0.7042 → 0.4871 | 1.0318 → 0.0803 | +-| val emergency-stop 200 | 0.7240 → 0.4877 | 0.7574 → 1.0299 | +- +-每条完整预测与人体 GT 分别做 GMR,再与原始 G1 作三组比较;主对照共 36 条带音轨视频。人体指标先于 retarget 计算,GT→G1 的差异单独保留,不能把它当作可直接相减的误差下界。视频为姿态回放,不是物理跟踪或任务成功率评测。 +- +-加上两组历史/CFG诊断共42条视频,全部逐帧解码通过,共16,812帧,音轨也全部可解码。校验清单在评测目录 `video_checks.json`。这次20步DDIM的单轮规划平均约0.46–0.65秒,尚未满足每0.4秒执行窗口的严格实时要求;当前验收针对离线推理。 +- +-## 急停样本与历史诊断 +- +-训练急停样本的 GT 先行走约 5 米再停下,微调闭环预测大多停留在起点附近。它的问题主要是没有启动该段行走,不能简单解释成“听到停止后仍一直向前”。该样本原始动作描述为 `stop fast walk forward`,随后释放 `Pause for a moment.`;没有为改善结果改写文本。 +- +-| train emergency-stop 36诊断 | 关节RMSE(rad) | 根位置RMSE(m) | +-| --- | ---: | ---: | +-| 预测历史,CFG2,主结果 | 0.3624 | 2.0289 | +-| GT历史,CFG2,仅诊断 | 0.0896 | 0.0910 | +-| 预测历史,CFG1,仅诊断 | 0.4779 | 2.0421 | +- +-这说明模型在给定正确历史时能较好地拟合该轨迹,但闭环可能停留在静止状态;降低 CFG 未解决问题。GT历史结果不能替代主闭环指标,也不能独立证明唯一原因。验证急停样本的根误差还从 0.757m 升到 1.030m,因此不能宣称已经学会急停。 +- +-结论是微调框架与梯度路径可用,小规模训练已改善多数样本;当前仍有闭环启动、状态切换和根运动不足的问题。下一阶段应扩大训练覆盖,增加连续多轮训练/预测历史扰动,并单列启动与停止指标,再决定正式训练预算;只延长这六条 session 的训练不能保证解决上述问题。纯文本输入在词释放前也没有语音起点信息。 +- +-## 留存 +- +-运行目录保留配置、数据/文本审计、源码与环境快照、loss 曲线和完整 resume 检查。当前 `best.pt` / `last.pt` 在 `/dev/shm/kimodo-english-overfit-20261021`,属于临时测试权重,重启不保留;两份独立恢复测试的重复 checkpoint 已清理,日志仍保留。GenMo 的 GPU1 正式训练未停止。 +diff --git a/docs/KIMODO_HFD_TRAINING.md b/docs/KIMODO_HFD_TRAINING.md +deleted file mode 100644 +index 4810ba7..0000000 +--- a/docs/KIMODO_HFD_TRAINING.md ++++ /dev/null +@@ -1,112 +0,0 @@ +-# Kimodo HFD 全量微调与小规模验收 +- +-本地入口保留 Kimodo-SMPLX-RP-v1 的 root→body 两阶段网络与官方初始化,增加监督训练、完整断点恢复、连续推理、人体指标、整段 GMR 和视频评测。上游代码固定在 `e1c219c99b4b94ba48be1fb6a1cbc64ecee4c954`,不修改上游实现。 +- +-2026-10-24:修复版已从官方初始化完成新一轮100K训练;当前best与last均为100K,新正式权重保留。根骨干200个参数张量仍与官方初始化逐位一致,梯度和时序响应正常,未重现旧恒定输出失效。六段验证的G1根位置RMSE为0.264m,但后退与部分指令跟随仍不足。完整数值和有声视频见[本轮正式评测](HUMAN_FORMAL_VALIDATION_20261024.md)。 +- +-训练方式见[训练修复与验收](KIMODO_TRAINING_REPAIR.md):冻结根 Transformer 骨干,仅微调根输入/输出投影,身体保持训练;增加实际解码运动连续性监督,分阶段学习率及梯度裁剪。此前失败正式训练的run和权重已经删除,当前同名目录属于新训练。以下早期过拟合结果保留为历史对照。 +- +-历史混合闭环过拟合未通过。较好的早期诊断路线为归一化x0目标、2,000步初训加1,000步GT历史微调:行走样本达到约0.056rad/0.040m,但其他模式相位与泛化仍不足。历史实验见[静止退化复查](KIMODO_MOTION_RECOVERY.md),完整对照见[闭环稳定性报告](KIMODO_CLOSED_LOOP_STABILITY.md)。`kimodo_overfit.yaml`和`kimodo_refine_probe.yaml`复现这条诊断路线;`kimodo.yaml`现为全量正式训练配置。未默认启用在线历史混入。旧缓存需按当前编码重新准备。 +- +-## 输入、目标与因果边界 +- +-- 原生英文 human 配对样本,按实际 session 组合操作拼接人体动作;G1 只用于配对和评测参考,不进入 Kimodo 输入或损失。小规模验收固定 train/val 各六个 session;正式入口覆盖完整英文原生索引,按需组装配对人体 session。 +-- 文本沿用共同协议:对齐后的前景英文 GT 单词,词结束加 0.1s 后等待麦克风 `0.6+0.4k` 释放,保留最近 30s。动作时刻 `t` 对应麦克风 `t+0.2s`。无音频、翻译、未来文本、交互类型标签输入。 +-- 冻结 LLM2Vec/Llama-3-8B-Instruct,MNTP 适配器合并,再加载 supervised 适配器;逐项核对权重。原生 mean pooling 得到 `1×4096`,网络内部的 50 个文本 token 由其自身投影构造。文档预算 400 token,超长保留最近文本;模板总预算 512,编码内部 batch=1。 +-- 30Hz,历史60帧、默认预测60帧、执行12帧。12帧预测仅用于诊断,未采纳为默认。历史mask锁定203维姿态,70维速度/接触辅助量不作为已知约束。当前配置使用因果root特征分解,使历史在输入和监督中一致,不改变物理人体动作;首个已知姿态确定heading与平移坐标系。 +-- 输出原生 273D 特征,由原生 inverse 解码到 22 个 SMPL-X 关节和根轨迹,再逆变换到 AMASS;手部中性。保留 session 体型,显式处理其 pelvis 偏移。完整输出与人体 GT 分别 GMR 到 G1。 +- +-当前`training.objective: normalized_x0`在归一化273D特征上做x0均方误差,root与body分别取均值后相加;仅未来有效帧计分,末端无下一帧时速度/接触不计分。保留原生两阶段之间的detach,两阶段各有直接梯度。诊断配置使用50% t=999、其余均匀t,以及70%运动窗口抽样。原七项物理量SmoothL1配方通过`physical_huber`保留作对照,YAML的`training.weights`只在该模式生效;不宣称复现官方未公开训练配方。 +- +-推理从一个已知初始姿态启动,后续历史来自已执行预测;CFG两支保留相同历史。历史70维未观察辅助量在每个DDIM步恢复为训练时的纯噪声边缘分布,避免把未监督的历史预测继续反馈;`legacy`仅用于旧规则对照。默认DDIM20、CFG2,无轨迹后处理。`--teacher-forced`仅诊断使用。 +- +-## 环境 +- +-动作使用 `/root/autodl-tmp/envs/kimodo/bin/python`,保留上游 Transformers 5.1.0。原生 LLM2Vec 的双向 Llama 代码需要独立旧版依赖,在 `/root/autodl-tmp/envs/kimodo-text-compat` 保存 `requirements/kimodo-text.txt` 所列包;仅文本进程将其加入 Python 路径。可用 `KIMODO_TEXT_COMPAT` 指向其他安装目录。 +- +-重新安装文本 overlay 时,应先用用户要求的 `/root/.local/bin/turbo` 调用 pip,并用 `--no-deps --target <目录> -r requirements/kimodo-text.txt`,复用动作环境已有的 Torch 等依赖。不要直接将该 requirements 安装进动作环境。文本 worker 自检未来 token 能影响前面 hidden state,防止误用单向注意力;小规模离线缓存完成后动作训练不加载 8B 编码器。全量训练采用常驻独立文本进程,与动作训练共用所选 GPU;缓存命中时不重复编码。 +- +-## 全量正式训练(2026-10-23 修复版) +- +-```bash +-cd /root/autodl-tmp/hfd-baseline-portable +-/root/autodl-tmp/envs/kimodo/bin/python -u train_kimodo.py --config configs/kimodo.yaml +-``` +- +-中断恢复使用相同命令追加 `--resume`,读取完整 `last.pt`,恢复模型、Adam、余弦调度器、全局 RNG、抽样 RNG 和每个 batch 槽位的 session;100K 是总步数。配置/实现/数据身份变化会拒绝恢复。 +- +-- GPU 0;batch 8;100K 步;BF16。根仅训练输入/输出投影,AdamW 2e-6→2e-7,在20K衰减到底;身体AdamW 1e-5→1e-6,在100K衰减到底;500步warmup,两阶段分别裁剪到1。因果root、normalized x0、10%强制t=999(其余均匀t)、文本dropout 0.1;执行12帧权重1,剩余规划帧0.25。增加原始根历史重建及实际解码速度/加速度监督;根注意力采用数学后端,运行配置随checkpoint保存并恢复。全量所有合法窗口均匀边缘抽样,不启用在线历史课程。具体权重以`configs/kimodo.yaml`为准。 +-- train 原生英文 155,117 个 session;val 独立原生英文 4,719 个 session;没有6–20秒筛选、没有每模式数量上限、无第三方G1动作、无中文。初始化统计会打印真实覆盖数。 +-- 不预拼155K个session。每个 batch 槽位独立按保守长度上界拒绝抽样,再均匀选0.4s决策点;训练槽位复用16个batch,验证复用4个batch。所有session有被抽到的机会,预算内不保证每个session都被访问;100K不是一个epoch。 +-- paired SMPL-X 动态拼装与 GenMo 相同来源和时间轴;本地512MiB窗口LRU、每split512MiB原始sessionLRU;冻结LLM2Vec独立进程按需编码,每次内部batch1,SQLite缓存本地 `/root/autodl-tmp/cache/kimodo-text-full`,特征payload上限4GiB,清缓存不改变抽样顺序。编码器不反传;主模型正常反传。 +-- 文本仍是按单词释放的最近30s英文GT,60帧历史,60帧目标,12帧执行,无音频或未来语义输入。正式训练依然用GT历史;闭环推理用预测历史。小样本泛化差属于质量风险,继续通过全量独立验证和闭环视频观察。 +-- 首次启动要构建train/val原生索引、加载8B文本模型并编码首次抽样文本,启动阶段可能数分钟。`FULL_DATA`、`DATA`是正常准备进度。每10步打印`TRAIN`;`seconds`含数据等待和更新,`data_seconds`单独显示组装/文本等待。 +-- 每2K步执行固定种子16×8个验证窗口,并记录同种子训练诊断;这些窗口从完整val索引抽取,绝非全val遍历。噪声损失用于选`best.pt`,不能替代闭环动作评测。 +-- 日志/配置在 `/root/autodl-tmp/runs/hfd-kimodo-english-train`;权重在 `/root/autodl-tmp/models/kimodo/hfd-english-train`,保留完整last和最近2个完整快照,以及仅模型best。本地唯一临时文件完成flush/fsync后原子替换;写入失败最多重试3次,保留上一断点。快照用硬链接避免重复写入。 +- +-训练结束后,使用正式run选择六模式验证session,执行预测历史→整段GMR→带音频视频: +- +-```bash +-/root/autodl-tmp/envs/kimodo/bin/python -u evaluate_kimodo.py \ +- --run /root/autodl-tmp/runs/hfd-kimodo-english-train \ +- --output /root/autodl-tmp/runs/kimodo-full-validation \ +- --splits val --variants trained --gpu 0 +-``` +- +-全量评测按每模式第一个满足6–20s的原生英文session作固定可视化子集,动作G1仅评测导出,训练不加载。推理可额外使用`--checkpoint`选择其他完整快照。评测GPU需要容纳动作模型和独立文本模型;勿与该GPU上的训练并发。批量全val指标可后续扩大选择,当前命令只展示六例。 +- +-启动验收:44项单元测试通过;实际全量入口完成2步、保存完整断点后恢复到3步,核对下一批位置、session槽位、抽样RNG、Adam和调度器。运行统计身份误变问题已修复。临时测试权重已删除;这不是100K质量验收。报告见 [`kimodo_full_training_ready_20261021.json`](../reports/kimodo_full_training_ready_20261021.json)。 +- +-## 小规模诊断命令 +- +-```bash +-cd /root/autodl-tmp/hfd-baseline-portable +-PY=/root/autodl-tmp/envs/kimodo/bin/python +-$PY prepare_kimodo.py --config configs/kimodo_overfit.yaml +-$PY cache_kimodo_text.py --run /root/autodl-tmp/cache/kimodo-overfit --gpu 0 +-$PY train_kimodo.py --config configs/kimodo_overfit.yaml --offline-check +-$PY train_kimodo.py --config configs/kimodo_overfit.yaml +-# 中断后恢复,max_steps 是总预算: +-# $PY train_kimodo.py --config configs/kimodo_overfit.yaml --resume auto +-``` +- +-复现第二阶段时,复用首阶段的冻结输入,再从其权重warm-start(首阶段权重需暂时保留): +- +-```bash +-$PY - <<'PY' +-from pathlib import Path +-from omegaconf import OmegaConf +-base = OmegaConf.load('configs/kimodo_overfit.yaml') +-refine = OmegaConf.load('configs/kimodo_refine_probe.yaml') +-destination = Path(refine.output) +-destination.mkdir(parents=True, exist_ok=True) +-for name in ['data.json', 'text_features.pt', 'text_features.json', 'texts.json']: +- target = destination / name +- if not target.exists(): +- target.symlink_to(Path(base.output) / name) +-OmegaConf.save(refine, destination / 'resolved_config.yaml') +-PY +-$PY train_kimodo.py --config configs/kimodo_refine_probe.yaml +-``` +- +-完整checkpoint包含网络、AdamW、cosine scheduler、全局step、Torch/CUDA/Python/NumPy随机状态和窗口采样器;启用在线历史时还包含历史缓存、对应policy step与混入选择随机数。恢复核对数据、缓存文本、代码、官方权重/统计及配置。`--stop-after 4`可用于检查。`best.pt`按固定验证去噪损失选,`last.pt`保存最后状态;闭环质量另行评测。逐位复现需显式启用确定性CUDA配置,见最新报告。 +- +-```bash +-$PY evaluate_kimodo.py \ +- --run /root/autodl-tmp/cache/kimodo-overfit-refine \ +- --output /root/autodl-tmp/cache/kimodo-evaluation \ +- --checkpoint /dev/shm/hfd-kimodo-overfit-refine/last.pt --gpu 0 \ +- --modes turn-taking interruption emergency-stop backchannel hesitation pause +-``` +- +-默认分别报告训练/验证的 turn-taking、interruption、emergency-stop,每个对照官方初始化与微调模型。先计算人体根/旋转误差,再按每个完整轨迹做 GMR,保存 predicted human↔retargeted GT、predicted human↔原始 G1、retargeted GT↔原始 G1 三组比较。视频含原始音轨和文本释放记录。两种模型共用 seed、采样步数、CFG 与历史策略;验证集用于诊断,不作为独立测试集。 +- +-首阶段YAML为2,000步、batch8、AdamW 3e-5→3e-6;`kimodo_refine_probe.yaml`再从首阶段模型权重初始化,重置Adam,以1e-5→1e-6训练1,000步。第二阶段不是`--resume`,而是显式warm-start。工作数据和日志在cache,临时权重在/dev/shm,runs只存视频。正式训练仍需另行实现更大session范围并验证预算与泛化。 +- +-## 验收产物 +- +-cache中的工作目录记录数据、文本审计、完整模型梯度检查与训练指标;runs只保存筛选后的MP4。首轮失败诊断见[过拟合验收](KIMODO_HFD_OVERFIT_RESULTS.md),最新对照、视频与保留范围见[闭环稳定性](KIMODO_CLOSED_LOOP_STABILITY.md)。 +- +-2026-10-22 保存故障修复:旧训练在64K保存时无法在共享盘创建临时文件,62K完整断点及best已归档到 `/root/autodl-fs/pangxinlei/models/kimodo/archive/hfd-english-train-failed-20261022`。默认正式配置改用本地盘保存新权重,旧失败run的日志等已清理。标准启动命令从头训练;旧归档含匹配原断点契约的训练脚本副本,新代码不会自动绕过旧断点的实现校验。见[故障报告](../reports/kimodo_checkpoint_failure_20261022.json)。 +- +-2026-10-23 旧一轮正式训练完成 100K 步。六类 held-out session 的历史数值与 24 段有声视频见[正式训练验证](KIMODO_FORMAL_VALIDATION.md)。进一步确认 100K 根网络退化为近似常数;替换早期根权重也尚未恢复正确后退。随后加入的退化检测和自动停止已按用户要求全部撤回。本轮改为修复训练策略,旧失败run及其权重已全部删除,历史视频和分析数据保留;具体原因见[训练失效机制](KIMODO_ROOT_TRAINING_CAUSE_ANALYSIS.md)。 +diff --git a/docs/KIMODO_MOTION_RECOVERY.md b/docs/KIMODO_MOTION_RECOVERY.md +deleted file mode 100644 +index 6cc4edb..0000000 +--- a/docs/KIMODO_MOTION_RECOVERY.md ++++ /dev/null +@@ -1,78 +0,0 @@ +-# Kimodo 静止退化复查与恢复实验 +- +-本文保留历史恢复实验。后续在线历史课程、因果root表示修正和同预算预测时长对照见[最新闭环报告](KIMODO_CLOSED_LOOP_STABILITY.md);混合闭环验收仍未通过。 +- +-用户观看视频后指出预测几乎不动。复查确认:[原1,000步实验](KIMODO_HFD_OVERFIT_RESULTS.md)的闭环过拟合**没有通过**。宏平均误差下降不足以证明动作复现;先前报告对此解释不充分。 +- +-## 直接证据 +- +-三条原训练评测的预测关节运动速度仅为GT的32%–36%。`train-emergency-stop-000036` 的根路径为0.83m,GT为5.29m;一直保持初始姿态的关节/根误差为0.3607rad/2.0503m,模型为0.3624rad/2.0289m,几乎没有超过静止基线。 +- +-新增 `audit_kimodo_motion.py --inference <目录>` 同时报告原人体误差、静止初始姿态基线、根路径、关节角速度、GT活跃区间误差和持续运动启动时间。运动幅度不能单独证明正确性:抖动和错误动作也可能增加路径/角速度。启动检测仅为诊断,使用根速度>0.15m/s持续0.2s,不是通用任务成功判据。 +- +-## 训练改动与对照 +- +-输入协议保持不变:已释放英文文本、60帧历史、60帧未来、每轮执行12帧;无音频、GT阶段、未来姿态或剩余时长条件。文本缓存和预训练动作参数均沿用原实验。 +- +-`hfd_bridge/kimodo_recovery.py` 增加可选训练组合: +- +-- 在root阶段输出传给body阶段之前锁定已知历史root,保持两阶段条件一致。旧实现仅在两阶段完成后锁定最终输出。单独给旧权重加此约束,根误差仅2.0289→2.0265m,**它本身没有解决静止问题**。 +-- 原速度损失针对独立的velocity特征块,不能保证实际解码姿态在动。增加由解码root、FK关节和旋转增量计算的时序监督,权重为1、2、0.5;不是只拟合辅助速度输出。 +-- 每批50%样本使用t=999,增加从近纯噪声依靠文字/历史生成的训练;其余扩散步仍均匀采样。 +-- 训练按GT运动标签将70%抽样分给明显运动窗口,保留30%等待/停止窗口。六条原始session有83个活跃窗口、122个其余窗口。标签只用于训练抽样,不输入模型。 +- +-## 单样本、同预算对照 +- +-均从原六条训练的1,000步模型初始化,只训练同一条急停/行走session,新增1,000步、batch8、LR3e-5、相同cosine预算、text dropout=0、同seed。原配方对照保留七项损失与均匀窗口/扩散步;恢复配方同时启用上述改动。单条实验是定位问题,不是泛化评测。 +- +-| 配方(预测历史、CFG2) | 关节RMSE(rad) | 根RMSE(m) | +-| --- | ---: | ---: | +-| 原六条1,000步模型 | 0.3624 | 2.0289 | +-| 原配方单条新增1,000步 | 0.3164 | 0.5658 | +-| 恢复组合单条新增1,000步 | 0.0427 | 0.1332 | +- +-恢复组合开始真正行走,持续运动起点约3.17s,GT约3.33s;关节速度是GT的1.23倍。不过根路径7.31m仍高于GT5.29m,有抖动/多余位移。CFG1诊断为0.0437rad/0.0759m,根路径7.08m;不能只看较低的位置误差忽略多余路径。原配方更充分训练也改善根误差,且恢复实验同时改变多项因素,不能将改善完全归因于某一项损失。 +- +-[单条真实闭环视频](/root/autodl-tmp/runs/kimodo-motion-recovery-20261021/probe1000/evaluation/retargeted_human_comparison.mp4),仅第一帧GT初始化;[动作/位移曲线](/root/autodl-tmp/runs/kimodo-motion-recovery-20261021/single_motion_comparison.png)。 +- +-## 六条混合实验 +- +-`configs/kimodo_motion_recovery_six.yaml` 从原1,000步模型继续微调3,000步,恢复10%文本dropout,默认CFG2。六种训练模式均用模型自己历史做完整rollout。 +- +-| 训练模式 | 关节RMSE(rad) | 预测/GT关节角速度比 | +-| --- | ---: | ---: | +-| turn-taking | 0.3158 | 0.360 | +-| interruption | 0.3138 | 1.175 | +-| emergency-stop | 0.0912 | 1.354 | +-| backchannel | 0.3228 | 1.110 | +-| hesitation | 0.5201 | 0.269 | +-| pause | 0.4859 | 0.568 | +- +-因此单条恢复**没有自动推广到六条混合过拟合**。行走改善,部分舞蹈仍运动不足;幅度接近GT的样本也可能动作/相位错误。hesitation给GT历史时关节误差0.0630rad,而闭环为0.5201rad,说明历史状态分布差异仍然很大。GT历史只用于诊断,不计入验收结果。 +- +-## 预测历史训练 +- +-`prepare_kimodo_predicted_history.py` 仅对六条训练session增加一份由已完成模型闭环生成的历史视图,验证集保持原样。每个训练窗口的60帧输入来自缓存的真实预测历史;未来监督来自GT。专家相对位移从当前预测根位置出发,避免要求瞬移回已错过的GT位置;前6帧身体旋转用0.2s过渡接向专家姿态。此过渡属于训练目标构造,不是推理后处理。 +- +-原GT历史视图与预测历史视图各205个窗口,训练均使用最近已释放文字。数据记录输入预测的哈希与plan ID,并拒绝GT历史rollout;`augmentation`视图不当作额外独立评测session。新增 `configs/kimodo_predicted_history.yaml`,从六条3,000步模型初始化,再训练2,000步。不能把六条的两种历史视图称为12条独立训练session。 +- +-所有阶段均为固定小规模数据诊断,不代表全库训练已经准备就绪。后续应按闭环动作幅度、与静止基线的比较、动作一致性与启动/停止共同判断,不能以去噪loss或平均姿态误差单独宣布通过。 +- +-## 预测历史实验结果与结论 +- +-完成新增2,000步后,仍使用相同seed、DDIM20、CFG2、一个GT初始姿态和后续完整预测历史,评估原始六条训练session(不把增强视图当成新样本)。 +- +-| 训练模式 | 关节RMSE(rad) | 预测/GT关节角速度比 | +-| --- | ---: | ---: | +-| turn-taking | 0.4061 | 1.458 | +-| interruption | 0.3091 | 1.217 | +-| emergency-stop | 0.1933 | 1.910 | +-| backchannel | 0.4659 | 1.615 | +-| hesitation | 0.3127 | 1.197 | +-| pause | 0.3637 | 1.403 | +- +-角速度比的精确数值以各session的 `motion_audit.json` 和汇总 `motion_recovery_comparison.json` 为准。相比上阶段,hesitation/pause的静止退化缓解,但有些模式更不准确。急停根位置误差0.7804m、速度RMSE3.9659m/s,明显抖动;turn-taking虽然根位置误差只有0.0595m,关节误差仍0.4061rad。这再次说明移动更多或根误差更小,都不能代替动作复现。 +- +-**当前六条混合闭环过拟合仍未通过。** 不将这两个混合恢复版本替换为正式配方,不推荐直接扩大全量训练。单条真实动作已经可以学到,但混合数据的闭环启动、状态切换、动作相位和轨迹稳定性尚不可靠。实验支持历史分布差异的重要性,没有证明它是唯一原因;缓存一轮预测历史也没有充分解决问题。现阶段继续盲目增加正式训练步数,无法保证排除静止或错误动作。 +- +-本次30项回归测试通过;预测历史缓存的因果边界、输入预测哈希和已知物理历史一致性检查通过。链路正确性与动作验收分开报告。所有新实验保留数据盘日志和视频,测试权重仍暂存在`/dev/shm`;原GenMo GPU1训练未停止。 +diff --git a/docs/KIMODO_ROOT_COLLAPSE_DIAGNOSIS.md b/docs/KIMODO_ROOT_COLLAPSE_DIAGNOSIS.md +deleted file mode 100644 +index 6fa6855..0000000 +--- a/docs/KIMODO_ROOT_COLLAPSE_DIAGNOSIS.md ++++ /dev/null +@@ -1,55 +0,0 @@ +-2026-10-23:后退失败的直接原因已定位到 **100K 权重中的根轨迹网络退化为近似常数输出**。 +-窗口坐标转换与已知历史保留检查通过。单纯换回早期根网络仍未恢复正确后退,因此还需修复根网络训练稳定性与闭环动作续接,不能将本轮替换对照视为可用修复。 +- +-证据来自已有 validation 样本 `val-backchannel-000057`,未用该样本训练: +- +-- 输入正确:3.0 秒已释放完整 `walking backwards with crutches`,7.8 秒才追加 `Wow.`。去掉 `Wow.` 后仍反向漂移。 +-- GT 的人体坐标编码→解码→导出最大平移误差约 0.0000083 m;采样末尾的已知历史解码误差为 0。 +-- 在决策帧 0、132、240、396,使用预测历史或 GT 历史,100K 根网络都输出几乎相同的未来位置。 +-- 在几乎无噪声的输入(t=0,提供加噪后的 GT 未来)上也如此。t=526/999、FP32/BF16、train/eval 均复现,排除了仅由半精度或评估模式引起的异常。 +-- 官方初始化和 2K 网络在相同输入下输出随时间变化的根轨迹。 +- +-下面对八个 GT 存在明显位移的窗口进行额外检查,混合 t=0/526/999。运动 RMS 定义为有效未来帧相对其时间均值的三维位置偏差 RMS,使用根阶段的 smooth-root 通道,不让身体阶段的髋部残差掩盖根阶段退化。 +- +-| 网络 | 预测根轨迹运动 RMS (m) | GT RMS (m) | 近似常数窗口 | +-|---|---:|---:|---:| +-| 官方初始化 | 0.1594 | 0.1958 | 0 / 8 | +-| 2K | 0.2381 | 0.1958 | 0 / 8 | +-| 100K | **0.000057** | 0.1958 | **8 / 8** | +- +-现有窗口以已知历史第一帧确定原点/朝向,编码解码的逆变换本身成立。 +-但退化网络不断在该原点附近输出几乎相同的位置,而不是延续历史末尾位置。 +-当历史以 60 帧长度滚动时,原点跟随过去预测变化,从而出现约两秒一次的位置脉冲。 +-这是“近似常数预测 + 滚动窗口原点”共同产生的现象,不能简单认定为坐标轴翻转或平移加错一次。 +- +-本轮进行了仅替换根网络、保留 100K 身体网络的闭环消融,均保持原始因果文本、随机种子、窗口和采样配置,不添加轨迹修正或 GT 历史: +- +-| 组合 | 沿 GT 后退方向的净位移 (m) | 人体位置 RMSE (m) | 人体姿态 RMSE (rad) | +-|---|---:|---:|---:| +-| GT | +3.498 | — | — | +-| 100K 整模 | -0.780 | 1.513 | 0.541 | +-| 2K 根 + 100K 身体 | -0.075 | 1.300 | 0.570 | +-| 官方根 + 100K 身体 | +0.183 | 1.132 | 0.706 | +- +-替换缓解了漂移,但没有完成约 3.5 m 的后退;身体姿态误差还增大。 +-这表明常数根网络确实是问题之一,现有早期根权重也不足以解决完整 HFD 动作续接。 +-两个替换对照均已整段 retarget 并保存有声视频: +- +-- `/root/autodl-tmp/runs/kimodo-root-diagnosis-20261023/root2k-body100k/backchannel.mp4` +-- `/root/autodl-tmp/runs/kimodo-root-diagnosis-20261023/root-official-body100k/backchannel.mp4` +- +-训练日志提供了时间线索:1–6K 的平均 normalized-root loss 约 0.042,6.5–7K 升至约 0.149,7–8K 约 0.220,而身体 loss 仍在缓慢下降。 +-没有保留这个区间的逐步权重,无法确定精确退化时刻,也未分别消融学习率、高噪声比例或样本分布来确定触发因素。 +-可以确定训练后的根阶段失效;不能把“原因就是学习率过大”等猜测当作已验证结论。 +- +-用户随后明确要求分析原因,不加入退化检测。上一轮加入的检测模块、测试、配置和自动停止已全部撤回,训练脚本恢复为本次 100K 正式训练时的原版本。原因分析见[根网络训练失效机制](KIMODO_ROOT_TRAINING_CAUSE_ANALYSIS.md)。 +- +-正式模型权重与已有视频均保留。本轮没有追加训练,没有生成临时训练 checkpoint。 +-旧 100K 训练脚本快照保存在 `/root/autodl-tmp/models/kimodo/hfd-english-train/source-100k-20261023/train_kimodo.py`。 +-训练脚本已恢复到旧正式训练契约的相同内容;诊断推理仍可正常加载全部旧权重。 +- +-下一轮训练应从健康根权重恢复,分别控制根阶段与身体阶段的更新,加入实际解码位移/接缝监督,并用固定闭环验证选择权重。 +-100K 身体阶段的姿态收益可作为初始化候选;当前不能把 100K 整模或上述混合模型直接作为已经修复的正式 baseline。 +- +-[完整检查数据](../reports/kimodo_root_collapse_20261023.json)。 +-逐窗口特征、模式/精度对照与诊断推理保存在 `/root/autodl-tmp/cache/kimodo-formal-validation-20261023/root-probe/`。 +diff --git a/docs/KIMODO_ROOT_TRAINING_CAUSE_ANALYSIS.md b/docs/KIMODO_ROOT_TRAINING_CAUSE_ANALYSIS.md +deleted file mode 100644 +index aa08523..0000000 +--- a/docs/KIMODO_ROOT_TRAINING_CAUSE_ANALYSIS.md ++++ /dev/null +@@ -1,59 +0,0 @@ +-2026-10-23:本轮分析的核心结论是,微调后的根网络丢失了时序和条件信息,随后落入前层梯度接近零的常数回归状态。上一轮所说的“根网络退化”只是输出表现;这里补充网络内部如何失效、为何训练无法自行恢复,以及哪些训练设置最值得怀疑。 +- +-本文记录旧100K训练的原因分析。分析阶段按用户要求撤回全部退化检测及自动停止,恢复了旧训练代码和配置。用户随后授权修复训练框架并删除失败目录;当前代码和配置已按[训练修复](KIMODO_TRAINING_REPAIR.md)更新,失败run及其全部权重已删除,既有数值、分析缓存和评测视频保留。以下机制测量均来自删除前保存的诊断结果。 +- +-已确认的内部机制: +- +-1. 根 Transformer 的不同时间帧逐层变得相同。在同一 GT 历史、同一接近无噪声未来输入下,2K 网络最后一层的帧间特征 RMS 为 0.0489,100K 只有约 1.08e-7。100K 第 1 层仍有约 0.0601 的差异,到第 8 层约 0.000854,第 13 层约 3.56e-7。信息在 Transformer 内部被抹平,早于坐标解码。 +-2. 深层注意力同时趋于均匀。该输入共 172 个 token;均匀注意力的熵为 ln(172)=5.14749,100K 最后一层实测约 5.14749,2K 约 1.0776。不同帧的同质化与近似全局平均相互强化,最后输出不再区分时间或条件。 +-3. 前层几乎无法再获得有效梯度。以同一根 MSE 对照,2K 对输入的梯度范数为 0.0170,100K 为 4.76e-12,下降约 35.8 亿倍。100K 输入投影权重梯度约 9.50e-11,但输出层权重梯度仍约 8.14。因此“根分支总梯度非零”不代表整个根网络仍能学习;末端有更新,前层已经很难恢复时序信息。 +-4. 在这种状态下,当前 MSE 的最佳行为是调整末端,使固定输出接近训练目标的平均值。实际抽样验证了这一点,见下表。这是“进入近似常数状态后为什么会维持这个平均答案”的解释;不代表 MSE 的全局最优只能是常数。 +- +-网络内部的这些测量使用 FP32、普通数学注意力实现,避免把加速后端的额外差异混入梯度结论。之前的 train/eval、FP32/BF16 对照也都确认 100K 输出近似常数,因此最终失效本身并非单一评估模式的现象。 +- +-对原正式训练采样器以 seed 42 抽取前 1,024 个窗口,覆盖 64 个 session;维持 batch8、16 批刷新一次的原约定: +- +-| 项目 | 结果 | +-|---|---:| +-| 根轨迹运动 RMS <1 cm 的窗口 | 49.3% | +-| 根轨迹运动 RMS ≥3 cm 的窗口 | 39.3% | +-| 尚无已释放文本的窗口 | 6.7% | +-| 仅输出样本目标均值的最优常数根 MSE | 0.22739 | +-| 100K 模型固定输出在同一批目标上的根 MSE | 0.22990 | +- +-100K 的固定答案只比“完全忽略输入、始终回答平均值”的最优常数解差约 1.1%。这与均值回归状态相符。 +-同时约 39% 窗口有明显位移,2K 网络也能利用输入得到更低误差,不能把问题简单归结为“数据全是静止”或“输入没有动作信息”。这只是 64 个 session 的抽样,不是完整数据集比例普查。 +- +-当前训练实现让这种状态难以被其他目标纠正: +- +-- 根阶段和身体阶段是两个 Transformer。原生 `TwostageDenoiser.forward` 在训练时以 `torch.no_grad()` 将根预测转成身体条件,并显式 `detach()`。所以身体阶段即使发现行走不对,其损失也不能传回根网络。 +-- 正式目标是 `normalized_x0`:根阶段五个特征的 MSE,加身体阶段特征的 MSE。YAML 中 `weights.root_position`、`joint_velocity` 等物理损失权重属于另一个 objective,本次没有启用。 +-- 身体输出中的 `velocities` 是独立辅助通道。导出人体运动时实际位置由 smooth-root 和髋部位置残差解码,未通过积分这些速度通道生成。辅助速度通道学对,不等于实际导出的根位移正确。 +-- 本次未配置 `recovery.temporal_weights`,因此没有针对实际解码轨迹速度、接缝速度或加速度的额外监督。 +-- 已知历史在输出端会被覆盖回观测值,且不计入未来损失。因而“历史输出保留准确”并不能证明 Transformer 内部仍在利用历史:根网络内部可以已经忘记历史,输出检查仍通过。 +- +-这些结构事实解释了为什么身体 loss 可以继续下降、整体梯度也非零,而根网络的坏状态得不到修正。 +-它们并不单独证明某一条代码必然使模型退化;健康 2K 权重已经说明同一结构可以暂时正常学习。 +- +-最值得怀疑的触发设置,是较强的根网络全参数更新与高噪声目标组合: +- +-- 根和身体共享 AdamW、3e-5 初始学习率和按 100K 步拉长的 cosine 调度,没有根阶段独立的更新预算。 +-- 2K 时学习率仍为 2.997e-5,6.5K 为 2.972e-5,8K 为 2.958e-5,几乎仍处于初始强度。 +-- 日志中 1–6K 的平均根 loss 约 0.042;6.5–7K 约 0.149,7–8K 约 0.220。身体 loss 同期仍缓慢下降。这是优化恶化的信号,不是仅在未见过的验证集上变差。 +-- `high_noise_fraction=0.5` 将约 50.05% 样本设为最大噪声 t=999;均匀采样时这一比例只有 0.1%。这大幅增加了无法依赖未来 noisy motion、必须依赖历史和文本的样本权重。 +-- 最大噪声时历史和文本仍然存在,因此高噪声本身不要求模型输出常数。但一旦条件表征开始丢失,这部分目标对模型来说就变成高不确定性 MSE 回归,平均答案成为容易维持的方向。 +- +-最符合现有证据的解释是:在这套长程微调设置下,根网络的时序/条件表示被后续更新破坏,深层特征和注意力进入近似均匀状态;前层梯度随之消失;MSE 将尚能更新的输出端推向平均答案,而隔离的身体损失无法救回根网络。 +-随后滚动窗口以历史开头为原点,重复生成同一个非零局部位置偏差,把根网络的均值偏差转成约两秒一次的世界坐标漂移。 +- +-严格的因果边界:目前没有保存 6.5–8K 附近的连续权重,也没有在相同全量数据顺序和随机种子下,分别改变根学习率、最大噪声比例与实际轨迹监督重训。因此不能宣称“唯一原因就是学习率”或“唯一原因就是高噪声”。已实测的是内部信息同质化、梯度消失、均值回归状态及其不能被身体梯度纠正的机制。 +- +-补充数值排查:本环境 PyTorch 2.7.1+cu128、RTX 5090。FP32 的 efficient attention 在特定高噪声输入上出现明显异常的大梯度;普通数学后端没有这一现象。 +-但正式训练使用 BF16,profile 确认使用 efficient attention。对六个既有训练 session 重建的 32 批、每批八个窗口,按原噪声/dropout 策略对照 BF16 efficient 与 math,没有复现那种量级的梯度爆炸:全参数梯度范数比最大约 1.061。 +-这不足以覆盖整次训练或排除稀有数值问题,但目前不能把 FP32 的异常直接当作本次 BF16 训练退化的已证实触发原因。 +- +-参数替换对照中,仅恢复输入投影、文本投影或输出层都不能恢复时序变化。 +-仅将注意力参数恢复为 2K,而保留其他 100K 参数时,输入梯度从 4.75e-12 恢复到约 0.00407,根轨迹运动 RMS 从约 6e-7 m 提高到 0.00626 m。这是一个直接的内部干预,支持注意力参数更新与信息/梯度消失存在联系;但该组合仍不能生成完整正确轨迹,不能据此宣称已经修复。 +-这支持失效发生在 Transformer 的内部条件处理,而不是某一个坐标输出 bias 写错。本轮所有替换仅用于只读诊断,没有生成新的训练权重。 +- +-[完整原因分析数据](../reports/kimodo_root_training_cause_20261023.json)。原始逐层、梯度、参数替换、数据采样及后端记录均在 `/root/autodl-tmp/cache/kimodo-formal-validation-20261023/root-probe/`。 +diff --git a/docs/KIMODO_TRAINING_REPAIR.md b/docs/KIMODO_TRAINING_REPAIR.md +deleted file mode 100644 +index 013ccff..0000000 +--- a/docs/KIMODO_TRAINING_REPAIR.md ++++ /dev/null +@@ -1,50 +0,0 @@ +-2026-10-23:修复 Kimodo 的训练方式,针对根 Transformer 在此前长程微调中失去时序差异的问题。输入仍为按时间释放的英文 GT 文本、最近 30 秒上下文和已执行人体历史;没有新增未来文本、阶段标签、GT 推理历史或轨迹修正。 +- +-这次修改训练本身,未加入退化检测、阈值触发、自动停止或自动替换权重。 +- +-- 根阶段保留官方预训练的注意力、FFN、归一化、文本和时间嵌入,只训练输入/输出投影,共 565,253 个参数;身体阶段训练 141,529,356 个参数。冻结的根骨干仍参与前向和对输入投影的反向传播。 +-- 根和身体分组优化、分别裁剪梯度。根学习率 2e-6→2e-7,在 20K 步衰减到底;身体 1e-5→1e-6,在 100K 步衰减;均有 500 步 warmup。末端根学习率在 20K 后保持下限。 +-- 最大噪声强制采样比例从 0.5 降至 0.1,其余 timestep 仍均匀采样。 +-- 保留有效未来帧的 normalized-x0 监督,前 12 帧执行段权重为 1,剩余规划段为 0.25。 +-- 加入原始根网络对已知历史的重建损失,读取硬覆盖历史之前的网络输出。目标只来自已知过去,不使用未来作为观测。 +-- 对实际解码的根速度、关节速度、旋转变化、首帧接缝速度和根加速度监督;原来的独立速度辅助通道不能替代这些目标。 +-- 根注意力使用数学后端,身体保留原加速实现。运行配置保存进 checkpoint,推理加载时自动恢复。 +-- 完整 checkpoint 保留模型、分组 AdamW、调度器、随机状态及 native session 槽位,继续支持 `--resume`。 +- +-正式配置仍为全量 native 英文人体配对数据、batch8、100K 步、GPU0。缓存与官方预训练资产复用;新训练不继承失败的 100K 微调权重。 +- +-```bash +-cd /root/autodl-tmp/hfd-baseline-portable +-/root/autodl-tmp/envs/kimodo/bin/python -u train_kimodo.py --config configs/kimodo.yaml +-``` +- +-中断后在相同命令末尾追加 `--resume`。进程可在 screen 中前台运行。 +- +-用户要求删除的旧目录已经完整移除,包括其中的 best、last、98K/100K 快照、日志和脚本副本: +- +-- `/root/autodl-tmp/runs/hfd-kimodo-english-train` +-- `/root/autodl-tmp/models/kimodo/hfd-english-train` +- +-这两个路径后来由修复后的正式训练重新创建,当前包含新一轮100K产物。删除旧失败训练时,官方模型、原始数据、共用文本缓存、GenMo活跃训练及既有诊断视频均保留。 +- +-验收报告:[kimodo_training_repair_20261023.json](../reports/kimodo_training_repair_20261023.json)。 +- +-- 30 项单元测试通过,覆盖真实两阶段小网络的冻结/反传、原始历史监督、分阶段优化与恢复、因果表示、解码运动损失和 session 槽位恢复。全尺寸官方网络离线检查通过,根/身体均有有效梯度、参数实际更新、文本条件有效。 +-- 全量入口实际读取 155,117 个 train、4,719 个 val session,完成2步保存,再恢复到4步;AdamW和调度器均到第4步,8个native session槽位随断点恢复。短测试尚未触发2K周期验证,`best_validation`在正式JSON报告中记为null,表示尚未选best。 +-- 六段 train session 小样本训练到8K,1K时执行过完整恢复;六段val不参与更新。固定GT历史验证loss由0.577降至1K的0.222,8K为0.152。所有8,000步记录的loss有限。验证loss不是闭环动作成功率。 +-- 8K根骨干的200个参数张量与官方初始化逐位一致;四个根投影参数及204个身体参数张量均已更新。同一后退窗口、t=0时,根末层时序RMS为0.0364,输入梯度范数为0.00944;旧失败100K分别约1.08e-7、4.76e-12。本次检查未重现旧恒定输出失效,不能外推保证100K质量。 +-- 六类val和两段train完成单帧GT初始化后的全程预测历史生成,整段预测和人体GT分别GMR到G1;导出8段有声视频。帧数、30Hz、音轨及完整音视频解码全部通过,也抽帧查看了预测/参考和文本显示。 +- +-同六段val、同seed/DDIM20/CFG2的等权平均如下;1K只做人体空间诊断,未单独retarget: +- +-| 模型 | 人体姿态 RMSE(rad) | 人体根位置 RMSE(m) | 人体根速度 RMSE(m/s) | G1关节 RMSE(rad) | G1根位置 RMSE(m) | +-|---|---:|---:|---:|---:|---:| +-| 官方初始化+当前适配 | 0.627 | 0.502 | 0.269 | 0.464 | 0.433 | +-| 新训练1K | 0.411 | 0.476 | 0.150 | — | — | +-| 新训练8K | 0.418 | 0.602 | 0.328 | 0.367 | 0.519 | +- +-动作质量仍未通过完整闭环验收:8K姿态误差优于官方初始化,但根位置/速度相较1K恶化,行走与停止样本仍有轨迹漂移;此前后退任务也不能宣称已解决。该实验验证训练链路及旧根骨干失效是否重现,不把增加小样本训练步数视为质量保证。全量训练命令可执行,最终效果仍需独立闭环评测;这不是完整HFD benchmark。 +- +-视频目录:`/root/autodl-tmp/runs/kimodo-training-repair-20261023/`,每例只有`retargeted_human_comparison.mp4`,左侧预测、右侧同流程retarget的human GT。工作数据、曲线和校验日志位于`/root/autodl-tmp/cache/kimodo-training-repair-20261023/`。本次临时checkpoint验收后全部删除,未启动新的100K正式训练。 +- +-2026-10-24后续:用户已启动并完成修复版100K训练,新正式权重保留。根骨干及其时序/梯度复查通过;六类闭环、整段retarget及18段有声视频见[新100K评测](HUMAN_FORMAL_VALIDATION_20261024.md)。前面的8K数值属于修复时的小样本验收。 +diff --git a/docs/LOCAL_INDEX_TRAINING.md b/docs/LOCAL_INDEX_TRAINING.md +deleted file mode 100644 +index ba1d40e..0000000 +--- a/docs/LOCAL_INDEX_TRAINING.md ++++ /dev/null +@@ -1,71 +0,0 @@ +-# 本地索引镜像与训练性能(2026-10-19) +- +-优化已完成,两个旧训练进程已按用户授权停止,原输出目录已删除;screen 保留。旧日志和配置保存在本报告证据目录,新正式训练尚未自动启动。当前正常训练命令见文末。 +- +-## 原因与改动 +- +-64 次实际合法窗口采样的 CPU 剖析耗时 50.64 秒,其中 tar 索引查询累计 33.52 秒,SQLite execute/fetch 占 33.80 秒,实际 tar 字节读取仅约 2.88 秒。主要瓶颈是对共享盘 `index.sqlite` 的大量小查询,而不是显存不足或大文件吞吐。 +- +-新增 `hfd_bridge/index_cache.py`:把约 **433.59 MiB** 的原始只读索引镜像到数据盘 `cache/hfd/index-mirrors/<身份>/index.sqlite`。按源路径、大小、修改时间和 SQLite 头部身份区分版本,使用进程文件锁和原子替换;复制时检查源未变化,并核对源复制流和本地文件的 SHA256。明确指定 `HFD_INDEX_PATH` 时仍尊重该覆盖路径。 +- +-共享 tar 读取器和第三方 session provider 都使用同一个索引入口。原始 tar 保留共享盘;这是元数据索引的本地副本,不是把全量 session 或原始数据复制到内存/本地盘。它可以供后续使用相同读取层的 baseline 复用。 +- +-正式 OMG/TextOp 及 reference YAML 开启 `runtime.environment.HFD_LOCAL_INDEX_CACHE: '1'`;设为 `'0'` 可做旧路径对照。正式 batch512、刷新频率、语言/文本协议、学习率和100K/300K步计划保持原配置。当前优化没有更改原生组装算法、训练循环或数据采样。 +- +-## 对照证据 +- +-CPU ABBA 对照(远端、本地、本地、远端)同64个样本依次为 **48.81、3.95、3.61、10.46 秒**。两次远端时长差异说明文件/页缓存有明显影响;即便采用已经变热的最后一轮远端,本地仍更快。四轮采样ID、动作、音频、控制标签及 reference segment 内容哈希一致。 +- +-随后使用两个真实原生训练循环,batch128,每种路径经历3次数据池准备。OMG48步,TextOp96步;同模型两种路径使用相同初始化、seed、更新预算、采样ID与窗口帧。以下为训练段耗时,包含组池、更新和有界验证,排除索引/模型初始化: +- +-| 模型 | 远端索引秒 | 本地索引秒 | 加速比 | +-| --- | ---: | ---: | ---: | +-| OMG | 251.6 | 33.9 | 7.42× | +-| TEXTOP | 219.9 | 35.6 | 6.18× | +- +-OMG两条loss曲线逐项相同,TextOp最大绝对差为2.98e-8。两种路径的采样ID与窗口帧完全一致。第二种路径复用了原始CPU索引,故不能把包含冷启动的总墙钟差异全算成优化收益。此轮与旧正式训练并行运行,存在共享资源干扰;这些加速比只描述该短测,不能直接当成100K/300K完整训练加速承诺。 +- +-## 正式 batch512 复核 +- +-确认上述加速和135项回归通过后,先停止并清理旧正式训练,再独立运行两张卡各512次更新、每模型两个完整数据池刷新周期。采用当前正式batch512与刷新频率,短学习率/阶段预算和有界验证用于测速,权重不作为质量结论。 +- +-| 模型 | 训练段秒 | 两次组池秒 | 池内秒/步中位数 | 训练段实例/秒 | +-| --- | ---: | --- | ---: | ---: | +-| OMG | 283.8 | 38.0, 53.5 | 0.271 | 924 | +-| TEXTOP | 183.9 | 41.4, 40.3 | 0.143 | 1426 | +- +-训练段包含组池与最终验证/周期保存,初始化墙钟另记在JSON中。TextOp的实例按每次更新的primitive样本计,OMG为动作窗;两模型绝对实例/秒不能横向解释为架构优劣。池内中位数也不包含完整的冷组池时间,长程耗时应使用整体训练段与后续正式验证成本估算。 +- +-两模型各记录512次更新、有限训练loss,保存的Adam状态计数均为512,且恰好准备2个训练池,无多余尾池。临时测速checkpoint已清理,仅保留配置、采样、指标和日志。性能变好不代表动作质量变好;英文前景oracle文本及其已有局限保持原协议,仍需正式验证曲线与连续回放判断质量。 +- +-## 停止、清理与重启 +- +-仅停止已核验命令行和工作目录的旧OMG PID322001、TextOp PID322362。OMG收到SIGTERM后未及时退出,再使用SIGKILL;TextOp由SIGTERM结束。以下两个目录已删除: +- +-- `/root/autodl-tmp/runs/hfd-omg-english-train` +-- `/root/autodl-tmp/runs/hfd-textop-english-train` +- +-小体积日志、配置、数据manifest和采样记录保存在 [old-training](/root/autodl-tmp/runs/archive/legacy-20261020/data-pipeline-optimization-20261019/old-training),checkpoint未保留;原始数据、官方权重、其他历史run和screen未清理。测试用索引副本已去重,保留一份当前有效的本地镜像。 +- +-在两个screen中分别从头启动,不加 `--resume`: +- +-```bash +-cd /root/autodl-tmp/hfd-baseline-portable +-/root/autodl-tmp/envs/omg/bin/python -u train.py --config configs/omg.yaml +-``` +- +-```bash +-cd /root/autodl-tmp/hfd-baseline-portable +-/root/autodl-tmp/envs/textop/bin/python -u train.py --config configs/textop.yaml +-``` +- +-## 验证文件 +- +-- [加速与输入/loss对照](/root/autodl-tmp/runs/archive/legacy-20261020/data-pipeline-optimization-20261019/speed-gate.json) +-- [CPU ABBA与输入哈希](/root/autodl-tmp/runs/archive/legacy-20261020/data-pipeline-optimization-20261019/index-abba.json) +-- [CPU剖析](/root/autodl-tmp/runs/archive/legacy-20261020/data-pipeline-optimization-20261019/provider-profile.txt) +-- [OMG batch512](/root/autodl-tmp/runs/archive/legacy-20261020/data-pipeline-optimization-20261019/omg-full512.json) +-- [TextOp batch512](/root/autodl-tmp/runs/archive/legacy-20261020/data-pipeline-optimization-20261019/textop-full512.json) +-- [旧训练停止与目录删除记录](/root/autodl-tmp/runs/archive/legacy-20261020/data-pipeline-optimization-20261019/old-training-cleanup.json) +- +-分环境135项回归测试与两套正式配置dry-run通过。性能探针保存在仓库外的本报告证据目录,没有重新堆入日常 `tools/`。 +diff --git a/docs/MIGRATION.md b/docs/MIGRATION.md +deleted file mode 100644 +index d2d72fe..0000000 +--- a/docs/MIGRATION.md ++++ /dev/null +@@ -1,5 +0,0 @@ +-# 迁移入口 +- +-当前完整操作见[训练复现指南](PORTABLE_TRAINING.md):四模型环境、外部资源、原始tar、英文词对齐、路径生成和正式训练命令。 +- +-本版本发布训练代码、配置、资源版本元数据和文档。本机测试/评测在被忽略的`test/`。历史ASR、双语和过拟合配置不作为当前默认。旧迁移说明的本地副本在`test/docs/`,未随Git发布。 +diff --git a/docs/OMG_ENGLISH_VALIDATION.md b/docs/OMG_ENGLISH_VALIDATION.md +deleted file mode 100644 +index 24fe908..0000000 +--- a/docs/OMG_ENGLISH_VALIDATION.md ++++ /dev/null +@@ -1,56 +0,0 @@ +-# OMG 英文对齐文本训练后的评测 +- +-`evaluate_omg.py` 从正式训练的 `resolved_config.yaml`、`data_manifest.json` 和 +-`val_native_manifest.json` 恢复验证协议,调用相同的原生及第三方数据 provider。 +-它只支持当前英文 `aligned_reference_context` / `microphone` OMG 配置。 +- +-默认选六类原生场景各一条6–30秒英文 session,另选三条第三方英文验证 session。 +-原生按索引取首个符合条件的样本,第三方在合格索引中等距取样;选择发生在生成之前。 +-这不是按类别比例采样的全量验证估计。`test` 是 `val` 的别名,验证数据用于过训练监测。 +- +-```bash +-cd /root/autodl-tmp/hfd-baseline-portable +-/root/autodl-tmp/envs/omg/bin/python -u evaluate_omg.py \ +- --trained-run /root/autodl-tmp/runs/hfd-omg-english-train \ +- --output /root/autodl-tmp/runs/omg-english-validation-20261020 \ +- --gpu 1 +-``` +- +-默认对比官方初始化与 `best/best.ckpt`。可用 `--checkpoint PATH` 显式选择新权重, +-`--per-mode N` / `--third-party N` 控制样本数,`--stage prepare|run|summarize` +-分阶段运行。相同参数重跑会复用完整结果;不完整推理目录须检查后另选输出目录或移走, +-不会静默覆盖。不会启动训练、识别、翻译、强制对齐或下载。 +- +-## 与训练一致的输入 +- +-- 相同英文过滤、配方、原生种子、验证有序索引及第三方 selection。 +-- 原始混合 PCM;过去2秒音频。词级离线对齐结果来自训练用缓存。 +-- 截至当前释放点的最近30秒 GT 前景文本,0.1秒 guard,0.6+0.4k 秒更新。 +-- 原生 T5-Base 保留最近50 token;实际 token ID、截断方向、PCM截止和文字释放时间记录在规划日志。 +-- 动作时间到麦克风时间偏移0.2秒,音视频对预测和GT共同前置0.2秒显示。 +- +-评测冻结时逐值校验原生导出和 compact provider 的音频、动作、标签及有效起点数, +-逐个50Hz状态帧比较导出 sidecar 和训练 provider 的文字条件。第三方直接导出训练 provider 数组。 +-标签、完整原文和未来动作不传入规划器;标签仅用于计分分层,事件元数据只用于视频标注。 +- +-## 连续推理与指标 +- +-两份权重用相同噪声seed12345、DDIM10、joint_one、剩余24帧触发的原生异步缓冲。 +-公平对照固定规划延迟为4个30Hz帧(约133ms),同时单独记录实际规划耗时。 +-这不是包含前端和部署调度的实测端到端延迟。仅开头13帧动作作GT初始化,之后反馈已执行预测, +-不按窗口重置GT。初始化及共同显示前缀不计分。 +- +-统计29关节RMSE、根xyz分量RMSE、根朝向误差、FK链接位置误差,以及listen/motion/idle分层误差。 +-标签索引先减去视频共同前缀,再转换到50Hz原始动作时钟,避免使用未来标签。 +-没有DTW或事后根位置对齐;这些误差衡量参考跟随,不等于语义成功率。 +- +-OMG直接输出G1的29关节qpos,不需要人体retarget。渲染为运动学预测/GT对照,带原始混合音轨, +-不是物理tracker。所有视频完整解码,音轨核验;两份权重的GT和音频必须相同。 +- +-输出包含 `README.md` 视频索引、`metrics.csv`、`summary.json`、`verification.json`、 +-权重SHA256、冻结样本与对齐sidecar、逐次推理/渲染日志。 +-官方预训练重合、第三方与原生动作源重合未审计;小样本单seed不能替代完整评测。 +- +-视频封装保留完整H.264视频流,不再用 `-shortest`:实测相同长度PCM与含B帧的视频流封装时, +-该参数可能裁掉末尾视频包(508帧样本曾变为505帧)。新增实际编码/全帧解码回归测试; +-渲染中间视频只在最终MP4帧数、帧率和音轨通过检查后删除。 +diff --git a/docs/OMG_TRAINING_REVIEW.md b/docs/OMG_TRAINING_REVIEW.md +deleted file mode 100644 +index 9002c7c..0000000 +--- a/docs/OMG_TRAINING_REVIEW.md ++++ /dev/null +@@ -1,105 +0,0 @@ +-# OMG 正式训练复查与吞吐优化(2026-10-17) +- +-清理说明:本文记录的一次性实验脚本已归档,复现前请按[脚本恢复说明](WORKSPACE_CLEANUP.md)取回;历史指标与日志保留。 +- +-本轮输出:`/root/autodl-tmp/runs/archive/legacy-20261020/hfd-omg-train-20261010-v2`。已正常完成 **100,000 次优化器更新**,不是被 OOM 或异常中断。screen 的 `Trainer.fit stopped: max_steps=100000 reached` 和 checkpoint 的 `global_step=100000` 一致。训练循环耗时 108 小时 33 分 7 秒,整个命令约 108 小时 45 分,约 4.53 天。完成时间为 2026-10-15 01:30(Asia/Taipei)。 +- +-## 本轮结果 +- +-| 指标 | 结果 | +-| --- | --- | +-| 模型 | OMG 原生 50m 配方;denoiser 参数约 54.14M,另有冻结的 T5 等模块 | +-| 训练 | RTX 5090 × 2;每卡 batch 512,总 batch 1024;bf16 mixed | +-| 训练 loss,前 1,000 步平均 | 0.04943 | +-| 训练 loss,最后 10,000 步平均 | 0.02315 | +-| 首次验证,2,000 步 | 0.0365847 | +-| 最低日志验证 loss,90,000 步 | 0.0220401 | +-| 最后验证,100,000 步 | 0.0234458 | +-| 最终权重检查 | 可加载,394 个 state tensors 均无 NaN/Inf,含 optimizer/scheduler 状态 | +- +-![训练与验证曲线](../reports/omg_training_20261017.png) +- +-训练 loss 是 rank 0 的原生扩散损失;验证 loss 由两卡同步归约。它们不是厘米级动作误差,也不是任务成功率。`val/motion_loss=0` 来自原生 `diffusion_only` 配方,不代表动作没有监督;动作特征是扩散损失的预测目标。这里只确认训练收敛趋势和权重健康,尚未完成论文独立测试集或连续 rollout 评测。 +- +-最终文件是 `last.ckpt`。保留的周期权重是 `checkpoints/step-00096000.ckpt`、`step-00098000.ckpt` 和 `step-00100000.ckpt`;`checkpoints/last.ckpt` 是周期保存的链接。原策略按步数保留最近三份,**90,000 步的最低 loss 权重已被删除,不能恢复为 best**。现存三份记录的验证 loss 分别是 0.0222620、0.0223724、0.0234458;要据此选模型,应在同一固定验证样本上重新比较。 +- +-旧验证并不固定:每次验证每卡 16 个 batch,池每累计 256 个验证 batch 换一次。因此 2k、34k、66k、98k 步分别开始使用不同的验证 session 池;池内窗口和扩散噪声也不断变化。最后两个 loss 不能直接与 96k 比较后认定过拟合。 +- +-本轮 manifest 记录 train=322,243、val=9,272 个原生候选 session,`command_interruption_session_variants=2`。两卡记录合计访问 207,173 个不同训练 session,约占候选的 64.3%;有放回随机窗口训练不等于完整遍历一个 epoch。**本轮不包含后来加入的 third-party 包**:当前 `configs/sessions.yaml` 已有 `include_third_party: true`,不能用今天的配置文件反向描述旧实验。复查性能时使用关闭 third-party 的临时配方,重放旧 run 的 session 索引;当前正式数据配置未被改回旧值。 +- +-## 是否使用 OMG 原生脚本 +- +-入口是 portable 的 `train.py::omg`,没有直接调用上游 `omg.cli.generation.train.main`。入口组合上游 `exp=50m`,加载官方模型定义及初始化权重,再构造 `pytorch_lightning.Trainer`,训练、验证、优化器和 scheduler 仍调用上游 `MotionGenerator` 的实现。替换的是 HFD 数据适配器和训练管理 callback。准确表述是“使用 OMG 原生模型及训练步骤,通过 HFD 适配入口训练”。 +- +-本轮保持原生 100k 更新、2k warmup、6e-5 学习率及 cosine schedule。数据输入是过去两秒已释放音频 + 因果 ASR 当前指令,完整 session 动作监督;未改变动作映射、音频截止时间或模型目标。 +- +-## 慢在哪里 +- +-1. 每卡 batch=512 的 session 池每 256 个 batch 刷新。本轮 391 次刷新,取每次两个 rank 中较慢的准备时间累加约 **59.22 小时**,约占训练循环 54.6%。这包含共享 tar 读取、原生 session 组装、ASR、运动预处理和首个 batch,不能全算成磁盘 I/O。rank 0 的刷新中位数 493.6 秒。 +-2. 四次验证池刷新另需约 1.10 小时。完整训练平均每步 3.91 秒,旧日志没有每步数据/计算细分计时。 +-3. ASR SQLite 有 209,704 条记录、约 1.18 GiB 压缩有效内容。旧实现每插入一条都 `SUM(length(data))` 扫全表;复查时同一只读扫描缓存热后仍约 0.20–0.29 秒。这是可去除的元数据开销,不是模型计算。 +-4. 旧普通 batch 逐个调用 512 次 motion codec,许多小 CPU tensor 运算串行执行。音频特征此前已做批处理,动作特征仍有批处理空间。 +-5. tar `Member.readinto` 对小读请求仍取一个 256 KiB 块,且本地 tar 路径没有块复用。`BufferedReader`/NPZ 的小读和回读可反复读取同一块。 +- +-`num_workers=63/127` 是 Lightning 通用提示。当前数据对象持有 SQLite、状态采样器以及同卡 Whisper,且明确禁止 DataLoader worker;仅修改 `num_workers` 会被配置校验拒绝。若做并行预取,需要把 CPU 读盘/组装与 GPU ASR 分开,并管理 rank/worker 的采样状态,不能直接照提示填值。 +- +-## 已落地的改动 +- +-- 动作预处理:用同一个 OMG 原生 codec 一次计算整批 10 帧历史与 60 帧目标。`data.batched_motion=true`,可设为 false 回退逐样本实现。测试比较全部 batch 字段,容差 `rtol=atol=1e-6`,保持动作定义与时间对齐。 +-- ASR SQLite:增加由数据库 trigger 原子维护的容量计数,首次迁移扫描一次,此后不再每次插入扫全表。兼容已有 ASR 内容与 key,两卡连接共同维护;访问时间最多每分钟写一次,减少热缓存写锁。原有 2 GiB 有效内容上限仍生效;该上限不等于 SQLite 文件/WAL 的物理上限。 +-- tar I/O:每个打开的成员保留最多四个 256 KiB 块,减少小读放大;关闭时释放。原始 tar 留在共享盘,不解压,不复制完整数据集。 +-- 训练管理:固定验证窗口与扩散随机数,验证结束恢复训练 RNG;单独保存 `best/best.ckpt`,不受最近三份周期权重清理影响;记录每步等待/计算时间及最终训练摘要。时间字段是主机墙钟,`data_wait_seconds` 可能包含周期 checkpoint 等 callback 开销,不等于纯磁盘 I/O。 +-- 可选早停:原生 Lightning `EarlyStopping`,监测 `val/loss`,从 30k 步开始计数,`patience=10`,`min_delta=0.0002`。每 2k 步验证,因此对应连续 20k 步没有足够改进。patience 计算的是验证次数,见 [Lightning 官方说明](https://lightning.ai/docs/pytorch/latest/optional-api/api/lightning.pytorch.callbacks.EarlyStopping)。 +- +-默认 `training.fixed_validation=true`、`training.save_best=true`,但 `training.early_stopping.enabled=false`,保留原生固定 100k 预算。要选择早停训练,在原训练命令上加 `--set training.early_stopping.enabled=true paths.output_dir=/root/autodl-tmp/runs/新的目录`。这会成为不同的训练预算,论文中须记录。 +- +-固定验证复用同一组每卡 16 个 batch(双卡 16,384 次窗口抽样,窗口允许重复),仍只是监测子集。不能按旧随机验证日志选择最有利的早停阈值:事后机械套 patience=10 会在 88k 停,随后 90k 还有改善,说明不能保证早停明显省时且完全不损失质量。优化每步吞吐比激进缩短训练更有依据。 +- +-## 验证和性能测量 +- +-单元测试覆盖原生动作 codec 对照、固定验证样本、验证 RNG 恢复、真实 Lightning 早停/最佳权重保存、旧 SQLite 迁移与双连接容量计数、tar 小读/seek/EOF 等。性能测量仅重放旧训练池、运行限定步数,不保存更新后的训练权重,也不改写已完成 run。 +- +-本机测量记录集中在 `/root/autodl-tmp/cache/hfd/review-20261017/`,不提交原始数据或大文件。47 项相关测试与 12 个 subtests 通过,配置 dry-run 和 `git diff --check` 通过。 +- +-同一旧 run 的 512-session 池,4 CPU 线程,比较结果: +- +-| 测量口径 | 优化前 | 优化后 | +-| --- | ---: | ---: | +-| session 已缓存、仍有新音频窗口的 batch 数据准备 | 1.68 秒 | 0.79–0.80 秒 | +-| 同一进程、同一批窗口且音频已缓存,逐样本/批量动作 codec 对照 | 0.91–0.92 秒 | 0.048–0.078 秒 | +-| 209,704 条 ASR 记录的容量查询 | 约 0.16 秒/次全表扫描 | 约 2–44 微秒/次读计数 | +- +-真实 batch 对照比较了所有 tensor、caption 和索引,`rtol=atol=1e-6` 通过。热 session、新音频窗口的数据准备约快 **2.1 倍**;音频全命中的对照展示了消除逐样本动作计算后的收益,不能把该倍数当成整轮训练加速。 +- +-两次首次装入同一池分别耗时约 417 和 209 秒,数组读取/原生组装分别占约 407 和 200 秒。它们共享文件系统缓存且执行有重叠,不是受控冷盘对照,不能归因于代码修改。新 tar 块缓存有读取一致性与读放大量测试,但尚无隔离的整轮冷读测速。 +- +-双卡端到端短测已完成:加载本轮最终 checkpoint,在真实 HFD 数据上每卡 batch=512,各完成 6 次前向、反向与优化器更新;所有记录的 loss 有限,denoiser 确认为训练模式。第 2–6 步完整迭代为 **1.25–1.32 秒/步,均值约 1.28 秒**,模型计算约 0.42–0.50 秒/步。每卡 PyTorch 峰值 allocated 显存约 16.42 GiB;本次 ASR 已命中缓存,没有同时运行缺失文本的 Whisper 识别,不能据此提高正式 batch。 +- +-首次准备在 rank 0/1 分别是 68.9/371.2 秒;rank 0 第一轮又等待约 304 秒,而后两卡进入上述稳定吞吐。这个对照直接展示了共享盘读取/原生组装的不均衡及 DDP 等待。短测没有跨过 256-step 池刷新边界,不能用 1.28 秒直接乘 100k 预测完整训练时间。检查已结束,两张 GPU 已释放;没有启动新的正式训练,也没有覆盖本轮 checkpoint。 +- +-下一阶段若刷新仍占主导,优先将因果 ASR 从主训练循环拆成可复用的准备阶段,再考虑有界异步预取。单纯把 `refresh_batches` 从 256 加到 1024 虽然摊薄刷新,却减少相同步数下接触的新 session、增加相邻 batch 相关性,本次没有采用。当前数据盘不足以额外复制约 49.8 GiB 的全部原始 tar,因此也没有把“全部搬到本地”当作立即可行的方案。 +- +-## 异步预取与并行 I/O(2026-10-17 后续) +- +-增加 `data.prefetch_batches`(0 或 1)。开启后,OMG 在 GPU 训练当前 batch 时,用一个后台线程准备下一个已缓存的 CPU batch,包括动作编码和音频特征。最多额外保留一个 CPU batch;validation 仍在主线程执行。 +- +-后台线程不构造新 session、不访问 SQLite、不运行 Whisper,也不执行 GPU 工作。`WindowPool` 保留当前槽位的合法时刻范围,避免热 batch 为查询长度重新组装数组;缓存未就绪或到达刷新边界时,自动回到主线程。固定 seed 下,测试跨多个刷新边界逐一比较动作、音频、caption、采样时刻,并验证训练 RNG 不变。后台异常传回主线程,关闭 iterator 会结束任务。实现使用标准库 [ThreadPoolExecutor](https://docs.python.org/3.10/library/concurrent.futures.html)。 +- +-另外增加本地 tar 成员的有界并行预读:`runtime.environment.HFD_IO_READAHEAD_BLOCKS=4`,`runtime.environment.HFD_IO_WORKERS=4`。线程仅执行本地文件读取,不共享 SQLite 连接,不进行网络下载。大型/连续读取才触发预读,seek 会取消旧预读,每个成员的未来块数有界;原有四块缓存仍保留。它作用于原生 `ArchivePath/Member` 读取路径,third-party 已组装 WAV/NPZ 的直接大块读取路径没有替换。 +- +-8 个实际音频成员的同步→异步→异步→同步对照,全部 SHA256 相同,时间为 0.704、0.094、0.071、0.052 秒。第一次同步读取受冷缓存影响;最后一次同步比异步更快,因此不能把第一次与第二次相除宣称并行 I/O 提升。目前并行 tar 预读**默认关闭**,保留为不同文件系统上的可测选项。 +- +-CPU batch 异步预取则使用双卡 64 步的同步→异步→异步→同步分段对照;每段排除开头两步,以减少流水线切换影响。该测试仍重放旧训练池,不包含 third-party,也不执行新的正式训练。它测的是池内准备与 GPU 计算的重叠,不包含下一池的后台组装/ASR,不能视为解决全部刷新等待。 +- +-最终采用 **OMG 一个热 CPU batch 预取,tar 并行预读默认关闭**。两卡各 64 步均完成,所有 loss 有限,denoiser 确认为 train mode。每段排除两步后,按每一步较慢的 rank 汇总: +- +-| 模式 | 平均每步 | 分段结果(rank 0) | +-| --- | ---: | --- | +-| 同步 | 1.072 秒 | 第一段 1.244 秒;末段 0.897 秒 | +-| CPU batch 异步预取 | 0.617 秒 | 中间两段 0.656 / 0.572 秒 | +- +-该对照约提高池内吞吐 1.74 倍,单步时间减少约 42%。音频缓存会逐步变热,ABBA 仅减轻该影响,不能彻底消除;异步中段也比更晚、更热的同步末段快。CPU/GPU 并发时模型阶段本身略慢,但整体迭代更快。每卡 PyTorch allocated 峰值约 16.42 GiB。首次装载两卡分别约 227/223 秒,此部分未被热 batch 预取隐藏。 +- +-当前默认 `configs/omg.yaml` 中 `data.prefetch_batches: 1`,同步回退为 0;并行 I/O 的 `HFD_IO_READAHEAD_BLOCKS` 默认仍为 0。52 项相关测试及 12 个 subtests 通过,另有真实双卡 64 步检查。这里只运行了有限步数的性能探针,不保存更新权重;GPU 已释放,旧正式训练权重未改写。原始测量在 `ddp-async-abba/{rank0,rank1,summary}.json` 与 `tar-async-abba.json`(位于本机 review 缓存目录)。 +- +-复查命令可使用 `tools/benchmark_omg_training.py --steps 64 --compare-prefetch`(双卡需 torchrun,配合实际配置、`--replay-dir` 和 `--report-dir`);单独 tar 对照使用 `tools/benchmark_omg_io.py --members 8 --report /数据盘/新的测量.json`。这些命令是测试工具,不是正式训练或完整训练时长评估。 +diff --git a/docs/OMG_VALIDATION_DIAGNOSIS.md b/docs/OMG_VALIDATION_DIAGNOSIS.md +deleted file mode 100644 +index 9898751..0000000 +--- a/docs/OMG_VALIDATION_DIAGNOSIS.md ++++ /dev/null +@@ -1,77 +0,0 @@ +-# OMG 验证视频:提前动作与听完不执行的诊断 +- +-清理说明:本文记录的一次性实验脚本已归档,复现前请按[脚本恢复说明](WORKSPACE_CLEANUP.md)取回;历史指标与日志保留。 +- +-本次仅检查冻结的 100k checkpoint,没有训练、改权重、替换模型输入或覆盖原视频。诊断产物位于 `/root/autodl-tmp/runs/archive/legacy-20261020/omg-100k-validation/diagnostics/`。 +- +-## 1. 已确认的中文文本语义丢失 +- +-在实际使用的 `/root/autodl-tmp/models/OMG/t5-base-local` tokenizer 上,以下实际规划文本的 token IDs 完全相同: +- +-| 实际文本 | token IDs | token | +-| --- | --- | --- | +-| 慢跑向前 | `[3, 2, 1]` | `▁, , ` | +-| 小禮母 | `[3, 2, 1]` | 同上 | +-| 慢跑向前跳越中身手架 | `[3, 2, 1]` | 同上 | +-| 慢跑向前跳越中身手构想后跳 | `[3, 2, 1]` | 同上 | +- +-原生 `FrozenT5TextEncoder` 只根据这些 token IDs 和 mask 计算冻结 T5 特征,因此这些输入的文本条件相同,无法区分任务。打断样本第二、第三条中文指令到来后,ASR 字符串虽然变化,模型文本分支却收不到对应的语义变化。35D 声学分支包含频带能量等声学特征,不能假定它会自动补回丢失的指令语义。 +- +-这不只是 Whisper 的中文识别不准;即使正确识别“慢跑向前”,当前文本入口仍会丢失含义。之前只验证长度和截断,没有验证未知词覆盖,是适配验收的遗漏。保留完整历史 ASR 也不能修复该 tokenizer 问题;继续增加同配方训练步数不会恢复已经被映射成同一个 token 的指令信息。 +- +-`clock-and-token-audit.json` 保存六个视频每次规划实际文本及 token。用户指出的 `interruption-stable_latest_sentence.mp4` 对应 `val-000064-seed42`: +- +-- 视频任务语音起点为 1.6、9.2、14.8 秒。 +-- 进入模型的后续文本含“慢跑向前跳越中身手架”等 ASR 错误,且中文段均发生上述 token 合并。 +-- 预测根位置相对开场的最大水平位移仅 0.0242 m,GT 为 4.0543 m;模型没有跟随参考里的明显移动。 +- +-这些证据确认输入适配有问题,不能直接把该样本的失败归因为训练步数不足。 +- +-## 2. 提前动作包含生成模型自身的提前起步 +- +-六条视频的首次明显姿态变化几乎都在 1.6–1.7 秒,尽管语音起点不同。这里的“明显变化”定义为相对开场姿态、29 关节 RMS 偏移首次超过 0.05 rad,不是一个训练出的倾听分类器。 +- +-尤其是英文紧急停止样本 `val-000111-seed42`: +- +-- 视频指令语音到 2.4 秒才开始;GT 姿态超过阈值在约 3.47 秒。 +-- 预测在约 1.63 秒已经超过阈值,呈视频中的倾听/准备样式。 +-- 产生这些姿态的首轮规划只读到 PCM 1.0 秒,实际任务语音到 PCM 2.6 秒才开始;首轮 ASR 为背景上的幻觉 `you`。 +- +-对同一 checkpoint、同一开场历史、同一 seed,重新生成首个 60 帧计划,实际条件输出与保存视频的已执行部分匹配。分别去文本、去音频、两者都去掉,仍会产生相近幅度的动作: +- +-| 首轮条件 | 最大关节姿态偏移 RMS | +-| --- | ---: | +-| 实际音频+文本 | 0.5723 rad | +-| 去文本 | 0.5752 rad | +-| 去音频 | 0.5722 rad | +-| 两者都去掉 | 0.5756 rad | +- +-这是首轮计划诊断,不是完整 session 消融,也不表示屏蔽条件时原生扩散模型本应静止。它表明这些动作无需当前语音指令触发即可生成。当前模型只预测姿态,没有显式“检测到指令后才能进入倾听/执行”的门控;HFD 所需的等待和触发行为尚未被可靠学到。不能将其解释为模型提前听懂了后面的指令。 +- +-数值见 `silent-first-plan-audit.json`,轨迹见 `silent-first-plans.npz` 和 `early-motion.png`。 +- +-## 3. 现有动作轴与音频释放轴还会影响视频观感 +- +-现有评测沿用训练的 HFD 时钟,动作组 `k` 使用截止 `0.6+0.4*k` 秒的 PCM;视频按动作时间显示,播放的音轨却取原 PCM 的 `t+0.2` 秒位置。这不是统一麦克风墙钟的真实在线系统。 +- +-按实际计划交接记录换算后,打断视频有新计划开始执行时、输入音频内容比视频当时播放位置领先约 0.167 秒的区间;停止视频最大约 0.233 秒。它会影响“提前/延迟”的直观判断,当前视频不应直接用于真实响应延迟评估。 +- +-这个小幅度时钟领先不能解释停止样本中约 0.77 秒的提前起步:首轮规划根本未包含任务语音,且无条件首轮仍产生动作。两件事应分开处理。修正评测墙钟后需要重新核对动作标签和训练时的释放映射,不能只把音轨随意平移来美化视频。 +- +-## 4. 仍有生成/闭环能力问题,但尚未完全归因 +- +-英文跳跃样本文本能正常编码;预测根高度范围仅约 0.0633 m,GT 为约 0.3907 m。因此中文入口并非全部问题。现有模型、训练窗口和闭环规划组合仍有动作幅度与阶段推进不足。 +- +-异步规划执行旧缓冲、替换未执行的计划后半段,模型是否反复把动作推迟到未来,以及真实历史与预测历史的差异,尚需固定条件、固定交接时刻和历史模式对照才能分解。不能仅凭当前视频断言优化器错误、模型架构无能力或某个单一训练参数是根因。 +- +-后续优先级:先修复中文语义入口并审计五个模型的实际 tokenizer 覆盖;再统一评测墙钟;最后区分模型等待/触发学习和闭环规划问题,再决定重训配方。完整历史 ASR 仍是用户确认的共同策略,但不能把“完整文本存在”当作“模型获得了文本语义”。 +- +-复现主诊断: +- +-```bash +-# diagnostics 目录必须尚不存在;保留原诊断时改用新的评测根目录副本。 +-/root/autodl-tmp/envs/omg/bin/python tools/diagnose_omg_validation.py --root /root/autodl-tmp/runs/archive/legacy-20261020/omg-100k-validation +-``` +- +- +-后续输入与时钟修复已实现,旧诊断结果保持不变,见 [修复记录](SEMANTICS_CLOCK_FIX.md)。 +diff --git a/docs/OMG_VALIDATION_ROLLOUT.md b/docs/OMG_VALIDATION_ROLLOUT.md +deleted file mode 100644 +index a7cff48..0000000 +--- a/docs/OMG_VALIDATION_ROLLOUT.md ++++ /dev/null +@@ -1,67 +0,0 @@ +-# 已完成训练的 OMG:六类验证 session 连续推理 +- +-清理说明:本文记录的一次性实验脚本已归档,复现前请按[脚本恢复说明](WORKSPACE_CLEANUP.md)取回;历史指标与日志保留。 +- +-结果目录:`/root/autodl-tmp/runs/archive/legacy-20261020/omg-100k-validation/`。 +- +-后续诊断确认了 **T5 中文输入映射为相同 `` token、无任务语音时提前起步,以及视频/释放时钟差异**。详见 [提前动作与不执行诊断](OMG_VALIDATION_DIAGNOSIS.md)。本轮指标包含这些输入适配问题,不能单独用于判断动作模型能力。 +- +-## 数据和权重 +- +-使用 `/root/autodl-tmp/runs/archive/legacy-20261020/hfd-omg-train-20261010-v2/last.ckpt`,确认 `global_step=100000`。没有重新训练或改写权重,评测目录仅用软链接引用原 checkpoint。`evaluation.json` 保存 SHA256;每次推理使用原生模型的严格权重加载。 +- +-仓库有 HFD 原生 train/val 构造入口,数据来自本机 `/root/autodl-fs/pangxinlei/hfd_data` 的 tar 分片。旧训练对应 322243 条 train、9272 条 val session。任务动作源按哈希隔离;验证集曾用于训练监测。**源码中的 `test` 是 val 的别名,没有第三份独立、未使用过的 test 划分。** 公共 idle/listen 素材可能共用;官方 OMG 预训练数据与这些动作的重合没有审计,不能声称模型整个训练过程从未见过它们。 +- +-本次重建的 val 索引指纹 `dc6cb1533a00d2fd983d220628a999cad60023339fbf5abba106cd704d2421dd` 与训练保存的 manifest 一致,任务源哈希划分验证通过。每类选择第一条估计时长 6–30 秒的 session,按类别交替优先选择英文和中文;选择完成后才运行模型,没有根据输出挑样本。具体索引和动作身份见 `selection.json`。 +- +-| 场景 | session index | 语言 | 长度 | +-| --- | ---: | --- | ---: | +-| 轮流指令 | 3 | 英文 | 10.8 s | +-| 任务打断 | 64 | 中文 | 21.6 s | +-| 紧急停止 | 111 | 英文 | 16.8 s | +-| 附和 | 60 | 中文 | 15.6 s | +-| 犹豫 | 28 | 英文 | 16.8 s | +-| 停顿 | 41 | 中文 | 27.2 s | +- +-## 推理设置 +- +-旧权重采用 `stable_latest_sentence` 训练,因此主评测保持原策略;另以同一 checkpoint 测试当前默认 `causal_asr_context`。后者属于输入分布变化对照,不代表已用完整历史重训。 +- +-两组共用相同的因果 ASR sidecar、音频、GT 和 seed=12345。ASR 使用本地 Whisper-tiny、auto 语言、最多 30 秒历史、batch 32;每条输入切片的哈希与释放边界均核验。OMG 输入过去两秒已释放音频及文本,保持原生 10 帧动作历史、60 帧未来、30Hz、DDIM 10 步、joint_one CFG、剩余 24 帧触发的 async 缓冲。异步规划耗时实测,旧新策略的耗时和交接时刻可能不同;ASR 识别计算耗时不计入。 +- +-开场 13 帧 GT bootstrap 不计分,之后始终反馈已执行预测,没有逐窗口 GT 重置。listen/motion/idle 标签仅用于评测分层,不进入推理。输出直接是 G1 qpos,无需人体 retarget;视频仅为运动学回放,没有物理 tracker。 +- +-## 结果 +- +-关节 RMSE 为原生 29 关节,根坐标 RMSE 为 xyz 分量 RMSE;它不是逐帧根位置欧氏距离的均值。所有误差按共同时间轴直接计算,没有 DTW 或事后根位置对齐。 +- +-| 场景 | 原稳定句关节 RMSE | 完整历史关节 RMSE | 原稳定句根坐标 RMSE | +-| --- | ---: | ---: | ---: | +-| 轮流指令 | 0.2501 rad | 0.2264 rad | 0.0915 m | +-| 任务打断 | 0.3253 rad | 0.3244 rad | 1.6924 m | +-| 紧急停止 | 0.4356 rad | 0.4614 rad | 0.6104 m | +-| 附和 | 0.2860 rad | 0.2625 rad | 0.0133 m | +-| 犹豫 | 0.4297 rad | 0.4278 rad | 0.0437 m | +-| 停顿 | 0.2919 rad | 0.2695 rad | 0.5386 m | +-| 六条等权均值 | **0.3364 rad** | **0.3287 rad** | **0.4983 m** | +- +-按计分帧合并的关节 RMSE 为 0.3436 / 0.3393 rad;两种汇总的含义不同,完整记录在 `summary.json`。另保存根朝向误差、FK 链接位置误差、listen/motion/idle 分层误差及实际规划文本非空比例。 +- +-这些验证动作的参考跟随偏差明显高于之前单条 session 的过拟合结果。完整历史 ASR 的平均改善很小,紧急停止样本的关节误差反而上升。部分中文转写有明显识别偏差,例如打断样本的末次文本为“曼跑向前跳躍中身手架我想后跳”;保留实际转写,不用 GT 修正。这是 ASR+OMG 链路的结果,不能把所有偏差归因于动作模型。 +- +-这是六条验证样本、一个采样 seed 的初步检查,不是全量测试或任务成功率评测。不同动作也可能满足同一句话,因此 GT RMSE 不能单独判断语义是否正确。未提供停止响应延迟、任务成功率或物理稳定性结论。 +- +-## 查看和核验 +- +-打开结果目录的 `README.md` 可逐类访问两种策略的 12 条视频,均为左侧预测、右侧 GT,带原始混合音轨;`comparison.png`、`metrics.csv` 和 `summary.json` 分别为图表、表格和完整数值。`video_contact_sheet.jpg` 是中段视频帧概览。部分中文视频字幕受现有字体支持限制显示为问号,实际 Unicode 输入保存在各条 `report.json` 和 ASR sidecar 中。 +- +-12 条视频共 6528 帧完整解码,音轨检查通过;两种策略的 GT 数组逐值一致,ASR PCM 切片哈希通过。13 项相关缓冲、关节映射、文本因果性测试通过,记录在 `verification.json`。此次只读取本地 tar,没有下载数据。 +- +-## 复现 +- +-```bash +-# output 必须为新目录;从原训练记录恢复验证配方 +-/root/autodl-tmp/envs/omg/bin/python tools/prepare_omg_validation.py --trained-run /root/autodl-tmp/runs/archive/legacy-20261020/hfd-omg-train-20261010-v2 --output /root/autodl-tmp/runs/archive/legacy-20261020/omg-100k-validation +-/root/autodl-tmp/envs/omg/bin/python tools/run_omg_validation.py --root /root/autodl-tmp/runs/archive/legacy-20261020/omg-100k-validation +-/root/autodl-tmp/envs/omg/bin/python tools/summarize_omg_validation.py --root /root/autodl-tmp/runs/archive/legacy-20261020/omg-100k-validation +-``` +diff --git a/docs/PORTABLE_TRAINING.md b/docs/PORTABLE_TRAINING.md +deleted file mode 100644 +index c73bd69..0000000 +--- a/docs/PORTABLE_TRAINING.md ++++ /dev/null +@@ -1,167 +0,0 @@ +-# 在另一台机器复现训练 +- +-当前四个正式入口及配方见README。HFD/OMG/TextOp原生源码在`upstream/`;GenMo和Kimodo使用`UPSTREAM_LOCK.json`指定的外部源码。模型、原始数据和离线词时间缓存单独准备。本版本不发布测试/评测脚本、正式权重或日志。 +- +-以下以`/data/hfd_work`为工作盘、`/data/hfd_assets`为外部资产盘。已验证环境是Linux、Python3.10、PyTorch2.7.1+cu128、RTX5090 32GB。GenMo训练峰值约25GB;Kimodo还需要同卡常驻的8B冻结文本worker。已有资源可由路径参数复用,不必搬迁。 +- +-## 代码及环境 +- +-AutoDL下载/安装命令先加`/root/.local/bin/turbo`,不要全局source代理。下面下载命令展示该前缀;其他机器使用自身已授权的网络方式。HF失败可使用已配置的`hfmirror`,不能自动转用个人代理。 +- +-```bash +-/root/.local/bin/turbo git clone https://github.com/starstone3/Dual-hfd-portable.git /workspace/hfd-baseline-portable +-cd /workspace/hfd-baseline-portable +-mkdir -p /data/hfd_work/envs /data/hfd_work/hfd-human-models +-``` +- +-创建独立Python3.10环境。完整冻结依赖在`requirements/`;GenMo、Kimodo和forced-align的Torch版本另列于`runtime-versions.json`: +- +-```bash +-python3.10 -m venv /data/hfd_work/envs/genmo +-/root/.local/bin/turbo /data/hfd_work/envs/genmo/bin/python -m pip install \ +- --index-url https://download.pytorch.org/whl/cu128 torch==2.7.1 torchvision==0.22.1 +-/root/.local/bin/turbo /data/hfd_work/envs/genmo/bin/python -m pip install -r requirements/genmo.txt +- +-python3.10 -m venv /data/hfd_work/envs/kimodo +-/root/.local/bin/turbo /data/hfd_work/envs/kimodo/bin/python -m pip install \ +- --index-url https://download.pytorch.org/whl/cu128 torch==2.7.1 torchvision==0.22.1 +-/root/.local/bin/turbo /data/hfd_work/envs/kimodo/bin/python -m pip install -r requirements/kimodo.txt +-/root/.local/bin/turbo /data/hfd_work/envs/kimodo/bin/python -m pip install --no-deps \ +- --target /data/hfd_work/envs/kimodo-text-compat -r requirements/kimodo-text.txt +- +-python3.10 -m venv /data/hfd_work/envs/omg +-/root/.local/bin/turbo /data/hfd_work/envs/omg/bin/python -m pip install \ +- --extra-index-url https://download.pytorch.org/whl/cu128 -r requirements/omg.txt +-/root/.local/bin/turbo /data/hfd_work/envs/omg/bin/python -m pip install --no-deps -e upstream/omg +- +-python3.10 -m venv /data/hfd_work/envs/textop +-/root/.local/bin/turbo /data/hfd_work/envs/textop/bin/python -m pip install \ +- --extra-index-url https://download.pytorch.org/whl/cu128 -r requirements/textop.txt +-/root/.local/bin/turbo /data/hfd_work/envs/textop/bin/python -m pip install --no-deps \ +- -e upstream/textop/deps/CLIP -e upstream/textop/deps/isaac_utils -e upstream/textop/TextOpRobotMDAR +-``` +- +-固定外部代码revision,不换成最新分支: +- +-```bash +-/root/.local/bin/turbo git clone https://github.com/NVlabs/GENMO.git /data/hfd_work/hfd-human-models/genmo +-git -C /data/hfd_work/hfd-human-models/genmo checkout 16bebf402d8893184249ee206d957b8248cd8310 +-/root/.local/bin/turbo git clone https://github.com/nv-tlabs/kimodo.git /data/hfd_work/hfd-human-models/kimodo +-git -C /data/hfd_work/hfd-human-models/kimodo checkout e1c219c99b4b94ba48be1fb6a1cbc64ecee4c954 +-/root/.local/bin/turbo /data/hfd_work/envs/genmo/bin/python -m pip install --no-deps -e /data/hfd_work/hfd-human-models/genmo +-env SKIP_MOTION_CORRECTION_IN_SETUP=1 /root/.local/bin/turbo /data/hfd_work/envs/kimodo/bin/python -m pip install \ +- --no-deps --no-build-isolation -e /data/hfd_work/hfd-human-models/kimodo +-``` +- +-## 外部资源 +- +-完整身份清单在`TRAINING_ASSETS.json`。HFD索引与人体分片固定为`93ba4a7a9c66b3e1b7041e5f4ea038ef2418b948`,不能再用不含SMPL-X的旧版本。先准备已授权的Llama base及SMPL-X主体;其余缺失资产可按清单下载,不覆盖已有文件: +- +-```bash +-/root/.local/bin/turbo /data/hfd_work/envs/omg/bin/python tools/training_assets.py \ +- --models-root /data/hfd_assets/models --raw-root /data/hfd_assets/hfd_data \ +- --models omg textop genmo kimodo --omg-sizes 50m 100m 300m 500m --download +-/root/.local/bin/turbo /data/hfd_work/envs/omg/bin/python tools/download_hfd_tars.py \ +- --root /data/hfd_assets/hfd_data +-``` +- +-已有资产去掉`--download`进行SHA256核验;`--quick`只检查大小,`--plan`只列出所需文件及大小。HF源可使用授权的`hfmirror`重试;CLIP的Azure下载不能借此自动转用个人代理。只训练某一模型可缩小`--models`/`--omg-sizes`。不要在缺少原始tar时直接启动训练,训练缺分片会报错,不再隐式联网补齐。若复用既有完整资产,直接核验即可,不必重新下载。 +- +-准备下面的相对布局;权重和数据不在Git中: +- +-```text +-/data/hfd_assets/ +- hfd_data/ +- index.sqlite, MANIFEST.json, PATH_MAP.json +- metadata/seed_g1_smplx_pairs.parquet +- 原始tar分片(保留,不批量解压) +- models/ +- body_models/smplx/SMPLX_NEUTRAL.npz +- genmo/gem_smpl.ckpt +- genmo/t5-3b-local/ # 完整tokenizer/config/权重 +- kimodo/Kimodo-SMPLX-RP-v1/ # model.safetensors、config.yaml及完整stats/ +- kimodo/Meta-Llama-3-8B-Instruct/ # 完整base及分片 +- kimodo/LLM2Vec-Meta-Llama-3-8B-Instruct-mntp/ +- kimodo/LLM2Vec-Meta-Llama-3-8B-Instruct-mntp-supervised/ +- OMG/official/checkpoints/paper/50m/sstep=090000.ckpt +- OMG/t5-base-local/ +- TextOp-Data/TextOpRobotMDAR/logs/pretrained/checkpoint/{ckpt_200000.pth,vae.pth} +- TextOp-Data/TextOpRobotMDAR/dataset/PRIVATE-DATA/meanstd.pkl +- TextOp-Data/TextOpRobotMDAR/description/robots/g1/ +- TextOp-Data/clip/ViT-B-32.pt +- asr/Qwen3-ForcedAligner-0.6B/ # 重建对齐时需要 +-``` +- +-GenMo权重来自`nvidia/GEM-X` revision `5b9a69d5ed4713ac4ba2a63b335e66db87279a3e`的`gem_smpl.ckpt`。Kimodo权重来自`nvidia/Kimodo-SMPLX-RP-v1` revision `1419ba56b734c48bbafb41fefa84088ca94583b5`;两个LLM2Vec适配器revision和SHA256在`UPSTREAM_LOCK.json`。Llama本机来源为ModelScope的`LLM-Research/Meta-Llama-3-8B-Instruct`,也可按已有授权获取对应HF base,不能用量化版或不同base替代。SMPL-X主体需自行从已授权来源获取;本机neutral-2020文件SHA256为`bdf06146e27d92022fe5dadad3b9203373f6879eca8e4d8235359ee3ec6a5a74`。 +- +-GenMo需要四个额外身体回归/顶点表,URL、大小、revision和hash已固定: +- +-```bash +-/root/.local/bin/turbo /data/hfd_work/envs/genmo/bin/python tools/fetch_genmo_helpers.py \ +- --genmo-root /data/hfd_work/hfd-human-models/genmo +-/data/hfd_work/envs/genmo/bin/python tools/prepare_human_runtime.py \ +- --models-root /data/hfd_assets/models --genmo-root /data/hfd_work/hfd-human-models/genmo +-``` +- +-Kimodo安装时跳过setup中的MotionCorrection自动构建,避免其另行联网拉取构建依赖;随后由准备器使用已安装的Eigen和pybind11构建,并关闭CMake联网。准备器也建立本地base/adapter链接,不复制权重。构建需要C++17编译器和CMake>=3.15: +- +-```bash +-/data/hfd_work/envs/kimodo/bin/python tools/prepare_human_runtime.py \ +- --models-root /data/hfd_assets/models --kimodo-root /data/hfd_work/hfd-human-models/kimodo \ +- --build-correction --build-dir /data/hfd_work/cache/kimodo-motion-correction-build +-``` +- +-## 目标机器配置 +- +-```bash +-/data/hfd_work/envs/omg/bin/python tools/configure_training.py \ +- --work-root /data/hfd_work --assets-root /data/hfd_assets \ +- --raw-root /data/hfd_assets/hfd_data \ +- --external-code-root /data/hfd_work/hfd-human-models \ +- --env-root /data/hfd_work/envs --gpu 0 +-source /data/hfd_work/training-configs/training.env +-``` +- +-生成器按所选配方保留batch、预算、损失、调度、随机种子及数据规则,改成本机路径和GPU。默认四模型均使用GPU0,分别运行。已有OMG/TextOp资源可用`--robot-models-root /existing/models`复用;`--output`可改配置目录。它还生成`omg-{50m,100m,300m,500m}.yaml`和`genmo-{v2,legacy}.yaml`;`--omg-size`和`--genmo-recipe`选择默认项。默认OMG仍为50M,GenMo改为v2;更大OMG的微批次/累积与各原生损失见[复现矩阵](TRAINING_REPRODUCTION_MATRIX.md)。 +- +-## 英文词时间缓存 +- +-优先把现有`cache/hfd/reference-alignments.sqlite`迁移到新工作盘。它是源语音词对齐及第三方recipe索引,session仍动态组装。复制应在写入进程结束后保留完整SQLite状态。 +- +-没有缓存时创建独立Python3.10环境`forced-align`,按`runtime-versions.json`安装Torch,再安装`requirements/forced-align.txt`。准备`Qwen/Qwen3-ForcedAligner-0.6B` revision `c7cbfc2048c462b0d63a45797104fc9db3ad62b7`后运行: +- +-```bash +-/data/hfd_work/envs/forced-align/bin/python tools/align_reference_corpus.py \ +- --root /data/hfd_work/cache/reference-preparation \ +- --cache /data/hfd_work/cache/hfd/reference-alignments.sqlite \ +- --model /data/hfd_assets/models/asr/Qwen3-ForcedAligner-0.6B \ +- --language en --gpus 0 --batch-size 16 +-/data/hfd_work/envs/forced-align/bin/python tools/index_reference_recipes.py \ +- --cache /data/hfd_work/cache/hfd/reference-alignments.sqlite --language en \ +- --report /data/hfd_work/cache/reference-recipes.json +-``` +- +-双GPU可用`--gpus 0 1`。完成后应有61,860条英文源资产、缺失0;OMG/TextOp另用选定第三方英文recipe,GenMo/Kimodo仅用具有人体配对的原生英文session。重新对齐可能有数值差异;恢复原完整断点时应保留原缓存及代码/配置身份。原始tar、索引、PATH_MAP和配对metadata需完整;`tools/download_hfd_tars.py`可按已有manifest补齐分片。 +- +-## 预检、训练、恢复 +- +-从仓库根目录执行: +- +-```bash +-/data/hfd_work/envs/omg/bin/python train.py --config /data/hfd_work/training-configs/omg.yaml --dry-run +-/data/hfd_work/envs/textop/bin/python train.py --config /data/hfd_work/training-configs/textop.yaml --dry-run +-/data/hfd_work/envs/genmo/bin/python train_genmo_full.py --config /data/hfd_work/training-configs/genmo.yaml --check-only +-``` +- +-上述预检不更新参数。Kimodo的`--offline-check`会执行一次丢弃的梯度更新,加载全量索引及冻结文本worker。正式命令分别为: +- +-```bash +-/data/hfd_work/envs/omg/bin/python -u train.py --config /data/hfd_work/training-configs/omg.yaml +-/data/hfd_work/envs/textop/bin/python -u train.py --config /data/hfd_work/training-configs/textop.yaml +-/data/hfd_work/envs/genmo/bin/python -u train_genmo_full.py --config /data/hfd_work/training-configs/genmo.yaml +-/data/hfd_work/envs/kimodo/bin/python -u train_kimodo.py --config /data/hfd_work/training-configs/kimodo.yaml +-``` +- +-可在screen中前台执行。恢复时加`--resume auto`,不改变总预算;版本、路径或数据身份变化会拒绝伪装成同一次训练。`TRAINING_SNAPSHOT.json`与Git历史保留迁移前实现。跨机器从头复现使用生成配置;逐位复现还取决于缓存、完整随机状态、CUDA确定性及硬件,不能只凭seed保证。 +- +-GenMo旧权重的倾听姿态问题见[诊断说明](GENMO_LISTENING_DIAGNOSIS.md)。当前默认采用[GenMo v2训练策略](GENMO_V2_TRAINING.md),旧配方和原代码快照仍保留,未覆盖旧正式权重。新策略完整恢复使用`--resume auto`;旧权重迁入新策略使用`--init-checkpoint`,这是新run的权重初始化。 +diff --git a/docs/REFERENCE_DYNAMIC_TRAINING.md b/docs/REFERENCE_DYNAMIC_TRAINING.md +deleted file mode 100644 +index 99edc89..0000000 +--- a/docs/REFERENCE_DYNAMIC_TRAINING.md ++++ /dev/null +@@ -1,84 +0,0 @@ +-# 离线对齐与动态 session 训练 +- +-清理说明:本文记录的一次性实验脚本已归档,复现前请按[脚本恢复说明](WORKSPACE_CLEANUP.md)取回;历史指标与日志保留。 +- +-2026-10-19。当前实现保留原生 HFD 动态组装流程:缓存原始语音资产的词时间戳,训练时按实际组装位置平移时间戳。无需把全部 session 导出为 NPZ,也无需在线运行 ASR。 +- +-## 已落实的数据链路 +- +-1. `tools/align_reference_corpus.py` 从本地 tar 索引枚举原始任务、停止、附和、犹豫与修正语音,按唯一资产对齐。最初完成 123,720 个双语资产;按英文训练决策,当前库仅保留 61,860 个英文资产,英文缺失 0。没有批量解压或复制原始音频。 +-2. `/root/autodl-tmp/cache/hfd/reference-alignments.sqlite` 保存资产路径、原始音频 SHA256、16 kHz 长度、原文与词时间戳。训练时只读查表;不是把全部 session 常驻内存。进程内仍有受容量限制的 session/特征缓存。 +-3. 原生组装器在实际音频拼接时记录源资产和起止采样点,包括犹豫、停顿、修正、停止及附和。该元数据不改变动作、PCM、控制监督或随机组装流程。 +-4. 第三方已组装 session 按本地 `session_recipes.jsonl` 的前景目标语音位置映射,检查原始音频哈希与长度。背景音频保留,背景说话者的文本不作为条件。索引含 27,584 个 recipe、82,590 个目标片段;实际 train/val 仍以原选择清单的 26,737/783 个 session 为准。 +-5. 词结束后加实验性的 0.1 秒 guard,再向上取整到 `0.6, 1.0, 1.4, ...` 秒释放。模型只看到当前时刻已释放、过去 30 秒内的文字。原生动作时钟与麦克风时钟的 0.2 秒偏移按 `clock_policy: microphone` 统一处理。 +- +-24 个双语冻结样例的动态 PCM、动作与控制监督逐元素相同;动态生成的完整文本 records 也与先前导出的 `.reference.json` 完全相同。该检查覆盖 train/val、中英及六种交互类型,属于代表性回归检查,不表示已逐一实例化所有原生 session。 +- +-这属于 **oracle 前景转写条件**,不能作为真实 ASR 识别准确率实验。强制对齐也是时间估计,0.1 秒不是已证明的误差上界。历史双语库 8,342/123,720(6.74%)个资产存在无效跨度;当前保留的英文资产中为 6,608/61,860(10.68%)。这些资产保守地整段结束后释放文字;不猜测缺失词边界,不丢弃文本。 +- +-## 模型实际输入 +- +-| 模型 | 条件 | 监督与输出 | +-| --- | --- | --- | +-| OMG | 当前可用的过去 2 秒音频、已释放的最近 30 秒文字、原生动作历史 | 原生 G1 动作表示与原生训练损失 | +-| TextOp | 已释放的最近 30 秒文字、原生动作历史 | 原生 G1 动作表示、四 primitive 的原生三阶段 DAR 训练 | +- +-不输入控制标签、未来音频或未来原文。当前只采样英文前景 session,`text_conditioning: none`,不加载 ASR 或翻译模型。背景音频仍原样保留。中文翻译实验属于历史对照。 +- +-“保留 30 秒文字”是前端上下文约定,不代表编码器具有无限 token 容量。TextOp 的 CLIP 保留预算内的最近文本后缀;OMG 原生 T5 为 50 token,新 reference 配置通过适配工厂设置 `truncation_side=left`,超长时保留最近文字。架构、长度、参数与严格 checkpoint 加载不变;旧 run 不追溯修改。 +- +-## 配置和复查 +- +-新配置为 `configs/omg_reference_train.yaml`、`configs/textop_reference_train.yaml`,使用 `native_ondemand` 和 `aligned_reference_context`。`omg.yaml`、`textop.yaml` 也已同步为英文 reference 默认;旧 ASR 对照移至 `configs/history/`。OMG 使用物理 GPU 1,TextOp 使用物理 GPU 0。 +- +-```bash +-/root/autodl-tmp/envs/omg/bin/python train.py --config configs/omg_reference_train.yaml --dry-run +-/root/autodl-tmp/envs/textop/bin/python train.py --config configs/textop_reference_train.yaml --dry-run +-``` +- +-`--dry-run` 检查完整对齐索引、第三方 recipe 表和权重路径;完整读取与训练仍需短测。构建索引约需数分钟 CPU,首次组池和共享盘读取不能算作纯 GPU 更新吞吐。当前英文路线不做翻译。 +- +-本轮证据目录:`/root/autodl-tmp/runs/archive/legacy-20261020/reference-full-preparation-20261019/`。其中 `coverage.json`、`alignment-quality.json`、`dynamic-verification.json`、`provider-verification.json`、`translation-*.json` 分别记录覆盖、回退、动态文本、实际 provider 和翻译诊断。配置中的 100K/300K 步仍是旧预算;batch 改变后不能直接当作已验证的推荐预算。 +- +-TextOp 旧单 session 过拟合的 teacher-forced 短窗误差下降,但连续闭环误差变差。当前同 seed 引导系数 1 的补充检查仍是初始 0.3004 → 训练后 0.3121 rad;单纯把系数从 5 改成 1 没有解决。该证据不支持宣称全双工或长程动作能力已经验证。 +- +-全量正式训练尚未启动。当前英文短测报告将保存在 [纯英文短测报告](ENGLISH_ONLY_PROBE.md);历史双语结果不能直接与英文验证损失比较,必须使用同一英文回放集合。 +- +-## 历史双语短测结果 +- +-两模型都从官方权重初始化,batch=64,使用完整原生加第三方集合的随机窗口采样;不是逐一训练完所有 session。 +- +-| 检查 | OMG | TextOp | +-| --- | --- | --- | +-| 更新预算 | 512 步 | 768 步,三阶段各 256 步 | +-| 实际训练 session | 512 个槽位、510 个不同 ID | 192 个槽位、192 个不同 ID;每组四 primitive | +-| 原生验证损失 | 0.07212 → 0.04290,固定窗口和噪声 | 0.14213 → 0.14530,原生随机验证 | +-| 权重检查 | 512 次优化更新、梯度和损失有限、参数变化、严格重载通过 | 记录的梯度范数和损失有限,108 个状态张量变化,严格重载通过;未另挂 hook 独立计数 optimizer.step | +- +-OMG 池内单步中位耗时 0.115 秒,而包含冷池、验证和保存的训练阶段共约 1,064 秒;未包含此前索引与模型启动,不能用池内速度直接外推。OMG 每个 64-session 池准备中位约 103 秒;TextOp 三个池约 105/97/125 秒。当前耗时主要仍在共享盘读取、session 组装、首次翻译及特征准备,尚未证明正式 batch512 的端到端效率。 +- +-OMG 最后一轮因框架 lookahead 多准备了一个未训练的池,约 94 秒。这项开销已通过训练/验证 batch 上限修正;真实 CPU Lightning/DataLoader 的四步边界检查确认只请求四个源 batch,未触发下一冷池。57 项相关测试通过。本轮真实模型曲线和耗时来自修复前,未人为扣除这段耗时,也未重跑整个短测。 +- +-孤立批量翻译探针:64 条真实中文释放文本,batch16 与逐条结果 64/64 完全相同,纯模型翻译约 0.49 秒对 2.41 秒。这是后续优化依据,尚未改训练默认翻译策略,也没有解决译错动作的问题。 +- +-证据:运行目录下 `summary.json`、`pilot-curves.png`、`gpu-utilization.png`、`textop-checkpoint-verification.json`、`lightning-batch-bound-verification.json`;原生 run 的配置、采样记录、日志、验证曲线及 checkpoint 均保留。短测固定验证损失下降不等于连续闭环动作质量通过;旧 100K/300K 预算仍待校准。 +- +-## 正确中文直接输入的检查 +- +-2026-10-19 补充:文本转写正确与编码器支持中文是两项独立条件。原生冻结编码器实测如下: +- +-- OMG 的 T5-Base 将“向前走”“向后走”“向左转”“向右转”“停止动作”“继续动作”等 16 条中文对照指令全部编码为 `[3, 2, 1]`,其中 2 为 unknown token。相同 token 与 mask 的文本条件相同,后续仅训练动作网络无法恢复丢失的指令区别;音频及动作历史仍可能驱动输出。直接解冻这个编码器也不能区分已经被 tokenizer 合并的输入。 +-- TextOp 的 CLIP 字节 BPE 能保留不同中文字符,不能因此断言直接中文训练不可能。但当前冻结 CLIP 在这组手写 16 类中英文本匹配上仅 1/16 top1 正确。这是编码空间的小型诊断,不是中文动作准确率。 +- +-用上述短测 checkpoint,在同样两条中文验证 session(打断、急停)上保持种子、PCM、初始动作历史及文本释放时间相同,各回放中文原文、OPUS 英文和空文本,共 12 次。两条 session 的等权平均关节 RMSE: +- +-| 模型 | 直接中文 | 翻译英文 | 空文本 | +-| --- | ---: | ---: | ---: | +-| OMG | 0.4406 | 0.5121 | 0.4069 | +-| TextOp | 0.4355 | 0.3804 | 0.4307 | +- +-单位 rad,单种子 12345。OMG 的中文误差确实低于译文,但空文本更低;结合 token 坍缩,不能把这一结果解释成读懂中文。TextOp 译文在两个样例中均比中文误差低,但闭环关节误差仍不等于执行指令语义的准确率。 +- +-这只是已按英文条件训练的 checkpoint 的推理消融,尚未做直接中文与翻译英文的等预算训练比较。OMG 如需原生读懂中文,必须调整 tokenizer/encoder;TextOp 可单独做直接中文训练实验。更换中文/多语编码器还需要对齐原动作条件空间,并应标为独立模型变体。正式配置本轮未改变。 +- +-复现工具 `tools/probe_direct_chinese.py`;证据目录 `/root/autodl-tmp/runs/archive/legacy-20261020/direct-chinese-probe-20261019/`,其中 `encoders.json`、`omg-rollout.json`、`textop-rollout.json` 和 `summary.json` 保留完整结果。 +diff --git a/docs/ROLLOUT_DEBUG.md b/docs/ROLLOUT_DEBUG.md +deleted file mode 100644 +index fc94a97..0000000 +--- a/docs/ROLLOUT_DEBUG.md ++++ /dev/null +@@ -1,82 +0,0 @@ +-# 视频问题调试:输入、监督范围与执行节奏 +- +-清理说明:本文记录的一次性实验脚本已归档,复现前请按[脚本恢复说明](WORKSPACE_CLEANUP.md)取回;历史指标与日志保留。 +- +-本轮使用同一个原生双任务 session:接取别人递来的物品尚未完成时,输入捡起地上物品的新指令。数据未手工拼接,混合录音、背景音、ASR 与关节映射未改变。所有输出位于 `runs/hfd-final/`;主视频目录只展示两条当前结果,所有对照集中在 `debug/`。 +- +-## 结论与已采用修改 +- +-1. **监听姿态缺失首先是监督范围问题。** 之前把“不输入 listen 等控制标签”扩大成了“不训练这些帧的关节动作”。两个模型没有控制分类头,但都能拟合这些关节姿态。经用户确认改成 `full_session`:完整动作进入原生损失,控制标签仍不作为模型输入。原来 TextOP 仅 35 条四段连续训练链,OMG 仅 84 个两秒窗口;补回后分别为 88 条与 366 个窗口。监听、起步和收尾不再被过滤。 +-2. **OMG 重复动作与缺少收尾监督有关;卡顿还与异步计划交接有关。** 原先深弯腰在约 2.07、7.93、10.30、12.77 秒出现四次,当前主结果只有 9.57 秒一次,与真值相符。异步策略计算时执行旧计划,完成后跳过新计划前缀;两个计划在交接时的姿态不保证一致。最大跳变原来正好落在交接帧,并非播放器掉帧。该策略参照原生 `upstream/omg/src/omg/cli/pipeline/modes.py` 的 bridge/skip 流程,直接参考姿态回放没有物理跟踪器去缓和突变。 +-3. **TextOP 连续生成仍不合格,不能靠拉长文本窗口修好。** 它执行全部原生 8 帧(50 Hz 下 0.16 秒),且 2 帧动作历史是预测反馈,没有每个窗口重新初始化。固定真值历史时误差很小,但用自己的历史后逐渐停滞或延迟。完整监督后能出现监听姿态,却长时间停留在那里;捡取弯腰峰值约 11.66 秒,真值为 9.58 秒。延长训练反馈链也没有可靠解决。 +- +-当前 OMG 采用完整动作监督、过去两秒音频+持续因果 ASR、原生 DDIM 10 步、`joint_one` CFG、60 帧计划、剩余 24 帧时异步预取,使用实际测得规划耗时。修改的是原生采样参数,没有换网络、损失或添加手工平滑。TextOP 保留完整监督和原生 0.16 秒执行节奏,当前结果明确标记尚未收敛。 +- +-## 同 checkpoint 的对照 +- +-所有主对照固定 seed=12345。RMSE 为整段连续关节误差,排除开场真值初始化。TextOP 统计其 23 个关节,OMG 统计 29 个,不能把二者绝对值当作直接模型排名。 +- +-| 旧 motion-only checkpoint 的推理变化 | 连续 RMSE (rad) | 说明 | +-| --- | ---: | --- | +-| OMG 原推理 | 0.3294 | 50 步、测量延迟、异步 | +-| OMG 假设零延迟 | 0.3510 | 仅诊断;最大跳变从 0.4288 降到 0.1854 | +-| OMG 每 0.4 秒截断重规划 | 0.3708 | 不模拟计算耗时,仍差 | +-| OMG 执行完整 2 秒计划 | 0.3226 | 不模拟计算耗时,末尾仍重复 | +-| OMG 音频改为 0.6 秒 | 0.3204 | 零延迟,与训练 2 秒输入不匹配,仅诊断 | +-| OMG 音频改为完整前缀 | 0.3389 | 零延迟,前缀压成 60 个声学位置;仍无改善 | +-| OMG 去掉 ASR | 0.3867 | 零延迟,保持音频;文本条件有作用 | +-| OMG 去掉音频 | 0.3615 | 零延迟,保持 ASR;不能替代联合输入 | +-| TextOP 原推理 | 0.2434 | 原生 8 帧输出 | +-| TextOP 文本每 1.92 秒刷新 | 0.2441 | 仍执行原生 8 帧,改善几乎为零 | +- +-音频窗口和条件屏蔽是同 checkpoint 的敏感性检查;由于改变了训练时的输入分布,不能据此决定新训练的最佳窗口,更不能将诊断零延迟作为实时结果。 +- +-## 从官方权重重训,仅改为完整监督 +- +-TextOP 3200 次更新,阶段 `[800,1600,800]`,每次原生训练链 4 段;OMG 1600 次更新。其余参数与旧对应 run 保持一致。两者都通过有限 loss/梯度、参数更新、原生 checkpoint 严格加载检查。 +- +-| 完整监督后的设置 | 连续 RMSE (rad) | 最大逐帧跳变 RMS (rad) | +-| --- | ---: | ---: | +-| OMG 50 步+实际延迟 | 0.1865 | 0.3166 | +-| OMG 50 步+假设零延迟 | 0.0886 | 0.1242 | +-| OMG 10 步+实际延迟(当前) | 0.0619 | 0.1160 | +-| OMG 20 步+实际延迟 | 0.0622 | 0.1431 | +-| OMG 50 步+固定 5 帧延迟 | 0.0795 | 0.0976 | +-| OMG 10 步+固定 15 帧延迟 | 0.0753 | 0.2395 | +-| OMG 每次完整执行 2 秒,不模拟耗时 | 0.2293 | 0.0839 | +-| OMG 每 0.4 秒截断,不模拟耗时 | 0.2995 | 0.2400 | +-| TextOP 原生连续反馈(当前) | 0.2992 | 0.0197 | +-| TextOP 每块提供真值历史(oracle) | 0.0125 | 不作为连续性能 | +-| TextOP guidance=1 | 0.3247 | — | +-| TextOP 原生默认站姿初始化 | 0.4385 | — | +-| TextOP 去掉 ASR | 0.3260 | — | +-| TextOP 连续训练链从 4 段加到 16 段 | 0.2735 | 0.3357 | +- +-固定延迟交叉对照表明,采样轨迹与交接延迟都会影响结果,不能把全部误差改善只归因于速度。10 步采样平均规划约 0.157 秒,20 步约 0.230 秒,先前 50 步约 0.5 秒。没有手工插值、滤波或利用真值重置来美化当前视频。 +- +-OMG 10 步在 seed=12345/12346/12347 下 RMSE 分别是 **0.0619 / 0.1538 / 0.0585 rad**。主视频沿用原先 seed=12345,没有替换成最优种子。仍有随机性和残余接缝;真值最大逐帧跳变是 0.0355 rad,当前 OMG 的 0.1160 仍大于真值。 +- +-## 为什么 TextOP 的窗口很准,完整视频却不对 +- +-同一 ASR、同一个种子下,使用任务中的真值历史,8 帧误差约 0.0037–0.0054 rad;使用开场 HFD 站姿,预测的短期动作幅度约 0.003 rad。该旧 checkpoint 从原生默认站姿可以产生更大动作,但完整回放并未因此变好,说明不是简单换初始化就能解决。 +- +-完整监督的 TextOP 在每块真值历史诊断中误差 0.0125 rad,预测历史闭环却为 0.2992 rad,支持“训练短片段可拟合,闭环历史分布与任务阶段跟踪不稳定”的判断。缺少 KV cache 不是这里已经证实的唯一原因。TextOP 只有 ASR 条件:同一稳定句子可跨监听、执行、完成多个阶段;ASR 本身不显式提供动作是否已完成。保留完整标签用于动作监督不会自动增加阶段记忆。 +- +-16 段连续训练(2.56 秒)比 4 段(0.64 秒)误差略低,却引入大跳变,因此未采用。当前 TextOP 仍需更系统的原生训练与闭环验证;此次不将扩充模型状态或增添控制头偷偷并入 baseline。 +- +-## 复现与查看 +- +-当前配置:`configs/omg_full_session.yaml`、`configs/textop_full_session.yaml`。重训时改 `paths.output_dir` 为新目录。正式模板也已改为 `full_session`。当前 OMG 配置保存 `inference.omg_sampling_steps: 10`;TextOP smoke 保留此前诊断的 `static_prob=0.05`,正式模板仍是上游设置,应在论文里独立记录这一超参数。 +- +-```bash +-# 在对应模型环境运行;输出目录需尚不存在 +-python infer_overfit.py --run /root/autodl-tmp/runs/archive/legacy-20261020/hfd-final/omg --output-dir /data/runs/omg-replay --continuous-only --skip-window-eval --task-seconds 0 14.8 +-python infer_overfit.py --run /root/autodl-tmp/runs/archive/legacy-20261020/hfd-final/textop --output-dir /data/runs/textop-replay --continuous-only --task-seconds 0 14.8 +-# 仅作反事实诊断:--omg-latency-frames 0、--textop-history-mode teacher_forced +-# 其他对照:--omg-sampling-steps 50、--omg-execute-frames 60 --omg-replan-mode receding +-python tools/summarize_rollout_debug.py --root /root/autodl-tmp/runs/archive/legacy-20261020/hfd-final +-``` +- +-主视频:`videos/omg_full_session_playable.mp4`、`videos/textop_full_session_playable.mp4`。曲线:`debug/trajectory_comparison.png`。全部数值:`debug/summary.json`;旧结果在 `debug/previous_motion_only/`。这些均为有原始音频的关节参考姿态回放,没有物理跟踪器。 +- +-本轮沿用 HFD 名义动作时钟和音频释放映射,ASR 为预计算的因果 sidecar,未加入实际 ASR 计算延迟。OMG 仅计入本次声学特征、文本编码和模型规划耗时,不等同于完整麦克风到机器人端到端实时系统。所有测试来自同一个训练 session,不能作为独立测试集或论文泛化结论。 +diff --git a/docs/SEMANTICS_CLOCK_FIX.md b/docs/SEMANTICS_CLOCK_FIX.md +deleted file mode 100644 +index 82611b7..0000000 +--- a/docs/SEMANTICS_CLOCK_FIX.md ++++ /dev/null +@@ -1,59 +0,0 @@ +-# 中文语义与麦克风时钟修复 +- +-清理说明:本文记录的一次性实验脚本已归档,复现前请按[脚本恢复说明](WORKSPACE_CLEANUP.md)取回;历史指标与日志保留。 +- +-2026-10-18。修复已接入新 OMG/TextOp 训练配置、共享按需/third-party 数据入口、GenMo 文本缓存与窗口、连续推理和视频评测。未重训模型;历史 run、权重、ASR 和旧视频保持不变。HY-Motion/Kimodo 继续以同一输入契约为后续适配目标,其 HFD 训练器尚未实现。 +- +-## 语义条件 +- +-保留 Whisper-tiny `task=transcribe` 的原始 ASR,先按 `causal_asr_context` 获取已释放的完整假设,再通过冻结的 [Helsinki-NLP/opus-mt-zh-en](https://huggingface.co/Helsinki-NLP/opus-mt-zh-en) 转成英文。纯英文和空文本原样通过,仍由模型各自的原生文本编码器编码。没有使用 GT 指令或动作标签修正文本。 +- +-`慢跑向前` → `Jogging forward.`;`向后跳` → `Jump back.`。原先这些中文在 OMG T5-Base 中容易合并为 `[3,2,1]`,其中 `2` 是 UNK。六条指令的回归测试验证了 T5 token 可区分、无 UNK,保留左右手、否定和次序。实际三个样本共 42 次规划,翻译后文本均无 UNK,最长 19 token,未触发原生 50 token 截断。 +- +-实现:`hfd_bridge/semantics.py`。模型 eval、无梯度、beam=4、无采样,逐个 ASR 假设独立转换;长输入按 128 源 token 分块,输出遇到长度限制显式失败,不静默丢弃历史。初始化保护训练/采样 RNG。SQLite 缓存按模型身份、解码参数、规范化器版本和源字符串区分;缓存有容量限制,不跨释放点积累翻译记忆。 +- +-模型唯一存放在 `/root/autodl-fs/pangxinlei/models/translation/opus-mt-zh-en`,权重 312,087,009 字节。HF metadata 经 AutoDL 获取,权重在 AutoDL 超时、HF 镜像限流后从 ModelScope 国内直连下载(`trust_env=False`),未回退个人代理;本地 `download_identity.json` 记录版本和 SHA256。OMG/TextOp 依赖锁新增所需 tokenizer/规范化依赖;GenMo 环境已安装 `sacremoses==0.2.0`,其余依赖沿用现有环境。可用 `HFD_TRANSLATOR_ROOT` 指定相同身份的本地模型。 +- +-试验过 Whisper 直接 `task=translate`,但本例翻译误差明显,已放弃。输出目录下 `asr-en/` 是被拒绝的候选,三个修复视频均使用原始 `omg-100k-validation/asr` 的 transcribe sidecar,再做上述独立文本翻译。 +- +-## 信息可用时刻与视频 +- +-HFD 原始动作时刻 `t_motion` 对应麦克风时刻 `t_mic=t_motion+0.2s`。原始音频窗口 `[0.4k,0.4k+0.6]` 在 `r_k=0.6+0.4k` 到达。新 `microphone` 策略只读取满足 `r_k<=t_mic` 的最近完整窗口和 ASR。首个窗口还未到达时输入为空文本及静音。音频先裁剪再做归一化、重采样和卷积特征,未来 PCM 不参与这些运算。 +- +-例如原生动作 0.4s 请求对应麦克风 0.6s,只能用到 PCM 0.6s;历史 `legacy_hfd` 会选择到 1.0s。共享实现为 `hfd_bridge/clocks.py`,30Hz/50Hz 数据与推理共用;原始 PCM 和 sidecar 不需要重做。 +- +-完整视频从麦克风 0s 播放,预测与 GT 同时前置 0.2s 首帧(OMG/GenMo 6 帧、TextOp 10 帧)。这是显示时间变换,原生参考轨迹和计分范围保持不变;不是单独平移预测去匹配 GT。GenMo 仍只对完整预测和完整人体 GT 各 retarget 一次,原生 `retargeted.npz` 保持原始时间,显示前缀只写入视频比较文件。 +- +-新配置显式记录: +- +-```yaml +-caption_policy: causal_asr_context +-clock_policy: microphone +-text_conditioning: asr_to_english +-``` +- +-OMG/TextOp 位于 `data` 下,GenMo 位于顶层。无新字段的旧 run 保留 `legacy_hfd`、不翻译;新 GenMo 缓存记录翻译策略并包含启动所需空文本 embedding,不允许将旧缓存误当成新输入。 +- +-**延迟口径:** ASR/翻译为逐释放点因果离线缓存,计算耗时没有加入回放时间;OMG 规划仍使用实测生成耗时。信息因果性已验证,但不代表测得了包含前端的端到端实时延迟。 +- +-## 验证与同权重回放 +- +-输出:`/root/autodl-tmp/runs/archive/legacy-20261020/omg-semantics-clock-fix`。使用原始 100k OMG 权重、seed12345、DDIM10、native async、过去 2s PCM;对照旧 `causal_asr_context` 视频,同时修复语义与时钟,不能用这组三样本区分两项修复各自的效果。 +- +-| 样本 | 旧关节 RMSE(rad) | 新关节 RMSE(rad) | 旧根位置 RMSE(m) | 新根位置 RMSE(m) | +-| --- | ---: | ---: | ---: | ---: | +-| interruption | 0.3244 | 0.3539 | 1.6918 | 1.5484 | +-| backchannel | 0.2625 | 0.2776 | 0.0100 | 0.0112 | +-| emergency-stop | 0.4614 | 0.3864 | 0.4912 | 0.1547 | +- +-检查结果: +- +-- 70 项共享回归测试通过,另有 11 项 GenMo 几何/音频测试通过。 +-- OMG/TextOp 的真实数据契约检查通过:未来动作、未来 ASR/标签、未来 PCM 不影响当前条件,标量/批量音频特征一致。 +-- 三个视频共 1,638 帧完整解码且有音轨;NPZ 音频严格等于原始 PCM 从 0 开始的相应片段。 +-- 新视频去掉共同显示前缀后,GT 与旧评测逐元素完全相同;三个样本均链接同一原始 checkpoint。 +-- 42 次规划的 PCM 截止和 ASR 释放时刻均不晚于请求的麦克风时刻;保存每次原始 ASR、英文条件及实际 T5 token。 +- +-可复查 `verification.json`、`tests.log`、`genmo-tests.log`、`omg-contract.log`、`textop-contract.log`。逐次验证脚本为 `tools/verify_semantics_clock.py`。 +- +-**当前限制:** 输入已能区分中文,但 Whisper 错字、静音幻觉和翻译误差仍在;完整 ASR 也不保证模型理解纠错。旧权重曾在中文 UNK/旧时钟下训练,此次没有重训,动作仍有明显偏差,指标并非全面改善。不能据此宣称提前倾听、不执行指令等模型行为已经修好。 +diff --git a/docs/TEXTOP_ENGLISH_VALIDATION.md b/docs/TEXTOP_ENGLISH_VALIDATION.md +deleted file mode 100644 +index ecfd00c..0000000 +--- a/docs/TEXTOP_ENGLISH_VALIDATION.md ++++ /dev/null +@@ -1,83 +0,0 @@ +-# TextOp 300K 英文验证评测(2026-10-21) +- +-训练目录:`/root/autodl-tmp/runs/hfd-textop-english-train`,使用完整训练后的 +-`ckpt_300000.pth`,与官方初始化checkpoint做同条件对照。没有继续训练或改动已有权重。 +- +-结果根目录:[textop-english-validation-20261021](/root/autodl-tmp/runs/textop-english-validation-20261021/README.md)。 +-包含18条有原混合音轨的视频、原始qpos、文本/token审计、指标CSV、汇总JSON、比较图及验证记录。 +- +-## 样本与协议 +- +-- 六类原生英文验证各一条:普通指令、打断、急停、附和、犹豫、暂停;另取三条第三方验证。 +-- 使用训练时相同的val provider、配方、seed42、原生与第三方选择摘要和离线文本库; +- 导出动作、音频及逐帧文字与动态provider逐值相同,原生任务source hash划分核对通过。 +-- 样本在生成前选定:各类首条估计6–30秒的合格英文session,第三方按合格ID等距选取。 +- 不是按结果挑选。当前没有独立test,val曾用于训练监测。 +-- 两组生成seed均12345;相同一次初始GT历史/根锚点,之后只反馈预测。 +- 原生2帧57D历史、8帧primitive、50Hz、CFG5,使用训练配置中的5步原生扩散。 +- 覆盖所选session内的完整primitive,末尾不足8帧不额外补齐或重置GT。 +-- 文本为`aligned_reference_context`:已释放最近30秒英文GT前景文本,CLIP77最近后缀; +- microphone时钟、0.1秒guard,不运行ASR/翻译,不输入音频或控制标签。 +-- TextOp直接生成G1,不做人体系retarget;未建模腕关节固定,主关节指标只算原生23个关节。 +-- 视频左预测右GT,保留0.2秒共同microphone前置帧;评分排除前置及GT bootstrap, +- 不做DTW或事后根对齐。同步运动学回放,不把实测规划耗时模拟为控制器延迟。 +- +-## 结果 +- +-以下按session等权平均,关节单位rad,根xyz分量RMSE单位m: +- +-| 范围 | 初始关节RMSE | 300K关节RMSE | 初始根RMSE | 300K根RMSE | +-| --- | ---: | ---: | ---: | ---: | +-| 全部9条 | 0.4380 | 0.4152 | 1.2243 | 1.4877 | +-| 原生6条 | 0.4481 | 0.4569 | 0.7812 | 0.9121 | +-| 第三方3条 | 0.4178 | 0.3318 | 2.1105 | 2.6390 | +- +-整体关节RMSE降低约5.2%,根RMSE增加约21.5%;原生六类未整体改善。 +-附和、暂停样本改善,急停和打断仍有明显问题,不能以局部或平均关节误差下降宣称任务能力提升。 +- +-| 原生场景 | 初始关节RMSE | 300K关节RMSE | 初始根RMSE | 300K根RMSE | +-| --- | ---: | ---: | ---: | ---: | +-| 普通指令 | 0.4346 | 0.4541 | 0.2770 | 0.1805 | +-| 打断 | 0.4084 | 0.3946 | 1.3258 | 2.6228 | +-| 急停 | 0.5908 | 0.6995 | 1.4652 | 2.1198 | +-| 附和 | 0.3017 | 0.2429 | 1.5385 | 0.5089 | +-| 犹豫 | 0.4440 | 0.5042 | 0.0399 | 0.0328 | +-| 暂停 | 0.5091 | 0.4462 | 0.0410 | 0.0079 | +- +-优先查看: +- +-- [急停300K](/root/autodl-tmp/runs/textop-english-validation-20261021/videos/emergency-stop-trained.mp4) +-- [打断300K](/root/autodl-tmp/runs/textop-english-validation-20261021/videos/interruption-trained.mp4) +-- [附和300K](/root/autodl-tmp/runs/textop-english-validation-20261021/videos/backchannel-trained.mp4) +-- [全部对照与指标](/root/autodl-tmp/runs/textop-english-validation-20261021/README.md) +- +-急停样本`val-000111-seed42`中,`Stop.`在麦克风10.6秒已经释放且实际送入CLIP; +-12–16.88秒区间,预测平均根速度约1.008m/s、GT约0.014m/s,模型仍持续运动。 +-这是单条样例的停止后运动诊断,不是经过统一定义的停止延迟指标;可以排除该例是因 +-停止文字未释放、被token截断或ASR漏识别造成。不能仅凭此确定训练目标、历史依赖或文本理解 +-中哪一项是主要原因。详见结果目录的`emergency-stop-diagnostic.json`。 +- +-训练日志最终`eval_total=0.10740`;最后100K阶段最低约0.10626(250K)。 +-训练验证样本/阶段历史分布与连续闭环不同,不能把该loss解释为上述视频误差或用全程曲线 +-单独证明过拟合。完整抽取的150个验证点见结果目录`training-curve.json`。 +- +-## 复现与验证 +- +-```bash +-cd /root/autodl-tmp/hfd-baseline-portable +-/root/autodl-tmp/envs/textop/bin/python -u evaluate_textop.py \ +- --trained-run /root/autodl-tmp/runs/hfd-textop-english-train \ +- --output /root/autodl-tmp/runs/textop-english-validation-NEW \ +- --gpu 0 +-``` +- +-入口支持`--checkpoint`、`--stage prepare/run/summarize`。原生模型、冻结VAE、CLIP、统计量 +-均复用本地资产;不重新下载,不把验证样本写入训练数据。 +-推理入口已支持显式TextOp checkpoint,并可跳过无关训练窗口评测;每个primitive记录原文、 +-实际编码后缀、CLIP token ID、条件帧、释放时刻和规划耗时。 +-视频长文本按列宽换行,过长仅显示最近150字符;完整模型输入仍保留在report,不改变条件。 +- +-18条视频共16,504帧,完整解码、原音频一致、两权重GT一致、计划时钟/文本/token一致、 +-CLIP77与释放因果边界检查通过。配套25项数据/时钟/文本回归检查通过。 +-这些验证说明评测链路一致,不证明语义成功或统计显著改善。 +-共享stationary池、第三方与原生的源重叠及上游预训练重叠未排除,不能称完全未见动作泛化。 +diff --git a/docs/TRAINING_REPRODUCTION_MATRIX.md b/docs/TRAINING_REPRODUCTION_MATRIX.md +deleted file mode 100644 +index e420551..0000000 +--- a/docs/TRAINING_REPRODUCTION_MATRIX.md ++++ /dev/null +@@ -1,44 +0,0 @@ +-# 训练复现覆盖范围 +- +-当前发布四种 baseline 的训练入口,HY-Motion 不在已完成训练框架范围内。源码、正式配置、依赖版本、外部源码 revision 和必要资产身份均在仓库;数据与模型文件按单独授权准备,不上传权重/日志/测试脚本。 +- +-## 现有配方和 OMG 多个初始化权重 +- +-| 配方 | 入口/生成配置 | 训练预算 | 微批次 × 累积 | 说明 | +-|---|---|---:|---:|---| +-| OMG 50M paper | `train.py` / `omg-50m.yaml` | 100K optimizer steps | 512 × 1 | 已有 HFD 正式实验使用此初始化 | +-| OMG 100M paper | `train.py` / `omg-100m.yaml` | 100K | 64 × 8 | 独立输出目录;有效 batch 512 | +-| OMG 300M paper | `train.py` / `omg-300m.yaml` | 100K | 32 × 16 | 独立输出目录;有效 batch 512 | +-| OMG 500M paper | `train.py` / `omg-500m.yaml` | 100K | 16 × 32 | 独立输出目录;有效 batch 512 | +-| TextOp | `train.py` / `textop.yaml` | 300K,三阶段各 100K | 512 × 1 | DAR/VAE/CLIP/57D 统计及机器人资源配套 | +-| GenMo legacy | `train_genmo_full.py` / `genmo-legacy.yaml` | 100K | 32 × 1 | 保留原正式配方及旧代码快照 | +-| GenMo v2 | `train_genmo_full.py` / `genmo-v2.yaml` | 100K | 32 × 1 | 新恢复课程、执行段损失、闭环选权重 | +-| Kimodo | `train_kimodo.py` / `kimodo.yaml` | 100K | 8 × 1 | 当前因果动作表示、根锚点约束、预测历史恢复及冻结文本 worker | +- +-其他 OMG 尺寸是可复现的初始化与微调配置,不代表已经各自完成 HFD 全量训练。各尺寸的实际文件名和 SHA256 见 `configs/omg_variants.yaml`,下载来源固定为 `THU-MARS/OMG` revision `06a88419da9f9125d0b5480cc10966741a189358`。100M/500M 需显式打开与权重相符的音频/人体分支结构;人体条件的可用性mask仍为false,HFD适配不输入未来人体GT。500M 仅移除已知的 52 个废弃 `visual_film` 参数,其他缺失/多余参数仍拒绝加载。保留各原生 experiment 的损失配置,100M/500M 的 full loss 与 50M/300M 的 diffusion-only loss 不能混称完全相同的训练目标。 +- +-上述 OMG 微批次按单卡 32GB 保守设置,均保持有效 batch 512;实际峰值还受文本长度与数据影响,不能把单次检查当作最大 batch 调优。50M/100M/300M/500M 均已用对应微批次完成严格权重匹配和一次丢弃的参数更新。历史 HFD run 中的 `best`、98K、100K 等是同一次训练的不同 checkpoint,均由原生验证与保留策略产生;它们不是新的模型尺寸,也不在 Git 中。 +- +-## 生成与运行 +- +-按便携指南准备资产后: +- +-```bash +-/data/hfd_work/envs/omg/bin/python tools/configure_training.py \ +- --work-root /data/hfd_work --assets-root /data/hfd_assets \ +- --raw-root /data/hfd_assets/hfd_data --omg-size 300m --genmo-recipe v2 --gpu 0 +-source /data/hfd_work/training-configs/training.env +-/data/hfd_work/envs/omg/bin/python -u train.py --config /data/hfd_work/training-configs/omg-300m.yaml +-``` +- +-生成器始终输出四种 OMG 配置及两种 GenMo 配方;`omg.yaml` 和 `genmo.yaml` 保存所选默认项。所有配置默认单卡执行,不能把多条命令同时塞到同一块卡。 +- +-## 外部资产、数据与环境 +- +-`TRAINING_ASSETS.json` 记录必要模型、统计、机器人资源、SMPL-X 主体、原始索引和配对表的相对路径、大小、SHA256 以及已确认的下载来源。`tools/training_assets.py` 默认只验证;显式 `--download` 才会下载缺失文件,不覆盖已有文件。Llama 需按记录哈希获取同一份已授权 base,SMPL-X 主体仍需自行获取。派生 runtime adapter 路径与本机原生扩展由准备器重建。 +- +-HFD 原始数据版本必须使用 `93ba4a7a9c66b3e1b7041e5f4ea038ef2418b948`;旧 `7745110...` 缺少新增人体元数据/分片。先下载同一版本的索引、PATH_MAP、manifest 和人体配对表,再按 manifest 补齐 tar,保留 tar 不解压。训练不自动联网补缺失文件。 +- +-源文本时间戳使用 `reference-alignments.sqlite`。迁移已完成的数据库并校验身份可保持既有释放时间;也可按固定 ForcedAligner、依赖和离线对齐命令重建。重建可能有数值差异,不应声称与旧缓存逐位相同。 +- +-原始数据、外部模型和离线缓存齐备后,训练不再依赖当前服务器的 `setup/`、其他适配仓库、测试目录或运行日志。四模型的完整断点均包含优化器/调度/采样或课程所需状态。严格恢复仍要求相同代码、配方和资产;Git 历史保留原训练快照。完整操作见 [PORTABLE_TRAINING.md](PORTABLE_TRAINING.md),GenMo v2 见 [GENMO_V2_TRAINING.md](GENMO_V2_TRAINING.md)。 +diff --git a/docs/TRAINING_RESUME.md b/docs/TRAINING_RESUME.md +deleted file mode 100644 +index ba9c43b..0000000 +--- a/docs/TRAINING_RESUME.md ++++ /dev/null +@@ -1,43 +0,0 @@ +-# OMG / TextOp 训练续训 +- +-2026-10-19:两个原生训练入口已接入完整 checkpoint 恢复,适用于当前一卡 OMG、一卡 TextOp。正式配置 `training.resume: null` 表示新训练;已有输出目录在没有显式 resume 时仍会拒绝覆盖。 +- +-从仓库根目录运行,自动选择原输出目录的 checkpoint: +- +-```bash +-/root/autodl-tmp/envs/omg/bin/python train.py --config configs/omg.yaml --resume auto +-/root/autodl-tmp/envs/textop/bin/python train.py --config configs/textop.yaml --resume auto +-``` +- +-`--resume` 不带参数等价于 `--resume auto`。找不到文件会报错,不会退回从头训练。OMG 在根目录 `last.ckpt`、`checkpoints/last.ckpt` 和周期 checkpoint 中按保存的步数选择最新文件,避免旧的根目录 last 覆盖较新的周期保存点;TextOp 同样按保存步数选择。忽略原子保存中的 `.partial` 文件。 +- +-显式指定恢复点: +- +-```bash +-/root/autodl-tmp/envs/omg/bin/python train.py --config configs/omg.yaml --resume /path/to/step-00002000.ckpt +-/root/autodl-tmp/envs/textop/bin/python train.py --config configs/textop.yaml --resume /path/to/ckpt_2000.pth +-``` +- +-也可在 YAML 设置 `training.resume: auto` 或 checkpoint 绝对路径。显式命令行参数优先。可以追加 `--dry-run` 先检查 checkpoint、配置与输出目录。 +- +-## 恢复内容与边界 +- +-- OMG 由 Lightning 恢复模型、优化器、scheduler、训练步数、precision 和原生 callback 状态。 +-- TextOp 恢复 denoiser、AdamW、全局步数及启用时的 EMA;学习率退火和三阶段位置由原生逻辑按已恢复步数继续计算。 +-- 新 checkpoint 额外保存 Python、NumPy、Torch/CUDA 随机状态、动态池的 session 索引、批次位置和采样 RNG。OMG 保存实际已消费 batch 的池快照,不使用可能已被预取推进的现场状态。 +-- 不保存大型音频/动作内存缓存。恢复时重建当前池所需内容,首次读取可能仍慢。 +-- TextOp 必须在完整 primitive 链边界恢复;配置检查要求保存间隔能被 `num_primitive` 整除。 +- +-`max_steps` 是原训练计划的总目标,不是恢复后再增加的步数。续训会检查模型、数据语言、文本时钟、batch、学习率、总步数、scheduler/阶段等配置一致;允许更换单张 GPU、调整保存/验证频率,或指定新的空输出目录。当前没有接入多 GPU 的完整 HFD 采样恢复。 +- +-官方预训练权重使用 `weights.init_checkpoint` 初始化,不能当作此次 HFD 续训点。历史 HFD checkpoint 如缺少新增随机/数据池状态,仍可恢复其完整优化器和步数,但会明确提示采样不能精确接续。新 checkpoint 的训练恢复也不承诺 GPU 逐位一致;原生验证的执行与随机数消耗会影响后续数值轨迹。 +- +-只能恢复到最近成功保存的位置,断电前尚未保存的更新不会恢复。两模型正式配置现在均每 **2000 步**保存,保留最近 **3** 份;OMG 另存最佳验证 checkpoint。100K/300K 总预算未改动,正式长训练未启动。 +- +-## 日志与检查 +- +-恢复后标量日志追加写入;原始 `resolved_config.yaml` 保留。每次恢复在 `resume_history//` 保存本次配置、checkpoint 来源、代码摘要、数据清单及恢复前的原生配置/训练摘要。显式回退到较早 checkpoint 时,标量日志可能出现重复步数,可用恢复记录区分。 +- +-已通过两个真实 GPU 原生训练入口的 **4 步停止 → 恢复到 8 步**检查:恢复后各完成4次更新,Adam状态计数从4到8,OMG scheduler接续,TextOp跨阶段接续,有限损失/梯度与严格重载通过。该 smoke 使用已导出的英文 session;动态池与预取的一致性另用 CPU Lightning 回归覆盖,不代表正式全数据训练已完成。 +- +-分环境共 **132 项回归测试**通过,包括预取启用下的消费位置、优化器与 scheduler 连续性、TextOp 原生 manager 跨阶段、RNG 和采样池恢复。证据见 [验证记录](/root/autodl-tmp/runs/archive/legacy-20261020/resume-validation-20261019/verification.json)。本次验证生成的临时 checkpoint 已清理,配置、日志和结果保留。 +diff --git a/docs/WORKSPACE_CLEANUP.md b/docs/WORKSPACE_CLEANUP.md +deleted file mode 100644 +index 1551b42..0000000 +--- a/docs/WORKSPACE_CLEANUP.md ++++ /dev/null +@@ -1,78 +0,0 @@ +-# 工作区清理(2026-10-19) +- +-检查了 `git diff --stat`、逐文件 diff、`git status --short` 的未跟踪文件,以及运行时代码/测试的引用。仅用 `git diff` 会漏掉大量未跟踪脚本。新增的 bridge、动态数据读取、对齐、GenMo 和 retarget 模块仍是实际功能,不能按“未跟踪”直接删除。 +- +-已将 **43 个一次性实验脚本**移出工作区;包括旧 ASR/翻译探针、不同阶段的预算和吞吐 benchmark、中英 pilot、已完成过拟合与验证的专用汇总脚本。旧的 `omg_asr_context_overfit.yaml`、`textop_asr_context_overfit.yaml`、`omg_budget60k.yaml` 三份配置移入 `configs/history/`,引用已更新。未重置或提交用户已有改动。 +- +-以下保留在日常工作区: +- +-- 正式训练入口、当前配置与 resume 实现。 +-- 动态 session/缓存适配、文本时钟、音频和动作特征、映射等运行时模块。 +-- OMG/TextOp/GenMo 推理、retarget、视频评测入口及必要回归测试。 +-- 原始数据、模型、已有实验权重、日志和报告。 +-- 七个数据准备/审查工具:`align_reference_corpus.py`、`audit_reference_text.py`、`download_hfd_tars.py`、`index_reference_recipes.py`、`prepare_reference_sessions.py`、`verify_dynamic_reference.py`、`verify_reference_release.py`。 +- +-## 归档与恢复 +- +-脚本移出前已打包,并逐文件验证归档内容 SHA256,清单保存原路径、大小和哈希。归档在服务器工作区外,若迁移后还需复现实验,应同时复制该归档。 +- +-- [归档包](/root/autodl-tmp/hfd-code-archive/20261019-resume-cleanup/historical-experiments.tar.gz) +-- [逐文件清单](/root/autodl-tmp/hfd-code-archive/20261019-resume-cleanup/cleanup.json) +-- [清理前 Git 状态](/root/autodl-tmp/hfd-code-archive/20261019-resume-cleanup/git-status-before.txt) +-- [清理前已跟踪文件 diff](/root/autodl-tmp/hfd-code-archive/20261019-resume-cleanup/git-diff-before.patch) +- +-按需恢复某个脚本,例如: +- +-```bash +-tar -xzf /root/autodl-tmp/hfd-code-archive/20261019-resume-cleanup/historical-experiments.tar.gz -C /root/autodl-tmp/hfd-baseline-portable tools/run_english_training_probe.py +-``` +- +-历史报告中的脚本路径属于当时的实验记录,已添加归档提示;当前报告、指标和视频链接保留。归档 SHA256:`1fc40fe878610f4c37610d77a56aee9a5db70bfb32ed9bf7c30d25ae71fba06c`。 +- +-## 移出工作区的脚本 +- +-| 原路径 | 字节 | +-| --- | ---: | +-| `tools/analyze_omg_budget.py` | 5110 | +-| `tools/audit_asr_context.py` | 3829 | +-| `tools/audit_genmo_audio.py` | 4258 | +-| `tools/audit_genmo_inputs.py` | 3353 | +-| `tools/audit_reference_translation.py` | 3706 | +-| `tools/benchmark_omg_data.py` | 6149 | +-| `tools/benchmark_omg_io.py` | 2526 | +-| `tools/benchmark_omg_resident.py` | 4317 | +-| `tools/benchmark_omg_training.py` | 8946 | +-| `tools/benchmark_textop_batch.py` | 7413 | +-| `tools/benchmark_translation_batch.py` | 2800 | +-| `tools/check_genmo_ambiguity.py` | 1781 | +-| `tools/compare_asr_models.py` | 7332 | +-| `tools/compare_english_checkpoints.py` | 6201 | +-| `tools/diagnose_omg_validation.py` | 6358 | +-| `tools/eval_bilingual_pilot.py` | 3519 | +-| `tools/eval_reference_overfit.py` | 3580 | +-| `tools/initial_bilingual_validation.py` | 2183 | +-| `tools/prepare_bilingual_pilot.py` | 3325 | +-| `tools/prepare_omg_validation.py` | 4446 | +-| `tools/probe_asr_latency.py` | 2183 | +-| `tools/probe_bilingual_condition.py` | 4904 | +-| `tools/probe_direct_chinese.py` | 7190 | +-| `tools/probe_reference_alignment.py` | 15248 | +-| `tools/prune_chinese_caches.py` | 6225 | +-| `tools/render_bilingual_pilot.py` | 2202 | +-| `tools/run_bilingual_pilot.py` | 4289 | +-| `tools/run_english_training_probe.py` | 1537 | +-| `tools/run_omg_validation.py` | 5170 | +-| `tools/summarize_asr_context.py` | 5454 | +-| `tools/summarize_asr_throughput.py` | 6969 | +-| `tools/summarize_bilingual_pilot.py` | 8655 | +-| `tools/summarize_english_probe.py` | 7373 | +-| `tools/summarize_genmo_audio.py` | 5254 | +-| `tools/summarize_genmo_overfit.py` | 4932 | +-| `tools/summarize_omg_validation.py` | 9952 | +-| `tools/summarize_reference_overfit.py` | 9336 | +-| `tools/summarize_reference_preparation.py` | 6832 | +-| `tools/summarize_rollout_debug.py` | 4354 | +-| `tools/summarize_textop_batch.py` | 5570 | +-| `tools/verify_bilingual_pilot.py` | 3282 | +-| `tools/verify_native_batch_limit.py` | 2224 | +-| `tools/verify_semantics_clock.py` | 4106 | +diff --git a/docs/history/BASELINE_METHOD_BEFORE_SEMANTIC_FIX.md b/docs/history/BASELINE_METHOD_BEFORE_SEMANTIC_FIX.md +deleted file mode 100644 +index 0ea8331..0000000 +--- a/docs/history/BASELINE_METHOD_BEFORE_SEMANTIC_FIX.md ++++ /dev/null +@@ -1,95 +0,0 @@ +-# 最终采用的 baseline 输入与推理方式 +- +-本轮完成了用户提出的累计音频、原生推理节奏对照。两模型仍使用原生网络、损失、训练循环和官方初始化权重;没有新增 KV 模块、控制分类头、任务进度真值或动作平滑。测试样本为原生双任务 interruption session `train-002203-seed42`:接取物品未完成时输入捡起地上物品的新指令。原始声音、背景音和动作时间线保留。 +- +-## 最终选择 +- +-| 模型 | 输入 | 动作生成与衔接 | +-| --- | --- | --- | +-| OMG | 保存原始录音;模型每次编码截至 HFD 当前释放时刻的最近两秒 PCM。开头左侧补零,保留原生 60 个声学条件位置 | 每次预测原生 60 帧 / 两秒;模拟原生异步参考缓冲,剩余 24 个 30 Hz 参考帧(0.8 s)时请求新计划。计算期间继续执行旧参考,新计划跳过已耗时对应帧后接入 | +-| TextOP | 由已释放音频生成的因果 ASR;当前实验保持最新稳定 ASR 句子,文字内容连续两次一致才更新,标点变化不算新命令;继续使用原生 CLIP token 上限 | 原生 8 帧 / 0.16 s primitive,约 6.25 Hz;持续反馈预测历史,换文本时不重置姿态;原生 guidance=5 与三阶段训练 | +- +-音频到达频率、语义条件变化、动作生成频率是三个独立事件。音频仍按 HFD 0.6 秒窗口 / 0.4 秒释放间隔提供可用边界;模型可以重新编码更长的已听历史,但不能读取边界之后的录音。OMG 不强制每 0.4 秒抛弃未来计划,也不等整个任务做完才处理新指令。 +- +-OMG 缓冲参数对应原生 tracker 50 Hz 下剩余 40 帧的 0.8 秒时间,当前在 30 Hz 的参考轨迹上实现为 24 帧。两秒计划减去 0.8 秒缓冲,对应大约 1.2 秒的规划请求间隔;新计划真正接入还受计算时间影响。 +- +-## 为什么没有采用“全历史直接输入” +- +-累计录音已经保留。另一个 OMG 模型从官方初始化,用 `hfd_full_prefix` 在同一 session 上训练 1600 步:每个动作点读 `[0,当前释放截止)` 的全部 PCM,再把原生声学特征线性插值到固定 60 个条件位置。训练和推理策略一致,未来录音不会影响当前特征。 +- +-这个适配没有带来更好的结果,且随着历史增长降低时间分辨率、增加重新编码成本。OMG 原生音频编码为 35D 声学特征,不是语言语义编码器。此结果只能否定本次“全历史压缩到固定位置”的适配效果,不能概括为完整音频历史没有价值。 +- +-最终采用两秒历史,是为了保留原生特征的时间排列和一致的训练/推理接口。超过两秒的指令仍可能被截断,这是明确的 baseline 上下文限制。改成可变长语义编码器将涉及网络改造,本轮没有这样做。 +- +-TextOP 输入的是文字,不是原始 PCM。此 14.8 秒样本短于 ASR 的 30 秒历史上限,ASR 本来就在每个截止时刻重新识别从开始到当时的全部录音。真正改变的是供给 CLIP 的指令组织和稳定保持方式。`stable_latest_sentence` 是词法预处理,不使用 GT turn、事件标签或真值任务文本;它也不是可靠的语音任务边界检测器。多句复合指令、复杂犹豫或 ASR 幻觉可能使末句不代表完整任务;可用 `rolling_transcript` 做保留完整转写的对照,正式训练需在既定验证集核对这一限制。 +- +-## 结果 +- +-所有误差来自该训练 session 的连续生成,开头真值初始化帧被排除;不代表独立测试集或完整全双工评测。TextOP 指标为 23 个建模关节,OMG 为 29 个,不能直接横向排名。 +- +-| 实验 | 连续关节 RMSE(rad) | root RMSE(m) | +-| --- | --- | --- | +-| OMG 两秒历史,0.4 s 重规划(前轮) | 0.25026 | 0.04641 | +-| OMG 两秒历史,同步两秒块执行(前轮) | 0.18776 | 0.03093 | +-| OMG 两秒历史,模拟原生异步缓冲,本轮最终测时 | 0.09436 | 0.01764 | +-| OMG 全历史压缩,模拟原生异步缓冲 | 0.22073 | 0.03660 | +-| TextOP 滚动转写、原生周期、guidance=5(前轮) | 0.40908 | 0.03739 | +-| TextOP 稳定最新句、原生周期、guidance=5 | 0.30987 | 0.02933 | +- +-OMG 两秒历史模型已在前轮完成同预算 1600 步训练,本轮复用其唯一最终权重测试缓冲;累计历史模型和新 TextOP 均重新从官方初始化训练 1600 步。没有为不同执行周期挑选不同检查点。 +- +-TextOP 真实历史窗口 RMSE 为 0.00700 rad,但连续生成仍未充分拟合。相同静止历史下切换两条指令,8 帧输出差异约 0.00066 rad;相同文本下换成真实动作历史,输出差异约 0.22 / 0.47 rad。这说明此次小样本拟合仍主要依赖姿态续接,不能把稳定文本和原生周期描述为已经解决任务启动问题。 +- +-OMG 固定时延敏感性对照:异步缓冲用零计算时延时为 0.10596 rad,固定 0.2 秒为 0.08031 rad,最终实测时延为 0.09436 rad。人工固定时延仅用于诊断,最终没有选择最有利的 0.2 秒去改善指标。 +- +-## 推理耗时与时钟边界 +- +-本机 RTX 5090、FP32 原生 PyTorch,模型加载/预热后: +- +-| 输入 / 模型 | 每次规划平均耗时 | P95 | +-| --- | --- | --- | +-| OMG 两秒历史 | 0.479 s | 0.489 s | +-| OMG 累计历史 | 0.523 s | 0.555 s | +-| TextOP primitive | 0.066 s | 0.112 s | +- +-OMG 计时包含本轮声学特征、历史 FK/编码和原生生成,并在 CPU 取回预测后结束计时。TextOP 包含当前 CLIP 条件读取/编码及原生生成,不包含之前生成 ASR sidecar 的 Whisper 计算。渲染与真实机器人控制器不计入。 +- +-在当前未优化后端,OMG 本身已超过 0.4 秒预算,不能宣称 0.4 秒内完成完整推理和执行。异步缓冲允许旧参考在规划期间继续输出,最终实测无缓冲耗尽。模型输出回到缓冲时,按实测时间向上取整到 30 Hz 帧数,跳过新计划中已过去的帧;旧参考提供这段 bridge,而不是把未执行的新预测当作实际历史。 +- +-这是离线源时间参考轨迹上的执行模拟,使用真实测得的规划耗时;视频仍是原始音频内容对齐的姿态回放。它不是实时麦克风、ASR、机器人 tracker 的完整墙钟部署,不从视频直接推导论文中的硬件反应时延。HFD 音频释放截止、源动作时间、真实计算耗时分别记录在报告中。 +- +-## 配置、复现与迁移 +- +-采用的短训练配置:`configs/history/omg_adopted.yaml`、`configs/history/textop_adopted.yaml`。正式训练模板 `configs/omg.yaml`、`configs/textop.yaml` 也记录相同输入和推理策略,但独立数据划分、正式训练和论文评测尚未完成。按根目录 README 安装两个环境并准备外部权重、ASR、机器人 assets;更新 YAML 中路径、GPU 和 batch size 后训练到新的 output_dir。 +- +-```bash +-python train.py --config configs/history/omg_adopted.yaml +-python train.py --config configs/history/textop_adopted.yaml +-python infer_overfit.py --run NEW_OMG_RUN --continuous-only --task-seconds 0 14.8 +-python infer_overfit.py --run NEW_TEXTOP_RUN --continuous-only --task-seconds 0 14.8 +-``` +- +-当前已验证权重与输出: +- +-- OMG 权重:`/root/autodl-tmp/runs/hfd-omg-context-debug/omg-past2-train/last.ckpt`;推理输出 `hfd-native-context-final/omg-past2-adopted`。 +-- TextOP 权重:`/root/autodl-tmp/runs/hfd-native-context-final/textop-stable-command/ckpt_1600.pth`;推理输出该 run 的 `inference`。 +- +-复现最终 OMG 用 `--omg-replan-mode native_async --output-dir NEW_OUTPUT`,因为该保留权重的旧 resolved_config 仍记录前轮重规划设置。新 adopted 配置把这些推理设置放在 `inference` 中,无需 CLI 覆盖。 +- +-重要配置项:`data.audio_policy`、`data.caption_policy`、`inference.omg_replan_mode`、`inference.omg_buffer_remaining_frames`、`inference.textop_guidance_scale`。`--omg-latency-frames` 是诊断覆盖,正式采用实测时间,不在配置中固定有利的延迟。 +- +-全历史对照用 `configs/history/omg_native_context.yaml`,其中 `hfd_full_prefix` 不作为最终默认。缓存仅用于训练时重复窗口的确定性声学特征,单 source 最多 128 组;推理不复用未接收的未来条件。缓存与直接提取逐项一致。 +- +-## 原生依据与核验 +- +-- [TextOP 原生部署论文](https://arxiv.org/html/2602.07439v1#S3.SS5):生成器 6.25 Hz,tracker 50 Hz。 +-- [OMG 原生执行文档](https://github.com/Tsinghua-MARS-Lab/OMG/blob/main/docs/generation.md):同步与异步参考缓冲。 +-- 本地固定版 `upstream/omg/src/omg/cli/pipeline/modes.py`:旧参考 bridge、新参考跳过 bridge 时段;原生 tracker 参数剩余 40 帧 / 50 Hz。 +- +-新两个训练的优化器更新、参数变化、有限 loss/梯度和严格检查点重载均通过。数据合同检查和 37 项回归测试通过,覆盖累计音频未来不可见、缓存一致性、ASR 稳定策略的因果性、缓冲 bridge/跳过位置与 underrun 检测。全历史模型的三个等价重复检查点已清理,唯一最终权重与对照记录保留。 +- +-## 当前语义保持方案(2026-10-07) +- +-OMG 的正式 smoke 配置使用 `motion_only` 监督,并把原生 ASR 作为因果文本条件接入。每个推理时刻仍使用最近 2 秒 PCM;ASR 的最后一个已稳定句子在后续背景音阶段持续保留,因此任务描述结束后不会因为音频窗口滑走而丢失目标。ASR 文本只使用当时已经释放的转写,不能看到未来文本或真值动作。 +- +-OMG 代码保留两种 guidance:`joint_one` 是当前可复现实验默认值(单次采样,适合异步缓冲);`native` 启用原生文本/音频分支引导。后者在本机一次规划约需 1.37 秒,在 0.8 秒异步余量下会缓冲欠载,已记录为诊断结果,不能宣称实时可用。TextOP 的 3200-step 起步诊断连续误差为 0.2434 rad,窗口误差为 0.0094 rad;这说明训练链路有效,但整段闭环仍未达到理想过拟合。 +diff --git a/docs/history/DUPLEX_TASK.md b/docs/history/DUPLEX_TASK.md +deleted file mode 100644 +index 8ba0556..0000000 +--- a/docs/history/DUPLEX_TASK.md ++++ /dev/null +@@ -1,79 +0,0 @@ +-> 历史实验记录;当前入口见 [README](../../README.md)。旧运行目录已移动到 `/root/autodl-tmp/hfd-runs-archive-20261007/`。 +- +-# 两个正常任务的完整全双工 session +- +-本例选用原生 HFD `train-002203-seed42`:任务一是接取别人递来的物品(`neutral_item_take_R_001__A545`),任务二是弯腰捡起地上的物品(`neutral_item_pick_up_floor_R_002__A544`)。英文女声 `female_02_en`;保留原始录音与空调背景音。两项任务都包含实质动作,原有 listen/idle 过渡仍保留在时间线上,标签没有作为模型的额外控制输入。 +- +-这是原生 interruption session,第二条指令到达时第一项动作尚未完成。第一项动作会被中断,随后完成第二项动作;不是声称两个动作都完整执行结束。两任务之间的姿态过渡由原生 HFD 构造器处理,没有人为把两个独立视频拼在一起,也没有对预测做平滑。 +- +-## 时间线 +- +-运动时间线长 14.8 秒,50 Hz 共 740 帧。 +- +-| 时间 | 原生内容 | +-| --- | --- | +-| 0–1.2 s | 开场 idle;第一条录音实际在运动时间 0.4–2.16 s 可听见 | +-| 1.2–3.2 s | 第一条指令的 listen/姿态过渡 | +-| 3.2–6.0 s | 第一项动作;第二条录音在运动时间 5.2 s 到达,动作仍在执行 | +-| 6.0–8.0 s | 第二条指令的 listen/姿态过渡 | +-| 8.0–12.0 s | 第二项实质动作:捡起地上的物品 | +-| 12.0–14.8 s | 收尾 idle | +- +-`turn_speech_start_frames=[20,260]`。音频数组相对运动有原生 0.2 秒前置偏移:未裁剪音频中的实际发声时间为 0.6–2.36 s、5.4–7.16 s。导出的视频按此偏移对齐,所以视频中是 0.4–2.16 s、5.2–6.96 s。原生中断记录显示第一项动作剩余 86 帧,即 1.72 秒;不是由视频外观猜测。 +- +-## 训练与视频口径 +- +-两个模型从官方初始化权重开始,用各自原生训练框架在此 session 上训练 1600 步。训练目标窗口覆盖该 session 的时间范围 `[0,730)`,只跳过尾部 0.2 秒 idle;`max_windows: null`,不局限于此前八个稀疏窗口。TextOP batch size 4、原生四 primitive 链;OMG batch size 8、原生 60 帧未来动作窗口。OMG 使用完整目标区间音频,属于之前确认的离线口径;TextOP 使用原始录音的因果 ASR,不使用真值指令文本。少样本选择跳过正式训练的大规模交互配比,不改变原生构造算法。 +- +-视频覆盖整个 session,从开场连续推理到收尾。只在开头用一次真值历史初始化,随后使用预测历史,没有在第二条指令到达时重置姿态。TextOP 以原生 8 帧 primitive 为单位,可能略少于最后一个不完整 primitive;OMG 保留整段 30 Hz 时间线。前缀真值初始化单独标注,并从连续推理误差中排除。左右分别为预测和真值,视频只是关节姿态播放,不包含物理控制和实际物体交互。 +- +-视频统一为 H.264 + MP3 的 MP4,使用 `_playable.mp4` 新文件名。主入口为 `/root/autodl-tmp/runs/hfd-overfit-videos-jointfixed/README.md`。结果指标和音频验证写入 `duplex_task_summary.json`;窗口指标使用真值历史,连续整段指标使用预测历史,二者不能混为模型完全学会全双工任务。 +- +-## 迁移后复现 +- +-先按根目录 README 准备两个环境、官方权重、外部机器人 assets 与 HFD 原始 tar 分片路径,再更新 YAML 中绝对路径。依次执行(工作目录是代码文件夹): +- +-```bash +-/root/autodl-tmp/envs/omg/bin/python prepare_duplex_task.py --output /root/autodl-tmp/runs/hfd-duplex-task-overfit/sessions +-/root/autodl-tmp/envs/omg/bin/python transcribe_sessions.py /root/autodl-tmp/runs/hfd-duplex-task-overfit/sessions/train-002203-seed42.npz --device cuda:0 --language auto --output-dir /root/autodl-tmp/runs/hfd-duplex-task-overfit/asr +-/root/autodl-tmp/envs/textop/bin/python train.py --config configs/history/textop_duplex_task.yaml +-/root/autodl-tmp/envs/omg/bin/python train.py --config configs/history/omg_duplex_task.yaml +-/root/autodl-tmp/envs/textop/bin/python infer_overfit.py --run /root/autodl-tmp/runs/hfd-duplex-task-overfit/textop --continuous-only --task-seconds 0 14.8 --textop-guidance-scale 1 +-/root/autodl-tmp/envs/omg/bin/python infer_overfit.py --run /root/autodl-tmp/runs/hfd-duplex-task-overfit/omg --continuous-only --task-seconds 0 14.8 +-``` +- +-渲染每个 `RUN/inference`: +- +-```bash +-/root/autodl-tmp/envs/omg/bin/python render_overfit.py --inference RUN/inference --assets-root /root/autodl-tmp/datasets/eval_textop_omg/OMG_code_20260927/OMG/assets --gpu 0 +-``` +- +-选择脚本默认构建 512 条原始 source 的候选池,筛选指定两动作的原生 interruption session,保存 `selection.json` 和原生 metadata。输出目录必须不存在,防止混用旧样本和检查点。 +- +-## 清理检查点 +- +-`cleanup_checkpoints.py` 只清理 OMG 同一训练步、每一个 state_dict 张量完全相同的重复保存。正式初始化权重、唯一训练结果、TextOP 所需 VAE 都保留。先检查,然后执行: +- +-```bash +-/root/autodl-tmp/envs/omg/bin/python cleanup_checkpoints.py RUN +-/root/autodl-tmp/envs/omg/bin/python cleanup_checkpoints.py RUN --apply +-``` +- +-每个 run 保留 `last.ckpt`,删除等价 `pipeline.ckpt` 和 `checkpoints/*.ckpt`,记录 `checkpoint_cleanup.json` 并更新训练检查报告。不同参数或不同 global_step 会拒绝删除。可复现代码目录不包含这些大文件。 +- +-## 本次检查结果 +- +-两模型的 1600 次优化更新、参数变化、有限 loss/梯度和严格检查点重新加载均通过。23 项回归测试通过。 +- +-| 模型 | 训练窗口初始化关节 RMSE | 训练窗口训练后关节 RMSE | 整段连续关节 RMSE | 整段 root RMSE | +-| --- | --- | --- | --- | --- | +-| TextOP | 0.03169 rad | 0.00516 rad | 0.32428 rad | 0.02954 m | +-| OMG | 0.30846 rad | 0.10780 rad | 0.10088 rad | 0.01108 m | +- +-TextOP 的短窗口明显拟合,但整段预测仍未跟随两项动作。OMG 在整段中能看到接取和弯腰动作,但仍有姿态误差。这符合训练链路测试的用途,不作为两个 baseline 充分拟合或正式评测的结论。TextOP 的指标只计其建模的 23 关节,OMG 计 29 关节,不能直接据此排名。视频初始化分别占 2 帧 / 13 帧,已从整段指标中排除。 +- +-视频全部帧和音轨均完整解码通过,配音在编码前与原生 session 对应音频切片逐样本完全相等。两条视频音轨均非静音。TextOP 为 738 帧、14.76 s;OMG 为 444 帧、14.8 s。 +- +-本次共移除三个 OMG run 中的 9 份等价重复检查点,释放 9,535,716,429 bytes(约 8.88 GiB)。每个 run 的唯一最终 `last.ckpt` 都保留。清理记录汇总位于 `/root/autodl-tmp/runs/diagnostics/checkpoint_cleanup_summary.json`。 +- +-本页旧 TextOP 数值和视频使用 guidance=1。当前默认推理已补齐原生 guidance=5;复现旧结果须显式指定 `--textop-guidance-scale 1`。当前窗口输入与新检查见 [HFD_WINDOW_PROTOCOL.md](HFD_WINDOW_PROTOCOL.md)。 +diff --git a/docs/history/HFD_HUMAN_BASELINE_INPUT_PROTOCOL_RESEARCH_20261017.md b/docs/history/HFD_HUMAN_BASELINE_INPUT_PROTOCOL_RESEARCH_20261017.md +deleted file mode 100644 +index 4c560a7..0000000 +--- a/docs/history/HFD_HUMAN_BASELINE_INPUT_PROTOCOL_RESEARCH_20261017.md ++++ /dev/null +@@ -1,318 +0,0 @@ +-> 历史调研快照,归档于2026-10-21,包含2026-10-17起调研及后续历史补充。当前约定见[最新输入协议](../HFD_HUMAN_BASELINE_INPUT_PROTOCOL.md),以下旧默认与待实现状态不用于当前训练。 +- +-# HFD 人体 baseline:模型结构、因果输入与训练约定 +- +-当前输入修复:新训练配置使用 `clock_policy: microphone` 和 `text_conditioning: asr_to_english`,保留 `causal_asr_context`。按 `r_k <= t_motion+0.2s` 选择已释放 PCM/ASR,中文 ASR 经冻结翻译后交给原生文本编码器;视频完整保留麦克风开头。无新字段的历史 run 仍保留旧行为。详见 [中文语义与时钟修复](../SEMANTICS_CLOCK_FIX.md)。 +- +-调研日期:2026-10-17(Asia/Taipei)。状态:**代码与论文调研完成;以下新模型配方是待实现、待验证的建议,不是已经跑通的训练链路。** +- +-2026-10-18 实现进展:GenMo 的单条成对 session、原生训练桥接、连续推理、GMR 与视频链路已实现;短/长历史实验及限制见 [GENMO_HFD_OVERFIT.md](../GENMO_HFD_OVERFIT.md)。下文保留调研时的原始建议,不代表全部任务已验收。 +- +-2026-10-18 用户确认:五个 baseline 默认文本策略统一为 **`causal_asr_context`**;OMG/TextOp 已完成同 session 过拟合对照。此决定覆盖下文保留的历史建议。当前共同规则见 [五模型文本条件约定](../BASELINE_TEXT_CONDITION.md)。旧训练权重和结果仍按其原始 manifest 标识。 +- +-## 1. 推荐结论 +- +-先冻结同一份全双工 session 计划,生成经过验收的人体动作时间轴;再做各模型的表示转换。**统一信息可用时间,不强求模型接收相同形状的张量。** 0.6 秒是 HFD 每次新增的音频窗口,不是所有 baseline 的语义记忆长度,也不是动作预测长度。 +- +-第一版建议保留各模型的文本编码器,以与现有 TextOP/OMG 一致的因果 ASR 提供语义;用当前已知动作提供连续性。原生支持掩码条件的模型沿用该接口;没有该接口的模型必须明确标注为适配版本。不要把三者改造成同一个 Qwen 音频前端模型后仍称作原生 baseline。 +- +-| 模型 | 核实到的原生结构与条件 | 建议的 HFD 输入 | 适配性质 | +-| --- | --- | --- | --- | +-| 当前 HFD | Qwen 流式语音上下文;动作端有状态/历史和 RTC | 每次 0.6s PCM、0.4s 更新,session 内积累语义历史 | 参照系统;不等于独立短窗分类 | +-| 当前 TextOP | 文本条件的动作 primitive 生成 | 最近 30s 完整因果 ASR+原生 2 帧动作历史 | 本地已实现;50Hz,8 帧未来 | +-| 当前 OMG | 文本/声学条件的动作生成 | 同一因果 ASR+过去 2s 已释放 PCM+原生 10 帧历史 | 本地已实现;30Hz,60 帧未来 | +-| GenMo / GEM-SMPL | Transformer 扩散;T5 文本、逐帧音频/音乐/视觉条件;有 3D motion overwrite 接口 | 主实验:ASR 文本+已知动作前缀,视觉/音乐/语音分支标为缺失;另做 ASR+过去音频实验 | 有原生训练入口,但前缀训练分布需适配验证 | +-| HY-Motion 1.0 | 双流/单流 MMDiT,flow matching;Qwen3 token 特征+CLIP-L 全局特征 | ASR 文本经两个原生编码器;流式版另加显式动作前缀条件 | 没有公开的原生历史动作入口;流式版是方法扩展 | +-| Kimodo | root→body 两阶段 Transformer 扩散;LLM2Vec 文本+observed motion+mask | ASR 文本+历史姿态约束及 mask;未来无真值约束 | 原生约束接口可用;连续前缀微调和训练循环需实现 | +- +-建议先实现 **Kimodo-SMPLX 的输入转换与短样本闭环**,其次 GEM,最后 HY-Motion 的流式扩展。Kimodo 的 SMPLX 路线转换最直接,但官方提示该版本有 retargeting artifacts;这里只把它作为接入起点,不预先认定质量最好。若后续采用 SOMA checkpoint,需要单独验证人体到 SOMA 的转换,不能把当前 SMPL-X 数组换名当 SOMA。 +- +-## 2. 本次依据与版本 +- +-用户引用的两个线程已从本机保存记录核对: +- +-- `codex://threads/01a13a26-2519-7861-a188-c7adaa9667c1`:人体源动作交接与下载。 +-- `codex://threads/01a101b8-f059-72d1-90b4-9749bf524531`:OMG/TextOP 训练、third-party、I/O。 +- +-本地现行约定以 [SESSION_FORMAT.md](../../SESSION_FORMAT.md)、[README.md](../../README.md)、[SMPLX_HANDOFF.md](/root/autodl-fs/pangxinlei/hfd_data/SMPLX_HANDOFF.md) 为准,不从历史实验文档恢复旧策略。 +- +-官方源码通过 AutoDL 下载到 `/root/autodl-tmp/downloads/hfd-human-research-20261017/`,跳过 LFS 内容和子模块;未安装环境、下载模型权重或启动训练。 +- +-| 上游 | 本次固定 commit | 论文/官方说明 | +-| --- | --- | --- | +-| [NVlabs/GENMO](https://github.com/NVlabs/GENMO) | `16bebf402d8893184249ee206d957b8248cd8310` | [GENMO,ICCV 2025](https://arxiv.org/abs/2505.01425);仓库已改称 GEM | +-| [Tencent-Hunyuan/HY-Motion-1.0](https://github.com/Tencent-Hunyuan/HY-Motion-1.0) | `4e426f5a1021cbcf7f375458c37b840ee7225229` | [HY-Motion 1.0](https://arxiv.org/abs/2512.23464) | +-| [nv-tlabs/kimodo](https://github.com/nv-tlabs/kimodo) | `e1c219c99b4b94ba48be1fb6a1cbc64ecee4c954` | [Kimodo](https://arxiv.org/abs/2603.15546)、[官方文档](https://research.nvidia.com/labs/sil/projects/kimodo/docs/) | +- +-逐文件 SHA256 和源码定位另见 [调研来源清单](../HFD_HUMAN_BASELINE_SOURCES.json)。本次是静态核对,不声称已验证 checkpoint 兼容性、显存或推理速度。 +- +-## 3. 先构造正确的人体 session +- +-### 3.1 当前 human 包的边界 +- +-交接包提供 142,220 条成对 SMPL-X 源动作,而非已经验证的完整全双工人体训练集。本地看到 128 个 `seed_smplx/*.tar`,本次未重新做全量分片哈希审计。 +- +-按动作身份连接:本机实际读取的 `metadata/seed_g1_smplx_pairs.parquet` 使用 `motion_id` 列,对应 G1 metadata 的 `filename`;人体路径直接取表中的 `smplx_path`。保留原 split、八音色、背景、Stop、backchannel、hesitation、pause 等音频原料。禁止按行号、G1 帧下标或文件名字串猜对应关系。 +- +-人体源含 `trans[T,3]`、`root_orient[T,3]`、`pose_body[T,63]`、手/脸、`betas[16]` 和 `sample_times`。`sample_times` 是时间依据;源帧率并非严格 30Hz。`trans` 是 AMASS 模型平移,不能直接解释成世界 pelvis 位置。 +- +-### 3.2 两种身体共用计划,各自合成几何 +- +-应拆成 `SessionPlan → render_g1(plan) / render_human(plan)`。计划至少固定:split、motion/source ID、每段源时间映射、1–4 轮顺序、音色、背景偏移与增益、语音事件、动作开始/截断/保持/恢复、idle/listen 素材、过渡长度,以及所有随机选择。仅同 seed 或同 source index 不足以保证两版成对。 +- +-人体 renderer 采用共同 50Hz session 时间轴;根平移按秒插值,旋转用 SO(3)/SLERP。根据当前人体末位置和朝向拼接下一段,使用一致体型;不能复制机器人根高度、关节限制或保持姿态数组。idle/listen 和 Stop 过渡也必须重新构造人体几何。 +- +-验收后再把**已合成的整条人体 session**转到模型的 30Hz 时间网格。派生输入的最后一个历史采样点不得通过插值使用决策边界后的 GT 帧:在已知前缀上单独插值/边界保持;未来标签可按完整 GT 重采样。记录两者的取样规则。 +- +-### 3.3 third-party 不能自动宣称已有人体版 +- +-当前 portable 已接入 `hfd_third_party_multi_v5`,但其读取器加载的是生产包的 `[T,36] states`,即 G1 动作。新 SMPL-X 源动作不自动补出这些生产 session 的人体几何。必须恢复其完整源片段/时间映射,或从冻结计划同时渲染两版;不能从 G1 36D 直接构造 SMPL-X 真值。 +- +-若暂时无法恢复,应在所有参加该轮对比的模型上使用共同的、已验证的人体 session 子集,注明不含该部分 third-party。不要用新模型的较小数据集与既有 OMG 训练结果直接归因模型优劣。旁人说话的 PCM 应保留;数据层不能用旁人 GT 标签替模型过滤输入。 +- +-## 4. 因果边界与共同数据契约 +- +-### 4.1 三个时钟分开保存 +- +-当前 portable 对第 `k` 个语义组规定: +- +-```text +-新增音频窗口:W_k = audio[6400*k : 6400*k + 9600] # 16kHz +-音频释放时刻:r_k = 0.6 + 0.4*k # 音频轴,秒 +-原生动作起点:i = 20*k + j,j=0..19 # 50Hz 动作索引 +-该组内条件:仅使用截至 r_k 释放的音频/ASR +-``` +- +-这来自 `hfd_bridge/sessions.py::SessionWindows`。`i/50`、`r_k`、实际墙钟不是可随意替换的同一时间。接入人体系时必须保留这张映射;不要直接以 `audio[:i/50]` 替代现行约定,也不要因为 `r_k` 数值大于动作名义时间就擅自移位动作监督。原始语音与动作内容的约 0.2 秒偏移,也不能当作这里的释放延迟使用。 +- +-如果另建“麦克风物理时间零点”的真实在线评测,需明确初始预缓冲和动作轴到墙钟的映射,所有模型共用;它与复现现行 portable 是不同的时间协议。 +- +-### 4.2 共同可用信息 +- +-在一次决策处,允许输入:已释放的混合 PCM、由这些 PCM 得到的 ASR、已知当前姿态及过去已执行姿态、固定的身体/坐标定义,以及事先指定的计算窗口长度。 +- +-不允许输入:未来音频、整句 GT 指令、最终纠正后的文本、GT turn ID、listen/motion/idle 标签、未来 Stop 时刻、未来动作长度/进度、未来轨迹/落脚点、从未来 GT 动作算出的历史特征。GT 标签可用于监督、分层抽样和评估,不可变成推理时不可得的条件。 +- +-噪声化的未来动作是扩散/flow matching 的训练变量;这属于目标学习过程。问题在于把干净未来特征送入条件,或在评估时提供它们。未来生成 token 相互注意本身不违反在线输入因果性,因此不应把所有模型强制改成逐帧 causal attention。 +- +-### 4.3 建议缓存字段 +- +-```text +-session_id, split, plan_hash, recipe_hash, pcm_hash +-decision_index_50hz, semantic_group, audio_release_sample +-motion_sample_times, known_motion_end, native_fps +-asr_cache_hash, asr_policy, asr_text, asr_audio_end_sample +-body_model_id, shape_id, coordinate_transform, representation_id +-history_features, history_feature_mask, valid_history_mask +-future_target_features, valid_future_mask +-``` +- +-训练专用的目标和控制标签与输入字段隔开;ASR 缓存绑定 PCM 和识别配置,不绑定某种模型的姿态。G1/human 共用 PCM 后可以复用 ASR;人体特征缓存必须独立。 +- +-## 5. 文本条件、语义记忆与全双工限制 +- +-### 5.1 五个 baseline 默认采用完整历史 ASR +- +-复用现有 Whisper-tiny、`language=auto`,每个语义组仅识别截至 `r_k` 的最多 30 秒历史混合音频。原始 sidecar 保留中英文和背景,不翻译、不从 metadata 修正;当前模型条件另经冻结中文→英文转换。离线预计算必须逐释放点截断后识别;先识别完整 session 再按词时间戳切分会利用未来声学上下文。 +- +-用户已确认五个 baseline 统一采用 `causal_asr_context`:每次使用截至当前释放点、最近最多 30 秒音频的完整识别结果;非空立即更新,空结果沿用上一份非空文本,session 结束清空。不要把历次重识别结果不断拼接,也不要用未来识别结果回填过去。最后一句允许是不完整的当前假设,不再等待连续两次一致。 +- +-五个模型共用同一字符串,再各自通过原生 tokenizer/encoder:OMG 的 T5、TextOp 的 CLIP、GenMo 的 T5-3B、HY-Motion 的 Qwen3/CLIP、Kimodo 的 LLM2Vec。embedding 不共享,长度预算按原生配置,缓存并报告截断;不得静默换成摘要、GT 指令或“当前任务”。共同实现为 `hfd_bridge.textop_condition.caption_contexts`。 +- +-### 5.2 对照结果与历史兼容 +- +-OMG/TextOp 的相同 session 对照见 [完整历史 ASR 过拟合测试](../ASR_CONTEXT_OVERFIT.md)。TextOp 有改善但仍不稳定,OMG 没有整体改善。两种策略同时改变了历史覆盖范围和两次确认延迟,不能只归因于上下文长度;保留多个句子也不保证理解否定、指令替换或第三方话语。 +- +-`stable_latest_sentence` 和 `rolling_transcript` 只供显式历史复现/消融使用,不能混称新默认。已保存的旧权重、结果和缓存不改写。GenMo 新缓存记录策略;当前 runner 复用缓存时必须匹配请求策略,不匹配则要求重新生成。HY-Motion 和 Kimodo 的 HFD 训练适配仍待实现,但文本入口必须遵循此默认约定。 +- +-不建议第一版新增 GT 驱动的“当前任务选择器”,也不把所有 PCM 统一变成 Qwen hidden 后接新头。需要意图/说话人解析时,另立适配系统实验,单独报告其训练数据、错误和延迟。 +- +-## 6. GenMo / GEM:原生结构决定如何接入 +- +-2026-10-18:ASR+过去 2s 原生 PCM 分支已实现并完成小样本匹配实验,当前未优于 ASR-only;实际因果边界、维度、BN 诊断和结果见 [GENMO_PAST_AUDIO_TEST.md](../GENMO_PAST_AUDIO_TEST.md)。 +- +-### 6.1 已核实的能力 +- +-本次针对官方 **GEM-SMPL**,不是视频生成模型 Mochi。原生 `gem_smpl` 配方有 regression/diffusion 双模式,16 层、1024 隐藏宽度的 RoPE Transformer,文本通过 T5-3B 的 1024D token 特征做 cross attention。检查点和实际配置优先于论文中的概念表示。 +- +-公开编码器的动作输入/预测核心是 151D:身体 21×6D 旋转、10D betas、camera-view 根旋转 6D、gravity-view 根旋转 6D、根局部位移 3D。末三维虽叫 velocity,代码计算的是相邻帧位移,**没有乘 FPS**;不能直接套米/秒速度。见 [原生 encoder](https://github.com/NVlabs/GENMO/blob/16bebf402d8893184249ee206d957b8248cd8310/gem/network/endecoder.py#L122)、[位移计算](https://github.com/NVlabs/GENMO/blob/16bebf402d8893184249ee206d957b8248cd8310/gem/utils/motion_utils.py#L25)。 +- +-PCM 分支是多层卷积,输出逐帧 128D 特征,再投影进动作条件;BEAT2 loader 把音频重采样至 **18kHz**、动作 30Hz。它是逐帧对齐语音条件,不能把 HFD 16kHz 波形原样塞入并假定时长不变,也不能把 music 的 35D 特征当语音语义。见 [音频编码](https://github.com/NVlabs/GENMO/blob/16bebf402d8893184249ee206d957b8248cd8310/gem/gem.py#L175)、[BEAT2 loader](https://github.com/NVlabs/GENMO/blob/16bebf402d8893184249ee206d957b8248cd8310/gem/datasets/beat2/beat2.py)。 +- +-denoiser 有 `observed_motion_3d`、`motion_mask_3d`,可把已知动作写入 noisy motion。但这不等于已具备 HFD 的 KV/RTC。默认 `dual_mode` 的 `in_attr` 没有启用 observed-motion embedder;代码有相关分支引用,也不能仅凭该引用就认定实际 checkpoint 含对应参数。见 [overwrite 接口](https://github.com/NVlabs/GENMO/blob/16bebf402d8893184249ee206d957b8248cd8310/gem/network/gem_denoiser.py#L214)、[默认条件配置](https://github.com/NVlabs/GENMO/blob/16bebf402d8893184249ee206d957b8248cd8310/configs/pipeline/dual_mode.yaml)。 +- +-### 6.2 推荐训练构造 +- +-主实验命名 `GEM-ASR-prefix`。30Hz 下建议起始设定为 12 帧已知历史+60 帧未来;12 帧包含当前状态,所有 future 条件只使用当前决策的文本。未使用的视觉、2D、camera-motion、music、speech 条件必须通过原生 availability mask 标缺失,不从 GT 人体投影出“视觉输入”。保留固定虚拟相机/重力坐标以完成表示转换,不根据未来 GT 轨迹设置相机。 +- +-历史经原生 encoder 转成 151D,送入现成 overwrite 接口;未知未来 observed 值清零、mask=0。首版不为缺失的 observed-motion embedder 虚构权重,也不自动扩展 `in_attr`。先严格加载原始 checkpoint,再检查前缀分支和梯度;如必须增加 mask embedding,明确命名为额外结构扩展。 +- +-保留原生训练入口和双模式执行关系。其 `forward_train(mode='diffusion')` 要求 `regression_outputs`,不能直接删除 regression 调用而声称仍用原训练流程。HFD 有完整 3D 监督,无需伪造 2D-only 样本或启用无来源的相机重投影监督。采用原生噪声过程与 3D loss;将已知前缀/补齐帧从未来预测指标中排除,并审计每个损失的有效 mask,避免已知帧复制降低表观 loss。见 [训练路径](https://github.com/NVlabs/GENMO/blob/16bebf402d8893184249ee206d957b8248cd8310/gem/network/gem_diffusion.py#L73)。 +- +-### 6.3 语音增强实验 +- +-命名 `GEM-ASR-past-audio`,与主实验分开。只截过去 2 秒已释放 PCM,再重采样至 18kHz;编码前就截断,防止卷积跨边界。将过去音频的逐帧特征放在窗口的历史位置,未知未来位置设 availability=0,不能把未来动作对应的真实音频给模型。 +- +-如果仅保留 0.4 秒动作历史而希望使用整段 2 秒音频,现有逐帧对齐接口需要额外设计。第一版增强实验建议使用 **60 帧过去上下文+60 帧未来**,已知动作范围同步扩为两秒;另做相同两秒动作历史但关闭音频的对照。否则不能把收益单独归因音频。音频有效长度、卷积输出帧数、填充边界必须用真实 forward 验证,不能假设输出恰好逐点对齐。 +- +-## 7. HY-Motion:文本条件可以复用,流式历史需要扩展 +- +-### 7.1 原生能力与公开程度 +- +-原生动作表示是 SMPL-H 身体 22 关节的 **201D**:根平移 3、根旋转 6、21 个局部旋转 126、22 个局部位置 66。它不等于 HumanML3D 263D。网络结合双流/单流 Transformer,语义条件为 Qwen3-8B token 特征和 CLIP-L 全局向量,flow matching 预测速度。论文中的局部 attention 是双向局部性,不是在线语义记忆。见 [论文模型章节](https://arxiv.org/html/2512.23464v1#S3)、[文本编码器](https://github.com/Tencent-Hunyuan/HY-Motion-1.0/blob/4e426f5a1021cbcf7f375458c37b840ee7225229/hymotion/network/text_encoders/text_encoder.py)。 +- +-核查的 `generate()` 接口接收文本、seed、duration;`hidden_state_dict` 是预编码文本特征,并非历史动作状态。`x_mask_temporal` 表示有效序列位置,也不是 known-history mask。源码有 `mask_mode='causal'` 分支同样不等于已训练的流式预测器。公开仓库以推理为主,本次未发现可直接套用的完整官方训练循环。见 [生成接口](https://github.com/Tencent-Hunyuan/HY-Motion-1.0/blob/4e426f5a1021cbcf7f375458c37b840ee7225229/hymotion/pipeline/motion_diffusion.py#L477)、[forward](https://github.com/Tencent-Hunyuan/HY-Motion-1.0/blob/4e426f5a1021cbcf7f375458c37b840ee7225229/hymotion/network/hymotion_mmdit.py#L409)。 +- +-### 7.2 两种结果必须分开命名 +- +-`HY-native-window`:每次使用当前 ASR 文本、固定两秒计算窗口做文本到动作。保持原生网络,用于判断文本/动作表示是否接通;它没有历史姿态条件,不能冒充连续全双工 baseline,也不能靠 GT 对齐每个预测片段掩盖跳变。 +- +-`HY-ASR-prefix`:推荐的流式适配。采用 12 帧历史+60 帧未来、30Hz;增加显式 known-mask 和历史条件通路,例如零初始化的 mask 投影,保留原有 201D 输入投影权重。未来真值只用于 flow target。新增参数及加载差异必须写入 manifest;明确这是我们实现的扩展。 +- +-一种可实现的训练约定是:已知前缀始终为干净历史;未知部分按 `x_s=(1-s)ε+s*x_GT` 插值;网络得到 known mask 和当前 ASR,预测未知部分的 `x_GT-ε`,只在未知有效帧上计算 flow loss。采样每个 ODE 子步锁定历史,未知位置从噪声积分,结束后只执行未来段。训练/采样必须采用同一种前缀参数化。它改变了原始无前缀训练分布,需通过微调和闭环验证,不能只改推理器硬覆盖就宣称完成适配。 +- +-禁用 duration estimation 和 prompt rewrite,计算时域预先固定;不能用 GT 总动作长度或下一事件时间决定输出长度。保留双编码器,不把 Qwen 单独替换成 CLIP。可先冻结文本编码器、只训练 motion 模型及新增投影;显存/LoRA 策略在真实加载后决定。 +- +-原生解码默认包含旋转/平移平滑和按整段最低顶点落地。流式版若对整个“历史+预测”执行这些操作,会修改历史和世界根高度。需锁定已执行前缀、固定 session 地面/世界原点;后处理只作用于未执行计划,另外报告开启/关闭结果。利用自己预测的未来做后处理可以是因果规划,但不能回写已执行帧或利用 GT 未来。 +- +-## 8. Kimodo:优先复用原生约束机制 +- +-### 8.1 结构与版本选择 +- +-Kimodo 先预测 root,再把 root 转为局部表示供 body denoiser 使用。`motion_mask_mode='concat'` 时,已知值覆盖 noisy motion,特征 mask 与动作拼接进入网络;这是真正的约束接口。见 [两阶段 denoiser](https://github.com/nv-tlabs/kimodo/blob/e1c219c99b4b94ba48be1fb6a1cbc64ecee4c954/kimodo/model/twostage_denoiser.py#L73)。 +- +-文本默认是 LLM2Vec/Llama-3-8B 系列特征,不是 CLIP/T5;本次看到的配置可加载本地编码器或编码服务。训练实验应固定权重版本并本地缓存已释放文本的 embedding,服务不可静默换模型。 +- +-官方分别提供 SOMA、G1、SMPLX checkpoint。SOMA 内部 30 关节、公共输出可为 77 关节;SMPLX 为身体 22 关节。本批数据先对接固定的 `nvidia/Kimodo-SMPLX-RP-v1`,不使用会随版本变化的默认短名。官方说明 SMPLX 版本可能有较明显重定向伪影。见 [骨架说明](https://research.nvidia.com/labs/sil/projects/kimodo/docs/key_concepts/skeleton.html)。 +- +-源码的 AMASS→Kimodo 转换包含 pelvis offset、Y-up/Z-up 与额外朝向旋转;优先使用 `amass_arrays_to_kimodo_motion()` 并做往返检验,不能直接照搬 GEM 的坐标矩阵。源 16D betas 与 Kimodo 固定骨架体型也需核对;必要的形体标准化必须训练/验证一致,不能丢掉 betas 后声称几何完全相同。见 [官方转换](https://github.com/nv-tlabs/kimodo/blob/e1c219c99b4b94ba48be1fb6a1cbc64ecee4c954/kimodo/exports/smplx.py#L74)。 +- +-### 8.2 推荐输入与训练 +- +-命名 `Kimodo-SMPLX-ASR-prefix`。起始采用 12 帧历史+60 帧未来、30Hz;历史来自当前可知的姿态。优先按原生 full-body/末端约束格式构造历史,未来所有约束 mask=0;用 `first_heading_angle` 和可知根位置承接世界状态。不能输入未来全身关键帧、真实根路径、未来接触标签,即使原生 API 支持这些条件。 +- +-原生 multi-prompt 代码已经用上一段生成结果构造 transition constraints,说明无需另起一套条件网络;但原 UI/批处理可以看到全部 prompt,在线 wrapper 必须在每次释放后单步调用,不一次传入整条未来 prompt 时间线。原实现还会重新生成拼接重叠区,在线版不能修改已经执行的部分。见 [transition 构造](https://github.com/nv-tlabs/kimodo/blob/e1c219c99b4b94ba48be1fb6a1cbc64ecee4c954/kimodo/model/kimodo_model.py#L140)。 +- +-沿用 checkpoint 的扩散 schedule、两阶段网络和运动表示,新增 HFD 数据训练循环。论文目标是 clean-motion 分量的加权 smooth-L1 与 FK 一致性,不能用简单的统一 MSE 替换后称原生损失复现。连续历史前缀也不是原稀疏关键帧分布,需在微调中覆盖。历史姿态约束下的未约束辅助量仍可能需要损失;具体 mask 按语义/维度确定,不笼统把前缀全部当成未知目标。见 [论文训练目标](https://arxiv.org/html/2603.15546v1#S4.SS3)。 +- +-公开代码提供 denoiser、噪声过程与采样,不等于已提供完整训练器;优化器、EMA、loss 权重和约束采样需按论文/配置实现并标明复现范围。本轮不承诺两卡 batch 或训练时长。 +- +-### 8.3 平滑根与速度的未来泄漏 +- +-Kimodo 原生表示包含平滑根、heading、局部位置、全局旋转、速度和接触。SMPLX22 的该表示按源码组成是 `3+2+22*3+22*6+22*3+4=273D`,实际仍须以 checkpoint 的 `motion_rep_dim` 断言,不能套用 SOMA 的维数。见 [运动表示](https://github.com/nv-tlabs/kimodo/blob/e1c219c99b4b94ba48be1fb6a1cbc64ecee4c954/kimodo/motion_rep/reps/kimodo_motionrep.py)。 +- +-`get_smooth_root_pos()` 是整段轨迹平滑;速度是前向差分。**禁止对“GT 历史+GT 未来”整体编码后,把前 12 帧当输入。** 原生平滑甚至会影响历史段中更早的位置,单独屏蔽最后一帧速度不足以消除泄漏。见 [平滑根](https://github.com/nv-tlabs/kimodo/blob/e1c219c99b4b94ba48be1fb6a1cbc64ecee4c954/kimodo/motion_rep/smooth_root.py#L202)、[速度计算](https://github.com/nv-tlabs/kimodo/blob/e1c219c99b4b94ba48be1fb6a1cbc64ecee4c954/kimodo/motion_rep/feature_utils.py#L36)。 +- +-建议从已知前缀独立构造 full-body/end-effector constraints,而不是锁定整段 GT 编码的所有 273D 特征。若使用完整历史特征,只能在过去前缀上编码,边界速度用最近已知位移延拓;推理可以保留前次已生成计划的特征,但必须记录来源为模型预测。未来目标可以离线整体编码;输入中的平滑根与目标平滑根可能不同,需通过可观测姿态约束而非强求两者逐值相等来处理。 +- +-## 9. 新模型共同的窗口与 rollout 建议 +- +-以下是起始实验参数,不是三篇论文的原始训练配置: +- +-| 项目 | 第一版建议 | +-| --- | --- | +-| 新人体模型动作 FPS | 30Hz,严格按时间重采样 | +-| 动作历史 | 0.4s / 12 帧,包含当前状态;不足时重复首姿态并记录有效范围 | +-| 未来计划 | 2s / 60 帧;不依赖 GT 剩余动作长度 | +-| 名义条件更新/执行 | 0.4s / 12 帧后可重规划;每次仅用当时已释放条件 | +-| 语义条件 | 五者共享 `causal_asr_context`;旧稳定句仅显式历史对照 | +-| 训练覆盖 | `full_session`,包含 listen、idle、保持、过渡及事件前后 | +-| 外部控制 | 第一版不添加 GT stop gate、事件识别头或手工姿态重置 | +- +-预测两秒不表示必须播放完两秒。新语义组到来后可重规划,避免 Stop 要等完整计划结束。理想零计算延迟回放和实测延迟回放分别报告;如果规划/ASR 耗时超过 0.4 秒,不能声称达到该响应频率。 +- +-训练窗口在 session 内按合法起点取样,不能跨 session 连接;未来段即使覆盖下一条指令,也只提供当前可用文本。未来事件尚不可知时,监督天然存在多解,应保留并报告这种不确定性,不以未来 ASR 消除它。可增加只监督近期执行段的消融,但它改变训练损失,不能混为原生方法。 +- +-第一阶段使用 GT 过去姿态做 teacher-forcing,未来按模型原目标训练;正式评估必须用自身实际执行结果作为下一步历史,只允许 session 初始姿态来自 GT,不能每组重新喂 GT。后续可加入 rollout/noisy-history 微调,单独标注训练配方,不把它当作已经实现的能力。 +- +-统计归一化先沿用官方 checkpoint 的对应统计以保留初始化语义;人体分布若明显失配,再设计重估和权重/输入变换的成套方案。统计只能来自 train,不能把 SMPL-X 原始轴角、G1 36D、TextOP/OMG 的统计互相套用。 +- +-## 10. 实现前后的验收与比较口径 +- +-1. **人体 session**:逐任务覆盖 1–4 轮;与 G1 版比较计划、PCM 哈希、事件边界、split。人体单独检查旋转合法、无 NaN、拼接连续、固定体型与带声音视频。当前尚未完成。 +-2. **表示往返**:已知姿态→原生特征→解码,检查左右手、根朝向、位移单位、pelvis offset、帧率与固定形体;预先记录转换误差。所有模型按共同身体 22 关节评估,手指不纳入主体指标。 +-3. **因果反事实**:固定当前 PCM/历史,仅替换释放点后的音频,当前输入必须不变;仅替换当前姿态后的 GT 动作,历史特征/文本/约束必须不变;修改未来事件元数据也不得改变当前输入。 +-4. **原生模型接通**:严格加载官方 checkpoint,核对特征维数与实际前向;检查 loss、有效 mask、有限梯度、参数更新与 checkpoint 恢复。HY 新增参数单独列出;Kimodo 新训练器说明 loss 复现范围。 +-5. **短样本闭环**:包含普通指令、Stop、backchannel、hesitation、pause、third-party 的已验证人体样例;比较 teacher-forcing 与自主 rollout,检查“听完才动”、停止延迟、被打断后保持和下一指令切换。 +-6. **正式比较**:固定 session manifest、共同数据范围、ASR 版本、初始状态和随机种子;报告源动作/事件覆盖、训练更新数及有效目标帧,不强求不同架构拥有相同 batch/epoch。固定验证集,测试集不参与调参。 +-7. **结果**:同时报告动作质量、事件响应、语义错误、边界跳变、长时漂移、ASR 稳定等待、模型规划延迟和执行缓冲。文本输入不能完全覆盖声学/说话人信息,ASR 系统与原始音频 HFD 属于不同输入管线,表格应标明。 +- +-上游预训练数据与 HFD 源动作也要核查重叠。Kimodo 当前 README 明确说明部分旧 SOMA checkpoint 存在 BONES-SEED benchmark overlap,v1.2 修正;这不自动证明 SMPLX-v1 对 HFD 测试完全独立。无可核实的预训练划分时,写明“下游 session/source 隔离,预训练重叠未排除”,不能宣称完全未见动作泛化。 +- +-本说明不改变现行 OMG/TextOP 配置。下一步需要实现的是 session 计划冻结与 human renderer,再是三个 adapter;不是直接把新源动作交给既有训练命令。 +- +-## 11. 2026-10-18:官方与社区训练脚本综合核查 +- +-本节回答“是否有可用训练脚本”,不代表已经接通 HFD 训练。核查了官方当前源码树、官方训练代码 issue、定向搜索得到的社区仓库及实际训练代码;没有穷举所有 fork。固定版本、文件哈希与小规模 probe 的记录见 `reports/human_training_scripts_20261018.json`。选取的上游文本源码快照保存在 `/root/autodl-fs/pangxinlei/downloads/hfd-training-script-research-20261018/`,未替换三个原有模型仓库。 +- +-### 11.1 结论与选择 +- +-| 模型/候选 | 实际公开内容 | 能否用于当前 HFD 逐帧监督微调 | 验证程度 | +-| --- | --- | --- | --- | +-| GenMo/GEM 官方 | Lightning/Hydra 训练器、数据模块、双模式 diffusion/regression、损失、优化器与 checkpoint | **优先复用官方框架**,补 HFD 数据/条件 bridge 和专用配置 | 源码核对;现有环境及小型原生网络反向已通过;未跑真实 HFD 官方训练步 | +-| HY-Motion 官方 | 201D 模型、CLIP/Qwen 文本编码、采样、评测 | **没有官方训练入口**,监督 flow-matching 循环仍需补 | 官方源码与作者 issue 回复核对 | +-| HY-Motion:MotionRFT | 对 Full/Lite 的 LoRA、可微采样、奖励模型与微调循环 | **可参考训练组件,但不是逐帧监督训练的直接替代品** | 静态核对训练/梯度代码;未运行其完整奖励训练 | +-| Kimodo:Calix-L 后训练项目 | 特征数据集、文本缓存、训练步、优化器、梯度累积、checkpoint、七项结构化损失 | **适合作为轻量监督训练起点**;默认 SOMA 配置和离线约束采样必须替换 | 七项损失及其原样 train_step 已对 tiny 官方 SMPLX22 两阶段网络做反向 probe | +-| Kimodo:windydauy 的 scq 分支 | 完整 G1 微调目录、DDP/AMP、EMA、AdamAtan2、七项损失;另有蒸馏工作流 | **可复用训练设施**,G1 数据适配与损失须重审;不直接采用改变层数的蒸馏版本 | 损失模块在当前 SMPLX22 表示上反向通过;完整训练循环仅静态核对 | +- +-“可复用”指可以基于现有代码补齐当前任务,**不是修改数据路径即可开训**。共同缺口仍是第 4 节的冻结 SessionPlan 与 human renderer,以及因果 ASR/姿态历史、闭环推理和第 10 节验收。 +- +-### 11.2 GenMo/GEM:官方训练框架最完整 +- +-来源:[NVlabs/GENMO](https://github.com/NVlabs/GENMO),本地与远端核对版本均为 `16bebf402d8893184249ee206d957b8248cd8310`。这里是 NVIDIA 人体动作项目,不能误用同名 Genmo/Mochi 视频生成训练器,也不把 GEM-X 的另一套网络混入当前 baseline。 +- +-主要入口: +- +-- `scripts/train.py`:实例化模型和 datamodule,调用 Lightning `Trainer.fit`,支持断点恢复、日志和验证。 +-- `gem/gem.py`:`training_step`、`configure_optimizers`,原生 regression/diffusion 的执行次序。 +-- `gem/pipeline/gem_pipeline.py`、`gem/network/gem_diffusion.py`:扩散训练及几何损失。 +-- `configs/exp/gem_smpl.yaml`:包含文本/音频生成的完整模型配方;`gem_smpl_regression` 是回归配方,不能拿它代替文生动作训练。 +- +-官方命令 `python scripts/train.py exp=gem_smpl` 是框架入口示例;原配方会读取多套原始训练/评测集,并非 HFD 命令。默认 exp 引用缺失的 mixed,需显式指定 exp。模型构造还依赖主体模型、归一化与辅助资产,CLI --help 成功不证明这些运行资产已完整准备。 +- +-接 HFD 时应保留官方模型、噪声过程和合适的 3D 损失,接入 HFD datamodule 与原生条件 availability/overwrite。不要给模型伪造未来视觉观测;世界/重力/虚拟相机变换、body shape、语音特征与文本缓存仍须按第 6 节核对。diffusion 路径依赖 regression 输出,首轮不直接裁掉 regression 调用。输入历史的速度边界和未来 mask 需通过反事实检查。 +- +-### 11.3 HY-Motion:已有社区训练,但主要是奖励微调 +- +-官方 [issue #2](https://github.com/Tencent-Hunyuan/HY-Motion-1.0/issues/2) 中,协作者于 2025-12-31 明确回复没有开源该项目训练代码的计划。当前官方版本 `4e426f5a1021cbcf7f375458c37b840ee7225229` 的源码树也未找到训练入口。README 的预训练→高质量微调→RL 是方法说明,不能当成已发布训练器。网上部分教程给出 `train.py`/`configs/finetune.yaml` 却没有对应仓库文件,不作为可执行证据。 +- +-找到的主要可用社区实现是 [Xiaofeng-Tan/MotionRFT](https://github.com/Xiaofeng-Tan/MotionRFT),固定版本 `84543efec6bf8c0d819bfa75665f5c820c35daae`: +- +-- `RFT_HY/motionrft_hy.py` 包含 `LoRALayer`、`apply_lora_to_model`、`DifferentiableFlowMatchingSamplerV2`、训练器和 checkpoint;支持 Full/Lite 配置。 +-- 它先无梯度采样到随机时间,再通过一个可微末步产生动作,使用 `loss = -combined_reward` 更新 LoRA;不是标准配对动作的 flow-matching 速度场监督。 +-- 模型采样仍使用原生 **201D**。奖励/匹配模型使用的 **135D** 是取根平移与旋转部分的另一条数据通路,不可把 Mean_135/Std_135 或 HumanML263 统计替代 HY 的 201D 模型统计。 +-- 依赖 HumanML3D 风格数据、SPM/Sentence-T5 与可选 MotionReward checkpoint。代码即使选择 MotionReward 也会构造 SPM;不能假定仅准备 HY 权重即可启动。改为 HFD ASR/全 session 还需修改数据、目标和时间条件。 +- +-建议把它作为 LoRA 注入、可微网络调用和保存恢复的参考;第一版 HFD 过拟合仍采用原生 201D 与配对 GT 的监督 flow-matching 目标,不额外引入奖励模型作为第一阶段前提。若采用 MotionRFT 的奖励目标,应独立标为后训练实验,并评估事件响应是否被全片段奖励掩盖。原生 HY 接口没有经过训练的姿态历史约束,MotionRFT 也没有直接补齐我们的因果前缀接口。 +- +-另有 [FlowHMR](https://github.com/flowhmr/flowhmr/blob/main/docs/training.md) 复用 HY MMDiT backbone,提供 `train_v2m.py` 的监督训练与 GRPO;但它是视频到动作任务,数据要求相机、bbox 和 SAM 3D Body 特征,不能将其训练器/权重原样称为 HY 文本动作微调。另查了 5 个 HY mirror/streaming 候选,未发现额外可直接用的监督训练器;具体结果见快照 `hy-other-candidates.json`。 +- +-### 11.4 Kimodo:社区监督训练确实可以复用,但配方有差别 +- +-官方当前版本 `e1c219c99b4b94ba48be1fb6a1cbc64ecee4c954` 仍主要提供推理、约束与评测。[训练代码请求 #25](https://github.com/nv-tlabs/kimodo/issues/25) 的社区回复只描述尝试微调,没有给出官方训练发行版。 +- +-**候选 A:Calix-L/Kimodo-SOMA-RP-final-showcase**,固定版本 `2ece5366f8cee2a54999763f372c19bf3e20f5e2`。 +- +-- [train/train.py](https://github.com/Calix-L/Kimodo-SOMA-RP-final-showcase/blob/main/train/train.py) 包含完整的加噪→原生 denoiser→loss→backward→optimizer→保存循环,不只是展示网页。 +-- `train/kimodo_seed_repro/{data,kimodo_io,train_utils}.py` 提供特征读取、缓存文本、加载与恢复、AdamW/AdamAtan2、masked MSE/结构化 Smooth-L1;还有 `scripts/train_with_val.py`。 +-- 2026-10-18 本地实测:使用当前 Kimodo **SMPLX22 / 273D 官方统计**、tiny 随机两阶段 denoiser、synthetic cached text,原样调用其 `train_step`,打开关键帧条件与七项结构化损失;root 和 body 两分支均得到有限、非零梯度。另单独验证其 loss 反向。 +-- 该 probe 只验证接口和基本梯度流,不加载 full-size 预训练权重,不使用真实 HFD,不验证收敛、混合精度、checkpoint 恢复或官方论文复现。 +- +-必须修改/复核的点:现有 YAML 默认 SOMA、150 帧离线裁剪;关键帧可从完整 GT 有效区间随机抽取,并包含整体编码的 smooth root,不能直接用于 HFD 因果历史。`first_heading_angle` 固定零需与规范化策略一致。缓存文本路径下,`text_dropout` 仅修改文本字符串,已有缓存向量不会因此置空,使用缓存 CFG 时必须单独修正。基本 train.py 未见 EMA 更新;若声称沿用论文配方需要补齐。FK loss 当前比较 `FK(pred)` 与 `FK(target)`,没有直接使用目标的 position feature;应与论文的 FK→目标位置定义核对,不能只因七个权重相同就称一致。 +- +-**候选 B:[windydauy/kimodo_my,scq 分支](https://github.com/windydauy/kimodo_my/tree/scq/kimodo/training)**,固定版本 `59e12423f4a313aba8339f065bba7b326e7e7db2`。 +- +-其 `kimodo/training/train.py`、`dataset.py`、`loss.py`、`optimizers.py`、`ema.py` 已形成训练工作流,包含 DDP/AMP、断点、CFG dropout、EMA、AdamAtan2 和关键帧课程;比 Calix 的最小循环具备更多训练设施。七项 loss 在当前 SMPLX22 表示上的 CPU 合成反向 probe 通过,但完整循环还未执行。 +- +-它默认是 G1 CSV/机器人配置,不能直接读取 SMPL-X;七项 loss 实现使用 **普通 L1**,与当前约定的 Smooth-L1 不完全一致。默认 loss 权重与 YAML 指定权重也不同,运行必须显式锁定。其 `distillation` 目录另有 16→8 层、100→20 步等学生训练,改变了架构/目标;基础 human baseline 不直接采用。对应官方 [PR #44](https://github.com/nv-tlabs/kimodo/pull/44) 已关闭且未合入,不能标注官方训练代码。 +- +-其他容易误判的线索:`LavaTap/qwen-kimodo-train-fork` 的 `train_text_projection.py` 训练的是 Qwen→LLM2Vec 投影,不是动作模型。Blender/ComfyUI 接入器主要是推理或 retarget。近期 [Controllable-Motion-Exaggeration](https://github.com/AHHHZ975/Controllable-Motion-Exaggeration) 论文有训练适配,但当前仓库仅 README、标记 coming soon,暂没有可拿来运行的训练器。 +- +-### 11.5 对当前任务的执行建议 +- +-1. **GenMo**:保留官方 Lightning/Hydra,写 HFD bridge 与单独训练配方;不另造通用训练框架。 +-2. **Kimodo**:按用户既定的优先顺序,先用 Calix 的已验证训练步作为参考,沿用官方两阶段网络/扩散/统计,吸收另一项目的 EMA/DDP 设施;先统一损失、缓存 CFG、初始朝向与仅过去约束。所有复用的第三方代码保留来源与许可,不覆盖官方原文件。 +-3. **HY-Motion**:复用官方网络和编码器,参考 MotionRFT 的可训练调用/LoRA,补最小监督 flow-matching 循环;不要把奖励微调当过拟合测试,不能声称用了不存在的官方训练框架。 +-4. 三者接同一个冻结的 HFD session plan,分别编码为原生表示;训练和评测独立保留 teacher-forced / predicted-history。评测统一走已有 SMPL-X→G1 链路和 human GT retarget 对照。 +- +-本轮只完成调研、源码快照与上述合成 probe,没有启动正式训练或修改三者的实际训练配方。 +diff --git a/docs/history/HFD_WINDOW_PROTOCOL.md b/docs/history/HFD_WINDOW_PROTOCOL.md +deleted file mode 100644 +index 61916a7..0000000 +--- a/docs/history/HFD_WINDOW_PROTOCOL.md ++++ /dev/null +@@ -1,68 +0,0 @@ +-> 历史实验记录;当前入口见 [README](../../README.md)。旧运行目录已移动到 `/root/autodl-tmp/hfd-runs-archive-20261007/`。 +- +-# HFD 当前音频窗口口径与 TextOP 连续推理排查 +- +-旧 `omg_duplex_session_playable.mp4` 使用此前确认的 `offline_target_aligned`:动作目标未来 2 秒与这 2 秒对应音频一起输入模型。它可用于离线训练链路检查,不能据此宣称实时全双工能力;连续关节 RMSE 为 0.10088 rad,并非完美拟合。旧结果保留为离线对照。 +- +-## 新 OMG 输入时序 +- +-`data.audio_policy: hfd_current_window`。HFD 原始动作时钟为 50 Hz,每 20 个动作点使用一个 0.6 秒音频窗口;窗口每 0.4 秒更新: +- +-- 原始动作帧 `u` 的语义组 `k = u // 20`。 +-- 当前 PCM 范围 `[0.4*k, 0.4*k+0.6)` 秒,16 kHz 时为 `[6400*k, 6400*k+9600)`。 +-- 音频可用截止为 `0.6+0.4*k`。这沿用 HFD 释放时钟,与运动/音频内容偏移 0.2 秒不同,也不表示零延迟在线系统。 +-- OMG 保留原生 30 Hz 动作格式;状态帧 `t` 对应组 `k=floor(t/12)`。训练在每个可用 30 Hz 动作点取样,同组复用同一窗口。 +- +-OMG 当前音频接口需要每个预测位置一个 35D 声学特征,不接收原始 PCM,也没有 Qwen 的语言语义 KV 缓存。转换器只对当前 9600 个样本调用原生 `current35` 提取器,再将得到的特征序列线性插值到原生 60 个条件位置。**这 60 个位置都编码同一个当前窗口,不能解释为读取了未来 2 秒音频。** 包括峰值归一化在内的全部特征运算仅访问该窗口。没有额外的语义缓存前向,原生声学 embedding 和扩散采样仍然执行。 +- +-训练仍用原生 Lightning、125D 动作 codec、60 帧未来目标和 10 帧状态历史。未来动作是真值监督,未来音频不是输入。修改当前可用窗口以外的任何 PCM 不会改变输入;测试与数据合同检查覆盖此约束。 +- +-## 连续推理 +- +-只在开头使用一次真值历史。每轮用当前 0.6 秒音频窗口预测原生 60 帧(2 秒),执行前 12 帧(0.4 秒),再用已执行的预测姿态更新历史,读取新窗口并重新规划。未执行的预测丢弃,不能拿未执行的未来预测初始化下一轮。 +- +-不能把一整段 session 的各个未来音频窗口一次性传给 `generate(num_frames=整段)`:这会让当前规划提前接触之后才可用的条件。新推理代码逐轮调用原生生成器,并在报告中保存每轮 PCM 起止位置、状态帧和执行帧数。 +- +-这是沿用 HFD 音频释放规则的 OMG 适配,保留 OMG 原生预测长度;不是宣称 OMG 拥有 HFD 的语义记忆。当前窗口可能只包含长指令的一部分,缺少 KV 历史确实可能降低表现。音频在正式训练中也保留中文和背景音,不根据真值指令/任务边界提取音频。 +- +-## TextOP 为什么几乎不动 +- +-旧小样本配置 `stages: [1600]` 只进入原生 stage 0。虽然 `use_rollout: true`,`DARManager.should_rollout()` 在 stage 0 始终返回 False,所以训练全程使用每个 primitive 的真实历史。整段视频则只在开头给一次真实历史,随后反馈模型预测,训练和这种验证存在历史分布差异。 +- +-旧 TextOP 训练窗口关节 RMSE 为 0.00516 rad,整段连续 RMSE 为 0.32428 rad。两者不矛盾:从真实动作姿态继续生成 8 帧比从静止姿态独立启动并持续任务容易。文本在 listen/执行/收尾阶段也可能保持相同,而 TextOP 没有额外的真值控制标签或任务进度输入,容易停留在接近静止的历史状态。这个现象不能仅凭一条 loss 断定是映射问题;映射沿用已独立验证的关节名称转换。 +- +-本次保留原生训练循环与损失,用 `stages: [400,800,400]` 进入原生反馈预测历史的训练阶段。原生正式训练模板本来就有三阶段;此次是修正小样本实验只运行首阶段的局限。新配置为 `configs/history/textop_duplex_causal.yaml`,新 OMG 配置为 `configs/history/omg_duplex_causal.yaml`。两者均从官方初始化重训,没有复用旧离线 OMG 权重。 +- +-## 复现 +- +-已有原生 `train-002203-seed42` 与完整因果 ASR 后,先更新配置中的环境、session、权重和输出路径,再执行: +- +-```bash +-/root/autodl-tmp/envs/textop/bin/python train.py --config configs/history/textop_duplex_causal.yaml +-/root/autodl-tmp/envs/omg/bin/python train.py --config configs/history/omg_duplex_causal.yaml +-/root/autodl-tmp/envs/textop/bin/python infer_overfit.py --run /root/autodl-tmp/runs/hfd-duplex-causal/textop --continuous-only --task-seconds 0 14.8 +-/root/autodl-tmp/envs/omg/bin/python infer_overfit.py --run /root/autodl-tmp/runs/hfd-duplex-causal/omg --continuous-only --task-seconds 0 14.8 +-``` +- +-用原有 `render_overfit.py` 渲染两目录的 inference,默认 H.264/MP3 MP4。本页对应的旧 OMG 实验使用 `hfd_current_window`;后续正式模板已统一为 `hfd_past_2s`,详见 [OMG_CONTEXT_DEBUG.md](OMG_CONTEXT_DEBUG.md);旧离线 smoke/duplex 配置保留用于复现旧结果。 +- +-## 推理设置核验与本次结果 +- +-进一步对照原生 TextOP 可视化发现,旧导出使用普通条件采样(相当于 CFG guidance=1),原生 `robotmdar/config/vis_dar.yaml` 默认 guidance=5。当前 `infer_overfit.py` 已使用原生 `ClassifierFreeWrapper` 并默认 guidance=5;可用配置 `inference.textop_guidance_scale` 或 CLI `--textop-guidance-scale` 指定。旧视频/诊断的 guidance=1 设置明确保留为对照,不与新设置的误差直接混用。 +- +-同一静止历史、同一噪声种子,替换实际已释放的两条 ASR 条件:guidance=1 的输出差异约 0.00014 rad;guidance=5 约 0.00069 rad。guidance=5 下,保持文本不变而换成真实动作历史时,输出差异约 0.22 / 0.47 rad,真实历史窗口的误差约 0.00575 / 0.01026 rad。说明此次小样本拟合主要依赖动作历史续接,文本对从静止姿态启动的作用仍很弱。 +- +-三阶段加原生 guidance=5 后,TextOP 整段仍未拟合,不能将补齐这些设置描述为已解决几乎不动。旧单阶段配置是实验局限,guidance=1 是与原生默认可视化不同的设置;二者都不是静止现象的充分解释。诊断只针对本次两个任务和小样本预算,不断言整个 TextOP 架构没有能力。 +- +-OMG 当前窗口训练后窗口 RMSE 为 0.12468 rad,整段预测为 0.30954 rad;旧离线整段为 0.10088 rad。新旧同时改变音频范围、窗口取样密度与重新规划方式,这一对比不能单独量化未来音频的因果贡献。新口径不会向当前预测提供之后才释放的 PCM,而缺少语义缓存仍可能限制长指令处理。 +- +-两个 1600 步原生训练链路、严格权重重载与两个数据合同检查通过,27 项回归测试通过。新视频均保留原始音轨,音频和全部视频帧完整解码通过。OMG 推理每轮的窗口边界和执行历史记录在 `hfd_window_summary.json`。新 OMG run 的三份等价重复检查点已经清理,释放约 2.96 GiB,最终 `last.ckpt` 保留。 +- +-新视频入口: +- +-- `/root/autodl-tmp/runs/hfd-overfit-videos-jointfixed/textop_duplex_rollout_playable.mp4`:三阶段、原生 guidance=5、已释放 ASR。 +-- `/root/autodl-tmp/runs/hfd-overfit-videos-jointfixed/omg_duplex_hfd_window_playable.mp4`:当前 0.6 秒音频、0.4 秒重新规划。 +- +-诊断可复现:`python diagnose_textop.py --run RUN --frames 242 490 --guidance-scale 5`。这只比较同种子下原生 8 帧生成结果,不使用真值任务文本;真值姿态仅用于明确标注的历史对照。 +- +-最终原生 guidance=5 的 TextOP 训练窗口 RMSE 为 0.00687 rad,整段为 0.40908 rad(root 0.03739 m)。这再次表明短窗口拟合不能证明整段任务能力;新 OMG 整段为 0.30954 rad(root 0.06111 m)。两模型的关节维度不同,仍不能据此排名。 +diff --git a/docs/history/NORMAL_TASK.md b/docs/history/NORMAL_TASK.md +deleted file mode 100644 +index 90a97d6..0000000 +--- a/docs/history/NORMAL_TASK.md ++++ /dev/null +@@ -1,73 +0,0 @@ +-> 历史实验记录;当前入口见 [README](../../README.md)。旧运行目录已移动到 `/root/autodl-tmp/hfd-runs-archive-20261007/`。 +- +-# 正常单任务的小样本检查 +- +-推荐观看当前视频目录中的 `textop_complete_task_playable.mp4` 和 `omg_complete_task_playable.mp4`。这是连续播放的完整单任务,不是 `*_fitted_windows_playable.mp4` 的窗口拼接。 +- +-## 样本与任务 +- +-原始样本为 `sessions/train-000094-seed42.npz`,由 HFD 原生 session 构造器生成。元数据确认 `interaction_mode=turn-taking`、`turn_count=1`,没有打断、停止或后续任务。动作名是 `grab_walk_ff_180_002__A552_M`,即向前走并作抓取动作。英文女声为 `female_02_en`,ASR 识别为 “Character walks forward while grabbing something.”。没有额外场景物体或抓取物理模拟,播放的是机器人参考动作。 +- +-完整 session 长 17.6 秒: +- +-| 动作时间 | 内容 | +-| --- | --- | +-| 0.00–1.20 s | 初始待机;指令开始 | +-| 1.20–4.40 s | 原生听指令阶段与过渡 | +-| 4.40–16.00 s | 连续向前走并抓取,共 11.6 秒 | +-| 16.00–17.60 s | 动作结束后的待机 | +- +-原始麦克风语音约在 0.60–3.32 秒;与动作时钟相差 0.2 秒,按适配器既有时钟关系播放。数据中的听指令姿态也保留,但两个模型不接收 `listen/motion/idle` 控制标签。初始真实历史用于模型启动并在画面中标注;之后自由反馈预测。初始化画面不计入连续推理误差。 +- +-## 训练与导出 +- +-两模型从各自官方初始化权重重新训练 800 步,使用原生 DAR / Lightning 框架。配置选择整个 session 的 `[0,870)` 原始 50 Hz 帧,并保留这一连续区间内所有合法窗口,`max_windows=null`;不再均匀抽八个分散窗口。实际视频结尾根据模型帧网格取整到约 17.3–17.4 秒,只省略末尾少量待机。 +- +-TextOP 使用这个 session 新生成的逐释放窗口 ASR;OMG 使用完整离线目标段音频。音频、原始轨迹和关节映射保持不变,没有平滑、改速度、补帧或拼接其他任务来改善观感。 +- +-`textop-single/` 和 `omg-single/` 保存有效检查点、配置、指标与推理结果;`sessions/` 保存原始单轮任务及选择依据;`asr/` 保存新生成的完整转写;`summary.json` 保存本次训练和推理指标。视频和任务说明在 `/root/autodl-tmp/runs/hfd-overfit-videos-jointfixed`。 +- +-窗口拟合和自由连续推理分别报告。800 步只是训练链路的小样本检查,不保证完整任务的自由推理已拟合;模型偏离真值时不能当作视频拼接问题。右侧真值用来检查正常任务的实际动作连续性。 +- +-## 之前的视频为什么跳跃 +- +-旧 `textop_fitted_windows_playable.mp4` 拼接八段,每段仅 0.64 秒,总长 5.12 秒。八段中约 82% 的帧标为 motion、约 18% 为 idle,没有 listen 帧。但采样覆盖很偏:第一段是初始待机,其余几乎来自另一条长 session,没有覆盖原先向前走抓取任务的完整执行。训练窗口从不同时间点直接连接,必然出现时间与姿态跳变,不能当作完整任务视频。 +- +-新样本选择、训练和导出可以复现: +- +-```bash +-# 所有 output 必须指向未使用的新目录;选择仍使用原生构造,仅跳过小样本不适用的大规模最终配比。 +-python prepare_normal_task.py --output /data/normal-task/sessions --motion grab_walk_ff --limit-sources 128 +-# 转写选中的 NPZ,并按新路径修改 configs/*_normal_task.yaml。 +-python transcribe_sessions.py /data/normal-task/sessions/sample.npz --device cuda:0 --output-dir /data/normal-task/asr +-python train.py --config configs/history/textop_normal_task.yaml +-python train.py --config configs/history/omg_normal_task.yaml +-python infer_overfit.py --run /data/normal-task/textop --continuous-only --task-seconds 0 17.4 --textop-guidance-scale 1 +-python infer_overfit.py --run /data/normal-task/omg --continuous-only --task-seconds 0 17.4 +-python render_overfit.py --inference /data/normal-task/textop/inference --assets-root /data/OMG/assets --gpu 0 +-python render_overfit.py --inference /data/normal-task/omg/inference --assets-root /data/OMG/assets --gpu 1 +-``` +- +-`data.frame_range` 是原始 50 Hz 的 `[start,end)` 采样边界,仅允许 smoke 模式、单个 session,用于明确小样本检查范围;正式训练仍走完整数据配置。相机跟随使用机器人身体尺寸计算距离,避免走得远时机器人越画越小;这只影响视角,不改变原始世界坐标和误差。 +- +-本次 800 步结果(rad / m): +- +-| 模型 | 训练窗口关节 RMSE:初始 → 训练后 | 完整自由推理关节 RMSE | 完整自由推理根位置 RMSE | +-| --- | --- | --- | --- | +-| TEXTOP | 0.0448 → 0.0089 | 0.3164 | 2.5040 | +-| OMG | 0.2926 → 0.2252 | 0.2522 | 0.8357 | +- +-两模型训练检查均通过,但这些连续误差说明完整任务自由推理尚未充分拟合,不能作为已成功复现任务的结论。 +- +-## 当前视频格式与 VS Code 播放 +- +-当前交付统一为 MP4 容器、H.264 视频、48 kHz 双声道 MP3 音轨;文件名以 `_playable.mp4` 结尾。根据官方文档选择 H.264、MP4 与 MP3,避开不支持的 AAC。此前 VP8 WebM 在用户当前预览中仍无法打开,因此已清理。 +- +-官方支持清单:https://github.com/microsoft/vscode/issues/180780 +-官方 Webview 说明:https://code.visualstudio.com/api/extension-guides/webview#supported-media-formats +- +-所有视频检查过完整声画解码、非静音音轨、帧数和帧率;H.264 码流与原始视频的 SHA256 一致,转换只更换音频编码和封装,没有改动作。检查记录见视频目录的 `playable_mp4_check.json`。远程服务器上的解码检查不等于本地 VS Code 实测,首次打开请使用新文件名。 +- +-后续 `render_overfit.py` 与 `remux_video_audio.py` 默认导出 H.264/MP3 MP4,无需 WebM 或额外的 `--vscode-preview` 参数。 +- +-本页旧 TextOP 数值和视频使用 guidance=1。当前默认推理已补齐原生 guidance=5;复现旧结果须显式指定 `--textop-guidance-scale 1`。当前窗口输入与新检查见 [HFD_WINDOW_PROTOCOL.md](HFD_WINDOW_PROTOCOL.md)。 +diff --git a/docs/history/OMG_CONTEXT_DEBUG.md b/docs/history/OMG_CONTEXT_DEBUG.md +deleted file mode 100644 +index 92bcc43..0000000 +--- a/docs/history/OMG_CONTEXT_DEBUG.md ++++ /dev/null +@@ -1,79 +0,0 @@ +-> 历史实验记录;当前入口见 [README](../../README.md)。旧运行目录已移动到 `/root/autodl-tmp/hfd-runs-archive-20261007/`。 +- +-# OMG 两秒历史上下文与动作启动调试 +- +-样本仍是原生 `train-002203-seed42`:接取物品尚未完成时,输入捡起地上物品的新指令。完整原生音轨、中文/英文与背景音处理策略没有改变。此样本的两条录音在视频时间 0.4–2.16 s、5.2–6.96 s,视频到 14.8 s。最后一句话结束后已有约 7.8 s,所以不能直接归因于视频结束太早。 +- +-## 输入策略 +- +-新策略 `data.audio_policy: hfd_past_2s`:以 HFD 当前组的释放截止 `hi=9600+6400*k` 为准,读取 `[max(0,hi-32000),hi)` 原生 PCM。不足两秒时左侧补零。原生 OMG 仍保持 30 Hz、125D 动作、10 帧历史、60 帧预测与 `current35` 音频特征。对完整两秒 PCM 提取原生逐帧特征,取前 60 行,不做旧 0.6 秒到 60 位置的插值。 +- +-训练和推理都采用同一策略。没有额外 KV 预填充,每次重新编码已听过的最近两秒;未来音频不输入。两秒是有限历史上下文,不保证所有长指令都完整保留,也不是完整的 ASR/语言理解编码器。开头补零和窗口内峰值归一化在训练/推理中一致。 +- +-新增测试检查:截止时刻保持 HFD 时钟、左侧补零长度、原生特征顺序、修改未释放音频不影响当前特征,以及之前的录音确实能影响两秒条件。数据合同检查覆盖动作/音频因果边界。 +- +-## 对照设计 +- +-所有实验使用同一个 native session、相同初始随机种子与一次真实历史初始化。训练时动作点和未来目标保持一致;新模型从官方 50m 初始化重新训练 1600 步,没有从旧离线模型续训。 +- +-| 实验 | 训练音频 | 推理音频 | 每轮执行 | +-| --- | --- | --- | --- | +-| A | 当前 0.6 s | 当前 0.6 s | 0.4 s | +-| B | 当前 0.6 s | 最近两秒 | 0.4 s | +-| B2 | 当前 0.6 s | 最近两秒 | 2 s | +-| C | 最近两秒 | 最近两秒 | 0.4 s | +-| C2 | 最近两秒 | 最近两秒 | 2 s | +- +-B 就是用户提出的“训练 0.6 秒,推理两秒”,保留为诊断对照。B/C 的音频时间都不超出当前 HFD 释放边界。B2/C2 检查原生两秒预测能否在后半段启动;它们两秒才更新一次规划,不能与 0.4 秒重新规划混为同一个实时反应口径。 +- +-每轮报告记录两秒预测前 0.4 秒与之后 1.6 秒相对上一已执行姿态的关节变化。不能只看执行片段而忽略计划后半段:反复重规划可能把稍后的启动动作丢弃。执行后的状态才反馈到下一轮,没有使用未执行未来姿态。 +- +-## 配置与复现 +- +-新配置 `configs/history/omg_duplex_past2s.yaml` 包含原生框架、官方初始化权重、dataset、GPU、batch size、训练步数、音频策略与 `inference.omg_execute_frames: 12`。迁移时按根目录 README 准备环境和外部权重,修改 YAML 路径后运行: +- +-```bash +-python train.py --config configs/history/omg_duplex_past2s.yaml +-python infer_overfit.py --run RUN --continuous-only --task-seconds 0 14.8 +-python infer_overfit.py --run RUN --continuous-only --task-seconds 0 14.8 --omg-execute-frames 60 --output-dir RUN/inference-execute2 +-``` +- +-复现 B:将 RUN 指向已有当前 0.6 秒训练结果,再加 `--omg-audio-policy hfd_past_2s --output-dir NEW_OUTPUT`。推理报告分别记录训练和推理策略,防止把输入不匹配的结果误认为匹配训练。 +- +-`--gpu-id` 只改变推理用的物理 GPU,不改训练配置或权重。输出目录必须不存在。渲染仍使用原有 `render_overfit.py`,交付为 H.264/MP3 MP4,保留对应原生音轨。 +- +-## 本次结果 +- +-| 实验 | 整段关节 RMSE(rad) | root RMSE(m) | +-| --- | --- | --- | +-| A:0.6 s 训练 / 0.6 s 推理 / 执行 0.4 s | 0.30954 | 0.06111 | +-| B:0.6 s 训练 / 2 s 推理 / 执行 0.4 s | 0.30978 | 0.04566 | +-| B2:0.6 s 训练 / 2 s 推理 / 执行 2 s | 0.26777 | 0.04292 | +-| C:2 s 训练 / 2 s 推理 / 执行 0.4 s | 0.25026 | 0.04641 | +-| C2:2 s 训练 / 2 s 推理 / 执行 2 s | 0.18776 | 0.03093 | +- +-新两秒上下文模型的真实历史训练窗口 RMSE 从初始化 0.31303 rad 降到 0.09066 rad。两种执行长度使用相同最终权重,不是分别训练或按较好视频选择检查点。 +- +-只把旧 0.6 秒训练模型的推理输入延长到两秒,没有解决几乎不动。匹配训练与推理上下文后,第一项动作能启动;0.4 秒重规划下第二项仍较弱。以原生两秒块执行时,两段实质动作都可见,但捡取动作晚于真值,仍未充分拟合。 +- +-计划检查表明,某些静止历史下模型将明显动作安排在预测后半段。例如 B 在 source 2.4 s 的计划:前 0.4 s 相对上一姿态变化约 0.00518 rad,后 1.6 s 约 0.42096 rad。频繁重新规划会丢掉这部分计划,是此样本启动不足的一个因素。这是当前样本的诊断,不把两秒执行视为更好的实时策略;两秒执行降低指令更新频率,正式论文必须明确执行周期与反应延迟。 +- +-新视频: +- +-- `omg_duplex_past2_exec04_playable.mp4`:训练/推理统一两秒历史,0.4 秒重规划,14.8 s。 +-- `omg_duplex_past2_exec2_playable.mp4`:相同权重、相同两秒历史,原生两秒块执行,14.8 s。 +-- `omg_duplex_past2_exec2_tail4_playable.mp4`:相同两秒块推理追加四秒静音,18.8 s。 +- +-视频都在 `/root/autodl-tmp/runs/hfd-overfit-videos-jointfixed`。左侧预测、右侧原生真值,保留原始任务录音和背景音,H.264/MP3 MP4。第三条 14.8 s 之后右侧仅保持最终姿态,不存在真实真值,画面有明确标注。 +- +-延长尾段时,原始 14.8 s 的预测与原音轨逐项完全不变;新增部分全为诊断静音,并从误差计算中排除。因此延长版仍为 0.18776 rad,而不能借延长真值待机降低指标。新增四秒主要是收尾/待机,没有补出新的任务动作。原始模型的几乎不动不能归因于没有多给四秒等待时间。 +- +-本次 1600 步原生训练、严格最终权重重载、数据合同和 29 项回归检查通过;三条新视频完整声画解码通过。新 run 三份等价重复检查点已清理,释放约 2.96 GiB,唯一 `last.ckpt` 保留。完整数字、每轮窗口边界和执行片段记录见 `omg_context_debug_summary.json`。 +- +-延长尾段只用于诊断,必须跟在完整原生 session 后;复现命令: +- +-```bash +-python infer_overfit.py --run RUN --continuous-only --task-seconds 0 14.8 --omg-execute-frames 60 --tail-seconds 4 --output-dir RUN/inference-tail4 +-``` +- +-当前训练/推理两秒上下文模板仍默认执行 12 帧(0.4 s);两秒块对照用 CLI `--omg-execute-frames 60`。窗口长度、预测长度、执行长度是三个不同参数,不能把扩大上下文和延长等待混为同一件事。 +diff --git a/docs/history/README.md b/docs/history/README.md +deleted file mode 100644 +index c6d32cb..0000000 +--- a/docs/history/README.md ++++ /dev/null +@@ -1,9 +0,0 @@ +-> 最新调试改为 full_session 动作监督;OMG 10 步采样,TextOP 连续生成仍未解决。以根目录 README 和 docs/ROLLOUT_DEBUG.md 为准。 +- +-# 历史记录 +- +-本目录和 `configs/history/`、`reports/history/` 保留过去实验的文档、配置与检查记录。当前方案以根目录 README 为准。 +- +-这些文件没有删除,也没有移出 Git 仓库。引用的旧运行路径属于当时证据;实际产物现归档到 `/root/autodl-tmp/hfd-runs-archive-20261007/`。 +- +-旧运行产物后来经用户确认永久删除;历史配置和报告仍保留,但旧产物路径不再可用。当前样本与模型保存在 `runs/hfd-final/`。 +diff --git a/docs/history/VIDEO_SAMPLES.md b/docs/history/VIDEO_SAMPLES.md +deleted file mode 100644 +index 5a47fd5..0000000 +--- a/docs/history/VIDEO_SAMPLES.md ++++ /dev/null +@@ -1,92 +0,0 @@ +-> 历史实验记录;当前入口见 [README](../../README.md)。旧运行目录已移动到 `/root/autodl-tmp/hfd-runs-archive-20261007/`。 +- +-推荐查看新的完整单任务视频 `textop_complete_task_playable.mp4`、`omg_complete_task_playable.mp4`,任务与训练说明见 [NORMAL_TASK.md](NORMAL_TASK.md)。下文是此前八窗口及事件短片的记录。 +- +-# 修正关节映射后的推理视频与任务说明 +- +-七个 MP4 均已重新导出声音,使用 H.264 画面、48 kHz 双声道 MP3 音轨;逐个完整解码,检查非零音量、音频时长和视频帧数。音轨来自原始 session 的混合录音,保留语音、背景音和静音间隔。这批视频只对播放声音施加固定增益(不削波),没有改变模型输入、预测动作或训练结果。逐文件增益、音量和哈希见 `audio_verification.json`。 +- +-除 `gt_before_after_playable.mp4` 外,左侧是模型预测,右侧是真值。真值对照视频左侧是旧的错误关节解释,右侧是正确解释;两侧均为同一段真值,没有模型预测。 +- +-## 原生 session 中的任务 +- +-任务描述依据 session 的原始动作名与事件元数据,便于理解动作。下面引用的普通任务英文为已保存 ASR 的识别结果,可能有识别误差;`Stop now.` 是原始控制事件中记录的口令。GT 描述只用于本文解释,没有作为额外模型输入。 +- +-| session | 任务与交互 | 声音 | +-| --- | --- | --- | +-| `train-000002-seed42` | 第一轮从站立进入吸烟动作(`idle_to_smoke_idle_R_001__A552`);第一任务尚未结束时,第二轮要求向前走并抓取物体(`grab_walk_ff_180_001__A549`)。新任务打断旧任务。 | 英文男声;ASR 为 “from standing to standing and smoking.”,随后 “Character walks forward while grabbing something.”;保留空调背景音。 | +-| `train-000014-seed42` | 四轮分别为转身、站立喝水后扔瓶、再次转身、向前走并抓取物体。第二轮喝水/扔瓶任务未完成时插入 `Stop now.`,然后再进行后续任务。 | 英文女声;第二轮 ASR 为 “Drinks from bottle while standing, then throws bottle to the left.”;停止口令来自原始元数据;保留空调背景音。 | +- +-原生数据支持中文和背景音,本次选中的这两个 session 恰好是英文,未把中文样本转成英文。动作名和 ASR 的左右方向可能不一致,因此描述不据 ASR 推断精确转向;真值动作以原始轨迹为准。 +- +-## 连续事件视频 +- +-以下时间范围均为原始 session 的动作时间。原生适配器的麦克风时间比动作时间多 0.2 秒;音频片段按这一关系裁切,与事件时间对齐。 +- +-| 文件 | 动作时间范围 | 关键指令在视频中的时间 | 要观察的任务 | +-| --- | --- | --- | --- | +-| `textop_interruption_playable.mp4` | 7.72–13.64 s | 第二轮开始于约 1.08 s | 吸烟任务进行中输入向前走、抓取的新任务;听到的是第二轮指令,第一轮完整指令在更早处。 | +-| `omg_interruption_playable.mp4` | 7.63–13.63 s | 第二轮开始于约 1.17 s | 同上。 | +-| `textop_emergency-stop_playable.mp4` | 81.00–86.92 s | `Stop now.` 开始于约 1.00 s | 喝水/扔瓶任务中的急停;第二轮完整任务指令位于录音 73.8–77.0 s,不包含在这一短片中。 | +-| `omg_emergency-stop_playable.mp4` | 80.83–86.83 s | `Stop now.` 开始于约 1.17 s | 同上。 | +-| `gt_before_after_playable.mp4` | 7.00–13.00 s | 第二轮开始于约 1.80 s | 同一段吸烟被新任务打断的真值,仅用于对照错误/正确的关节映射。现在也包含对应原始音频。 | +- +-指令前后短暂只有背景音是 session 原本的内容。连续推理开头使用一次真实历史,之后反馈模型预测。不是所有事件窗口都在八窗口训练子集中,因此这些片段不是完整的全双工能力评测。 +- +-## 训练窗口拼接视频 +- +-`*_fitted_windows_playable.mp4` 拼接八个用于训练检查的窗口,并非完整连续任务。TextOP 每段 0.64 秒,总长 5.12 秒;OMG 每段 2 秒,总长 16 秒。窗口常落在动作执行或待机阶段,可能只含背景音,也可能截到半句话;不添加额外解说或补播片段外的指令。 +- +-### TEXTOP:`textop_fitted_windows_playable.mp4` +- +-| 样本 | 视频时间 | 原始 session | 原始动作时间 | +-| --- | --- | --- | --- | +-| 0 | 0.00–0.64 s | `train-000002-seed42.npz` | 0.04–0.68 s | +-| 1 | 0.64–1.28 s | `train-000014-seed42.npz` | 4.20–4.84 s | +-| 2 | 1.28–1.92 s | `train-000014-seed42.npz` | 33.32–33.96 s | +-| 3 | 1.92–2.56 s | `train-000014-seed42.npz` | 62.44–63.08 s | +-| 4 | 2.56–3.20 s | `train-000014-seed42.npz` | 91.72–92.36 s | +-| 5 | 3.20–3.84 s | `train-000014-seed42.npz` | 120.84–121.48 s | +-| 6 | 3.84–4.48 s | `train-000014-seed42.npz` | 149.96–150.60 s | +-| 7 | 4.48–5.12 s | `train-000014-seed42.npz` | 179.24–179.88 s | +- +-### OMG:`omg_fitted_windows_playable.mp4` +- +-| 样本 | 视频时间 | 原始 session | 原始动作时间 | +-| --- | --- | --- | --- | +-| 0 | 0.00–2.00 s | `train-000002-seed42.npz` | 0.43–2.43 s | +-| 1 | 2.00–4.00 s | `train-000014-seed42.npz` | 5.63–7.63 s | +-| 2 | 4.00–6.00 s | `train-000014-seed42.npz` | 34.43–36.43 s | +-| 3 | 6.00–8.00 s | `train-000014-seed42.npz` | 62.83–64.83 s | +-| 4 | 8.00–10.00 s | `train-000014-seed42.npz` | 91.63–93.63 s | +-| 5 | 10.00–12.00 s | `train-000014-seed42.npz` | 120.03–122.03 s | +-| 6 | 12.00–14.00 s | `train-000014-seed42.npz` | 148.83–150.83 s | +-| 7 | 14.00–16.00 s | `train-000014-seed42.npz` | 177.63–179.63 s | +- +-## 结果与重现 +- +-两个模型从官方初始化权重重新训练 800 步;修复前关节映射错误的检查点、视频和动作指标作废。TextOP 预测 23 个关节,其余腕关节固定在初始化姿态;OMG 预测全部 29 个关节。OMG 使用离线目标段音频,TextOP 使用已释放窗口的 ASR。视频是参考姿态播放,不是物理跟踪执行。 +- +-训练、误差与映射核验位于数据盘的视频目录 `/root/autodl-tmp/runs/hfd-overfit-videos-jointfixed`,见 `summary.json`、`textop_report.json`、`omg_report.json`、`mapping_audit.json`;新的声音核验见 `audio_verification.json`。 +- +-后续导出使用代码仓库的 `render_overfit.py`,其默认生成 48 kHz 双声道 MP3 并实测解码音量。真值对照使用 `check_joint_mapping.py`,现在也导出原始音频。已有画面可用 `remux_video_audio.py` 重建音轨,例如: +- +-```bash +-python remux_video_audio.py \ +- --video /data/runs/textop/inference/interruption_playable.mp4 \ +- --audio-npz /data/runs/textop/inference/interruption.npz \ +- --output /data/runs/textop/inference/interruption_with_audio.mp4 +-``` +- +-若从完整 session NPZ 裁切音频,需传 `--audio-start-seconds`(动作起点 + 0.2 秒)。默认播放增益为 1;`--gain` 仅影响导出的声音,不影响模型。脚本在增益会造成削波时拒绝导出。 +- +-## 当前视频格式与 VS Code 播放 +- +-当前交付统一为 MP4 容器、H.264 视频、48 kHz 双声道 MP3 音轨;文件名以 `_playable.mp4` 结尾。根据官方文档选择 H.264、MP4 与 MP3,避开不支持的 AAC。此前 VP8 WebM 在用户当前预览中仍无法打开,因此已清理。 +- +-官方支持清单:https://github.com/microsoft/vscode/issues/180780 +-官方 Webview 说明:https://code.visualstudio.com/api/extension-guides/webview#supported-media-formats +- +-所有视频检查过完整声画解码、非静音音轨、帧数和帧率;H.264 码流与原始视频的 SHA256 一致,转换只更换音频编码和封装,没有改动作。检查记录见视频目录的 `playable_mp4_check.json`。远程服务器上的解码检查不等于本地 VS Code 实测,首次打开请使用新文件名。 +- +-后续 `render_overfit.py` 与 `remux_video_audio.py` 默认导出 H.264/MP3 MP4,无需 WebM 或额外的 `--vscode-preview` 参数。 +diff --git a/hfd_bridge/genmo_prefetch.py b/hfd_bridge/genmo_prefetch.py +index e9445d1..b37742f 100644 +--- a/hfd_bridge/genmo_prefetch.py ++++ b/hfd_bridge/genmo_prefetch.py +@@ -15,6 +15,8 @@ class CPUBatches(IterableDataset): + + def __iter__(self): + os.environ['CUDA_VISIBLE_DEVICES']='' ++ from .genmo_runtime import require_replay_runtime ++ require_replay_runtime() + torch.set_num_threads(2) + from .genmo_data import PairedNativeProvider,GenmoBatches + provider=PairedNativeProvider(**self.options) +diff --git a/hfd_bridge/genmo_training.py b/hfd_bridge/genmo_training.py +index 3f094ed..eba28ac 100644 +--- a/hfd_bridge/genmo_training.py ++++ b/hfd_bridge/genmo_training.py +@@ -13,6 +13,19 @@ def evaluation_mode(model): + for module,state in states.items():module.training=state + + ++@contextmanager ++def sampling_autocast(device_type): ++ """Do not cache no-grad weight casts inside the outer training autocast. ++ ++ Reusing those casts in the subsequent loss forward silently removes the ++ denoiser's parameter gradients. Preserve the caller's AMP dtype and state. ++ """ ++ with torch.autocast(device_type=device_type, ++ enabled=torch.is_autocast_enabled(device_type), ++ dtype=torch.get_autocast_dtype(device_type),cache_enabled=False): ++ yield ++ ++ + def execution_mask(mask,history,execute,weight): + """Preserve each feature's total loss mass while emphasizing executed frames.""" + result=mask.float().clone() +@@ -61,7 +74,7 @@ def consistent_translation_loss(inputs,outputs,pipeline): + def generate_parameters(model,batch): + """Generate from past parameters only. Future placeholders contain no GT.""" + from .genmo import HISTORY,FUTURE +- with evaluation_mode(model): ++ with sampling_autocast(batch['smpl_params_w']['transl'].device.type),evaluation_mode(model): + model.prepare_batch(batch,'regression') + model.create_condition_mask(batch,{},'diffusion',train=False) + result=model.pipeline.denoiser3d.forward_test(batch) +diff --git a/hfd_bridge/human_sessions.py b/hfd_bridge/human_sessions.py +index 52fb0c1..89175d9 100644 +--- a/hfd_bridge/human_sessions.py ++++ b/hfd_bridge/human_sessions.py +@@ -65,11 +65,11 @@ class NativeHumanObserver: + def __init__(self, ds, module, root): + # Share the bridge's verified local metadata mirror. Raw SMPL-X assets + # remain indexed in their original shared-storage tar shards. +- from .storage import ArchivePath ++ from .storage import ArchivePath,store + class IndexedReader: + def read(self, path):return ArchivePath(path).read_bytes() + self.store = IndexedReader() +- pairs = pd.read_parquet(Path(root)/'metadata/seed_g1_smplx_pairs.parquet') ++ pairs = pd.read_parquet(store().metadata_path('metadata/seed_g1_smplx_pairs.parquet',root)) + self.pairs = {r.motion_id:r for r in pairs.itertuples()} + self.ds, self.module = ds, module + self.rows = {} +diff --git a/hfd_bridge/kimodo.py b/hfd_bridge/kimodo.py +index 67cb18d..e9742a6 100644 +--- a/hfd_bridge/kimodo.py ++++ b/hfd_bridge/kimodo.py +@@ -25,6 +25,8 @@ def setup(gpu=0): + HFD_BODY_MODELS_ROOT=os.environ.get('HFD_BODY_MODELS_ROOT','/root/autodl-fs/pangxinlei/models/body_models'), + TEXT_ENCODERS_DIR=os.environ.get('TEXT_ENCODERS_DIR',str(Path(os.environ.get('KIMODO_MODELS_ROOT','/root/autodl-fs/pangxinlei/models/kimodo'))/'runtime-text-encoders')), + TEXT_ENCODER_DEVICE='cuda:0',OMP_NUM_THREADS='4',MKL_NUM_THREADS='4') ++ os.environ.setdefault('HFD_NATIVE_STARTUP_CACHE','1') ++ os.environ.setdefault('HFD_STARTUP_INDEX_CACHE_DIR',str(Path(os.environ.get('HFD_DATA_DISK','/root/autodl-tmp'))/'cache/hfd/startup-indexes')) + torch.set_num_threads(4) + + +diff --git a/hfd_bridge/loader.py b/hfd_bridge/loader.py +index 31321d0..124b42d 100644 +--- a/hfd_bridge/loader.py ++++ b/hfd_bridge/loader.py +@@ -30,7 +30,7 @@ def _instrument_returns(text,filename): + tree=Returns().visit(ast.parse(text,filename)) + return compile(ast.fix_missing_locations(tree),filename,'exec') + +-def load_hfd(): ++def load_hfd(replay_numpy2=False): + root=Path(os.environ['HFD_CODE_ROOT']);sys.path.insert(0,str(root/'src')) + source=root/'src/psi/data/seed/full_duplex_audio_command_dataset.py' + text=source.read_text() +@@ -74,5 +74,9 @@ def load_hfd(): + module._hfd_return_hooks=hooked;module._hfd_current_observer=None + module._hfd_observer_return=_return_hook + exec(_instrument_returns(text,str(source)) if hooked else compile(text,str(source),'exec'),module.__dict__) ++ if replay_numpy2: ++ from .replay_numerics import quat_to_yaw_wxyz,wrap_angle ++ module.quat_to_yaw_wxyz=quat_to_yaw_wxyz ++ module.wrap_angle=wrap_angle + module.SOURCE_SHA256=hashlib.sha256(source.read_bytes()).hexdigest() + return module +diff --git a/hfd_bridge/omg_checks.py b/hfd_bridge/omg_checks.py +index 17e6574..7784fa9 100644 +--- a/hfd_bridge/omg_checks.py ++++ b/hfd_bridge/omg_checks.py +@@ -37,8 +37,9 @@ class WarmupEarlyStopping(EarlyStopping): + + + class OMGTrainingObserver(pl.Callback): +- def __init__(self,output,check_pipeline=False,audio_policy='hfd_past_2s',use_asr_text=False,input_text_source='asr'): ++ def __init__(self,output,check_pipeline=False,audio_policy='hfd_past_2s',use_asr_text=False,input_text_source='asr',checkpoint_root=None): + self.output=Path(output);self.check_pipeline=check_pipeline;self.optimizer_updates=0;self.validation_batches=0 ++ self.checkpoint_root=Path(checkpoint_root) if checkpoint_root else self.output + self.audio_policy=audio_policy + self.use_asr_text=use_asr_text + self.input_text_source=input_text_source +@@ -52,6 +53,7 @@ class OMGTrainingObserver(pl.Callback): + self.optimizer_updates+=1 + def on_train_start(self,trainer,module): + self.initial_step=trainer.global_step ++ self.logged_step=trainer.global_step;self.micro_losses=[];self.micro_compute=[];self.micro_wait=[] + def on_before_optimizer_step(self,trainer,module,optimizer): + if self.check_pipeline: + grads=[p.grad for p in module.parameters() if p.grad is not None] +@@ -60,12 +62,18 @@ class OMGTrainingObserver(pl.Callback): + def on_train_batch_end(self,trainer,module,outputs,batch,batch_idx): + loss=outputs['loss'] if isinstance(outputs,dict) else outputs + if not torch.isfinite(loss).all():raise RuntimeError('Nonfinite OMG training loss') +- if trainer.is_global_zero: ++ self.micro_losses.append(float(loss.detach())*trainer.accumulate_grad_batches) ++ self.micro_compute.append(time.monotonic()-self.batch_started) ++ if self.data_wait_seconds is not None:self.micro_wait.append(self.data_wait_seconds) ++ if trainer.is_global_zero and trainer.global_step>self.logged_step: + with (self.output/'metrics.jsonl').open('a') as f: +- f.write(json.dumps(dict(step=trainer.global_step,loss=float(loss.detach()), ++ f.write(json.dumps(dict(step=trainer.global_step,loss=sum(self.micro_losses)/len(self.micro_losses), ++ micro_batches=len(self.micro_losses),learning_rate=trainer.optimizers[0].param_groups[0]['lr'], + elapsed_seconds=time.monotonic()-self.started, +- data_wait_seconds=self.data_wait_seconds, +- compute_seconds=time.monotonic()-self.batch_started))+'\n') ++ data_wait_seconds=sum(self.micro_wait),compute_seconds=sum(self.micro_compute), ++ peak_allocated_mib=torch.cuda.max_memory_allocated(module.device)/2**20 if module.device.type=='cuda' else 0))+'\n') ++ if trainer.global_step>self.logged_step: ++ self.logged_step=trainer.global_step;self.micro_losses=[];self.micro_compute=[];self.micro_wait=[] + self.previous_batch_end=time.monotonic() + def on_train_batch_start(self,trainer,module,batch,batch_idx): + self.batch_started=time.monotonic() +@@ -91,7 +99,7 @@ class OMGTrainingObserver(pl.Callback): + elapsed_seconds=time.monotonic()-self.started,best_checkpoints=best,early_stopping=early),indent=2)+'\n') + if not self.check_pipeline:return + changed=any(not torch.equal(self.initial[n],p.detach().cpu()) for n,p in module.named_parameters() if n in self.initial) +- file=self.output/'pipeline.ckpt' ++ file=self.checkpoint_root/'pipeline.ckpt' + trainer.save_checkpoint(str(file)) + trainer.strategy.barrier() + ckpt=torch.load(file,map_location='cpu',weights_only=False) +diff --git a/hfd_bridge/ondemand.py b/hfd_bridge/ondemand.py +index 0fd7144..4d00440 100644 +--- a/hfd_bridge/ondemand.py ++++ b/hfd_bridge/ondemand.py +@@ -35,8 +35,15 @@ def native_dataset(config,split): + os.environ[env]=str(resolve(c[field])) + os.environ['HFD_STARTUP_INDEX_CACHE']=os.environ.get('HFD_NATIVE_STARTUP_CACHE','0') + os.environ['HFD_RANGE_CACHE_MAX_BYTES']=str(1024*2**20) +- ds=build_dataset(c.limit_sources,split,OmegaConf.to_container(c.recipe_overrides),candidate_pool=c.candidate_pool) ++ selection=None ++ if c.get('final_release'): ++ from .final_split import FinalSelection ++ selection=FinalSelection(OmegaConf.to_container(c.final_release,resolve=True),split) ++ selection.identity['third_party_replay_source_sha256']=sha256(resolve(c.code_root)/'src/psi/data/seed/third_party_dataset.py') ++ upstream_split=selection.upstream_split if selection is not None else split ++ ds=build_dataset(c.limit_sources,upstream_split,OmegaConf.to_container(c.recipe_overrides),candidate_pool=c.candidate_pool,replay_numpy2=selection is not None) + ds.bridge_compact_output=True ++ if selection is not None:ds.bridge_final_selection=selection + _datasets[key]=ds + return _datasets[key],c + +@@ -113,6 +120,24 @@ def legal_times(arrays,model,num_primitive=4): + return range(12,max(12,stop)) + + ++def final_window_times(session,prefix_frames,model,num_primitive): ++ """Use published clocks for rejection tests, without replaying a candidate. ++ ++ Native dense starts stop prefix_frames+1 before the final motion frame. ++ The pinned third-party replay keeps that same complete native timeline. ++ This supplies the exact legal grid before loading any motion/audio arrays. ++ """ ++ count=int(session['dense_steps']);groups=int(session['semantic_windows']) ++ geometry=dict(valid_start_count=count, ++ motion_q29_xyz_quat=range(count+int(prefix_frames)+1), ++ audio=range((groups-1)*6400+9600)) ++ return legal_times(geometry,model,num_primitive) ++ ++ ++def final_window_count(session,prefix_frames,model,num_primitive): ++ return len(final_window_times(session,prefix_frames,model,num_primitive)) ++ ++ + class NativeSessionProvider: + def __init__(self,config,split,model,seed=42,num_primitive=4,memory_mib=512, + asr_cache=None,asr_cache_mib=2048,asr_model=None,asr_device='cuda:0',asr_batch_size=32, +@@ -121,6 +146,8 @@ class NativeSessionProvider: + if language not in ('all','en','zh'):raise ValueError('Unknown foreground language filter') + self.language=language + self.ds,c=native_dataset(config,split) ++ self.split=split ++ self.final_selection=getattr(self.ds,'bridge_final_selection',None) + if c.candidate_pool or c.limit_sources or c.count is not None: + raise ValueError('On-demand formal sampling requires the full balanced native recipe') + self.model=model;self.seed=int(seed);self.num_primitive=num_primitive +@@ -152,6 +179,10 @@ class NativeSessionProvider: + upper=np.asarray([self.ds._session_max_stream_steps(s)*20+self.ds.prefix_frames+1 for s in self.ds.sessions],dtype=np.int64) + self.bounds=(upper+7)//8 if model=='textop' else (upper*3+4)//5 + self.eligible=np.array([language=='all' or s[0].get('_session_voice_name','').endswith('_'+language) for s in self.ds.sessions],dtype=bool) ++ if self.final_selection is not None: ++ allowed=self.final_selection.native ++ if allowed and max(allowed)>=len(self.ds.sessions):raise ValueError('Final native index exceeds the pinned source inventory') ++ self.eligible &= np.array([i in allowed for i in range(len(self.ds.sessions))],dtype=bool) + if not self.eligible.any():raise ValueError('No native sessions in requested foreground language') + self.bounds[~self.eligible]=0 + self.rank=int(os.environ.get('RANK','0'));self.world_size=int(os.environ.get('WORLD_SIZE','1')) +@@ -174,27 +205,39 @@ class NativeSessionProvider: + for session in self.ds.sessions: + digest.update(json.dumps([(int(r['row_id']),r.get('_session_interaction_mode'),r.get('_session_voice_name'),r.get('_session_voice_variant')) for r in session],sort_keys=True).encode()) + self.audit['ordered_session_selection_sha256']=digest.hexdigest() ++ if self.final_selection is not None: ++ self.audit['final_release']=self.final_selection.identity ++ self.audit['ordered_session_selection_sha256']=self.final_selection.ordered_identity(language) + def arrays(self,index): + if hasattr(self,'eligible') and not self.eligible[int(index)]:raise ValueError('Session excluded by foreground language filter') + cached=self.cache.get(index) + if cached is not None:return cached + raw=fixed_session(self.ds,int(index),self.seed);arrays=raw['arrays'] + raw['times']=legal_times(arrays,self.model,self.num_primitive) ++ if getattr(self,'final_selection',None) is not None and self.model in ('textop','omg'): ++ if raw['times']!=self.window_times(index):raise ValueError('Final native window grid changed') + if len(raw['times'])>self.bounds[index]:raise RuntimeError('Native conservative length bound violated') + raw['size']=sum(v.nbytes for v in arrays.values()) + self.cache.put(index,raw,raw['size']);return raw ++ def window_times(self,index): ++ if getattr(self,'final_selection',None) is not None and self.model in ('textop','omg'): ++ return final_window_times(self.final_selection.native_record(index)['session'], ++ self.ds.prefix_frames,self.model,self.num_primitive) ++ return self.arrays(index)['times'] ++ def window_count(self,index): ++ return len(self.window_times(index)) + def choose(self): + for _ in range(10000): + slot=int(self.rng.integers(int(self.cumulative[-1]))) + index=int(np.searchsorted(self.cumulative,slot,side='right')) +- raw=self.arrays(index);n=len(raw['times']) ++ n=self.window_count(index) + if self.rng.random()*self.bounds[index]self.bounds[index]:raise RuntimeError('Replayed third-party length bound violated') ++ raw_item=dict(arrays=arrays,times=times,size=sum(x.nbytes for x in arrays.values()),source_id='third-party::'+sid) ++ self.cache.put(cache_key,raw_item,raw_item['size']);return raw_item + import io, wave + with np.load(io.BytesIO(self._asset('timelines',sid)),allow_pickle=False) as z: + raw=np.asarray(z['states'],dtype=np.float32) +@@ -347,7 +418,7 @@ class ThirdPartySessionProvider: + for _ in range(10000): + slot=int(self.rng.integers(int(self.cumulative[-1]))) + index=int(np.searchsorted(self.cumulative,slot,side='right')) +- n=len(self.arrays(index)['times']) ++ n=self.window_count(index) + if self.rng.random()*self.bounds[index]self.capacity:self.files.popitem(last=False) ++ data=os.pread(handle.fileno(),count,start) ++ while len(data)1 and not ondemand, +- val_check_interval=c.training.eval_every,check_val_every_n_epoch=None) ++ val_check_interval=validation_interval(c.training),check_val_every_n_epoch=None) ++ from hydra.utils import instantiate ++ trainer_options['plugins']=[instantiate(cfg.checkpoint_io)] + if ondemand: +- trainer_options['limit_val_batches']=int(c.training.validation_batches) ++ trainer_options['limit_val_batches']=len(val_loader.dataset) + elif not val: + trainer_options['limit_val_batches']=0 + if os.environ.get('LOCAL_RANK','0')=='0': +@@ -316,7 +333,7 @@ def omg(c, train, val, output): + model.train() + if model.text_encoder is not None:model.text_encoder.encoder.eval() + trainer.fit(model,train_dataloaders=train_loader,val_dataloaders=val_loader,ckpt_path=c.training.get('resume')) +- trainer.save_checkpoint(str(output/'last.ckpt')) ++ trainer.save_checkpoint(str(weights_output/'last.ckpt')) + + + def native_options(c): +diff --git a/train_genmo_full.py b/train_genmo_full.py +index cd3a7cc..43bb8f7 100644 +--- a/train_genmo_full.py ++++ b/train_genmo_full.py +@@ -37,6 +37,20 @@ def main(): + for name in ['init_checkpoint','text_encoder','body_models']: + if not Path(c[name]).exists():raise FileNotFoundError(c[name]) + out=Path(c.output).resolve();weights=Path(c.weights_output).resolve() ++ from hfd_bridge.genmo_runtime import require_replay_runtime ++ runtime=require_replay_runtime() ++ from hfd_bridge.genmo_replay_audit import read_status ++ audit_path=c.data.get('replay_audit_status') ++ if audit_path: ++ audit_status=read_status(audit_path,runtime, ++ require_risk=bool(c.data.get('require_replay_risk_gate',True))) ++ from hfd_bridge.asr import sha256 as config_sha256 ++ if audit_status['identity']['config_sha256']!=config_sha256(Path(c.data.config)): ++ raise ValueError('Replay audit uses another session configuration') ++ for name,digest in audit_status['identity']['implementation'].items(): ++ if config_sha256(Path(__file__).parent/name)!=digest: ++ raise ValueError('Replay audit implementation changed: '+name) ++ if audit_status.get('failed'):raise RuntimeError('Replay audit failed: '+str(audit_status.get('failure'))) + if a.check_only: + print(OmegaConf.to_yaml(c));return + os.environ.update(CUDA_VISIBLE_DEVICES=str(c.gpu),GENMO_ROOT=str(c.genmo_root), +@@ -69,9 +83,11 @@ def main(): + protocol=OmegaConf.to_container(c,resolve=True) + for key in ['output','weights_output','gpu','python']:protocol.pop(key,None) + protocol['training'].pop('resume',None) ++ protocol['replay_runtime']=runtime + protocol['data_implementation']={name:sha256(Path(__file__).parent/name) for name in + ['train_genmo_full.py','hfd_bridge/genmo.py','hfd_bridge/genmo_data.py','hfd_bridge/genmo_validation.py', + 'hfd_bridge/genmo_training.py','hfd_bridge/genmo_geometry.py','hfd_bridge/genmo_prefetch.py', ++ 'hfd_bridge/genmo_runtime.py','hfd_bridge/genmo_replay_audit.py','hfd_bridge/genmo_safety.py', + 'hfd_bridge/ondemand.py','hfd_bridge/training.py','hfd_bridge/textop_condition.py', + 'hfd_bridge/human_sessions.py','hfd_bridge/reference_text.py','hfd_bridge/reference_runtime.py', + 'hfd_bridge/loader.py','hfd_bridge/sessions.py','hfd_bridge/storage.py','hfd_bridge/clocks.py']} +@@ -149,11 +165,17 @@ def main(): + finally:Path(temporary).unlink(missing_ok=True) + + class Audit(Callback): +- def __init__(self):self.start=time.monotonic();self.grad_checks=0;self.last_end=None ++ def __init__(self):self.start=time.monotonic();self.grad_checks=0;self.denoiser_grad_checks=0;self.last_end=None + def on_train_batch_start(self,trainer,module,batch,batch_idx): + self.started=time.monotonic();self.wait=None if self.last_end is None else self.started-self.last_end + self.data_profile=batch.pop('_hfd_data_profile',{}) + def on_before_optimizer_step(self,trainer,module,optimizer): ++ # Recovery sampling must never silence the trainable Transformer. ++ core=module.pipeline.denoiser3d.denoiser.blocks ++ missing=[name for name,param in core.named_parameters() if param.requires_grad and param.grad is None] ++ if missing: ++ raise FloatingPointError(f'GenMo Transformer gradients missing for {len(missing)} tensors: {missing[:3]}') ++ self.denoiser_grad_checks+=1 + if trainer.global_step>=4 and trainer.global_step%100 and self.grad_checks>=4:return + gradients=[v.grad for v in module.parameters() if v.requires_grad and v.grad is not None] + if not gradients or not all(torch.isfinite(g).all() for g in gradients):raise FloatingPointError('Missing/nonfinite gradients') +@@ -180,11 +202,13 @@ def main(): + def on_save_checkpoint(self,trainer,module,state):state['hfd_contract']=protocol + + audit=Audit() ++ from hfd_bridge.genmo_safety import GenmoReplaySafety ++ replay_safety=GenmoReplaySafety(weights,c.training.get('early_checkpoint_steps',[]),audit_path,runtime) + from hfd_bridge.genmo_validation import GenmoClosedLoopValidation + closed=c.get('strategy',{}).get('closed_loop_validation',{}) + monitor='val/closed_loop_score' if closed.get('enabled',False) else 'val/loss' + callbacks=[HFDResumeState(train,val),GenmoValidationNoise(int(c.training.seed)+1000003), +- GenmoClosedLoopValidation(val.provider,text,closed,out),audit, ++ GenmoClosedLoopValidation(val.provider,text,closed,out),audit,replay_safety, + ModelCheckpoint(dirpath=str(weights/'checkpoints'),every_n_train_steps=int(c.training.save_every), + save_top_k=int(c.training.keep_last_checkpoints),monitor='step',mode='max',save_last='link', + filename='step-{step:08d}',auto_insert_metric_name=False,enable_version_counter=False), +@@ -204,9 +228,12 @@ def main(): + DataLoader(val,batch_size=None,num_workers=0,generator=val_generator), + ckpt_path=str(checkpoint) if checkpoint else None) + summary=dict(global_step=trainer.global_step,elapsed_seconds=time.monotonic()-audit.start, +- checked_gradient_updates=audit.grad_checks,weights_output=str(weights),gpu_peak_bytes=torch.cuda.max_memory_allocated()) ++ checked_gradient_updates=audit.grad_checks,checked_denoiser_updates=audit.denoiser_grad_checks, ++ weights_output=str(weights),gpu_peak_bytes=torch.cuda.max_memory_allocated(), ++ replay_runtime=runtime,replay_audit_failure=replay_safety.failure) + (out/'training_summary.json').write_text(json.dumps(summary,indent=2)+'\n') + print('COMPLETE',json.dumps(summary),flush=True) ++ if replay_safety.failure:raise RuntimeError('Stopped at a complete update after replay audit failure') + + + if __name__=='__main__':main() +diff --git a/train_kimodo.py b/train_kimodo.py +index 75219f0..aba64ab 100644 +--- a/train_kimodo.py ++++ b/train_kimodo.py +@@ -51,6 +51,8 @@ def main(): + p.add_argument('--resume',nargs='?',const='auto');p.add_argument('--stop-after',type=int) + p.add_argument('--offline-check',action='store_true',help='Full pretrained gradient/text/causality checks only') + a=p.parse_args();c=OmegaConf.load(a.config);out=Path(c.output);weights=Path(c.weights_output) ++ # Model imports can initialise CUDA; select the physical GPU before them. ++ os.environ['CUDA_VISIBLE_DEVICES']=str(int(c.gpu)) + if c.get('kimodo_root'):os.environ['KIMODO_ROOT']=str(c.kimodo_root) + if c.get('body_models'):os.environ['HFD_BODY_MODELS_ROOT']=str(c.body_models) + if c.get('text_models_root'):os.environ['KIMODO_MODELS_ROOT']=str(c.text_models_root) diff --git a/handoff-v1/inventory/omg-git-status.txt b/handoff-v1/inventory/omg-git-status.txt new file mode 100644 index 0000000000000000000000000000000000000000..e02f3ef5e8126ca9153363f0d1393f8349abf734 --- /dev/null +++ b/handoff-v1/inventory/omg-git-status.txt @@ -0,0 +1,109 @@ + D FINAL_BASELINE_METHOD.md + D JOINT_MAPPING.md + M MANIFEST.json + M README.md + D SESSION_FORMAT.md + M UPSTREAM_LOCK.json + M configs/genmo.yaml + M configs/genmo_legacy.yaml + M configs/kimodo.yaml + M configs/omg.yaml + M configs/omg_variants.yaml + M configs/sessions.yaml + M configs/textop.yaml + D docs/ASR_CONTEXT_OVERFIT.md + D docs/BASELINE_TEXT_CONDITION.md + D docs/ENGLISH_ONLY_PROBE.md + D docs/FORMAL_TRAINING_READINESS.md + D docs/GENMO_FORMAL_TRAINING.md + D docs/GENMO_HFD_OVERFIT.md + D docs/GENMO_LISTENING_DIAGNOSIS.md + D docs/GENMO_PAST_AUDIO_TEST.md + D docs/GENMO_V2_TRAINING.md + D docs/GENMO_V2_VALIDATION.md + D docs/HFD_HUMAN_BASELINE_INPUT_PROTOCOL.md + M docs/HFD_HUMAN_BASELINE_SOURCES.json + D docs/HUMAN_FORMAL_VALIDATION_20261024.md + D docs/HUMAN_RETARGET_EVALUATION.md + D docs/KIMODO_CLOSED_LOOP_STABILITY.md + D docs/KIMODO_FORMAL_VALIDATION.md + D docs/KIMODO_HFD_OVERFIT_RESULTS.md + D docs/KIMODO_HFD_TRAINING.md + D docs/KIMODO_MOTION_RECOVERY.md + D docs/KIMODO_ROOT_COLLAPSE_DIAGNOSIS.md + D docs/KIMODO_ROOT_TRAINING_CAUSE_ANALYSIS.md + D docs/KIMODO_TRAINING_REPAIR.md + D docs/LOCAL_INDEX_TRAINING.md + D docs/MIGRATION.md + D docs/OMG_ENGLISH_VALIDATION.md + D docs/OMG_TRAINING_REVIEW.md + D docs/OMG_VALIDATION_DIAGNOSIS.md + D docs/OMG_VALIDATION_ROLLOUT.md + D docs/PORTABLE_TRAINING.md + D docs/REFERENCE_DYNAMIC_TRAINING.md + D docs/ROLLOUT_DEBUG.md + D docs/SEMANTICS_CLOCK_FIX.md + D docs/TEXTOP_ENGLISH_VALIDATION.md + D docs/TRAINING_REPRODUCTION_MATRIX.md + D docs/TRAINING_RESUME.md + D docs/WORKSPACE_CLEANUP.md + D docs/history/BASELINE_METHOD_BEFORE_SEMANTIC_FIX.md + D docs/history/DUPLEX_TASK.md + D docs/history/HFD_HUMAN_BASELINE_INPUT_PROTOCOL_RESEARCH_20261017.md + D docs/history/HFD_WINDOW_PROTOCOL.md + D docs/history/NORMAL_TASK.md + D docs/history/OMG_CONTEXT_DEBUG.md + D docs/history/README.md + D docs/history/VIDEO_SAMPLES.md + M hfd_bridge/genmo_prefetch.py + M hfd_bridge/genmo_training.py + M hfd_bridge/human_sessions.py + M hfd_bridge/kimodo.py + M hfd_bridge/loader.py + M hfd_bridge/omg_checks.py + M hfd_bridge/ondemand.py + M hfd_bridge/ondemand_data.py + M hfd_bridge/resume.py + M hfd_bridge/sessions.py + M hfd_bridge/storage.py + M prepare_sessions.py + M requirements/genmo.txt + M requirements/runtime-versions.json + M tools/configure_training.py + M train.py + M train_genmo_full.py + M train_kimodo.py +?? configs/hymotion.yaml +?? configs/hymotion_c24.yaml +?? configs/omg_l.yaml +?? configs/omg_xl.yaml +?? docs/CURRENT_RUNS.md +?? docs/HYMOTION_ASSETS.json +?? hfd_bridge/final_split.py +?? hfd_bridge/final_third_party.py +?? hfd_bridge/genmo_replay_audit.py +?? hfd_bridge/genmo_runtime.py +?? hfd_bridge/genmo_safety.py +?? hfd_bridge/hot_assets.py +?? hfd_bridge/hymotion_autoregressive.py +?? hfd_bridge/hymotion_data.py +?? hfd_bridge/hymotion_evaluation.py +?? hfd_bridge/hymotion_metrics.py +?? hfd_bridge/hymotion_model.py +?? hfd_bridge/hymotion_rep.py +?? hfd_bridge/hymotion_rollout.py +?? hfd_bridge/hymotion_text.py +?? hfd_bridge/hymotion_training.py +?? hfd_bridge/hymotion_validation.py +?? hfd_bridge/omg_scaling.py +?? hfd_bridge/replay_numerics.py +?? hfd_bridge/textop_text_cache.py +?? patches/ +?? requirements/hymotion.txt +?? tools/audit_genmo_replay.py +?? tools/download_hymotion_assets.py +?? tools/index_hymotion_text_cache.py +?? tools/prepare_hot_assets.py +?? train_hymotion.py +?? train_hymotion_autoregressive.py +?? upstream/hfd/src/psi/data/seed/third_party_dataset.py diff --git a/handoff-v1/inventory/omg-git-submodules.txt b/handoff-v1/inventory/omg-git-submodules.txt new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/handoff-v1/inventory/omg-l-best.ckpt.json b/handoff-v1/inventory/omg-l-best.ckpt.json new file mode 100644 index 0000000000000000000000000000000000000000..a8dc18fc1741e8fbd60fa31403625492f2c70179 --- /dev/null +++ b/handoff-v1/inventory/omg-l-best.ckpt.json @@ -0,0 +1,3596 @@ +{ + "epoch": 0, + "global_step": 98000, + "pytorch-lightning_version": "2.6.6", + "state_dict": { + "denoiser.time_embed.proj.0.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.time_embed.proj.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.time_embed.proj.2.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.time_embed.proj.2.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.input_proj.weight": { + "shape": [ + 1152, + 1277 + ], + "dtype": "torch.float32" + }, + "denoiser.input_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.text_proj.weight": { + "shape": [ + 1152, + 768 + ], + "dtype": "torch.float32" + }, + "denoiser.text_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.extra_proj.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.extra_proj.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.extra_proj.1.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.extra_proj.1.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.norm_cross.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.norm_cross.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.cross_attn.in_proj_weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.cross_attn.in_proj_bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.cross_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.cross_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.norm_self.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.norm_self.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.self_attn.qkv.weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.self_attn.qkv.bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.self_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.self_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.norm_ff.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.norm_ff.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.ff.0.weight": { + "shape": [ + 4608, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.ff.0.bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.ff.3.weight": { + "shape": [ + 1152, + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.ff.3.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.audio_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.audio_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.audio_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.audio_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.human_motion_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.human_motion_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.human_motion_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.human_motion_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.norm_cross.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.norm_cross.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.cross_attn.in_proj_weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.cross_attn.in_proj_bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.cross_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.cross_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.norm_self.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.norm_self.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.self_attn.qkv.weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.self_attn.qkv.bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.self_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.self_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.norm_ff.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.norm_ff.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.ff.0.weight": { + "shape": [ + 4608, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.ff.0.bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.ff.3.weight": { + "shape": [ + 1152, + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.ff.3.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.audio_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.audio_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.audio_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.audio_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.human_motion_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.human_motion_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.human_motion_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.human_motion_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.norm_cross.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.norm_cross.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.cross_attn.in_proj_weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.cross_attn.in_proj_bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.cross_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.cross_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.norm_self.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.norm_self.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.self_attn.qkv.weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.self_attn.qkv.bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.self_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.self_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.norm_ff.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.norm_ff.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.ff.0.weight": { + "shape": [ + 4608, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.ff.0.bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.ff.3.weight": { + "shape": [ + 1152, + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.ff.3.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.audio_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.audio_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.audio_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.audio_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.human_motion_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.human_motion_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.human_motion_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.human_motion_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.norm_cross.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.norm_cross.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.cross_attn.in_proj_weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.cross_attn.in_proj_bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.cross_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.cross_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.norm_self.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.norm_self.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.self_attn.qkv.weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.self_attn.qkv.bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.self_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.self_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.norm_ff.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.norm_ff.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.ff.0.weight": { + "shape": [ + 4608, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.ff.0.bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.ff.3.weight": { + "shape": [ + 1152, + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.ff.3.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.audio_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.audio_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.audio_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.audio_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.human_motion_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.human_motion_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.human_motion_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.human_motion_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.norm_cross.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.norm_cross.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.cross_attn.in_proj_weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.cross_attn.in_proj_bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.cross_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.cross_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.norm_self.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.norm_self.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.self_attn.qkv.weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.self_attn.qkv.bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.self_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.self_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.norm_ff.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.norm_ff.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.ff.0.weight": { + "shape": [ + 4608, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.ff.0.bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.ff.3.weight": { + "shape": [ + 1152, + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.ff.3.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.audio_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.audio_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.audio_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.audio_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.human_motion_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.human_motion_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.human_motion_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.human_motion_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.norm_cross.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.norm_cross.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.cross_attn.in_proj_weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.cross_attn.in_proj_bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.cross_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.cross_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.norm_self.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.norm_self.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.self_attn.qkv.weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.self_attn.qkv.bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.self_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.self_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.norm_ff.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.norm_ff.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.ff.0.weight": { + "shape": [ + 4608, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.ff.0.bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.ff.3.weight": { + "shape": [ + 1152, + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.ff.3.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.audio_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.audio_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.audio_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.audio_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.human_motion_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.human_motion_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.human_motion_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.human_motion_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.norm_cross.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.norm_cross.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.cross_attn.in_proj_weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.cross_attn.in_proj_bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.cross_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.cross_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.norm_self.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.norm_self.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.self_attn.qkv.weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.self_attn.qkv.bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.self_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.self_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.norm_ff.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.norm_ff.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.ff.0.weight": { + "shape": [ + 4608, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.ff.0.bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.ff.3.weight": { + "shape": [ + 1152, + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.ff.3.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.audio_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.audio_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.audio_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.audio_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.human_motion_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.human_motion_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.human_motion_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.human_motion_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.norm_cross.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.norm_cross.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.cross_attn.in_proj_weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.cross_attn.in_proj_bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.cross_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.cross_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.norm_self.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.norm_self.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.self_attn.qkv.weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.self_attn.qkv.bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.self_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.self_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.norm_ff.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.norm_ff.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.ff.0.weight": { + "shape": [ + 4608, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.ff.0.bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.ff.3.weight": { + "shape": [ + 1152, + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.ff.3.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.audio_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.audio_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.audio_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.audio_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.human_motion_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.human_motion_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.human_motion_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.human_motion_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.norm_cross.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.norm_cross.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.cross_attn.in_proj_weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.cross_attn.in_proj_bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.cross_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.cross_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.norm_self.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.norm_self.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.self_attn.qkv.weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.self_attn.qkv.bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.self_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.self_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.norm_ff.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.norm_ff.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.ff.0.weight": { + "shape": [ + 4608, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.ff.0.bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.ff.3.weight": { + "shape": [ + 1152, + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.ff.3.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.audio_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.audio_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.audio_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.audio_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.human_motion_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.human_motion_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.human_motion_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.human_motion_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.norm_cross.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.norm_cross.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.cross_attn.in_proj_weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.cross_attn.in_proj_bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.cross_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.cross_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.norm_self.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.norm_self.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.self_attn.qkv.weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.self_attn.qkv.bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.self_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.self_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.norm_ff.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.norm_ff.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.ff.0.weight": { + "shape": [ + 4608, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.ff.0.bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.ff.3.weight": { + "shape": [ + 1152, + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.ff.3.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.audio_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.audio_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.audio_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.audio_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.human_motion_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.human_motion_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.human_motion_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.human_motion_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.norm_cross.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.norm_cross.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.cross_attn.in_proj_weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.cross_attn.in_proj_bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.cross_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.cross_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.norm_self.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.norm_self.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.self_attn.qkv.weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.self_attn.qkv.bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.self_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.self_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.norm_ff.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.norm_ff.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.ff.0.weight": { + "shape": [ + 4608, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.ff.0.bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.ff.3.weight": { + "shape": [ + 1152, + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.ff.3.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.audio_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.audio_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.audio_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.audio_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.human_motion_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.human_motion_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.human_motion_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.human_motion_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.norm_cross.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.norm_cross.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.cross_attn.in_proj_weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.cross_attn.in_proj_bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.cross_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.cross_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.norm_self.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.norm_self.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.self_attn.qkv.weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.self_attn.qkv.bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.self_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.self_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.norm_ff.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.norm_ff.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.ff.0.weight": { + "shape": [ + 4608, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.ff.0.bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.ff.3.weight": { + "shape": [ + 1152, + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.ff.3.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.audio_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.audio_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.audio_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.audio_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.human_motion_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.human_motion_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.human_motion_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.human_motion_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.norm_cross.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.norm_cross.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.cross_attn.in_proj_weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.cross_attn.in_proj_bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.cross_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.cross_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.norm_self.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.norm_self.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.self_attn.qkv.weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.self_attn.qkv.bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.self_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.self_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.norm_ff.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.norm_ff.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.ff.0.weight": { + "shape": [ + 4608, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.ff.0.bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.ff.3.weight": { + "shape": [ + 1152, + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.ff.3.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.audio_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.audio_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.audio_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.audio_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.human_motion_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.human_motion_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.human_motion_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.human_motion_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.norm_cross.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.norm_cross.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.cross_attn.in_proj_weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.cross_attn.in_proj_bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.cross_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.cross_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.norm_self.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.norm_self.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.self_attn.qkv.weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.self_attn.qkv.bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.self_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.self_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.norm_ff.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.norm_ff.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.ff.0.weight": { + "shape": [ + 4608, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.ff.0.bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.ff.3.weight": { + "shape": [ + 1152, + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.ff.3.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.audio_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.audio_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.audio_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.audio_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.human_motion_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.human_motion_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.human_motion_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.human_motion_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.norm.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.norm.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.output.weight": { + "shape": [ + 125, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.output.bias": { + "shape": [ + 125 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.shared.weight": { + "shape": [ + 32128, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.embed_tokens.weight": { + "shape": [ + 32128, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.0.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.0.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.0.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.0.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.0.layer.0.SelfAttention.relative_attention_bias.weight": { + "shape": [ + 32, + 12 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.0.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.0.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.0.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.0.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.1.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.1.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.1.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.1.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.1.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.1.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.1.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.1.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.2.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.2.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.2.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.2.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.2.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.2.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.2.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.2.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.3.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.3.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.3.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.3.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.3.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.3.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.3.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.3.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.4.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.4.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.4.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.4.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.4.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.4.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.4.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.4.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.5.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.5.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.5.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.5.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.5.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.5.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.5.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.5.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.6.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.6.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.6.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.6.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.6.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.6.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.6.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.6.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.7.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.7.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.7.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.7.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.7.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.7.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.7.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.7.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.8.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.8.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.8.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.8.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.8.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.8.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.8.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.8.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.9.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.9.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.9.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.9.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.9.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.9.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.9.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.9.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.10.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.10.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.10.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.10.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.10.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.10.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.10.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.10.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.11.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.11.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.11.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.11.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.11.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.11.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.11.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.11.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.final_layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "history_projector.0.weight": { + "shape": [ + 125 + ], + "dtype": "torch.float32" + }, + "history_projector.0.bias": { + "shape": [ + 125 + ], + "dtype": "torch.float32" + }, + "history_projector.1.weight": { + "shape": [ + 1152, + 125 + ], + "dtype": "torch.float32" + }, + "history_projector.1.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "history_projector.3.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "history_projector.3.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "audio_embedder.0.weight": { + "shape": [ + 35 + ], + "dtype": "torch.float32" + }, + "audio_embedder.0.bias": { + "shape": [ + 35 + ], + "dtype": "torch.float32" + }, + "audio_embedder.1.weight": { + "shape": [ + 2304, + 35 + ], + "dtype": "torch.float32" + }, + "audio_embedder.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "audio_embedder.4.weight": { + "shape": [ + 1152, + 2304 + ], + "dtype": "torch.float32" + }, + "audio_embedder.4.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "human_motion_embedder.0.weight": { + "shape": [ + 66 + ], + "dtype": "torch.float32" + }, + "human_motion_embedder.0.bias": { + "shape": [ + 66 + ], + "dtype": "torch.float32" + }, + "human_motion_embedder.1.weight": { + "shape": [ + 2304, + 66 + ], + "dtype": "torch.float32" + }, + "human_motion_embedder.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "human_motion_embedder.4.weight": { + "shape": [ + 1152, + 2304 + ], + "dtype": "torch.float32" + }, + "human_motion_embedder.4.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + } + }, + "loops": { + "fit_loop": { + "state_dict": { + "keys": [], + "count": 0 + }, + "epoch_loop.state_dict": { + "keys": [ + "_batches_that_stepped" + ], + "count": 1 + }, + "epoch_loop.batch_progress": { + "keys": [ + "total", + "current", + "is_last_batch" + ], + "count": 3 + }, + "epoch_loop.scheduler_progress": { + "keys": [ + "total", + "current" + ], + "count": 2 + }, + "epoch_loop.automatic_optimization.state_dict": { + "keys": [], + "count": 0 + }, + "epoch_loop.automatic_optimization.optim_progress": { + "keys": [ + "optimizer" + ], + "count": 1 + }, + "epoch_loop.manual_optimization.state_dict": { + "keys": [], + "count": 0 + }, + "epoch_loop.manual_optimization.optim_step_progress": { + "keys": [ + "total", + "current" + ], + "count": 2 + }, + "epoch_loop.val_loop.state_dict": { + "keys": [], + "count": 0 + }, + "epoch_loop.val_loop.batch_progress": { + "keys": [ + "total", + "current", + "is_last_batch" + ], + "count": 3 + }, + "epoch_progress": { + "keys": [ + "total", + "current" + ], + "count": 2 + } + }, + "validate_loop": { + "state_dict": { + "keys": [], + "count": 0 + }, + "batch_progress": { + "keys": [ + "total", + "current", + "is_last_batch" + ], + "count": 3 + } + }, + "test_loop": { + "state_dict": { + "keys": [], + "count": 0 + }, + "batch_progress": { + "keys": [ + "total", + "current", + "is_last_batch" + ], + "count": 3 + } + }, + "predict_loop": { + "state_dict": { + "keys": [], + "count": 0 + }, + "batch_progress": { + "keys": [ + "total", + "current" + ], + "count": 2 + } + } + }, + "callbacks": { + "ModelCheckpoint{'monitor': 'step', 'mode': 'max', 'every_n_train_steps': 2000, 'every_n_epochs': 0, 'train_time_interval': None}": { + "monitor": "step", + "best_model_score": { + "shape": [], + "dtype": "torch.int64" + }, + "best_model_path": "/autodl-fs/data/pangxinlei/models/OMG/hfd-l-english-split20261026/checkpoints/step-00098000.ckpt", + "current_score": { + "shape": [], + "dtype": "torch.int64" + }, + "dirpath": "/autodl-fs/data/pangxinlei/models/OMG/hfd-l-english-split20261026/checkpoints", + "best_k_models": { + "keys": [ + "/autodl-fs/data/pangxinlei/models/OMG/hfd-l-english-split20261026/checkpoints/step-00096000.ckpt", + "/autodl-fs/data/pangxinlei/models/OMG/hfd-l-english-split20261026/checkpoints/step-00098000.ckpt" + ], + "count": 2 + }, + "kth_best_model_path": "/autodl-fs/data/pangxinlei/models/OMG/hfd-l-english-split20261026/checkpoints/step-00096000.ckpt", + "kth_value": { + "shape": [], + "dtype": "torch.int64" + }, + "last_model_path": "/autodl-fs/data/pangxinlei/models/OMG/hfd-l-english-split20261026/checkpoints/last.ckpt" + }, + "ModelCheckpoint{'monitor': 'val/loss', 'mode': 'min', 'every_n_train_steps': 0, 'every_n_epochs': 1, 'train_time_interval': None}": { + "monitor": "val/loss", + "best_model_score": { + "shape": [], + "dtype": "torch.float32" + }, + "best_model_path": "/autodl-fs/data/pangxinlei/models/OMG/hfd-l-english-split20261026/best/best.ckpt", + "current_score": { + "shape": [], + "dtype": "torch.float32" + }, + "dirpath": "/autodl-fs/data/pangxinlei/models/OMG/hfd-l-english-split20261026/best", + "best_k_models": { + "keys": [ + "/autodl-fs/data/pangxinlei/models/OMG/hfd-l-english-split20261026/best/best.ckpt" + ], + "count": 1 + }, + "kth_best_model_path": "/autodl-fs/data/pangxinlei/models/OMG/hfd-l-english-split20261026/best/best.ckpt", + "kth_value": { + "shape": [], + "dtype": "torch.float32" + }, + "last_model_path": "" + } + }, + "optimizer_states": { + "count": 1, + "sample": [ + { + "state": { + "keys": [ + "0", + "1", + "2", + "3", + "4", + "5", + "6", + "7", + "8", + "9", + "10", + "11", + "12", + "13", + "14", + "15", + "16", + "17", + "18", + "19", + "20", + "21", + "22", + "23", + "24", + "25", + "26", + "27", + "28", + "29" + ], + "count": 336 + }, + "param_groups": { + "count": 1, + "sample": [ + { + "keys": [ + "lr", + "betas", + "eps", + "weight_decay", + "amsgrad", + "maximize", + "foreach", + "capturable", + "differentiable", + "fused", + "decoupled_weight_decay", + "initial_lr", + "params" + ], + "count": 13 + } + ] + } + } + ] + }, + "lr_schedulers": { + "count": 1, + "sample": [ + { + "_milestones": { + "count": 1, + "sample": [ + 2000 + ] + }, + "last_epoch": 98000, + "_last_lr": { + "count": 1, + "sample": [ + 1.0606109138651e-06 + ] + }, + "_schedulers": { + "count": 2, + "sample": [ + { + "keys": [ + "start_factor", + "end_factor", + "total_iters", + "base_lrs", + "last_epoch", + "_step_count", + "_get_lr_called_within_step", + "_last_lr" + ], + "count": 8 + }, + { + "keys": [ + "T_max", + "eta_min", + "base_lrs", + "last_epoch", + "_step_count", + "_get_lr_called_within_step", + "_last_lr" + ], + "count": 7 + } + ] + } + } + ] + }, + "hfd_resume": { + "version": 1, + "rng": { + "python": { + "count": 3, + "sample": [ + 3, + { + "count": 625, + "sample": [ + 2147483648, + 3564348608 + ] + } + ] + }, + "numpy": { + "count": 5, + "sample": [ + "MT19937", + { + "count": 624, + "sample": [ + 42, + 3107752595 + ] + } + ] + }, + "torch": { + "shape": [ + 5056 + ], + "dtype": "torch.uint8" + }, + "cuda": { + "count": 1, + "sample": [ + { + "shape": [ + 16 + ], + "dtype": "torch.uint8" + } + ] + } + }, + "train_pool": { + "step": 98000, + "slots": { + "count": 512, + "sample": [ + 271745, + 295270 + ] + }, + "batch_size": 512, + "refresh": 256, + "rng": { + "keys": [ + "bit_generator", + "state", + "has_uint32", + "uinteger" + ], + "count": 4 + }, + "identity": { + "keys": [ + "sessions", + "native_source_sha256", + "asr", + "recipe_sha256", + "foreground_language", + "ordered_session_selection_sha256", + "final_release" + ], + "count": 7 + } + }, + "val_pool": { + "step": 16, + "slots": { + "count": 512, + "sample": [ + 6295, + 7193 + ] + }, + "batch_size": 512, + "refresh": 256, + "rng": { + "keys": [ + "bit_generator", + "state", + "has_uint32", + "uinteger" + ], + "count": 4 + }, + "identity": { + "keys": [ + "sessions", + "native_source_sha256", + "asr", + "recipe_sha256", + "foreground_language", + "ordered_session_selection_sha256", + "final_release" + ], + "count": 7 + } + } + }, + "omg_model_architecture": { + "format": "omg.model_architecture", + "version": 1, + "denoiser_type": "omg.generation.denoisers.transformer.MotionTransformerDenoiser", + "frame_cond_injection": "per_layer_film", + "attention": { + "rotary_self_attention_qk_norm": true, + "cross_attention_qk_norm": true + } + } +} \ No newline at end of file diff --git a/handoff-v1/inventory/omg-l-last.ckpt.json b/handoff-v1/inventory/omg-l-last.ckpt.json new file mode 100644 index 0000000000000000000000000000000000000000..76ed84c7b8508fd0e3731d03a6028aaf86ce345f --- /dev/null +++ b/handoff-v1/inventory/omg-l-last.ckpt.json @@ -0,0 +1,3596 @@ +{ + "epoch": 1, + "global_step": 100000, + "pytorch-lightning_version": "2.6.6", + "state_dict": { + "denoiser.time_embed.proj.0.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.time_embed.proj.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.time_embed.proj.2.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.time_embed.proj.2.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.input_proj.weight": { + "shape": [ + 1152, + 1277 + ], + "dtype": "torch.float32" + }, + "denoiser.input_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.text_proj.weight": { + "shape": [ + 1152, + 768 + ], + "dtype": "torch.float32" + }, + "denoiser.text_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.extra_proj.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.extra_proj.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.extra_proj.1.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.extra_proj.1.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.norm_cross.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.norm_cross.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.cross_attn.in_proj_weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.cross_attn.in_proj_bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.cross_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.cross_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.norm_self.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.norm_self.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.self_attn.qkv.weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.self_attn.qkv.bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.self_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.self_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.norm_ff.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.norm_ff.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.ff.0.weight": { + "shape": [ + 4608, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.ff.0.bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.ff.3.weight": { + "shape": [ + 1152, + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.ff.3.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.audio_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.audio_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.audio_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.audio_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.human_motion_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.human_motion_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.human_motion_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.human_motion_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.norm_cross.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.norm_cross.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.cross_attn.in_proj_weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.cross_attn.in_proj_bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.cross_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.cross_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.norm_self.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.norm_self.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.self_attn.qkv.weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.self_attn.qkv.bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.self_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.self_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.norm_ff.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.norm_ff.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.ff.0.weight": { + "shape": [ + 4608, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.ff.0.bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.ff.3.weight": { + "shape": [ + 1152, + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.ff.3.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.audio_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.audio_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.audio_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.audio_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.human_motion_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.human_motion_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.human_motion_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.human_motion_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.norm_cross.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.norm_cross.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.cross_attn.in_proj_weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.cross_attn.in_proj_bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.cross_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.cross_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.norm_self.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.norm_self.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.self_attn.qkv.weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.self_attn.qkv.bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.self_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.self_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.norm_ff.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.norm_ff.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.ff.0.weight": { + "shape": [ + 4608, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.ff.0.bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.ff.3.weight": { + "shape": [ + 1152, + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.ff.3.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.audio_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.audio_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.audio_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.audio_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.human_motion_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.human_motion_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.human_motion_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.human_motion_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.norm_cross.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.norm_cross.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.cross_attn.in_proj_weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.cross_attn.in_proj_bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.cross_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.cross_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.norm_self.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.norm_self.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.self_attn.qkv.weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.self_attn.qkv.bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.self_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.self_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.norm_ff.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.norm_ff.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.ff.0.weight": { + "shape": [ + 4608, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.ff.0.bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.ff.3.weight": { + "shape": [ + 1152, + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.ff.3.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.audio_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.audio_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.audio_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.audio_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.human_motion_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.human_motion_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.human_motion_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.human_motion_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.norm_cross.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.norm_cross.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.cross_attn.in_proj_weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.cross_attn.in_proj_bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.cross_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.cross_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.norm_self.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.norm_self.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.self_attn.qkv.weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.self_attn.qkv.bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.self_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.self_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.norm_ff.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.norm_ff.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.ff.0.weight": { + "shape": [ + 4608, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.ff.0.bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.ff.3.weight": { + "shape": [ + 1152, + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.ff.3.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.audio_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.audio_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.audio_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.audio_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.human_motion_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.human_motion_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.human_motion_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.human_motion_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.norm_cross.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.norm_cross.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.cross_attn.in_proj_weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.cross_attn.in_proj_bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.cross_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.cross_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.norm_self.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.norm_self.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.self_attn.qkv.weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.self_attn.qkv.bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.self_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.self_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.norm_ff.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.norm_ff.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.ff.0.weight": { + "shape": [ + 4608, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.ff.0.bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.ff.3.weight": { + "shape": [ + 1152, + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.ff.3.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.audio_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.audio_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.audio_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.audio_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.human_motion_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.human_motion_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.human_motion_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.human_motion_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.norm_cross.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.norm_cross.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.cross_attn.in_proj_weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.cross_attn.in_proj_bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.cross_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.cross_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.norm_self.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.norm_self.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.self_attn.qkv.weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.self_attn.qkv.bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.self_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.self_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.norm_ff.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.norm_ff.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.ff.0.weight": { + "shape": [ + 4608, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.ff.0.bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.ff.3.weight": { + "shape": [ + 1152, + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.ff.3.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.audio_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.audio_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.audio_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.audio_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.human_motion_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.human_motion_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.human_motion_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.human_motion_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.norm_cross.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.norm_cross.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.cross_attn.in_proj_weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.cross_attn.in_proj_bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.cross_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.cross_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.norm_self.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.norm_self.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.self_attn.qkv.weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.self_attn.qkv.bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.self_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.self_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.norm_ff.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.norm_ff.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.ff.0.weight": { + "shape": [ + 4608, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.ff.0.bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.ff.3.weight": { + "shape": [ + 1152, + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.ff.3.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.audio_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.audio_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.audio_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.audio_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.human_motion_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.human_motion_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.human_motion_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.human_motion_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.norm_cross.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.norm_cross.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.cross_attn.in_proj_weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.cross_attn.in_proj_bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.cross_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.cross_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.norm_self.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.norm_self.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.self_attn.qkv.weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.self_attn.qkv.bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.self_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.self_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.norm_ff.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.norm_ff.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.ff.0.weight": { + "shape": [ + 4608, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.ff.0.bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.ff.3.weight": { + "shape": [ + 1152, + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.ff.3.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.audio_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.audio_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.audio_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.audio_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.human_motion_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.human_motion_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.human_motion_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.human_motion_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.norm_cross.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.norm_cross.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.cross_attn.in_proj_weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.cross_attn.in_proj_bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.cross_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.cross_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.norm_self.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.norm_self.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.self_attn.qkv.weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.self_attn.qkv.bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.self_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.self_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.norm_ff.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.norm_ff.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.ff.0.weight": { + "shape": [ + 4608, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.ff.0.bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.ff.3.weight": { + "shape": [ + 1152, + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.ff.3.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.audio_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.audio_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.audio_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.audio_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.human_motion_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.human_motion_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.human_motion_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.human_motion_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.norm_cross.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.norm_cross.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.cross_attn.in_proj_weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.cross_attn.in_proj_bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.cross_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.cross_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.norm_self.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.norm_self.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.self_attn.qkv.weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.self_attn.qkv.bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.self_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.self_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.norm_ff.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.norm_ff.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.ff.0.weight": { + "shape": [ + 4608, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.ff.0.bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.ff.3.weight": { + "shape": [ + 1152, + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.ff.3.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.audio_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.audio_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.audio_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.audio_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.human_motion_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.human_motion_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.human_motion_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.human_motion_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.norm_cross.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.norm_cross.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.cross_attn.in_proj_weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.cross_attn.in_proj_bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.cross_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.cross_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.norm_self.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.norm_self.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.self_attn.qkv.weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.self_attn.qkv.bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.self_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.self_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.norm_ff.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.norm_ff.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.ff.0.weight": { + "shape": [ + 4608, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.ff.0.bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.ff.3.weight": { + "shape": [ + 1152, + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.ff.3.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.audio_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.audio_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.audio_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.audio_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.human_motion_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.human_motion_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.human_motion_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.human_motion_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.norm_cross.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.norm_cross.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.cross_attn.in_proj_weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.cross_attn.in_proj_bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.cross_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.cross_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.norm_self.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.norm_self.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.self_attn.qkv.weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.self_attn.qkv.bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.self_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.self_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.norm_ff.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.norm_ff.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.ff.0.weight": { + "shape": [ + 4608, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.ff.0.bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.ff.3.weight": { + "shape": [ + 1152, + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.ff.3.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.audio_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.audio_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.audio_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.audio_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.human_motion_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.human_motion_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.human_motion_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.human_motion_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.norm_cross.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.norm_cross.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.cross_attn.in_proj_weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.cross_attn.in_proj_bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.cross_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.cross_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.norm_self.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.norm_self.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.self_attn.qkv.weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.self_attn.qkv.bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.self_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.self_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.norm_ff.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.norm_ff.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.ff.0.weight": { + "shape": [ + 4608, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.ff.0.bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.ff.3.weight": { + "shape": [ + 1152, + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.ff.3.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.audio_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.audio_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.audio_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.audio_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.human_motion_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.human_motion_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.human_motion_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.human_motion_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.norm.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.norm.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.output.weight": { + "shape": [ + 125, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.output.bias": { + "shape": [ + 125 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.shared.weight": { + "shape": [ + 32128, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.embed_tokens.weight": { + "shape": [ + 32128, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.0.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.0.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.0.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.0.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.0.layer.0.SelfAttention.relative_attention_bias.weight": { + "shape": [ + 32, + 12 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.0.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.0.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.0.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.0.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.1.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.1.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.1.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.1.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.1.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.1.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.1.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.1.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.2.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.2.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.2.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.2.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.2.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.2.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.2.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.2.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.3.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.3.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.3.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.3.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.3.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.3.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.3.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.3.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.4.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.4.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.4.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.4.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.4.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.4.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.4.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.4.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.5.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.5.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.5.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.5.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.5.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.5.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.5.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.5.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.6.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.6.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.6.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.6.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.6.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.6.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.6.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.6.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.7.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.7.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.7.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.7.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.7.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.7.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.7.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.7.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.8.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.8.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.8.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.8.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.8.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.8.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.8.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.8.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.9.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.9.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.9.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.9.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.9.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.9.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.9.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.9.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.10.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.10.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.10.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.10.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.10.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.10.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.10.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.10.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.11.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.11.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.11.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.11.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.11.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.11.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.11.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.11.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.final_layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "history_projector.0.weight": { + "shape": [ + 125 + ], + "dtype": "torch.float32" + }, + "history_projector.0.bias": { + "shape": [ + 125 + ], + "dtype": "torch.float32" + }, + "history_projector.1.weight": { + "shape": [ + 1152, + 125 + ], + "dtype": "torch.float32" + }, + "history_projector.1.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "history_projector.3.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "history_projector.3.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "audio_embedder.0.weight": { + "shape": [ + 35 + ], + "dtype": "torch.float32" + }, + "audio_embedder.0.bias": { + "shape": [ + 35 + ], + "dtype": "torch.float32" + }, + "audio_embedder.1.weight": { + "shape": [ + 2304, + 35 + ], + "dtype": "torch.float32" + }, + "audio_embedder.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "audio_embedder.4.weight": { + "shape": [ + 1152, + 2304 + ], + "dtype": "torch.float32" + }, + "audio_embedder.4.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "human_motion_embedder.0.weight": { + "shape": [ + 66 + ], + "dtype": "torch.float32" + }, + "human_motion_embedder.0.bias": { + "shape": [ + 66 + ], + "dtype": "torch.float32" + }, + "human_motion_embedder.1.weight": { + "shape": [ + 2304, + 66 + ], + "dtype": "torch.float32" + }, + "human_motion_embedder.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "human_motion_embedder.4.weight": { + "shape": [ + 1152, + 2304 + ], + "dtype": "torch.float32" + }, + "human_motion_embedder.4.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + } + }, + "loops": { + "fit_loop": { + "state_dict": { + "keys": [], + "count": 0 + }, + "epoch_loop.state_dict": { + "keys": [ + "_batches_that_stepped" + ], + "count": 1 + }, + "epoch_loop.batch_progress": { + "keys": [ + "total", + "current", + "is_last_batch" + ], + "count": 3 + }, + "epoch_loop.scheduler_progress": { + "keys": [ + "total", + "current" + ], + "count": 2 + }, + "epoch_loop.automatic_optimization.state_dict": { + "keys": [], + "count": 0 + }, + "epoch_loop.automatic_optimization.optim_progress": { + "keys": [ + "optimizer" + ], + "count": 1 + }, + "epoch_loop.manual_optimization.state_dict": { + "keys": [], + "count": 0 + }, + "epoch_loop.manual_optimization.optim_step_progress": { + "keys": [ + "total", + "current" + ], + "count": 2 + }, + "epoch_loop.val_loop.state_dict": { + "keys": [], + "count": 0 + }, + "epoch_loop.val_loop.batch_progress": { + "keys": [ + "total", + "current", + "is_last_batch" + ], + "count": 3 + }, + "epoch_progress": { + "keys": [ + "total", + "current" + ], + "count": 2 + } + }, + "validate_loop": { + "state_dict": { + "keys": [], + "count": 0 + }, + "batch_progress": { + "keys": [ + "total", + "current", + "is_last_batch" + ], + "count": 3 + } + }, + "test_loop": { + "state_dict": { + "keys": [], + "count": 0 + }, + "batch_progress": { + "keys": [ + "total", + "current", + "is_last_batch" + ], + "count": 3 + } + }, + "predict_loop": { + "state_dict": { + "keys": [], + "count": 0 + }, + "batch_progress": { + "keys": [ + "total", + "current" + ], + "count": 2 + } + } + }, + "callbacks": { + "ModelCheckpoint{'monitor': 'step', 'mode': 'max', 'every_n_train_steps': 2000, 'every_n_epochs': 0, 'train_time_interval': None}": { + "monitor": "step", + "best_model_score": { + "shape": [], + "dtype": "torch.int64" + }, + "best_model_path": "/autodl-fs/data/pangxinlei/models/OMG/hfd-l-english-split20261026/checkpoints/step-00100000.ckpt", + "current_score": { + "shape": [], + "dtype": "torch.int64" + }, + "dirpath": "/autodl-fs/data/pangxinlei/models/OMG/hfd-l-english-split20261026/checkpoints", + "best_k_models": { + "keys": [ + "/autodl-fs/data/pangxinlei/models/OMG/hfd-l-english-split20261026/checkpoints/step-00098000.ckpt", + "/autodl-fs/data/pangxinlei/models/OMG/hfd-l-english-split20261026/checkpoints/step-00100000.ckpt" + ], + "count": 2 + }, + "kth_best_model_path": "/autodl-fs/data/pangxinlei/models/OMG/hfd-l-english-split20261026/checkpoints/step-00098000.ckpt", + "kth_value": { + "shape": [], + "dtype": "torch.int64" + }, + "last_model_path": "/autodl-fs/data/pangxinlei/models/OMG/hfd-l-english-split20261026/checkpoints/last.ckpt" + }, + "ModelCheckpoint{'monitor': 'val/loss', 'mode': 'min', 'every_n_train_steps': 0, 'every_n_epochs': 1, 'train_time_interval': None}": { + "monitor": "val/loss", + "best_model_score": { + "shape": [], + "dtype": "torch.float32" + }, + "best_model_path": "/autodl-fs/data/pangxinlei/models/OMG/hfd-l-english-split20261026/best/best.ckpt", + "current_score": { + "shape": [], + "dtype": "torch.float32" + }, + "dirpath": "/autodl-fs/data/pangxinlei/models/OMG/hfd-l-english-split20261026/best", + "best_k_models": { + "keys": [ + "/autodl-fs/data/pangxinlei/models/OMG/hfd-l-english-split20261026/best/best.ckpt" + ], + "count": 1 + }, + "kth_best_model_path": "/autodl-fs/data/pangxinlei/models/OMG/hfd-l-english-split20261026/best/best.ckpt", + "kth_value": { + "shape": [], + "dtype": "torch.float32" + }, + "last_model_path": "" + } + }, + "optimizer_states": { + "count": 1, + "sample": [ + { + "state": { + "keys": [ + "0", + "1", + "2", + "3", + "4", + "5", + "6", + "7", + "8", + "9", + "10", + "11", + "12", + "13", + "14", + "15", + "16", + "17", + "18", + "19", + "20", + "21", + "22", + "23", + "24", + "25", + "26", + "27", + "28", + "29" + ], + "count": 336 + }, + "param_groups": { + "count": 1, + "sample": [ + { + "keys": [ + "lr", + "betas", + "eps", + "weight_decay", + "amsgrad", + "maximize", + "foreach", + "capturable", + "differentiable", + "fused", + "decoupled_weight_decay", + "initial_lr", + "params" + ], + "count": 13 + } + ] + } + } + ] + }, + "lr_schedulers": { + "count": 1, + "sample": [ + { + "_milestones": { + "count": 1, + "sample": [ + 2000 + ] + }, + "last_epoch": 100000, + "_last_lr": { + "count": 1, + "sample": [ + 1e-06 + ] + }, + "_schedulers": { + "count": 2, + "sample": [ + { + "keys": [ + "start_factor", + "end_factor", + "total_iters", + "base_lrs", + "last_epoch", + "_step_count", + "_get_lr_called_within_step", + "_last_lr" + ], + "count": 8 + }, + { + "keys": [ + "T_max", + "eta_min", + "base_lrs", + "last_epoch", + "_step_count", + "_get_lr_called_within_step", + "_last_lr" + ], + "count": 7 + } + ] + } + } + ] + }, + "hfd_resume": { + "version": 1, + "rng": { + "python": { + "count": 3, + "sample": [ + 3, + { + "count": 625, + "sample": [ + 2147483648, + 3564348608 + ] + } + ] + }, + "numpy": { + "count": 5, + "sample": [ + "MT19937", + { + "count": 624, + "sample": [ + 42, + 3107752595 + ] + } + ] + }, + "torch": { + "shape": [ + 5056 + ], + "dtype": "torch.uint8" + }, + "cuda": { + "count": 1, + "sample": [ + { + "shape": [ + 16 + ], + "dtype": "torch.uint8" + } + ] + } + }, + "train_pool": { + "step": 100000, + "slots": { + "count": 512, + "sample": [ + 248743, + 347419 + ] + }, + "batch_size": 512, + "refresh": 256, + "rng": { + "keys": [ + "bit_generator", + "state", + "has_uint32", + "uinteger" + ], + "count": 4 + }, + "identity": { + "keys": [ + "sessions", + "native_source_sha256", + "asr", + "recipe_sha256", + "foreground_language", + "ordered_session_selection_sha256", + "final_release" + ], + "count": 7 + } + }, + "val_pool": { + "step": 16, + "slots": { + "count": 512, + "sample": [ + 6295, + 7193 + ] + }, + "batch_size": 512, + "refresh": 256, + "rng": { + "keys": [ + "bit_generator", + "state", + "has_uint32", + "uinteger" + ], + "count": 4 + }, + "identity": { + "keys": [ + "sessions", + "native_source_sha256", + "asr", + "recipe_sha256", + "foreground_language", + "ordered_session_selection_sha256", + "final_release" + ], + "count": 7 + } + } + }, + "omg_model_architecture": { + "format": "omg.model_architecture", + "version": 1, + "denoiser_type": "omg.generation.denoisers.transformer.MotionTransformerDenoiser", + "frame_cond_injection": "per_layer_film", + "attention": { + "rotary_self_attention_qk_norm": true, + "cross_attention_qk_norm": true + } + } +} \ No newline at end of file diff --git a/handoff-v1/inventory/omg-l-source-audit.json b/handoff-v1/inventory/omg-l-source-audit.json new file mode 100644 index 0000000000000000000000000000000000000000..3baa2b9d991ed4bec282e5d908aa3a4aaa3db203 --- /dev/null +++ b/handoff-v1/inventory/omg-l-source-audit.json @@ -0,0 +1,586 @@ +{ + "hfd_bridge/storage.py": { + "expected": "02c4fe4eed5f6d17bf140568af365b0abc8e5843df018d65553828890058a1fa", + "actual": "02c4fe4eed5f6d17bf140568af365b0abc8e5843df018d65553828890058a1fa" + }, + "hfd_bridge/__init__.py": { + "expected": "885528eb35015189a0bcd2dc62b60c61878efcf959caef68ff69e93501cafd5c", + "actual": "885528eb35015189a0bcd2dc62b60c61878efcf959caef68ff69e93501cafd5c" + }, + "hfd_bridge/loader.py": { + "expected": "13772137dc5cbb036bd5b5e4ee6ffe8cb45ba8d033ad580887b4dd63b253ac9b", + "actual": "13772137dc5cbb036bd5b5e4ee6ffe8cb45ba8d033ad580887b4dd63b253ac9b" + }, + "hfd_bridge/sessions.py": { + "expected": "c271aecf69e52acaee178b1381e465f22af82732f606d4e0bd5f891e2bb86076", + "actual": "c271aecf69e52acaee178b1381e465f22af82732f606d4e0bd5f891e2bb86076" + }, + "hfd_bridge/asr.py": { + "expected": "4baaf4faecfdce2054844a41326760220af301eb11d6fad892bed8c7380dceb0", + "actual": "4baaf4faecfdce2054844a41326760220af301eb11d6fad892bed8c7380dceb0" + }, + "hfd_bridge/training.py": { + "expected": "b513a632bcf010c83c23a35bcfd09e189149ff8c0d7d667d8966ec09e6e55970", + "actual": "b513a632bcf010c83c23a35bcfd09e189149ff8c0d7d667d8966ec09e6e55970" + }, + "hfd_bridge/native.py": { + "expected": "72ca9f675c2dfc7a2e304191da7b4d4016cb6bc226ccc22f9e7d0aeb2889c9f4", + "actual": "72ca9f675c2dfc7a2e304191da7b4d4016cb6bc226ccc22f9e7d0aeb2889c9f4" + }, + "hfd_bridge/paths.py": { + "expected": "2aaee82ecad571dcc1186c275e270af6c854c670db4ae354ee984d23a50b8bf3", + "actual": "2aaee82ecad571dcc1186c275e270af6c854c670db4ae354ee984d23a50b8bf3" + }, + "hfd_bridge/configured_data.py": { + "expected": "f916e0f9d4cdd2a4120fe94a9f1dfe59bbd24606d71ce1941023f04f97f181cd", + "actual": "f916e0f9d4cdd2a4120fe94a9f1dfe59bbd24606d71ce1941023f04f97f181cd" + }, + "hfd_bridge/textop_checks.py": { + "expected": "00662346781a638298cfa344b9ce8a58b6e8bba11261cd5783fdca956f4df7c8", + "actual": "00662346781a638298cfa344b9ce8a58b6e8bba11261cd5783fdca956f4df7c8" + }, + "hfd_bridge/omg_checks.py": { + "expected": "094d997043d4e0d868261b7f25b9452fca832dfa48572aeac4a3f2ba9f42d172", + "actual": "094d997043d4e0d868261b7f25b9452fca832dfa48572aeac4a3f2ba9f42d172" + }, + "hfd_bridge/textop_condition.py": { + "expected": "4fca0f4a02df3b843a908b28b64766877585aa3300977359d21fdb3adb5a3c28", + "actual": "4fca0f4a02df3b843a908b28b64766877585aa3300977359d21fdb3adb5a3c28" + }, + "hfd_bridge/joints.py": { + "expected": "7827c82ca3463a026db6007c41a66528e1294485b4acb081f27f09449da2d154", + "actual": "7827c82ca3463a026db6007c41a66528e1294485b4acb081f27f09449da2d154" + }, + "hfd_bridge/audio_condition.py": { + "expected": "cab6795bdddc36aa63d4f54c3c8e8aaee7958c86882abe35ae76f67a79b4b945", + "actual": "cab6795bdddc36aa63d4f54c3c8e8aaee7958c86882abe35ae76f67a79b4b945" + }, + "hfd_bridge/batch_asr.py": { + "expected": "9d4f0ba1d382f45c7a0adf562bc008485fce7cd00e86e29607528676f503f8f1", + "actual": "9d4f0ba1d382f45c7a0adf562bc008485fce7cd00e86e29607528676f503f8f1" + }, + "hfd_bridge/ondemand.py": { + "expected": "940339aa16a6fa25a27ace23a956017f7325813b0cd2a49f76e6f1fb1d014cc0", + "actual": "940339aa16a6fa25a27ace23a956017f7325813b0cd2a49f76e6f1fb1d014cc0" + }, + "hfd_bridge/ondemand_data.py": { + "expected": "24c61e36623eaca0bdc2f7bff2c8ce307946c4a5afef18f0427b96c06c9ed57e", + "actual": "24c61e36623eaca0bdc2f7bff2c8ce307946c4a5afef18f0427b96c06c9ed57e" + }, + "hfd_bridge/prefetch.py": { + "expected": "b227251cc6470f980231999bbecbeb1fc583c754eda56e70d4f3c84ed1845d64", + "actual": "b227251cc6470f980231999bbecbeb1fc583c754eda56e70d4f3c84ed1845d64" + }, + "hfd_bridge/human_sessions.py": { + "expected": "0461ea66b1c324ec38e9e833ee8eb19b95a24c3f478b094fb600dcdc3c53cdea", + "actual": "0461ea66b1c324ec38e9e833ee8eb19b95a24c3f478b094fb600dcdc3c53cdea" + }, + "hfd_bridge/genmo.py": { + "expected": "3c898d2828267f38f8ba9f41aae62c4a52cbae87cf63dd2cef2b6634f7160060", + "actual": "3c898d2828267f38f8ba9f41aae62c4a52cbae87cf63dd2cef2b6634f7160060" + }, + "hfd_bridge/genmo_audio.py": { + "expected": "1cdfa6ebc19a75af7540b2280f5dc9c3ecc584ed1c386c434de32e3777020c6f", + "actual": "1cdfa6ebc19a75af7540b2280f5dc9c3ecc584ed1c386c434de32e3777020c6f" + }, + "hfd_bridge/clocks.py": { + "expected": "27054c0b2d3e7d4087eb8fc3c59c6bcf5a581ca2e3022341e306732dcbe92637", + "actual": "27054c0b2d3e7d4087eb8fc3c59c6bcf5a581ca2e3022341e306732dcbe92637" + }, + "hfd_bridge/semantics.py": { + "expected": "42ed59d58609095b8053d0598a9c575d52af9b8508e699162743cc11599f03b4", + "actual": "42ed59d58609095b8053d0598a9c575d52af9b8508e699162743cc11599f03b4" + }, + "hfd_bridge/reference_text.py": { + "expected": "50da93fed3be2bad05035bf978c9be41ba7b48aaf3ce7c43a879519e7386a121", + "actual": "50da93fed3be2bad05035bf978c9be41ba7b48aaf3ce7c43a879519e7386a121" + }, + "hfd_bridge/reference_runtime.py": { + "expected": "2b7fe129152d0f7ea3a4a4e3f5fc9ae799e805138a0d76487e9f59065fe7d975", + "actual": "2b7fe129152d0f7ea3a4a4e3f5fc9ae799e805138a0d76487e9f59065fe7d975" + }, + "hfd_bridge/resume.py": { + "expected": "e4e373802e0dd21e39fa2d899eff524080ce1735ca86eefbf4ef2080f170ffbf", + "actual": "e4e373802e0dd21e39fa2d899eff524080ce1735ca86eefbf4ef2080f170ffbf" + }, + "hfd_bridge/resume_lightning.py": { + "expected": "69e4c6bf11a199f3a979227eb551870c44e371fcae0af924ec64d2d58d3ce7a6", + "actual": "69e4c6bf11a199f3a979227eb551870c44e371fcae0af924ec64d2d58d3ce7a6" + }, + "hfd_bridge/index_cache.py": { + "expected": "d75cbcb0105448c61c841e525a1f17e63c55c83ad9201104e8118e22f7728b17", + "actual": "d75cbcb0105448c61c841e525a1f17e63c55c83ad9201104e8118e22f7728b17" + }, + "hfd_bridge/genmo_data.py": { + "expected": "4b102d1eecb0af9757a76d9600ede345c3f21f92166276a22cbf8e144fa9a38c", + "actual": "4b102d1eecb0af9757a76d9600ede345c3f21f92166276a22cbf8e144fa9a38c" + }, + "hfd_bridge/genmo_validation.py": { + "expected": "fa265cf9df51a4b36db977be945547dd78bdf105db9aed8ccba940b1a977f5e4", + "actual": "fa265cf9df51a4b36db977be945547dd78bdf105db9aed8ccba940b1a977f5e4" + }, + "hfd_bridge/kimodo_data.py": { + "expected": "35b3b7d8fc8f36fe65eb311e6e65e667fe1fdbcdfc23e58f37fb75c1c9e9feae", + "actual": "35b3b7d8fc8f36fe65eb311e6e65e667fe1fdbcdfc23e58f37fb75c1c9e9feae" + }, + "hfd_bridge/kimodo.py": { + "expected": "081a1047b853a2ee143aad2a5699f33cbecd0b4f93a07ad085760e16665f20ce", + "actual": "081a1047b853a2ee143aad2a5699f33cbecd0b4f93a07ad085760e16665f20ce" + }, + "hfd_bridge/kimodo_recovery.py": { + "expected": "b9177ff13b55b753430dd14d46778b723d2a03ce8927313c246c5d7ce780809b", + "actual": "b9177ff13b55b753430dd14d46778b723d2a03ce8927313c246c5d7ce780809b" + }, + "hfd_bridge/kimodo_closed_loop.py": { + "expected": "38d7e022b1450a715b35542cd23a70b1491a6dbd307dbb4be730b9f562100f91", + "actual": "38d7e022b1450a715b35542cd23a70b1491a6dbd307dbb4be730b9f562100f91" + }, + "hfd_bridge/kimodo_causal_rep.py": { + "expected": "c0c40003a3d36ec35e8490fa1165ac88723178b15a0e118d2d1d4c9f7ed7c537", + "actual": "c0c40003a3d36ec35e8490fa1165ac88723178b15a0e118d2d1d4c9f7ed7c537" + }, + "hfd_bridge/kimodo_diffusion_training.py": { + "expected": "078fe8cf1ee39d910a6ca5ef94d67f20c4398cfe505895a089acb80c05336539", + "actual": "078fe8cf1ee39d910a6ca5ef94d67f20c4398cfe505895a089acb80c05336539" + }, + "hfd_bridge/kimodo_text_worker.py": { + "expected": "a275c121a9364ae85b6233a67f3acb63ceddb2c9da490c2ea7d476f4cc710751", + "actual": "a275c121a9364ae85b6233a67f3acb63ceddb2c9da490c2ea7d476f4cc710751" + }, + "hfd_bridge/kimodo_full_data.py": { + "expected": "c62b4bb2673f1875792edb3403347f76bb52fa2360d92a9c1091eae10d43f905", + "actual": "c62b4bb2673f1875792edb3403347f76bb52fa2360d92a9c1091eae10d43f905" + }, + "hfd_bridge/kimodo_checkpoint.py": { + "expected": "b9cd24df225c91a6212c354983e3f51acb495373747f303c6096000af37f7d59", + "actual": "b9cd24df225c91a6212c354983e3f51acb495373747f303c6096000af37f7d59" + }, + "hfd_bridge/kimodo_finetuning.py": { + "expected": "a2c54f04af2b516c14d00025c8b695093ff09da6f04b0b73eb3269817bc0a544", + "actual": "a2c54f04af2b516c14d00025c8b695093ff09da6f04b0b73eb3269817bc0a544" + }, + "hfd_bridge/reference_assets.py": { + "expected": "e977a1d804807629fe925d511acdd25b488a11acc3767c0234020af8f699a43c", + "actual": "e977a1d804807629fe925d511acdd25b488a11acc3767c0234020af8f699a43c" + }, + "hfd_bridge/omg_model_loader.py": { + "expected": "f972080dd6140e1574124be615c9351001b51a73c5774b63bd1c063671bd56aa", + "actual": "f972080dd6140e1574124be615c9351001b51a73c5774b63bd1c063671bd56aa" + }, + "hfd_bridge/genmo_geometry.py": { + "expected": "4d3e912a858b8d3d4ddf73941c3120d5304ed70187c02fc969730816e155b2cb", + "actual": "4d3e912a858b8d3d4ddf73941c3120d5304ed70187c02fc969730816e155b2cb" + }, + "hfd_bridge/genmo_training.py": { + "expected": "556100608e466ca801abd623d24af30eebe7f9e3ac598c6709f813f07d5d45e9", + "actual": "556100608e466ca801abd623d24af30eebe7f9e3ac598c6709f813f07d5d45e9" + }, + "hfd_bridge/genmo_prefetch.py": { + "expected": "0f00c9c5c951f53a0b55cad0c71d470c9312381e75ec31f756f606cf20309141", + "actual": "0f00c9c5c951f53a0b55cad0c71d470c9312381e75ec31f756f606cf20309141" + }, + "hfd_bridge/hymotion_rep.py": { + "expected": "ee2f8282720d311b98d3ceb8cc72f63219622b5f61a138af6eb3221181a89f7a", + "actual": "ee2f8282720d311b98d3ceb8cc72f63219622b5f61a138af6eb3221181a89f7a" + }, + "hfd_bridge/hymotion_model.py": { + "expected": "ed29d2d03b247ed681a09dc230697885c5ab8fa8b30f82c4a827aa8fe0b41597", + "actual": "ed29d2d03b247ed681a09dc230697885c5ab8fa8b30f82c4a827aa8fe0b41597" + }, + "hfd_bridge/hymotion_data.py": { + "expected": "79503602a9a4de4f67931fa944a198d6f438b531a363ffbe545ecc4dd54f97e7", + "actual": "79503602a9a4de4f67931fa944a198d6f438b531a363ffbe545ecc4dd54f97e7" + }, + "hfd_bridge/hymotion_text.py": { + "expected": "f0238ffc3656e2a90a26360c5e9bc8c6de3d2d6be0af6e87787c83237914e98b", + "actual": "f0238ffc3656e2a90a26360c5e9bc8c6de3d2d6be0af6e87787c83237914e98b" + }, + "hfd_bridge/hymotion_rollout.py": { + "expected": "6e27c4e24bd42a211df7be2a0271337af4e658d66e5d303633912bbd40313517", + "actual": "6e27c4e24bd42a211df7be2a0271337af4e658d66e5d303633912bbd40313517" + }, + "hfd_bridge/hymotion_validation.py": { + "expected": "6043090f19315c9fde677168fed06dca1fb390b559a7b9668e0ba46799074cd2", + "actual": "6043090f19315c9fde677168fed06dca1fb390b559a7b9668e0ba46799074cd2" + }, + "hfd_bridge/hymotion_training.py": { + "expected": "792b76cfc35f58b0c7c94b40a660f00c4971fff1d978ffd0c923b9f26a5fa733", + "actual": "792b76cfc35f58b0c7c94b40a660f00c4971fff1d978ffd0c923b9f26a5fa733" + }, + "hfd_bridge/human_eval.py": { + "expected": "c943c332315f16f41cb53beee8303f2d94a4b522e678cfd346e8a6c9ec203b71", + "actual": "c943c332315f16f41cb53beee8303f2d94a4b522e678cfd346e8a6c9ec203b71" + }, + "hfd_bridge/omg_rollout.py": { + "expected": "24ecdec4a2c698bd8cdc2b4ac6e1c0ad960c252b5a9f9a15d57a7716545fc4d7", + "actual": "24ecdec4a2c698bd8cdc2b4ac6e1c0ad960c252b5a9f9a15d57a7716545fc4d7" + }, + "hfd_bridge/final_split.py": { + "expected": "d0f0adef291d17d3b73477258b5de74bb8ed477ca1952c9b83ddb115efd34305", + "actual": "d0f0adef291d17d3b73477258b5de74bb8ed477ca1952c9b83ddb115efd34305" + }, + "hfd_bridge/final_third_party.py": { + "expected": "3926a0f72d419bb5006d8a62136a8f5ffd5ddca439c39dd7768f7cf03422d716", + "actual": "3926a0f72d419bb5006d8a62136a8f5ffd5ddca439c39dd7768f7cf03422d716" + }, + "hfd_bridge/replay_numerics.py": { + "expected": "827c70aa9b2308ac33b6fccee2c4dc88df58d6b89d902d8c215adfdf918ab231", + "actual": "827c70aa9b2308ac33b6fccee2c4dc88df58d6b89d902d8c215adfdf918ab231" + }, + "hfd_bridge/textop_text_cache.py": { + "expected": "3d5c8a187e461144aa97cc539661e6e90b475800379c0c096b243bcd337a0e38", + "actual": "3d5c8a187e461144aa97cc539661e6e90b475800379c0c096b243bcd337a0e38" + }, + "hfd_bridge/hot_assets.py": { + "expected": "43830324054a551ef4ed364ad45d9b2dfd8830b7dde2661f4404b1ffc8cabe10", + "actual": "43830324054a551ef4ed364ad45d9b2dfd8830b7dde2661f4404b1ffc8cabe10" + }, + "hfd_bridge/hymotion_autoregressive.py": { + "expected": "cfbcc9329760364b22c1871fc69c42b77787ed0ab622a2c5fc1a6bed93479d86", + "actual": "cfbcc9329760364b22c1871fc69c42b77787ed0ab622a2c5fc1a6bed93479d86" + }, + "hfd_bridge/genmo_runtime.py": { + "expected": "b4bc26373388974ca50a5e76d382b210950b805ef19b514fcc90adbf555bdc2e", + "actual": "b4bc26373388974ca50a5e76d382b210950b805ef19b514fcc90adbf555bdc2e" + }, + "hfd_bridge/genmo_replay_audit.py": { + "expected": "fb8ac4ac51489ebe9faca4b08f6f2159aea4bb33040ea6c04c7c99c6641c4ae2", + "actual": "fb8ac4ac51489ebe9faca4b08f6f2159aea4bb33040ea6c04c7c99c6641c4ae2" + }, + "hfd_bridge/genmo_safety.py": { + "expected": "f5f20e57cef26e62ec4bb5c0eb1e9d74d30e814b1740cef9f36008936e234158", + "actual": "f5f20e57cef26e62ec4bb5c0eb1e9d74d30e814b1740cef9f36008936e234158" + }, + "hfd_bridge/hymotion_metrics.py": { + "expected": "48d79ed03b833fc948eeb466890aefe6bdb66ed88860f6fbd49b1a5ea0c2c04f", + "actual": "48d79ed03b833fc948eeb466890aefe6bdb66ed88860f6fbd49b1a5ea0c2c04f" + }, + "hfd_bridge/hymotion_evaluation.py": { + "expected": "fe72667281f04a55bc8c88375b7ca3cee1f1b6de2c23fcd7d45f1e73c1530a99", + "actual": "fe72667281f04a55bc8c88375b7ca3cee1f1b6de2c23fcd7d45f1e73c1530a99" + }, + "hfd_bridge/omg_scaling.py": { + "expected": "0ed0b1e0a8c72e039e5c461b193595b4e50da4bfc2f155ec3d4754923af51360", + "actual": "0ed0b1e0a8c72e039e5c461b193595b4e50da4bfc2f155ec3d4754923af51360" + }, + "upstream/omg/download_data.py": { + "expected": "a247118ce874864affc5153a23d94788871d3b6d327bcc33ceeba45279d7732c", + "actual": "a247118ce874864affc5153a23d94788871d3b6d327bcc33ceeba45279d7732c" + }, + "upstream/omg/download_models.py": { + "expected": "811132811b35321174adc69a754cf64f9437eeba5d19f6ad02625ef9d9f73b57", + "actual": "811132811b35321174adc69a754cf64f9437eeba5d19f6ad02625ef9d9f73b57" + }, + "upstream/omg/a.py": { + "expected": "66e1c68c0a631d6fef0d7d13948299671baf72ec813fe025c1cac1e222fd4c63", + "actual": "66e1c68c0a631d6fef0d7d13948299671baf72ec813fe025c1cac1e222fd4c63" + }, + "upstream/omg/smoke.py": { + "expected": "8008c82e92250f148512e3127b939df3aded7025a89473738d024444d21edcfd", + "actual": "8008c82e92250f148512e3127b939df3aded7025a89473738d024444d21edcfd" + }, + "upstream/omg/validate_model.py": { + "expected": "36023f739cd8dfe24179f9247b5de04ca425d7a0d73c99b166bbeaea778a975f", + "actual": "36023f739cd8dfe24179f9247b5de04ca425d7a0d73c99b166bbeaea778a975f" + }, + "upstream/omg/validate_setup.py": { + "expected": "4e086d4615c621386cb54885aad0ede341c24a71a1668c3e5386f368c02f3741", + "actual": "4e086d4615c621386cb54885aad0ede341c24a71a1668c3e5386f368c02f3741" + }, + "upstream/omg/paper_table.py": { + "expected": "97c16c1c044639362f46cac439159384a1d79dc80a90c839c5e2ce471d8060d2", + "actual": "97c16c1c044639362f46cac439159384a1d79dc80a90c839c5e2ce471d8060d2" + }, + "upstream/omg/paper_main_table.py": { + "expected": "8f95f11b00882e4574df37e2f44d5843a15073004f0b9f78450895c65939196f", + "actual": "8f95f11b00882e4574df37e2f44d5843a15073004f0b9f78450895c65939196f" + }, + "upstream/omg/scripts/prepare_omg_inputs.py": { + "expected": "99bbc364b5f380ae7c2cc0fa53657722f896d3613d8b98937881d06197beccb4", + "actual": "99bbc364b5f380ae7c2cc0fa53657722f896d3613d8b98937881d06197beccb4" + }, + "upstream/omg/scripts/extract_omg_paper_row.py": { + "expected": "f64cb31b5a19efba6f3e1b1771d3d5175f6610392a588d118014e98790b82503", + "actual": "f64cb31b5a19efba6f3e1b1771d3d5175f6610392a588d118014e98790b82503" + }, + "upstream/omg/scripts/psi0_omg_generate.py": { + "expected": "38ccbd4c8f1d62b470193ae6e3598445f22408e2df6edc6b267ed46b4873f579", + "actual": "38ccbd4c8f1d62b470193ae6e3598445f22408e2df6edc6b267ed46b4873f579" + }, + "upstream/omg/scripts/verify_omg_artifact_match.py": { + "expected": "46815bd446b956583ae7feeb0ed03a58b6dde8cc78ecb4a1d4fd1a3a6b737ee7", + "actual": "46815bd446b956583ae7feeb0ed03a58b6dde8cc78ecb4a1d4fd1a3a6b737ee7" + }, + "upstream/omg/src/omg/__init__.py": { + "expected": "e06eff3bd6a146e0082518eb791164e26b80b548f9bdcd21c9f0aca1220bca34", + "actual": "e06eff3bd6a146e0082518eb791164e26b80b548f9bdcd21c9f0aca1220bca34" + }, + "upstream/omg/src/omg/callbacks/__init__.py": { + "expected": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", + "actual": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855" + }, + "upstream/omg/src/omg/callbacks/divergence_guard.py": { + "expected": "c1ec2773dddda3bd800247bca0abd3eca46aeae3c78ebae8050b98739292b61a", + "actual": "c1ec2773dddda3bd800247bca0abd3eca46aeae3c78ebae8050b98739292b61a" + }, + "upstream/omg/src/omg/callbacks/global_step_progress.py": { + "expected": "bd21180de0140a38fa2601f51b56c94aef346c757c140786e88b907720f3eb15", + "actual": "bd21180de0140a38fa2601f51b56c94aef346c757c140786e88b907720f3eb15" + }, + "upstream/omg/src/omg/callbacks/grad_norm.py": { + "expected": "c41e7e0bac01c984c85c38baf6de4b9e10ded95875a79ddc1c9d85ded5b9b742", + "actual": "c41e7e0bac01c984c85c38baf6de4b9e10ded95875a79ddc1c9d85ded5b9b742" + }, + "upstream/omg/src/omg/callbacks/speed_timer.py": { + "expected": "29ae0ec76a0545576bc31f240a4beab74f5c5e9ae50a0078c1ed46b8326fc360", + "actual": "29ae0ec76a0545576bc31f240a4beab74f5c5e9ae50a0078c1ed46b8326fc360" + }, + "upstream/omg/src/omg/callbacks/training_diagnostics.py": { + "expected": "6e348e5cca9e690379cdb7a3364678b666d7c53d0a7405d5269ca713574a6917", + "actual": "6e348e5cca9e690379cdb7a3364678b666d7c53d0a7405d5269ca713574a6917" + }, + "upstream/omg/src/omg/cli/__init__.py": { + "expected": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", + "actual": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855" + }, + "upstream/omg/src/omg/cli/data/__init__.py": { + "expected": "052be341b9c1b05585d3a27073a741d768974cfdb2a91659a687ed1d3d6291f9", + "actual": "052be341b9c1b05585d3a27073a741d768974cfdb2a91659a687ed1d3d6291f9" + }, + "upstream/omg/src/omg/cli/data/materialize_episode_cache.py": { + "expected": "c8e61cbcf17f55e17b47ad2a3e5c2aa0919d9f5bcc2a93edbf5a519e79f9793a", + "actual": "c8e61cbcf17f55e17b47ad2a3e5c2aa0919d9f5bcc2a93edbf5a519e79f9793a" + }, + "upstream/omg/src/omg/cli/data/validate_episode_cache.py": { + "expected": "6840e28c774964583d88c6402c201ab36c0150ec67139403a66782364173c878", + "actual": "6840e28c774964583d88c6402c201ab36c0150ec67139403a66782364173c878" + }, + "upstream/omg/src/omg/cli/generation/__init__.py": { + "expected": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", + "actual": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855" + }, + "upstream/omg/src/omg/cli/generation/compute_stats.py": { + "expected": "849ed1ac16612e23e09a887e5a030c5e0c496610e9afe882ce98e99d570bd6ca", + "actual": "849ed1ac16612e23e09a887e5a030c5e0c496610e9afe882ce98e99d570bd6ca" + }, + "upstream/omg/src/omg/cli/generation/train.py": { + "expected": "a39cc0c6cfccd7b6660bbd9ab4f88f90bc634bb95ce30e22492c8a6cb74a7c8c", + "actual": "a39cc0c6cfccd7b6660bbd9ab4f88f90bc634bb95ce30e22492c8a6cb74a7c8c" + }, + "upstream/omg/src/omg/cli/generation/benchmark.py": { + "expected": "ff813f344953da4b02388837f5a510230f8d4430393a88f205261faaf28c5dee", + "actual": "ff813f344953da4b02388837f5a510230f8d4430393a88f205261faaf28c5dee" + }, + "upstream/omg/src/omg/cli/generation/physical_benchmark.py": { + "expected": "f48b2b9d2d312c44d0e33eebf0c7fd95d8f2ce65a871fe253ef8ff9ceee30dd0", + "actual": "f48b2b9d2d312c44d0e33eebf0c7fd95d8f2ce65a871fe253ef8ff9ceee30dd0" + }, + "upstream/omg/src/omg/cli/generation/generate.py": { + "expected": "66eda839732e76a3b0fc39a767e6a238aa93bbf13ef2f3ddb6e1f9021948bc4a", + "actual": "66eda839732e76a3b0fc39a767e6a238aa93bbf13ef2f3ddb6e1f9021948bc4a" + }, + "upstream/omg/src/omg/cli/generation/export_onnx.py": { + "expected": "9cacee46fea5e18f8a9f9ffd52b25110f89e80be20d21acce85fe066392c00c1", + "actual": "9cacee46fea5e18f8a9f9ffd52b25110f89e80be20d21acce85fe066392c00c1" + }, + "upstream/omg/src/omg/core/__init__.py": { + "expected": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", + "actual": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855" + }, + "upstream/omg/src/omg/core/logging.py": { + "expected": "86c57564f7d8614d71d95e0ce377724642c8c3aa44501d9c317c33bf9ba7dce7", + "actual": "86c57564f7d8614d71d95e0ce377724642c8c3aa44501d9c317c33bf9ba7dce7" + }, + "upstream/omg/src/omg/core/paths.py": { + "expected": "6bb254b7219e70b58fc8581479857bfb7ee77a9f8937faf5dcec784dd1173a4d", + "actual": "6bb254b7219e70b58fc8581479857bfb7ee77a9f8937faf5dcec784dd1173a4d" + }, + "upstream/omg/src/omg/core/tensor.py": { + "expected": "998e6f24948983c15f3257f055a452e07b96cc0f13b357cb4881f4545bddb7d0", + "actual": "998e6f24948983c15f3257f055a452e07b96cc0f13b357cb4881f4545bddb7d0" + }, + "upstream/omg/src/omg/data/__init__.py": { + "expected": "4e8cd9d824b5b834a673011d1cd95cf0227a1deb60ee917ba85ec364d51db21f", + "actual": "4e8cd9d824b5b834a673011d1cd95cf0227a1deb60ee917ba85ec364d51db21f" + }, + "upstream/omg/src/omg/data/datamodule.py": { + "expected": "94cbaf84b019d559dd7810494694c156803d318f23a5ec95549315c59ea795d6", + "actual": "94cbaf84b019d559dd7810494694c156803d318f23a5ec95549315c59ea795d6" + }, + "upstream/omg/src/omg/data/episode_cache.py": { + "expected": "96639991cb75a3db602c608d929e4f742caef615166198e66731c406f068207e", + "actual": "96639991cb75a3db602c608d929e4f742caef615166198e66731c406f068207e" + }, + "upstream/omg/src/omg/data/episode_cache_inspect.py": { + "expected": "b7f43c2a5f7a1c63dcb9c243e79018ebff3ecce056fba22ac9b1948e754feb20", + "actual": "b7f43c2a5f7a1c63dcb9c243e79018ebff3ecce056fba22ac9b1948e754feb20" + }, + "upstream/omg/src/omg/data/lerobot_dataset.py": { + "expected": "78172eb344a37cab8457133a957eb25691ad78054f71db0c39171f241be12831", + "actual": "78172eb344a37cab8457133a957eb25691ad78054f71db0c39171f241be12831" + }, + "upstream/omg/src/omg/data/windowing.py": { + "expected": "88710f0c24d6c696670b702b9ce77609c59acd2106389fa769a00d39a98f2afb", + "actual": "88710f0c24d6c696670b702b9ce77609c59acd2106389fa769a00d39a98f2afb" + }, + "upstream/omg/src/omg/generation/__init__.py": { + "expected": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", + "actual": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855" + }, + "upstream/omg/src/omg/generation/architecture.py": { + "expected": "f77548e52394c78b813ad36dcd5fb8c8981ba393e5c723126491df6c171aeabf", + "actual": "f77548e52394c78b813ad36dcd5fb8c8981ba393e5c723126491df6c171aeabf" + }, + "upstream/omg/src/omg/generation/conditions/__init__.py": { + "expected": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", + "actual": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855" + }, + "upstream/omg/src/omg/generation/conditions/t5.py": { + "expected": "67c0f834106ff860e612a978bdba6ef503d2cf0dfb24e831af0bba723e95c890", + "actual": "67c0f834106ff860e612a978bdba6ef503d2cf0dfb24e831af0bba723e95c890" + }, + "upstream/omg/src/omg/generation/denoisers/__init__.py": { + "expected": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", + "actual": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855" + }, + "upstream/omg/src/omg/generation/denoisers/interface.py": { + "expected": "c1f8b5f2a93e6cf0a4b5df730f9f744c95d9b58bc5245e3cb15176a61065e1dd", + "actual": "c1f8b5f2a93e6cf0a4b5df730f9f744c95d9b58bc5245e3cb15176a61065e1dd" + }, + "upstream/omg/src/omg/generation/denoisers/transformer.py": { + "expected": "ae968a16ba6193607721f5c0c1910122818797407df3ee0f0fafa3728c0e398a", + "actual": "ae968a16ba6193607721f5c0c1910122818797407df3ee0f0fafa3728c0e398a" + }, + "upstream/omg/src/omg/generation/diffusion/__init__.py": { + "expected": "fab4af0ec5b7b38892ffb309d6073841de00f8f37e5b4c2da1657e1b30c92e1c", + "actual": "fab4af0ec5b7b38892ffb309d6073841de00f8f37e5b4c2da1657e1b30c92e1c" + }, + "upstream/omg/src/omg/generation/diffusion/forcing.py": { + "expected": "935b2f38db1d1d3c496df0b03cadaeb032ef1539522169f2731b87a8724362b5", + "actual": "935b2f38db1d1d3c496df0b03cadaeb032ef1539522169f2731b87a8724362b5" + }, + "upstream/omg/src/omg/generation/diffusion/guided.py": { + "expected": "3fd42c7ae9d01497223b08421ae3becc0a7be81846c1cf08c8bb944d8949b870", + "actual": "3fd42c7ae9d01497223b08421ae3becc0a7be81846c1cf08c8bb944d8949b870" + }, + "upstream/omg/src/omg/generation/diffusion/objectives.py": { + "expected": "db05bcbc32268e5d21298e19130e9e8f4b23eff7d1b679886306cd7c94195675", + "actual": "db05bcbc32268e5d21298e19130e9e8f4b23eff7d1b679886306cd7c94195675" + }, + "upstream/omg/src/omg/generation/diffusion/schedules.py": { + "expected": "a06c3f3ab1d81e07f680d44f8fb9ef56e8bd2a6e4c5e4ac911c316217f6e92e0", + "actual": "a06c3f3ab1d81e07f680d44f8fb9ef56e8bd2a6e4c5e4ac911c316217f6e92e0" + }, + "upstream/omg/src/omg/generation/export/__init__.py": { + "expected": "0ee7ed60aa43848dad0a4ba0e5dd29d815aab6dc37f4ae1c46d8e56a2ebc1770", + "actual": "0ee7ed60aa43848dad0a4ba0e5dd29d815aab6dc37f4ae1c46d8e56a2ebc1770" + }, + "upstream/omg/src/omg/generation/export/onnx.py": { + "expected": "168ad808867669b69cd69a419d5b8d2c3fd124c69c20daeca53b6fc59a0a21dd", + "actual": "168ad808867669b69cd69a419d5b8d2c3fd124c69c20daeca53b6fc59a0a21dd" + }, + "upstream/omg/src/omg/generation/losses/__init__.py": { + "expected": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", + "actual": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855" + }, + "upstream/omg/src/omg/generation/losses/contact.py": { + "expected": "75927b00fc5584988f2ecedbcfbba9af3ba80206966b0a91642ec133db5e4267", + "actual": "75927b00fc5584988f2ecedbcfbba9af3ba80206966b0a91642ec133db5e4267" + }, + "upstream/omg/src/omg/generation/losses/fk.py": { + "expected": "75927b00fc5584988f2ecedbcfbba9af3ba80206966b0a91642ec133db5e4267", + "actual": "75927b00fc5584988f2ecedbcfbba9af3ba80206966b0a91642ec133db5e4267" + }, + "upstream/omg/src/omg/generation/losses/motion.py": { + "expected": "00ee200046658dafcae2b9ec439ed0c3203ab533f4dd9d4a8475541df74c2fa5", + "actual": "00ee200046658dafcae2b9ec439ed0c3203ab533f4dd9d4a8475541df74c2fa5" + }, + "upstream/omg/src/omg/generation/losses/terrain.py": { + "expected": "75927b00fc5584988f2ecedbcfbba9af3ba80206966b0a91642ec133db5e4267", + "actual": "75927b00fc5584988f2ecedbcfbba9af3ba80206966b0a91642ec133db5e4267" + }, + "upstream/omg/src/omg/generation/metrics/__init__.py": { + "expected": "f686c86d57bab24aa1dd27cfae2bbab88757d7909e74cd7d4b9525a892ba025e", + "actual": "f686c86d57bab24aa1dd27cfae2bbab88757d7909e74cd7d4b9525a892ba025e" + }, + "upstream/omg/src/omg/generation/metrics/physical.py": { + "expected": "7aa2d8fc7d29b9d3ec836c3d23d0b5a7fc64a1eea5359c7300f6e6e191904f29", + "actual": "7aa2d8fc7d29b9d3ec836c3d23d0b5a7fc64a1eea5359c7300f6e6e191904f29" + }, + "upstream/omg/src/omg/generation/models/__init__.py": { + "expected": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", + "actual": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855" + }, + "upstream/omg/src/omg/generation/models/motion_generator.py": { + "expected": "05947c8e73646f634e9bbedf0a60f2becaf1ded21b9d1093f1084eaaefef2a3b", + "actual": "05947c8e73646f634e9bbedf0a60f2becaf1ded21b9d1093f1084eaaefef2a3b" + }, + "upstream/omg/src/omg/motion/__init__.py": { + "expected": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", + "actual": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855" + }, + "upstream/omg/src/omg/motion/feature_codec.py": { + "expected": "3d148816baf239dbc938210c959954c3c2fac4c6d9914ad48b851976ed7e2960", + "actual": "3d148816baf239dbc938210c959954c3c2fac4c6d9914ad48b851976ed7e2960" + }, + "upstream/omg/src/omg/motion/representation.py": { + "expected": "457328078b2a2652e06c39e5fa06c8a0d8ba9223dcf24e508a33cdd1806f45a7", + "actual": "457328078b2a2652e06c39e5fa06c8a0d8ba9223dcf24e508a33cdd1806f45a7" + }, + "upstream/omg/src/omg/pipeline/__init__.py": { + "expected": "3806048000a9bdc918ace57c68aabcd3c56bb503c4998b103d1593553a23a895", + "actual": "3806048000a9bdc918ace57c68aabcd3c56bb503c4998b103d1593553a23a895" + }, + "upstream/omg/src/omg/pipeline/audio.py": { + "expected": "d88a0b51fd72d80fdf7da1344082e7799f481fb114a246ed4e66c0b9745f3d09", + "actual": "d88a0b51fd72d80fdf7da1344082e7799f481fb114a246ed4e66c0b9745f3d09" + }, + "upstream/omg/src/omg/pipeline/condition_sequence.py": { + "expected": "e831393b0b843219520ee8a836b7cbc0b406295e60947cba1a443f1f78a81377", + "actual": "e831393b0b843219520ee8a836b7cbc0b406295e60947cba1a443f1f78a81377" + }, + "upstream/omg/src/omg/pipeline/human.py": { + "expected": "d5be4338bc69308450c18331da4cf9a93fc9e02bd5596dce92a3e58248ff14d0", + "actual": "d5be4338bc69308450c18331da4cf9a93fc9e02bd5596dce92a3e58248ff14d0" + }, + "upstream/omg/src/omg/pipeline/planner.py": { + "expected": "fa7aa07eebf9f56287301a6ab13fb7a42bf8b310e152b0916033019d7d55ce0c", + "actual": "fa7aa07eebf9f56287301a6ab13fb7a42bf8b310e152b0916033019d7d55ce0c" + }, + "upstream/omg/src/omg/robots/__init__.py": { + "expected": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", + "actual": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855" + }, + "upstream/omg/src/omg/robots/g1/__init__.py": { + "expected": "dc46237fccb843612112274586e9e00dee6a0e472f4ce6b93da267b64f8f4c19", + "actual": "dc46237fccb843612112274586e9e00dee6a0e472f4ce6b93da267b64f8f4c19" + }, + "upstream/omg/src/omg/robots/g1/constants.py": { + "expected": "1f2d0fd495de8765349ad24bdb8e8b327ffdf80c16b13c665fabfe8e9b1a39ea", + "actual": "1f2d0fd495de8765349ad24bdb8e8b327ffdf80c16b13c665fabfe8e9b1a39ea" + }, + "upstream/omg/src/omg/robots/g1/kinematics.py": { + "expected": "c8adbae0d96ef85ba948e27f1fb16bed35d8529c08a6d71d1943d31644585db4", + "actual": "c8adbae0d96ef85ba948e27f1fb16bed35d8529c08a6d71d1943d31644585db4" + }, + "upstream/omg/src/omg/runtime/__init__.py": { + "expected": "9728869f63d93ada09a0f77fc10e98f2add27f4ab84e999d228becca102315fe", + "actual": "9728869f63d93ada09a0f77fc10e98f2add27f4ab84e999d228becca102315fe" + }, + "upstream/omg/src/omg/runtime/checkpoint_io.py": { + "expected": "1b9ce178d3e5c1bba39c77d3098ea7e64788d8c81c48acdebaec09cc1be88e5e", + "actual": "1b9ce178d3e5c1bba39c77d3098ea7e64788d8c81c48acdebaec09cc1be88e5e" + }, + "upstream/omg/src/omg/runtime/onnx_providers.py": { + "expected": "5ed87e442ebf9c56b4065fc011c89d7745ac5bf163158fbf0732db160d0e5673", + "actual": "5ed87e442ebf9c56b4065fc011c89d7745ac5bf163158fbf0732db160d0e5673" + }, + "upstream/omg/src/omg/utils/rotation_conversions.py": { + "expected": "627ead38f2e15b0001eea501f6be09e7b6c309c21f1938ef419f65f8e1eb7ea5", + "actual": "627ead38f2e15b0001eea501f6be09e7b6c309c21f1938ef419f65f8e1eb7ea5" + }, + "train.py": { + "expected": "61f49739e14275a671dc3343afcf2d6b225da547a952f963037d545afe680dd4", + "actual": "61f49739e14275a671dc3343afcf2d6b225da547a952f963037d545afe680dd4" + } +} \ No newline at end of file diff --git a/handoff-v1/inventory/omg-l-step-00098000.ckpt.json b/handoff-v1/inventory/omg-l-step-00098000.ckpt.json new file mode 100644 index 0000000000000000000000000000000000000000..a8dc18fc1741e8fbd60fa31403625492f2c70179 --- /dev/null +++ b/handoff-v1/inventory/omg-l-step-00098000.ckpt.json @@ -0,0 +1,3596 @@ +{ + "epoch": 0, + "global_step": 98000, + "pytorch-lightning_version": "2.6.6", + "state_dict": { + "denoiser.time_embed.proj.0.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.time_embed.proj.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.time_embed.proj.2.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.time_embed.proj.2.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.input_proj.weight": { + "shape": [ + 1152, + 1277 + ], + "dtype": "torch.float32" + }, + "denoiser.input_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.text_proj.weight": { + "shape": [ + 1152, + 768 + ], + "dtype": "torch.float32" + }, + "denoiser.text_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.extra_proj.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.extra_proj.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.extra_proj.1.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.extra_proj.1.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.norm_cross.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.norm_cross.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.cross_attn.in_proj_weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.cross_attn.in_proj_bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.cross_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.cross_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.norm_self.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.norm_self.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.self_attn.qkv.weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.self_attn.qkv.bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.self_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.self_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.norm_ff.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.norm_ff.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.ff.0.weight": { + "shape": [ + 4608, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.ff.0.bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.ff.3.weight": { + "shape": [ + 1152, + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.ff.3.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.audio_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.audio_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.audio_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.audio_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.human_motion_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.human_motion_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.human_motion_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.human_motion_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.norm_cross.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.norm_cross.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.cross_attn.in_proj_weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.cross_attn.in_proj_bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.cross_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.cross_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.norm_self.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.norm_self.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.self_attn.qkv.weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.self_attn.qkv.bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.self_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.self_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.norm_ff.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.norm_ff.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.ff.0.weight": { + "shape": [ + 4608, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.ff.0.bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.ff.3.weight": { + "shape": [ + 1152, + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.ff.3.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.audio_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.audio_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.audio_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.audio_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.human_motion_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.human_motion_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.human_motion_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.human_motion_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.norm_cross.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.norm_cross.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.cross_attn.in_proj_weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.cross_attn.in_proj_bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.cross_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.cross_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.norm_self.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.norm_self.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.self_attn.qkv.weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.self_attn.qkv.bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.self_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.self_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.norm_ff.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.norm_ff.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.ff.0.weight": { + "shape": [ + 4608, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.ff.0.bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.ff.3.weight": { + "shape": [ + 1152, + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.ff.3.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.audio_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.audio_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.audio_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.audio_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.human_motion_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.human_motion_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.human_motion_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.human_motion_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.norm_cross.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.norm_cross.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.cross_attn.in_proj_weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.cross_attn.in_proj_bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.cross_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.cross_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.norm_self.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.norm_self.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.self_attn.qkv.weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.self_attn.qkv.bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.self_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.self_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.norm_ff.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.norm_ff.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.ff.0.weight": { + "shape": [ + 4608, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.ff.0.bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.ff.3.weight": { + "shape": [ + 1152, + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.ff.3.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.audio_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.audio_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.audio_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.audio_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.human_motion_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.human_motion_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.human_motion_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.human_motion_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.norm_cross.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.norm_cross.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.cross_attn.in_proj_weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.cross_attn.in_proj_bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.cross_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.cross_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.norm_self.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.norm_self.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.self_attn.qkv.weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.self_attn.qkv.bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.self_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.self_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.norm_ff.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.norm_ff.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.ff.0.weight": { + "shape": [ + 4608, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.ff.0.bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.ff.3.weight": { + "shape": [ + 1152, + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.ff.3.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.audio_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.audio_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.audio_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.audio_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.human_motion_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.human_motion_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.human_motion_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.human_motion_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.norm_cross.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.norm_cross.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.cross_attn.in_proj_weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.cross_attn.in_proj_bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.cross_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.cross_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.norm_self.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.norm_self.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.self_attn.qkv.weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.self_attn.qkv.bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.self_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.self_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.norm_ff.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.norm_ff.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.ff.0.weight": { + "shape": [ + 4608, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.ff.0.bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.ff.3.weight": { + "shape": [ + 1152, + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.ff.3.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.audio_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.audio_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.audio_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.audio_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.human_motion_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.human_motion_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.human_motion_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.human_motion_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.norm_cross.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.norm_cross.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.cross_attn.in_proj_weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.cross_attn.in_proj_bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.cross_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.cross_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.norm_self.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.norm_self.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.self_attn.qkv.weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.self_attn.qkv.bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.self_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.self_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.norm_ff.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.norm_ff.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.ff.0.weight": { + "shape": [ + 4608, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.ff.0.bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.ff.3.weight": { + "shape": [ + 1152, + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.ff.3.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.audio_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.audio_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.audio_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.audio_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.human_motion_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.human_motion_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.human_motion_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.human_motion_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.norm_cross.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.norm_cross.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.cross_attn.in_proj_weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.cross_attn.in_proj_bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.cross_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.cross_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.norm_self.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.norm_self.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.self_attn.qkv.weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.self_attn.qkv.bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.self_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.self_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.norm_ff.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.norm_ff.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.ff.0.weight": { + "shape": [ + 4608, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.ff.0.bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.ff.3.weight": { + "shape": [ + 1152, + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.ff.3.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.audio_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.audio_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.audio_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.audio_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.human_motion_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.human_motion_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.human_motion_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.human_motion_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.norm_cross.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.norm_cross.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.cross_attn.in_proj_weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.cross_attn.in_proj_bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.cross_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.cross_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.norm_self.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.norm_self.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.self_attn.qkv.weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.self_attn.qkv.bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.self_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.self_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.norm_ff.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.norm_ff.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.ff.0.weight": { + "shape": [ + 4608, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.ff.0.bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.ff.3.weight": { + "shape": [ + 1152, + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.ff.3.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.audio_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.audio_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.audio_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.audio_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.human_motion_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.human_motion_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.human_motion_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.human_motion_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.norm_cross.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.norm_cross.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.cross_attn.in_proj_weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.cross_attn.in_proj_bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.cross_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.cross_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.norm_self.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.norm_self.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.self_attn.qkv.weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.self_attn.qkv.bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.self_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.self_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.norm_ff.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.norm_ff.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.ff.0.weight": { + "shape": [ + 4608, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.ff.0.bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.ff.3.weight": { + "shape": [ + 1152, + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.ff.3.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.audio_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.audio_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.audio_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.audio_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.human_motion_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.human_motion_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.human_motion_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.human_motion_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.norm_cross.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.norm_cross.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.cross_attn.in_proj_weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.cross_attn.in_proj_bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.cross_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.cross_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.norm_self.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.norm_self.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.self_attn.qkv.weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.self_attn.qkv.bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.self_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.self_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.norm_ff.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.norm_ff.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.ff.0.weight": { + "shape": [ + 4608, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.ff.0.bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.ff.3.weight": { + "shape": [ + 1152, + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.ff.3.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.audio_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.audio_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.audio_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.audio_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.human_motion_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.human_motion_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.human_motion_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.human_motion_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.norm_cross.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.norm_cross.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.cross_attn.in_proj_weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.cross_attn.in_proj_bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.cross_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.cross_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.norm_self.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.norm_self.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.self_attn.qkv.weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.self_attn.qkv.bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.self_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.self_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.norm_ff.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.norm_ff.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.ff.0.weight": { + "shape": [ + 4608, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.ff.0.bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.ff.3.weight": { + "shape": [ + 1152, + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.ff.3.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.audio_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.audio_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.audio_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.audio_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.human_motion_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.human_motion_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.human_motion_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.human_motion_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.norm_cross.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.norm_cross.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.cross_attn.in_proj_weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.cross_attn.in_proj_bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.cross_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.cross_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.norm_self.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.norm_self.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.self_attn.qkv.weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.self_attn.qkv.bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.self_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.self_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.norm_ff.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.norm_ff.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.ff.0.weight": { + "shape": [ + 4608, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.ff.0.bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.ff.3.weight": { + "shape": [ + 1152, + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.ff.3.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.audio_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.audio_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.audio_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.audio_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.human_motion_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.human_motion_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.human_motion_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.human_motion_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.norm_cross.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.norm_cross.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.cross_attn.in_proj_weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.cross_attn.in_proj_bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.cross_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.cross_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.norm_self.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.norm_self.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.self_attn.qkv.weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.self_attn.qkv.bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.self_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.self_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.norm_ff.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.norm_ff.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.ff.0.weight": { + "shape": [ + 4608, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.ff.0.bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.ff.3.weight": { + "shape": [ + 1152, + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.ff.3.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.audio_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.audio_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.audio_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.audio_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.human_motion_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.human_motion_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.human_motion_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.human_motion_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.norm.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.norm.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.output.weight": { + "shape": [ + 125, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.output.bias": { + "shape": [ + 125 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.shared.weight": { + "shape": [ + 32128, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.embed_tokens.weight": { + "shape": [ + 32128, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.0.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.0.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.0.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.0.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.0.layer.0.SelfAttention.relative_attention_bias.weight": { + "shape": [ + 32, + 12 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.0.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.0.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.0.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.0.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.1.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.1.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.1.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.1.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.1.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.1.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.1.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.1.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.2.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.2.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.2.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.2.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.2.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.2.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.2.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.2.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.3.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.3.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.3.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.3.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.3.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.3.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.3.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.3.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.4.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.4.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.4.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.4.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.4.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.4.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.4.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.4.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.5.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.5.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.5.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.5.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.5.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.5.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.5.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.5.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.6.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.6.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.6.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.6.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.6.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.6.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.6.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.6.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.7.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.7.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.7.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.7.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.7.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.7.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.7.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.7.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.8.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.8.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.8.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.8.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.8.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.8.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.8.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.8.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.9.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.9.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.9.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.9.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.9.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.9.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.9.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.9.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.10.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.10.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.10.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.10.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.10.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.10.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.10.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.10.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.11.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.11.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.11.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.11.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.11.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.11.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.11.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.11.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.final_layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "history_projector.0.weight": { + "shape": [ + 125 + ], + "dtype": "torch.float32" + }, + "history_projector.0.bias": { + "shape": [ + 125 + ], + "dtype": "torch.float32" + }, + "history_projector.1.weight": { + "shape": [ + 1152, + 125 + ], + "dtype": "torch.float32" + }, + "history_projector.1.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "history_projector.3.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "history_projector.3.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "audio_embedder.0.weight": { + "shape": [ + 35 + ], + "dtype": "torch.float32" + }, + "audio_embedder.0.bias": { + "shape": [ + 35 + ], + "dtype": "torch.float32" + }, + "audio_embedder.1.weight": { + "shape": [ + 2304, + 35 + ], + "dtype": "torch.float32" + }, + "audio_embedder.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "audio_embedder.4.weight": { + "shape": [ + 1152, + 2304 + ], + "dtype": "torch.float32" + }, + "audio_embedder.4.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "human_motion_embedder.0.weight": { + "shape": [ + 66 + ], + "dtype": "torch.float32" + }, + "human_motion_embedder.0.bias": { + "shape": [ + 66 + ], + "dtype": "torch.float32" + }, + "human_motion_embedder.1.weight": { + "shape": [ + 2304, + 66 + ], + "dtype": "torch.float32" + }, + "human_motion_embedder.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "human_motion_embedder.4.weight": { + "shape": [ + 1152, + 2304 + ], + "dtype": "torch.float32" + }, + "human_motion_embedder.4.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + } + }, + "loops": { + "fit_loop": { + "state_dict": { + "keys": [], + "count": 0 + }, + "epoch_loop.state_dict": { + "keys": [ + "_batches_that_stepped" + ], + "count": 1 + }, + "epoch_loop.batch_progress": { + "keys": [ + "total", + "current", + "is_last_batch" + ], + "count": 3 + }, + "epoch_loop.scheduler_progress": { + "keys": [ + "total", + "current" + ], + "count": 2 + }, + "epoch_loop.automatic_optimization.state_dict": { + "keys": [], + "count": 0 + }, + "epoch_loop.automatic_optimization.optim_progress": { + "keys": [ + "optimizer" + ], + "count": 1 + }, + "epoch_loop.manual_optimization.state_dict": { + "keys": [], + "count": 0 + }, + "epoch_loop.manual_optimization.optim_step_progress": { + "keys": [ + "total", + "current" + ], + "count": 2 + }, + "epoch_loop.val_loop.state_dict": { + "keys": [], + "count": 0 + }, + "epoch_loop.val_loop.batch_progress": { + "keys": [ + "total", + "current", + "is_last_batch" + ], + "count": 3 + }, + "epoch_progress": { + "keys": [ + "total", + "current" + ], + "count": 2 + } + }, + "validate_loop": { + "state_dict": { + "keys": [], + "count": 0 + }, + "batch_progress": { + "keys": [ + "total", + "current", + "is_last_batch" + ], + "count": 3 + } + }, + "test_loop": { + "state_dict": { + "keys": [], + "count": 0 + }, + "batch_progress": { + "keys": [ + "total", + "current", + "is_last_batch" + ], + "count": 3 + } + }, + "predict_loop": { + "state_dict": { + "keys": [], + "count": 0 + }, + "batch_progress": { + "keys": [ + "total", + "current" + ], + "count": 2 + } + } + }, + "callbacks": { + "ModelCheckpoint{'monitor': 'step', 'mode': 'max', 'every_n_train_steps': 2000, 'every_n_epochs': 0, 'train_time_interval': None}": { + "monitor": "step", + "best_model_score": { + "shape": [], + "dtype": "torch.int64" + }, + "best_model_path": "/autodl-fs/data/pangxinlei/models/OMG/hfd-l-english-split20261026/checkpoints/step-00098000.ckpt", + "current_score": { + "shape": [], + "dtype": "torch.int64" + }, + "dirpath": "/autodl-fs/data/pangxinlei/models/OMG/hfd-l-english-split20261026/checkpoints", + "best_k_models": { + "keys": [ + "/autodl-fs/data/pangxinlei/models/OMG/hfd-l-english-split20261026/checkpoints/step-00096000.ckpt", + "/autodl-fs/data/pangxinlei/models/OMG/hfd-l-english-split20261026/checkpoints/step-00098000.ckpt" + ], + "count": 2 + }, + "kth_best_model_path": "/autodl-fs/data/pangxinlei/models/OMG/hfd-l-english-split20261026/checkpoints/step-00096000.ckpt", + "kth_value": { + "shape": [], + "dtype": "torch.int64" + }, + "last_model_path": "/autodl-fs/data/pangxinlei/models/OMG/hfd-l-english-split20261026/checkpoints/last.ckpt" + }, + "ModelCheckpoint{'monitor': 'val/loss', 'mode': 'min', 'every_n_train_steps': 0, 'every_n_epochs': 1, 'train_time_interval': None}": { + "monitor": "val/loss", + "best_model_score": { + "shape": [], + "dtype": "torch.float32" + }, + "best_model_path": "/autodl-fs/data/pangxinlei/models/OMG/hfd-l-english-split20261026/best/best.ckpt", + "current_score": { + "shape": [], + "dtype": "torch.float32" + }, + "dirpath": "/autodl-fs/data/pangxinlei/models/OMG/hfd-l-english-split20261026/best", + "best_k_models": { + "keys": [ + "/autodl-fs/data/pangxinlei/models/OMG/hfd-l-english-split20261026/best/best.ckpt" + ], + "count": 1 + }, + "kth_best_model_path": "/autodl-fs/data/pangxinlei/models/OMG/hfd-l-english-split20261026/best/best.ckpt", + "kth_value": { + "shape": [], + "dtype": "torch.float32" + }, + "last_model_path": "" + } + }, + "optimizer_states": { + "count": 1, + "sample": [ + { + "state": { + "keys": [ + "0", + "1", + "2", + "3", + "4", + "5", + "6", + "7", + "8", + "9", + "10", + "11", + "12", + "13", + "14", + "15", + "16", + "17", + "18", + "19", + "20", + "21", + "22", + "23", + "24", + "25", + "26", + "27", + "28", + "29" + ], + "count": 336 + }, + "param_groups": { + "count": 1, + "sample": [ + { + "keys": [ + "lr", + "betas", + "eps", + "weight_decay", + "amsgrad", + "maximize", + "foreach", + "capturable", + "differentiable", + "fused", + "decoupled_weight_decay", + "initial_lr", + "params" + ], + "count": 13 + } + ] + } + } + ] + }, + "lr_schedulers": { + "count": 1, + "sample": [ + { + "_milestones": { + "count": 1, + "sample": [ + 2000 + ] + }, + "last_epoch": 98000, + "_last_lr": { + "count": 1, + "sample": [ + 1.0606109138651e-06 + ] + }, + "_schedulers": { + "count": 2, + "sample": [ + { + "keys": [ + "start_factor", + "end_factor", + "total_iters", + "base_lrs", + "last_epoch", + "_step_count", + "_get_lr_called_within_step", + "_last_lr" + ], + "count": 8 + }, + { + "keys": [ + "T_max", + "eta_min", + "base_lrs", + "last_epoch", + "_step_count", + "_get_lr_called_within_step", + "_last_lr" + ], + "count": 7 + } + ] + } + } + ] + }, + "hfd_resume": { + "version": 1, + "rng": { + "python": { + "count": 3, + "sample": [ + 3, + { + "count": 625, + "sample": [ + 2147483648, + 3564348608 + ] + } + ] + }, + "numpy": { + "count": 5, + "sample": [ + "MT19937", + { + "count": 624, + "sample": [ + 42, + 3107752595 + ] + } + ] + }, + "torch": { + "shape": [ + 5056 + ], + "dtype": "torch.uint8" + }, + "cuda": { + "count": 1, + "sample": [ + { + "shape": [ + 16 + ], + "dtype": "torch.uint8" + } + ] + } + }, + "train_pool": { + "step": 98000, + "slots": { + "count": 512, + "sample": [ + 271745, + 295270 + ] + }, + "batch_size": 512, + "refresh": 256, + "rng": { + "keys": [ + "bit_generator", + "state", + "has_uint32", + "uinteger" + ], + "count": 4 + }, + "identity": { + "keys": [ + "sessions", + "native_source_sha256", + "asr", + "recipe_sha256", + "foreground_language", + "ordered_session_selection_sha256", + "final_release" + ], + "count": 7 + } + }, + "val_pool": { + "step": 16, + "slots": { + "count": 512, + "sample": [ + 6295, + 7193 + ] + }, + "batch_size": 512, + "refresh": 256, + "rng": { + "keys": [ + "bit_generator", + "state", + "has_uint32", + "uinteger" + ], + "count": 4 + }, + "identity": { + "keys": [ + "sessions", + "native_source_sha256", + "asr", + "recipe_sha256", + "foreground_language", + "ordered_session_selection_sha256", + "final_release" + ], + "count": 7 + } + } + }, + "omg_model_architecture": { + "format": "omg.model_architecture", + "version": 1, + "denoiser_type": "omg.generation.denoisers.transformer.MotionTransformerDenoiser", + "frame_cond_injection": "per_layer_film", + "attention": { + "rotary_self_attention_qk_norm": true, + "cross_attention_qk_norm": true + } + } +} \ No newline at end of file diff --git a/handoff-v1/inventory/omg-l-step-00100000.ckpt.json b/handoff-v1/inventory/omg-l-step-00100000.ckpt.json new file mode 100644 index 0000000000000000000000000000000000000000..22e563d41fb00e94e4e871e25e9c54fa2ca53f59 --- /dev/null +++ b/handoff-v1/inventory/omg-l-step-00100000.ckpt.json @@ -0,0 +1,3596 @@ +{ + "epoch": 0, + "global_step": 100000, + "pytorch-lightning_version": "2.6.6", + "state_dict": { + "denoiser.time_embed.proj.0.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.time_embed.proj.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.time_embed.proj.2.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.time_embed.proj.2.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.input_proj.weight": { + "shape": [ + 1152, + 1277 + ], + "dtype": "torch.float32" + }, + "denoiser.input_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.text_proj.weight": { + "shape": [ + 1152, + 768 + ], + "dtype": "torch.float32" + }, + "denoiser.text_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.extra_proj.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.extra_proj.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.extra_proj.1.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.extra_proj.1.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.norm_cross.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.norm_cross.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.cross_attn.in_proj_weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.cross_attn.in_proj_bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.cross_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.cross_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.norm_self.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.norm_self.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.self_attn.qkv.weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.self_attn.qkv.bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.self_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.self_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.norm_ff.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.norm_ff.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.ff.0.weight": { + "shape": [ + 4608, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.ff.0.bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.ff.3.weight": { + "shape": [ + 1152, + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.ff.3.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.audio_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.audio_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.audio_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.audio_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.human_motion_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.human_motion_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.human_motion_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.human_motion_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.norm_cross.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.norm_cross.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.cross_attn.in_proj_weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.cross_attn.in_proj_bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.cross_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.cross_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.norm_self.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.norm_self.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.self_attn.qkv.weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.self_attn.qkv.bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.self_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.self_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.norm_ff.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.norm_ff.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.ff.0.weight": { + "shape": [ + 4608, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.ff.0.bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.ff.3.weight": { + "shape": [ + 1152, + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.ff.3.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.audio_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.audio_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.audio_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.audio_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.human_motion_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.human_motion_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.human_motion_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.human_motion_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.norm_cross.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.norm_cross.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.cross_attn.in_proj_weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.cross_attn.in_proj_bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.cross_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.cross_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.norm_self.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.norm_self.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.self_attn.qkv.weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.self_attn.qkv.bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.self_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.self_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.norm_ff.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.norm_ff.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.ff.0.weight": { + "shape": [ + 4608, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.ff.0.bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.ff.3.weight": { + "shape": [ + 1152, + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.ff.3.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.audio_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.audio_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.audio_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.audio_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.human_motion_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.human_motion_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.human_motion_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.human_motion_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.norm_cross.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.norm_cross.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.cross_attn.in_proj_weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.cross_attn.in_proj_bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.cross_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.cross_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.norm_self.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.norm_self.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.self_attn.qkv.weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.self_attn.qkv.bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.self_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.self_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.norm_ff.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.norm_ff.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.ff.0.weight": { + "shape": [ + 4608, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.ff.0.bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.ff.3.weight": { + "shape": [ + 1152, + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.ff.3.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.audio_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.audio_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.audio_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.audio_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.human_motion_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.human_motion_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.human_motion_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.human_motion_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.norm_cross.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.norm_cross.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.cross_attn.in_proj_weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.cross_attn.in_proj_bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.cross_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.cross_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.norm_self.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.norm_self.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.self_attn.qkv.weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.self_attn.qkv.bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.self_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.self_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.norm_ff.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.norm_ff.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.ff.0.weight": { + "shape": [ + 4608, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.ff.0.bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.ff.3.weight": { + "shape": [ + 1152, + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.ff.3.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.audio_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.audio_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.audio_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.audio_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.human_motion_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.human_motion_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.human_motion_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.human_motion_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.norm_cross.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.norm_cross.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.cross_attn.in_proj_weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.cross_attn.in_proj_bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.cross_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.cross_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.norm_self.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.norm_self.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.self_attn.qkv.weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.self_attn.qkv.bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.self_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.self_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.norm_ff.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.norm_ff.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.ff.0.weight": { + "shape": [ + 4608, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.ff.0.bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.ff.3.weight": { + "shape": [ + 1152, + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.ff.3.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.audio_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.audio_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.audio_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.audio_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.human_motion_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.human_motion_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.human_motion_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.human_motion_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.norm_cross.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.norm_cross.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.cross_attn.in_proj_weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.cross_attn.in_proj_bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.cross_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.cross_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.norm_self.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.norm_self.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.self_attn.qkv.weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.self_attn.qkv.bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.self_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.self_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.norm_ff.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.norm_ff.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.ff.0.weight": { + "shape": [ + 4608, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.ff.0.bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.ff.3.weight": { + "shape": [ + 1152, + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.ff.3.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.audio_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.audio_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.audio_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.audio_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.human_motion_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.human_motion_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.human_motion_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.human_motion_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.norm_cross.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.norm_cross.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.cross_attn.in_proj_weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.cross_attn.in_proj_bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.cross_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.cross_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.norm_self.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.norm_self.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.self_attn.qkv.weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.self_attn.qkv.bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.self_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.self_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.norm_ff.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.norm_ff.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.ff.0.weight": { + "shape": [ + 4608, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.ff.0.bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.ff.3.weight": { + "shape": [ + 1152, + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.ff.3.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.audio_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.audio_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.audio_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.audio_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.human_motion_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.human_motion_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.human_motion_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.human_motion_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.norm_cross.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.norm_cross.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.cross_attn.in_proj_weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.cross_attn.in_proj_bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.cross_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.cross_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.norm_self.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.norm_self.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.self_attn.qkv.weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.self_attn.qkv.bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.self_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.self_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.norm_ff.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.norm_ff.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.ff.0.weight": { + "shape": [ + 4608, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.ff.0.bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.ff.3.weight": { + "shape": [ + 1152, + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.ff.3.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.audio_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.audio_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.audio_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.audio_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.human_motion_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.human_motion_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.human_motion_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.human_motion_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.norm_cross.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.norm_cross.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.cross_attn.in_proj_weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.cross_attn.in_proj_bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.cross_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.cross_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.norm_self.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.norm_self.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.self_attn.qkv.weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.self_attn.qkv.bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.self_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.self_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.norm_ff.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.norm_ff.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.ff.0.weight": { + "shape": [ + 4608, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.ff.0.bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.ff.3.weight": { + "shape": [ + 1152, + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.ff.3.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.audio_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.audio_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.audio_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.audio_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.human_motion_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.human_motion_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.human_motion_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.human_motion_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.norm_cross.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.norm_cross.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.cross_attn.in_proj_weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.cross_attn.in_proj_bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.cross_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.cross_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.norm_self.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.norm_self.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.self_attn.qkv.weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.self_attn.qkv.bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.self_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.self_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.norm_ff.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.norm_ff.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.ff.0.weight": { + "shape": [ + 4608, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.ff.0.bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.ff.3.weight": { + "shape": [ + 1152, + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.ff.3.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.audio_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.audio_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.audio_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.audio_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.human_motion_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.human_motion_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.human_motion_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.human_motion_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.norm_cross.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.norm_cross.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.cross_attn.in_proj_weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.cross_attn.in_proj_bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.cross_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.cross_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.norm_self.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.norm_self.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.self_attn.qkv.weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.self_attn.qkv.bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.self_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.self_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.norm_ff.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.norm_ff.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.ff.0.weight": { + "shape": [ + 4608, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.ff.0.bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.ff.3.weight": { + "shape": [ + 1152, + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.ff.3.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.audio_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.audio_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.audio_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.audio_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.human_motion_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.human_motion_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.human_motion_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.human_motion_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.norm_cross.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.norm_cross.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.cross_attn.in_proj_weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.cross_attn.in_proj_bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.cross_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.cross_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.norm_self.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.norm_self.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.self_attn.qkv.weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.self_attn.qkv.bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.self_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.self_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.norm_ff.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.norm_ff.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.ff.0.weight": { + "shape": [ + 4608, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.ff.0.bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.ff.3.weight": { + "shape": [ + 1152, + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.ff.3.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.audio_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.audio_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.audio_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.audio_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.human_motion_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.human_motion_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.human_motion_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.human_motion_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.norm_cross.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.norm_cross.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.cross_attn.in_proj_weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.cross_attn.in_proj_bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.cross_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.cross_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.norm_self.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.norm_self.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.self_attn.qkv.weight": { + "shape": [ + 3456, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.self_attn.qkv.bias": { + "shape": [ + 3456 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.self_attn.out_proj.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.self_attn.out_proj.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.norm_ff.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.norm_ff.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.ff.0.weight": { + "shape": [ + 4608, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.ff.0.bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.ff.3.weight": { + "shape": [ + 1152, + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.ff.3.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.audio_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.audio_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.audio_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.audio_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.human_motion_film.0.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.human_motion_film.0.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.human_motion_film.1.weight": { + "shape": [ + 2304, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.13.human_motion_film.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "denoiser.norm.weight": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.norm.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.output.weight": { + "shape": [ + 125, + 1152 + ], + "dtype": "torch.float32" + }, + "denoiser.output.bias": { + "shape": [ + 125 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.shared.weight": { + "shape": [ + 32128, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.embed_tokens.weight": { + "shape": [ + 32128, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.0.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.0.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.0.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.0.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.0.layer.0.SelfAttention.relative_attention_bias.weight": { + "shape": [ + 32, + 12 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.0.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.0.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.0.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.0.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.1.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.1.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.1.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.1.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.1.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.1.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.1.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.1.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.2.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.2.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.2.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.2.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.2.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.2.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.2.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.2.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.3.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.3.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.3.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.3.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.3.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.3.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.3.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.3.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.4.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.4.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.4.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.4.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.4.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.4.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.4.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.4.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.5.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.5.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.5.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.5.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.5.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.5.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.5.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.5.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.6.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.6.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.6.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.6.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.6.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.6.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.6.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.6.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.7.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.7.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.7.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.7.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.7.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.7.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.7.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.7.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.8.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.8.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.8.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.8.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.8.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.8.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.8.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.8.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.9.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.9.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.9.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.9.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.9.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.9.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.9.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.9.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.10.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.10.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.10.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.10.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.10.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.10.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.10.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.10.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.11.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.11.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.11.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.11.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.11.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.11.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.11.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.11.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.final_layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "history_projector.0.weight": { + "shape": [ + 125 + ], + "dtype": "torch.float32" + }, + "history_projector.0.bias": { + "shape": [ + 125 + ], + "dtype": "torch.float32" + }, + "history_projector.1.weight": { + "shape": [ + 1152, + 125 + ], + "dtype": "torch.float32" + }, + "history_projector.1.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "history_projector.3.weight": { + "shape": [ + 1152, + 1152 + ], + "dtype": "torch.float32" + }, + "history_projector.3.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "audio_embedder.0.weight": { + "shape": [ + 35 + ], + "dtype": "torch.float32" + }, + "audio_embedder.0.bias": { + "shape": [ + 35 + ], + "dtype": "torch.float32" + }, + "audio_embedder.1.weight": { + "shape": [ + 2304, + 35 + ], + "dtype": "torch.float32" + }, + "audio_embedder.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "audio_embedder.4.weight": { + "shape": [ + 1152, + 2304 + ], + "dtype": "torch.float32" + }, + "audio_embedder.4.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + }, + "human_motion_embedder.0.weight": { + "shape": [ + 66 + ], + "dtype": "torch.float32" + }, + "human_motion_embedder.0.bias": { + "shape": [ + 66 + ], + "dtype": "torch.float32" + }, + "human_motion_embedder.1.weight": { + "shape": [ + 2304, + 66 + ], + "dtype": "torch.float32" + }, + "human_motion_embedder.1.bias": { + "shape": [ + 2304 + ], + "dtype": "torch.float32" + }, + "human_motion_embedder.4.weight": { + "shape": [ + 1152, + 2304 + ], + "dtype": "torch.float32" + }, + "human_motion_embedder.4.bias": { + "shape": [ + 1152 + ], + "dtype": "torch.float32" + } + }, + "loops": { + "fit_loop": { + "state_dict": { + "keys": [], + "count": 0 + }, + "epoch_loop.state_dict": { + "keys": [ + "_batches_that_stepped" + ], + "count": 1 + }, + "epoch_loop.batch_progress": { + "keys": [ + "total", + "current", + "is_last_batch" + ], + "count": 3 + }, + "epoch_loop.scheduler_progress": { + "keys": [ + "total", + "current" + ], + "count": 2 + }, + "epoch_loop.automatic_optimization.state_dict": { + "keys": [], + "count": 0 + }, + "epoch_loop.automatic_optimization.optim_progress": { + "keys": [ + "optimizer" + ], + "count": 1 + }, + "epoch_loop.manual_optimization.state_dict": { + "keys": [], + "count": 0 + }, + "epoch_loop.manual_optimization.optim_step_progress": { + "keys": [ + "total", + "current" + ], + "count": 2 + }, + "epoch_loop.val_loop.state_dict": { + "keys": [], + "count": 0 + }, + "epoch_loop.val_loop.batch_progress": { + "keys": [ + "total", + "current", + "is_last_batch" + ], + "count": 3 + }, + "epoch_progress": { + "keys": [ + "total", + "current" + ], + "count": 2 + } + }, + "validate_loop": { + "state_dict": { + "keys": [], + "count": 0 + }, + "batch_progress": { + "keys": [ + "total", + "current", + "is_last_batch" + ], + "count": 3 + } + }, + "test_loop": { + "state_dict": { + "keys": [], + "count": 0 + }, + "batch_progress": { + "keys": [ + "total", + "current", + "is_last_batch" + ], + "count": 3 + } + }, + "predict_loop": { + "state_dict": { + "keys": [], + "count": 0 + }, + "batch_progress": { + "keys": [ + "total", + "current" + ], + "count": 2 + } + } + }, + "callbacks": { + "ModelCheckpoint{'monitor': 'step', 'mode': 'max', 'every_n_train_steps': 2000, 'every_n_epochs': 0, 'train_time_interval': None}": { + "monitor": "step", + "best_model_score": { + "shape": [], + "dtype": "torch.int64" + }, + "best_model_path": "/autodl-fs/data/pangxinlei/models/OMG/hfd-l-english-split20261026/checkpoints/step-00100000.ckpt", + "current_score": { + "shape": [], + "dtype": "torch.int64" + }, + "dirpath": "/autodl-fs/data/pangxinlei/models/OMG/hfd-l-english-split20261026/checkpoints", + "best_k_models": { + "keys": [ + "/autodl-fs/data/pangxinlei/models/OMG/hfd-l-english-split20261026/checkpoints/step-00098000.ckpt", + "/autodl-fs/data/pangxinlei/models/OMG/hfd-l-english-split20261026/checkpoints/step-00100000.ckpt" + ], + "count": 2 + }, + "kth_best_model_path": "/autodl-fs/data/pangxinlei/models/OMG/hfd-l-english-split20261026/checkpoints/step-00098000.ckpt", + "kth_value": { + "shape": [], + "dtype": "torch.int64" + }, + "last_model_path": "/autodl-fs/data/pangxinlei/models/OMG/hfd-l-english-split20261026/checkpoints/last.ckpt" + }, + "ModelCheckpoint{'monitor': 'val/loss', 'mode': 'min', 'every_n_train_steps': 0, 'every_n_epochs': 1, 'train_time_interval': None}": { + "monitor": "val/loss", + "best_model_score": { + "shape": [], + "dtype": "torch.float32" + }, + "best_model_path": "/autodl-fs/data/pangxinlei/models/OMG/hfd-l-english-split20261026/best/best.ckpt", + "current_score": { + "shape": [], + "dtype": "torch.float32" + }, + "dirpath": "/autodl-fs/data/pangxinlei/models/OMG/hfd-l-english-split20261026/best", + "best_k_models": { + "keys": [ + "/autodl-fs/data/pangxinlei/models/OMG/hfd-l-english-split20261026/best/best.ckpt" + ], + "count": 1 + }, + "kth_best_model_path": "/autodl-fs/data/pangxinlei/models/OMG/hfd-l-english-split20261026/best/best.ckpt", + "kth_value": { + "shape": [], + "dtype": "torch.float32" + }, + "last_model_path": "" + } + }, + "optimizer_states": { + "count": 1, + "sample": [ + { + "state": { + "keys": [ + "0", + "1", + "2", + "3", + "4", + "5", + "6", + "7", + "8", + "9", + "10", + "11", + "12", + "13", + "14", + "15", + "16", + "17", + "18", + "19", + "20", + "21", + "22", + "23", + "24", + "25", + "26", + "27", + "28", + "29" + ], + "count": 336 + }, + "param_groups": { + "count": 1, + "sample": [ + { + "keys": [ + "lr", + "betas", + "eps", + "weight_decay", + "amsgrad", + "maximize", + "foreach", + "capturable", + "differentiable", + "fused", + "decoupled_weight_decay", + "initial_lr", + "params" + ], + "count": 13 + } + ] + } + } + ] + }, + "lr_schedulers": { + "count": 1, + "sample": [ + { + "_milestones": { + "count": 1, + "sample": [ + 2000 + ] + }, + "last_epoch": 100000, + "_last_lr": { + "count": 1, + "sample": [ + 1e-06 + ] + }, + "_schedulers": { + "count": 2, + "sample": [ + { + "keys": [ + "start_factor", + "end_factor", + "total_iters", + "base_lrs", + "last_epoch", + "_step_count", + "_get_lr_called_within_step", + "_last_lr" + ], + "count": 8 + }, + { + "keys": [ + "T_max", + "eta_min", + "base_lrs", + "last_epoch", + "_step_count", + "_get_lr_called_within_step", + "_last_lr" + ], + "count": 7 + } + ] + } + } + ] + }, + "hfd_resume": { + "version": 1, + "rng": { + "python": { + "count": 3, + "sample": [ + 3, + { + "count": 625, + "sample": [ + 2147483648, + 3564348608 + ] + } + ] + }, + "numpy": { + "count": 5, + "sample": [ + "MT19937", + { + "count": 624, + "sample": [ + 42, + 3107752595 + ] + } + ] + }, + "torch": { + "shape": [ + 5056 + ], + "dtype": "torch.uint8" + }, + "cuda": { + "count": 1, + "sample": [ + { + "shape": [ + 16 + ], + "dtype": "torch.uint8" + } + ] + } + }, + "train_pool": { + "step": 100000, + "slots": { + "count": 512, + "sample": [ + 248743, + 347419 + ] + }, + "batch_size": 512, + "refresh": 256, + "rng": { + "keys": [ + "bit_generator", + "state", + "has_uint32", + "uinteger" + ], + "count": 4 + }, + "identity": { + "keys": [ + "sessions", + "native_source_sha256", + "asr", + "recipe_sha256", + "foreground_language", + "ordered_session_selection_sha256", + "final_release" + ], + "count": 7 + } + }, + "val_pool": { + "step": 16, + "slots": { + "count": 512, + "sample": [ + 6295, + 7193 + ] + }, + "batch_size": 512, + "refresh": 256, + "rng": { + "keys": [ + "bit_generator", + "state", + "has_uint32", + "uinteger" + ], + "count": 4 + }, + "identity": { + "keys": [ + "sessions", + "native_source_sha256", + "asr", + "recipe_sha256", + "foreground_language", + "ordered_session_selection_sha256", + "final_release" + ], + "count": 7 + } + } + }, + "omg_model_architecture": { + "format": "omg.model_architecture", + "version": 1, + "denoiser_type": "omg.generation.denoisers.transformer.MotionTransformerDenoiser", + "frame_cond_injection": "per_layer_film", + "attention": { + "rotary_self_attention_qk_norm": true, + "cross_attention_qk_norm": true + } + } +} \ No newline at end of file diff --git a/handoff-v1/inventory/omg-runtime.txt b/handoff-v1/inventory/omg-runtime.txt new file mode 100644 index 0000000000000000000000000000000000000000..5401cdf377df1c55867a02d9f666e448d8b4746d --- /dev/null +++ b/handoff-v1/inventory/omg-runtime.txt @@ -0,0 +1 @@ +{"python": "3.10.21 (main, Aug 27 2026, 14:42:07) [GCC 14.3.0]", "torch": "2.7.1+cu128", "cuda": "12.8", "cudnn": 90701} diff --git a/handoff-v1/inventory/omg-xl-best.ckpt.json b/handoff-v1/inventory/omg-xl-best.ckpt.json new file mode 100644 index 0000000000000000000000000000000000000000..43497b017c14fb0e75eefa8dbb976412fd9e4d7c --- /dev/null +++ b/handoff-v1/inventory/omg-xl-best.ckpt.json @@ -0,0 +1,3432 @@ +{ + "epoch": 0, + "global_step": 88000, + "pytorch-lightning_version": "2.6.6", + "state_dict": { + "denoiser.time_embed.proj.0.weight": { + "shape": [ + 1536, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.time_embed.proj.0.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.time_embed.proj.2.weight": { + "shape": [ + 1536, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.time_embed.proj.2.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.input_proj.weight": { + "shape": [ + 1536, + 1661 + ], + "dtype": "torch.float32" + }, + "denoiser.input_proj.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.text_proj.weight": { + "shape": [ + 1536, + 768 + ], + "dtype": "torch.float32" + }, + "denoiser.text_proj.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.extra_proj.0.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.extra_proj.0.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.extra_proj.1.weight": { + "shape": [ + 1536, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.extra_proj.1.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.norm_cross.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.norm_cross.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.cross_attn.in_proj_weight": { + "shape": [ + 4608, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.cross_attn.in_proj_bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.cross_attn.out_proj.weight": { + "shape": [ + 1536, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.cross_attn.out_proj.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.norm_self.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.norm_self.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.self_attn.qkv.weight": { + "shape": [ + 4608, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.self_attn.qkv.bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.self_attn.out_proj.weight": { + "shape": [ + 1536, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.self_attn.out_proj.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.norm_ff.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.norm_ff.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.ff.0.weight": { + "shape": [ + 6144, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.ff.0.bias": { + "shape": [ + 6144 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.ff.3.weight": { + "shape": [ + 1536, + 6144 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.ff.3.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.audio_film.0.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.audio_film.0.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.audio_film.1.weight": { + "shape": [ + 3072, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.audio_film.1.bias": { + "shape": [ + 3072 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.human_motion_film.0.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.human_motion_film.0.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.human_motion_film.1.weight": { + "shape": [ + 3072, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.human_motion_film.1.bias": { + "shape": [ + 3072 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.norm_cross.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.norm_cross.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.cross_attn.in_proj_weight": { + "shape": [ + 4608, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.cross_attn.in_proj_bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.cross_attn.out_proj.weight": { + "shape": [ + 1536, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.cross_attn.out_proj.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.norm_self.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.norm_self.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.self_attn.qkv.weight": { + "shape": [ + 4608, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.self_attn.qkv.bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.self_attn.out_proj.weight": { + "shape": [ + 1536, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.self_attn.out_proj.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.norm_ff.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.norm_ff.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.ff.0.weight": { + "shape": [ + 6144, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.ff.0.bias": { + "shape": [ + 6144 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.ff.3.weight": { + "shape": [ + 1536, + 6144 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.ff.3.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.audio_film.0.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.audio_film.0.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.audio_film.1.weight": { + "shape": [ + 3072, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.audio_film.1.bias": { + "shape": [ + 3072 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.human_motion_film.0.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.human_motion_film.0.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.human_motion_film.1.weight": { + "shape": [ + 3072, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.human_motion_film.1.bias": { + "shape": [ + 3072 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.norm_cross.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.norm_cross.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.cross_attn.in_proj_weight": { + "shape": [ + 4608, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.cross_attn.in_proj_bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.cross_attn.out_proj.weight": { + "shape": [ + 1536, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.cross_attn.out_proj.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.norm_self.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.norm_self.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.self_attn.qkv.weight": { + "shape": [ + 4608, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.self_attn.qkv.bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.self_attn.out_proj.weight": { + "shape": [ + 1536, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.self_attn.out_proj.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.norm_ff.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.norm_ff.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.ff.0.weight": { + "shape": [ + 6144, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.ff.0.bias": { + "shape": [ + 6144 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.ff.3.weight": { + "shape": [ + 1536, + 6144 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.ff.3.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.audio_film.0.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.audio_film.0.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.audio_film.1.weight": { + "shape": [ + 3072, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.audio_film.1.bias": { + "shape": [ + 3072 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.human_motion_film.0.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.human_motion_film.0.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.human_motion_film.1.weight": { + "shape": [ + 3072, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.human_motion_film.1.bias": { + "shape": [ + 3072 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.norm_cross.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.norm_cross.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.cross_attn.in_proj_weight": { + "shape": [ + 4608, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.cross_attn.in_proj_bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.cross_attn.out_proj.weight": { + "shape": [ + 1536, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.cross_attn.out_proj.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.norm_self.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.norm_self.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.self_attn.qkv.weight": { + "shape": [ + 4608, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.self_attn.qkv.bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.self_attn.out_proj.weight": { + "shape": [ + 1536, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.self_attn.out_proj.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.norm_ff.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.norm_ff.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.ff.0.weight": { + "shape": [ + 6144, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.ff.0.bias": { + "shape": [ + 6144 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.ff.3.weight": { + "shape": [ + 1536, + 6144 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.ff.3.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.audio_film.0.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.audio_film.0.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.audio_film.1.weight": { + "shape": [ + 3072, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.audio_film.1.bias": { + "shape": [ + 3072 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.human_motion_film.0.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.human_motion_film.0.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.human_motion_film.1.weight": { + "shape": [ + 3072, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.human_motion_film.1.bias": { + "shape": [ + 3072 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.norm_cross.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.norm_cross.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.cross_attn.in_proj_weight": { + "shape": [ + 4608, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.cross_attn.in_proj_bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.cross_attn.out_proj.weight": { + "shape": [ + 1536, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.cross_attn.out_proj.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.norm_self.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.norm_self.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.self_attn.qkv.weight": { + "shape": [ + 4608, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.self_attn.qkv.bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.self_attn.out_proj.weight": { + "shape": [ + 1536, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.self_attn.out_proj.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.norm_ff.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.norm_ff.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.ff.0.weight": { + "shape": [ + 6144, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.ff.0.bias": { + "shape": [ + 6144 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.ff.3.weight": { + "shape": [ + 1536, + 6144 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.ff.3.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.audio_film.0.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.audio_film.0.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.audio_film.1.weight": { + "shape": [ + 3072, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.audio_film.1.bias": { + "shape": [ + 3072 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.human_motion_film.0.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.human_motion_film.0.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.human_motion_film.1.weight": { + "shape": [ + 3072, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.human_motion_film.1.bias": { + "shape": [ + 3072 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.norm_cross.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.norm_cross.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.cross_attn.in_proj_weight": { + "shape": [ + 4608, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.cross_attn.in_proj_bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.cross_attn.out_proj.weight": { + "shape": [ + 1536, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.cross_attn.out_proj.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.norm_self.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.norm_self.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.self_attn.qkv.weight": { + "shape": [ + 4608, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.self_attn.qkv.bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.self_attn.out_proj.weight": { + "shape": [ + 1536, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.self_attn.out_proj.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.norm_ff.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.norm_ff.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.ff.0.weight": { + "shape": [ + 6144, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.ff.0.bias": { + "shape": [ + 6144 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.ff.3.weight": { + "shape": [ + 1536, + 6144 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.ff.3.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.audio_film.0.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.audio_film.0.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.audio_film.1.weight": { + "shape": [ + 3072, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.audio_film.1.bias": { + "shape": [ + 3072 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.human_motion_film.0.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.human_motion_film.0.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.human_motion_film.1.weight": { + "shape": [ + 3072, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.human_motion_film.1.bias": { + "shape": [ + 3072 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.norm_cross.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.norm_cross.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.cross_attn.in_proj_weight": { + "shape": [ + 4608, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.cross_attn.in_proj_bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.cross_attn.out_proj.weight": { + "shape": [ + 1536, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.cross_attn.out_proj.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.norm_self.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.norm_self.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.self_attn.qkv.weight": { + "shape": [ + 4608, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.self_attn.qkv.bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.self_attn.out_proj.weight": { + "shape": [ + 1536, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.self_attn.out_proj.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.norm_ff.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.norm_ff.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.ff.0.weight": { + "shape": [ + 6144, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.ff.0.bias": { + "shape": [ + 6144 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.ff.3.weight": { + "shape": [ + 1536, + 6144 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.ff.3.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.audio_film.0.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.audio_film.0.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.audio_film.1.weight": { + "shape": [ + 3072, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.audio_film.1.bias": { + "shape": [ + 3072 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.human_motion_film.0.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.human_motion_film.0.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.human_motion_film.1.weight": { + "shape": [ + 3072, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.human_motion_film.1.bias": { + "shape": [ + 3072 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.norm_cross.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.norm_cross.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.cross_attn.in_proj_weight": { + "shape": [ + 4608, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.cross_attn.in_proj_bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.cross_attn.out_proj.weight": { + "shape": [ + 1536, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.cross_attn.out_proj.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.norm_self.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.norm_self.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.self_attn.qkv.weight": { + "shape": [ + 4608, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.self_attn.qkv.bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.self_attn.out_proj.weight": { + "shape": [ + 1536, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.self_attn.out_proj.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.norm_ff.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.norm_ff.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.ff.0.weight": { + "shape": [ + 6144, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.ff.0.bias": { + "shape": [ + 6144 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.ff.3.weight": { + "shape": [ + 1536, + 6144 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.ff.3.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.audio_film.0.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.audio_film.0.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.audio_film.1.weight": { + "shape": [ + 3072, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.audio_film.1.bias": { + "shape": [ + 3072 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.human_motion_film.0.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.human_motion_film.0.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.human_motion_film.1.weight": { + "shape": [ + 3072, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.human_motion_film.1.bias": { + "shape": [ + 3072 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.norm_cross.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.norm_cross.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.cross_attn.in_proj_weight": { + "shape": [ + 4608, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.cross_attn.in_proj_bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.cross_attn.out_proj.weight": { + "shape": [ + 1536, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.cross_attn.out_proj.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.norm_self.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.norm_self.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.self_attn.qkv.weight": { + "shape": [ + 4608, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.self_attn.qkv.bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.self_attn.out_proj.weight": { + "shape": [ + 1536, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.self_attn.out_proj.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.norm_ff.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.norm_ff.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.ff.0.weight": { + "shape": [ + 6144, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.ff.0.bias": { + "shape": [ + 6144 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.ff.3.weight": { + "shape": [ + 1536, + 6144 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.ff.3.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.audio_film.0.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.audio_film.0.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.audio_film.1.weight": { + "shape": [ + 3072, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.audio_film.1.bias": { + "shape": [ + 3072 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.human_motion_film.0.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.human_motion_film.0.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.human_motion_film.1.weight": { + "shape": [ + 3072, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.human_motion_film.1.bias": { + "shape": [ + 3072 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.norm_cross.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.norm_cross.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.cross_attn.in_proj_weight": { + "shape": [ + 4608, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.cross_attn.in_proj_bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.cross_attn.out_proj.weight": { + "shape": [ + 1536, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.cross_attn.out_proj.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.norm_self.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.norm_self.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.self_attn.qkv.weight": { + "shape": [ + 4608, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.self_attn.qkv.bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.self_attn.out_proj.weight": { + "shape": [ + 1536, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.self_attn.out_proj.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.norm_ff.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.norm_ff.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.ff.0.weight": { + "shape": [ + 6144, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.ff.0.bias": { + "shape": [ + 6144 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.ff.3.weight": { + "shape": [ + 1536, + 6144 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.ff.3.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.audio_film.0.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.audio_film.0.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.audio_film.1.weight": { + "shape": [ + 3072, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.audio_film.1.bias": { + "shape": [ + 3072 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.human_motion_film.0.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.human_motion_film.0.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.human_motion_film.1.weight": { + "shape": [ + 3072, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.human_motion_film.1.bias": { + "shape": [ + 3072 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.norm_cross.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.norm_cross.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.cross_attn.in_proj_weight": { + "shape": [ + 4608, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.cross_attn.in_proj_bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.cross_attn.out_proj.weight": { + "shape": [ + 1536, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.cross_attn.out_proj.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.norm_self.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.norm_self.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.self_attn.qkv.weight": { + "shape": [ + 4608, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.self_attn.qkv.bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.self_attn.out_proj.weight": { + "shape": [ + 1536, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.self_attn.out_proj.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.norm_ff.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.norm_ff.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.ff.0.weight": { + "shape": [ + 6144, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.ff.0.bias": { + "shape": [ + 6144 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.ff.3.weight": { + "shape": [ + 1536, + 6144 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.ff.3.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.audio_film.0.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.audio_film.0.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.audio_film.1.weight": { + "shape": [ + 3072, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.audio_film.1.bias": { + "shape": [ + 3072 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.human_motion_film.0.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.human_motion_film.0.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.human_motion_film.1.weight": { + "shape": [ + 3072, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.human_motion_film.1.bias": { + "shape": [ + 3072 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.norm_cross.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.norm_cross.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.cross_attn.in_proj_weight": { + "shape": [ + 4608, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.cross_attn.in_proj_bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.cross_attn.out_proj.weight": { + "shape": [ + 1536, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.cross_attn.out_proj.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.norm_self.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.norm_self.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.self_attn.qkv.weight": { + "shape": [ + 4608, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.self_attn.qkv.bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.self_attn.out_proj.weight": { + "shape": [ + 1536, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.self_attn.out_proj.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.norm_ff.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.norm_ff.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.ff.0.weight": { + "shape": [ + 6144, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.ff.0.bias": { + "shape": [ + 6144 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.ff.3.weight": { + "shape": [ + 1536, + 6144 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.ff.3.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.audio_film.0.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.audio_film.0.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.audio_film.1.weight": { + "shape": [ + 3072, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.audio_film.1.bias": { + "shape": [ + 3072 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.human_motion_film.0.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.human_motion_film.0.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.human_motion_film.1.weight": { + "shape": [ + 3072, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.human_motion_film.1.bias": { + "shape": [ + 3072 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.norm_cross.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.norm_cross.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.cross_attn.in_proj_weight": { + "shape": [ + 4608, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.cross_attn.in_proj_bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.cross_attn.out_proj.weight": { + "shape": [ + 1536, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.cross_attn.out_proj.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.norm_self.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.norm_self.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.self_attn.qkv.weight": { + "shape": [ + 4608, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.self_attn.qkv.bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.self_attn.out_proj.weight": { + "shape": [ + 1536, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.self_attn.out_proj.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.norm_ff.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.norm_ff.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.ff.0.weight": { + "shape": [ + 6144, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.ff.0.bias": { + "shape": [ + 6144 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.ff.3.weight": { + "shape": [ + 1536, + 6144 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.ff.3.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.audio_film.0.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.audio_film.0.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.audio_film.1.weight": { + "shape": [ + 3072, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.audio_film.1.bias": { + "shape": [ + 3072 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.human_motion_film.0.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.human_motion_film.0.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.human_motion_film.1.weight": { + "shape": [ + 3072, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.human_motion_film.1.bias": { + "shape": [ + 3072 + ], + "dtype": "torch.float32" + }, + "denoiser.norm.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.norm.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.output.weight": { + "shape": [ + 125, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.output.bias": { + "shape": [ + 125 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.shared.weight": { + "shape": [ + 32128, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.embed_tokens.weight": { + "shape": [ + 32128, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.0.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.0.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.0.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.0.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.0.layer.0.SelfAttention.relative_attention_bias.weight": { + "shape": [ + 32, + 12 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.0.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.0.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.0.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.0.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.1.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.1.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.1.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.1.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.1.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.1.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.1.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.1.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.2.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.2.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.2.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.2.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.2.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.2.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.2.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.2.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.3.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.3.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.3.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.3.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.3.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.3.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.3.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.3.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.4.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.4.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.4.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.4.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.4.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.4.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.4.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.4.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.5.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.5.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.5.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.5.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.5.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.5.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.5.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.5.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.6.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.6.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.6.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.6.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.6.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.6.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.6.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.6.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.7.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.7.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.7.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.7.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.7.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.7.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.7.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.7.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.8.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.8.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.8.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.8.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.8.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.8.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.8.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.8.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.9.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.9.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.9.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.9.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.9.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.9.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.9.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.9.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.10.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.10.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.10.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.10.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.10.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.10.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.10.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.10.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.11.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.11.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.11.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.11.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.11.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.11.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.11.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.11.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.final_layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "history_projector.0.weight": { + "shape": [ + 125 + ], + "dtype": "torch.float32" + }, + "history_projector.0.bias": { + "shape": [ + 125 + ], + "dtype": "torch.float32" + }, + "history_projector.1.weight": { + "shape": [ + 1536, + 125 + ], + "dtype": "torch.float32" + }, + "history_projector.1.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "history_projector.3.weight": { + "shape": [ + 1536, + 1536 + ], + "dtype": "torch.float32" + }, + "history_projector.3.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "audio_embedder.0.weight": { + "shape": [ + 35 + ], + "dtype": "torch.float32" + }, + "audio_embedder.0.bias": { + "shape": [ + 35 + ], + "dtype": "torch.float32" + }, + "audio_embedder.1.weight": { + "shape": [ + 3072, + 35 + ], + "dtype": "torch.float32" + }, + "audio_embedder.1.bias": { + "shape": [ + 3072 + ], + "dtype": "torch.float32" + }, + "audio_embedder.4.weight": { + "shape": [ + 1536, + 3072 + ], + "dtype": "torch.float32" + }, + "audio_embedder.4.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "human_motion_embedder.0.weight": { + "shape": [ + 66 + ], + "dtype": "torch.float32" + }, + "human_motion_embedder.0.bias": { + "shape": [ + 66 + ], + "dtype": "torch.float32" + }, + "human_motion_embedder.1.weight": { + "shape": [ + 3072, + 66 + ], + "dtype": "torch.float32" + }, + "human_motion_embedder.1.bias": { + "shape": [ + 3072 + ], + "dtype": "torch.float32" + }, + "human_motion_embedder.4.weight": { + "shape": [ + 1536, + 3072 + ], + "dtype": "torch.float32" + }, + "human_motion_embedder.4.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + } + }, + "loops": { + "fit_loop": { + "state_dict": { + "keys": [], + "count": 0 + }, + "epoch_loop.state_dict": { + "keys": [ + "_batches_that_stepped" + ], + "count": 1 + }, + "epoch_loop.batch_progress": { + "keys": [ + "total", + "current", + "is_last_batch" + ], + "count": 3 + }, + "epoch_loop.scheduler_progress": { + "keys": [ + "total", + "current" + ], + "count": 2 + }, + "epoch_loop.automatic_optimization.state_dict": { + "keys": [], + "count": 0 + }, + "epoch_loop.automatic_optimization.optim_progress": { + "keys": [ + "optimizer" + ], + "count": 1 + }, + "epoch_loop.manual_optimization.state_dict": { + "keys": [], + "count": 0 + }, + "epoch_loop.manual_optimization.optim_step_progress": { + "keys": [ + "total", + "current" + ], + "count": 2 + }, + "epoch_loop.val_loop.state_dict": { + "keys": [], + "count": 0 + }, + "epoch_loop.val_loop.batch_progress": { + "keys": [ + "total", + "current", + "is_last_batch" + ], + "count": 3 + }, + "epoch_progress": { + "keys": [ + "total", + "current" + ], + "count": 2 + } + }, + "validate_loop": { + "state_dict": { + "keys": [], + "count": 0 + }, + "batch_progress": { + "keys": [ + "total", + "current", + "is_last_batch" + ], + "count": 3 + } + }, + "test_loop": { + "state_dict": { + "keys": [], + "count": 0 + }, + "batch_progress": { + "keys": [ + "total", + "current", + "is_last_batch" + ], + "count": 3 + } + }, + "predict_loop": { + "state_dict": { + "keys": [], + "count": 0 + }, + "batch_progress": { + "keys": [ + "total", + "current" + ], + "count": 2 + } + } + }, + "callbacks": { + "ModelCheckpoint{'monitor': 'step', 'mode': 'max', 'every_n_train_steps': 2000, 'every_n_epochs': 0, 'train_time_interval': None}": { + "monitor": "step", + "best_model_score": { + "shape": [], + "dtype": "torch.int64" + }, + "best_model_path": "/autodl-fs/data/pangxinlei/models/OMG/hfd-xl-english-split20261026/checkpoints/step-00088000.ckpt", + "current_score": { + "shape": [], + "dtype": "torch.int64" + }, + "dirpath": "/autodl-fs/data/pangxinlei/models/OMG/hfd-xl-english-split20261026/checkpoints", + "best_k_models": { + "keys": [ + "/autodl-fs/data/pangxinlei/models/OMG/hfd-xl-english-split20261026/checkpoints/step-00086000.ckpt", + "/autodl-fs/data/pangxinlei/models/OMG/hfd-xl-english-split20261026/checkpoints/step-00088000.ckpt" + ], + "count": 2 + }, + "kth_best_model_path": "/autodl-fs/data/pangxinlei/models/OMG/hfd-xl-english-split20261026/checkpoints/step-00086000.ckpt", + "kth_value": { + "shape": [], + "dtype": "torch.int64" + }, + "last_model_path": "/autodl-fs/data/pangxinlei/models/OMG/hfd-xl-english-split20261026/checkpoints/last.ckpt" + }, + "ModelCheckpoint{'monitor': 'val/loss', 'mode': 'min', 'every_n_train_steps': 0, 'every_n_epochs': 1, 'train_time_interval': None}": { + "monitor": "val/loss", + "best_model_score": { + "shape": [], + "dtype": "torch.float32" + }, + "best_model_path": "/autodl-fs/data/pangxinlei/models/OMG/hfd-xl-english-split20261026/best/best.ckpt", + "current_score": { + "shape": [], + "dtype": "torch.float32" + }, + "dirpath": "/autodl-fs/data/pangxinlei/models/OMG/hfd-xl-english-split20261026/best", + "best_k_models": { + "keys": [ + "/autodl-fs/data/pangxinlei/models/OMG/hfd-xl-english-split20261026/best/best.ckpt" + ], + "count": 1 + }, + "kth_best_model_path": "/autodl-fs/data/pangxinlei/models/OMG/hfd-xl-english-split20261026/best/best.ckpt", + "kth_value": { + "shape": [], + "dtype": "torch.float32" + }, + "last_model_path": "" + } + }, + "optimizer_states": { + "count": 1, + "sample": [ + { + "state": { + "keys": [ + "0", + "1", + "2", + "3", + "4", + "5", + "6", + "7", + "8", + "9", + "10", + "11", + "12", + "13", + "14", + "15", + "16", + "17", + "18", + "19", + "20", + "21", + "22", + "23", + "24", + "25", + "26", + "27", + "28", + "29" + ], + "count": 314 + }, + "param_groups": { + "count": 1, + "sample": [ + { + "keys": [ + "lr", + "betas", + "eps", + "weight_decay", + "amsgrad", + "maximize", + "foreach", + "capturable", + "differentiable", + "fused", + "decoupled_weight_decay", + "initial_lr", + "params" + ], + "count": 13 + } + ] + } + } + ] + }, + "lr_schedulers": { + "count": 1, + "sample": [ + { + "_milestones": { + "count": 1, + "sample": [ + 2000 + ] + }, + "last_epoch": 88000, + "_last_lr": { + "count": 1, + "sample": [ + 3.1559556582929402e-06 + ] + }, + "_schedulers": { + "count": 2, + "sample": [ + { + "keys": [ + "start_factor", + "end_factor", + "total_iters", + "base_lrs", + "last_epoch", + "_step_count", + "_get_lr_called_within_step", + "_last_lr" + ], + "count": 8 + }, + { + "keys": [ + "T_max", + "eta_min", + "base_lrs", + "last_epoch", + "_step_count", + "_get_lr_called_within_step", + "_last_lr" + ], + "count": 7 + } + ] + } + } + ] + }, + "hfd_resume": { + "version": 1, + "rng": { + "python": { + "count": 3, + "sample": [ + 3, + { + "count": 625, + "sample": [ + 2147483648, + 3564348608 + ] + } + ] + }, + "numpy": { + "count": 5, + "sample": [ + "MT19937", + { + "count": 624, + "sample": [ + 42, + 3107752595 + ] + } + ] + }, + "torch": { + "shape": [ + 5056 + ], + "dtype": "torch.uint8" + }, + "cuda": { + "count": 1, + "sample": [ + { + "shape": [ + 16 + ], + "dtype": "torch.uint8" + } + ] + } + }, + "train_pool": { + "step": 88000, + "slots": { + "count": 512, + "sample": [ + 313022, + 201301 + ] + }, + "batch_size": 512, + "refresh": 256, + "rng": { + "keys": [ + "bit_generator", + "state", + "has_uint32", + "uinteger" + ], + "count": 4 + }, + "identity": { + "keys": [ + "sessions", + "native_source_sha256", + "asr", + "recipe_sha256", + "foreground_language", + "ordered_session_selection_sha256", + "final_release" + ], + "count": 7 + } + }, + "val_pool": { + "step": 16, + "slots": { + "count": 512, + "sample": [ + 6295, + 7193 + ] + }, + "batch_size": 512, + "refresh": 256, + "rng": { + "keys": [ + "bit_generator", + "state", + "has_uint32", + "uinteger" + ], + "count": 4 + }, + "identity": { + "keys": [ + "sessions", + "native_source_sha256", + "asr", + "recipe_sha256", + "foreground_language", + "ordered_session_selection_sha256", + "final_release" + ], + "count": 7 + } + } + }, + "omg_model_architecture": { + "format": "omg.model_architecture", + "version": 1, + "denoiser_type": "omg.generation.denoisers.transformer.MotionTransformerDenoiser", + "frame_cond_injection": "per_layer_film", + "attention": { + "rotary_self_attention_qk_norm": true, + "cross_attention_qk_norm": true + } + } +} \ No newline at end of file diff --git a/handoff-v1/inventory/omg-xl-source-audit.json b/handoff-v1/inventory/omg-xl-source-audit.json new file mode 100644 index 0000000000000000000000000000000000000000..3baa2b9d991ed4bec282e5d908aa3a4aaa3db203 --- /dev/null +++ b/handoff-v1/inventory/omg-xl-source-audit.json @@ -0,0 +1,586 @@ +{ + "hfd_bridge/storage.py": { + "expected": "02c4fe4eed5f6d17bf140568af365b0abc8e5843df018d65553828890058a1fa", + "actual": "02c4fe4eed5f6d17bf140568af365b0abc8e5843df018d65553828890058a1fa" + }, + "hfd_bridge/__init__.py": { + "expected": "885528eb35015189a0bcd2dc62b60c61878efcf959caef68ff69e93501cafd5c", + "actual": "885528eb35015189a0bcd2dc62b60c61878efcf959caef68ff69e93501cafd5c" + }, + "hfd_bridge/loader.py": { + "expected": "13772137dc5cbb036bd5b5e4ee6ffe8cb45ba8d033ad580887b4dd63b253ac9b", + "actual": "13772137dc5cbb036bd5b5e4ee6ffe8cb45ba8d033ad580887b4dd63b253ac9b" + }, + "hfd_bridge/sessions.py": { + "expected": "c271aecf69e52acaee178b1381e465f22af82732f606d4e0bd5f891e2bb86076", + "actual": "c271aecf69e52acaee178b1381e465f22af82732f606d4e0bd5f891e2bb86076" + }, + "hfd_bridge/asr.py": { + "expected": "4baaf4faecfdce2054844a41326760220af301eb11d6fad892bed8c7380dceb0", + "actual": "4baaf4faecfdce2054844a41326760220af301eb11d6fad892bed8c7380dceb0" + }, + "hfd_bridge/training.py": { + "expected": "b513a632bcf010c83c23a35bcfd09e189149ff8c0d7d667d8966ec09e6e55970", + "actual": "b513a632bcf010c83c23a35bcfd09e189149ff8c0d7d667d8966ec09e6e55970" + }, + "hfd_bridge/native.py": { + "expected": "72ca9f675c2dfc7a2e304191da7b4d4016cb6bc226ccc22f9e7d0aeb2889c9f4", + "actual": "72ca9f675c2dfc7a2e304191da7b4d4016cb6bc226ccc22f9e7d0aeb2889c9f4" + }, + "hfd_bridge/paths.py": { + "expected": "2aaee82ecad571dcc1186c275e270af6c854c670db4ae354ee984d23a50b8bf3", + "actual": "2aaee82ecad571dcc1186c275e270af6c854c670db4ae354ee984d23a50b8bf3" + }, + "hfd_bridge/configured_data.py": { + "expected": "f916e0f9d4cdd2a4120fe94a9f1dfe59bbd24606d71ce1941023f04f97f181cd", + "actual": "f916e0f9d4cdd2a4120fe94a9f1dfe59bbd24606d71ce1941023f04f97f181cd" + }, + "hfd_bridge/textop_checks.py": { + "expected": "00662346781a638298cfa344b9ce8a58b6e8bba11261cd5783fdca956f4df7c8", + "actual": "00662346781a638298cfa344b9ce8a58b6e8bba11261cd5783fdca956f4df7c8" + }, + "hfd_bridge/omg_checks.py": { + "expected": "094d997043d4e0d868261b7f25b9452fca832dfa48572aeac4a3f2ba9f42d172", + "actual": "094d997043d4e0d868261b7f25b9452fca832dfa48572aeac4a3f2ba9f42d172" + }, + "hfd_bridge/textop_condition.py": { + "expected": "4fca0f4a02df3b843a908b28b64766877585aa3300977359d21fdb3adb5a3c28", + "actual": "4fca0f4a02df3b843a908b28b64766877585aa3300977359d21fdb3adb5a3c28" + }, + "hfd_bridge/joints.py": { + "expected": "7827c82ca3463a026db6007c41a66528e1294485b4acb081f27f09449da2d154", + "actual": "7827c82ca3463a026db6007c41a66528e1294485b4acb081f27f09449da2d154" + }, + "hfd_bridge/audio_condition.py": { + "expected": "cab6795bdddc36aa63d4f54c3c8e8aaee7958c86882abe35ae76f67a79b4b945", + "actual": "cab6795bdddc36aa63d4f54c3c8e8aaee7958c86882abe35ae76f67a79b4b945" + }, + "hfd_bridge/batch_asr.py": { + "expected": "9d4f0ba1d382f45c7a0adf562bc008485fce7cd00e86e29607528676f503f8f1", + "actual": "9d4f0ba1d382f45c7a0adf562bc008485fce7cd00e86e29607528676f503f8f1" + }, + "hfd_bridge/ondemand.py": { + "expected": "940339aa16a6fa25a27ace23a956017f7325813b0cd2a49f76e6f1fb1d014cc0", + "actual": "940339aa16a6fa25a27ace23a956017f7325813b0cd2a49f76e6f1fb1d014cc0" + }, + "hfd_bridge/ondemand_data.py": { + "expected": "24c61e36623eaca0bdc2f7bff2c8ce307946c4a5afef18f0427b96c06c9ed57e", + "actual": "24c61e36623eaca0bdc2f7bff2c8ce307946c4a5afef18f0427b96c06c9ed57e" + }, + "hfd_bridge/prefetch.py": { + "expected": "b227251cc6470f980231999bbecbeb1fc583c754eda56e70d4f3c84ed1845d64", + "actual": "b227251cc6470f980231999bbecbeb1fc583c754eda56e70d4f3c84ed1845d64" + }, + "hfd_bridge/human_sessions.py": { + "expected": "0461ea66b1c324ec38e9e833ee8eb19b95a24c3f478b094fb600dcdc3c53cdea", + "actual": "0461ea66b1c324ec38e9e833ee8eb19b95a24c3f478b094fb600dcdc3c53cdea" + }, + "hfd_bridge/genmo.py": { + "expected": "3c898d2828267f38f8ba9f41aae62c4a52cbae87cf63dd2cef2b6634f7160060", + "actual": "3c898d2828267f38f8ba9f41aae62c4a52cbae87cf63dd2cef2b6634f7160060" + }, + "hfd_bridge/genmo_audio.py": { + "expected": "1cdfa6ebc19a75af7540b2280f5dc9c3ecc584ed1c386c434de32e3777020c6f", + "actual": "1cdfa6ebc19a75af7540b2280f5dc9c3ecc584ed1c386c434de32e3777020c6f" + }, + "hfd_bridge/clocks.py": { + "expected": "27054c0b2d3e7d4087eb8fc3c59c6bcf5a581ca2e3022341e306732dcbe92637", + "actual": "27054c0b2d3e7d4087eb8fc3c59c6bcf5a581ca2e3022341e306732dcbe92637" + }, + "hfd_bridge/semantics.py": { + "expected": "42ed59d58609095b8053d0598a9c575d52af9b8508e699162743cc11599f03b4", + "actual": "42ed59d58609095b8053d0598a9c575d52af9b8508e699162743cc11599f03b4" + }, + "hfd_bridge/reference_text.py": { + "expected": "50da93fed3be2bad05035bf978c9be41ba7b48aaf3ce7c43a879519e7386a121", + "actual": "50da93fed3be2bad05035bf978c9be41ba7b48aaf3ce7c43a879519e7386a121" + }, + "hfd_bridge/reference_runtime.py": { + "expected": "2b7fe129152d0f7ea3a4a4e3f5fc9ae799e805138a0d76487e9f59065fe7d975", + "actual": "2b7fe129152d0f7ea3a4a4e3f5fc9ae799e805138a0d76487e9f59065fe7d975" + }, + "hfd_bridge/resume.py": { + "expected": "e4e373802e0dd21e39fa2d899eff524080ce1735ca86eefbf4ef2080f170ffbf", + "actual": "e4e373802e0dd21e39fa2d899eff524080ce1735ca86eefbf4ef2080f170ffbf" + }, + "hfd_bridge/resume_lightning.py": { + "expected": "69e4c6bf11a199f3a979227eb551870c44e371fcae0af924ec64d2d58d3ce7a6", + "actual": "69e4c6bf11a199f3a979227eb551870c44e371fcae0af924ec64d2d58d3ce7a6" + }, + "hfd_bridge/index_cache.py": { + "expected": "d75cbcb0105448c61c841e525a1f17e63c55c83ad9201104e8118e22f7728b17", + "actual": "d75cbcb0105448c61c841e525a1f17e63c55c83ad9201104e8118e22f7728b17" + }, + "hfd_bridge/genmo_data.py": { + "expected": "4b102d1eecb0af9757a76d9600ede345c3f21f92166276a22cbf8e144fa9a38c", + "actual": "4b102d1eecb0af9757a76d9600ede345c3f21f92166276a22cbf8e144fa9a38c" + }, + "hfd_bridge/genmo_validation.py": { + "expected": "fa265cf9df51a4b36db977be945547dd78bdf105db9aed8ccba940b1a977f5e4", + "actual": "fa265cf9df51a4b36db977be945547dd78bdf105db9aed8ccba940b1a977f5e4" + }, + "hfd_bridge/kimodo_data.py": { + "expected": "35b3b7d8fc8f36fe65eb311e6e65e667fe1fdbcdfc23e58f37fb75c1c9e9feae", + "actual": "35b3b7d8fc8f36fe65eb311e6e65e667fe1fdbcdfc23e58f37fb75c1c9e9feae" + }, + "hfd_bridge/kimodo.py": { + "expected": "081a1047b853a2ee143aad2a5699f33cbecd0b4f93a07ad085760e16665f20ce", + "actual": "081a1047b853a2ee143aad2a5699f33cbecd0b4f93a07ad085760e16665f20ce" + }, + "hfd_bridge/kimodo_recovery.py": { + "expected": "b9177ff13b55b753430dd14d46778b723d2a03ce8927313c246c5d7ce780809b", + "actual": "b9177ff13b55b753430dd14d46778b723d2a03ce8927313c246c5d7ce780809b" + }, + "hfd_bridge/kimodo_closed_loop.py": { + "expected": "38d7e022b1450a715b35542cd23a70b1491a6dbd307dbb4be730b9f562100f91", + "actual": "38d7e022b1450a715b35542cd23a70b1491a6dbd307dbb4be730b9f562100f91" + }, + "hfd_bridge/kimodo_causal_rep.py": { + "expected": "c0c40003a3d36ec35e8490fa1165ac88723178b15a0e118d2d1d4c9f7ed7c537", + "actual": "c0c40003a3d36ec35e8490fa1165ac88723178b15a0e118d2d1d4c9f7ed7c537" + }, + "hfd_bridge/kimodo_diffusion_training.py": { + "expected": "078fe8cf1ee39d910a6ca5ef94d67f20c4398cfe505895a089acb80c05336539", + "actual": "078fe8cf1ee39d910a6ca5ef94d67f20c4398cfe505895a089acb80c05336539" + }, + "hfd_bridge/kimodo_text_worker.py": { + "expected": "a275c121a9364ae85b6233a67f3acb63ceddb2c9da490c2ea7d476f4cc710751", + "actual": "a275c121a9364ae85b6233a67f3acb63ceddb2c9da490c2ea7d476f4cc710751" + }, + "hfd_bridge/kimodo_full_data.py": { + "expected": "c62b4bb2673f1875792edb3403347f76bb52fa2360d92a9c1091eae10d43f905", + "actual": "c62b4bb2673f1875792edb3403347f76bb52fa2360d92a9c1091eae10d43f905" + }, + "hfd_bridge/kimodo_checkpoint.py": { + "expected": "b9cd24df225c91a6212c354983e3f51acb495373747f303c6096000af37f7d59", + "actual": "b9cd24df225c91a6212c354983e3f51acb495373747f303c6096000af37f7d59" + }, + "hfd_bridge/kimodo_finetuning.py": { + "expected": "a2c54f04af2b516c14d00025c8b695093ff09da6f04b0b73eb3269817bc0a544", + "actual": "a2c54f04af2b516c14d00025c8b695093ff09da6f04b0b73eb3269817bc0a544" + }, + "hfd_bridge/reference_assets.py": { + "expected": "e977a1d804807629fe925d511acdd25b488a11acc3767c0234020af8f699a43c", + "actual": "e977a1d804807629fe925d511acdd25b488a11acc3767c0234020af8f699a43c" + }, + "hfd_bridge/omg_model_loader.py": { + "expected": "f972080dd6140e1574124be615c9351001b51a73c5774b63bd1c063671bd56aa", + "actual": "f972080dd6140e1574124be615c9351001b51a73c5774b63bd1c063671bd56aa" + }, + "hfd_bridge/genmo_geometry.py": { + "expected": "4d3e912a858b8d3d4ddf73941c3120d5304ed70187c02fc969730816e155b2cb", + "actual": "4d3e912a858b8d3d4ddf73941c3120d5304ed70187c02fc969730816e155b2cb" + }, + "hfd_bridge/genmo_training.py": { + "expected": "556100608e466ca801abd623d24af30eebe7f9e3ac598c6709f813f07d5d45e9", + "actual": "556100608e466ca801abd623d24af30eebe7f9e3ac598c6709f813f07d5d45e9" + }, + "hfd_bridge/genmo_prefetch.py": { + "expected": "0f00c9c5c951f53a0b55cad0c71d470c9312381e75ec31f756f606cf20309141", + "actual": "0f00c9c5c951f53a0b55cad0c71d470c9312381e75ec31f756f606cf20309141" + }, + "hfd_bridge/hymotion_rep.py": { + "expected": "ee2f8282720d311b98d3ceb8cc72f63219622b5f61a138af6eb3221181a89f7a", + "actual": "ee2f8282720d311b98d3ceb8cc72f63219622b5f61a138af6eb3221181a89f7a" + }, + "hfd_bridge/hymotion_model.py": { + "expected": "ed29d2d03b247ed681a09dc230697885c5ab8fa8b30f82c4a827aa8fe0b41597", + "actual": "ed29d2d03b247ed681a09dc230697885c5ab8fa8b30f82c4a827aa8fe0b41597" + }, + "hfd_bridge/hymotion_data.py": { + "expected": "79503602a9a4de4f67931fa944a198d6f438b531a363ffbe545ecc4dd54f97e7", + "actual": "79503602a9a4de4f67931fa944a198d6f438b531a363ffbe545ecc4dd54f97e7" + }, + "hfd_bridge/hymotion_text.py": { + "expected": "f0238ffc3656e2a90a26360c5e9bc8c6de3d2d6be0af6e87787c83237914e98b", + "actual": "f0238ffc3656e2a90a26360c5e9bc8c6de3d2d6be0af6e87787c83237914e98b" + }, + "hfd_bridge/hymotion_rollout.py": { + "expected": "6e27c4e24bd42a211df7be2a0271337af4e658d66e5d303633912bbd40313517", + "actual": "6e27c4e24bd42a211df7be2a0271337af4e658d66e5d303633912bbd40313517" + }, + "hfd_bridge/hymotion_validation.py": { + "expected": "6043090f19315c9fde677168fed06dca1fb390b559a7b9668e0ba46799074cd2", + "actual": "6043090f19315c9fde677168fed06dca1fb390b559a7b9668e0ba46799074cd2" + }, + "hfd_bridge/hymotion_training.py": { + "expected": "792b76cfc35f58b0c7c94b40a660f00c4971fff1d978ffd0c923b9f26a5fa733", + "actual": "792b76cfc35f58b0c7c94b40a660f00c4971fff1d978ffd0c923b9f26a5fa733" + }, + "hfd_bridge/human_eval.py": { + "expected": "c943c332315f16f41cb53beee8303f2d94a4b522e678cfd346e8a6c9ec203b71", + "actual": "c943c332315f16f41cb53beee8303f2d94a4b522e678cfd346e8a6c9ec203b71" + }, + "hfd_bridge/omg_rollout.py": { + "expected": "24ecdec4a2c698bd8cdc2b4ac6e1c0ad960c252b5a9f9a15d57a7716545fc4d7", + "actual": "24ecdec4a2c698bd8cdc2b4ac6e1c0ad960c252b5a9f9a15d57a7716545fc4d7" + }, + "hfd_bridge/final_split.py": { + "expected": "d0f0adef291d17d3b73477258b5de74bb8ed477ca1952c9b83ddb115efd34305", + "actual": "d0f0adef291d17d3b73477258b5de74bb8ed477ca1952c9b83ddb115efd34305" + }, + "hfd_bridge/final_third_party.py": { + "expected": "3926a0f72d419bb5006d8a62136a8f5ffd5ddca439c39dd7768f7cf03422d716", + "actual": "3926a0f72d419bb5006d8a62136a8f5ffd5ddca439c39dd7768f7cf03422d716" + }, + "hfd_bridge/replay_numerics.py": { + "expected": "827c70aa9b2308ac33b6fccee2c4dc88df58d6b89d902d8c215adfdf918ab231", + "actual": "827c70aa9b2308ac33b6fccee2c4dc88df58d6b89d902d8c215adfdf918ab231" + }, + "hfd_bridge/textop_text_cache.py": { + "expected": "3d5c8a187e461144aa97cc539661e6e90b475800379c0c096b243bcd337a0e38", + "actual": "3d5c8a187e461144aa97cc539661e6e90b475800379c0c096b243bcd337a0e38" + }, + "hfd_bridge/hot_assets.py": { + "expected": "43830324054a551ef4ed364ad45d9b2dfd8830b7dde2661f4404b1ffc8cabe10", + "actual": "43830324054a551ef4ed364ad45d9b2dfd8830b7dde2661f4404b1ffc8cabe10" + }, + "hfd_bridge/hymotion_autoregressive.py": { + "expected": "cfbcc9329760364b22c1871fc69c42b77787ed0ab622a2c5fc1a6bed93479d86", + "actual": "cfbcc9329760364b22c1871fc69c42b77787ed0ab622a2c5fc1a6bed93479d86" + }, + "hfd_bridge/genmo_runtime.py": { + "expected": "b4bc26373388974ca50a5e76d382b210950b805ef19b514fcc90adbf555bdc2e", + "actual": "b4bc26373388974ca50a5e76d382b210950b805ef19b514fcc90adbf555bdc2e" + }, + "hfd_bridge/genmo_replay_audit.py": { + "expected": "fb8ac4ac51489ebe9faca4b08f6f2159aea4bb33040ea6c04c7c99c6641c4ae2", + "actual": "fb8ac4ac51489ebe9faca4b08f6f2159aea4bb33040ea6c04c7c99c6641c4ae2" + }, + "hfd_bridge/genmo_safety.py": { + "expected": "f5f20e57cef26e62ec4bb5c0eb1e9d74d30e814b1740cef9f36008936e234158", + "actual": "f5f20e57cef26e62ec4bb5c0eb1e9d74d30e814b1740cef9f36008936e234158" + }, + "hfd_bridge/hymotion_metrics.py": { + "expected": "48d79ed03b833fc948eeb466890aefe6bdb66ed88860f6fbd49b1a5ea0c2c04f", + "actual": "48d79ed03b833fc948eeb466890aefe6bdb66ed88860f6fbd49b1a5ea0c2c04f" + }, + "hfd_bridge/hymotion_evaluation.py": { + "expected": "fe72667281f04a55bc8c88375b7ca3cee1f1b6de2c23fcd7d45f1e73c1530a99", + "actual": "fe72667281f04a55bc8c88375b7ca3cee1f1b6de2c23fcd7d45f1e73c1530a99" + }, + "hfd_bridge/omg_scaling.py": { + "expected": "0ed0b1e0a8c72e039e5c461b193595b4e50da4bfc2f155ec3d4754923af51360", + "actual": "0ed0b1e0a8c72e039e5c461b193595b4e50da4bfc2f155ec3d4754923af51360" + }, + "upstream/omg/download_data.py": { + "expected": "a247118ce874864affc5153a23d94788871d3b6d327bcc33ceeba45279d7732c", + "actual": "a247118ce874864affc5153a23d94788871d3b6d327bcc33ceeba45279d7732c" + }, + "upstream/omg/download_models.py": { + "expected": "811132811b35321174adc69a754cf64f9437eeba5d19f6ad02625ef9d9f73b57", + "actual": "811132811b35321174adc69a754cf64f9437eeba5d19f6ad02625ef9d9f73b57" + }, + "upstream/omg/a.py": { + "expected": "66e1c68c0a631d6fef0d7d13948299671baf72ec813fe025c1cac1e222fd4c63", + "actual": "66e1c68c0a631d6fef0d7d13948299671baf72ec813fe025c1cac1e222fd4c63" + }, + "upstream/omg/smoke.py": { + "expected": "8008c82e92250f148512e3127b939df3aded7025a89473738d024444d21edcfd", + "actual": "8008c82e92250f148512e3127b939df3aded7025a89473738d024444d21edcfd" + }, + "upstream/omg/validate_model.py": { + "expected": "36023f739cd8dfe24179f9247b5de04ca425d7a0d73c99b166bbeaea778a975f", + "actual": "36023f739cd8dfe24179f9247b5de04ca425d7a0d73c99b166bbeaea778a975f" + }, + "upstream/omg/validate_setup.py": { + "expected": "4e086d4615c621386cb54885aad0ede341c24a71a1668c3e5386f368c02f3741", + "actual": "4e086d4615c621386cb54885aad0ede341c24a71a1668c3e5386f368c02f3741" + }, + "upstream/omg/paper_table.py": { + "expected": "97c16c1c044639362f46cac439159384a1d79dc80a90c839c5e2ce471d8060d2", + "actual": "97c16c1c044639362f46cac439159384a1d79dc80a90c839c5e2ce471d8060d2" + }, + "upstream/omg/paper_main_table.py": { + "expected": "8f95f11b00882e4574df37e2f44d5843a15073004f0b9f78450895c65939196f", + "actual": "8f95f11b00882e4574df37e2f44d5843a15073004f0b9f78450895c65939196f" + }, + "upstream/omg/scripts/prepare_omg_inputs.py": { + "expected": "99bbc364b5f380ae7c2cc0fa53657722f896d3613d8b98937881d06197beccb4", + "actual": "99bbc364b5f380ae7c2cc0fa53657722f896d3613d8b98937881d06197beccb4" + }, + "upstream/omg/scripts/extract_omg_paper_row.py": { + "expected": "f64cb31b5a19efba6f3e1b1771d3d5175f6610392a588d118014e98790b82503", + "actual": "f64cb31b5a19efba6f3e1b1771d3d5175f6610392a588d118014e98790b82503" + }, + "upstream/omg/scripts/psi0_omg_generate.py": { + "expected": "38ccbd4c8f1d62b470193ae6e3598445f22408e2df6edc6b267ed46b4873f579", + "actual": "38ccbd4c8f1d62b470193ae6e3598445f22408e2df6edc6b267ed46b4873f579" + }, + "upstream/omg/scripts/verify_omg_artifact_match.py": { + "expected": "46815bd446b956583ae7feeb0ed03a58b6dde8cc78ecb4a1d4fd1a3a6b737ee7", + "actual": "46815bd446b956583ae7feeb0ed03a58b6dde8cc78ecb4a1d4fd1a3a6b737ee7" + }, + "upstream/omg/src/omg/__init__.py": { + "expected": "e06eff3bd6a146e0082518eb791164e26b80b548f9bdcd21c9f0aca1220bca34", + "actual": "e06eff3bd6a146e0082518eb791164e26b80b548f9bdcd21c9f0aca1220bca34" + }, + "upstream/omg/src/omg/callbacks/__init__.py": { + "expected": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", + "actual": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855" + }, + "upstream/omg/src/omg/callbacks/divergence_guard.py": { + "expected": "c1ec2773dddda3bd800247bca0abd3eca46aeae3c78ebae8050b98739292b61a", + "actual": "c1ec2773dddda3bd800247bca0abd3eca46aeae3c78ebae8050b98739292b61a" + }, + "upstream/omg/src/omg/callbacks/global_step_progress.py": { + "expected": "bd21180de0140a38fa2601f51b56c94aef346c757c140786e88b907720f3eb15", + "actual": "bd21180de0140a38fa2601f51b56c94aef346c757c140786e88b907720f3eb15" + }, + "upstream/omg/src/omg/callbacks/grad_norm.py": { + "expected": "c41e7e0bac01c984c85c38baf6de4b9e10ded95875a79ddc1c9d85ded5b9b742", + "actual": "c41e7e0bac01c984c85c38baf6de4b9e10ded95875a79ddc1c9d85ded5b9b742" + }, + "upstream/omg/src/omg/callbacks/speed_timer.py": { + "expected": "29ae0ec76a0545576bc31f240a4beab74f5c5e9ae50a0078c1ed46b8326fc360", + "actual": "29ae0ec76a0545576bc31f240a4beab74f5c5e9ae50a0078c1ed46b8326fc360" + }, + "upstream/omg/src/omg/callbacks/training_diagnostics.py": { + "expected": "6e348e5cca9e690379cdb7a3364678b666d7c53d0a7405d5269ca713574a6917", + "actual": "6e348e5cca9e690379cdb7a3364678b666d7c53d0a7405d5269ca713574a6917" + }, + "upstream/omg/src/omg/cli/__init__.py": { + "expected": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", + "actual": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855" + }, + "upstream/omg/src/omg/cli/data/__init__.py": { + "expected": "052be341b9c1b05585d3a27073a741d768974cfdb2a91659a687ed1d3d6291f9", + "actual": "052be341b9c1b05585d3a27073a741d768974cfdb2a91659a687ed1d3d6291f9" + }, + "upstream/omg/src/omg/cli/data/materialize_episode_cache.py": { + "expected": "c8e61cbcf17f55e17b47ad2a3e5c2aa0919d9f5bcc2a93edbf5a519e79f9793a", + "actual": "c8e61cbcf17f55e17b47ad2a3e5c2aa0919d9f5bcc2a93edbf5a519e79f9793a" + }, + "upstream/omg/src/omg/cli/data/validate_episode_cache.py": { + "expected": "6840e28c774964583d88c6402c201ab36c0150ec67139403a66782364173c878", + "actual": "6840e28c774964583d88c6402c201ab36c0150ec67139403a66782364173c878" + }, + "upstream/omg/src/omg/cli/generation/__init__.py": { + "expected": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", + "actual": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855" + }, + "upstream/omg/src/omg/cli/generation/compute_stats.py": { + "expected": "849ed1ac16612e23e09a887e5a030c5e0c496610e9afe882ce98e99d570bd6ca", + "actual": "849ed1ac16612e23e09a887e5a030c5e0c496610e9afe882ce98e99d570bd6ca" + }, + "upstream/omg/src/omg/cli/generation/train.py": { + "expected": "a39cc0c6cfccd7b6660bbd9ab4f88f90bc634bb95ce30e22492c8a6cb74a7c8c", + "actual": "a39cc0c6cfccd7b6660bbd9ab4f88f90bc634bb95ce30e22492c8a6cb74a7c8c" + }, + "upstream/omg/src/omg/cli/generation/benchmark.py": { + "expected": "ff813f344953da4b02388837f5a510230f8d4430393a88f205261faaf28c5dee", + "actual": "ff813f344953da4b02388837f5a510230f8d4430393a88f205261faaf28c5dee" + }, + "upstream/omg/src/omg/cli/generation/physical_benchmark.py": { + "expected": "f48b2b9d2d312c44d0e33eebf0c7fd95d8f2ce65a871fe253ef8ff9ceee30dd0", + "actual": "f48b2b9d2d312c44d0e33eebf0c7fd95d8f2ce65a871fe253ef8ff9ceee30dd0" + }, + "upstream/omg/src/omg/cli/generation/generate.py": { + "expected": "66eda839732e76a3b0fc39a767e6a238aa93bbf13ef2f3ddb6e1f9021948bc4a", + "actual": "66eda839732e76a3b0fc39a767e6a238aa93bbf13ef2f3ddb6e1f9021948bc4a" + }, + "upstream/omg/src/omg/cli/generation/export_onnx.py": { + "expected": "9cacee46fea5e18f8a9f9ffd52b25110f89e80be20d21acce85fe066392c00c1", + "actual": "9cacee46fea5e18f8a9f9ffd52b25110f89e80be20d21acce85fe066392c00c1" + }, + "upstream/omg/src/omg/core/__init__.py": { + "expected": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", + "actual": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855" + }, + "upstream/omg/src/omg/core/logging.py": { + "expected": "86c57564f7d8614d71d95e0ce377724642c8c3aa44501d9c317c33bf9ba7dce7", + "actual": "86c57564f7d8614d71d95e0ce377724642c8c3aa44501d9c317c33bf9ba7dce7" + }, + "upstream/omg/src/omg/core/paths.py": { + "expected": "6bb254b7219e70b58fc8581479857bfb7ee77a9f8937faf5dcec784dd1173a4d", + "actual": "6bb254b7219e70b58fc8581479857bfb7ee77a9f8937faf5dcec784dd1173a4d" + }, + "upstream/omg/src/omg/core/tensor.py": { + "expected": "998e6f24948983c15f3257f055a452e07b96cc0f13b357cb4881f4545bddb7d0", + "actual": "998e6f24948983c15f3257f055a452e07b96cc0f13b357cb4881f4545bddb7d0" + }, + "upstream/omg/src/omg/data/__init__.py": { + "expected": "4e8cd9d824b5b834a673011d1cd95cf0227a1deb60ee917ba85ec364d51db21f", + "actual": "4e8cd9d824b5b834a673011d1cd95cf0227a1deb60ee917ba85ec364d51db21f" + }, + "upstream/omg/src/omg/data/datamodule.py": { + "expected": "94cbaf84b019d559dd7810494694c156803d318f23a5ec95549315c59ea795d6", + "actual": "94cbaf84b019d559dd7810494694c156803d318f23a5ec95549315c59ea795d6" + }, + "upstream/omg/src/omg/data/episode_cache.py": { + "expected": "96639991cb75a3db602c608d929e4f742caef615166198e66731c406f068207e", + "actual": "96639991cb75a3db602c608d929e4f742caef615166198e66731c406f068207e" + }, + "upstream/omg/src/omg/data/episode_cache_inspect.py": { + "expected": "b7f43c2a5f7a1c63dcb9c243e79018ebff3ecce056fba22ac9b1948e754feb20", + "actual": "b7f43c2a5f7a1c63dcb9c243e79018ebff3ecce056fba22ac9b1948e754feb20" + }, + "upstream/omg/src/omg/data/lerobot_dataset.py": { + "expected": "78172eb344a37cab8457133a957eb25691ad78054f71db0c39171f241be12831", + "actual": "78172eb344a37cab8457133a957eb25691ad78054f71db0c39171f241be12831" + }, + "upstream/omg/src/omg/data/windowing.py": { + "expected": "88710f0c24d6c696670b702b9ce77609c59acd2106389fa769a00d39a98f2afb", + "actual": "88710f0c24d6c696670b702b9ce77609c59acd2106389fa769a00d39a98f2afb" + }, + "upstream/omg/src/omg/generation/__init__.py": { + "expected": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", + "actual": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855" + }, + "upstream/omg/src/omg/generation/architecture.py": { + "expected": "f77548e52394c78b813ad36dcd5fb8c8981ba393e5c723126491df6c171aeabf", + "actual": "f77548e52394c78b813ad36dcd5fb8c8981ba393e5c723126491df6c171aeabf" + }, + "upstream/omg/src/omg/generation/conditions/__init__.py": { + "expected": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", + "actual": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855" + }, + "upstream/omg/src/omg/generation/conditions/t5.py": { + "expected": "67c0f834106ff860e612a978bdba6ef503d2cf0dfb24e831af0bba723e95c890", + "actual": "67c0f834106ff860e612a978bdba6ef503d2cf0dfb24e831af0bba723e95c890" + }, + "upstream/omg/src/omg/generation/denoisers/__init__.py": { + "expected": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", + "actual": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855" + }, + "upstream/omg/src/omg/generation/denoisers/interface.py": { + "expected": "c1f8b5f2a93e6cf0a4b5df730f9f744c95d9b58bc5245e3cb15176a61065e1dd", + "actual": "c1f8b5f2a93e6cf0a4b5df730f9f744c95d9b58bc5245e3cb15176a61065e1dd" + }, + "upstream/omg/src/omg/generation/denoisers/transformer.py": { + "expected": "ae968a16ba6193607721f5c0c1910122818797407df3ee0f0fafa3728c0e398a", + "actual": "ae968a16ba6193607721f5c0c1910122818797407df3ee0f0fafa3728c0e398a" + }, + "upstream/omg/src/omg/generation/diffusion/__init__.py": { + "expected": "fab4af0ec5b7b38892ffb309d6073841de00f8f37e5b4c2da1657e1b30c92e1c", + "actual": "fab4af0ec5b7b38892ffb309d6073841de00f8f37e5b4c2da1657e1b30c92e1c" + }, + "upstream/omg/src/omg/generation/diffusion/forcing.py": { + "expected": "935b2f38db1d1d3c496df0b03cadaeb032ef1539522169f2731b87a8724362b5", + "actual": "935b2f38db1d1d3c496df0b03cadaeb032ef1539522169f2731b87a8724362b5" + }, + "upstream/omg/src/omg/generation/diffusion/guided.py": { + "expected": "3fd42c7ae9d01497223b08421ae3becc0a7be81846c1cf08c8bb944d8949b870", + "actual": "3fd42c7ae9d01497223b08421ae3becc0a7be81846c1cf08c8bb944d8949b870" + }, + "upstream/omg/src/omg/generation/diffusion/objectives.py": { + "expected": "db05bcbc32268e5d21298e19130e9e8f4b23eff7d1b679886306cd7c94195675", + "actual": "db05bcbc32268e5d21298e19130e9e8f4b23eff7d1b679886306cd7c94195675" + }, + "upstream/omg/src/omg/generation/diffusion/schedules.py": { + "expected": "a06c3f3ab1d81e07f680d44f8fb9ef56e8bd2a6e4c5e4ac911c316217f6e92e0", + "actual": "a06c3f3ab1d81e07f680d44f8fb9ef56e8bd2a6e4c5e4ac911c316217f6e92e0" + }, + "upstream/omg/src/omg/generation/export/__init__.py": { + "expected": "0ee7ed60aa43848dad0a4ba0e5dd29d815aab6dc37f4ae1c46d8e56a2ebc1770", + "actual": "0ee7ed60aa43848dad0a4ba0e5dd29d815aab6dc37f4ae1c46d8e56a2ebc1770" + }, + "upstream/omg/src/omg/generation/export/onnx.py": { + "expected": "168ad808867669b69cd69a419d5b8d2c3fd124c69c20daeca53b6fc59a0a21dd", + "actual": "168ad808867669b69cd69a419d5b8d2c3fd124c69c20daeca53b6fc59a0a21dd" + }, + "upstream/omg/src/omg/generation/losses/__init__.py": { + "expected": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", + "actual": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855" + }, + "upstream/omg/src/omg/generation/losses/contact.py": { + "expected": "75927b00fc5584988f2ecedbcfbba9af3ba80206966b0a91642ec133db5e4267", + "actual": "75927b00fc5584988f2ecedbcfbba9af3ba80206966b0a91642ec133db5e4267" + }, + "upstream/omg/src/omg/generation/losses/fk.py": { + "expected": "75927b00fc5584988f2ecedbcfbba9af3ba80206966b0a91642ec133db5e4267", + "actual": "75927b00fc5584988f2ecedbcfbba9af3ba80206966b0a91642ec133db5e4267" + }, + "upstream/omg/src/omg/generation/losses/motion.py": { + "expected": "00ee200046658dafcae2b9ec439ed0c3203ab533f4dd9d4a8475541df74c2fa5", + "actual": "00ee200046658dafcae2b9ec439ed0c3203ab533f4dd9d4a8475541df74c2fa5" + }, + "upstream/omg/src/omg/generation/losses/terrain.py": { + "expected": "75927b00fc5584988f2ecedbcfbba9af3ba80206966b0a91642ec133db5e4267", + "actual": "75927b00fc5584988f2ecedbcfbba9af3ba80206966b0a91642ec133db5e4267" + }, + "upstream/omg/src/omg/generation/metrics/__init__.py": { + "expected": "f686c86d57bab24aa1dd27cfae2bbab88757d7909e74cd7d4b9525a892ba025e", + "actual": "f686c86d57bab24aa1dd27cfae2bbab88757d7909e74cd7d4b9525a892ba025e" + }, + "upstream/omg/src/omg/generation/metrics/physical.py": { + "expected": "7aa2d8fc7d29b9d3ec836c3d23d0b5a7fc64a1eea5359c7300f6e6e191904f29", + "actual": "7aa2d8fc7d29b9d3ec836c3d23d0b5a7fc64a1eea5359c7300f6e6e191904f29" + }, + "upstream/omg/src/omg/generation/models/__init__.py": { + "expected": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", + "actual": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855" + }, + "upstream/omg/src/omg/generation/models/motion_generator.py": { + "expected": "05947c8e73646f634e9bbedf0a60f2becaf1ded21b9d1093f1084eaaefef2a3b", + "actual": "05947c8e73646f634e9bbedf0a60f2becaf1ded21b9d1093f1084eaaefef2a3b" + }, + "upstream/omg/src/omg/motion/__init__.py": { + "expected": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", + "actual": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855" + }, + "upstream/omg/src/omg/motion/feature_codec.py": { + "expected": "3d148816baf239dbc938210c959954c3c2fac4c6d9914ad48b851976ed7e2960", + "actual": "3d148816baf239dbc938210c959954c3c2fac4c6d9914ad48b851976ed7e2960" + }, + "upstream/omg/src/omg/motion/representation.py": { + "expected": "457328078b2a2652e06c39e5fa06c8a0d8ba9223dcf24e508a33cdd1806f45a7", + "actual": "457328078b2a2652e06c39e5fa06c8a0d8ba9223dcf24e508a33cdd1806f45a7" + }, + "upstream/omg/src/omg/pipeline/__init__.py": { + "expected": "3806048000a9bdc918ace57c68aabcd3c56bb503c4998b103d1593553a23a895", + "actual": "3806048000a9bdc918ace57c68aabcd3c56bb503c4998b103d1593553a23a895" + }, + "upstream/omg/src/omg/pipeline/audio.py": { + "expected": "d88a0b51fd72d80fdf7da1344082e7799f481fb114a246ed4e66c0b9745f3d09", + "actual": "d88a0b51fd72d80fdf7da1344082e7799f481fb114a246ed4e66c0b9745f3d09" + }, + "upstream/omg/src/omg/pipeline/condition_sequence.py": { + "expected": "e831393b0b843219520ee8a836b7cbc0b406295e60947cba1a443f1f78a81377", + "actual": "e831393b0b843219520ee8a836b7cbc0b406295e60947cba1a443f1f78a81377" + }, + "upstream/omg/src/omg/pipeline/human.py": { + "expected": "d5be4338bc69308450c18331da4cf9a93fc9e02bd5596dce92a3e58248ff14d0", + "actual": "d5be4338bc69308450c18331da4cf9a93fc9e02bd5596dce92a3e58248ff14d0" + }, + "upstream/omg/src/omg/pipeline/planner.py": { + "expected": "fa7aa07eebf9f56287301a6ab13fb7a42bf8b310e152b0916033019d7d55ce0c", + "actual": "fa7aa07eebf9f56287301a6ab13fb7a42bf8b310e152b0916033019d7d55ce0c" + }, + "upstream/omg/src/omg/robots/__init__.py": { + "expected": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", + "actual": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855" + }, + "upstream/omg/src/omg/robots/g1/__init__.py": { + "expected": "dc46237fccb843612112274586e9e00dee6a0e472f4ce6b93da267b64f8f4c19", + "actual": "dc46237fccb843612112274586e9e00dee6a0e472f4ce6b93da267b64f8f4c19" + }, + "upstream/omg/src/omg/robots/g1/constants.py": { + "expected": "1f2d0fd495de8765349ad24bdb8e8b327ffdf80c16b13c665fabfe8e9b1a39ea", + "actual": "1f2d0fd495de8765349ad24bdb8e8b327ffdf80c16b13c665fabfe8e9b1a39ea" + }, + "upstream/omg/src/omg/robots/g1/kinematics.py": { + "expected": "c8adbae0d96ef85ba948e27f1fb16bed35d8529c08a6d71d1943d31644585db4", + "actual": "c8adbae0d96ef85ba948e27f1fb16bed35d8529c08a6d71d1943d31644585db4" + }, + "upstream/omg/src/omg/runtime/__init__.py": { + "expected": "9728869f63d93ada09a0f77fc10e98f2add27f4ab84e999d228becca102315fe", + "actual": "9728869f63d93ada09a0f77fc10e98f2add27f4ab84e999d228becca102315fe" + }, + "upstream/omg/src/omg/runtime/checkpoint_io.py": { + "expected": "1b9ce178d3e5c1bba39c77d3098ea7e64788d8c81c48acdebaec09cc1be88e5e", + "actual": "1b9ce178d3e5c1bba39c77d3098ea7e64788d8c81c48acdebaec09cc1be88e5e" + }, + "upstream/omg/src/omg/runtime/onnx_providers.py": { + "expected": "5ed87e442ebf9c56b4065fc011c89d7745ac5bf163158fbf0732db160d0e5673", + "actual": "5ed87e442ebf9c56b4065fc011c89d7745ac5bf163158fbf0732db160d0e5673" + }, + "upstream/omg/src/omg/utils/rotation_conversions.py": { + "expected": "627ead38f2e15b0001eea501f6be09e7b6c309c21f1938ef419f65f8e1eb7ea5", + "actual": "627ead38f2e15b0001eea501f6be09e7b6c309c21f1938ef419f65f8e1eb7ea5" + }, + "train.py": { + "expected": "61f49739e14275a671dc3343afcf2d6b225da547a952f963037d545afe680dd4", + "actual": "61f49739e14275a671dc3343afcf2d6b225da547a952f963037d545afe680dd4" + } +} \ No newline at end of file diff --git a/handoff-v1/inventory/omg-xl-step-00088000.ckpt.json b/handoff-v1/inventory/omg-xl-step-00088000.ckpt.json new file mode 100644 index 0000000000000000000000000000000000000000..43497b017c14fb0e75eefa8dbb976412fd9e4d7c --- /dev/null +++ b/handoff-v1/inventory/omg-xl-step-00088000.ckpt.json @@ -0,0 +1,3432 @@ +{ + "epoch": 0, + "global_step": 88000, + "pytorch-lightning_version": "2.6.6", + "state_dict": { + "denoiser.time_embed.proj.0.weight": { + "shape": [ + 1536, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.time_embed.proj.0.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.time_embed.proj.2.weight": { + "shape": [ + 1536, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.time_embed.proj.2.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.input_proj.weight": { + "shape": [ + 1536, + 1661 + ], + "dtype": "torch.float32" + }, + "denoiser.input_proj.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.text_proj.weight": { + "shape": [ + 1536, + 768 + ], + "dtype": "torch.float32" + }, + "denoiser.text_proj.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.extra_proj.0.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.extra_proj.0.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.extra_proj.1.weight": { + "shape": [ + 1536, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.extra_proj.1.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.norm_cross.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.norm_cross.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.cross_attn.in_proj_weight": { + "shape": [ + 4608, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.cross_attn.in_proj_bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.cross_attn.out_proj.weight": { + "shape": [ + 1536, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.cross_attn.out_proj.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.norm_self.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.norm_self.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.self_attn.qkv.weight": { + "shape": [ + 4608, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.self_attn.qkv.bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.self_attn.out_proj.weight": { + "shape": [ + 1536, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.self_attn.out_proj.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.norm_ff.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.norm_ff.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.ff.0.weight": { + "shape": [ + 6144, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.ff.0.bias": { + "shape": [ + 6144 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.ff.3.weight": { + "shape": [ + 1536, + 6144 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.ff.3.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.audio_film.0.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.audio_film.0.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.audio_film.1.weight": { + "shape": [ + 3072, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.audio_film.1.bias": { + "shape": [ + 3072 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.human_motion_film.0.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.human_motion_film.0.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.human_motion_film.1.weight": { + "shape": [ + 3072, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.human_motion_film.1.bias": { + "shape": [ + 3072 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.norm_cross.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.norm_cross.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.cross_attn.in_proj_weight": { + "shape": [ + 4608, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.cross_attn.in_proj_bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.cross_attn.out_proj.weight": { + "shape": [ + 1536, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.cross_attn.out_proj.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.norm_self.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.norm_self.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.self_attn.qkv.weight": { + "shape": [ + 4608, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.self_attn.qkv.bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.self_attn.out_proj.weight": { + "shape": [ + 1536, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.self_attn.out_proj.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.norm_ff.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.norm_ff.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.ff.0.weight": { + "shape": [ + 6144, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.ff.0.bias": { + "shape": [ + 6144 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.ff.3.weight": { + "shape": [ + 1536, + 6144 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.ff.3.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.audio_film.0.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.audio_film.0.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.audio_film.1.weight": { + "shape": [ + 3072, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.audio_film.1.bias": { + "shape": [ + 3072 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.human_motion_film.0.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.human_motion_film.0.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.human_motion_film.1.weight": { + "shape": [ + 3072, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.human_motion_film.1.bias": { + "shape": [ + 3072 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.norm_cross.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.norm_cross.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.cross_attn.in_proj_weight": { + "shape": [ + 4608, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.cross_attn.in_proj_bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.cross_attn.out_proj.weight": { + "shape": [ + 1536, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.cross_attn.out_proj.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.norm_self.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.norm_self.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.self_attn.qkv.weight": { + "shape": [ + 4608, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.self_attn.qkv.bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.self_attn.out_proj.weight": { + "shape": [ + 1536, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.self_attn.out_proj.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.norm_ff.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.norm_ff.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.ff.0.weight": { + "shape": [ + 6144, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.ff.0.bias": { + "shape": [ + 6144 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.ff.3.weight": { + "shape": [ + 1536, + 6144 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.ff.3.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.audio_film.0.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.audio_film.0.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.audio_film.1.weight": { + "shape": [ + 3072, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.audio_film.1.bias": { + "shape": [ + 3072 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.human_motion_film.0.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.human_motion_film.0.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.human_motion_film.1.weight": { + "shape": [ + 3072, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.human_motion_film.1.bias": { + "shape": [ + 3072 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.norm_cross.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.norm_cross.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.cross_attn.in_proj_weight": { + "shape": [ + 4608, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.cross_attn.in_proj_bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.cross_attn.out_proj.weight": { + "shape": [ + 1536, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.cross_attn.out_proj.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.norm_self.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.norm_self.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.self_attn.qkv.weight": { + "shape": [ + 4608, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.self_attn.qkv.bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.self_attn.out_proj.weight": { + "shape": [ + 1536, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.self_attn.out_proj.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.norm_ff.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.norm_ff.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.ff.0.weight": { + "shape": [ + 6144, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.ff.0.bias": { + "shape": [ + 6144 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.ff.3.weight": { + "shape": [ + 1536, + 6144 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.ff.3.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.audio_film.0.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.audio_film.0.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.audio_film.1.weight": { + "shape": [ + 3072, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.audio_film.1.bias": { + "shape": [ + 3072 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.human_motion_film.0.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.human_motion_film.0.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.human_motion_film.1.weight": { + "shape": [ + 3072, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.human_motion_film.1.bias": { + "shape": [ + 3072 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.norm_cross.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.norm_cross.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.cross_attn.in_proj_weight": { + "shape": [ + 4608, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.cross_attn.in_proj_bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.cross_attn.out_proj.weight": { + "shape": [ + 1536, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.cross_attn.out_proj.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.norm_self.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.norm_self.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.self_attn.qkv.weight": { + "shape": [ + 4608, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.self_attn.qkv.bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.self_attn.out_proj.weight": { + "shape": [ + 1536, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.self_attn.out_proj.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.norm_ff.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.norm_ff.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.ff.0.weight": { + "shape": [ + 6144, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.ff.0.bias": { + "shape": [ + 6144 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.ff.3.weight": { + "shape": [ + 1536, + 6144 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.ff.3.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.audio_film.0.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.audio_film.0.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.audio_film.1.weight": { + "shape": [ + 3072, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.audio_film.1.bias": { + "shape": [ + 3072 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.human_motion_film.0.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.human_motion_film.0.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.human_motion_film.1.weight": { + "shape": [ + 3072, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.human_motion_film.1.bias": { + "shape": [ + 3072 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.norm_cross.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.norm_cross.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.cross_attn.in_proj_weight": { + "shape": [ + 4608, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.cross_attn.in_proj_bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.cross_attn.out_proj.weight": { + "shape": [ + 1536, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.cross_attn.out_proj.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.norm_self.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.norm_self.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.self_attn.qkv.weight": { + "shape": [ + 4608, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.self_attn.qkv.bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.self_attn.out_proj.weight": { + "shape": [ + 1536, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.self_attn.out_proj.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.norm_ff.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.norm_ff.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.ff.0.weight": { + "shape": [ + 6144, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.ff.0.bias": { + "shape": [ + 6144 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.ff.3.weight": { + "shape": [ + 1536, + 6144 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.ff.3.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.audio_film.0.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.audio_film.0.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.audio_film.1.weight": { + "shape": [ + 3072, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.audio_film.1.bias": { + "shape": [ + 3072 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.human_motion_film.0.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.human_motion_film.0.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.human_motion_film.1.weight": { + "shape": [ + 3072, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.human_motion_film.1.bias": { + "shape": [ + 3072 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.norm_cross.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.norm_cross.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.cross_attn.in_proj_weight": { + "shape": [ + 4608, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.cross_attn.in_proj_bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.cross_attn.out_proj.weight": { + "shape": [ + 1536, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.cross_attn.out_proj.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.norm_self.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.norm_self.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.self_attn.qkv.weight": { + "shape": [ + 4608, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.self_attn.qkv.bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.self_attn.out_proj.weight": { + "shape": [ + 1536, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.self_attn.out_proj.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.norm_ff.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.norm_ff.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.ff.0.weight": { + "shape": [ + 6144, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.ff.0.bias": { + "shape": [ + 6144 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.ff.3.weight": { + "shape": [ + 1536, + 6144 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.ff.3.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.audio_film.0.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.audio_film.0.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.audio_film.1.weight": { + "shape": [ + 3072, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.audio_film.1.bias": { + "shape": [ + 3072 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.human_motion_film.0.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.human_motion_film.0.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.human_motion_film.1.weight": { + "shape": [ + 3072, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.human_motion_film.1.bias": { + "shape": [ + 3072 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.norm_cross.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.norm_cross.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.cross_attn.in_proj_weight": { + "shape": [ + 4608, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.cross_attn.in_proj_bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.cross_attn.out_proj.weight": { + "shape": [ + 1536, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.cross_attn.out_proj.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.norm_self.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.norm_self.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.self_attn.qkv.weight": { + "shape": [ + 4608, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.self_attn.qkv.bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.self_attn.out_proj.weight": { + "shape": [ + 1536, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.self_attn.out_proj.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.norm_ff.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.norm_ff.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.ff.0.weight": { + "shape": [ + 6144, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.ff.0.bias": { + "shape": [ + 6144 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.ff.3.weight": { + "shape": [ + 1536, + 6144 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.ff.3.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.audio_film.0.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.audio_film.0.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.audio_film.1.weight": { + "shape": [ + 3072, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.audio_film.1.bias": { + "shape": [ + 3072 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.human_motion_film.0.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.human_motion_film.0.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.human_motion_film.1.weight": { + "shape": [ + 3072, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.human_motion_film.1.bias": { + "shape": [ + 3072 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.norm_cross.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.norm_cross.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.cross_attn.in_proj_weight": { + "shape": [ + 4608, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.cross_attn.in_proj_bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.cross_attn.out_proj.weight": { + "shape": [ + 1536, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.cross_attn.out_proj.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.norm_self.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.norm_self.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.self_attn.qkv.weight": { + "shape": [ + 4608, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.self_attn.qkv.bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.self_attn.out_proj.weight": { + "shape": [ + 1536, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.self_attn.out_proj.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.norm_ff.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.norm_ff.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.ff.0.weight": { + "shape": [ + 6144, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.ff.0.bias": { + "shape": [ + 6144 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.ff.3.weight": { + "shape": [ + 1536, + 6144 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.ff.3.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.audio_film.0.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.audio_film.0.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.audio_film.1.weight": { + "shape": [ + 3072, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.audio_film.1.bias": { + "shape": [ + 3072 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.human_motion_film.0.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.human_motion_film.0.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.human_motion_film.1.weight": { + "shape": [ + 3072, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.human_motion_film.1.bias": { + "shape": [ + 3072 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.norm_cross.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.norm_cross.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.cross_attn.in_proj_weight": { + "shape": [ + 4608, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.cross_attn.in_proj_bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.cross_attn.out_proj.weight": { + "shape": [ + 1536, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.cross_attn.out_proj.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.norm_self.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.norm_self.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.self_attn.qkv.weight": { + "shape": [ + 4608, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.self_attn.qkv.bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.self_attn.out_proj.weight": { + "shape": [ + 1536, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.self_attn.out_proj.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.norm_ff.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.norm_ff.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.ff.0.weight": { + "shape": [ + 6144, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.ff.0.bias": { + "shape": [ + 6144 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.ff.3.weight": { + "shape": [ + 1536, + 6144 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.ff.3.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.audio_film.0.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.audio_film.0.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.audio_film.1.weight": { + "shape": [ + 3072, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.audio_film.1.bias": { + "shape": [ + 3072 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.human_motion_film.0.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.human_motion_film.0.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.human_motion_film.1.weight": { + "shape": [ + 3072, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.human_motion_film.1.bias": { + "shape": [ + 3072 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.norm_cross.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.norm_cross.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.cross_attn.in_proj_weight": { + "shape": [ + 4608, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.cross_attn.in_proj_bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.cross_attn.out_proj.weight": { + "shape": [ + 1536, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.cross_attn.out_proj.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.norm_self.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.norm_self.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.self_attn.qkv.weight": { + "shape": [ + 4608, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.self_attn.qkv.bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.self_attn.out_proj.weight": { + "shape": [ + 1536, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.self_attn.out_proj.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.norm_ff.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.norm_ff.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.ff.0.weight": { + "shape": [ + 6144, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.ff.0.bias": { + "shape": [ + 6144 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.ff.3.weight": { + "shape": [ + 1536, + 6144 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.ff.3.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.audio_film.0.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.audio_film.0.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.audio_film.1.weight": { + "shape": [ + 3072, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.audio_film.1.bias": { + "shape": [ + 3072 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.human_motion_film.0.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.human_motion_film.0.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.human_motion_film.1.weight": { + "shape": [ + 3072, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.human_motion_film.1.bias": { + "shape": [ + 3072 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.norm_cross.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.norm_cross.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.cross_attn.in_proj_weight": { + "shape": [ + 4608, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.cross_attn.in_proj_bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.cross_attn.out_proj.weight": { + "shape": [ + 1536, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.cross_attn.out_proj.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.norm_self.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.norm_self.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.self_attn.qkv.weight": { + "shape": [ + 4608, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.self_attn.qkv.bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.self_attn.out_proj.weight": { + "shape": [ + 1536, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.self_attn.out_proj.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.norm_ff.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.norm_ff.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.ff.0.weight": { + "shape": [ + 6144, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.ff.0.bias": { + "shape": [ + 6144 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.ff.3.weight": { + "shape": [ + 1536, + 6144 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.ff.3.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.audio_film.0.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.audio_film.0.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.audio_film.1.weight": { + "shape": [ + 3072, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.audio_film.1.bias": { + "shape": [ + 3072 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.human_motion_film.0.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.human_motion_film.0.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.human_motion_film.1.weight": { + "shape": [ + 3072, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.human_motion_film.1.bias": { + "shape": [ + 3072 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.norm_cross.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.norm_cross.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.cross_attn.in_proj_weight": { + "shape": [ + 4608, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.cross_attn.in_proj_bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.cross_attn.out_proj.weight": { + "shape": [ + 1536, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.cross_attn.out_proj.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.norm_self.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.norm_self.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.self_attn.qkv.weight": { + "shape": [ + 4608, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.self_attn.qkv.bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.self_attn.out_proj.weight": { + "shape": [ + 1536, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.self_attn.out_proj.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.norm_ff.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.norm_ff.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.ff.0.weight": { + "shape": [ + 6144, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.ff.0.bias": { + "shape": [ + 6144 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.ff.3.weight": { + "shape": [ + 1536, + 6144 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.ff.3.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.audio_film.0.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.audio_film.0.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.audio_film.1.weight": { + "shape": [ + 3072, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.audio_film.1.bias": { + "shape": [ + 3072 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.human_motion_film.0.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.human_motion_film.0.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.human_motion_film.1.weight": { + "shape": [ + 3072, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.human_motion_film.1.bias": { + "shape": [ + 3072 + ], + "dtype": "torch.float32" + }, + "denoiser.norm.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.norm.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.output.weight": { + "shape": [ + 125, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.output.bias": { + "shape": [ + 125 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.shared.weight": { + "shape": [ + 32128, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.embed_tokens.weight": { + "shape": [ + 32128, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.0.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.0.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.0.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.0.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.0.layer.0.SelfAttention.relative_attention_bias.weight": { + "shape": [ + 32, + 12 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.0.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.0.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.0.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.0.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.1.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.1.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.1.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.1.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.1.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.1.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.1.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.1.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.2.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.2.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.2.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.2.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.2.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.2.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.2.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.2.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.3.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.3.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.3.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.3.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.3.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.3.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.3.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.3.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.4.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.4.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.4.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.4.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.4.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.4.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.4.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.4.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.5.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.5.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.5.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.5.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.5.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.5.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.5.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.5.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.6.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.6.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.6.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.6.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.6.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.6.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.6.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.6.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.7.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.7.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.7.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.7.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.7.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.7.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.7.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.7.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.8.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.8.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.8.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.8.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.8.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.8.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.8.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.8.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.9.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.9.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.9.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.9.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.9.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.9.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.9.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.9.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.10.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.10.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.10.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.10.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.10.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.10.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.10.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.10.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.11.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.11.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.11.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.11.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.11.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.11.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.11.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.11.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.final_layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "history_projector.0.weight": { + "shape": [ + 125 + ], + "dtype": "torch.float32" + }, + "history_projector.0.bias": { + "shape": [ + 125 + ], + "dtype": "torch.float32" + }, + "history_projector.1.weight": { + "shape": [ + 1536, + 125 + ], + "dtype": "torch.float32" + }, + "history_projector.1.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "history_projector.3.weight": { + "shape": [ + 1536, + 1536 + ], + "dtype": "torch.float32" + }, + "history_projector.3.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "audio_embedder.0.weight": { + "shape": [ + 35 + ], + "dtype": "torch.float32" + }, + "audio_embedder.0.bias": { + "shape": [ + 35 + ], + "dtype": "torch.float32" + }, + "audio_embedder.1.weight": { + "shape": [ + 3072, + 35 + ], + "dtype": "torch.float32" + }, + "audio_embedder.1.bias": { + "shape": [ + 3072 + ], + "dtype": "torch.float32" + }, + "audio_embedder.4.weight": { + "shape": [ + 1536, + 3072 + ], + "dtype": "torch.float32" + }, + "audio_embedder.4.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "human_motion_embedder.0.weight": { + "shape": [ + 66 + ], + "dtype": "torch.float32" + }, + "human_motion_embedder.0.bias": { + "shape": [ + 66 + ], + "dtype": "torch.float32" + }, + "human_motion_embedder.1.weight": { + "shape": [ + 3072, + 66 + ], + "dtype": "torch.float32" + }, + "human_motion_embedder.1.bias": { + "shape": [ + 3072 + ], + "dtype": "torch.float32" + }, + "human_motion_embedder.4.weight": { + "shape": [ + 1536, + 3072 + ], + "dtype": "torch.float32" + }, + "human_motion_embedder.4.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + } + }, + "loops": { + "fit_loop": { + "state_dict": { + "keys": [], + "count": 0 + }, + "epoch_loop.state_dict": { + "keys": [ + "_batches_that_stepped" + ], + "count": 1 + }, + "epoch_loop.batch_progress": { + "keys": [ + "total", + "current", + "is_last_batch" + ], + "count": 3 + }, + "epoch_loop.scheduler_progress": { + "keys": [ + "total", + "current" + ], + "count": 2 + }, + "epoch_loop.automatic_optimization.state_dict": { + "keys": [], + "count": 0 + }, + "epoch_loop.automatic_optimization.optim_progress": { + "keys": [ + "optimizer" + ], + "count": 1 + }, + "epoch_loop.manual_optimization.state_dict": { + "keys": [], + "count": 0 + }, + "epoch_loop.manual_optimization.optim_step_progress": { + "keys": [ + "total", + "current" + ], + "count": 2 + }, + "epoch_loop.val_loop.state_dict": { + "keys": [], + "count": 0 + }, + "epoch_loop.val_loop.batch_progress": { + "keys": [ + "total", + "current", + "is_last_batch" + ], + "count": 3 + }, + "epoch_progress": { + "keys": [ + "total", + "current" + ], + "count": 2 + } + }, + "validate_loop": { + "state_dict": { + "keys": [], + "count": 0 + }, + "batch_progress": { + "keys": [ + "total", + "current", + "is_last_batch" + ], + "count": 3 + } + }, + "test_loop": { + "state_dict": { + "keys": [], + "count": 0 + }, + "batch_progress": { + "keys": [ + "total", + "current", + "is_last_batch" + ], + "count": 3 + } + }, + "predict_loop": { + "state_dict": { + "keys": [], + "count": 0 + }, + "batch_progress": { + "keys": [ + "total", + "current" + ], + "count": 2 + } + } + }, + "callbacks": { + "ModelCheckpoint{'monitor': 'step', 'mode': 'max', 'every_n_train_steps': 2000, 'every_n_epochs': 0, 'train_time_interval': None}": { + "monitor": "step", + "best_model_score": { + "shape": [], + "dtype": "torch.int64" + }, + "best_model_path": "/autodl-fs/data/pangxinlei/models/OMG/hfd-xl-english-split20261026/checkpoints/step-00088000.ckpt", + "current_score": { + "shape": [], + "dtype": "torch.int64" + }, + "dirpath": "/autodl-fs/data/pangxinlei/models/OMG/hfd-xl-english-split20261026/checkpoints", + "best_k_models": { + "keys": [ + "/autodl-fs/data/pangxinlei/models/OMG/hfd-xl-english-split20261026/checkpoints/step-00086000.ckpt", + "/autodl-fs/data/pangxinlei/models/OMG/hfd-xl-english-split20261026/checkpoints/step-00088000.ckpt" + ], + "count": 2 + }, + "kth_best_model_path": "/autodl-fs/data/pangxinlei/models/OMG/hfd-xl-english-split20261026/checkpoints/step-00086000.ckpt", + "kth_value": { + "shape": [], + "dtype": "torch.int64" + }, + "last_model_path": "/autodl-fs/data/pangxinlei/models/OMG/hfd-xl-english-split20261026/checkpoints/last.ckpt" + }, + "ModelCheckpoint{'monitor': 'val/loss', 'mode': 'min', 'every_n_train_steps': 0, 'every_n_epochs': 1, 'train_time_interval': None}": { + "monitor": "val/loss", + "best_model_score": { + "shape": [], + "dtype": "torch.float32" + }, + "best_model_path": "/autodl-fs/data/pangxinlei/models/OMG/hfd-xl-english-split20261026/best/best.ckpt", + "current_score": { + "shape": [], + "dtype": "torch.float32" + }, + "dirpath": "/autodl-fs/data/pangxinlei/models/OMG/hfd-xl-english-split20261026/best", + "best_k_models": { + "keys": [ + "/autodl-fs/data/pangxinlei/models/OMG/hfd-xl-english-split20261026/best/best.ckpt" + ], + "count": 1 + }, + "kth_best_model_path": "/autodl-fs/data/pangxinlei/models/OMG/hfd-xl-english-split20261026/best/best.ckpt", + "kth_value": { + "shape": [], + "dtype": "torch.float32" + }, + "last_model_path": "" + } + }, + "optimizer_states": { + "count": 1, + "sample": [ + { + "state": { + "keys": [ + "0", + "1", + "2", + "3", + "4", + "5", + "6", + "7", + "8", + "9", + "10", + "11", + "12", + "13", + "14", + "15", + "16", + "17", + "18", + "19", + "20", + "21", + "22", + "23", + "24", + "25", + "26", + "27", + "28", + "29" + ], + "count": 314 + }, + "param_groups": { + "count": 1, + "sample": [ + { + "keys": [ + "lr", + "betas", + "eps", + "weight_decay", + "amsgrad", + "maximize", + "foreach", + "capturable", + "differentiable", + "fused", + "decoupled_weight_decay", + "initial_lr", + "params" + ], + "count": 13 + } + ] + } + } + ] + }, + "lr_schedulers": { + "count": 1, + "sample": [ + { + "_milestones": { + "count": 1, + "sample": [ + 2000 + ] + }, + "last_epoch": 88000, + "_last_lr": { + "count": 1, + "sample": [ + 3.1559556582929402e-06 + ] + }, + "_schedulers": { + "count": 2, + "sample": [ + { + "keys": [ + "start_factor", + "end_factor", + "total_iters", + "base_lrs", + "last_epoch", + "_step_count", + "_get_lr_called_within_step", + "_last_lr" + ], + "count": 8 + }, + { + "keys": [ + "T_max", + "eta_min", + "base_lrs", + "last_epoch", + "_step_count", + "_get_lr_called_within_step", + "_last_lr" + ], + "count": 7 + } + ] + } + } + ] + }, + "hfd_resume": { + "version": 1, + "rng": { + "python": { + "count": 3, + "sample": [ + 3, + { + "count": 625, + "sample": [ + 2147483648, + 3564348608 + ] + } + ] + }, + "numpy": { + "count": 5, + "sample": [ + "MT19937", + { + "count": 624, + "sample": [ + 42, + 3107752595 + ] + } + ] + }, + "torch": { + "shape": [ + 5056 + ], + "dtype": "torch.uint8" + }, + "cuda": { + "count": 1, + "sample": [ + { + "shape": [ + 16 + ], + "dtype": "torch.uint8" + } + ] + } + }, + "train_pool": { + "step": 88000, + "slots": { + "count": 512, + "sample": [ + 313022, + 201301 + ] + }, + "batch_size": 512, + "refresh": 256, + "rng": { + "keys": [ + "bit_generator", + "state", + "has_uint32", + "uinteger" + ], + "count": 4 + }, + "identity": { + "keys": [ + "sessions", + "native_source_sha256", + "asr", + "recipe_sha256", + "foreground_language", + "ordered_session_selection_sha256", + "final_release" + ], + "count": 7 + } + }, + "val_pool": { + "step": 16, + "slots": { + "count": 512, + "sample": [ + 6295, + 7193 + ] + }, + "batch_size": 512, + "refresh": 256, + "rng": { + "keys": [ + "bit_generator", + "state", + "has_uint32", + "uinteger" + ], + "count": 4 + }, + "identity": { + "keys": [ + "sessions", + "native_source_sha256", + "asr", + "recipe_sha256", + "foreground_language", + "ordered_session_selection_sha256", + "final_release" + ], + "count": 7 + } + } + }, + "omg_model_architecture": { + "format": "omg.model_architecture", + "version": 1, + "denoiser_type": "omg.generation.denoisers.transformer.MotionTransformerDenoiser", + "frame_cond_injection": "per_layer_film", + "attention": { + "rotary_self_attention_qk_norm": true, + "cross_attention_qk_norm": true + } + } +} \ No newline at end of file diff --git a/handoff-v1/inventory/omg-xl-step-00090000.ckpt.json b/handoff-v1/inventory/omg-xl-step-00090000.ckpt.json new file mode 100644 index 0000000000000000000000000000000000000000..410f2671b17eccd6e3ce9f558a5b378c27e4a5e4 --- /dev/null +++ b/handoff-v1/inventory/omg-xl-step-00090000.ckpt.json @@ -0,0 +1,3432 @@ +{ + "epoch": 0, + "global_step": 90000, + "pytorch-lightning_version": "2.6.6", + "state_dict": { + "denoiser.time_embed.proj.0.weight": { + "shape": [ + 1536, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.time_embed.proj.0.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.time_embed.proj.2.weight": { + "shape": [ + 1536, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.time_embed.proj.2.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.input_proj.weight": { + "shape": [ + 1536, + 1661 + ], + "dtype": "torch.float32" + }, + "denoiser.input_proj.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.text_proj.weight": { + "shape": [ + 1536, + 768 + ], + "dtype": "torch.float32" + }, + "denoiser.text_proj.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.extra_proj.0.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.extra_proj.0.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.extra_proj.1.weight": { + "shape": [ + 1536, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.extra_proj.1.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.norm_cross.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.norm_cross.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.cross_attn.in_proj_weight": { + "shape": [ + 4608, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.cross_attn.in_proj_bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.cross_attn.out_proj.weight": { + "shape": [ + 1536, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.cross_attn.out_proj.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.norm_self.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.norm_self.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.self_attn.qkv.weight": { + "shape": [ + 4608, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.self_attn.qkv.bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.self_attn.out_proj.weight": { + "shape": [ + 1536, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.self_attn.out_proj.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.norm_ff.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.norm_ff.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.ff.0.weight": { + "shape": [ + 6144, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.ff.0.bias": { + "shape": [ + 6144 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.ff.3.weight": { + "shape": [ + 1536, + 6144 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.ff.3.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.audio_film.0.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.audio_film.0.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.audio_film.1.weight": { + "shape": [ + 3072, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.audio_film.1.bias": { + "shape": [ + 3072 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.human_motion_film.0.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.human_motion_film.0.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.human_motion_film.1.weight": { + "shape": [ + 3072, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.0.human_motion_film.1.bias": { + "shape": [ + 3072 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.norm_cross.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.norm_cross.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.cross_attn.in_proj_weight": { + "shape": [ + 4608, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.cross_attn.in_proj_bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.cross_attn.out_proj.weight": { + "shape": [ + 1536, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.cross_attn.out_proj.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.norm_self.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.norm_self.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.self_attn.qkv.weight": { + "shape": [ + 4608, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.self_attn.qkv.bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.self_attn.out_proj.weight": { + "shape": [ + 1536, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.self_attn.out_proj.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.norm_ff.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.norm_ff.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.ff.0.weight": { + "shape": [ + 6144, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.ff.0.bias": { + "shape": [ + 6144 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.ff.3.weight": { + "shape": [ + 1536, + 6144 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.ff.3.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.audio_film.0.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.audio_film.0.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.audio_film.1.weight": { + "shape": [ + 3072, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.audio_film.1.bias": { + "shape": [ + 3072 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.human_motion_film.0.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.human_motion_film.0.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.human_motion_film.1.weight": { + "shape": [ + 3072, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.1.human_motion_film.1.bias": { + "shape": [ + 3072 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.norm_cross.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.norm_cross.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.cross_attn.in_proj_weight": { + "shape": [ + 4608, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.cross_attn.in_proj_bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.cross_attn.out_proj.weight": { + "shape": [ + 1536, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.cross_attn.out_proj.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.norm_self.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.norm_self.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.self_attn.qkv.weight": { + "shape": [ + 4608, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.self_attn.qkv.bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.self_attn.out_proj.weight": { + "shape": [ + 1536, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.self_attn.out_proj.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.norm_ff.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.norm_ff.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.ff.0.weight": { + "shape": [ + 6144, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.ff.0.bias": { + "shape": [ + 6144 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.ff.3.weight": { + "shape": [ + 1536, + 6144 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.ff.3.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.audio_film.0.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.audio_film.0.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.audio_film.1.weight": { + "shape": [ + 3072, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.audio_film.1.bias": { + "shape": [ + 3072 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.human_motion_film.0.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.human_motion_film.0.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.human_motion_film.1.weight": { + "shape": [ + 3072, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.2.human_motion_film.1.bias": { + "shape": [ + 3072 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.norm_cross.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.norm_cross.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.cross_attn.in_proj_weight": { + "shape": [ + 4608, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.cross_attn.in_proj_bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.cross_attn.out_proj.weight": { + "shape": [ + 1536, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.cross_attn.out_proj.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.norm_self.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.norm_self.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.self_attn.qkv.weight": { + "shape": [ + 4608, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.self_attn.qkv.bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.self_attn.out_proj.weight": { + "shape": [ + 1536, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.self_attn.out_proj.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.norm_ff.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.norm_ff.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.ff.0.weight": { + "shape": [ + 6144, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.ff.0.bias": { + "shape": [ + 6144 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.ff.3.weight": { + "shape": [ + 1536, + 6144 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.ff.3.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.audio_film.0.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.audio_film.0.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.audio_film.1.weight": { + "shape": [ + 3072, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.audio_film.1.bias": { + "shape": [ + 3072 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.human_motion_film.0.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.human_motion_film.0.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.human_motion_film.1.weight": { + "shape": [ + 3072, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.3.human_motion_film.1.bias": { + "shape": [ + 3072 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.norm_cross.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.norm_cross.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.cross_attn.in_proj_weight": { + "shape": [ + 4608, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.cross_attn.in_proj_bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.cross_attn.out_proj.weight": { + "shape": [ + 1536, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.cross_attn.out_proj.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.norm_self.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.norm_self.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.self_attn.qkv.weight": { + "shape": [ + 4608, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.self_attn.qkv.bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.self_attn.out_proj.weight": { + "shape": [ + 1536, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.self_attn.out_proj.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.norm_ff.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.norm_ff.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.ff.0.weight": { + "shape": [ + 6144, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.ff.0.bias": { + "shape": [ + 6144 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.ff.3.weight": { + "shape": [ + 1536, + 6144 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.ff.3.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.audio_film.0.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.audio_film.0.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.audio_film.1.weight": { + "shape": [ + 3072, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.audio_film.1.bias": { + "shape": [ + 3072 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.human_motion_film.0.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.human_motion_film.0.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.human_motion_film.1.weight": { + "shape": [ + 3072, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.4.human_motion_film.1.bias": { + "shape": [ + 3072 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.norm_cross.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.norm_cross.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.cross_attn.in_proj_weight": { + "shape": [ + 4608, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.cross_attn.in_proj_bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.cross_attn.out_proj.weight": { + "shape": [ + 1536, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.cross_attn.out_proj.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.norm_self.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.norm_self.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.self_attn.qkv.weight": { + "shape": [ + 4608, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.self_attn.qkv.bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.self_attn.out_proj.weight": { + "shape": [ + 1536, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.self_attn.out_proj.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.norm_ff.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.norm_ff.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.ff.0.weight": { + "shape": [ + 6144, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.ff.0.bias": { + "shape": [ + 6144 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.ff.3.weight": { + "shape": [ + 1536, + 6144 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.ff.3.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.audio_film.0.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.audio_film.0.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.audio_film.1.weight": { + "shape": [ + 3072, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.audio_film.1.bias": { + "shape": [ + 3072 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.human_motion_film.0.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.human_motion_film.0.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.human_motion_film.1.weight": { + "shape": [ + 3072, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.5.human_motion_film.1.bias": { + "shape": [ + 3072 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.norm_cross.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.norm_cross.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.cross_attn.in_proj_weight": { + "shape": [ + 4608, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.cross_attn.in_proj_bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.cross_attn.out_proj.weight": { + "shape": [ + 1536, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.cross_attn.out_proj.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.norm_self.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.norm_self.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.self_attn.qkv.weight": { + "shape": [ + 4608, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.self_attn.qkv.bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.self_attn.out_proj.weight": { + "shape": [ + 1536, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.self_attn.out_proj.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.norm_ff.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.norm_ff.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.ff.0.weight": { + "shape": [ + 6144, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.ff.0.bias": { + "shape": [ + 6144 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.ff.3.weight": { + "shape": [ + 1536, + 6144 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.ff.3.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.audio_film.0.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.audio_film.0.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.audio_film.1.weight": { + "shape": [ + 3072, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.audio_film.1.bias": { + "shape": [ + 3072 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.human_motion_film.0.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.human_motion_film.0.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.human_motion_film.1.weight": { + "shape": [ + 3072, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.6.human_motion_film.1.bias": { + "shape": [ + 3072 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.norm_cross.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.norm_cross.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.cross_attn.in_proj_weight": { + "shape": [ + 4608, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.cross_attn.in_proj_bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.cross_attn.out_proj.weight": { + "shape": [ + 1536, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.cross_attn.out_proj.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.norm_self.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.norm_self.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.self_attn.qkv.weight": { + "shape": [ + 4608, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.self_attn.qkv.bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.self_attn.out_proj.weight": { + "shape": [ + 1536, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.self_attn.out_proj.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.norm_ff.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.norm_ff.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.ff.0.weight": { + "shape": [ + 6144, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.ff.0.bias": { + "shape": [ + 6144 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.ff.3.weight": { + "shape": [ + 1536, + 6144 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.ff.3.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.audio_film.0.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.audio_film.0.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.audio_film.1.weight": { + "shape": [ + 3072, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.audio_film.1.bias": { + "shape": [ + 3072 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.human_motion_film.0.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.human_motion_film.0.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.human_motion_film.1.weight": { + "shape": [ + 3072, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.7.human_motion_film.1.bias": { + "shape": [ + 3072 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.norm_cross.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.norm_cross.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.cross_attn.in_proj_weight": { + "shape": [ + 4608, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.cross_attn.in_proj_bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.cross_attn.out_proj.weight": { + "shape": [ + 1536, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.cross_attn.out_proj.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.norm_self.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.norm_self.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.self_attn.qkv.weight": { + "shape": [ + 4608, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.self_attn.qkv.bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.self_attn.out_proj.weight": { + "shape": [ + 1536, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.self_attn.out_proj.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.norm_ff.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.norm_ff.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.ff.0.weight": { + "shape": [ + 6144, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.ff.0.bias": { + "shape": [ + 6144 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.ff.3.weight": { + "shape": [ + 1536, + 6144 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.ff.3.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.audio_film.0.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.audio_film.0.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.audio_film.1.weight": { + "shape": [ + 3072, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.audio_film.1.bias": { + "shape": [ + 3072 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.human_motion_film.0.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.human_motion_film.0.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.human_motion_film.1.weight": { + "shape": [ + 3072, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.8.human_motion_film.1.bias": { + "shape": [ + 3072 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.norm_cross.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.norm_cross.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.cross_attn.in_proj_weight": { + "shape": [ + 4608, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.cross_attn.in_proj_bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.cross_attn.out_proj.weight": { + "shape": [ + 1536, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.cross_attn.out_proj.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.norm_self.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.norm_self.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.self_attn.qkv.weight": { + "shape": [ + 4608, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.self_attn.qkv.bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.self_attn.out_proj.weight": { + "shape": [ + 1536, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.self_attn.out_proj.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.norm_ff.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.norm_ff.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.ff.0.weight": { + "shape": [ + 6144, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.ff.0.bias": { + "shape": [ + 6144 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.ff.3.weight": { + "shape": [ + 1536, + 6144 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.ff.3.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.audio_film.0.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.audio_film.0.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.audio_film.1.weight": { + "shape": [ + 3072, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.audio_film.1.bias": { + "shape": [ + 3072 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.human_motion_film.0.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.human_motion_film.0.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.human_motion_film.1.weight": { + "shape": [ + 3072, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.9.human_motion_film.1.bias": { + "shape": [ + 3072 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.norm_cross.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.norm_cross.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.cross_attn.in_proj_weight": { + "shape": [ + 4608, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.cross_attn.in_proj_bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.cross_attn.out_proj.weight": { + "shape": [ + 1536, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.cross_attn.out_proj.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.norm_self.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.norm_self.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.self_attn.qkv.weight": { + "shape": [ + 4608, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.self_attn.qkv.bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.self_attn.out_proj.weight": { + "shape": [ + 1536, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.self_attn.out_proj.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.norm_ff.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.norm_ff.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.ff.0.weight": { + "shape": [ + 6144, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.ff.0.bias": { + "shape": [ + 6144 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.ff.3.weight": { + "shape": [ + 1536, + 6144 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.ff.3.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.audio_film.0.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.audio_film.0.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.audio_film.1.weight": { + "shape": [ + 3072, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.audio_film.1.bias": { + "shape": [ + 3072 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.human_motion_film.0.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.human_motion_film.0.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.human_motion_film.1.weight": { + "shape": [ + 3072, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.10.human_motion_film.1.bias": { + "shape": [ + 3072 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.norm_cross.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.norm_cross.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.cross_attn.in_proj_weight": { + "shape": [ + 4608, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.cross_attn.in_proj_bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.cross_attn.out_proj.weight": { + "shape": [ + 1536, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.cross_attn.out_proj.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.norm_self.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.norm_self.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.self_attn.qkv.weight": { + "shape": [ + 4608, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.self_attn.qkv.bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.self_attn.out_proj.weight": { + "shape": [ + 1536, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.self_attn.out_proj.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.norm_ff.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.norm_ff.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.ff.0.weight": { + "shape": [ + 6144, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.ff.0.bias": { + "shape": [ + 6144 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.ff.3.weight": { + "shape": [ + 1536, + 6144 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.ff.3.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.audio_film.0.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.audio_film.0.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.audio_film.1.weight": { + "shape": [ + 3072, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.audio_film.1.bias": { + "shape": [ + 3072 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.human_motion_film.0.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.human_motion_film.0.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.human_motion_film.1.weight": { + "shape": [ + 3072, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.11.human_motion_film.1.bias": { + "shape": [ + 3072 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.norm_cross.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.norm_cross.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.cross_attn.in_proj_weight": { + "shape": [ + 4608, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.cross_attn.in_proj_bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.cross_attn.out_proj.weight": { + "shape": [ + 1536, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.cross_attn.out_proj.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.norm_self.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.norm_self.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.self_attn.qkv.weight": { + "shape": [ + 4608, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.self_attn.qkv.bias": { + "shape": [ + 4608 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.self_attn.out_proj.weight": { + "shape": [ + 1536, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.self_attn.out_proj.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.norm_ff.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.norm_ff.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.ff.0.weight": { + "shape": [ + 6144, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.ff.0.bias": { + "shape": [ + 6144 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.ff.3.weight": { + "shape": [ + 1536, + 6144 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.ff.3.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.audio_film.0.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.audio_film.0.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.audio_film.1.weight": { + "shape": [ + 3072, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.audio_film.1.bias": { + "shape": [ + 3072 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.human_motion_film.0.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.human_motion_film.0.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.human_motion_film.1.weight": { + "shape": [ + 3072, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.layers.12.human_motion_film.1.bias": { + "shape": [ + 3072 + ], + "dtype": "torch.float32" + }, + "denoiser.norm.weight": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.norm.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.output.weight": { + "shape": [ + 125, + 1536 + ], + "dtype": "torch.float32" + }, + "denoiser.output.bias": { + "shape": [ + 125 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.shared.weight": { + "shape": [ + 32128, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.embed_tokens.weight": { + "shape": [ + 32128, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.0.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.0.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.0.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.0.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.0.layer.0.SelfAttention.relative_attention_bias.weight": { + "shape": [ + 32, + 12 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.0.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.0.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.0.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.0.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.1.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.1.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.1.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.1.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.1.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.1.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.1.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.1.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.2.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.2.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.2.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.2.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.2.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.2.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.2.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.2.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.3.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.3.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.3.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.3.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.3.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.3.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.3.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.3.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.4.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.4.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.4.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.4.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.4.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.4.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.4.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.4.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.5.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.5.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.5.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.5.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.5.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.5.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.5.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.5.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.6.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.6.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.6.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.6.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.6.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.6.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.6.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.6.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.7.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.7.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.7.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.7.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.7.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.7.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.7.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.7.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.8.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.8.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.8.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.8.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.8.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.8.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.8.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.8.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.9.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.9.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.9.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.9.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.9.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.9.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.9.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.9.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.10.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.10.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.10.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.10.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.10.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.10.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.10.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.10.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.11.layer.0.SelfAttention.q.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.11.layer.0.SelfAttention.k.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.11.layer.0.SelfAttention.v.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.11.layer.0.SelfAttention.o.weight": { + "shape": [ + 768, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.11.layer.0.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.11.layer.1.DenseReluDense.wi.weight": { + "shape": [ + 3072, + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.11.layer.1.DenseReluDense.wo.weight": { + "shape": [ + 768, + 3072 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.block.11.layer.1.layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "text_encoder.encoder.encoder.final_layer_norm.weight": { + "shape": [ + 768 + ], + "dtype": "torch.float32" + }, + "history_projector.0.weight": { + "shape": [ + 125 + ], + "dtype": "torch.float32" + }, + "history_projector.0.bias": { + "shape": [ + 125 + ], + "dtype": "torch.float32" + }, + "history_projector.1.weight": { + "shape": [ + 1536, + 125 + ], + "dtype": "torch.float32" + }, + "history_projector.1.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "history_projector.3.weight": { + "shape": [ + 1536, + 1536 + ], + "dtype": "torch.float32" + }, + "history_projector.3.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "audio_embedder.0.weight": { + "shape": [ + 35 + ], + "dtype": "torch.float32" + }, + "audio_embedder.0.bias": { + "shape": [ + 35 + ], + "dtype": "torch.float32" + }, + "audio_embedder.1.weight": { + "shape": [ + 3072, + 35 + ], + "dtype": "torch.float32" + }, + "audio_embedder.1.bias": { + "shape": [ + 3072 + ], + "dtype": "torch.float32" + }, + "audio_embedder.4.weight": { + "shape": [ + 1536, + 3072 + ], + "dtype": "torch.float32" + }, + "audio_embedder.4.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "human_motion_embedder.0.weight": { + "shape": [ + 66 + ], + "dtype": "torch.float32" + }, + "human_motion_embedder.0.bias": { + "shape": [ + 66 + ], + "dtype": "torch.float32" + }, + "human_motion_embedder.1.weight": { + "shape": [ + 3072, + 66 + ], + "dtype": "torch.float32" + }, + "human_motion_embedder.1.bias": { + "shape": [ + 3072 + ], + "dtype": "torch.float32" + }, + "human_motion_embedder.4.weight": { + "shape": [ + 1536, + 3072 + ], + "dtype": "torch.float32" + }, + "human_motion_embedder.4.bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + } + }, + "loops": { + "fit_loop": { + "state_dict": { + "keys": [], + "count": 0 + }, + "epoch_loop.state_dict": { + "keys": [ + "_batches_that_stepped" + ], + "count": 1 + }, + "epoch_loop.batch_progress": { + "keys": [ + "total", + "current", + "is_last_batch" + ], + "count": 3 + }, + "epoch_loop.scheduler_progress": { + "keys": [ + "total", + "current" + ], + "count": 2 + }, + "epoch_loop.automatic_optimization.state_dict": { + "keys": [], + "count": 0 + }, + "epoch_loop.automatic_optimization.optim_progress": { + "keys": [ + "optimizer" + ], + "count": 1 + }, + "epoch_loop.manual_optimization.state_dict": { + "keys": [], + "count": 0 + }, + "epoch_loop.manual_optimization.optim_step_progress": { + "keys": [ + "total", + "current" + ], + "count": 2 + }, + "epoch_loop.val_loop.state_dict": { + "keys": [], + "count": 0 + }, + "epoch_loop.val_loop.batch_progress": { + "keys": [ + "total", + "current", + "is_last_batch" + ], + "count": 3 + }, + "epoch_progress": { + "keys": [ + "total", + "current" + ], + "count": 2 + } + }, + "validate_loop": { + "state_dict": { + "keys": [], + "count": 0 + }, + "batch_progress": { + "keys": [ + "total", + "current", + "is_last_batch" + ], + "count": 3 + } + }, + "test_loop": { + "state_dict": { + "keys": [], + "count": 0 + }, + "batch_progress": { + "keys": [ + "total", + "current", + "is_last_batch" + ], + "count": 3 + } + }, + "predict_loop": { + "state_dict": { + "keys": [], + "count": 0 + }, + "batch_progress": { + "keys": [ + "total", + "current" + ], + "count": 2 + } + } + }, + "callbacks": { + "ModelCheckpoint{'monitor': 'step', 'mode': 'max', 'every_n_train_steps': 2000, 'every_n_epochs': 0, 'train_time_interval': None}": { + "monitor": "step", + "best_model_score": { + "shape": [], + "dtype": "torch.int64" + }, + "best_model_path": "/autodl-fs/data/pangxinlei/models/OMG/hfd-xl-english-split20261026/checkpoints/step-00090000.ckpt", + "current_score": { + "shape": [], + "dtype": "torch.int64" + }, + "dirpath": "/autodl-fs/data/pangxinlei/models/OMG/hfd-xl-english-split20261026/checkpoints", + "best_k_models": { + "keys": [ + "/autodl-fs/data/pangxinlei/models/OMG/hfd-xl-english-split20261026/checkpoints/step-00088000.ckpt", + "/autodl-fs/data/pangxinlei/models/OMG/hfd-xl-english-split20261026/checkpoints/step-00090000.ckpt" + ], + "count": 2 + }, + "kth_best_model_path": "/autodl-fs/data/pangxinlei/models/OMG/hfd-xl-english-split20261026/checkpoints/step-00088000.ckpt", + "kth_value": { + "shape": [], + "dtype": "torch.int64" + }, + "last_model_path": "/autodl-fs/data/pangxinlei/models/OMG/hfd-xl-english-split20261026/checkpoints/last.ckpt" + }, + "ModelCheckpoint{'monitor': 'val/loss', 'mode': 'min', 'every_n_train_steps': 0, 'every_n_epochs': 1, 'train_time_interval': None}": { + "monitor": "val/loss", + "best_model_score": { + "shape": [], + "dtype": "torch.float32" + }, + "best_model_path": "/autodl-fs/data/pangxinlei/models/OMG/hfd-xl-english-split20261026/best/best.ckpt", + "current_score": { + "shape": [], + "dtype": "torch.float32" + }, + "dirpath": "/autodl-fs/data/pangxinlei/models/OMG/hfd-xl-english-split20261026/best", + "best_k_models": { + "keys": [ + "/autodl-fs/data/pangxinlei/models/OMG/hfd-xl-english-split20261026/best/best.ckpt" + ], + "count": 1 + }, + "kth_best_model_path": "/autodl-fs/data/pangxinlei/models/OMG/hfd-xl-english-split20261026/best/best.ckpt", + "kth_value": { + "shape": [], + "dtype": "torch.float32" + }, + "last_model_path": "" + } + }, + "optimizer_states": { + "count": 1, + "sample": [ + { + "state": { + "keys": [ + "0", + "1", + "2", + "3", + "4", + "5", + "6", + "7", + "8", + "9", + "10", + "11", + "12", + "13", + "14", + "15", + "16", + "17", + "18", + "19", + "20", + "21", + "22", + "23", + "24", + "25", + "26", + "27", + "28", + "29" + ], + "count": 314 + }, + "param_groups": { + "count": 1, + "sample": [ + { + "keys": [ + "lr", + "betas", + "eps", + "weight_decay", + "amsgrad", + "maximize", + "foreach", + "capturable", + "differentiable", + "fused", + "decoupled_weight_decay", + "initial_lr", + "params" + ], + "count": 13 + } + ] + } + } + ] + }, + "lr_schedulers": { + "count": 1, + "sample": [ + { + "_milestones": { + "count": 1, + "sample": [ + 2000 + ] + }, + "last_epoch": 90000, + "_last_lr": { + "count": 1, + "sample": [ + 2.502855463185791e-06 + ] + }, + "_schedulers": { + "count": 2, + "sample": [ + { + "keys": [ + "start_factor", + "end_factor", + "total_iters", + "base_lrs", + "last_epoch", + "_step_count", + "_get_lr_called_within_step", + "_last_lr" + ], + "count": 8 + }, + { + "keys": [ + "T_max", + "eta_min", + "base_lrs", + "last_epoch", + "_step_count", + "_get_lr_called_within_step", + "_last_lr" + ], + "count": 7 + } + ] + } + } + ] + }, + "hfd_resume": { + "version": 1, + "rng": { + "python": { + "count": 3, + "sample": [ + 3, + { + "count": 625, + "sample": [ + 2147483648, + 3564348608 + ] + } + ] + }, + "numpy": { + "count": 5, + "sample": [ + "MT19937", + { + "count": 624, + "sample": [ + 42, + 3107752595 + ] + } + ] + }, + "torch": { + "shape": [ + 5056 + ], + "dtype": "torch.uint8" + }, + "cuda": { + "count": 1, + "sample": [ + { + "shape": [ + 16 + ], + "dtype": "torch.uint8" + } + ] + } + }, + "train_pool": { + "step": 90000, + "slots": { + "count": 512, + "sample": [ + 12530, + 216623 + ] + }, + "batch_size": 512, + "refresh": 256, + "rng": { + "keys": [ + "bit_generator", + "state", + "has_uint32", + "uinteger" + ], + "count": 4 + }, + "identity": { + "keys": [ + "sessions", + "native_source_sha256", + "asr", + "recipe_sha256", + "foreground_language", + "ordered_session_selection_sha256", + "final_release" + ], + "count": 7 + } + }, + "val_pool": { + "step": 16, + "slots": { + "count": 512, + "sample": [ + 6295, + 7193 + ] + }, + "batch_size": 512, + "refresh": 256, + "rng": { + "keys": [ + "bit_generator", + "state", + "has_uint32", + "uinteger" + ], + "count": 4 + }, + "identity": { + "keys": [ + "sessions", + "native_source_sha256", + "asr", + "recipe_sha256", + "foreground_language", + "ordered_session_selection_sha256", + "final_release" + ], + "count": 7 + } + } + }, + "omg_model_architecture": { + "format": "omg.model_architecture", + "version": 1, + "denoiser_type": "omg.generation.denoisers.transformer.MotionTransformerDenoiser", + "frame_cond_injection": "per_layer_film", + "attention": { + "rotary_self_attention_qk_norm": true, + "cross_attention_qk_norm": true + } + } +} \ No newline at end of file diff --git a/handoff-v1/inventory/quality-archive-verification.json b/handoff-v1/inventory/quality-archive-verification.json new file mode 100644 index 0000000000000000000000000000000000000000..ec0e6b6613cb762d828f1483bc14f478671bb0a6 --- /dev/null +++ b/handoff-v1/inventory/quality-archive-verification.json @@ -0,0 +1,42 @@ +[ + { + "archive": "original-installed-workspaces.tar.gz", + "entries": 775, + "status": "all_entry_hashes_and_symlinks_verified" + }, + { + "archive": "genmo-complete-quality-results.tar.gz", + "entries": 6301, + "status": "all_entry_hashes_and_symlinks_verified" + }, + { + "archive": "hymotion-complete-quality-results.tar.gz", + "entries": 6301, + "status": "all_entry_hashes_and_symlinks_verified" + }, + { + "archive": "kimodo-complete-quality-results.tar.gz", + "entries": 6301, + "status": "all_entry_hashes_and_symlinks_verified" + }, + { + "archive": "textop-complete-quality-results.tar.gz", + "entries": 4254, + "status": "all_entry_hashes_and_symlinks_verified" + }, + { + "archive": "omg-b-complete-quality-results.tar.gz", + "entries": 4254, + "status": "all_entry_hashes_and_symlinks_verified" + }, + { + "archive": "omg-l-complete-quality-results.tar.gz", + "entries": 4248, + "status": "all_entry_hashes_and_symlinks_verified" + }, + { + "archive": "omg-xl-complete-quality-results.tar.gz", + "entries": 4252, + "status": "all_entry_hashes_and_symlinks_verified" + } +] \ No newline at end of file diff --git a/handoff-v1/inventory/recovery-capabilities.json b/handoff-v1/inventory/recovery-capabilities.json new file mode 100644 index 0000000000000000000000000000000000000000..e73cf81fa0271d3eacd9741adcfe41e5789539b1 --- /dev/null +++ b/handoff-v1/inventory/recovery-capabilities.json @@ -0,0 +1,114 @@ +{ + "omg-b": { + "actually_trained": true, + "representation": "G1 robot (not SMPL/SMPL-X)", + "latest_complete_step": 100000, + "best_complete_step": 100000, + "best_missing": false, + "optimizer": true, + "scheduler": "saved Lightning scheduler", + "scaler": "not used (BF16 OMG / FP32 TextOp)", + "ema": "not separately enabled", + "rank_count": 1, + "rng": [ + "Python", + "NumPy", + "torch CPU", + "CUDA rank0" + ], + "data_state": "consumed HFD train/val pool state at primitive-chain / logical-batch boundary", + "epoch": "Lightning loop/epoch saved", + "post_original_budget_extension_validated": false, + "short_resume_acceptance": "not performed: retained checkpoints at budget limit", + "original_shell_environment": "not fully retained; resolved config and current acceptance command/env supplied", + "clean_container_rebuild_validated": false, + "cross_hardware_bitwise_guarantee": false, + "raw_data_in_package": false, + "raw_data_access": "fixed revision plus 675 files sizes/SHA, exact indexes and preprocessing included", + "frozen_external_models": "exact revision/URL/size/SHA in resources; download needed" + }, + "omg-l": { + "actually_trained": true, + "representation": "G1 robot (not SMPL/SMPL-X)", + "latest_complete_step": 100000, + "best_complete_step": 98000, + "best_missing": false, + "optimizer": true, + "scheduler": "saved Lightning scheduler", + "scaler": "not used (BF16 OMG / FP32 TextOp)", + "ema": "not separately enabled", + "rank_count": 1, + "rng": [ + "Python", + "NumPy", + "torch CPU", + "CUDA rank0" + ], + "data_state": "consumed HFD train/val pool state at primitive-chain / logical-batch boundary", + "epoch": "Lightning loop/epoch saved", + "post_original_budget_extension_validated": false, + "short_resume_acceptance": "see acceptance-report.json", + "original_shell_environment": "not fully retained; resolved config and current acceptance command/env supplied", + "clean_container_rebuild_validated": false, + "cross_hardware_bitwise_guarantee": false, + "raw_data_in_package": false, + "raw_data_access": "fixed revision plus 675 files sizes/SHA, exact indexes and preprocessing included", + "frozen_external_models": "exact revision/URL/size/SHA in resources; download needed" + }, + "omg-xl": { + "actually_trained": true, + "representation": "G1 robot (not SMPL/SMPL-X)", + "latest_complete_step": 90000, + "best_complete_step": 88000, + "best_missing": false, + "optimizer": true, + "scheduler": "saved Lightning scheduler", + "scaler": "not used (BF16 OMG / FP32 TextOp)", + "ema": "not separately enabled", + "rank_count": 1, + "rng": [ + "Python", + "NumPy", + "torch CPU", + "CUDA rank0" + ], + "data_state": "consumed HFD train/val pool state at primitive-chain / logical-batch boundary", + "epoch": "Lightning loop/epoch saved", + "post_original_budget_extension_validated": false, + "short_resume_acceptance": "see acceptance-report.json", + "original_shell_environment": "not fully retained; resolved config and current acceptance command/env supplied", + "clean_container_rebuild_validated": false, + "cross_hardware_bitwise_guarantee": false, + "raw_data_in_package": false, + "raw_data_access": "fixed revision plus 675 files sizes/SHA, exact indexes and preprocessing included", + "frozen_external_models": "exact revision/URL/size/SHA in resources; download needed" + }, + "textop": { + "actually_trained": true, + "representation": "G1 robot (not SMPL/SMPL-X)", + "latest_complete_step": 300000, + "best_complete_step": null, + "best_missing": true, + "optimizer": true, + "scheduler": "no independent scheduler; LR derived by original manager from step/stages", + "scaler": "not used (BF16 OMG / FP32 TextOp)", + "ema": "not separately enabled", + "rank_count": 1, + "rng": [ + "Python", + "NumPy", + "torch CPU", + "CUDA rank0" + ], + "data_state": "consumed HFD train/val pool state at primitive-chain / logical-batch boundary", + "epoch": "iteration-based training; no epoch concept", + "post_original_budget_extension_validated": false, + "short_resume_acceptance": "see acceptance-report.json", + "original_shell_environment": "not fully retained; resolved config and current acceptance command/env supplied", + "clean_container_rebuild_validated": false, + "cross_hardware_bitwise_guarantee": false, + "raw_data_in_package": false, + "raw_data_access": "fixed revision plus 675 files sizes/SHA, exact indexes and preprocessing included", + "frozen_external_models": "exact revision/URL/size/SHA in resources; download needed" + } +} \ No newline at end of file diff --git a/handoff-v1/inventory/textop-EXTERNAL_ASSETS.json b/handoff-v1/inventory/textop-EXTERNAL_ASSETS.json new file mode 100644 index 0000000000000000000000000000000000000000..426dc4d881b0436b594d364f791855b52a9ab2fb --- /dev/null +++ b/handoff-v1/inventory/textop-EXTERNAL_ASSETS.json @@ -0,0 +1,1229 @@ +{ + "version": 1, + "description": "External file identities only. Model groups use models-root; hfd-metadata uses raw-root. Llama and licensed body model require authorized manual provisioning.", + "files": [ + { + "group": "textop", + "path": "TextOp-Data/TextOpRobotMDAR/dataset/PRIVATE-DATA/meanstd.pkl", + "bytes": 1816, + "sha256": "570e0f4fe961f949bf2fca0520233c9de77dbc863325684d49fe0ea48b3f4c6d", + "source": { + "url": "https://huggingface.co/datasets/Yochish/TextOp-Data/resolve/bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9/TextOpRobotMDAR/dataset/PRIVATE-DATA/meanstd.pkl", + "repo": "Yochish/TextOp-Data", + "revision": "bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9", + "repo_type": "dataset", + "filename": "TextOpRobotMDAR/dataset/PRIVATE-DATA/meanstd.pkl", + "local_prefix": "TextOp-Data" + } + }, + { + "group": "textop", + "path": "TextOp-Data/TextOpRobotMDAR/dataset/PRIVATE-DATA/statistics.yaml", + "bytes": 32, + "sha256": "3605cd9d1ba39fcc7528a3d24a48141f9e827291f6ca7bb71d41a41914c1df11", + "source": { + "url": "https://huggingface.co/datasets/Yochish/TextOp-Data/resolve/bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9/TextOpRobotMDAR/dataset/PRIVATE-DATA/statistics.yaml", + "repo": "Yochish/TextOp-Data", + "revision": "bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9", + "repo_type": "dataset", + "filename": "TextOpRobotMDAR/dataset/PRIVATE-DATA/statistics.yaml", + "local_prefix": "TextOp-Data" + } + }, + { + "group": "textop", + "path": "TextOp-Data/TextOpRobotMDAR/description/robots/g1/g1_23dof_lock_wrist.xml", + "bytes": 26891, + "sha256": "13bde8d8f33546b810b7f3350a6f2b74a43773a293c5d458a53df1807b6eeb8b", + "source": { + "url": "https://huggingface.co/datasets/Yochish/TextOp-Data/resolve/bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9/TextOpRobotMDAR/description/robots/g1/g1_23dof_lock_wrist.xml", + "repo": "Yochish/TextOp-Data", + "revision": "bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9", + "repo_type": "dataset", + "filename": "TextOpRobotMDAR/description/robots/g1/g1_23dof_lock_wrist.xml", + "local_prefix": "TextOp-Data" + } + }, + { + "group": "textop", + "path": "TextOp-Data/TextOpRobotMDAR/description/robots/g1/g1_23dof_lock_wrist_ankle4p.urdf", + "bytes": 36169, + "sha256": "adcd71a002ccfec18243445c4c1b4eae2bcf744c0628aa65a987fa59d2e2cf57", + "source": { + "url": "https://huggingface.co/datasets/Yochish/TextOp-Data/resolve/bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9/TextOpRobotMDAR/description/robots/g1/g1_23dof_lock_wrist_ankle4p.urdf", + "repo": "Yochish/TextOp-Data", + "revision": "bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9", + "repo_type": "dataset", + "filename": "TextOpRobotMDAR/description/robots/g1/g1_23dof_lock_wrist_ankle4p.urdf", + "local_prefix": "TextOp-Data" + } + }, + { + "group": "textop", + "path": "TextOp-Data/TextOpRobotMDAR/description/robots/g1/g1_23dof_lock_wrist_ankle4p.xml", + "bytes": 26891, + "sha256": "13bde8d8f33546b810b7f3350a6f2b74a43773a293c5d458a53df1807b6eeb8b", + "source": { + "url": "https://huggingface.co/datasets/Yochish/TextOp-Data/resolve/bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9/TextOpRobotMDAR/description/robots/g1/g1_23dof_lock_wrist_ankle4p.xml", + "repo": "Yochish/TextOp-Data", + "revision": "bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9", + "repo_type": "dataset", + "filename": "TextOpRobotMDAR/description/robots/g1/g1_23dof_lock_wrist_ankle4p.xml", + "local_prefix": "TextOp-Data" + } + }, + { + "group": "textop", + "path": "TextOp-Data/TextOpRobotMDAR/description/robots/g1/g1_23dof_lock_wrist_fitmotionONLY.xml", + "bytes": 20464, + "sha256": "e1a4260fe73077afd47518417b2d6609148b086438003d1956656ce133ae5101", + "source": { + "url": "https://huggingface.co/datasets/Yochish/TextOp-Data/resolve/bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9/TextOpRobotMDAR/description/robots/g1/g1_23dof_lock_wrist_fitmotionONLY.xml", + "repo": "Yochish/TextOp-Data", + "revision": "bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9", + "repo_type": "dataset", + "filename": "TextOpRobotMDAR/description/robots/g1/g1_23dof_lock_wrist_fitmotionONLY.xml", + "local_prefix": "TextOp-Data" + } + }, + { + "group": "textop", + "path": "TextOp-Data/TextOpRobotMDAR/description/robots/g1/g1_29dof.xml", + "bytes": 25850, + "sha256": "37395dd8f0efe315db8c6fa702f6ffd5e9fd4d8ddb9eaf0ad94a3855a383b183", + "source": { + "url": "https://huggingface.co/datasets/Yochish/TextOp-Data/resolve/bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9/TextOpRobotMDAR/description/robots/g1/g1_29dof.xml", + "repo": "Yochish/TextOp-Data", + "revision": "bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9", + "repo_type": "dataset", + "filename": "TextOpRobotMDAR/description/robots/g1/g1_29dof.xml", + "local_prefix": "TextOp-Data" + } + }, + { + "group": "textop", + "path": "TextOp-Data/TextOpRobotMDAR/description/robots/g1/g1_29dof_rev_1_0.xml", + "bytes": 25850, + "sha256": "b790a977e0f1f17d3478f33af6c0c70db9ee85d116745423a1e66a49c1712a3f", + "source": { + "url": "https://huggingface.co/datasets/Yochish/TextOp-Data/resolve/bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9/TextOpRobotMDAR/description/robots/g1/g1_29dof_rev_1_0.xml", + "repo": "Yochish/TextOp-Data", + "revision": "bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9", + "repo_type": "dataset", + "filename": "TextOpRobotMDAR/description/robots/g1/g1_29dof_rev_1_0.xml", + "local_prefix": "TextOp-Data" + } + }, + { + "group": "textop", + "path": "TextOp-Data/TextOpRobotMDAR/description/robots/g1/g1_29dof_rev_1_0_with_toe.xml", + "bytes": 26938, + "sha256": "aad6757582d8aff7219a91c5aa0eafd9d63937cd1e7b370d2033d20176de6118", + "source": { + "url": "https://huggingface.co/datasets/Yochish/TextOp-Data/resolve/bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9/TextOpRobotMDAR/description/robots/g1/g1_29dof_rev_1_0_with_toe.xml", + "repo": "Yochish/TextOp-Data", + "revision": "bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9", + "repo_type": "dataset", + "filename": "TextOpRobotMDAR/description/robots/g1/g1_29dof_rev_1_0_with_toe.xml", + "local_prefix": "TextOp-Data" + } + }, + { + "group": "textop", + "path": "TextOp-Data/TextOpRobotMDAR/description/robots/g1/meshes/head_link.STL", + "bytes": 932784, + "sha256": "005fb67fbd3eff94aa8bf4a6e83238174e9f91b6721f7111594322f223724411", + "source": { + "url": "https://huggingface.co/datasets/Yochish/TextOp-Data/resolve/bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9/TextOpRobotMDAR/description/robots/g1/meshes/head_link.STL", + "repo": "Yochish/TextOp-Data", + "revision": "bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9", + "repo_type": "dataset", + "filename": "TextOpRobotMDAR/description/robots/g1/meshes/head_link.STL", + "local_prefix": "TextOp-Data" + } + }, + { + "group": "textop", + "path": "TextOp-Data/TextOpRobotMDAR/description/robots/g1/meshes/left_ankle_pitch_link.STL", + "bytes": 71184, + "sha256": "d49e3abc6f5b12e532062cd575b87b5ef40cd2a3fc18f54a1ca5bba4f773d51d", + "source": { + "url": "https://huggingface.co/datasets/Yochish/TextOp-Data/resolve/bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9/TextOpRobotMDAR/description/robots/g1/meshes/left_ankle_pitch_link.STL", + "repo": "Yochish/TextOp-Data", + "revision": "bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9", + "repo_type": "dataset", + "filename": "TextOpRobotMDAR/description/robots/g1/meshes/left_ankle_pitch_link.STL", + "local_prefix": "TextOp-Data" + } + }, + { + "group": "textop", + "path": "TextOp-Data/TextOpRobotMDAR/description/robots/g1/meshes/left_ankle_roll_link.STL", + "bytes": 653384, + "sha256": "c4092af943141d4d9f74232f3cfa345afc6565f46a077793b8ae0e68b39dc33f", + "source": { + "url": "https://huggingface.co/datasets/Yochish/TextOp-Data/resolve/bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9/TextOpRobotMDAR/description/robots/g1/meshes/left_ankle_roll_link.STL", + "repo": "Yochish/TextOp-Data", + "revision": "bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9", + "repo_type": "dataset", + "filename": "TextOpRobotMDAR/description/robots/g1/meshes/left_ankle_roll_link.STL", + "local_prefix": "TextOp-Data" + } + }, + { + "group": "textop", + "path": "TextOp-Data/TextOpRobotMDAR/description/robots/g1/meshes/left_elbow_link.STL", + "bytes": 88784, + "sha256": "fa752198accd104d5c4c3a01382e45165b944fbbc5acce085059223324e5bed3", + "source": { + "url": "https://huggingface.co/datasets/Yochish/TextOp-Data/resolve/bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9/TextOpRobotMDAR/description/robots/g1/meshes/left_elbow_link.STL", + "repo": "Yochish/TextOp-Data", + "revision": "bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9", + "repo_type": "dataset", + "filename": "TextOpRobotMDAR/description/robots/g1/meshes/left_elbow_link.STL", + "local_prefix": "TextOp-Data" + } + }, + { + "group": "textop", + "path": "TextOp-Data/TextOpRobotMDAR/description/robots/g1/meshes/left_elbow_link_merge.STL", + "bytes": 3136784, + "sha256": "73c8adc6346b900eb6a2ae7159e25dfda1be60cd98affe13536e01ab33a6c1de", + "source": { + "url": "https://huggingface.co/datasets/Yochish/TextOp-Data/resolve/bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9/TextOpRobotMDAR/description/robots/g1/meshes/left_elbow_link_merge.STL", + "repo": "Yochish/TextOp-Data", + "revision": "bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9", + "repo_type": "dataset", + "filename": "TextOpRobotMDAR/description/robots/g1/meshes/left_elbow_link_merge.STL", + "local_prefix": "TextOp-Data" + } + }, + { + "group": "textop", + "path": "TextOp-Data/TextOpRobotMDAR/description/robots/g1/meshes/left_hand_index_0_link.STL", + "bytes": 475984, + "sha256": "6b35f2f77211d5a366f0b9a4e47c4ee35e536731266f1a34e9efa12db579b892", + "source": { + "url": "https://huggingface.co/datasets/Yochish/TextOp-Data/resolve/bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9/TextOpRobotMDAR/description/robots/g1/meshes/left_hand_index_0_link.STL", + "repo": "Yochish/TextOp-Data", + "revision": "bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9", + "repo_type": "dataset", + "filename": "TextOpRobotMDAR/description/robots/g1/meshes/left_hand_index_0_link.STL", + "local_prefix": "TextOp-Data" + } + }, + { + "group": "textop", + "path": "TextOp-Data/TextOpRobotMDAR/description/robots/g1/meshes/left_hand_index_1_link.STL", + "bytes": 1521784, + "sha256": "9e315ebc8a7a0cb98e033985b586b20d81cf8aa761181ae61ce56fcb14077a06", + "source": { + "url": "https://huggingface.co/datasets/Yochish/TextOp-Data/resolve/bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9/TextOpRobotMDAR/description/robots/g1/meshes/left_hand_index_1_link.STL", + "repo": "Yochish/TextOp-Data", + "revision": "bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9", + "repo_type": "dataset", + "filename": "TextOpRobotMDAR/description/robots/g1/meshes/left_hand_index_1_link.STL", + "local_prefix": "TextOp-Data" + } + }, + { + "group": "textop", + "path": "TextOp-Data/TextOpRobotMDAR/description/robots/g1/meshes/left_hand_middle_0_link.STL", + "bytes": 475984, + "sha256": "6b35f2f77211d5a366f0b9a4e47c4ee35e536731266f1a34e9efa12db579b892", + "source": { + "url": "https://huggingface.co/datasets/Yochish/TextOp-Data/resolve/bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9/TextOpRobotMDAR/description/robots/g1/meshes/left_hand_middle_0_link.STL", + "repo": "Yochish/TextOp-Data", + "revision": "bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9", + "repo_type": "dataset", + "filename": "TextOpRobotMDAR/description/robots/g1/meshes/left_hand_middle_0_link.STL", + "local_prefix": "TextOp-Data" + } + }, + { + "group": "textop", + "path": "TextOp-Data/TextOpRobotMDAR/description/robots/g1/meshes/left_hand_middle_1_link.STL", + "bytes": 1521784, + "sha256": "9e315ebc8a7a0cb98e033985b586b20d81cf8aa761181ae61ce56fcb14077a06", + "source": { + "url": "https://huggingface.co/datasets/Yochish/TextOp-Data/resolve/bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9/TextOpRobotMDAR/description/robots/g1/meshes/left_hand_middle_1_link.STL", + "repo": "Yochish/TextOp-Data", + "revision": "bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9", + "repo_type": "dataset", + "filename": "TextOpRobotMDAR/description/robots/g1/meshes/left_hand_middle_1_link.STL", + "local_prefix": "TextOp-Data" + } + }, + { + "group": "textop", + "path": "TextOp-Data/TextOpRobotMDAR/description/robots/g1/meshes/left_hand_palm_link.STL", + "bytes": 2140184, + "sha256": "23a486b75bd78a9bf03cec25d84d87f97f3dae038cf21a743b6d469b337e4004", + "source": { + "url": "https://huggingface.co/datasets/Yochish/TextOp-Data/resolve/bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9/TextOpRobotMDAR/description/robots/g1/meshes/left_hand_palm_link.STL", + "repo": "Yochish/TextOp-Data", + "revision": "bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9", + "repo_type": "dataset", + "filename": "TextOpRobotMDAR/description/robots/g1/meshes/left_hand_palm_link.STL", + "local_prefix": "TextOp-Data" + } + }, + { + "group": "textop", + "path": "TextOp-Data/TextOpRobotMDAR/description/robots/g1/meshes/left_hand_thumb_0_link.STL", + "bytes": 8884, + "sha256": "a90c721661c0622685488a3c74d1e122c7da89242d3a1daef75edb83422d05e0", + "source": { + "url": "https://huggingface.co/datasets/Yochish/TextOp-Data/resolve/bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9/TextOpRobotMDAR/description/robots/g1/meshes/left_hand_thumb_0_link.STL", + "repo": "Yochish/TextOp-Data", + "revision": "bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9", + "repo_type": "dataset", + "filename": "TextOpRobotMDAR/description/robots/g1/meshes/left_hand_thumb_0_link.STL", + "local_prefix": "TextOp-Data" + } + }, + { + "group": "textop", + "path": "TextOp-Data/TextOpRobotMDAR/description/robots/g1/meshes/left_hand_thumb_1_link.STL", + "bytes": 475984, + "sha256": "445c54a45bc13ce36001556f66bc0f49c83cb40321205ae4d676bb2874325684", + "source": { + "url": "https://huggingface.co/datasets/Yochish/TextOp-Data/resolve/bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9/TextOpRobotMDAR/description/robots/g1/meshes/left_hand_thumb_1_link.STL", + "repo": "Yochish/TextOp-Data", + "revision": "bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9", + "repo_type": "dataset", + "filename": "TextOpRobotMDAR/description/robots/g1/meshes/left_hand_thumb_1_link.STL", + "local_prefix": "TextOp-Data" + } + }, + { + "group": "textop", + "path": "TextOp-Data/TextOpRobotMDAR/description/robots/g1/meshes/left_hand_thumb_2_link.STL", + "bytes": 1521784, + "sha256": "3d8dbe5085acfc213d21aa8b0782e89cd79084e9678f3a85fc7b04a86b029db5", + "source": { + "url": "https://huggingface.co/datasets/Yochish/TextOp-Data/resolve/bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9/TextOpRobotMDAR/description/robots/g1/meshes/left_hand_thumb_2_link.STL", + "repo": "Yochish/TextOp-Data", + "revision": "bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9", + "repo_type": "dataset", + "filename": "TextOpRobotMDAR/description/robots/g1/meshes/left_hand_thumb_2_link.STL", + "local_prefix": "TextOp-Data" + } + }, + { + "group": "textop", + "path": "TextOp-Data/TextOpRobotMDAR/description/robots/g1/meshes/left_hip_pitch_link.STL", + "bytes": 181684, + "sha256": "4725168105ee768ee31638ef22b53f6be2d7641bfd7cfefe803488d884776fa4", + "source": { + "url": "https://huggingface.co/datasets/Yochish/TextOp-Data/resolve/bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9/TextOpRobotMDAR/description/robots/g1/meshes/left_hip_pitch_link.STL", + "repo": "Yochish/TextOp-Data", + "revision": "bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9", + "repo_type": "dataset", + "filename": "TextOpRobotMDAR/description/robots/g1/meshes/left_hip_pitch_link.STL", + "local_prefix": "TextOp-Data" + } + }, + { + "group": "textop", + "path": "TextOp-Data/TextOpRobotMDAR/description/robots/g1/meshes/left_hip_roll_link.STL", + "bytes": 192184, + "sha256": "91f25922ee8a7c3152790051bebad17b4d9cd243569c38fe340285ff93a97acf", + "source": { + "url": "https://huggingface.co/datasets/Yochish/TextOp-Data/resolve/bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9/TextOpRobotMDAR/description/robots/g1/meshes/left_hip_roll_link.STL", + "repo": "Yochish/TextOp-Data", + "revision": "bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9", + "repo_type": "dataset", + "filename": "TextOpRobotMDAR/description/robots/g1/meshes/left_hip_roll_link.STL", + "local_prefix": "TextOp-Data" + } + }, + { + "group": "textop", + "path": "TextOp-Data/TextOpRobotMDAR/description/robots/g1/meshes/left_hip_yaw_link.STL", + "bytes": 296284, + "sha256": "a16d88aa6ddac8083aa7ad55ed317bea44b1fa003d314fba88965b7ed0f3b55b", + "source": { + "url": "https://huggingface.co/datasets/Yochish/TextOp-Data/resolve/bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9/TextOpRobotMDAR/description/robots/g1/meshes/left_hip_yaw_link.STL", + "repo": "Yochish/TextOp-Data", + "revision": "bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9", + "repo_type": "dataset", + "filename": "TextOpRobotMDAR/description/robots/g1/meshes/left_hip_yaw_link.STL", + "local_prefix": "TextOp-Data" + } + }, + { + "group": "textop", + "path": "TextOp-Data/TextOpRobotMDAR/description/robots/g1/meshes/left_knee_link.STL", + "bytes": 854884, + "sha256": "8d92b9e3d3a636761150bb8025e32514c4602b91c7028d523ee42b3e632de477", + "source": { + "url": "https://huggingface.co/datasets/Yochish/TextOp-Data/resolve/bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9/TextOpRobotMDAR/description/robots/g1/meshes/left_knee_link.STL", + "repo": "Yochish/TextOp-Data", + "revision": "bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9", + "repo_type": "dataset", + "filename": "TextOpRobotMDAR/description/robots/g1/meshes/left_knee_link.STL", + "local_prefix": "TextOp-Data" + } + }, + { + "group": "textop", + "path": "TextOp-Data/TextOpRobotMDAR/description/robots/g1/meshes/left_rubber_hand.STL", + "bytes": 2287484, + "sha256": "cff2221a690fa69303f61fce68f2d155c1517b52efb6ca9262dd56e0bc6e70fe", + "source": { + "url": "https://huggingface.co/datasets/Yochish/TextOp-Data/resolve/bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9/TextOpRobotMDAR/description/robots/g1/meshes/left_rubber_hand.STL", + "repo": "Yochish/TextOp-Data", + "revision": "bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9", + "repo_type": "dataset", + "filename": "TextOpRobotMDAR/description/robots/g1/meshes/left_rubber_hand.STL", + "local_prefix": "TextOp-Data" + } + }, + { + "group": "textop", + "path": "TextOp-Data/TextOpRobotMDAR/description/robots/g1/meshes/left_shoulder_pitch_link.STL", + "bytes": 176784, + "sha256": "f0d1cfd02fcf0d42f95e678eeca33da3afbcc366ffba5c052847773ec4f31d52", + "source": { + "url": "https://huggingface.co/datasets/Yochish/TextOp-Data/resolve/bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9/TextOpRobotMDAR/description/robots/g1/meshes/left_shoulder_pitch_link.STL", + "repo": "Yochish/TextOp-Data", + "revision": "bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9", + "repo_type": "dataset", + "filename": "TextOpRobotMDAR/description/robots/g1/meshes/left_shoulder_pitch_link.STL", + "local_prefix": "TextOp-Data" + } + }, + { + "group": "textop", + "path": "TextOp-Data/TextOpRobotMDAR/description/robots/g1/meshes/left_shoulder_roll_link.STL", + "bytes": 400284, + "sha256": "fb9df21687773522598dc384f1a2945c7519f11cbc8bd372a49170316d6eee88", + "source": { + "url": "https://huggingface.co/datasets/Yochish/TextOp-Data/resolve/bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9/TextOpRobotMDAR/description/robots/g1/meshes/left_shoulder_roll_link.STL", + "repo": "Yochish/TextOp-Data", + "revision": "bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9", + "repo_type": "dataset", + "filename": "TextOpRobotMDAR/description/robots/g1/meshes/left_shoulder_roll_link.STL", + "local_prefix": "TextOp-Data" + } + }, + { + "group": "textop", + "path": "TextOp-Data/TextOpRobotMDAR/description/robots/g1/meshes/left_shoulder_yaw_link.STL", + "bytes": 249184, + "sha256": "1aa97e9748e924336567992181f78c7cd0652fd52a4afcca3df6b2ef6f9e712e", + "source": { + "url": "https://huggingface.co/datasets/Yochish/TextOp-Data/resolve/bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9/TextOpRobotMDAR/description/robots/g1/meshes/left_shoulder_yaw_link.STL", + "repo": "Yochish/TextOp-Data", + "revision": "bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9", + "repo_type": "dataset", + "filename": "TextOpRobotMDAR/description/robots/g1/meshes/left_shoulder_yaw_link.STL", + "local_prefix": "TextOp-Data" + } + }, + { + "group": "textop", + "path": "TextOp-Data/TextOpRobotMDAR/description/robots/g1/meshes/left_wrist_pitch_link.STL", + "bytes": 85984, + "sha256": "b251d8e05047f695d0f536cd78c11973cfa4e78d08cfe82759336cc3471de3a9", + "source": { + "url": "https://huggingface.co/datasets/Yochish/TextOp-Data/resolve/bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9/TextOpRobotMDAR/description/robots/g1/meshes/left_wrist_pitch_link.STL", + "repo": "Yochish/TextOp-Data", + "revision": "bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9", + "repo_type": "dataset", + "filename": "TextOpRobotMDAR/description/robots/g1/meshes/left_wrist_pitch_link.STL", + "local_prefix": "TextOp-Data" + } + }, + { + "group": "textop", + "path": "TextOp-Data/TextOpRobotMDAR/description/robots/g1/meshes/left_wrist_roll_link.STL", + "bytes": 356184, + "sha256": "edc387c9a0ba8c2237e9b296d32531426fabeb6f53e58df45c76106bca74148c", + "source": { + "url": "https://huggingface.co/datasets/Yochish/TextOp-Data/resolve/bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9/TextOpRobotMDAR/description/robots/g1/meshes/left_wrist_roll_link.STL", + "repo": "Yochish/TextOp-Data", + "revision": "bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9", + "repo_type": "dataset", + "filename": "TextOpRobotMDAR/description/robots/g1/meshes/left_wrist_roll_link.STL", + "local_prefix": "TextOp-Data" + } + }, + { + "group": "textop", + "path": "TextOp-Data/TextOpRobotMDAR/description/robots/g1/meshes/left_wrist_roll_rubber_hand.STL", + "bytes": 3484884, + "sha256": "e81030abd023bd9e4a308ef376d814a2c12d684d8a7670c335bbd5cd7809c909", + "source": { + "url": "https://huggingface.co/datasets/Yochish/TextOp-Data/resolve/bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9/TextOpRobotMDAR/description/robots/g1/meshes/left_wrist_roll_rubber_hand.STL", + "repo": "Yochish/TextOp-Data", + "revision": "bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9", + "repo_type": "dataset", + "filename": "TextOpRobotMDAR/description/robots/g1/meshes/left_wrist_roll_rubber_hand.STL", + "local_prefix": "TextOp-Data" + } + }, + { + "group": "textop", + "path": "TextOp-Data/TextOpRobotMDAR/description/robots/g1/meshes/left_wrist_yaw_link.STL", + "bytes": 318684, + "sha256": "83f8fb3a726bf9613d65dd14f0f447cb918c3c95b3938042a0c9c09749267d3b", + "source": { + "url": "https://huggingface.co/datasets/Yochish/TextOp-Data/resolve/bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9/TextOpRobotMDAR/description/robots/g1/meshes/left_wrist_yaw_link.STL", + "repo": "Yochish/TextOp-Data", + "revision": "bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9", + "repo_type": "dataset", + "filename": "TextOpRobotMDAR/description/robots/g1/meshes/left_wrist_yaw_link.STL", + "local_prefix": "TextOp-Data" + } + }, + { + "group": "textop", + "path": "TextOp-Data/TextOpRobotMDAR/description/robots/g1/meshes/logo_link.STL", + "bytes": 243384, + "sha256": "8571a0a19bc4916fa55f91449f51d5fdefd751000054865a842449429d5f155b", + "source": { + "url": "https://huggingface.co/datasets/Yochish/TextOp-Data/resolve/bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9/TextOpRobotMDAR/description/robots/g1/meshes/logo_link.STL", + "repo": "Yochish/TextOp-Data", + "revision": "bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9", + "repo_type": "dataset", + "filename": "TextOpRobotMDAR/description/robots/g1/meshes/logo_link.STL", + "local_prefix": "TextOp-Data" + } + }, + { + "group": "textop", + "path": "TextOp-Data/TextOpRobotMDAR/description/robots/g1/meshes/pelvis.STL", + "bytes": 1060884, + "sha256": "5ba6bbc888e630550140d3c26763f10206da8c8bd30ed886b8ede41c61f57a31", + "source": { + "url": "https://huggingface.co/datasets/Yochish/TextOp-Data/resolve/bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9/TextOpRobotMDAR/description/robots/g1/meshes/pelvis.STL", + "repo": "Yochish/TextOp-Data", + "revision": "bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9", + "repo_type": "dataset", + "filename": "TextOpRobotMDAR/description/robots/g1/meshes/pelvis.STL", + "local_prefix": "TextOp-Data" + } + }, + { + "group": "textop", + "path": "TextOp-Data/TextOpRobotMDAR/description/robots/g1/meshes/pelvis_contour_link.STL", + "bytes": 1805184, + "sha256": "5cc5c2c7a312329e3feeb2b03d3fc09fc29705bd01864f6767e51be959662420", + "source": { + "url": "https://huggingface.co/datasets/Yochish/TextOp-Data/resolve/bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9/TextOpRobotMDAR/description/robots/g1/meshes/pelvis_contour_link.STL", + "repo": "Yochish/TextOp-Data", + "revision": "bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9", + "repo_type": "dataset", + "filename": "TextOpRobotMDAR/description/robots/g1/meshes/pelvis_contour_link.STL", + "local_prefix": "TextOp-Data" + } + }, + { + "group": "textop", + "path": "TextOp-Data/TextOpRobotMDAR/description/robots/g1/meshes/right_ankle_pitch_link.STL", + "bytes": 71184, + "sha256": "15be426539ec1be70246d4d82a168806db64a41301af8b35c197a33348c787a9", + "source": { + "url": "https://huggingface.co/datasets/Yochish/TextOp-Data/resolve/bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9/TextOpRobotMDAR/description/robots/g1/meshes/right_ankle_pitch_link.STL", + "repo": "Yochish/TextOp-Data", + "revision": "bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9", + "repo_type": "dataset", + "filename": "TextOpRobotMDAR/description/robots/g1/meshes/right_ankle_pitch_link.STL", + "local_prefix": "TextOp-Data" + } + }, + { + "group": "textop", + "path": "TextOp-Data/TextOpRobotMDAR/description/robots/g1/meshes/right_ankle_roll_link.STL", + "bytes": 653784, + "sha256": "4b66222ea56653e627711b56d0a8949b4920da5df091da0ceb343f54e884e3a5", + "source": { + "url": "https://huggingface.co/datasets/Yochish/TextOp-Data/resolve/bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9/TextOpRobotMDAR/description/robots/g1/meshes/right_ankle_roll_link.STL", + "repo": "Yochish/TextOp-Data", + "revision": "bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9", + "repo_type": "dataset", + "filename": "TextOpRobotMDAR/description/robots/g1/meshes/right_ankle_roll_link.STL", + "local_prefix": "TextOp-Data" + } + }, + { + "group": "textop", + "path": "TextOp-Data/TextOpRobotMDAR/description/robots/g1/meshes/right_elbow_link.STL", + "bytes": 88784, + "sha256": "1be925d7aa268bb8fddf5362b9173066890c7d32092c05638608126e59d1e2ab", + "source": { + "url": "https://huggingface.co/datasets/Yochish/TextOp-Data/resolve/bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9/TextOpRobotMDAR/description/robots/g1/meshes/right_elbow_link.STL", + "repo": "Yochish/TextOp-Data", + "revision": "bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9", + "repo_type": "dataset", + "filename": "TextOpRobotMDAR/description/robots/g1/meshes/right_elbow_link.STL", + "local_prefix": "TextOp-Data" + } + }, + { + "group": "textop", + "path": "TextOp-Data/TextOpRobotMDAR/description/robots/g1/meshes/right_elbow_link_merge.STL", + "bytes": 3058284, + "sha256": "e6def7100e5bbb276c7498c5b7a8c8812f560ca35d6bd09c376ebab93595be7f", + "source": { + "url": "https://huggingface.co/datasets/Yochish/TextOp-Data/resolve/bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9/TextOpRobotMDAR/description/robots/g1/meshes/right_elbow_link_merge.STL", + "repo": "Yochish/TextOp-Data", + "revision": "bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9", + "repo_type": "dataset", + "filename": "TextOpRobotMDAR/description/robots/g1/meshes/right_elbow_link_merge.STL", + "local_prefix": "TextOp-Data" + } + }, + { + "group": "textop", + "path": "TextOp-Data/TextOpRobotMDAR/description/robots/g1/meshes/right_hand_index_0_link.STL", + "bytes": 475984, + "sha256": "12ab4300c95e437e834f9aef772b3b431c671bf34338930b52ad11aef73bbd0d", + "source": { + "url": "https://huggingface.co/datasets/Yochish/TextOp-Data/resolve/bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9/TextOpRobotMDAR/description/robots/g1/meshes/right_hand_index_0_link.STL", + "repo": "Yochish/TextOp-Data", + "revision": "bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9", + "repo_type": "dataset", + "filename": "TextOpRobotMDAR/description/robots/g1/meshes/right_hand_index_0_link.STL", + "local_prefix": "TextOp-Data" + } + }, + { + "group": "textop", + "path": "TextOp-Data/TextOpRobotMDAR/description/robots/g1/meshes/right_hand_index_1_link.STL", + "bytes": 1521784, + "sha256": "c9c34efce4563cacdcfd29fc838a982976d40f5442c71219811dcbbf3923a33d", + "source": { + "url": "https://huggingface.co/datasets/Yochish/TextOp-Data/resolve/bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9/TextOpRobotMDAR/description/robots/g1/meshes/right_hand_index_1_link.STL", + "repo": "Yochish/TextOp-Data", + "revision": "bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9", + "repo_type": "dataset", + "filename": "TextOpRobotMDAR/description/robots/g1/meshes/right_hand_index_1_link.STL", + "local_prefix": "TextOp-Data" + } + }, + { + "group": "textop", + "path": "TextOp-Data/TextOpRobotMDAR/description/robots/g1/meshes/right_hand_middle_0_link.STL", + "bytes": 475984, + "sha256": "12ab4300c95e437e834f9aef772b3b431c671bf34338930b52ad11aef73bbd0d", + "source": { + "url": "https://huggingface.co/datasets/Yochish/TextOp-Data/resolve/bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9/TextOpRobotMDAR/description/robots/g1/meshes/right_hand_middle_0_link.STL", + "repo": "Yochish/TextOp-Data", + "revision": "bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9", + "repo_type": "dataset", + "filename": "TextOpRobotMDAR/description/robots/g1/meshes/right_hand_middle_0_link.STL", + "local_prefix": "TextOp-Data" + } + }, + { + "group": "textop", + "path": "TextOp-Data/TextOpRobotMDAR/description/robots/g1/meshes/right_hand_middle_1_link.STL", + "bytes": 1521784, + "sha256": "c9c34efce4563cacdcfd29fc838a982976d40f5442c71219811dcbbf3923a33d", + "source": { + "url": "https://huggingface.co/datasets/Yochish/TextOp-Data/resolve/bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9/TextOpRobotMDAR/description/robots/g1/meshes/right_hand_middle_1_link.STL", + "repo": "Yochish/TextOp-Data", + "revision": "bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9", + "repo_type": "dataset", + "filename": "TextOpRobotMDAR/description/robots/g1/meshes/right_hand_middle_1_link.STL", + "local_prefix": "TextOp-Data" + } + }, + { + "group": "textop", + "path": "TextOp-Data/TextOpRobotMDAR/description/robots/g1/meshes/right_hand_palm_link.STL", + "bytes": 2140184, + "sha256": "86c0b231cc44477d64a6493e5a427ba16617a00738112dd187c652675b086fb9", + "source": { + "url": "https://huggingface.co/datasets/Yochish/TextOp-Data/resolve/bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9/TextOpRobotMDAR/description/robots/g1/meshes/right_hand_palm_link.STL", + "repo": "Yochish/TextOp-Data", + "revision": "bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9", + "repo_type": "dataset", + "filename": "TextOpRobotMDAR/description/robots/g1/meshes/right_hand_palm_link.STL", + "local_prefix": "TextOp-Data" + } + }, + { + "group": "textop", + "path": "TextOp-Data/TextOpRobotMDAR/description/robots/g1/meshes/right_hand_thumb_0_link.STL", + "bytes": 8884, + "sha256": "544298d0ea1088f5b276a10cc6a6a9e533efdd91594955fdc956c46211d07f83", + "source": { + "url": "https://huggingface.co/datasets/Yochish/TextOp-Data/resolve/bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9/TextOpRobotMDAR/description/robots/g1/meshes/right_hand_thumb_0_link.STL", + "repo": "Yochish/TextOp-Data", + "revision": "bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9", + "repo_type": "dataset", + "filename": "TextOpRobotMDAR/description/robots/g1/meshes/right_hand_thumb_0_link.STL", + "local_prefix": "TextOp-Data" + } + }, + { + "group": "textop", + "path": "TextOp-Data/TextOpRobotMDAR/description/robots/g1/meshes/right_hand_thumb_1_link.STL", + "bytes": 475984, + "sha256": "0a9a820da8dd10f298778b714f1364216e8a5976f4fd3a05689ea26327d44bf6", + "source": { + "url": "https://huggingface.co/datasets/Yochish/TextOp-Data/resolve/bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9/TextOpRobotMDAR/description/robots/g1/meshes/right_hand_thumb_1_link.STL", + "repo": "Yochish/TextOp-Data", + "revision": "bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9", + "repo_type": "dataset", + "filename": "TextOpRobotMDAR/description/robots/g1/meshes/right_hand_thumb_1_link.STL", + "local_prefix": "TextOp-Data" + } + }, + { + "group": "textop", + "path": "TextOp-Data/TextOpRobotMDAR/description/robots/g1/meshes/right_hand_thumb_2_link.STL", + "bytes": 1521784, + "sha256": "3f1bfb37668e8f61801c8d25f171fa1949e08666be86c67acad7e0079937cc45", + "source": { + "url": "https://huggingface.co/datasets/Yochish/TextOp-Data/resolve/bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9/TextOpRobotMDAR/description/robots/g1/meshes/right_hand_thumb_2_link.STL", + "repo": "Yochish/TextOp-Data", + "revision": "bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9", + "repo_type": "dataset", + "filename": "TextOpRobotMDAR/description/robots/g1/meshes/right_hand_thumb_2_link.STL", + "local_prefix": "TextOp-Data" + } + }, + { + "group": "textop", + "path": "TextOp-Data/TextOpRobotMDAR/description/robots/g1/meshes/right_hip_pitch_link.STL", + "bytes": 181284, + "sha256": "e4f3c99d7f4a7d34eadbef9461fc66e3486cb5442db1ec50c86317d459f1a9c6", + "source": { + "url": "https://huggingface.co/datasets/Yochish/TextOp-Data/resolve/bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9/TextOpRobotMDAR/description/robots/g1/meshes/right_hip_pitch_link.STL", + "repo": "Yochish/TextOp-Data", + "revision": "bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9", + "repo_type": "dataset", + "filename": "TextOpRobotMDAR/description/robots/g1/meshes/right_hip_pitch_link.STL", + "local_prefix": "TextOp-Data" + } + }, + { + "group": "textop", + "path": "TextOp-Data/TextOpRobotMDAR/description/robots/g1/meshes/right_hip_roll_link.STL", + "bytes": 192684, + "sha256": "4c254ef66a356f492947f360dd931965477b631e3fcc841f91ccc46d945d54f6", + "source": { + "url": "https://huggingface.co/datasets/Yochish/TextOp-Data/resolve/bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9/TextOpRobotMDAR/description/robots/g1/meshes/right_hip_roll_link.STL", + "repo": "Yochish/TextOp-Data", + "revision": "bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9", + "repo_type": "dataset", + "filename": "TextOpRobotMDAR/description/robots/g1/meshes/right_hip_roll_link.STL", + "local_prefix": "TextOp-Data" + } + }, + { + "group": "textop", + "path": "TextOp-Data/TextOpRobotMDAR/description/robots/g1/meshes/right_hip_yaw_link.STL", + "bytes": 296284, + "sha256": "e479c2936ca2057e9eb2f7dff6c189b7419d7b8484dea0b298cbb36a2a6aa668", + "source": { + "url": "https://huggingface.co/datasets/Yochish/TextOp-Data/resolve/bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9/TextOpRobotMDAR/description/robots/g1/meshes/right_hip_yaw_link.STL", + "repo": "Yochish/TextOp-Data", + "revision": "bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9", + "repo_type": "dataset", + "filename": "TextOpRobotMDAR/description/robots/g1/meshes/right_hip_yaw_link.STL", + "local_prefix": "TextOp-Data" + } + }, + { + "group": "textop", + "path": "TextOp-Data/TextOpRobotMDAR/description/robots/g1/meshes/right_knee_link.STL", + "bytes": 852284, + "sha256": "63c4008449c9bbe701a6e2b557b7a252e90cf3a5abcf54cee46862b9a69f8ec8", + "source": { + "url": "https://huggingface.co/datasets/Yochish/TextOp-Data/resolve/bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9/TextOpRobotMDAR/description/robots/g1/meshes/right_knee_link.STL", + "repo": "Yochish/TextOp-Data", + "revision": "bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9", + "repo_type": "dataset", + "filename": "TextOpRobotMDAR/description/robots/g1/meshes/right_knee_link.STL", + "local_prefix": "TextOp-Data" + } + }, + { + "group": "textop", + "path": "TextOp-Data/TextOpRobotMDAR/description/robots/g1/meshes/right_rubber_hand.STL", + "bytes": 2192684, + "sha256": "99533b778bca6246144fa511bb9d4e555e075c641f2a0251e04372869cd99d67", + "source": { + "url": "https://huggingface.co/datasets/Yochish/TextOp-Data/resolve/bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9/TextOpRobotMDAR/description/robots/g1/meshes/right_rubber_hand.STL", + "repo": "Yochish/TextOp-Data", + "revision": "bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9", + "repo_type": "dataset", + "filename": "TextOpRobotMDAR/description/robots/g1/meshes/right_rubber_hand.STL", + "local_prefix": "TextOp-Data" + } + }, + { + "group": "textop", + "path": "TextOp-Data/TextOpRobotMDAR/description/robots/g1/meshes/right_shoulder_pitch_link.STL", + "bytes": 176784, + "sha256": "24cdb387e0128dfe602770a81c56cdce3a0181d34d039a11d1aaf8819b7b8c02", + "source": { + "url": "https://huggingface.co/datasets/Yochish/TextOp-Data/resolve/bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9/TextOpRobotMDAR/description/robots/g1/meshes/right_shoulder_pitch_link.STL", + "repo": "Yochish/TextOp-Data", + "revision": "bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9", + "repo_type": "dataset", + "filename": "TextOpRobotMDAR/description/robots/g1/meshes/right_shoulder_pitch_link.STL", + "local_prefix": "TextOp-Data" + } + }, + { + "group": "textop", + "path": "TextOp-Data/TextOpRobotMDAR/description/robots/g1/meshes/right_shoulder_roll_link.STL", + "bytes": 401884, + "sha256": "962b97c48f9ce9e8399f45dd9522e0865d19aa9fd299406b2d475a8fc4a53e81", + "source": { + "url": "https://huggingface.co/datasets/Yochish/TextOp-Data/resolve/bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9/TextOpRobotMDAR/description/robots/g1/meshes/right_shoulder_roll_link.STL", + "repo": "Yochish/TextOp-Data", + "revision": "bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9", + "repo_type": "dataset", + "filename": "TextOpRobotMDAR/description/robots/g1/meshes/right_shoulder_roll_link.STL", + "local_prefix": "TextOp-Data" + } + }, + { + "group": "textop", + "path": "TextOp-Data/TextOpRobotMDAR/description/robots/g1/meshes/right_shoulder_yaw_link.STL", + "bytes": 249984, + "sha256": "a0b76489271da0c72461a344c9ffb0f0c6e64f019ea5014c1624886c442a2fe5", + "source": { + "url": "https://huggingface.co/datasets/Yochish/TextOp-Data/resolve/bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9/TextOpRobotMDAR/description/robots/g1/meshes/right_shoulder_yaw_link.STL", + "repo": "Yochish/TextOp-Data", + "revision": "bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9", + "repo_type": "dataset", + "filename": "TextOpRobotMDAR/description/robots/g1/meshes/right_shoulder_yaw_link.STL", + "local_prefix": "TextOp-Data" + } + }, + { + "group": "textop", + "path": "TextOp-Data/TextOpRobotMDAR/description/robots/g1/meshes/right_wrist_pitch_link.STL", + "bytes": 79584, + "sha256": "d22f8f3b3127f15a63e5be1ee273cd5075786c3142f1c3d9f76cbf43d2a26477", + "source": { + "url": "https://huggingface.co/datasets/Yochish/TextOp-Data/resolve/bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9/TextOpRobotMDAR/description/robots/g1/meshes/right_wrist_pitch_link.STL", + "repo": "Yochish/TextOp-Data", + "revision": "bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9", + "repo_type": "dataset", + "filename": "TextOpRobotMDAR/description/robots/g1/meshes/right_wrist_pitch_link.STL", + "local_prefix": "TextOp-Data" + } + }, + { + "group": "textop", + "path": "TextOp-Data/TextOpRobotMDAR/description/robots/g1/meshes/right_wrist_roll_link.STL", + "bytes": 356084, + "sha256": "a7ee9212ff5b94d6cb7f52bb1bbf3f352194d5b598acff74f4c77d340c5b344f", + "source": { + "url": "https://huggingface.co/datasets/Yochish/TextOp-Data/resolve/bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9/TextOpRobotMDAR/description/robots/g1/meshes/right_wrist_roll_link.STL", + "repo": "Yochish/TextOp-Data", + "revision": "bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9", + "repo_type": "dataset", + "filename": "TextOpRobotMDAR/description/robots/g1/meshes/right_wrist_roll_link.STL", + "local_prefix": "TextOp-Data" + } + }, + { + "group": "textop", + "path": "TextOp-Data/TextOpRobotMDAR/description/robots/g1/meshes/right_wrist_roll_rubber_hand.STL", + "bytes": 3481584, + "sha256": "0729aff1ac4356f9314de13a46906267642e58bc47f0d8a7f17f6590a6242ccf", + "source": { + "url": "https://huggingface.co/datasets/Yochish/TextOp-Data/resolve/bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9/TextOpRobotMDAR/description/robots/g1/meshes/right_wrist_roll_rubber_hand.STL", + "repo": "Yochish/TextOp-Data", + "revision": "bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9", + "repo_type": "dataset", + "filename": "TextOpRobotMDAR/description/robots/g1/meshes/right_wrist_roll_rubber_hand.STL", + "local_prefix": "TextOp-Data" + } + }, + { + "group": "textop", + "path": "TextOp-Data/TextOpRobotMDAR/description/robots/g1/meshes/right_wrist_yaw_link.STL", + "bytes": 341484, + "sha256": "bc9dece2d12509707e0057ba2e48df8f3d56db0c79410212963a25e8a50f61a6", + "source": { + "url": "https://huggingface.co/datasets/Yochish/TextOp-Data/resolve/bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9/TextOpRobotMDAR/description/robots/g1/meshes/right_wrist_yaw_link.STL", + "repo": "Yochish/TextOp-Data", + "revision": "bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9", + "repo_type": "dataset", + "filename": "TextOpRobotMDAR/description/robots/g1/meshes/right_wrist_yaw_link.STL", + "local_prefix": "TextOp-Data" + } + }, + { + "group": "textop", + "path": "TextOp-Data/TextOpRobotMDAR/description/robots/g1/meshes/torso_constraint_L_link.STL", + "bytes": 203584, + "sha256": "82be7f93e85b3d303a1d1e1847e2c916939bd61c424ed1ebd28691ec33909dd1", + "source": { + "url": "https://huggingface.co/datasets/Yochish/TextOp-Data/resolve/bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9/TextOpRobotMDAR/description/robots/g1/meshes/torso_constraint_L_link.STL", + "repo": "Yochish/TextOp-Data", + "revision": "bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9", + "repo_type": "dataset", + "filename": "TextOpRobotMDAR/description/robots/g1/meshes/torso_constraint_L_link.STL", + "local_prefix": "TextOp-Data" + } + }, + { + "group": "textop", + "path": "TextOp-Data/TextOpRobotMDAR/description/robots/g1/meshes/torso_constraint_L_rod_link.STL", + "bytes": 74884, + "sha256": "c10de1effa7ea797ac268006aa2a739036c7e1f326b2012d711ee2c20c5a6e96", + "source": { + "url": "https://huggingface.co/datasets/Yochish/TextOp-Data/resolve/bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9/TextOpRobotMDAR/description/robots/g1/meshes/torso_constraint_L_rod_link.STL", + "repo": "Yochish/TextOp-Data", + "revision": "bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9", + "repo_type": "dataset", + "filename": "TextOpRobotMDAR/description/robots/g1/meshes/torso_constraint_L_rod_link.STL", + "local_prefix": "TextOp-Data" + } + }, + { + "group": "textop", + "path": "TextOp-Data/TextOpRobotMDAR/description/robots/g1/meshes/torso_constraint_R_link.STL", + "bytes": 203584, + "sha256": "54ded433a3a0c76027365856fdbd55215643de88846f7d436598a4071e682725", + "source": { + "url": "https://huggingface.co/datasets/Yochish/TextOp-Data/resolve/bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9/TextOpRobotMDAR/description/robots/g1/meshes/torso_constraint_R_link.STL", + "repo": "Yochish/TextOp-Data", + "revision": "bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9", + "repo_type": "dataset", + "filename": "TextOpRobotMDAR/description/robots/g1/meshes/torso_constraint_R_link.STL", + "local_prefix": "TextOp-Data" + } + }, + { + "group": "textop", + "path": "TextOp-Data/TextOpRobotMDAR/description/robots/g1/meshes/torso_constraint_R_rod_link.STL", + "bytes": 74884, + "sha256": "cb83fd38a9f06c99e3f301c70f12d94ae770a8f0cf9501f83580a27f908990b4", + "source": { + "url": "https://huggingface.co/datasets/Yochish/TextOp-Data/resolve/bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9/TextOpRobotMDAR/description/robots/g1/meshes/torso_constraint_R_rod_link.STL", + "repo": "Yochish/TextOp-Data", + "revision": "bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9", + "repo_type": "dataset", + "filename": "TextOpRobotMDAR/description/robots/g1/meshes/torso_constraint_R_rod_link.STL", + "local_prefix": "TextOp-Data" + } + }, + { + "group": "textop", + "path": "TextOp-Data/TextOpRobotMDAR/description/robots/g1/meshes/torso_link.STL", + "bytes": 2232984, + "sha256": "e96d023f0368a4e3450b86ca5d4f10227d8141156a373e7da8cb3c93266523e0", + "source": { + "url": "https://huggingface.co/datasets/Yochish/TextOp-Data/resolve/bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9/TextOpRobotMDAR/description/robots/g1/meshes/torso_link.STL", + "repo": "Yochish/TextOp-Data", + "revision": "bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9", + "repo_type": "dataset", + "filename": "TextOpRobotMDAR/description/robots/g1/meshes/torso_link.STL", + "local_prefix": "TextOp-Data" + } + }, + { + "group": "textop", + "path": "TextOp-Data/TextOpRobotMDAR/description/robots/g1/meshes/torso_link_23dof_rev_1_0.STL", + "bytes": 7825434, + "sha256": "3cd0d56fde14b73c1623304684805029971c4f84b596f9914e823ca70a107fd2", + "source": { + "url": "https://huggingface.co/datasets/Yochish/TextOp-Data/resolve/bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9/TextOpRobotMDAR/description/robots/g1/meshes/torso_link_23dof_rev_1_0.STL", + "repo": "Yochish/TextOp-Data", + "revision": "bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9", + "repo_type": "dataset", + "filename": "TextOpRobotMDAR/description/robots/g1/meshes/torso_link_23dof_rev_1_0.STL", + "local_prefix": "TextOp-Data" + } + }, + { + "group": "textop", + "path": "TextOp-Data/TextOpRobotMDAR/description/robots/g1/meshes/torso_link_rev_1_0.STL", + "bytes": 2570584, + "sha256": "11ddb46f2098efbbd8816b1d65893632d8e78be936376c7cdcd6771899ccc723", + "source": { + "url": "https://huggingface.co/datasets/Yochish/TextOp-Data/resolve/bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9/TextOpRobotMDAR/description/robots/g1/meshes/torso_link_rev_1_0.STL", + "repo": "Yochish/TextOp-Data", + "revision": "bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9", + "repo_type": "dataset", + "filename": "TextOpRobotMDAR/description/robots/g1/meshes/torso_link_rev_1_0.STL", + "local_prefix": "TextOp-Data" + } + }, + { + "group": "textop", + "path": "TextOp-Data/TextOpRobotMDAR/description/robots/g1/meshes/waist_constraint_L.STL", + "bytes": 114684, + "sha256": "8ebafdcb4de6871113f0ca2c356618d6e46b1d50f6c0bf9e37f47b9d8e100d99", + "source": { + "url": "https://huggingface.co/datasets/Yochish/TextOp-Data/resolve/bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9/TextOpRobotMDAR/description/robots/g1/meshes/waist_constraint_L.STL", + "repo": "Yochish/TextOp-Data", + "revision": "bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9", + "repo_type": "dataset", + "filename": "TextOpRobotMDAR/description/robots/g1/meshes/waist_constraint_L.STL", + "local_prefix": "TextOp-Data" + } + }, + { + "group": "textop", + "path": "TextOp-Data/TextOpRobotMDAR/description/robots/g1/meshes/waist_constraint_R.STL", + "bytes": 114684, + "sha256": "791902a291ffbd35ab97383b7b44ea5d975de7c80eef838797c970790b382ca9", + "source": { + "url": "https://huggingface.co/datasets/Yochish/TextOp-Data/resolve/bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9/TextOpRobotMDAR/description/robots/g1/meshes/waist_constraint_R.STL", + "repo": "Yochish/TextOp-Data", + "revision": "bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9", + "repo_type": "dataset", + "filename": "TextOpRobotMDAR/description/robots/g1/meshes/waist_constraint_R.STL", + "local_prefix": "TextOp-Data" + } + }, + { + "group": "textop", + "path": "TextOp-Data/TextOpRobotMDAR/description/robots/g1/meshes/waist_roll_link.STL", + "bytes": 24184, + "sha256": "34f0aa73f41131230be4d25876c944fdf6c24d62553f199ff8b980c15e8913df", + "source": { + "url": "https://huggingface.co/datasets/Yochish/TextOp-Data/resolve/bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9/TextOpRobotMDAR/description/robots/g1/meshes/waist_roll_link.STL", + "repo": "Yochish/TextOp-Data", + "revision": "bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9", + "repo_type": "dataset", + "filename": "TextOpRobotMDAR/description/robots/g1/meshes/waist_roll_link.STL", + "local_prefix": "TextOp-Data" + } + }, + { + "group": "textop", + "path": "TextOp-Data/TextOpRobotMDAR/description/robots/g1/meshes/waist_roll_link_rev_1_0.STL", + "bytes": 85884, + "sha256": "b67c347a05abc3e8ddae600d98a082bee273bb39f8e651647708b0a7140a8a97", + "source": { + "url": "https://huggingface.co/datasets/Yochish/TextOp-Data/resolve/bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9/TextOpRobotMDAR/description/robots/g1/meshes/waist_roll_link_rev_1_0.STL", + "repo": "Yochish/TextOp-Data", + "revision": "bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9", + "repo_type": "dataset", + "filename": "TextOpRobotMDAR/description/robots/g1/meshes/waist_roll_link_rev_1_0.STL", + "local_prefix": "TextOp-Data" + } + }, + { + "group": "textop", + "path": "TextOp-Data/TextOpRobotMDAR/description/robots/g1/meshes/waist_support_link.STL", + "bytes": 150484, + "sha256": "1fae9e1bb609848a1667d32eed8d6083ae443538a306843056a2a660f1b2926a", + "source": { + "url": "https://huggingface.co/datasets/Yochish/TextOp-Data/resolve/bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9/TextOpRobotMDAR/description/robots/g1/meshes/waist_support_link.STL", + "repo": "Yochish/TextOp-Data", + "revision": "bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9", + "repo_type": "dataset", + "filename": "TextOpRobotMDAR/description/robots/g1/meshes/waist_support_link.STL", + "local_prefix": "TextOp-Data" + } + }, + { + "group": "textop", + "path": "TextOp-Data/TextOpRobotMDAR/description/robots/g1/meshes/waist_yaw_link.STL", + "bytes": 336284, + "sha256": "2883f20e03f09b669b5b4ce10677ee6b5191c0934b584d7cbaef2d0662856ffb", + "source": { + "url": "https://huggingface.co/datasets/Yochish/TextOp-Data/resolve/bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9/TextOpRobotMDAR/description/robots/g1/meshes/waist_yaw_link.STL", + "repo": "Yochish/TextOp-Data", + "revision": "bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9", + "repo_type": "dataset", + "filename": "TextOpRobotMDAR/description/robots/g1/meshes/waist_yaw_link.STL", + "local_prefix": "TextOp-Data" + } + }, + { + "group": "textop", + "path": "TextOp-Data/TextOpRobotMDAR/description/robots/g1/meshes/waist_yaw_link_rev_1_0.STL", + "bytes": 619984, + "sha256": "ec6db442b11f25eed898b5add07940c85d804f300de24dcbd264ccd8be7d554c", + "source": { + "url": "https://huggingface.co/datasets/Yochish/TextOp-Data/resolve/bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9/TextOpRobotMDAR/description/robots/g1/meshes/waist_yaw_link_rev_1_0.STL", + "repo": "Yochish/TextOp-Data", + "revision": "bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9", + "repo_type": "dataset", + "filename": "TextOpRobotMDAR/description/robots/g1/meshes/waist_yaw_link_rev_1_0.STL", + "local_prefix": "TextOp-Data" + } + }, + { + "group": "textop", + "path": "TextOp-Data/TextOpRobotMDAR/description/robots/g1/phc_g1_23dof.yaml", + "bytes": 4470, + "sha256": "839cfe4b7ef92a2ff7daea4cf51a99407507e128dcfab0c49c0cf6f4b596cd87", + "source": { + "url": "https://huggingface.co/datasets/Yochish/TextOp-Data/resolve/bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9/TextOpRobotMDAR/description/robots/g1/phc_g1_23dof.yaml", + "repo": "Yochish/TextOp-Data", + "revision": "bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9", + "repo_type": "dataset", + "filename": "TextOpRobotMDAR/description/robots/g1/phc_g1_23dof.yaml", + "local_prefix": "TextOp-Data" + } + }, + { + "group": "textop", + "path": "TextOp-Data/TextOpRobotMDAR/description/robots/g1/smpl_humanoid.xml", + "bytes": 21377, + "sha256": "32aae7088bca8576bf3d4385ab764be265f649eaf26e7cd7dd49d0de0941483f", + "source": { + "url": "https://huggingface.co/datasets/Yochish/TextOp-Data/resolve/bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9/TextOpRobotMDAR/description/robots/g1/smpl_humanoid.xml", + "repo": "Yochish/TextOp-Data", + "revision": "bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9", + "repo_type": "dataset", + "filename": "TextOpRobotMDAR/description/robots/g1/smpl_humanoid.xml", + "local_prefix": "TextOp-Data" + } + }, + { + "group": "textop", + "path": "TextOp-Data/TextOpRobotMDAR/logs/pretrained/checkpoint/.hydra/config.yaml", + "bytes": 10976, + "sha256": "de4bd07f5516ec54ab87f992ee631445fb7a8715a597b1d093187968d082d700", + "source": { + "url": "https://huggingface.co/datasets/Yochish/TextOp-Data/resolve/bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9/TextOpRobotMDAR/logs/pretrained/checkpoint/.hydra/config.yaml", + "repo": "Yochish/TextOp-Data", + "revision": "bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9", + "repo_type": "dataset", + "filename": "TextOpRobotMDAR/logs/pretrained/checkpoint/.hydra/config.yaml", + "local_prefix": "TextOp-Data" + } + }, + { + "group": "textop", + "path": "TextOp-Data/TextOpRobotMDAR/logs/pretrained/checkpoint/.hydra/hydra.yaml", + "bytes": 4469, + "sha256": "dc974fcc068bfea4d84df1f3dcca37a39a9f3c777a9d85ff8f6019812a6777d8", + "source": { + "url": "https://huggingface.co/datasets/Yochish/TextOp-Data/resolve/bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9/TextOpRobotMDAR/logs/pretrained/checkpoint/.hydra/hydra.yaml", + "repo": "Yochish/TextOp-Data", + "revision": "bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9", + "repo_type": "dataset", + "filename": "TextOpRobotMDAR/logs/pretrained/checkpoint/.hydra/hydra.yaml", + "local_prefix": "TextOp-Data" + } + }, + { + "group": "textop", + "path": "TextOp-Data/TextOpRobotMDAR/logs/pretrained/checkpoint/.hydra/overrides.yaml", + "bytes": 611, + "sha256": "7d923ad4a73600a57e60bd41237f7f94e5d003a26b1e5bf674ca910a062b11d3", + "source": { + "url": "https://huggingface.co/datasets/Yochish/TextOp-Data/resolve/bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9/TextOpRobotMDAR/logs/pretrained/checkpoint/.hydra/overrides.yaml", + "repo": "Yochish/TextOp-Data", + "revision": "bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9", + "repo_type": "dataset", + "filename": "TextOpRobotMDAR/logs/pretrained/checkpoint/.hydra/overrides.yaml", + "local_prefix": "TextOp-Data" + } + }, + { + "group": "textop", + "path": "TextOp-Data/TextOpRobotMDAR/logs/pretrained/checkpoint/cfg.yaml", + "bytes": 39765, + "sha256": "a8ad43077c0b8512985beb6993b54ea5b4392f7d7d0e79ca00cecbf1b479bab7", + "source": { + "url": "https://huggingface.co/datasets/Yochish/TextOp-Data/resolve/bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9/TextOpRobotMDAR/logs/pretrained/checkpoint/cfg.yaml", + "repo": "Yochish/TextOp-Data", + "revision": "bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9", + "repo_type": "dataset", + "filename": "TextOpRobotMDAR/logs/pretrained/checkpoint/cfg.yaml", + "local_prefix": "TextOp-Data" + } + }, + { + "group": "textop", + "path": "TextOp-Data/TextOpRobotMDAR/logs/pretrained/checkpoint/ckpt_200000.pth", + "bytes": 223633892, + "sha256": "9c514d8e61d6f24b9e5b6b278b9987c9d768c7c86883e834f51e0f0c2a18f00f", + "source": { + "url": "https://huggingface.co/datasets/Yochish/TextOp-Data/resolve/bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9/TextOpRobotMDAR/logs/pretrained/checkpoint/ckpt_200000.pth", + "repo": "Yochish/TextOp-Data", + "revision": "bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9", + "repo_type": "dataset", + "filename": "TextOpRobotMDAR/logs/pretrained/checkpoint/ckpt_200000.pth", + "local_prefix": "TextOp-Data" + } + }, + { + "group": "textop", + "path": "TextOp-Data/TextOpRobotMDAR/logs/pretrained/checkpoint/vae.pth", + "bytes": 513354942, + "sha256": "6522ab17b8904d1ea6a1dcdef241090000706b485312a319007dc4bb3eac8955", + "source": { + "url": "https://huggingface.co/datasets/Yochish/TextOp-Data/resolve/bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9/TextOpRobotMDAR/logs/pretrained/checkpoint/vae.pth", + "repo": "Yochish/TextOp-Data", + "revision": "bb9604ce36ad4c0a77b5cccde89ebbb7c3ca93c9", + "repo_type": "dataset", + "filename": "TextOpRobotMDAR/logs/pretrained/checkpoint/vae.pth", + "local_prefix": "TextOp-Data" + } + }, + { + "group": "textop", + "path": "TextOp-Data/clip/ViT-B-32.pt", + "bytes": 353976522, + "sha256": "40d365715913c9da98579312b702a82c18be219cc2a73407c4526f58eba950af", + "source": { + "url": "https://openaipublic.azureedge.net/clip/models/40d365715913c9da98579312b702a82c18be219cc2a73407c4526f58eba950af/ViT-B-32.pt" + } + }, + { + "group": "hfd-metadata", + "path": "MANIFEST.json", + "bytes": 283191, + "sha256": "a7d91a52dabd142d73a8e93ed86403a5763385b3115ff2217777bf82d2bb492e", + "source": { + "repo": "Tinghong-Ye22/hfd_data", + "revision": "93ba4a7a9c66b3e1b7041e5f4ea038ef2418b948", + "repo_type": "dataset", + "filename": "MANIFEST.json", + "local_prefix": "" + } + }, + { + "group": "hfd-metadata", + "path": "PATH_MAP.json", + "bytes": 1151, + "sha256": "0e89066ad6758bfbd2811ba497e541aaf2d8f51aeeda3959379b765930f53098", + "source": { + "repo": "Tinghong-Ye22/hfd_data", + "revision": "93ba4a7a9c66b3e1b7041e5f4ea038ef2418b948", + "repo_type": "dataset", + "filename": "PATH_MAP.json", + "local_prefix": "" + } + }, + { + "group": "hfd-metadata", + "path": "index.sqlite", + "bytes": 454651904, + "sha256": "e41b7ba92d5ccd57092b7c3a190828cc0e8a545c79a8bce575633a68f2bb19bf", + "source": { + "repo": "Tinghong-Ye22/hfd_data", + "revision": "93ba4a7a9c66b3e1b7041e5f4ea038ef2418b948", + "repo_type": "dataset", + "filename": "index.sqlite", + "local_prefix": "" + } + }, + { + "group": "hfd-metadata", + "path": "metadata/seed_g1_smplx_pairs.parquet", + "bytes": 26580927, + "sha256": "63fcd6eb1deb1c270d93a7944f9f2fe95a4433c3bf557cb4ca8671bd18197b10", + "source": { + "repo": "Tinghong-Ye22/hfd_data", + "revision": "93ba4a7a9c66b3e1b7041e5f4ea038ef2418b948", + "repo_type": "dataset", + "filename": "metadata/seed_g1_smplx_pairs.parquet", + "local_prefix": "" + } + } + ] +} \ No newline at end of file diff --git a/handoff-v1/inventory/textop-ckpt_299776.pth.json b/handoff-v1/inventory/textop-ckpt_299776.pth.json new file mode 100644 index 0000000000000000000000000000000000000000..8fe8846d02a5b43be1486ccfb6ee83ddc6cd421a --- /dev/null +++ b/handoff-v1/inventory/textop-ckpt_299776.pth.json @@ -0,0 +1,1711 @@ +{ + "denoiser": { + "sequence_pos_encoder.pe": { + "shape": [ + 5000, + 1, + 512 + ], + "dtype": "torch.float32" + }, + "embed_timestep.sequence_pos_encoder.pe": { + "shape": [ + 5000, + 1, + 512 + ], + "dtype": "torch.float32" + }, + "embed_timestep.time_embed.0.weight": { + "shape": [ + 512, + 512 + ], + "dtype": "torch.float32" + }, + "embed_timestep.time_embed.0.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "embed_timestep.time_embed.2.weight": { + "shape": [ + 512, + 512 + ], + "dtype": "torch.float32" + }, + "embed_timestep.time_embed.2.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "embed_text.weight": { + "shape": [ + 512, + 512 + ], + "dtype": "torch.float32" + }, + "embed_text.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "embed_history.weight": { + "shape": [ + 512, + 57 + ], + "dtype": "torch.float32" + }, + "embed_history.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "embed_noise.weight": { + "shape": [ + 512, + 128 + ], + "dtype": "torch.float32" + }, + "embed_noise.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.0.self_attn.in_proj_weight": { + "shape": [ + 1536, + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.0.self_attn.in_proj_bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.0.self_attn.out_proj.weight": { + "shape": [ + 512, + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.0.self_attn.out_proj.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.0.linear1.weight": { + "shape": [ + 1024, + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.0.linear1.bias": { + "shape": [ + 1024 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.0.linear2.weight": { + "shape": [ + 512, + 1024 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.0.linear2.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.0.norm1.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.0.norm1.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.0.norm2.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.0.norm2.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.1.self_attn.in_proj_weight": { + "shape": [ + 1536, + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.1.self_attn.in_proj_bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.1.self_attn.out_proj.weight": { + "shape": [ + 512, + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.1.self_attn.out_proj.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.1.linear1.weight": { + "shape": [ + 1024, + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.1.linear1.bias": { + "shape": [ + 1024 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.1.linear2.weight": { + "shape": [ + 512, + 1024 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.1.linear2.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.1.norm1.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.1.norm1.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.1.norm2.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.1.norm2.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.2.self_attn.in_proj_weight": { + "shape": [ + 1536, + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.2.self_attn.in_proj_bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.2.self_attn.out_proj.weight": { + "shape": [ + 512, + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.2.self_attn.out_proj.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.2.linear1.weight": { + "shape": [ + 1024, + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.2.linear1.bias": { + "shape": [ + 1024 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.2.linear2.weight": { + "shape": [ + 512, + 1024 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.2.linear2.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.2.norm1.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.2.norm1.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.2.norm2.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.2.norm2.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.3.self_attn.in_proj_weight": { + "shape": [ + 1536, + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.3.self_attn.in_proj_bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.3.self_attn.out_proj.weight": { + "shape": [ + 512, + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.3.self_attn.out_proj.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.3.linear1.weight": { + "shape": [ + 1024, + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.3.linear1.bias": { + "shape": [ + 1024 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.3.linear2.weight": { + "shape": [ + 512, + 1024 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.3.linear2.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.3.norm1.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.3.norm1.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.3.norm2.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.3.norm2.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.4.self_attn.in_proj_weight": { + "shape": [ + 1536, + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.4.self_attn.in_proj_bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.4.self_attn.out_proj.weight": { + "shape": [ + 512, + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.4.self_attn.out_proj.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.4.linear1.weight": { + "shape": [ + 1024, + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.4.linear1.bias": { + "shape": [ + 1024 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.4.linear2.weight": { + "shape": [ + 512, + 1024 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.4.linear2.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.4.norm1.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.4.norm1.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.4.norm2.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.4.norm2.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.5.self_attn.in_proj_weight": { + "shape": [ + 1536, + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.5.self_attn.in_proj_bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.5.self_attn.out_proj.weight": { + "shape": [ + 512, + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.5.self_attn.out_proj.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.5.linear1.weight": { + "shape": [ + 1024, + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.5.linear1.bias": { + "shape": [ + 1024 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.5.linear2.weight": { + "shape": [ + 512, + 1024 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.5.linear2.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.5.norm1.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.5.norm1.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.5.norm2.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.5.norm2.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.6.self_attn.in_proj_weight": { + "shape": [ + 1536, + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.6.self_attn.in_proj_bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.6.self_attn.out_proj.weight": { + "shape": [ + 512, + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.6.self_attn.out_proj.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.6.linear1.weight": { + "shape": [ + 1024, + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.6.linear1.bias": { + "shape": [ + 1024 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.6.linear2.weight": { + "shape": [ + 512, + 1024 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.6.linear2.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.6.norm1.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.6.norm1.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.6.norm2.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.6.norm2.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.7.self_attn.in_proj_weight": { + "shape": [ + 1536, + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.7.self_attn.in_proj_bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.7.self_attn.out_proj.weight": { + "shape": [ + 512, + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.7.self_attn.out_proj.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.7.linear1.weight": { + "shape": [ + 1024, + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.7.linear1.bias": { + "shape": [ + 1024 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.7.linear2.weight": { + "shape": [ + 512, + 1024 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.7.linear2.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.7.norm1.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.7.norm1.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.7.norm2.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.7.norm2.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "output_process.weight": { + "shape": [ + 128, + 512 + ], + "dtype": "torch.float32" + }, + "output_process.bias": { + "shape": [ + 128 + ], + "dtype": "torch.float32" + } + }, + "optimizer": { + "state": { + "0": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "1": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "2": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "3": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "4": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "5": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "6": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "7": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "8": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "9": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "10": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "11": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "12": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "13": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "14": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "15": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "16": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "17": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "18": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "19": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "20": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "21": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "22": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "23": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "24": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "25": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "26": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "27": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "28": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "29": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "30": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "31": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "32": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "33": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "34": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "35": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "36": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "37": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "38": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "39": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "40": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "41": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "42": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "43": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "44": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "45": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "46": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "47": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "48": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "49": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "50": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "51": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "52": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "53": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "54": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "55": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "56": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "57": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "58": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "59": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "60": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "61": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "62": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "63": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "64": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "65": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "66": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "67": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "68": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "69": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "70": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "71": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "72": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "73": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "74": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "75": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "76": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "77": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "78": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "79": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "80": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "81": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "82": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "83": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "84": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "85": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "86": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "87": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "88": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "89": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "90": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "91": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "92": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "93": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "94": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "95": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "96": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "97": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "98": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "99": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "100": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "101": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "102": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "103": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "104": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "105": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "106": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "107": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + } + }, + "param_groups": { + "count": 1, + "sample": [ + { + "keys": [ + "lr", + "betas", + "eps", + "weight_decay", + "amsgrad", + "maximize", + "foreach", + "capturable", + "differentiable", + "fused", + "decoupled_weight_decay", + "params" + ], + "count": 12 + } + ] + } + }, + "step": 299776, + "hfd_resume": { + "version": 1, + "rng": { + "python": { + "count": 3, + "sample": [ + 3, + { + "count": 625, + "sample": [ + 2147483648, + 3564348608 + ] + } + ] + }, + "numpy": { + "count": 5, + "sample": [ + "MT19937", + { + "count": 624, + "sample": [ + 196112892, + 4094558418 + ] + } + ] + }, + "torch": { + "shape": [ + 5056 + ], + "dtype": "torch.uint8" + }, + "cuda": { + "count": 1, + "sample": [ + { + "shape": [ + 16 + ], + "dtype": "torch.uint8" + } + ] + } + }, + "train_pool": { + "step": 74944, + "slots": { + "count": 512, + "sample": [ + 21577, + 285324 + ] + }, + "batch_size": 512, + "refresh": 64, + "rng": { + "keys": [ + "bit_generator", + "state", + "has_uint32", + "uinteger" + ], + "count": 4 + }, + "identity": { + "keys": [ + "sessions", + "native_source_sha256", + "asr", + "recipe_sha256", + "foreground_language", + "ordered_session_selection_sha256", + "final_release" + ], + "count": 7 + } + }, + "val_pool": { + "step": 447, + "slots": { + "count": 512, + "sample": [ + 322, + 4914 + ] + }, + "batch_size": 512, + "refresh": 64, + "rng": { + "keys": [ + "bit_generator", + "state", + "has_uint32", + "uinteger" + ], + "count": 4 + }, + "identity": { + "keys": [ + "sessions", + "native_source_sha256", + "asr", + "recipe_sha256", + "foreground_language", + "ordered_session_selection_sha256", + "final_release" + ], + "count": 7 + } + } + } +} \ No newline at end of file diff --git a/handoff-v1/inventory/textop-ckpt_300000.pth.json b/handoff-v1/inventory/textop-ckpt_300000.pth.json new file mode 100644 index 0000000000000000000000000000000000000000..6e1258a7d0f3800c5587459a4d69429c45af95e0 --- /dev/null +++ b/handoff-v1/inventory/textop-ckpt_300000.pth.json @@ -0,0 +1,1711 @@ +{ + "denoiser": { + "sequence_pos_encoder.pe": { + "shape": [ + 5000, + 1, + 512 + ], + "dtype": "torch.float32" + }, + "embed_timestep.sequence_pos_encoder.pe": { + "shape": [ + 5000, + 1, + 512 + ], + "dtype": "torch.float32" + }, + "embed_timestep.time_embed.0.weight": { + "shape": [ + 512, + 512 + ], + "dtype": "torch.float32" + }, + "embed_timestep.time_embed.0.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "embed_timestep.time_embed.2.weight": { + "shape": [ + 512, + 512 + ], + "dtype": "torch.float32" + }, + "embed_timestep.time_embed.2.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "embed_text.weight": { + "shape": [ + 512, + 512 + ], + "dtype": "torch.float32" + }, + "embed_text.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "embed_history.weight": { + "shape": [ + 512, + 57 + ], + "dtype": "torch.float32" + }, + "embed_history.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "embed_noise.weight": { + "shape": [ + 512, + 128 + ], + "dtype": "torch.float32" + }, + "embed_noise.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.0.self_attn.in_proj_weight": { + "shape": [ + 1536, + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.0.self_attn.in_proj_bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.0.self_attn.out_proj.weight": { + "shape": [ + 512, + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.0.self_attn.out_proj.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.0.linear1.weight": { + "shape": [ + 1024, + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.0.linear1.bias": { + "shape": [ + 1024 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.0.linear2.weight": { + "shape": [ + 512, + 1024 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.0.linear2.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.0.norm1.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.0.norm1.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.0.norm2.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.0.norm2.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.1.self_attn.in_proj_weight": { + "shape": [ + 1536, + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.1.self_attn.in_proj_bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.1.self_attn.out_proj.weight": { + "shape": [ + 512, + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.1.self_attn.out_proj.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.1.linear1.weight": { + "shape": [ + 1024, + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.1.linear1.bias": { + "shape": [ + 1024 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.1.linear2.weight": { + "shape": [ + 512, + 1024 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.1.linear2.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.1.norm1.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.1.norm1.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.1.norm2.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.1.norm2.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.2.self_attn.in_proj_weight": { + "shape": [ + 1536, + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.2.self_attn.in_proj_bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.2.self_attn.out_proj.weight": { + "shape": [ + 512, + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.2.self_attn.out_proj.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.2.linear1.weight": { + "shape": [ + 1024, + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.2.linear1.bias": { + "shape": [ + 1024 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.2.linear2.weight": { + "shape": [ + 512, + 1024 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.2.linear2.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.2.norm1.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.2.norm1.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.2.norm2.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.2.norm2.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.3.self_attn.in_proj_weight": { + "shape": [ + 1536, + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.3.self_attn.in_proj_bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.3.self_attn.out_proj.weight": { + "shape": [ + 512, + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.3.self_attn.out_proj.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.3.linear1.weight": { + "shape": [ + 1024, + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.3.linear1.bias": { + "shape": [ + 1024 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.3.linear2.weight": { + "shape": [ + 512, + 1024 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.3.linear2.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.3.norm1.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.3.norm1.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.3.norm2.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.3.norm2.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.4.self_attn.in_proj_weight": { + "shape": [ + 1536, + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.4.self_attn.in_proj_bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.4.self_attn.out_proj.weight": { + "shape": [ + 512, + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.4.self_attn.out_proj.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.4.linear1.weight": { + "shape": [ + 1024, + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.4.linear1.bias": { + "shape": [ + 1024 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.4.linear2.weight": { + "shape": [ + 512, + 1024 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.4.linear2.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.4.norm1.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.4.norm1.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.4.norm2.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.4.norm2.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.5.self_attn.in_proj_weight": { + "shape": [ + 1536, + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.5.self_attn.in_proj_bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.5.self_attn.out_proj.weight": { + "shape": [ + 512, + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.5.self_attn.out_proj.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.5.linear1.weight": { + "shape": [ + 1024, + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.5.linear1.bias": { + "shape": [ + 1024 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.5.linear2.weight": { + "shape": [ + 512, + 1024 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.5.linear2.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.5.norm1.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.5.norm1.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.5.norm2.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.5.norm2.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.6.self_attn.in_proj_weight": { + "shape": [ + 1536, + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.6.self_attn.in_proj_bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.6.self_attn.out_proj.weight": { + "shape": [ + 512, + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.6.self_attn.out_proj.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.6.linear1.weight": { + "shape": [ + 1024, + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.6.linear1.bias": { + "shape": [ + 1024 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.6.linear2.weight": { + "shape": [ + 512, + 1024 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.6.linear2.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.6.norm1.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.6.norm1.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.6.norm2.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.6.norm2.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.7.self_attn.in_proj_weight": { + "shape": [ + 1536, + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.7.self_attn.in_proj_bias": { + "shape": [ + 1536 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.7.self_attn.out_proj.weight": { + "shape": [ + 512, + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.7.self_attn.out_proj.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.7.linear1.weight": { + "shape": [ + 1024, + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.7.linear1.bias": { + "shape": [ + 1024 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.7.linear2.weight": { + "shape": [ + 512, + 1024 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.7.linear2.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.7.norm1.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.7.norm1.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.7.norm2.weight": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "seqTransEncoder.layers.7.norm2.bias": { + "shape": [ + 512 + ], + "dtype": "torch.float32" + }, + "output_process.weight": { + "shape": [ + 128, + 512 + ], + "dtype": "torch.float32" + }, + "output_process.bias": { + "shape": [ + 128 + ], + "dtype": "torch.float32" + } + }, + "optimizer": { + "state": { + "0": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "1": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "2": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "3": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "4": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "5": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "6": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "7": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "8": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "9": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "10": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "11": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "12": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "13": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "14": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "15": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "16": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "17": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "18": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "19": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "20": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "21": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "22": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "23": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "24": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "25": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "26": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "27": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "28": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "29": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "30": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "31": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "32": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "33": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "34": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "35": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "36": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "37": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "38": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "39": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "40": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "41": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "42": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "43": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "44": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "45": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "46": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "47": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "48": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "49": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "50": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "51": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "52": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "53": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "54": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "55": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "56": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "57": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "58": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "59": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "60": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "61": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "62": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "63": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "64": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "65": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "66": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "67": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "68": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "69": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "70": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "71": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "72": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "73": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "74": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "75": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "76": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "77": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "78": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "79": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "80": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "81": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "82": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "83": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "84": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "85": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "86": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "87": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "88": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "89": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "90": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "91": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "92": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "93": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "94": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "95": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "96": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "97": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "98": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "99": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "100": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "101": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "102": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "103": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "104": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "105": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "106": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + }, + "107": { + "keys": [ + "step", + "exp_avg", + "exp_avg_sq" + ], + "count": 3 + } + }, + "param_groups": { + "count": 1, + "sample": [ + { + "keys": [ + "lr", + "betas", + "eps", + "weight_decay", + "amsgrad", + "maximize", + "foreach", + "capturable", + "differentiable", + "fused", + "decoupled_weight_decay", + "params" + ], + "count": 12 + } + ] + } + }, + "step": 300000, + "hfd_resume": { + "version": 1, + "rng": { + "python": { + "count": 3, + "sample": [ + 3, + { + "count": 625, + "sample": [ + 2147483648, + 3564348608 + ] + } + ] + }, + "numpy": { + "count": 5, + "sample": [ + "MT19937", + { + "count": 624, + "sample": [ + 275961101, + 2652905238 + ] + } + ] + }, + "torch": { + "shape": [ + 5056 + ], + "dtype": "torch.uint8" + }, + "cuda": { + "count": 1, + "sample": [ + { + "shape": [ + 16 + ], + "dtype": "torch.uint8" + } + ] + } + }, + "train_pool": { + "step": 75000, + "slots": { + "count": 512, + "sample": [ + 346648, + 294349 + ] + }, + "batch_size": 512, + "refresh": 64, + "rng": { + "keys": [ + "bit_generator", + "state", + "has_uint32", + "uinteger" + ], + "count": 4 + }, + "identity": { + "keys": [ + "sessions", + "native_source_sha256", + "asr", + "recipe_sha256", + "foreground_language", + "ordered_session_selection_sha256", + "final_release" + ], + "count": 7 + } + }, + "val_pool": { + "step": 447, + "slots": { + "count": 512, + "sample": [ + 322, + 4914 + ] + }, + "batch_size": 512, + "refresh": 64, + "rng": { + "keys": [ + "bit_generator", + "state", + "has_uint32", + "uinteger" + ], + "count": 4 + }, + "identity": { + "keys": [ + "sessions", + "native_source_sha256", + "asr", + "recipe_sha256", + "foreground_language", + "ordered_session_selection_sha256", + "final_release" + ], + "count": 7 + } + } + } +} \ No newline at end of file diff --git a/handoff-v1/inventory/textop-environment-lock.txt b/handoff-v1/inventory/textop-environment-lock.txt new file mode 100644 index 0000000000000000000000000000000000000000..9689c38db0c2cb7ba5361a8b6d3b4afe7e105cf2 --- /dev/null +++ b/handoff-v1/inventory/textop-environment-lock.txt @@ -0,0 +1,98 @@ +absl-py==2.5.0 +antlr4-python3-runtime==4.9.3 +appdirs==1.4.4 +certifi==2026.7.22 +cffi==2.1.1 +charset-normalizer==3.5.1 +click==8.5.0 +clip @ file:///root/autodl-tmp/setup/textop/CLIP +cloudpickle==3.1.2 +contourpy==1.3.2 +cycler==0.12.1 +docker-pycreds==0.4.0 +easydict==1.13 +einops==0.8.1 +etils==1.13.0 +filelock==3.32.3 +fonttools==4.65.0 +fsspec==2026.7.0 +ftfy==6.3.1 +gitdb==4.0.12 +GitPython==3.1.62 +glfw==2.10.2 +grpcio==1.84.0 +hf-xet==1.7.0 +huggingface_hub==0.36.2 +hydra-core==1.3.2 +idna==3.20 +importlib_resources==7.1.0 +# Editable install with no version control (isaac_utils==0.0.1) +-e /root/autodl-tmp/datasets/eval_textop_omg/TextOp_code_20260927/TextOp/deps/isaac_utils +Jinja2==3.1.6 +joblib==1.6.0 +kiwisolver==1.5.1 +loguru==0.7.3 +loralib==0.1.2 +Markdown==3.10.3 +MarkupSafe==3.0.3 +matplotlib==3.10.9 +mpmath==1.3.0 +mujoco==3.2.3 +networkx==3.4.2 +numpy==1.24.4 +nvidia-cublas-cu12==12.8.3.14 +nvidia-cuda-cupti-cu12==12.8.57 +nvidia-cuda-nvrtc-cu12==12.8.61 +nvidia-cuda-runtime-cu12==12.8.57 +nvidia-cudnn-cu12==9.7.1.26 +nvidia-cufft-cu12==11.3.3.41 +nvidia-cufile-cu12==1.13.0.11 +nvidia-curand-cu12==10.3.9.55 +nvidia-cusolver-cu12==11.7.2.55 +nvidia-cusparse-cu12==12.5.7.53 +nvidia-cusparselt-cu12==0.6.3 +nvidia-nccl-cu12==2.26.2 +nvidia-nvjitlink-cu12==12.8.61 +nvidia-nvtx-cu12==12.8.55 +omegaconf==2.3.0 +packaging==26.3 +pandas==2.0.3 +pillow==12.3.0 +protobuf==4.25.9 +psutil==7.2.2 +pycparser==3.0 +PyOpenGL==3.1.10 +pyparsing==3.3.2 +python-dateutil==2.9.0.post0 +pytz==2026.3.post1 +PyYAML==6.0.3 +regex==2026.9.10 +requests==2.34.2 +# Editable install with no version control (robotmdar==0.1.0) +-e /root/autodl-tmp/datasets/eval_textop_omg/TextOp_code_20260927/TextOp/TextOpRobotMDAR +sacremoses==0.2.0 +safetensors==0.8.0 +scipy==1.10.1 +sentencepiece==0.2.2 +sentry-sdk==2.69.2 +setproctitle==1.3.7 +six==1.17.0 +smmap==5.0.3 +soundfile==0.14.0 +sympy==1.14.0 +tensorboard==2.20.0 +tensorboard-data-server==0.7.2 +tokenizers==0.22.2 +toolz==1.0.0 +torch==2.7.1+cu128 +torchvision==0.22.1+cu128 +tqdm==4.66.4 +transformers==4.57.6 +triton==3.3.1 +typing_extensions==4.16.0 +tzdata==2026.4 +urllib3==2.8.0 +wandb==0.16.6 +wcwidth==0.8.4 +Werkzeug==3.1.9 +zipp==4.1.0 diff --git a/handoff-v1/inventory/textop-evaluation-checkpoint-index.json b/handoff-v1/inventory/textop-evaluation-checkpoint-index.json new file mode 100644 index 0000000000000000000000000000000000000000..f624b67d77370e3741bca49382c902ae58ce0948 --- /dev/null +++ b/handoff-v1/inventory/textop-evaluation-checkpoint-index.json @@ -0,0 +1,1570 @@ +[ + { + "original_file": "/root/autodl-tmp/runs/hfd-textop-english-split20261026/code_sha256.json", + "archive_file": "root/autodl-tmp/runs/hfd-textop-english-split20261026/code_sha256.json", + "sha256": "d3bca6a3cdf75c36c0f7c85ab2fc5eddb576f2f78afef8ce4353a57eb4022ea0", + "references": [ + { + "key": "/hfd_bridge/configured_data.py", + "value": "f916e0f9d4cdd2a4120fe94a9f1dfe59bbd24606d71ce1941023f04f97f181cd" + }, + { + "key": "/hfd_bridge/kimodo_checkpoint.py", + "value": "b9cd24df225c91a6212c354983e3f51acb495373747f303c6096000af37f7d59" + }, + { + "key": "/upstream/textop/TextOpRobotMDAR/robotmdar/config/__init__.py", + "value": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855" + }, + { + "key": "/upstream/textop/TextOpRobotMDAR/robotmdar/dtype/seed.py", + "value": "1de944913a0d814a42e2224300c9ba716a02326eb473a3cf663b209414850675" + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/hfd-textop-english-split20261026/data_manifest.json", + "archive_file": "root/autodl-tmp/runs/hfd-textop-english-split20261026/data_manifest.json", + "sha256": "10ddacf33963af29f076cad1be4f416d6d558adb92b1cdaf413565b0026e378d", + "references": [ + { + "key": "/native_options/config", + "value": "/root/autodl-tmp/hfd-baseline-portable/configs/sessions.yaml" + }, + { + "key": "/native_options/seed", + "value": 42 + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/hfd-textop-english-split20261026/resume_history/20261028T104547297446Z/code_sha256.json", + "archive_file": "root/autodl-tmp/runs/hfd-textop-english-split20261026/resume_history/20261028T104547297446Z/code_sha256.json", + "sha256": "d3bca6a3cdf75c36c0f7c85ab2fc5eddb576f2f78afef8ce4353a57eb4022ea0", + "references": [ + { + "key": "/hfd_bridge/configured_data.py", + "value": "f916e0f9d4cdd2a4120fe94a9f1dfe59bbd24606d71ce1941023f04f97f181cd" + }, + { + "key": "/hfd_bridge/kimodo_checkpoint.py", + "value": "b9cd24df225c91a6212c354983e3f51acb495373747f303c6096000af37f7d59" + }, + { + "key": "/upstream/textop/TextOpRobotMDAR/robotmdar/config/__init__.py", + "value": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855" + }, + { + "key": "/upstream/textop/TextOpRobotMDAR/robotmdar/dtype/seed.py", + "value": "1de944913a0d814a42e2224300c9ba716a02326eb473a3cf663b209414850675" + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/hfd-textop-english-split20261026/resume_history/20261028T104547297446Z/data_manifest.json", + "archive_file": "root/autodl-tmp/runs/hfd-textop-english-split20261026/resume_history/20261028T104547297446Z/data_manifest.json", + "sha256": "10ddacf33963af29f076cad1be4f416d6d558adb92b1cdaf413565b0026e378d", + "references": [ + { + "key": "/native_options/config", + "value": "/root/autodl-tmp/hfd-baseline-portable/configs/sessions.yaml" + }, + { + "key": "/native_options/seed", + "value": 42 + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/hfd-textop-english-split20261026/resume_history/20261028T104547297446Z/resume.json", + "archive_file": "root/autodl-tmp/runs/hfd-textop-english-split20261026/resume_history/20261028T104547297446Z/resume.json", + "sha256": "cb593ff60d40172001c85a1d5e631bc601f81b663b9fc4c4fb35ce566142c538", + "references": [ + { + "key": "/checkpoint", + "value": "/root/autodl-tmp/runs/hfd-textop-english-split20261026/ckpt_256.pth" + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/hfd-textop-english-split20261026/resume_history/20261028T112850179940Z/code_sha256.json", + "archive_file": "root/autodl-tmp/runs/hfd-textop-english-split20261026/resume_history/20261028T112850179940Z/code_sha256.json", + "sha256": "215fcb308708082c9b15c4091303a888614cd6b3aac68a88234d915bd7ee48bd", + "references": [ + { + "key": "/hfd_bridge/configured_data.py", + "value": "f916e0f9d4cdd2a4120fe94a9f1dfe59bbd24606d71ce1941023f04f97f181cd" + }, + { + "key": "/hfd_bridge/kimodo_checkpoint.py", + "value": "b9cd24df225c91a6212c354983e3f51acb495373747f303c6096000af37f7d59" + }, + { + "key": "/upstream/textop/TextOpRobotMDAR/robotmdar/config/__init__.py", + "value": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855" + }, + { + "key": "/upstream/textop/TextOpRobotMDAR/robotmdar/dtype/seed.py", + "value": "1de944913a0d814a42e2224300c9ba716a02326eb473a3cf663b209414850675" + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/hfd-textop-english-split20261026/resume_history/20261028T112850179940Z/data_manifest.json", + "archive_file": "root/autodl-tmp/runs/hfd-textop-english-split20261026/resume_history/20261028T112850179940Z/data_manifest.json", + "sha256": "e81978233939244cdcecd4d67813c8096f49b5698352a7d1c5ee0e12ede81ab2", + "references": [ + { + "key": "/native_options/config", + "value": "/root/autodl-tmp/hfd-baseline-portable/configs/sessions.yaml" + }, + { + "key": "/native_options/seed", + "value": 42 + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/hfd-textop-english-split20261026/resume_history/20261028T112850179940Z/resume.json", + "archive_file": "root/autodl-tmp/runs/hfd-textop-english-split20261026/resume_history/20261028T112850179940Z/resume.json", + "sha256": "bc8f7f7053c9d2f5eac6aeecad6a58fe6689e6a4c8473ccba5d4c46abadabe40", + "references": [ + { + "key": "/checkpoint", + "value": "/root/autodl-tmp/runs/hfd-textop-english-split20261026/ckpt_1280.pth" + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/hfd-textop-english-split20261026/train_native_manifest.json", + "archive_file": "root/autodl-tmp/runs/hfd-textop-english-split20261026/train_native_manifest.json", + "sha256": "426235c49ee0f422dec6b4697f38ede28f5d63497bc4f4d6eb40c8fb50e1a7fd", + "references": [ + { + "key": "/seed", + "value": 42 + }, + { + "key": "/sampling_seed", + "value": 9000043 + }, + { + "key": "/native_config/audio_sample_step_frames", + "value": 20 + }, + { + "key": "/native_config/action_sample_step_frames", + "value": 1 + }, + { + "key": "/native_config/max_motion_steps", + "value": 0 + }, + { + "key": "/native_config/initial_idle_max_steps", + "value": 7 + }, + { + "key": "/native_config/control_post_idle_min_steps", + "value": 3 + }, + { + "key": "/native_config/control_post_idle_max_steps", + "value": 20 + }, + { + "key": "/native_config/max_speech_onset_steps", + "value": 3 + }, + { + "key": "/native_config/listen_reaction_steps", + "value": 1 + }, + { + "key": "/native_config/max_stream_steps", + "value": 1919 + }, + { + "key": "/final_release/seed_policy", + "value": "recorded_per_session" + }, + { + "key": "/third_party/final_release/seed_policy", + "value": "recorded_per_session" + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/hfd-textop-english-split20261026/val_native_manifest.json", + "archive_file": "root/autodl-tmp/runs/hfd-textop-english-split20261026/val_native_manifest.json", + "sha256": "ec1df1e779bf149962fdd6bf6a910372f51a027ae91201b458724dba5943c6d7", + "references": [ + { + "key": "/seed", + "value": 42 + }, + { + "key": "/sampling_seed", + "value": 9000049 + }, + { + "key": "/native_config/audio_sample_step_frames", + "value": 20 + }, + { + "key": "/native_config/action_sample_step_frames", + "value": 1 + }, + { + "key": "/native_config/max_motion_steps", + "value": 0 + }, + { + "key": "/native_config/initial_idle_max_steps", + "value": 7 + }, + { + "key": "/native_config/control_post_idle_min_steps", + "value": 3 + }, + { + "key": "/native_config/control_post_idle_max_steps", + "value": 20 + }, + { + "key": "/native_config/max_speech_onset_steps", + "value": 3 + }, + { + "key": "/native_config/listen_reaction_steps", + "value": 1 + }, + { + "key": "/native_config/max_stream_steps", + "value": 1919 + }, + { + "key": "/final_release/seed_policy", + "value": "recorded_per_session" + }, + { + "key": "/third_party/final_release/seed_policy", + "value": "recorded_per_session" + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/textop-test-split20261026-20261030/arguments.json", + "archive_file": "root/autodl-tmp/runs/textop-test-split20261026-20261030/arguments.json", + "sha256": "79cc1b3ca8e7c6146e6e1a6f9d60df145a6105467e1fb79651421b314610936f", + "references": [ + { + "key": "/checkpoint", + "value": "/root/autodl-tmp/runs/hfd-textop-english-split20261026/ckpt_300000.pth" + }, + { + "key": "/seed", + "value": 12345 + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/textop-test-split20261026-20261030/cases/test-000033-seed2314733246/initial/inference/report.json", + "archive_file": "root/autodl-tmp/runs/textop-test-split20261026-20261030/cases/test-000033-seed2314733246/initial/inference/report.json", + "sha256": "f783d1f85b606d09680bc82f9c0130421283a08f25803c1542e5f46026145b6f", + "references": [ + { + "key": "/seed", + "value": 12345 + }, + { + "key": "/checkpoint", + "value": "/root/autodl-tmp/models/TextOp-Data/TextOpRobotMDAR/logs/pretrained/checkpoint/ckpt_200000.pth" + }, + { + "key": "/omg_sampling_steps", + "value": null + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/textop-test-split20261026-20261030/cases/test-000033-seed2314733246/trained/inference/report.json", + "archive_file": "root/autodl-tmp/runs/textop-test-split20261026-20261030/cases/test-000033-seed2314733246/trained/inference/report.json", + "sha256": "61c4ad6988fd74213e5994097e11d2da91eada2bfc62ea78c01aab4c1c89bf60", + "references": [ + { + "key": "/seed", + "value": 12345 + }, + { + "key": "/checkpoint", + "value": "/root/autodl-tmp/runs/hfd-textop-english-split20261026/ckpt_300000.pth" + }, + { + "key": "/omg_sampling_steps", + "value": null + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/textop-test-split20261026-20261030/cases/test-000045-seed1597101952/initial/inference/report.json", + "archive_file": "root/autodl-tmp/runs/textop-test-split20261026-20261030/cases/test-000045-seed1597101952/initial/inference/report.json", + "sha256": "56cae4076bd86a8b44ba75de7035769e53175358d5c07734c3a215201b6e3229", + "references": [ + { + "key": "/seed", + "value": 12345 + }, + { + "key": "/checkpoint", + "value": "/root/autodl-tmp/models/TextOp-Data/TextOpRobotMDAR/logs/pretrained/checkpoint/ckpt_200000.pth" + }, + { + "key": "/omg_sampling_steps", + "value": null + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/textop-test-split20261026-20261030/cases/test-000045-seed1597101952/trained/inference/report.json", + "archive_file": "root/autodl-tmp/runs/textop-test-split20261026-20261030/cases/test-000045-seed1597101952/trained/inference/report.json", + "sha256": "205f6c379b60c07eacb92cbb904192ce51aed3f42c362f20d346ec987873a86d", + "references": [ + { + "key": "/seed", + "value": 12345 + }, + { + "key": "/checkpoint", + "value": "/root/autodl-tmp/runs/hfd-textop-english-split20261026/ckpt_300000.pth" + }, + { + "key": "/omg_sampling_steps", + "value": null + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/textop-test-split20261026-20261030/cases/test-000050-seed4281401434/initial/inference/report.json", + "archive_file": "root/autodl-tmp/runs/textop-test-split20261026-20261030/cases/test-000050-seed4281401434/initial/inference/report.json", + "sha256": "213c26b4820d311cae72b6d7a8e328c12e8aef9bd8ba0f39fd2d4b349108548e", + "references": [ + { + "key": "/seed", + "value": 12345 + }, + { + "key": "/checkpoint", + "value": "/root/autodl-tmp/models/TextOp-Data/TextOpRobotMDAR/logs/pretrained/checkpoint/ckpt_200000.pth" + }, + { + "key": "/omg_sampling_steps", + "value": null + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/textop-test-split20261026-20261030/cases/test-000050-seed4281401434/trained/inference/report.json", + "archive_file": "root/autodl-tmp/runs/textop-test-split20261026-20261030/cases/test-000050-seed4281401434/trained/inference/report.json", + "sha256": "8d6ed372504d5cd33b9133d34374e506d9b1020b7b9d128e3da46df0a9b2a742", + "references": [ + { + "key": "/seed", + "value": 12345 + }, + { + "key": "/checkpoint", + "value": "/root/autodl-tmp/runs/hfd-textop-english-split20261026/ckpt_300000.pth" + }, + { + "key": "/omg_sampling_steps", + "value": null + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/textop-test-split20261026-20261030/cases/test-000063-seed1287447995/initial/inference/report.json", + "archive_file": "root/autodl-tmp/runs/textop-test-split20261026-20261030/cases/test-000063-seed1287447995/initial/inference/report.json", + "sha256": "8ea5deeb357a5bafbc6d75433eddf375f2fa909729be3a4c8a990deca491e64d", + "references": [ + { + "key": "/seed", + "value": 12345 + }, + { + "key": "/checkpoint", + "value": "/root/autodl-tmp/models/TextOp-Data/TextOpRobotMDAR/logs/pretrained/checkpoint/ckpt_200000.pth" + }, + { + "key": "/omg_sampling_steps", + "value": null + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/textop-test-split20261026-20261030/cases/test-000063-seed1287447995/trained/inference/report.json", + "archive_file": "root/autodl-tmp/runs/textop-test-split20261026-20261030/cases/test-000063-seed1287447995/trained/inference/report.json", + "sha256": "f72db3dc8b32cdf873b8b5c281567ac16aa25b4682807206c6eab026a5110086", + "references": [ + { + "key": "/seed", + "value": 12345 + }, + { + "key": "/checkpoint", + "value": "/root/autodl-tmp/runs/hfd-textop-english-split20261026/ckpt_300000.pth" + }, + { + "key": "/omg_sampling_steps", + "value": null + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/textop-test-split20261026-20261030/cases/test-000111-seed3714435529/initial/inference/report.json", + "archive_file": "root/autodl-tmp/runs/textop-test-split20261026-20261030/cases/test-000111-seed3714435529/initial/inference/report.json", + "sha256": "1b7fccbe9083ff86a38f76a2be8ecfcbc6aece919ca2389f8ee57995c8d17a16", + "references": [ + { + "key": "/seed", + "value": 12345 + }, + { + "key": "/checkpoint", + "value": "/root/autodl-tmp/models/TextOp-Data/TextOpRobotMDAR/logs/pretrained/checkpoint/ckpt_200000.pth" + }, + { + "key": "/omg_sampling_steps", + "value": null + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/textop-test-split20261026-20261030/cases/test-000111-seed3714435529/trained/inference/report.json", + "archive_file": "root/autodl-tmp/runs/textop-test-split20261026-20261030/cases/test-000111-seed3714435529/trained/inference/report.json", + "sha256": "ea48c3549606f9e676397aad37c4ca4fffb6e6f49d285e40e2b82c4ce4ca0093", + "references": [ + { + "key": "/seed", + "value": 12345 + }, + { + "key": "/checkpoint", + "value": "/root/autodl-tmp/runs/hfd-textop-english-split20261026/ckpt_300000.pth" + }, + { + "key": "/omg_sampling_steps", + "value": null + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/textop-test-split20261026-20261030/cases/test-000421-seed1978152656/initial/inference/report.json", + "archive_file": "root/autodl-tmp/runs/textop-test-split20261026-20261030/cases/test-000421-seed1978152656/initial/inference/report.json", + "sha256": "32f3ee6675bd8a08cc925f2037442ca6030961a77178f22df0f0d5d27876677d", + "references": [ + { + "key": "/seed", + "value": 12345 + }, + { + "key": "/checkpoint", + "value": "/root/autodl-tmp/models/TextOp-Data/TextOpRobotMDAR/logs/pretrained/checkpoint/ckpt_200000.pth" + }, + { + "key": "/omg_sampling_steps", + "value": null + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/textop-test-split20261026-20261030/cases/test-000421-seed1978152656/trained/inference/report.json", + "archive_file": "root/autodl-tmp/runs/textop-test-split20261026-20261030/cases/test-000421-seed1978152656/trained/inference/report.json", + "sha256": "8edfbc7316f4d21daa8e5f206490f12d4a7d358977693532bfbac2df8b7c4f3a", + "references": [ + { + "key": "/seed", + "value": 12345 + }, + { + "key": "/checkpoint", + "value": "/root/autodl-tmp/runs/hfd-textop-english-split20261026/ckpt_300000.pth" + }, + { + "key": "/omg_sampling_steps", + "value": null + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/textop-test-split20261026-20261030/cases/third-test-35fb534d3cf361105fc557a2d0028b7da05d4a12828138ecbd30fb01c537c8cf/initial/inference/report.json", + "archive_file": "root/autodl-tmp/runs/textop-test-split20261026-20261030/cases/third-test-35fb534d3cf361105fc557a2d0028b7da05d4a12828138ecbd30fb01c537c8cf/initial/inference/report.json", + "sha256": "85b644742f42b88b198dd30cc8ebce6f145314bb028c6a9b64aefadcb6b94603", + "references": [ + { + "key": "/seed", + "value": 12345 + }, + { + "key": "/checkpoint", + "value": "/root/autodl-tmp/models/TextOp-Data/TextOpRobotMDAR/logs/pretrained/checkpoint/ckpt_200000.pth" + }, + { + "key": "/omg_sampling_steps", + "value": null + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/textop-test-split20261026-20261030/cases/third-test-35fb534d3cf361105fc557a2d0028b7da05d4a12828138ecbd30fb01c537c8cf/trained/inference/report.json", + "archive_file": "root/autodl-tmp/runs/textop-test-split20261026-20261030/cases/third-test-35fb534d3cf361105fc557a2d0028b7da05d4a12828138ecbd30fb01c537c8cf/trained/inference/report.json", + "sha256": "953bb51c3718bd22b1bbc72a64da33f380d89459808bd8151e3b838e8d6fd169", + "references": [ + { + "key": "/seed", + "value": 12345 + }, + { + "key": "/checkpoint", + "value": "/root/autodl-tmp/runs/hfd-textop-english-split20261026/ckpt_300000.pth" + }, + { + "key": "/omg_sampling_steps", + "value": null + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/textop-test-split20261026-20261030/cases/third-test-8933eb6de44a23953103048740c88aff70eac83163896dc0e9f83b6f88da8408/initial/inference/report.json", + "archive_file": "root/autodl-tmp/runs/textop-test-split20261026-20261030/cases/third-test-8933eb6de44a23953103048740c88aff70eac83163896dc0e9f83b6f88da8408/initial/inference/report.json", + "sha256": "0efa4d31e909e3766905a45d2f7a3f60f8936b8456cdef09eb856adb08b60cb0", + "references": [ + { + "key": "/seed", + "value": 12345 + }, + { + "key": "/checkpoint", + "value": "/root/autodl-tmp/models/TextOp-Data/TextOpRobotMDAR/logs/pretrained/checkpoint/ckpt_200000.pth" + }, + { + "key": "/omg_sampling_steps", + "value": null + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/textop-test-split20261026-20261030/cases/third-test-8933eb6de44a23953103048740c88aff70eac83163896dc0e9f83b6f88da8408/trained/inference/report.json", + "archive_file": "root/autodl-tmp/runs/textop-test-split20261026-20261030/cases/third-test-8933eb6de44a23953103048740c88aff70eac83163896dc0e9f83b6f88da8408/trained/inference/report.json", + "sha256": "f7d659f21e59e814d34c429a07ce5d316a1d5f0ef34a9c319ce064499065cd0b", + "references": [ + { + "key": "/seed", + "value": 12345 + }, + { + "key": "/checkpoint", + "value": "/root/autodl-tmp/runs/hfd-textop-english-split20261026/ckpt_300000.pth" + }, + { + "key": "/omg_sampling_steps", + "value": null + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/textop-test-split20261026-20261030/cases/third-test-f40ec4be3f9aaa51418929e3d258c25eb32216445dc2d867e5563891a59c4770/initial/inference/report.json", + "archive_file": "root/autodl-tmp/runs/textop-test-split20261026-20261030/cases/third-test-f40ec4be3f9aaa51418929e3d258c25eb32216445dc2d867e5563891a59c4770/initial/inference/report.json", + "sha256": "34c96caef75f4552ee64387b2a11013caa3f9c8fd3df2022e9390cb0b9e54f6d", + "references": [ + { + "key": "/seed", + "value": 12345 + }, + { + "key": "/checkpoint", + "value": "/root/autodl-tmp/models/TextOp-Data/TextOpRobotMDAR/logs/pretrained/checkpoint/ckpt_200000.pth" + }, + { + "key": "/omg_sampling_steps", + "value": null + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/textop-test-split20261026-20261030/cases/third-test-f40ec4be3f9aaa51418929e3d258c25eb32216445dc2d867e5563891a59c4770/trained/inference/report.json", + "archive_file": "root/autodl-tmp/runs/textop-test-split20261026-20261030/cases/third-test-f40ec4be3f9aaa51418929e3d258c25eb32216445dc2d867e5563891a59c4770/trained/inference/report.json", + "sha256": "a927b6d22931e20289a91f2f7e912447522036e742b6dae6509fb14fd9a5871b", + "references": [ + { + "key": "/seed", + "value": 12345 + }, + { + "key": "/checkpoint", + "value": "/root/autodl-tmp/runs/hfd-textop-english-split20261026/ckpt_300000.pth" + }, + { + "key": "/omg_sampling_steps", + "value": null + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/textop-test-split20261026-20261030/diagnostics-fixed-300k/summary.json", + "archive_file": "root/autodl-tmp/runs/textop-test-split20261026-20261030/diagnostics-fixed-300k/summary.json", + "sha256": "11db7f88cb82bacf5bad4fcd05c224bb11025a0699c365704fe1d11e8dabfda1", + "references": [ + { + "key": "/checkpoint", + "value": "/root/autodl-tmp/runs/hfd-textop-english-split20261026/ckpt_300000.pth" + }, + { + "key": "/seed", + "value": 12345 + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/textop-test-split20261026-20261030/diagnostics-fixed-300k/test-000033-seed2314733246/gt_history/report.json", + "archive_file": "root/autodl-tmp/runs/textop-test-split20261026-20261030/diagnostics-fixed-300k/test-000033-seed2314733246/gt_history/report.json", + "sha256": "23006952a56a3a3a8a55d8e11fe188fed08f1dfc6673709ce10669a5f5d38d3e", + "references": [ + { + "key": "/seed", + "value": 12345 + }, + { + "key": "/checkpoint", + "value": "/root/autodl-tmp/runs/hfd-textop-english-split20261026/ckpt_300000.pth" + }, + { + "key": "/omg_sampling_steps", + "value": null + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/textop-test-split20261026-20261030/diagnostics-fixed-300k/test-000033-seed2314733246/no_text/report.json", + "archive_file": "root/autodl-tmp/runs/textop-test-split20261026-20261030/diagnostics-fixed-300k/test-000033-seed2314733246/no_text/report.json", + "sha256": "0513fdd4b51c55c400eca39ece34adffc66aa0361ad2dab060cc52c362cc659b", + "references": [ + { + "key": "/seed", + "value": 12345 + }, + { + "key": "/checkpoint", + "value": "/root/autodl-tmp/runs/hfd-textop-english-split20261026/ckpt_300000.pth" + }, + { + "key": "/omg_sampling_steps", + "value": null + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/textop-test-split20261026-20261030/diagnostics-fixed-300k/test-000421-seed1978152656/gt_history/report.json", + "archive_file": "root/autodl-tmp/runs/textop-test-split20261026-20261030/diagnostics-fixed-300k/test-000421-seed1978152656/gt_history/report.json", + "sha256": "70e72f3caf42e9381070a56fc22bfe6f05bdb12f2268a844d6cc939fcc95aaab", + "references": [ + { + "key": "/seed", + "value": 12345 + }, + { + "key": "/checkpoint", + "value": "/root/autodl-tmp/runs/hfd-textop-english-split20261026/ckpt_300000.pth" + }, + { + "key": "/omg_sampling_steps", + "value": null + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/textop-test-split20261026-20261030/diagnostics-fixed-300k/test-000421-seed1978152656/no_text/report.json", + "archive_file": "root/autodl-tmp/runs/textop-test-split20261026-20261030/diagnostics-fixed-300k/test-000421-seed1978152656/no_text/report.json", + "sha256": "f23fae4e40de22297312459d959c6cb498f515b9eff580de18164477e3e539ab", + "references": [ + { + "key": "/seed", + "value": 12345 + }, + { + "key": "/checkpoint", + "value": "/root/autodl-tmp/runs/hfd-textop-english-split20261026/ckpt_300000.pth" + }, + { + "key": "/omg_sampling_steps", + "value": null + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/textop-test-split20261026-20261030/evaluation.json", + "archive_file": "root/autodl-tmp/runs/textop-test-split20261026-20261030/evaluation.json", + "sha256": "c0f682b02da31fe272008bf31e2b76c9849fad11066563f4244bfc1275d0aa38", + "references": [ + { + "key": "/checkpoints/initial/step", + "value": 200000 + }, + { + "key": "/checkpoints/trained/step", + "value": 300000 + }, + { + "key": "/seed", + "value": 12345 + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/textop-test-split20261026-20261030/selection.json", + "archive_file": "root/autodl-tmp/runs/textop-test-split20261026-20261030/selection.json", + "sha256": "7a4b19b1a256f6e3bf3766913901fd6bfd2435e1a8e926cda6d5cd21f3dd4001", + "references": [ + { + "key": "/provider_manifest/seed", + "value": 42 + }, + { + "key": "/provider_manifest/sampling_seed", + "value": 9000049 + }, + { + "key": "/provider_manifest/native_config/audio_sample_step_frames", + "value": 20 + }, + { + "key": "/provider_manifest/native_config/action_sample_step_frames", + "value": 1 + }, + { + "key": "/provider_manifest/native_config/max_motion_steps", + "value": 0 + }, + { + "key": "/provider_manifest/native_config/initial_idle_max_steps", + "value": 7 + }, + { + "key": "/provider_manifest/native_config/control_post_idle_min_steps", + "value": 3 + }, + { + "key": "/provider_manifest/native_config/control_post_idle_max_steps", + "value": 20 + }, + { + "key": "/provider_manifest/native_config/max_speech_onset_steps", + "value": 3 + }, + { + "key": "/provider_manifest/native_config/listen_reaction_steps", + "value": 1 + }, + { + "key": "/provider_manifest/native_config/max_stream_steps", + "value": 1919 + }, + { + "key": "/provider_manifest/final_release/seed_policy", + "value": "recorded_per_session" + }, + { + "key": "/provider_manifest/third_party/final_release/seed_policy", + "value": "recorded_per_session" + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/textop-test-split20261026-20261030/sessions/test-000033-seed2314733246.json", + "archive_file": "root/autodl-tmp/runs/textop-test-split20261026-20261030/sessions/test-000033-seed2314733246.json", + "sha256": "84417ec6730c196f8bbe62a47b2b55949f41fb30e2ccd01942a7936f8ff7140c", + "references": [ + { + "key": "/disfluency_events/0/gap_steps", + "value": 2 + }, + { + "key": "/terminal_idle_steps", + "value": 7 + }, + { + "key": "/idle_steps", + "value": 13 + }, + { + "key": "/listen_steps", + "value": 13 + }, + { + "key": "/motion_steps", + "value": 8 + }, + { + "key": "/semantic_steps", + "value": 34 + }, + { + "key": "/action_idle_steps", + "value": 249 + }, + { + "key": "/action_listen_steps", + "value": 260 + }, + { + "key": "/action_motion_steps", + "value": 160 + }, + { + "key": "/action_steps", + "value": 669 + }, + { + "key": "/sample_seed", + "value": 2314733246 + }, + { + "key": "/config/audio_sample_step_frames", + "value": 20 + }, + { + "key": "/config/action_sample_step_frames", + "value": 1 + }, + { + "key": "/config/max_motion_steps", + "value": 0 + }, + { + "key": "/config/initial_idle_max_steps", + "value": 7 + }, + { + "key": "/config/control_post_idle_min_steps", + "value": 3 + }, + { + "key": "/config/control_post_idle_max_steps", + "value": 20 + }, + { + "key": "/config/max_speech_onset_steps", + "value": 3 + }, + { + "key": "/config/listen_reaction_steps", + "value": 1 + }, + { + "key": "/config/max_stream_steps", + "value": 1919 + }, + { + "key": "/final_release/seed_policy", + "value": "recorded_per_session" + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/textop-test-split20261026-20261030/sessions/test-000045-seed1597101952.json", + "archive_file": "root/autodl-tmp/runs/textop-test-split20261026-20261030/sessions/test-000045-seed1597101952.json", + "sha256": "bbb575dcedd9f90f06a7b68c9ac4821043bbcfe319520856c643bbfae3fe8228", + "references": [ + { + "key": "/terminal_idle_steps", + "value": 5 + }, + { + "key": "/idle_steps", + "value": 10 + }, + { + "key": "/listen_steps", + "value": 11 + }, + { + "key": "/motion_steps", + "value": 25 + }, + { + "key": "/semantic_steps", + "value": 46 + }, + { + "key": "/action_idle_steps", + "value": 189 + }, + { + "key": "/action_listen_steps", + "value": 220 + }, + { + "key": "/action_motion_steps", + "value": 500 + }, + { + "key": "/action_steps", + "value": 909 + }, + { + "key": "/sample_seed", + "value": 1597101952 + }, + { + "key": "/config/audio_sample_step_frames", + "value": 20 + }, + { + "key": "/config/action_sample_step_frames", + "value": 1 + }, + { + "key": "/config/max_motion_steps", + "value": 0 + }, + { + "key": "/config/initial_idle_max_steps", + "value": 7 + }, + { + "key": "/config/control_post_idle_min_steps", + "value": 3 + }, + { + "key": "/config/control_post_idle_max_steps", + "value": 20 + }, + { + "key": "/config/max_speech_onset_steps", + "value": 3 + }, + { + "key": "/config/listen_reaction_steps", + "value": 1 + }, + { + "key": "/config/max_stream_steps", + "value": 1919 + }, + { + "key": "/final_release/seed_policy", + "value": "recorded_per_session" + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/textop-test-split20261026-20261030/sessions/test-000050-seed4281401434.json", + "archive_file": "root/autodl-tmp/runs/textop-test-split20261026-20261030/sessions/test-000050-seed4281401434.json", + "sha256": "e2ca769127679ed1679a2801ba8ec118bebbfa2396619b3c721d881c1a833ffd", + "references": [ + { + "key": "/disfluency_events/0/gap_steps", + "value": 4 + }, + { + "key": "/terminal_idle_steps", + "value": 4 + }, + { + "key": "/idle_steps", + "value": 8 + }, + { + "key": "/listen_steps", + "value": 12 + }, + { + "key": "/motion_steps", + "value": 19 + }, + { + "key": "/semantic_steps", + "value": 39 + }, + { + "key": "/action_idle_steps", + "value": 149 + }, + { + "key": "/action_listen_steps", + "value": 240 + }, + { + "key": "/action_motion_steps", + "value": 380 + }, + { + "key": "/action_steps", + "value": 769 + }, + { + "key": "/sample_seed", + "value": 4281401434 + }, + { + "key": "/config/audio_sample_step_frames", + "value": 20 + }, + { + "key": "/config/action_sample_step_frames", + "value": 1 + }, + { + "key": "/config/max_motion_steps", + "value": 0 + }, + { + "key": "/config/initial_idle_max_steps", + "value": 7 + }, + { + "key": "/config/control_post_idle_min_steps", + "value": 3 + }, + { + "key": "/config/control_post_idle_max_steps", + "value": 20 + }, + { + "key": "/config/max_speech_onset_steps", + "value": 3 + }, + { + "key": "/config/listen_reaction_steps", + "value": 1 + }, + { + "key": "/config/max_stream_steps", + "value": 1919 + }, + { + "key": "/final_release/seed_policy", + "value": "recorded_per_session" + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/textop-test-split20261026-20261030/sessions/test-000063-seed1287447995.json", + "archive_file": "root/autodl-tmp/runs/textop-test-split20261026-20261030/sessions/test-000063-seed1287447995.json", + "sha256": "c00cc1ab3dff8a4e7025a3b1b9a1a15cdbc6294679d6d639b62853ee8bb63dfb", + "references": [ + { + "key": "/terminal_idle_steps", + "value": 5 + }, + { + "key": "/idle_steps", + "value": 8 + }, + { + "key": "/listen_steps", + "value": 4 + }, + { + "key": "/motion_steps", + "value": 19 + }, + { + "key": "/semantic_steps", + "value": 31 + }, + { + "key": "/action_idle_steps", + "value": 149 + }, + { + "key": "/action_listen_steps", + "value": 80 + }, + { + "key": "/action_motion_steps", + "value": 380 + }, + { + "key": "/action_steps", + "value": 609 + }, + { + "key": "/sample_seed", + "value": 1287447995 + }, + { + "key": "/config/audio_sample_step_frames", + "value": 20 + }, + { + "key": "/config/action_sample_step_frames", + "value": 1 + }, + { + "key": "/config/max_motion_steps", + "value": 0 + }, + { + "key": "/config/initial_idle_max_steps", + "value": 7 + }, + { + "key": "/config/control_post_idle_min_steps", + "value": 3 + }, + { + "key": "/config/control_post_idle_max_steps", + "value": 20 + }, + { + "key": "/config/max_speech_onset_steps", + "value": 3 + }, + { + "key": "/config/listen_reaction_steps", + "value": 1 + }, + { + "key": "/config/max_stream_steps", + "value": 1919 + }, + { + "key": "/final_release/seed_policy", + "value": "recorded_per_session" + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/textop-test-split20261026-20261030/sessions/test-000111-seed3714435529.json", + "archive_file": "root/autodl-tmp/runs/textop-test-split20261026-20261030/sessions/test-000111-seed3714435529.json", + "sha256": "feac8344999acc57ff8208830b8bfff6226d7d70ad37dca570e0ed1f886a4ec7", + "references": [ + { + "key": "/terminal_idle_steps", + "value": 11 + }, + { + "key": "/idle_steps", + "value": 23 + }, + { + "key": "/listen_steps", + "value": 8 + }, + { + "key": "/motion_steps", + "value": 11 + }, + { + "key": "/semantic_steps", + "value": 42 + }, + { + "key": "/action_idle_steps", + "value": 449 + }, + { + "key": "/action_listen_steps", + "value": 160 + }, + { + "key": "/action_motion_steps", + "value": 220 + }, + { + "key": "/action_steps", + "value": 829 + }, + { + "key": "/sample_seed", + "value": 3714435529 + }, + { + "key": "/config/audio_sample_step_frames", + "value": 20 + }, + { + "key": "/config/action_sample_step_frames", + "value": 1 + }, + { + "key": "/config/max_motion_steps", + "value": 0 + }, + { + "key": "/config/initial_idle_max_steps", + "value": 7 + }, + { + "key": "/config/control_post_idle_min_steps", + "value": 3 + }, + { + "key": "/config/control_post_idle_max_steps", + "value": 20 + }, + { + "key": "/config/max_speech_onset_steps", + "value": 3 + }, + { + "key": "/config/listen_reaction_steps", + "value": 1 + }, + { + "key": "/config/max_stream_steps", + "value": 1919 + }, + { + "key": "/final_release/seed_policy", + "value": "recorded_per_session" + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/textop-test-split20261026-20261030/sessions/test-000421-seed1978152656.json", + "archive_file": "root/autodl-tmp/runs/textop-test-split20261026-20261030/sessions/test-000421-seed1978152656.json", + "sha256": "efef3818710808bc8bcaee37144d16eae71fe4c624119fda05468f26b2cc05fc", + "references": [ + { + "key": "/terminal_idle_steps", + "value": 3 + }, + { + "key": "/idle_steps", + "value": 12 + }, + { + "key": "/listen_steps", + "value": 19 + }, + { + "key": "/motion_steps", + "value": 31 + }, + { + "key": "/semantic_steps", + "value": 62 + }, + { + "key": "/action_idle_steps", + "value": 229 + }, + { + "key": "/action_listen_steps", + "value": 380 + }, + { + "key": "/action_motion_steps", + "value": 620 + }, + { + "key": "/action_steps", + "value": 1229 + }, + { + "key": "/sample_seed", + "value": 1978152656 + }, + { + "key": "/config/audio_sample_step_frames", + "value": 20 + }, + { + "key": "/config/action_sample_step_frames", + "value": 1 + }, + { + "key": "/config/max_motion_steps", + "value": 0 + }, + { + "key": "/config/initial_idle_max_steps", + "value": 7 + }, + { + "key": "/config/control_post_idle_min_steps", + "value": 3 + }, + { + "key": "/config/control_post_idle_max_steps", + "value": 20 + }, + { + "key": "/config/max_speech_onset_steps", + "value": 3 + }, + { + "key": "/config/listen_reaction_steps", + "value": 1 + }, + { + "key": "/config/max_stream_steps", + "value": 1919 + }, + { + "key": "/final_release/seed_policy", + "value": "recorded_per_session" + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/textop-test-split20261026-20261030/temporal-diagnostics.json", + "archive_file": "root/autodl-tmp/runs/textop-test-split20261026-20261030/temporal-diagnostics.json", + "sha256": "49dd3f066b33bada236c85b8402dd3451025654192ac5946c32116bd58d9af25", + "references": [ + { + "key": "/aggregate/initial/boundary_joint_step_rms_deg", + "value": 1.298330889807807 + }, + { + "key": "/aggregate/initial/inside_joint_step_rms_deg", + "value": 1.097119669119517 + }, + { + "key": "/aggregate/initial/boundary_p95_joint_step_rms_deg", + "value": 2.6945698764589094 + }, + { + "key": "/aggregate/trained/boundary_joint_step_rms_deg", + "value": 0.9475940333472358 + }, + { + "key": "/aggregate/trained/inside_joint_step_rms_deg", + "value": 0.7225109926528401 + }, + { + "key": "/aggregate/trained/boundary_p95_joint_step_rms_deg", + "value": 2.5304082242978936 + }, + { + "key": "/metrics/0/boundary_joint_step_rms_deg", + "value": 1.6781717538833618 + }, + { + "key": "/metrics/0/inside_joint_step_rms_deg", + "value": 1.452820897102356 + }, + { + "key": "/metrics/0/boundary_p95_joint_step_rms_deg", + "value": 3.9983274459838873 + }, + { + "key": "/metrics/1/boundary_joint_step_rms_deg", + "value": 0.9297308921813965 + }, + { + "key": "/metrics/1/inside_joint_step_rms_deg", + "value": 0.6802176833152771 + }, + { + "key": "/metrics/1/boundary_p95_joint_step_rms_deg", + "value": 2.9808097839355474 + }, + { + "key": "/metrics/2/boundary_joint_step_rms_deg", + "value": 0.9674554467201233 + }, + { + "key": "/metrics/2/inside_joint_step_rms_deg", + "value": 0.7487688660621643 + }, + { + "key": "/metrics/2/boundary_p95_joint_step_rms_deg", + "value": 2.2169342041015625 + }, + { + "key": "/metrics/3/boundary_joint_step_rms_deg", + "value": 1.0785208940505981 + }, + { + "key": "/metrics/3/inside_joint_step_rms_deg", + "value": 0.7557613849639893 + }, + { + "key": "/metrics/3/boundary_p95_joint_step_rms_deg", + "value": 3.171843957900999 + }, + { + "key": "/metrics/4/boundary_joint_step_rms_deg", + "value": 1.7991365194320679 + }, + { + "key": "/metrics/4/inside_joint_step_rms_deg", + "value": 1.486899733543396 + }, + { + "key": "/metrics/4/boundary_p95_joint_step_rms_deg", + "value": 3.929531002044674 + }, + { + "key": "/metrics/5/boundary_joint_step_rms_deg", + "value": 2.0738322734832764 + }, + { + "key": "/metrics/5/inside_joint_step_rms_deg", + "value": 1.7535502910614014 + }, + { + "key": "/metrics/5/boundary_p95_joint_step_rms_deg", + "value": 4.408130788803098 + }, + { + "key": "/metrics/6/boundary_joint_step_rms_deg", + "value": 1.6921218633651733 + }, + { + "key": "/metrics/6/inside_joint_step_rms_deg", + "value": 1.5675491094589233 + }, + { + "key": "/metrics/6/boundary_p95_joint_step_rms_deg", + "value": 2.6599767208099365 + }, + { + "key": "/metrics/7/boundary_joint_step_rms_deg", + "value": 1.2134593725204468 + }, + { + "key": "/metrics/7/inside_joint_step_rms_deg", + "value": 1.1091364622116089 + }, + { + "key": "/metrics/7/boundary_p95_joint_step_rms_deg", + "value": 2.8531829714775085 + }, + { + "key": "/metrics/8/boundary_joint_step_rms_deg", + "value": 0.9586306214332581 + }, + { + "key": "/metrics/8/inside_joint_step_rms_deg", + "value": 0.7441407442092896 + }, + { + "key": "/metrics/8/boundary_p95_joint_step_rms_deg", + "value": 2.463936507701874 + }, + { + "key": "/metrics/9/boundary_joint_step_rms_deg", + "value": 0.20458514988422394 + }, + { + "key": "/metrics/9/inside_joint_step_rms_deg", + "value": 0.10725241899490356 + }, + { + "key": "/metrics/9/boundary_p95_joint_step_rms_deg", + "value": 0.4962650015950203 + }, + { + "key": "/metrics/10/boundary_joint_step_rms_deg", + "value": 0.7561571598052979 + }, + { + "key": "/metrics/10/inside_joint_step_rms_deg", + "value": 0.5439569354057312 + }, + { + "key": "/metrics/10/boundary_p95_joint_step_rms_deg", + "value": 1.8875137567520137 + }, + { + "key": "/metrics/11/boundary_joint_step_rms_deg", + "value": 0.21896670758724213 + }, + { + "key": "/metrics/11/inside_joint_step_rms_deg", + "value": 0.11154558509588242 + }, + { + "key": "/metrics/11/boundary_p95_joint_step_rms_deg", + "value": 0.7258833050727795 + }, + { + "key": "/metrics/12/boundary_joint_step_rms_deg", + "value": 1.7376067638397217 + }, + { + "key": "/metrics/12/inside_joint_step_rms_deg", + "value": 1.585784673690796 + }, + { + "key": "/metrics/12/boundary_p95_joint_step_rms_deg", + "value": 3.4010719060897827 + }, + { + "key": "/metrics/13/boundary_joint_step_rms_deg", + "value": 1.6706026792526245 + }, + { + "key": "/metrics/13/inside_joint_step_rms_deg", + "value": 1.187457799911499 + }, + { + "key": "/metrics/13/boundary_p95_joint_step_rms_deg", + "value": 5.5677313804626465 + }, + { + "key": "/metrics/14/boundary_joint_step_rms_deg", + "value": 0.865028977394104 + }, + { + "key": "/metrics/14/inside_joint_step_rms_deg", + "value": 0.7003669142723083 + }, + { + "key": "/metrics/14/boundary_p95_joint_step_rms_deg", + "value": 1.57245934009552 + }, + { + "key": "/metrics/15/boundary_joint_step_rms_deg", + "value": 0.5459399223327637 + }, + { + "key": "/metrics/15/inside_joint_step_rms_deg", + "value": 0.41057950258255005 + }, + { + "key": "/metrics/15/boundary_p95_joint_step_rms_deg", + "value": 1.282839614152908 + }, + { + "key": "/metrics/16/boundary_joint_step_rms_deg", + "value": 1.2306689023971558 + }, + { + "key": "/metrics/16/inside_joint_step_rms_deg", + "value": 1.0437891483306885 + }, + { + "key": "/metrics/16/boundary_p95_joint_step_rms_deg", + "value": 2.121378004550934 + }, + { + "key": "/metrics/17/boundary_joint_step_rms_deg", + "value": 0.59270840883255 + }, + { + "key": "/metrics/17/inside_joint_step_rms_deg", + "value": 0.3870978057384491 + }, + { + "key": "/metrics/17/boundary_p95_joint_step_rms_deg", + "value": 1.2869872152805328 + } + ] + }, + { + "original_file": "/root/autodl-tmp/runs/textop-test-split20261026-20261030/verification.json", + "archive_file": "root/autodl-tmp/runs/textop-test-split20261026-20261030/verification.json", + "sha256": "d1a02cd783c4ee360cfc733e7fc23c0849d05bc09e260cceb03f2f3ebe4f2995", + "references": [ + { + "key": "/gt_equal_across_checkpoints", + "value": true + } + ] + } +] \ No newline at end of file diff --git a/handoff-v1/inventory/textop-final-resume-historical-source.json b/handoff-v1/inventory/textop-final-resume-historical-source.json new file mode 100644 index 0000000000000000000000000000000000000000..3ffaa7c5967205b0f2004a814184fb55bc6196c5 --- /dev/null +++ b/handoff-v1/inventory/textop-final-resume-historical-source.json @@ -0,0 +1,63 @@ +{ + "hfd_bridge/omg_checks.py": { + "expected": "1c80361412892538400cc225ab28aa03d5357ccab7a53e3a6d46deb5da9d32e8", + "recovered_commit": "ea269e3716e289e64fa3cec17f1366b57e3fab9b", + "used_by_this_model": true + }, + "hfd_bridge/resume.py": { + "expected": "be1b665778f3c3bc137f4242b0ac4d1f587903357954725ed509e712dac0ab8f", + "recovered_commit": "exact bytes recovered from prior stored local diff; SHA matches recorded final resume", + "used_by_this_model": true, + "recovered_sha256": "be1b665778f3c3bc137f4242b0ac4d1f587903357954725ed509e712dac0ab8f" + }, + "hfd_bridge/genmo_training.py": { + "expected": "8276d4f266b44e2c112e14021eb29c442753f2de07abf29e9adb321226dc87d8", + "recovered_commit": "ea269e3716e289e64fa3cec17f1366b57e3fab9b", + "used_by_this_model": false + }, + "hfd_bridge/genmo_prefetch.py": { + "expected": "3375352f7668dfd5c76ff629853aa8e4d0bcd2e2294a6e96932280f2eafa68e4", + "recovered_commit": "ea269e3716e289e64fa3cec17f1366b57e3fab9b", + "used_by_this_model": false + }, + "hfd_bridge/hymotion_model.py": { + "expected": "db0968384b87929c6495b87647a37e92e13729a94a34f6a5347f4d23ed05be5c", + "recovered_commit": null, + "used_by_this_model": false + }, + "hfd_bridge/hymotion_data.py": { + "expected": "06170526a18b60fa486f084864e2e65f2608274ed967ef367b29de269f8c9147", + "recovered_commit": null, + "used_by_this_model": false + }, + "hfd_bridge/hymotion_rollout.py": { + "expected": "a7fefe8d72f9e3614103632d68d1d63ce3ee072650739683131f5315d0a41181", + "recovered_commit": null, + "used_by_this_model": false + }, + "hfd_bridge/hymotion_validation.py": { + "expected": "cf6fd4b14fdbafc65a8a65abaade620b43d5b1f95e7a7a6194a21247f2ad4127", + "recovered_commit": null, + "used_by_this_model": false + }, + "hfd_bridge/hymotion_training.py": { + "expected": "67f40e79cc43950f6876f8c0c2c67fe1f606f1b67e692a7dd239457fa4b96a2c", + "recovered_commit": null, + "used_by_this_model": false + }, + "hfd_bridge/hymotion_clock.py": { + "expected": "1eaf969e9198f6902759457488ff8c80fc20c973b72388443b3ef0af7c8282a2", + "recovered_commit": null, + "used_by_this_model": false + }, + "hfd_bridge/hymotion_recovery.py": { + "expected": "e47b841f85dfc19b37a646a6cf6e865dd2642718073a3419a97195f16024c379", + "recovered_commit": null, + "used_by_this_model": false + }, + "train.py": { + "expected": "0724b3db6d374259db9f6378e9b663ef8438e5b969093f3b4d1e46ee56456ddc", + "recovered_commit": "ea269e3716e289e64fa3cec17f1366b57e3fab9b", + "used_by_this_model": true + } +} \ No newline at end of file diff --git a/handoff-v1/inventory/textop-git-commit.txt b/handoff-v1/inventory/textop-git-commit.txt new file mode 100644 index 0000000000000000000000000000000000000000..b497fa5066ab5a016e8f4eb188e2c2e2fbe8ce67 --- /dev/null +++ b/handoff-v1/inventory/textop-git-commit.txt @@ -0,0 +1 @@ +ea269e3716e289e64fa3cec17f1366b57e3fab9b diff --git a/handoff-v1/inventory/textop-git-diff.txt b/handoff-v1/inventory/textop-git-diff.txt new file mode 100644 index 0000000000000000000000000000000000000000..c9c545369cebbcf6d2a05f027ead2f705f40b6ea --- /dev/null +++ b/handoff-v1/inventory/textop-git-diff.txt @@ -0,0 +1,6411 @@ +diff --git a/FINAL_BASELINE_METHOD.md b/FINAL_BASELINE_METHOD.md +deleted file mode 100644 +index f388a06..0000000 +--- a/FINAL_BASELINE_METHOD.md ++++ /dev/null +@@ -1,16 +0,0 @@ +-# 当前保留的方案 +- +-本文为历史完整 session 方法记录。完整调试记录见 [视频问题调试](docs/ROLLOUT_DEBUG.md)。对应产物已归档到 `/root/autodl-tmp/runs/archive/legacy-20261020/hfd-final/`;该目录不随 portable 代码迁移。当前正式训练与评测见 [runs索引](/root/autodl-tmp/runs/README.md)。 +- +-| 模型 | 配置 | 当前连续 joint RMSE | 状态 | +-| --- | --- | ---: | --- | +-| OMG | `configs/omg_full_session.yaml` | 0.0619 rad | 监听与收尾恢复,重复弯腰消失,残余接缝仍在 | +-| TextOP | `configs/textop_full_session.yaml` | 0.2992 rad | 监听姿态出现,任务切换延迟,尚未完整拟合 | +- +-共同改为 `full_session` 关节监督,不向模型输入 listen/idle 等控制标签。训练仍使用原生框架和官方初始化。OMG 保持过去两秒音频+因果 ASR,采用原生 DDIM 10 步、CFG=1、60 帧规划和剩余 24 帧的异步预取;TextOP 保持原生 8 帧输出、CFG=5。 +- +-OMG 三个固定种子的误差为 0.0619、0.1538、0.0585 rad,当前视频沿用原 seed=12345。它仍未证明泛化或完美实时执行。TextOP 真值历史诊断为 0.0125 rad,但这不是连续推理性能;长训练反馈链对照尚未可靠修复闭环。 +- +-视频位于 `videos/omg_full_session_playable.mp4` 和 `videos/textop_full_session_playable.mp4`,均保留原始混合音频。完整报告在各模型的 `inference/`;动作曲线和同 checkpoint 的对照集中在 `debug/`。论文独立测试集评测仍待完成。 +- +-重训前修改配置的 `paths.output_dir`。完整迁移步骤见 [迁移说明](docs/MIGRATION.md);`configs/*_full_session.yaml` 中的 `paths.data_disk` 可通过 `HFD_DATA_DISK` 覆盖。 +diff --git a/JOINT_MAPPING.md b/JOINT_MAPPING.md +deleted file mode 100644 +index 8ebd645..0000000 +--- a/JOINT_MAPPING.md ++++ /dev/null +@@ -1,113 +0,0 @@ +-# HFD 到 TextOP / OMG 的动作映射与历史 bug +- +-## 原始数据和目标接口 +- +-session 的 `motion_q29_xyz_quat` 是 `[T, 36]` 数组,列布局为: +- +-| 列(从 0 开始) | 内容 | 顺序 | +-| --- | --- | --- | +-| `0:29` | G1 的 29 个关节角(rad) | HFD / IsaacLab 顺序 | +-| `29:32` | 根位置 | xyz | +-| `32:36` | 根四元数 | wxyz | +- +-IsaacLab 关节顺序交错排列左右侧和躯干。OMG 的 G1 qpos 使用机器人树顺序:左腿 6 个、右腿 6 个、腰部 3 个、左臂 7 个、右臂 7 个。两边都是 29 个角度,但同一列不一定对应同一个关节。这里不改变机器人关节定义,也不拟合角度变换;按相同关节名重新排列列即可。 +- +-session 保留原始顺序,不在导出阶段预先重排。新导出的 JSON 明确标记 `joint_order=isaaclab`、`quaternion_order=wxyz`;历史 v1 session 缺少这两个字段时按同一原始契约读取。适配器会拒绝显式声明其他顺序的数据,避免二次转换。 +- +-## 按名称转换 +- +-实现集中在 [`hfd_bridge/joints.py`](hfd_bridge/joints.py): +- +-```python +-HFD_TO_NATIVE = [HFD_JOINT_NAMES.index(name) for name in NATIVE_JOINT_NAMES] +-native_joints = raw[:, :29][:, HFD_TO_NATIVE] +-qpos = np.concatenate((raw[:, 29:36], native_joints), axis=1) +-``` +- +-`HFD_TO_NATIVE[i]` 表示“目标第 i 个关节应从原始哪一列取值”,所有下标从 0 开始。得到的 36D qpos 布局是 `[xyz3, quat_wxyz4, native_joints29]`。函数 `hfd_motion_to_qpos` 不改变根位置和四元数。 +- +-实际排列为: +- +-```text +-[0, 3, 6, 9, 13, 17, 1, 4, 7, 10, 14, 18, 2, 5, 8, 11, 15, 19, 21, 23, 25, 27, 12, 16, 20, 22, 24, 26, 28] +-``` +- +-| 目标关节下标 | 关节名 | HFD 原始列 | +-| --- | --- | --- | +-| 0 | `left_hip_pitch_joint` | 0 | +-| 1 | `left_hip_roll_joint` | 3 | +-| 2 | `left_hip_yaw_joint` | 6 | +-| 3 | `left_knee_joint` | 9 | +-| 4 | `left_ankle_pitch_joint` | 13 | +-| 5 | `left_ankle_roll_joint` | 17 | +-| 6 | `right_hip_pitch_joint` | 1 | +-| 7 | `right_hip_roll_joint` | 4 | +-| 8 | `right_hip_yaw_joint` | 7 | +-| 9 | `right_knee_joint` | 10 | +-| 10 | `right_ankle_pitch_joint` | 14 | +-| 11 | `right_ankle_roll_joint` | 18 | +-| 12 | `waist_yaw_joint` | 2 | +-| 13 | `waist_roll_joint` | 5 | +-| 14 | `waist_pitch_joint` | 8 | +-| 15 | `left_shoulder_pitch_joint` | 11 | +-| 16 | `left_shoulder_roll_joint` | 15 | +-| 17 | `left_shoulder_yaw_joint` | 19 | +-| 18 | `left_elbow_joint` | 21 | +-| 19 | `left_wrist_roll_joint` | 23 | +-| 20 | `left_wrist_pitch_joint` | 25 | +-| 21 | `left_wrist_yaw_joint` | 27 | +-| 22 | `right_shoulder_pitch_joint` | 12 | +-| 23 | `right_shoulder_roll_joint` | 16 | +-| 24 | `right_shoulder_yaw_joint` | 20 | +-| 25 | `right_elbow_joint` | 22 | +-| 26 | `right_wrist_roll_joint` | 24 | +-| 27 | `right_wrist_pitch_joint` | 26 | +-| 28 | `right_wrist_yaw_joint` | 28 | +- +-HFD 部署代码 [`upstream/hfd/scripts/deploy/run_sonic_sim_record_online.py`](upstream/hfd/scripts/deploy/run_sonic_sim_record_online.py) 使用相反方向的表,把 MuJoCo 关节值写成 IsaacLab 状态: +- +-```text +-[0, 6, 12, 1, 7, 13, 2, 8, 14, 3, 9, 15, 22, 4, 10, +- 16, 23, 5, 11, 17, 24, 18, 25, 19, 26, 20, 27, 21, 28] +-``` +- +-转换回原生顺序要取这个排列的逆排列,不能直接沿用同一张表。当前代码按关节名生成转换索引,并用部署表独立核对。 +- +-## 两个模型分别如何使用 +- +-**TextOP:** 原生接口只建模 23 个关节。适配器按模型的关节名称从 HFD 29 列中取值,保持 50 Hz,再交给原生 FeatureVersion=3 编码得到 57D 动作特征。选取的原生 29 关节下标为 `0:19` 和 `22:26`,即双腿、腰、双臂的肩和肘;六个腕关节不属于模型输出。实现可以直接使用 `TEXTOP_HFD_JOINTS` 从 raw 取 23 列,等价于先重排 29 列再选取 23 列。TextOP 根旋转接口使用 xyzw,所以输入 wxyz 会归一化并改为 `[x, y, z, w]`;根位置保留。视频中的六个腕关节使用初始化时的值,模型没有预测它们,也没有填入未来真值。 +- +-**OMG:** 原生接口使用全部 29 个关节。适配器将 HFD 的 50 Hz 动作重采样到 30 Hz:关节角和位置做线性插值,四元数做 SLERP;随后把关节列重排为原生顺序。qpos 根四元数保持 wxyz。再通过原生 G1 FK 和 codec 生成 125D 特征。按用户指定的离线测试口径,音频与目标动作窗口对齐。 +- +-渲染统一使用原生 29 关节的 qpos。真值也要经过同一份正确的 HFD 到原生转换,不能把 raw 的前 29 列直接写给渲染器。实现见 [`hfd_bridge/training.py`](hfd_bridge/training.py)、[`infer_overfit.py`](infer_overfit.py) 和 [`render_overfit.py`](render_overfit.py)。 +- +-## 之前的 bug 及影响 +- +-旧适配器误认为 HFD 的前 29 列已经是原生机器人树顺序,只调整了 root 字段的位置,没有重排关节列;TextOP 也按原生数字下标直接选择 HFD 列,因而取错关节。 +- +-例如原始第 1 列(从 0 开始)是 `right_hip_pitch_joint`,而原生第 1 个关节是 `left_hip_roll_joint`。错误代码把右髋 pitch 角写给左髋 roll。多个关节同时错位,会造成姿态明显异常。真值视频也采用了错误解释,因此即使完全没有模型预测,真值机器人仍会表现奇怪。 +- +-之前的往返测试用同一套错误排列做编码和解码,能够还原输入数字,只证明转换自洽,不能证明数字对应了正确的命名关节。错误动作空间内的训练损失下降,也不能证明真实机器人动作拟合正确。 +- +-以下结果不能用于论文中的动作质量或过拟合结论: +- +-- 修复前由旧适配器生成的动作特征、训练检查点、动作误差和预测视频。 +-- `hfd-video-overfit-textop-20261005`、`hfd-video-overfit-omg-20261005` 和 `hfd-overfit-videos-20261005` 中的旧结果,以及更早依赖同一错误排列的动作实验。 +- +-原始 HFD session、混合音频和 ASR 转写可以继续使用;错误动作特征必须重新生成。两个 baseline 已从官方初始化权重重新进行 800 步小样本训练,正确结果目录带 `jointfixed`,摘要见 [`reports/history/overfit_videos.json`](reports/history/overfit_videos.json)。旧视频、废弃适配脚本及旧实验检查点已清理;历史小型配置、日志和检查摘要集中保存在 `runs/diagnostics/historical/`,不能作为修复后模型的评测证据。 +- +-## 如何验证映射 +- +-```bash +-# 在 TextOP 环境运行已有回归测试 +-python -m unittest discover -s tests -t . +- +-# 在 OMG 环境核对独立 FK,并导出相同真值的错误/正确解释对照 +-# --output 必须是新的目录;外部 assets 需要包含 G1 URDF 和 mesh +-python check_joint_mapping.py \ +- --session /data/sessions/sample.npz \ +- --assets-root /data/OMG/assets \ +- --output /data/runs/joint-mapping-audit +-``` +- +-核验包括:部署表的逆排列与按名称转换一致;29 个关节逐个脉冲时仅到达同名电机;根位置和 wxyz 保持不变;50/30 Hz 公共时刻的姿态一致;OMG FK 与独立 MuJoCo 身体位置一致。已有样本核验的最大身体位置误差约 `8.0e-7 m`,见 [`reports/joint_mapping_audit.json`](reports/joint_mapping_audit.json)。这些检查确认数据与机器人接口对应正确;小样本训练确认训练链路可用,尚不构成完整全双工评测结论。 +diff --git a/MANIFEST.json b/MANIFEST.json +index 645a2a1..187efd9 100644 +--- a/MANIFEST.json ++++ b/MANIFEST.json +@@ -5,21 +5,9 @@ + "bytes": 4810, + "sha256": "df3bcbf6da74ff04850db5f301b6b9f84bfbea2ed68798787ad4fbb3277ec63f" + }, +- "FINAL_BASELINE_METHOD.md": { +- "bytes": 1728, +- "sha256": "014307754a7ae7d347ca5b64b6d30cb4b3831639e010ace4b6ca0f2f0b644200" +- }, +- "JOINT_MAPPING.md": { +- "bytes": 7338, +- "sha256": "a5047dc18dd36107ce385f6391460c7782fee6578fc565e7b627792c5096405f" +- }, + "README.md": { +- "bytes": 3805, +- "sha256": "8c55fba302e73f73e43ee5ad0096b852a775c23fd7ba01ac6fc068dedd4a3dd0" +- }, +- "SESSION_FORMAT.md": { +- "bytes": 5054, +- "sha256": "35c27ea0f7fe575cfbe236846ed540c3c95df432c0af8c6148091739a6c92869" ++ "bytes": 9402, ++ "sha256": "a85b9ee971e5a855b347a2b097ae5e4437b38369f1fb250faf6b135664d7e040" + }, + "TRAINING_ASSETS.json": { + "bytes": 87123, +@@ -81,181 +69,9 @@ + "bytes": 16378, + "sha256": "a4245020980dcbe6f8ba1527e96317549937ecaf128d43b8b2f24fef6405b365" + }, +- "docs/ASR_CONTEXT_OVERFIT.md": { +- "bytes": 6819, +- "sha256": "9838a5173dde9bfcffc3bab190282cd04e932b0933b52f1dbb6b7558078610ba" +- }, +- "docs/BASELINE_TEXT_CONDITION.md": { +- "bytes": 9685, +- "sha256": "1d291382f90539d20bc37e491d7c684ee888531fc2241fe53a688b7443363b40" +- }, +- "docs/ENGLISH_ONLY_PROBE.md": { +- "bytes": 9948, +- "sha256": "4ef5a6967a1e9edec4739e9882d97acc18d119a6e23342e599c0f9168d6077c2" +- }, +- "docs/FORMAL_TRAINING_READINESS.md": { +- "bytes": 14837, +- "sha256": "3676c5395bb74d0092327d2db1f8db8ba11ee4327785277b9769f8d196f88621" +- }, +- "docs/GENMO_FORMAL_TRAINING.md": { +- "bytes": 11336, +- "sha256": "eb3761334b649a5806fa5dc1e5029356e4b55e85524621b1e01642bf67a488f6" +- }, +- "docs/GENMO_HFD_OVERFIT.md": { +- "bytes": 12848, +- "sha256": "ca24702e3cfb4f359bfba677b24a013711cd0471f4255d2f8b619ab9ca327e93" +- }, +- "docs/GENMO_LISTENING_DIAGNOSIS.md": { +- "bytes": 3858, +- "sha256": "d54bdc5141d0bde62f1d76258f4ab353cbfe2d7fc713bffc6788326d322f58ee" +- }, +- "docs/GENMO_PAST_AUDIO_TEST.md": { +- "bytes": 9364, +- "sha256": "2a95909082c8ad998c542a205f1c2d3b4600bac04b9334753e5a16428356dbef" +- }, +- "docs/GENMO_V2_TRAINING.md": { +- "bytes": 6611, +- "sha256": "5c02cd209ccd1f9cf87f4f60b2d5fc54f00b412bde97ffa261856a3d96f4a397" +- }, +- "docs/GENMO_V2_VALIDATION.md": { +- "bytes": 5740, +- "sha256": "40d7cf46d333e194205dedf310954111d9fdf5ec014c4ca86dea4e3417824285" +- }, +- "docs/HFD_HUMAN_BASELINE_INPUT_PROTOCOL.md": { +- "bytes": 12323, +- "sha256": "dec2e8df1522e6740ef53ee1147e8e65cd318408c7835f65c507d56b5d28edd5" +- }, + "docs/HFD_HUMAN_BASELINE_SOURCES.json": { +- "bytes": 10598, +- "sha256": "14dae98d032442b7c86cd5f4375ff8aac2b8a99a8e8fa6840d6266ecf93f93a8" +- }, +- "docs/HUMAN_FORMAL_VALIDATION_20261024.md": { +- "bytes": 4630, +- "sha256": "286eec4ad74f26c93aa5ca5c91acbfb8b0fc0412c4ce743269a67bb074081ec1" +- }, +- "docs/HUMAN_RETARGET_EVALUATION.md": { +- "bytes": 7664, +- "sha256": "9efd14122d7f8461ccc9cf2db2d5843e02aac9b1b3b392b728b13517d3775a87" +- }, +- "docs/KIMODO_CLOSED_LOOP_STABILITY.md": { +- "bytes": 17001, +- "sha256": "7c29773be52bd16548208c4365cf4b31e4a9fa5b0b87efa9ee0eba6400d9f3f5" +- }, +- "docs/KIMODO_FORMAL_VALIDATION.md": { +- "bytes": 6733, +- "sha256": "d89622850760e9889549495e1b93dd30d4bc4f099ff9faf67f3384089527a2cf" +- }, +- "docs/KIMODO_HFD_OVERFIT_RESULTS.md": { +- "bytes": 6166, +- "sha256": "19dc1cc2dd3a0c6dba023b7b6ca979f0478b82c427e582b6386a9d94b512c091" +- }, +- "docs/KIMODO_HFD_TRAINING.md": { +- "bytes": 13291, +- "sha256": "4b7033db2727ded4d64a015c43d12b6fcd547a40a2e38662b2a33e2e3c9f0037" +- }, +- "docs/KIMODO_MOTION_RECOVERY.md": { +- "bytes": 7245, +- "sha256": "dc4f6192b533301946b218a51c0ab442572dd96996544e27b100f5594d7d2f22" +- }, +- "docs/KIMODO_ROOT_COLLAPSE_DIAGNOSIS.md": { +- "bytes": 4618, +- "sha256": "42a7a07344f523abb3bfb4d22214655141e4d4b38f7ca4d081f688ab57dc695d" +- }, +- "docs/KIMODO_ROOT_TRAINING_CAUSE_ANALYSIS.md": { +- "bytes": 8012, +- "sha256": "e2c119d0cc7e0cc65bf3c741142ad88c7844c41b5d33228189288848caac29a2" +- }, +- "docs/KIMODO_TRAINING_REPAIR.md": { +- "bytes": 5601, +- "sha256": "880ba0e4477735500471f7ac36fd075d63d874b4578b4ba8284045530670915e" +- }, +- "docs/LOCAL_INDEX_TRAINING.md": { +- "bytes": 6004, +- "sha256": "ac2dc3538b1214bf348bb672a79a6d6e1ea0b4d74cee907fa2416fe0f66405bd" +- }, +- "docs/MIGRATION.md": { +- "bytes": 426, +- "sha256": "1bbfc85ca07c183bdb6f25143d08db7dc1e526e9fb29993225a3232e74c689af" +- }, +- "docs/OMG_ENGLISH_VALIDATION.md": { +- "bytes": 3791, +- "sha256": "86a2a91471737d81fe9ef741aac5e224d3f727db9f6c08161c3bb98b78dee89b" +- }, +- "docs/OMG_TRAINING_REVIEW.md": { +- "bytes": 14254, +- "sha256": "d38b0b5049701c620935e0bf51bf0da77aa3cb2de675c33ba348b972039609d8" +- }, +- "docs/OMG_VALIDATION_DIAGNOSIS.md": { +- "bytes": 6133, +- "sha256": "5269c34a70f40977a9848106a59e05abbe9ee8486870ca4682e808cd0ac4259b" +- }, +- "docs/OMG_VALIDATION_ROLLOUT.md": { +- "bytes": 6132, +- "sha256": "ff8c7c8f249c3c89607f080ff281e15acc3d8fccc710c0f8563e417ecfa7d82e" +- }, +- "docs/PORTABLE_TRAINING.md": { +- "bytes": 11281, +- "sha256": "4b47034c1d5916537605e96153450ba5d8f3e89cb5e1fa4796e4a18b960ba973" +- }, +- "docs/REFERENCE_DYNAMIC_TRAINING.md": { +- "bytes": 9587, +- "sha256": "e55ead5d171134b3ef2a9238e4f28bd5315ace74733481f5033f6befe8b175b7" +- }, +- "docs/ROLLOUT_DEBUG.md": { +- "bytes": 8444, +- "sha256": "c45aefa40fe66dd1504465ec8254d4438aa4325b2f9687afda871e7c312e971d" +- }, +- "docs/SEMANTICS_CLOCK_FIX.md": { +- "bytes": 5914, +- "sha256": "66d1d57f18e69176b5ad91bf980a8f180350137b9c2b02d97ae1a71fa5d627cc" +- }, +- "docs/TEXTOP_ENGLISH_VALIDATION.md": { +- "bytes": 5350, +- "sha256": "c653a371cb97f690349d96c9750366cb570bd53cedb64695b283be9f67371374" +- }, +- "docs/TRAINING_REPRODUCTION_MATRIX.md": { +- "bytes": 4693, +- "sha256": "0c977387862ae82c48c2667e5d3f4520a9c522e6394a8d4e68cfa9db2a9919ad" +- }, +- "docs/TRAINING_RESUME.md": { +- "bytes": 4215, +- "sha256": "3caa601fdaef09d30ab79c195299fc5dfe52cd0b32be4daaf97e6bc7d7484253" +- }, +- "docs/WORKSPACE_CLEANUP.md": { +- "bytes": 4422, +- "sha256": "d10560c0c62a61f6cbee040c90c5db4fbe7167aecb286c2d8d637d161c72e90b" +- }, +- "docs/history/BASELINE_METHOD_BEFORE_SEMANTIC_FIX.md": { +- "bytes": 10085, +- "sha256": "60cfdeb3ee30d306e40af466723eaa98d73c2efa0a806b8b219a9b79f8013847" +- }, +- "docs/history/DUPLEX_TASK.md": { +- "bytes": 7084, +- "sha256": "75c02d717b23b000d95b35fbda0799326dece83f0e6ec893a368fe5e07ba420c" +- }, +- "docs/history/HFD_HUMAN_BASELINE_INPUT_PROTOCOL_RESEARCH_20261017.md": { +- "bytes": 40979, +- "sha256": "ad30933320fc00b863d50f7f102ab5c546c5a82ce4919ab4956af6c93997f702" +- }, +- "docs/history/HFD_WINDOW_PROTOCOL.md": { +- "bytes": 7986, +- "sha256": "4a5a25129aa0b5cd1173af7d2c9d365cdcb118a56856cb197f547f1e5e05c3d4" +- }, +- "docs/history/NORMAL_TASK.md": { +- "bytes": 6244, +- "sha256": "9a5600eeb00436d77603501666c910ca3ed536a93c696015ed706b5f5bff9c55" +- }, +- "docs/history/OMG_CONTEXT_DEBUG.md": { +- "bytes": 6920, +- "sha256": "34d50e904a8cfc721656e7939f6aff3c07021726d9a46b818ac853945d90134e" +- }, +- "docs/history/README.md": { +- "bytes": 670, +- "sha256": "09b277b367872ab7203e6ee1e6e48d31c75269e651982a1286f074b343b2ef43" +- }, +- "docs/history/VIDEO_SAMPLES.md": { +- "bytes": 7997, +- "sha256": "ff3489e07a6233176669aa5f2567284dbd880cb36e3f6089dbc60f3d9516295b" ++ "bytes": 10428, ++ "sha256": "2cc65718bc6bf9d0d089ab341db266f9bfc18fee1d6cdec4f80978ddf74f1493" + }, + "env.sh": { + "bytes": 631, +@@ -478,8 +294,8 @@ + "sha256": "1dc280ec6f2c008924b1a69079ee50d8e8281c0c2da2d291fc5c9f9c91a36769" + }, + "tools/configure_training.py": { +- "bytes": 6935, +- "sha256": "1519a086061cda3bf93aa1f7188bc0746c95efe2084cafb7feefc09405855308" ++ "bytes": 7506, ++ "sha256": "88ebaa442e230394b7929a85f26a696a309c5315fc4ce2f0b3307af730b7f7bc" + }, + "tools/download_hfd_tars.py": { + "bytes": 6701, +@@ -2348,6 +2164,42 @@ + "validate_sessions.py": { + "bytes": 1431, + "sha256": "8f7c045d7e9f75deeffc5777eefe31760c893c929fb25d2dffda8d8e520bd880" ++ }, ++ "docs/CURRENT_RUNS.md": { ++ "bytes": 15135, ++ "sha256": "a65e3eb72a8da86e705d134a9c43aa97be17aa51f53f4d3d1bf4c858faf7c2eb" ++ }, ++ "train_hymotion.py": { ++ "bytes": 3070, ++ "sha256": "c93a486c828963e568750da26ff81256804633fd34d28d9415d3535168c7c0b0" ++ }, ++ "train_hymotion_autoregressive.py": { ++ "bytes": 15136, ++ "sha256": "1438ed061573814dfa5fe9afe6914fb59df1e70d0215958dcbd9e1b114be839f" ++ }, ++ "hfd_bridge/hymotion_model.py": { ++ "bytes": 7571, ++ "sha256": "ed29d2d03b247ed681a09dc230697885c5ab8fa8b30f82c4a827aa8fe0b41597" ++ }, ++ "hfd_bridge/hymotion_data.py": { ++ "bytes": 2203, ++ "sha256": "79503602a9a4de4f67931fa944a198d6f438b531a363ffbe545ecc4dd54f97e7" ++ }, ++ "hfd_bridge/hymotion_training.py": { ++ "bytes": 3382, ++ "sha256": "792b76cfc35f58b0c7c94b40a660f00c4971fff1d978ffd0c923b9f26a5fa733" ++ }, ++ "hfd_bridge/hymotion_rollout.py": { ++ "bytes": 2186, ++ "sha256": "6e27c4e24bd42a211df7be2a0271337af4e658d66e5d303633912bbd40313517" ++ }, ++ "hfd_bridge/hymotion_evaluation.py": { ++ "bytes": 9348, ++ "sha256": "fe72667281f04a55bc8c88375b7ca3cee1f1b6de2c23fcd7d45f1e73c1530a99" ++ }, ++ "configs/hymotion_c24.yaml": { ++ "bytes": 1123, ++ "sha256": "258c03cd4ed31c41575c4aafe0e3ddd5758a0654a5f75a903833be6ff0359bca" + } + } + } +diff --git a/README.md b/README.md +index be11660..aac9867 100644 +--- a/README.md ++++ b/README.md +@@ -1,43 +1,861 @@ +-# HFD baseline training ++# HFD 五模型训练约定 + +-OMG、TextOp、GenMo和Kimodo在全双工HFD上的训练代码、配置及复现文档。仓库包含动态session组装、因果文本释放、原生训练、验证loss和完整断点恢复。HY-Motion尚无完整训练入口。 ++本文件维护当前训练约定;运行进度、有效权重、评测结论和待办见 [CURRENT_RUNS.md](docs/CURRENT_RUNS.md)。除非用户明确要求,不另写方案、诊断、历史报告等文档。 + +-| 模型 | 入口 | 正式配置 | 预算 | +-|---|---|---|---| +-| OMG | `train.py` | `configs/omg.yaml`,另支持50M/100M/300M/500M生成配置 | 100K,有效batch512 | +-| TextOp | `train.py` | `configs/textop.yaml` | 300K,batch512,三阶段各100K | +-| GenMo/GEM-SMPL v2 | `train_genmo_full.py` | `configs/genmo.yaml`,旧配方`configs/genmo_legacy.yaml` | 100K,batch32 | +-| Kimodo-SMPLX | `train_kimodo.py` | `configs/kimodo.yaml` | 100K,batch8 | ++## 数据与信息释放 + +-输入采用按词时间释放的英文前景GT文本、最近30秒上下文及各自的动作历史。OMG另输入已释放的过去2秒音频。session在训练时动态组装,不预导出全量session,不运行ASR或中文翻译。详见[输入协议](docs/HFD_HUMAN_BASELINE_INPUT_PROTOCOL.md)。 ++- 共用 [configs/sessions.yaml](configs/sessions.yaml) 的 final_release:profile=isolated_benchmark_1024_20261026;索引 Tinghong-Ye22/data_vqa_all_1025@bdca2fe7581c263e9ceb3a96204b314e199faddf,原料 Tinghong-Ye22/hfd_data@93ba4a7a9c66b3e1b7041e5f4ea038ef2418b948。 ++- 完整发布 train/val/test=258661/512/1024,含中文。当前仅英文:OMG/TextOp=126950/302/436(原生+第三方);GenMo/Kimodo/HY=117290/297/419(有配对SMPL-X的原生样本)。英文子集不等于完整1024 benchmark。 ++- 只回放动作session和原始配方,不下载或传入VQA问答、视频、问答配音;语音—动作与视频—音频完全解耦。保留严格session身份、动作长度和音频时钟检查。先按最终split/语言过滤,再按记录的upstream_split、dataset_index、seed回放,不重新随机划分。 ++- 英文文本用 aligned_reference_context、text_conditioning=none;读取本地预建GT词时间映射 /root/autodl-tmp/cache/hfd/reference-alignments.sqlite,不在线ASR、翻译或重做对齐。6608条严格对齐失败的源资产沿用已建立的整句结束后保守释放映射。 ++- 麦克风窗长0.6秒、步长0.4秒,词末加0.1秒guard后在下一个释放点可见;动作时刻t只访问麦克风时刻≤t+0.2的已释放文本,保留最近30秒,编码器另有token上限。session切换时清空。 ++- 不输入未来PCM、未释放文字、GT turn/listen/idle标签、未来时刻、根路径或姿态。带噪未来训练目标不是额外观测条件。人体标签从配对SMPL-X按原生事件配方构造,不以G1关节角代替。 + +-GenMo默认启用执行段监督、预测历史恢复与CPU预取,使用新的输出目录;旧正式权重保留。训练链路和短程检查不代表已完成新策略的全量质量验收,详见[GenMo v2](docs/GENMO_V2_TRAINING.md)。四模型及OMG多个初始化权重的覆盖范围见[复现矩阵](docs/TRAINING_REPRODUCTION_MATRIX.md)。 ++## 五模型默认训练 + +-本次速度、断点恢复及过拟合检查的实际结果见[GenMo v2验证记录](docs/GENMO_V2_VALIDATION.md),包括仍存在的动作不足问题。 ++分别使用 /root/autodl-tmp/envs/模型名/bin/python,共同seed=42;精确超参数以配置和代码为准。 + +-从[训练复现指南](docs/PORTABLE_TRAINING.md)开始,准备对应环境和外部资产后生成本机配置: ++| 模型/配置 | 条件与输出 | 正式训练预算 | ++|---|---|---| ++| OMG-B:[omg.yaml](configs/omg.yaml) | 冻结T5-Base,50 token;过去2秒已释放PCM;10帧历史@30Hz。原生125D、G1全29关节,计划60帧,异步缓冲剩24帧重规划 | 50M,100K,batch512,BF16;LR6e-5,2000 warmup后cosine至1e-6 | ++| OMG-L:[omg_l.yaml](configs/omg_l.yaml) | 与B相同的HFD数据、因果条件、125D表示和diffusion-only微调损失;官方300M/55K初始化 | GPU0,100K,microbatch64×累积8=512;BF16,LR6e-5,warmup2000+cosine | ++| OMG-XL:[omg_xl.yaml](configs/omg_xl.yaml) | 与B相同的HFD协议;官方500M/50K初始化;沿用严格加载器,仅排除官方旧版52个未使用visual_film张量 | GPU1,100K,microbatch32×累积16=512;BF16,LR6e-5,warmup2000+cosine | ++| TextOp:[textop.yaml](configs/textop.yaml) | 冻结CLIP ViT-B/32,77 token含边界;2帧57D历史@50Hz,无音频。128D latent经冻结VAE解码8帧,G1原生23关节,未建模腕关节固定 | 300K,三个100K阶段,batch512,FP32,LR1e-4;保留原生分阶段历史恢复 | ++| GenMo:[genmo.yaml](configs/genmo.yaml) | GEM-SMPL全量训练,冻结T5-3B,50 token;60帧人体历史,past_audio=false。151D,未来60/执行12帧@30Hz | 100K,batch32,BF16,LR2e-5→1e-6;regression+diffusion、执行段权重3、根速度1.0、旋转增量0.05;预测历史最高25%,1000 warmup+4000 ramp,DDIM10 | ++| Kimodo:[kimodo.yaml](configs/kimodo.yaml) | 冻结LLM2Vec,最近400正文token,独立文本worker;60帧人体历史,无音频。SMPLX22原生273D,因果root分解、root→body两阶段,未来60/执行12帧 | 100K,batch8,BF16;normalized_x0+解码连续性/根历史重建;根只训投影,冻结根注意力骨干;根/身体分开优化裁剪,LR上限2e-6/1e-5 | ++| HY-Motion:[hymotion_c24.yaml](configs/hymotion_c24.yaml) | 官方Lite+rank16/alpha32 LoRA;冻结Qwen3-8B 128正文token+CLIP-L,无音频。201D,历史60/未来60/执行12帧;10段预测反馈,Euler10/CFG1,反馈detach | C24完整训练100K,batch4,BF16,LR2e-5 cosine至1e-6;24帧连续恢复,根速度1.0/旋转增量0.05,前200更新渐启,接段权重3 | + +-```bash +-python tools/configure_training.py \ +- --work-root /data/hfd_work --assets-root /data/hfd_assets \ +- --raw-root /data/hfd_assets/hfd_data --gpu 0 +-source /data/hfd_work/training-configs/training.env ++TextOp每条训练短链有4个8帧primitive,共0.64秒;链首始终使用GT历史,后续按课程选择预测历史,每个新batch重新初始化短链。第三阶段不等于整条session完全自主展开。 + +-/data/hfd_work/envs/omg/bin/python train.py --config /data/hfd_work/training-configs/omg.yaml +-/data/hfd_work/envs/textop/bin/python train.py --config /data/hfd_work/training-configs/textop.yaml +-/data/hfd_work/envs/genmo/bin/python train_genmo_full.py --config /data/hfd_work/training-configs/genmo.yaml +-/data/hfd_work/envs/kimodo/bin/python train_kimodo.py --config /data/hfd_work/training-configs/kimodo.yaml +-``` ++关节顺序及TextOp关节选择以 [joints.py](hfd_bridge/joints.py) 为准;沿用原生统计、冻结编码器/VAE、体型和表示变换。依赖与资产由 [requirements](requirements)、[UPSTREAM_LOCK.json](UPSTREAM_LOCK.json)、[TRAINING_ASSETS.json](TRAINING_ASSETS.json)、[HYMOTION_ASSETS.json](docs/HYMOTION_ASSETS.json) 记录。 + +-命令分别运行;同一GPU顺序执行。中断后原命令加`--resume auto`,Kimodo也支持`--resume`。恢复要求配置、代码、数据身份一致,预算是总步数。 ++## 缓存、存储与恢复 + +-- `hfd_bridge/`:训练数据、动作表示、适配、文本worker、验证与断点管理。 +-- `upstream/`:按训练依赖裁剪的HFD、OMG、TextOp原生代码,保留许可证;独立评测与部署代码归入本地`test/`。 +-- `tools/`:数据下载/对齐、配置生成、外部资源连接与构建。 +-- `requirements/`、`UPSTREAM_LOCK.json`、`TRAINING_ASSETS.json`:冻结环境、外部源码revision及必要外部文件的大小、SHA256和来源。 +-- `TRAINING_SNAPSHOT.json`:迁移前当前训练源码hash;Git历史保留对应实现。 +-- `docs/`:训练协议、复现操作和诊断说明,历史实验页不作为当前默认命令。 ++- 原始tar在 /root/autodl-fs/pangxinlei/hfd_data,保留分片、不批量解压;换索引无需删除原料。新下载默认共享盘models/datasets/downloads,下载/装依赖先用 /root/.local/bin/turbo,回退严格遵守 /root/AGENTS.md。 ++- 热索引、词映射、文本embedding、派生session放tmp。热资产 /root/autodl-tmp/cache/hfd/hot-assets 已约16.6 GiB有效内容,不用默认4 GiB准备命令缩小;长尾人体原料仍按需读共享tar。 ++- OMG/TextOp原生session内存缓存8 GiB,当前分别每256/64 batch刷新。GenMo train/val缓存4/8 GiB、文本512 MiB、预取16、训练每16 batch刷新、CPU4/BLAS1。GPU瞬时空闲不代表适合加batch或并发任务。 ++- 日志用 /root/autodl-tmp/runs;OMG/TextOp权重同目录,人体模型正式权重用 /root/autodl-fs/pangxinlei/models/模型名/。tmp给checkpoint和输出留空间;共享df不代表账号配额。 ++- 完整续训须恢复模型、优化器、scheduler、已消费采样池和原总预算,并校验代码/数据/资产合同;仅加载权重须明确重置优化器与步数。训练运行期间不改其合同约束源码/配置。 ++- OMG-B/TextOp每256步完整断点、每2000步验证;GenMo/Kimodo每2000步保存并验证。GenMo另在128/256/512/1024步滚动保存一份完整早期断点,正常2000步保存成功后清理早期断点,后续保留最近2份及best。 ++- OMG-L/XL先采样完整512窗口逻辑批,再切成GPU小批;保留B的512个session槽、每256逻辑批刷新及相同随机采样顺序。每2000次优化器更新验证8192窗口(16×512)并保存完整断点,保留最近2份及验证best;日志loss按完整更新聚合。`val_check_interval`已换算为小批次数,避免累积梯度时过频验证。 ++- OMG-L/XL日志与运行合同分别在`/root/autodl-tmp/runs/hfd-omg-{l,xl}-english-split20261026`,完整权重/优化器在共享盘`/root/autodl-fs/pangxinlei/models/OMG/hfd-{l,xl}-english-split20261026`,本地checkpoints/best/last路径为兼容链接。使用原生DirectTorchCheckpointIO流式临时文件+原子替换,恢复同时检查配置、原运行归属、RNG和已消费的逻辑批状态。 ++- GenMo环境固定NumPy1.26.4、SciPy1.15.3、Pandas2.2.3,主进程/CPU生产进程检查版本并写入合同及权重;保留无梯度AMP采样cache_enabled=False修复、主干梯度检查及严格回放。CPU审计失败时保存最近完成更新并停止。 + +-本地单元测试、过拟合/探针、评测/渲染代码和原始报告集中到`test/`,由`.gitignore`忽略。旧入口的兼容链接也被忽略。GitHub不包含这些代码、模型权重、数据、日志、环境、视频或缓存;CLIP自带的压缩分词词表是必要训练资源。原生validation及断点完整性检查属于训练实现。 ++以下仅用于任务停止后续训;不要对运行中的目录重复启动: + +-进一步阅读:[迁移操作](docs/PORTABLE_TRAINING.md)、[GenMo训练](docs/GENMO_FORMAL_TRAINING.md)、[Kimodo训练](docs/KIMODO_HFD_TRAINING.md)、[文本时间轴](docs/REFERENCE_DYNAMIC_TRAINING.md)、[resume](docs/TRAINING_RESUME.md)、[GenMo倾听姿态诊断](docs/GENMO_LISTENING_DIAGNOSIS.md)。 ++~~~bash ++cd /root/autodl-tmp/hfd-baseline-portable ++/root/autodl-tmp/envs/omg/bin/python -u train.py --config configs/omg.yaml --resume auto --set 'runtime.gpu_ids=[1]' ++/root/autodl-tmp/envs/omg/bin/python -u train.py --config configs/omg_l.yaml --resume auto ++/root/autodl-tmp/envs/omg/bin/python -u train.py --config configs/omg_xl.yaml --resume auto ++/root/autodl-tmp/envs/textop/bin/python -u train.py --config configs/textop.yaml --resume auto --set 'runtime.gpu_ids=[0]' ++/root/autodl-tmp/envs/genmo/bin/python -u train_genmo_full.py --config configs/genmo.yaml --resume auto ++/root/autodl-tmp/envs/kimodo/bin/python -u train_kimodo.py --config configs/kimodo.yaml --resume auto ++CUDA_VISIBLE_DEVICES=1 /root/autodl-tmp/envs/hymotion/bin/python -u train_hymotion_autoregressive.py --config configs/hymotion_c24.yaml --set gpu=0 --set text.device=cuda:0 --resume auto ++~~~ ++ ++HY训练入口为 train_hymotion_autoregressive.py,当前唯一正式配方为configs/hymotion_c24.yaml。train_hymotion.py只提供模型/合同公共函数,不再包含旧训练循环。 ++ ++## HY C24完整训练约定 ++ ++- 用户于2026-10-30授权完整训练及HY专项清理。原始训练数据、split、英文筛选、样本选择、采样顺序和时间映射不变;不加入静止历史增强。恢复监督在训练目标构造中计算,不回写GT。 ++- 从C24验证集选出的best仅加载权重。该best与上一轮C的2000步模型完全一致,已独立复制到共享models/hymotion/hfd-c24-full-split20261026/initialization.pt;初始化来源、SHA256和选择理由保存在新运行目录initialization.json。优化器/步数重置,新增100K预算,不声称完整续接旧优化器。 ++- 保留原XY/yaw对齐;姿态SO(3)差和高度差在24帧内按w=1−3s²+2s³消除。保留flow,根速度权重1.0、旋转增量0.05,前200更新渐启;执行段接段第一帧权重3,FP32几何、无效尾帧屏蔽。 ++- 预测历史策略仍以每链GT历史启动,之后反馈预测;每链10段共4秒,不能描述成跨整个session完全自主训练。训练数据和实际采样约定不因代码清理改变。 ++- 初始和每2000步用6条未参与旧38K选优的validation、两个固定噪声seed、最长20秒自主闭环。best按含运动误差的score选择,并要求平均接段比≤3;best只是当前最优,不等于行为验收合格。 ++- last每256步、验证结束及最终步原子保存;保留initial/best/last,不累计中间断点。启动先完成17步完整断点,再自动恢复原100K预算,检查恢复链路。 ++- 旧split完整训练runs/hfd-hymotion-english及共享models/hymotion/hfd-english保留;近期HY实验目录和中间权重清理后,正式训练不依赖它们。结果摘要见CURRENT_RUNS,必需初始化和基线指标保存在当前运行目录。 ++- 保留数据/连续性/恢复/模型回归检查;清除旧时钟输入、静止/单窗恢复实验分支和一次性对照脚本。当前测试评测入口为test/evaluation/evaluate_hymotion.py,可通过--run/--weights指定运行;仅在用户要求评测时启动GPU评测。 ++ ++## 评测约定 ++ ++自主闭环只开头给GT姿态,之后用预测历史,不逐窗重置GT;不做视频平滑、rootfit或DTW。人体模型先生成完整原始序列,再做整段GMR。训练loss或短smoke通过不等于行为、连续性和根运动验收;验证集用于选权重/调参,测试集只报告固定模型。 ++ ++## 动作评测脚本入口 ++ ++- 移植来源为 Tinghong-Ye22/HFD-motion-evaluation-handoff-20261101@7901e042c9cf06e0081d7303a4b2c229136ec70b;下载原件在 /root/autodl-fs/pangxinlei/downloads/HFD-motion-evaluation-handoff-20261101,工作目录在 /root/autodl-tmp/hfd-motion-evaluation/HFD_bench。 ++- 命令入口 /root/autodl-tmp/hfd-motion-evaluation/bench,使用 configs/local_motion.json;只读看板入口为同目录 review --port 8770。增量环境 /root/autodl-tmp/envs/hfd-motion-eval 复用 OMG 基础环境,不修改原训练环境;实际依赖见工作目录 requirements-local.lock。 ++- 冻结识别器5c04cec、阈值、GT+0.7秒计分参照和原始结果均保留;模型预测不平移。交互衡量发布的motion reference,非实机执行。UNKNOWN、未适用和未响应删失分别保留。 ++- 质量入口 motion-score 调用指定官方源码与本机已有 evaluator/T5-3B/HoloMotion,包含原生29DOF到MuJoCo的置换。旧113587的14项质量分数仍因 invalid_input_joint_order 失效;测试样例通过不代表新模型完整质量评测完成。大型参考缓存、视频和971502训练权重未全部下载,不直接启动旧调度/恢复队列。 ++ ++## 五模型全英文1024条动作测试集 ++ ++- 2026-11-01用户指定用于OMG、TextOp、GenMo、Kimodo、HY-Motion。独立派生集在 `/root/autodl-tmp/datasets/hfd-test-english1024-v1`:保留原1024个test场景及1937个动作轮,共2229个motion事件;原436条英文前景保留,588条中文前景重建为对应英文。动作场景、原始session身份和训练/验证划分不变,没有从其他划分补动作样本。 ++- 文件入口为 `manifest.jsonl`、`events.jsonl`、`membership.jsonl`、`evaluation_contract.json`;每例 `sessions/test-en-NNNN.npz` 包含50Hz IsaacLab G1与配对159D SMPL-X、16维betas、16kHz麦克风和评分标签。原文件保留,新版本不覆盖原始test合同或旧分数。 ++- 生成器与读取接口在 `/root/autodl-tmp/hfd-motion-evaluation/english1024`。使用OMG环境Python;`dataset_loader.English1024.input_at(index, motion_frame, model)` 只给初始已知姿态及已释放文本,后续历史由模型预测;`scoring_targets` 单独返回GT,`paired_human_30hz` 复用原人体几何变换。五模型各用相同1024例,不沿用旧人体provider的419条筛选。 ++- 英文均复用已有审计语音。更换语言后重算麦克风、控制事件和动作时间轴,沿用已发布的30帧终端时钟。2例改用其他合格英文音色,37例因英文提示语更长而最小延长停顿;20例中文第三方干扰换为上游held-out英文干扰,其中7例原前景已为英文。保留干扰类型、事件数及同音色aside关系;重叠背景按不相交聚合片段移动一次,aside避开前景指令。 ++- 文本协议仍为离线前景reference alignment,30秒上下文、0.1秒guard、microphone clock;它是oracle参考转写,不是新ASR结果,guard不代表经认证的时间误差上限。原训练/validation配置和已运行模型均未修改。 ++- 全量核验见数据集 `verification.json`:2002个实际英文语音资产SHA匹配;全部1024例G1/人体/时钟/语言/事件通过;419条原生英文的音频、动作和标签经旧加载器重放逐项一致;436条原英文动作轮标注一致,429例完整motion标注一致(另7例替换了中文背景,浮点时钟按1ns容差比较);35组任务×模型输入检查及2项混音/aside回归通过。 ++- 用户随后确认保留流式输入与listen/idle动作,已启动下述完整流式质量评测。交互识别/评分尚未重跑;本次质量使用预先固定的GT执行窗口,不需要推断预测首尾,也不复用旧971502分数。 ++ ++## 完整流式动作质量评测 ++ ++- 入口 `/root/autodl-tmp/hfd-motion-evaluation/stream-quality`;`status` 查看进度,`run --gpus 0 1` 启动或恢复(运行中会拒绝重复启动),`collect` 汇总已有评分。2026-11-01已启动,脚本在 `/root/autodl-tmp/hfd-motion-evaluation/stream_quality`;模型和训练代码未改。 ++- 协议 `full_stream_fixed_gt_windows_v1`:五模型逐包/逐次读取现有因果输入,单个已知姿态启动,之后只有执行过的预测历史。完整保留listen/idle/执行输出,不按GT任务边界重启,不把完整目标文本提前交给生成器。OMG保留过去2秒PCM及native async缓冲;其余模型保留既有因果reference文本协议,非在线ASR。 ++- 每模型生成全部1024例。1937个GT动作窗口在模型生成前按event ID排序,官方batch32取1920个,其余17个明确列入excluded;五模型成员一致。按固定绝对时间切出预测,响应慢、未响应或仍为listen的输出照常保留;不寻找最像GT的片段,不平移、不DTW、不平滑。模型没有收到这些评分窗口。 ++- 使用现有验证集best:OMG100000、GenMo72000、Kimodo100000、HY14000;TextOp使用最终300000。精确路径/SHA/配置见评测输出run_config.json、checkpoint_metadata.json、preflight.json。人体完整流先整段GMR,再切评分窗口,GMR只在session开头重置。 ++- 大文件在 `/root/autodl-fs/pangxinlei/evaluations/hfd-english1024-stream-quality-v1`;每模型 `models//sessions/` 保存完整预测及因果审计,`phase_diagnostics.json` 单独保存listen/motion/idle的关节、根误差和根速度,`packed/` 保存官方输入,`scores/motion_quality.json` 保存14项结果。汇总自动写summary.json/csv,不编造一个总分;阶段诊断不冒充官方14项。 ++- 日志与进度在 `/root/autodl-tmp/runs/hfd-five-stream-quality-english1024-v1`;两张GPU动态排队,CPU并行整段GMR,模型失败会保留错误并阻止该模型缺样本评分。GPU评分调用指定原版OmniHM/HoloMotion、tracker-mode=run与原有分块FK执行器。 ++- 恢复时校验并复用已完成的1024条生成结果与packed输入。评分经独立`stream_quality/score.py`调用交接包同一`OmniHM.evaluate`接口,将输出写入用户指定共享盘;仍执行冻结配置、官方源码及输入SHA校验。GMR读取进度文件遇到短暂缺失时等待下次读取,其他错误照常报告;summary同步显示运行/失败阶段。 ++- 启动前已通过五模型8秒流式生成、三模型整段GMR、4项因果/固定时窗回归,29个互异关节值经过baseline及指定官方转换后逐项一致。短检查不是完整1024评分,最终状态及分数以实际产物为准。 ++- 2026-11-02已完成五模型全部评分:各1024条完整生成、各1920固定窗口、14项均有值,完整结果见上述summary.json/csv与docs/CURRENT_RUNS.md。跟踪MPJPE等比较HoloMotion执行与生成目标;Diversity的GT参考约1.17965。显示FID已×100、C-Slide已×10,引用时保留该缩放。 ++ ++ ++## 隔离恢复训练 pilot(2026-11-09) ++ ++首轮 HY 与 Kimodo 的独立 500 更新恢复试验已结束,均未通过闭环验收;首轮入口 `/root/autodl-tmp/hfd-motion-evaluation/HFD_bench/scripts/recovery_training/supervise_pilot.py --config /config.yaml`,分别使用对应原训练环境。运行目录 `` 为共享盘 `models/hfd-recovery-pilot-20261109/{hymotion,kimodo}`。必须先有实际容量预留收据;supervisor 先做 2 更新保存与完整恢复检查,再继续固定预算。新试验加入单姿态启动的预测历史链和真实历史回放,HY 去掉 C24 混合标签,Kimodo 保留独立 expert 速度标签;从旧 best 显式权重初始化,保持原推理 steps/CFG,旧正式训练及已有成绩不替换。 ++ ++权重、缓存、临时文件和日志都有独立预算与持续磁盘/inode 检查。验证新增动作幅度、进度和误差非退化门槛;只有合格闭环 checkpoint 可成为新 best。500 更新结束不自动延长,不把 loss 降低等同于 FID 提升。具体预算、选优门槛、实时状态路径见 docs/CURRENT_RUNS.md。若受控停训,先排查对应错误收据并重新验证可写空间,再严格恢复,不能用删除合同检查的方式续跑。 ++ ++### 当前第二轮恢复试验 ++ ++第二轮已启动,入口改为 `HFD_bench/scripts/recovery_training_v2/supervise_pilot.py --config /config.yaml`,运行目录为共享盘 `models/hfd-recovery-pilot-v2-20261109/{hymotion,kimodo}`。从各自原正式 best 重新初始化,各 250 更新,0/125/250 固定验证;首轮输出保留,不采用其退化 last 替换正式模型。历史采样、推理 steps/CFG 和验收条件保持。 ++ ++监督改为 XY/yaw 对齐、六帧 SO(3)/高度过渡,第六帧起回到对齐 GT;位置与导数 loss 使用同一目标。HY 连续性辅助项恢复,warmup 200 更新,root velocity 权重 1、rotation velocity 0.05、边界权重 3。仅修改训练标签,推理不加入 GT 或平滑。增加段内突变、根 jerk 诊断及三次完整验证轨迹;125 更新额外保留一份权重快照(Kimodo 为 model-only,完整恢复使用 last)。只有原定门槛全部通过才写 best,尚无 FID 改善结论。 ++ ++实际新增容量预留 HY 2 GiB、Kimodo 8 GiB 已通过;checkpoint 单次写入上限分别为 256/2560 MiB,其他资源阈值和断点保护沿用首轮。第二轮进度命令为对应 Python 执行 `HFD_bench/scripts/status_recovery_training.py --root /root/autodl-fs/pangxinlei/models/hfd-recovery-pilot-v2-20261109`。证据与当前状态见 docs/CURRENT_RUNS.md。 ++ ++迭代资源维护:每轮核查已失效临时文件和可再生的重复缓存,确认进程结束、无打开句柄和保留副本有效后及时清理;当前运行的容量预留、有效断点和诊断证据保留。日志轮转、缓存限额、受控结束释放剩余预留继续执行。 ++ ++用户已授权每小时自主检查 HY/Kimodo 恢复训练,完成后分析并持续迭代,及时清理无用中间产物;验证效果有明确希望且空间预算通过时,可直接展开分阶段正式训练,无需例行再次确认。当前会话心跳为每小时整点(Etc/UTC),复用原 auto-review 权限;具体任务、状态与操作范围见 HFD_bench/artifacts/recovery_hourly_task.json 及 recovery_hourly_task_request.json。只用 train/validation 选参,保留既定验收,更多 validation 复核后再升级正式训练,最终测试协议保持。 ++ ++ ++### 第三轮恢复试验(2026-11-09 11:14 UTC) ++ ++第二轮两模型均已完成 250 更新,均未通过门槛。HY 最终活跃动作幅度比 0.358、姿态误差 0.443 rad(初始 0.372,恶化约 19%),行走案例 961 仍未响应;Kimodo 幅度比 0.842、接段比 2.355,但根位置误差 0.527 m(初始 0.442,恶化约 19%)且行走结果不稳定。不能升级正式训练,也没有新的正式 FID。 ++ ++当前最新目录为 `/root/autodl-fs/pangxinlei/models/hfd-recovery-pilot-v3-20261109/{hymotion,kimodo}`,入口 `HFD_bench/scripts/recovery_training_v3/supervise_pilot.py`。各从原正式 best 初始化,仍为 250 更新及固定 0/125/250 验证。每模型只改变一个训练变量:HY 预测历史链概率 0.75→0.25,增加真实历史回放;Kimodo 根网络从仅投影训练改为完整解冻,学习率不变。保留 C6 标签、因果输入、两随机种子与全部验收门槛。监督器已启动,先执行初始验证及两更新保存/严格恢复检查,未通过不会继续剩余预算。 ++ ++实际写入/fsync 预算 HY 2 GiB、Kimodo 12 GiB 已通过,Kimodo 完整断点预计 3.2 GiB、单次原子写入上限 4 GiB。第二轮重复启动索引已清理 656.24 MiB,正式权重、恢复断点、125 快照和完整轨迹保留。逐例结果、配置差异、启动与清理清单见 HFD_bench/artifacts/recovery_hourly_20261109T11_analysis.json、recovery_v3_config_diff.json、recovery_v3_launch.json、recovery_cleanup_hourly_20261109T11.json。每小时任务的 latest_run_root 已更新为 v3。 ++ ++ ++用户最新要求(2026-11-09):任一模型正式训练成功启动并确认进程、日志、容量及恢复检查正常后,立即删除每小时心跳任务 `416e9f54`,不等待训练或最终评测完成,不自动重建。当前有界恢复 pilot 仍继续每小时检查;取消心跳不停止训练,也不关闭训练监督器的资源与断点保护。此要求覆盖此前“最终评测完成后取消”的旧条件。 ++ ++ ++### 每小时检查与第四轮 Kimodo(2026-11-09 12:04 UTC) ++ ++当前两模型轮次不同:HY 仍在 `/root/autodl-fs/pangxinlei/models/hfd-recovery-pilot-v3-20261109/hymotion`,Kimodo 最新为 `/root/autodl-fs/pangxinlei/models/hfd-recovery-pilot-v4-20261109/kimodo`,以 `artifacts/recovery_hourly_task.json` 的 latest_models 分别定位。尚未启动正式训练,小时任务保持;任一正式训练健康启动后立即取消任务的要求不变。 ++ ++HY v3 在本次检查期间推进至 168/250,loss/梯度有限、无错误,完整两更新恢复检查通过(断点 136,897,993 bytes)。125 验证幅度比 0.261、接段比 1.672、姿态误差 0.423 rad、根位置误差 0.487 m,未通过;维持原有 250 更新上限,不延长或重启。 ++ ++Kimodo v3 已完成 250,两更新完整断点 3,394,252,981 bytes、恢复校验通过,无数值或资源故障。最终幅度比 0.942、接段比 2.299、姿态误差 0.422 rad,但根位置误差从 0.442 增至 0.675 m(约 +53%),根 jerk RMS 从 357.30 增至 700.46 m/s³。行走案例 961 两种子进度比 -0.466/-0.055;案例 2327 最大根速度仍有 10.98/18.81 m/s,尖峰仍落在每段第二帧。125 同样不合格。已读取 0/125/250 全部逐例完整轨迹,不能将动作幅度达标当成恢复成功,也无新的正式 FID。 ++ ++Kimodo v4 从相同原正式 best 重新开始,相对 v3 只新增 stage1_root_velocity 权重 0.5:对反归一化第一阶段根位置求物理速度,与相同 C6 因果目标比较,包含历史/未来边界并屏蔽无效帧。目标是阻止第一阶段错误被身体骨盆偏移短暂抵消。保持完整根网络解冻、学习率、250 更新、固定 0/125/250 验证、两个种子和原验收门槛;不改变推理或加入 GT 条件。新损失通过真实表征检查:精确目标为零、补偿式跳变仍被识别、梯度有限且仅作用于根分量、无效尾帧不计分,见 artifacts/recovery_v4_objective_checks.json。源码入口 scripts/recovery_training_v4/supervise_pilot.py;准备与运行独立于 HY v3。 ++ ++第四轮 12 GiB 预算已实际写入/fsync,通过物理分配校验;4 GiB 原子写入上限、8 小时墙钟及原存储阈值不变。监督器已启动,先进行初始验证和两更新完整恢复检查,该检查尚待本轮执行完成。当前本地约 1.54 GiB 空闲、共享约 34,900 空闲 inode,小文件写入/fsync 通过;没有把共享 df 总空闲当账号配额。 ++ ++清理 v3 Kimodo 的 16 个重复启动索引及已分析且不再使用的 125 步落败快照,共 1,407.06 MiB。已核对进程结束、无打开句柄、无配置引用、缓存保留副本 SHA 相同;保留最终完整断点、原正式 best、必要旧对照及全部验证原始轨迹。清单 artifacts/recovery_cleanup_hourly_20261109T12.json。分析、配置差异与启动收据分别为 recovery_hourly_20261109T12_analysis.json、recovery_v4_config_diff.json、recovery_v4_launch.json。下一检查点为 HY 最终验证及 Kimodo v4 恢复/125 验证,不因平均指标改善放宽逐例行走门槛。 ++ ++ ++### 第五轮会话覆盖对照(2026-11-09 13:05 UTC) ++ ++最新两模型统一位于 `/root/autodl-fs/pangxinlei/models/hfd-recovery-pilot-v5-20261109/{hymotion,kimodo}`,入口 `HFD_bench/scripts/recovery_training_v5/supervise_pilot.py`。HY v3 和 Kimodo v4 均已完成 250 更新,完整恢复检查通过、训练数值有限、无资源错误,但未通过闭环门槛。尚未正式开训,小时检查继续;任一正式训练健康启动后立即取消心跳。 ++ ++HY v3 最终幅度比 0.263、接段比 1.686、姿态误差 0.427 rad、根位置误差 0.467 m、根 jerk 279.24 m/s³;两行走案例四种子组合进度比 0.336/-0.037/0.352/-0.010,均低于下限。Kimodo v4 最终幅度比 0.856、接段比 2.336、姿态误差 0.408 rad、根位置误差 0.564 m、根 jerk 376.07 m/s³。相对 v3 根 jerk 改善,但行走仍不可靠:案例 2327 两种子进度比 0.046/0.790,961 为 -0.168/0.653。v4 的 125 步姿态/位置非退化门槛通过,但两例进度未通过,故保留该快照诊断,不升级正式训练。已读取各轮初始/125/250 全部逐例完整轨迹;这些不是正式 FID。 ++ ++采样链路发现:WindowPool.refresh 按链起点批次计数,不按优化器更新计数。Kimodo v4 最终完整断点的 pool.step=6、refresh=16、slots=[32781,251404,112021,179604],证明 250 更新一直使用最初四个会话;HY v3 pool.step=7、refresh=2,日志实际覆盖 16 个不同会话。长链重复同一池显著限制本轮样本覆盖,但不能据此认定它是质量差的唯一原因。第五轮每模型只将 data.refresh_batches 改为 1(Kimodo 16→1,HY 2→1),每批新链重新抽取会话;仍保留最多 50 段闭环链、原概率、各自上一轮损失和学习率、同一原正式 best 初始化及全部验收门槛,各 250 更新,0/125/250 验证。没有读取 Test1024 选参。Kimodo 日志新增实际采样位置和池计数,便于后续核查覆盖。 ++ ++真实 WindowPool 的受控测试验证六批起点在 refresh16/2/1 时分别覆盖 4/12/24 个模拟会话,保存并恢复后下一批完全一致;实际覆盖仍由运行日志判断。检查和差异收据为 artifacts/recovery_v5_sampling_checks.json、recovery_v5_config_diff.json,启动源码 SHA 与进程在 recovery_v5_launch.json。原因果标签及推理实现未改;两更新完整断点/严格恢复仍由监督器先验证,第五轮目前处于该启动阶段,尚无训练后改善结论。 ++ ++HY 2 GiB、Kimodo 12 GiB 预算已重新实际写入/fsync 并确认物理分配;本地约 1.54 GiB 空闲,共享约 34,800 空闲 inode。保持原子替换预算、8 小时墙钟、目录项/缓存/日志上限及资源受控停止,不将共享 df 视为账号配额。清理结束轮次的重复启动索引及 HY v3 125 落败快照 786.77 MiB,已检查无进程占用、无必要配置引用、缓存 SHA 一致,保留当前完整断点、Kimodo v4 125 候选及全部轨迹。清单 recovery_cleanup_hourly_20261109T13.json,完整分析 recovery_hourly_20261109T13_analysis.json。状态脚本已修复缺少某模型目录时的报错,返回 present=false;没有修改任何旧训练合同源码。 ++ ++下一检查点为第五轮两模型恢复收据、实际会话覆盖以及 125/250 固定验证。失败时继续依据证据调整,不能仅延长当前失败配方或降低验收条件。 ++ ++ ++### 小时检查与 Kimodo 第六轮(2026-11-09 14:07 UTC) ++ ++当前 HY 仍为 `models/hfd-recovery-pilot-v5-20261109/hymotion`,Kimodo 最新为 `models/hfd-recovery-pilot-v6-20261109/kimodo`,根目录均在 `/root/autodl-fs/pangxinlei/`,以任务 JSON 的 latest_models 为准。没有正式训练启动;达到正式启动并核实健康条件后,仍应立即删除心跳 416e9f54。 ++ ++HY 本次检查从 138 推进到约 154/250;恢复检查通过(136,897,929 bytes)、训练数值有限、无错误,实际已记录 17 个训练会话。125 验证幅度比 0.252、接段比 1.661、姿态误差 0.424 rad、根位置误差 0.478 m,未通过。保持当前有界训练与原定最终验证,不抢占 GPU0。 ++ ++Kimodo v5 完成 250 更新,严格恢复通过(3,394,253,109 bytes),实际训练覆盖 25 个会话。最终幅度比 0.919、接段比 2.365、姿态误差 0.409 rad、根位置误差 0.371 m;相比原初始 0.468 rad/0.442 m,两种误差均改善。但行走案例 2327 两种子进度 0.496/0.519、961 为 0.547/0.227,仍有低于 0.5 的案例,不能四舍五入放宽门槛。根 jerk RMS 411.10 m/s³,高于初始 357.30,段首尖峰仍存在。125 的根误差 0.592 m、jerk 606.01,亦未通过。已读取 0/125/250 全部逐例原始轨迹;平均数改善不等于正式训练资格,也不是 FID 改善。 ++ ++完成同一 v5 final 权重下四个 validation 历史窗口、两种子、预测过去与真实过去的 16 次配对诊断,输出 `/root/autodl-fs/pangxinlei/models/hfd-recovery-history-diagnostics-v5-20261109/`。输入仅为截至 decision 的历史,无未来 GT;真实历史是诊断对照,不能作为正式验证或替代闭环成绩。案例 2327 decision396 的第二帧最大根速度从预测历史约 5.96/5.97 m/s 降至真实过去 0.12/0.32,但 decision504 真实过去仍约 2.85/2.63;961 decision276 真实过去仍有 2.32/4.12 m/s。说明历史累积参与部分异常,段首问题也能在真实过去下出现。低速度窗口可能只是静止,不能认作正确完成行走。 ++ ++第六轮从同一原正式 best 重新开始,保持第五轮采样刷新1、原链长和概率、完整根解冻、学习率、stage1 速度权重0.5、250 更新及两种子固定验证,只新增 stage1_boundary_weight=3:第一阶段根速度的历史末帧→首未来帧及首→第二未来帧两条差分加权3倍。推理无平滑、无 GT,验收不变。真实表征检查确认边界/内部误差比3、精确目标零、无效尾帧忽略、有限且仅作用于根分量的梯度;收据 recovery_v6_boundary_checks.json,配置差异 recovery_v6_config_diff.json,入口 scripts/recovery_training_v6/supervise_pilot.py。监督器已启动,初始验证和两更新完整恢复检查尚待完成,不声称本轮已改善。 ++ ++新增 12 GiB 预算已实际写入/fsync 并验证物理分配,保留4 GiB原子 checkpoint 上限、8小时墙钟及原资源门槛。本地约1.54 GiB空闲、共享约34,750空闲inode,小文件写入/fsync通过。已清理 v5 Kimodo 125 落败快照及结束诊断后的重复索引共1,407.06 MiB;无进程句柄和必要配置引用,缓存保留副本 SHA 一致。保留 v5 最终完整断点、v4 125候选、所有轨迹与活跃 HY 预留。诊断预留已释放。清单 recovery_cleanup_hourly_20261109T14.json;完整分析 recovery_hourly_20261109T14_analysis.json;启动收据 recovery_v6_launch.json。下一检查点是 HY 最终结果、Kimodo v6恢复收据与固定验证,而非直接延长未通过配方。 ++ ++ ++### 第七轮监督权重对照(2026-11-09 15:06 UTC) ++ ++当前两模型均进入 `/root/autodl-fs/pangxinlei/models/hfd-recovery-pilot-v7-20261109/{hymotion,kimodo}`,入口 `HFD_bench/scripts/recovery_training_v7/supervise_pilot.py`。上一轮 HY v5、Kimodo v6 均正常完成250更新,恢复检查通过、训练数值有限,未通过闭环验收;无正式训练,继续小时调度。任一正式训练成功启动并验证健康后立即取消心跳的要求保持。 ++ ++HY v5 最终幅度比0.264、接段比1.666、姿态误差0.429 rad、根位置误差0.487 m、根jerk285.94 m/s³。行走案例2327两种子进度0.279/0.289、961为-0.008/0.004,明显不合格。Kimodo v6 最终幅度比1.018、接段比2.283、姿态误差0.412 rad、根位置误差0.362 m、根jerk330.86 m/s³;虽然jerk低于原初始357.30,进度仍不可靠:2327为0.114/0.510,961为0.885/0.285。两个模型都已读取初始/125/250全部逐例完整轨迹,保存位置/姿态/速度误差、幅度、接段/段内突变、jerk与路径;中途候选也未通过,不用平均改善掩盖逐例失败,未计算新的正式FID。 ++ ++第七轮每模型仅改一个参数,保持原正式best初始化、固定训练种子、refresh1会话采样、原链长/概率、学习率、C6标签、推理及原验收不变,仍250更新、0/125/250固定验证。HY根速度辅助权重1→0.1,旋转速度权重0.05和warmup200保持;依据是v5后期根速度项数值均值约0.102,大于flow约0.071,同时幅度和姿态退化,这只是权重失衡假设,不能将损失数值比较当成梯度主导的证明。Kimodo closed_loop.tail_weight由0.25→1,完整监督原有60帧预测,不再将后48帧降权;检验局部接段改善却仍走不够是否与短时域拟合有关。执行仍只用前12帧,stage1_boundary_weight=3保留,不加推理平滑或未来GT。 ++ ++真实函数检查确认尾段权重1时原有效性掩码、历史条件和未定义末帧导数掩码完全保留、输入不被修改;HY调度只改变根速度系数。检查、配置差异、源码SHA与启动收据为 artifacts/recovery_v7_weighting_checks.json、recovery_v7_config_diff.json、recovery_v7_launch.json。两监督器真实进程已启动,先初始验证和两更新完整保存/严格恢复,本轮恢复检查尚未结束,不能声称训练效果已改善。本轮选参只用train/validation。 ++ ++已重新实际写入/fsync HY2 GiB与Kimodo12 GiB预算并验证物理分配,原子替换上限256 MiB/4 GiB、8小时墙钟、目录项/缓存/日志及空间inode护栏均保持。本地约1.54 GiB空闲,共享约34,690空闲inode,小文件实写/fsync通过,未把共享挂载总空闲视为账号配额。 ++ ++清理约4.98 GiB:已结束轮次的重复启动索引、HY v5及Kimodo v6的125步落败快照,以及不再需要的Kimodo v3最终完整断点。删除前核对运行结束、无打开句柄和配置输入依赖,缓存保留副本哈希一致。v3的配置/源码/完整原始轨迹/指标仍保留;保留当前HY v5、Kimodo v6完整断点、较好Kimodo v5完整候选、v4的125候选和监督审计所需旧断点,不无限累积落败权重。清单 recovery_cleanup_hourly_20261109T15.json,完整分析 recovery_hourly_20261109T15_analysis.json。下次先检查v7恢复与资源健康,任一完成即分析逐例,不机械延长失败配方。 ++ ++ ++### 小时检查与 Kimodo 原生初始化对照(2026-11-09 16 UTC) ++ ++HY v7仍运行,检查时194/250更新,数值有限、两步完整恢复检查通过、无错误。125验证幅度比0.269、姿态误差0.443 rad、根位置误差0.464 m,不合格,继续原定有界运行。 ++ ++Kimodo v7正常完成250更新、完整恢复通过,已读取初始/125/250全部逐例轨迹。最终幅度比0.770、接段比2.394、姿态误差0.422 rad、根位置误差0.376 m;原正式初始分别为幅度0.405、姿态0.468 rad、根位置0.442 m。行走案例2327两种子进度0.513/0.896、961为0.911/0.346;最后一个未达到0.5。根jerk511.39 m/s³高于初始357.30和v6最终330.86;段内姿态变化p95均值0.07275 rad,GT为0.05927。125根误差0.496 m、jerk643.45亦不合格。不能因平均误差改善晋升正式训练,也未计算新FID。 ++ ++第八轮仅Kimodo:`/root/autodl-fs/pangxinlei/models/hfd-recovery-pilot-v8-20261110/kimodo`,入口`HFD_bench/scripts/recovery_training_v8/supervise_pilot.py`。目录日期采用当前会话日期,检查收据时间来自机器UTC。只把initial_checkpoint从旧HFD正式best改为None,由上游配置加载原生Kimodo-SMPLX-RP-v1/model.safetensors。CPU逐张量精确核验408个张量、282790161个元素与原文件完全一致,非随机初始化。其余学习率、完整根训练、250更新、采样、C6标签、因果输入、推理及0/125/250两种子验证不变。复制原HFD初始baseline-validation.json且逐字节核对一致,误差门槛仍以原正式初始为准,不能因原生初始化改变验收。 ++ ++新监督器558793、子进程558794已核实存活,处于初始验证/两更新恢复检查阶段;本轮恢复尚未通过,效果未知。收据artifacts/recovery_v8_initialization_check.json、recovery_v8_config_diff.json、recovery_v8_launch.json。新预算12 GiB已实写/fsync并验证物理分配,4 GiB原子保存上限及原空间/inode/目录项/日志/8小时保护保持。本地约1.54 GiB可用、共享约34600空闲inode,本地及共享小文件实写通过;不将共享df当账号配额。 ++ ++已清理v7 Kimodo重复索引和125落败快照1407.06 MiB,核实结束进程、无打开句柄、快照无配置输入引用,索引与v8保留副本SHA一致。保留v7最终完整恢复断点、必要旧候选、配置与全部轨迹;活跃HY预留不动。清单artifacts/recovery_cleanup_hourly_20261109T16.json;逐例分析recovery_hourly_20261109T16_analysis.json。 ++ ++无正式训练启动,小时心跳416e9f54继续;一旦正式训练成功启动并验证健康,立即删除心跳,不停止训练监督器。下次先检查HY最终结果和Kimodo v8完整恢复、初始及训练后闭环验证。不把本次原生初始化假设当成已证实改善。 ++ ++ ++### 第九轮有界对照(2026-11-09 17 UTC) ++ ++HY v7、Kimodo v8均完成250更新,完整两更新保存/恢复通过,训练及已读取0/125/250全部逐例轨迹数值有限,无错误或资源受控停止。没有正式训练;小时心跳416e9f54保持,正式训练成功启动并核实健康后立即删除。 ++ ++HY v7降低根速度辅助权重后,最终幅度比0.272、接段比2.044、身体姿态误差0.450 rad、根位置误差0.478 m、jerk291.97 m/s³。原初始姿态0.372 rad,最终退化超过10%,且有1个幅度坍缩案例;行走2327两种子进度0.350/0.315,961约0/0.005。中途125亦未过关,不能将较低jerk解释为动作恢复。 ++ ++Kimodo v8原生初始化在当前因果适配接口下,初始幅度1.258、姿态0.603 rad、根位置0.608 m、jerk723.84;最终幅度1.545、接段比2.948、姿态0.601 rad、根位置0.569 m、jerk758.54。段内姿态变化p95均值0.10262 rad(GT0.05927);行走2327进度0.014/-0.019,961为-0.029/-0.036。125也不合格。原生权重增加运动幅度但未恢复正确进度,不支持“换回原生权重即可解决”,也不能据此否定原生模型在其原协议下的能力;当前结果包含因果表征、历史输入及任务域的适配影响。未计算新FID。 ++ ++下一轮根目录`/root/autodl-fs/pangxinlei/models/hfd-recovery-pilot-v9-20261110/{hymotion,kimodo}`,入口`HFD_bench/scripts/recovery_training_v9/supervise_pilot.py`。HY以v7为对照,仅initial_checkpoint改为None,build_model仍加载官方latest.ckpt。CPU核对302个backbone张量及null文本参数与官方文件完全一致,新增观察列为0、LoRA B为0,非随机网络;其余损失和配置不变。Kimodo以更好的v7为对照(不是从v8续训),保持原正式best初始化,仅新增stage1_root_displacement权重0.5:对有效未来帧的第一阶段物理根位置相对最后历史帧的位移,用beta0.1的SmoothL1监督并按score掩码归一。检验短时速度监督下仍走不够的问题;不改变推理、标签、历史构造或执行长度。测试确认精确目标损失0、无效尾段与空掩码忽略、整体平移不变、有限且仅根通道梯度。此项是假设检验,不预设改善。 ++ ++两模型仍250更新、0/125/250固定validation与两个种子,train/validation选参,Test1024未使用。原HFD初始baseline-validation.json逐字节保留,误差门槛不随新初始化放宽。配置差异仅指定变量与路径/元数据,收据artifacts/recovery_v9_config_diff.json;测试recovery_v9_displacement_check.json、recovery_v9_hy_initialization_check.json;启动源码SHA和进程分别recovery_v9_hymotion_launch.json、recovery_v9_kimodo_launch.json。两监督器及子进程已启动,目前初始验证/两更新恢复检查尚未结束,不声称本轮恢复或效果已通过。 ++ ++HY2 GiB、Kimodo12 GiB预算重新实写/fsync并验证物理分配,256 MiB/4 GiB原子checkpoint上限、8小时墙钟、空间/inode/日志/文件数限制保持。本地约1.54 GiB可用、共享约34500空闲inode,另做本地/共享小文件实写;共享df不代表账号配额。清理HY v7与Kimodo v8重复启动索引及125落败快照1865.72 MiB,核实进程结束、无打开句柄、缓存与v9副本SHA一致、快照无配置输入引用。保留最终完整断点、原正式权重、少量既有对照、全部轨迹。清单artifacts/recovery_cleanup_hourly_20261109T17.json,完整分析recovery_hourly_20261109T17_analysis.json。下次先检查v9恢复收据和资源健康,完成后按原门槛分析逐例。 ++ ++ ++### 小时检查与 Kimodo 第十轮(2026-11-09 18 UTC) ++ ++HY v9仍运行,检查时195→198/250更新,真实进程560419存活,GPU0约20.6 GiB,loss和已记录梯度有限,完整两步恢复通过,无错误或资源停止。125验证幅度1.952、接段比5.406、姿态误差0.925 rad、根位置0.717 m,明显未通过。保留当前有界实验等待最终验证,不重复启动或占用其GPU。 ++ ++Kimodo v9完成250,完整恢复检查通过,无非有限值,实际训练覆盖25会话,记录中预测历史槽位占比0.835(设定抽样概率0.75;长链会影响实际槽位占比)。已读取初始/125/250全部逐例完整轨迹及指标。最终幅度0.730、接段2.450、姿态0.404 rad、根位置0.396 m、jerk353.50 m/s³;与v7的0.422 rad、0.376 m、511.39相比,姿态与jerk改善,根位置略差但仍优于原初始0.442 m。段内姿态p95均值0.06497 rad,GT0.05927。没有幅度坍缩,但2327两种子进度0.332/1.090、961为0.517/0.261,仍有两个低于0.5。125根误差0.497 m、jerk642.08、961种子2026进度0.272,亦不合格。不能据此正式开训,不将诊断指标当FID。 ++ ++第十轮仅Kimodo,目录`/root/autodl-fs/pangxinlei/models/hfd-recovery-pilot-v10-20261110/kimodo`,入口`HFD_bench/scripts/recovery_training_v10/supervise_pilot.py`。以v9为对照,仅pilot.predicted_probability由0.75降至0.25;假设长预测历史反馈误差限制正确运动监督,增加真实过去训练占比进行检验。保持原正式best初始化、位移权重0.5、所有损失、学习率、采样刷新、链长、C6标签和250更新。运行源码除准备器外逐字节一致;推理完全闭环、无未来GT,不把真实过去诊断当正式成绩。原初始baseline-validation.json逐字节保留,验收与0/125/250两种子协议不变,Test1024未使用。配置核验artifacts/recovery_v10_config_diff.json,启动进程和源码SHA在recovery_v10_launch.json。 ++ ++12 GiB预算已重新实写/fsync且物理分配通过,保留4 GiB原子checkpoint上限、8小时墙钟、空间/inode/文件数/日志保护。监督器561515及子进程已核实启动,处于初始验证和两步恢复检查,尚无本轮改善结论。本地约1.54 GiB可用、共享约34480空闲inode,本地与共享小文件实写通过;不把共享df当账号配额。活跃HY预留不动。 ++ ++清理4.535 GiB:v9 Kimodo重复启动索引、125落败快照,以及已分析且无后续用途的v8 Kimodo原生初始化最终权重。删除前核实完成收据、进程结束、无打开句柄、无配置输入引用,缓存与v10副本SHA一致。保留v9当前完整断点、较优既有对照、原正式权重,以及v8配置/全部轨迹/指标/原生权重核验;不再无限积累落败完整权重。清单artifacts/recovery_cleanup_hourly_20261109T18.json,完整分析recovery_hourly_20261109T18_analysis.json。 ++ ++当前无正式训练,心跳416e9f54继续。下次先检查HY最终结果与Kimodo v10恢复收据、资源和逐例验证;任何正式训练健康启动后立即取消心跳,不终止监督器。 ++ ++ ++### 小时检查与固定权重推理诊断 v11(2026-11-09 19 UTC) ++ ++HY v9及Kimodo v10均正常完成250更新,完整两步恢复通过,训练loss/已记录梯度与0/125/250全部读取的完整轨迹有限,无错误、无资源停止、预留已释放。均未通过固定门槛,没有正式训练,小时心跳416e9f54保持。 ++ ++HY v9原生初始化最终幅度1.888、接段5.423、姿态0.882 rad、根位置0.710 m、jerk2306.62 m/s³;原生初始分别为幅度1.694、接段5.710、姿态1.061 rad、根位置0.811 m、jerk2498.46。训练有局部改善但仍远差于原HFD正式初始姿态0.372/根位置0.449,验收不随初始化放宽。行走2327进度0.024/0.116、961为-0.173/-0.014;无幅度坍缩不代表正确运动。段内姿态p95均值0.11329 rad,GT0.05927。125也不合格。 ++ ++Kimodo v10降低预测历史抽样后,最终幅度0.723、接段2.020、姿态0.460 rad、根位置0.429 m、jerk408.32;125分别0.616、2.144、0.466 rad、0.431 m、329.05。最终行走2327进度1.097/0.882,但961为-0.015/0.285;较v9的961进度0.517/0.261,第一种子退化。因此不接受“增加真实历史一定改善”的假设,也不凭更平滑中途结果正式开训。没有新的正式FID。 ++ ++连续单变量训练配方未带来稳定逐例通过,本轮先做有界推理诊断,暂不再追加新的损失权重或延长失败训练。目录`/root/autodl-fs/pangxinlei/models/hfd-recovery-inference-v11-20261110/{hymotion,kimodo}`,入口`HFD_bench/scripts/recovery_inference_v11/supervise_pilot.py`。HY固定原HFD正式best,比较10/50采样步;Kimodo固定v9最终完整权重,比较20/100步。各自CFG、相同6个validation案例、两个种子、因果历史、坐标转换、执行长度和原验收不变;不使用Test1024。先跑原步数对照用于复现已有结果,然后跑更细采样。先前Kimodo单窗口20/100步诊断不能替代这次完整闭环比较。 ++ ++诊断分支在优化器创建前返回,零训练更新、不保存训练断点,不计作正式训练或恢复续训;无需伪造两步恢复检查。复用原监督器空间/inode/目录项保护、日志轮转和受控终止,单次墙钟上限4小时,完成后释放预留。静态检查artifacts/recovery_v11_static_check.json,固定输入路径和SHA及基线一致性recovery_v11_inputs.json,启动源码SHA/进程recovery_v11_launch.json。输出为logs/diagnostic.jsonl及validation-traces/step-000010或000050等(此处编号是采样步数,不是训练更新);完成收据标记training_updates=0。状态脚本last_validation可能为空,应读取diagnostic.jsonl。 ++ ++两监督器562275/562278及子进程562279/562280已核实存活,无启动错误;结果尚待完成。每模型2 GiB预算已实写/fsync且验证物理分配,因无新训练checkpoint只预算有界缓存、日志和两套轨迹。本地约1.54 GiB可用、共享约34400空闲inode,本地/共享小文件实写通过,不将共享df当配额。 ++ ++清理1.822 GiB:HY v9和Kimodo v10已结束轮次的重复启动索引、125不合格快照;删除前核验进程结束、无打开句柄、快照无配置引用、缓存与诊断副本SHA一致。保留最终完整断点、原正式权重、诊断所用Kimodo v9完整权重、其他必要对照及全部原始轨迹。清单artifacts/recovery_cleanup_hourly_20261109T19.json;完整分析recovery_hourly_20261109T19_analysis.json。下次检查诊断完成/错误、原步数复现及增加步数的逐例影响,再据此决定训练改动;不将诊断直接当正式训练资格。任一正式训练健康启动后立即取消心跳的条件保持。 ++ ++ ++### 小时检查、HY精度修复与Kimodo v12(2026-11-09 20 UTC) ++ ++v11两模型诊断均正常完成,零训练更新,无错误/资源停止,预留释放。读取全部两套完整逐例轨迹,数值有限,输入权重SHA与启动前完全一致。没有正式训练,心跳416e9f54继续。 ++ ++Kimodo20步与v9最终全部预测及参考轨迹逐元素相同,确认复现。100步幅度1.004、接段2.255、姿态0.451 rad、根位置0.434 m、jerk542.36 m/s³;20步分别0.730、2.450、0.404 rad、0.396 m、353.50。100步行走2327进度0.670/0.863,961为-0.003/0.487,未达标且jerk升高,不能用更多采样步替代训练修复,不采用100步为新默认。 ++ ++HY v11发现诊断入口缺少原训练验证的BF16 autocast上下文。10步原协议复现失败,预测轨迹最大元素绝对差0.5620(混合位置/旋转通道,不作为单一物理量解释);其jerk255.71与原基线335.19不同。v11的10/50步结果仍保留作FP32诊断证据,但不能用于判断原BF16协议增加步数的收益,也不能与既有BF16指标混合统计。原因是新增诊断分支只包torch.no_grad,原validate同时包torch.autocast;旧训练结果未受该新增分支影响。 ++ ++HY v12复制独立源码到scripts/recovery_inference_v12,添加与原validate一致的BF16 autocast和validation seed;目录`/root/autodl-fs/pangxinlei/models/hfd-recovery-inference-v12-20261110/hymotion`。仍固定原正式best、10/50步、同一6案例与两个种子,零更新,4小时监督上限,原误差基线不变。监督器563273、子进程563276已核实存活。下一轮必须先验证10步复现;若仍不一致,继续排查文本缓存/随机上下文,不先宣布改善。诊断输出logs/diagnostic.jsonl、validation-traces,编号是采样步数。 ++ ++Kimodo v12以v9为基线,仅预测历史概率0.75→1.0。v10增加真实历史后961第一种子进度从v9的0.517退到-0.015,因此测试增加闭环暴露,而非延长失败预算。目录`/root/autodl-fs/pangxinlei/models/hfd-recovery-pilot-v12-20261110/kimodo`,入口scripts/recovery_training_v12/supervise_pilot.py。保持原正式best初始化、位移监督0.5、所有其他损失/学习率/采样刷新/链长/推理、250更新和0/125/250固定两种子验证;不把v10结果作为初始化,不改验收,不用Test1024。除准备器外训练与推理源码和v9逐字节一致,原baseline-validation.json逐字节一致。监督器563316及训练子进程已启动,目前初始验证/两更新完整恢复检查阶段,恢复检查尚未完成、效果未知。 ++ ++HY2 GiB、Kimodo12 GiB预算重新实写/fsync并确认物理分配,保留原资源保护;本地约1.54 GiB可用、共享约34340空闲inode,本地与共享写入探针通过,不把df总量当账号配额。清理v11已完成诊断的重复索引656.24 MiB,核对进程结束、无打开句柄和v12保留副本SHA一致,保留所有诊断轨迹/指标/配置与正式及必要候选权重。清单artifacts/recovery_cleanup_hourly_20261109T20.json;完整分析recovery_hourly_20261109T20_analysis.json;配置/静态核验recovery_v12_config_diff.json、recovery_v12_static_check.json;两启动收据recovery_v12_hymotion_launch.json、recovery_v12_kimodo_launch.json。任一正式训练健康启动后立即删除心跳的要求保持。 ++ ++ ++### 小时检查与文本条件诊断 v13(2026-11-09 21 UTC) ++ ++HY v12固定权重诊断和Kimodo v12训练均正常完成,预留已释放,无错误/资源停止。已读取所有完整轨迹且有限;Kimodo训练loss/记录梯度有限、完整两更新恢复通过。无正式训练,心跳416e9f54继续。 ++ ++HY修复BF16上下文后,10步所有预测/参考轨迹与原正式初始逐元素完全一致,确认复现问题解决。50步幅度0.489、接段1.324、姿态0.404 rad、根位置0.498 m、jerk304.67;10步分别0.291、1.290、0.372 rad、0.449 m、335.19。50步幅度仍低于0.5,根位置超过原初始10%上限,行走2327进度1.188/1.869、961为-0.222/-0.283,仍失败。不放宽门槛、不采用50步作为已通过方案。 ++ ++Kimodo v12全预测历史250更新最终幅度1.100、接段2.386、姿态0.439 rad、根位置0.441 m、jerk405.63。行走2327进度0.827/0.678,961为0.107/0.510,仍有一个失败;125时961为1.108/0.390,jerk890.09。中途和最终对同一案例不同种子不稳定,不能以均值改善晋升正式训练,也不机械延长250预算。所有诊断都不是新FID。 ++ ++读取既有validation历史诊断的实际因果文本:961在decision276/432是“turns left Hmm... walks forward”;2327在decision396/504是“interchanging an item between hands stop jog forward left diagonally”。Source.condition按已发布时钟取完整30秒参考转录上下文,未用GT turn或控制标签选新指令。文本中旧/新动作混合是可检验假设,尚不能断言根因或擅自丢弃旧指令。 ++ ++v13固定权重原文本/空文本对照已启动,目录`/root/autodl-fs/pangxinlei/models/hfd-recovery-text-v13-20261110/{hymotion,kimodo}`,入口scripts/recovery_text_v13/supervise_pilot.py。HY用原正式best、10步、BF16;Kimodo用v12最终完整权重、20步;同一6个validation案例、两种子、CFG、因果历史、原始验收和其余推理保持。两臂只在text.encode入口把已发布文本保留或替换为空字符串;空文本是编码空转录,并非假定Kimodo得到严格CFG无条件向量。没有未来文本、GT动作/标签选择,也不改变训练或评测门槛。原文本先跑以检验复现。该诊断测量文本敏感性,空文本即使分数变好也不能当任务成功或正式训练资格。 ++ ++零训练更新、无优化器更新、无新训练checkpoint,4小时监督上限;输出logs/diagnostic.jsonl的arm字段为original/empty,validation-traces/step-000000与step-000001对应两臂,不是训练更新。静态检查artifacts/recovery_v13_static_check.json,输入权重SHA/步数recovery_v13_inputs.json,源码SHA/进程recovery_v13_launch.json。监督器564572/564576及子进程已核实存活,尚无结果结论。 ++ ++每模型2 GiB预算实写/fsync验证通过,保持磁盘/inode/目录项/日志保护;本地约1.54 GiB可用、共享约34240空闲inode,写入探针通过,不将df当账号配额。清理4.856 GiB:HY v12诊断与Kimodo v12训练的重复索引、Kimodo v12不合格125快照,以及无后续用途的Kimodo v10最终权重。核对结束收据、进程已退出、无打开句柄、无配置输入引用,缓存与新副本哈希一致。保留当前完整断点、必要对照/正式权重及全部轨迹/配置/指标。清单artifacts/recovery_cleanup_hourly_20261109T21.json,完整分析recovery_hourly_20261109T21_analysis.json。 ++ ++下一轮先检验原文本组复现,再读取两臂完整逐例轨迹判断文本敏感性,结合行走进度/幅度/接段/jerk而非单一平均误差确定后续训练。未达到正式条件,不提前取消心跳;任一正式训练健康启动后立即取消的要求保持。 ++ ++ ++### 小时检查与引导强度诊断 v14(2026-11-09 22 UTC) ++ ++v13两模型文本对照正常完成,零更新,无错误/资源停止,预留释放。已读取全部两臂完整逐例轨迹并确认有限;original组与既有对应权重结果逐元素完全一致。原文本对照可复现,空文本效果不是入口精度差异导致。无正式训练,心跳416e9f54保持。 ++ ++HY原文本幅度0.291、姿态0.372 rad、根位置0.449 m、jerk335.19;空文本为0.269、0.387 rad、0.530 m、334.10。行走2327从0.483/0.680降至0.173/0.056;961仍近零。Kimodo原文本幅度1.100、姿态0.439 rad、根位置0.441 m、jerk405.63;空文本为1.176、0.439 rad、0.526 m、56.93。空文本行走2327仅0.067/0.035,961为-0.022/-0.030,原文本分别0.827/0.678和0.107/0.510。空文本低jerk来自缺少正确行走,不能作为改善或替代输入。两臂都未通过门槛;不能据此断言历史文本冲突是唯一根因,也不能宣称模型忽略文本。未测新FID。 ++ ++下一有界诊断v14只改变引导强度,目录`/root/autodl-fs/pangxinlei/models/hfd-recovery-guidance-v14-20261110/{hymotion,kimodo}`,入口scripts/recovery_guidance_v14/supervise_pilot.py。保持v13相同固定权重、完整因果文本、同一6个validation案例/两个种子、HY10步BF16与Kimodo20步、历史/坐标/执行长度及原门槛。HY比较guidance1→2,检验增强条件响应;Kimodo比较2→1,检验文本引导是否放大根运动抖动。假设未证实,不以分数改善预设正确;仍须检查逐例进度与幅度。无未来文本/GT选指令,Test1024未使用。 ++ ++两臂original/adjusted_guidance写入logs/diagnostic.jsonl,guidance字段记实际数值;validation-traces/step-000000和000001分别对应两臂,非训练步。零优化更新、不保存训练断点,4小时监督上限与资源受控停止保留。输入权重SHA与步数在artifacts/recovery_v14_inputs.json;源码SHA与进程recovery_v14_launch.json;静态检查recovery_v14_static_check.json。监督器565596/565599及子进程已核实存活,尚无本轮结果。 ++ ++各2 GiB预算重新实写/fsync并验证物理分配,本地约1.54 GiB空闲、共享约34170空闲inode,探针写入通过;不将df当账号配额。清理已完成v13的重复索引656.24 MiB,确认旧进程结束、无打开句柄、与v14副本SHA一致,保留全部轨迹/指标/源码/配置及正式与必要候选权重。清单artifacts/recovery_cleanup_hourly_20261109T22.json;完整分析recovery_hourly_20261109T22_analysis.json。下次先检验original组复现,再分析调整引导强度的逐例效果,为后续恢复训练选有证据的设置;没有通过全部门槛不正式开训。正式训练健康启动后立即取消小时任务的条件保持。 ++ ++ ++### 第十五轮因果文本适配恢复(2026-11-09 23 UTC) ++ ++v14两模型引导诊断正常完成,无错误/资源停止,预留释放;全部两臂完整轨迹已读取且有限,original组逐元素复现既有结果。HY guidance2幅度0.341、接段1.530、姿态0.379 rad、根位置0.575 m、jerk529.91;行走2327进度2.108/2.568过冲,961仍为-0.024/-0.008。Kimodo guidance1幅度1.095、接段2.412、姿态0.429 rad、根位置0.475 m、jerk143.44,但2327进度0.120/0.366、961为0.075/0.206。两者均未通过,低jerk不能代替任务运动。不采用这些推理设置,无正式训练。 ++ ++恢复实验v15目录`/root/autodl-fs/pangxinlei/models/hfd-recovery-pilot-v15-20261110/{hymotion,kimodo}`,入口scripts/recovery_training_v15/supervise_pilot.py。HY以v7、Kimodo以v12为对照,均从原正式best重新初始化,不从诊断输出续训;保持各自损失/学习率/预测历史概率/250更新/0、125、250两种子验证/原推理设置,仅加入训练和推理共享的因果文本适配。原baseline-validation.json逐字节保持,原误差门槛不因输入变换放宽。 ++ ++适配器causal_text.py只读取截至当前发布时刻的文本前缀和发布时间。新文本为旧文本词前缀的延续时,距上次新增词不足1.2秒则继续当前片段,否则从新增加的词开始;不变文本保持当前片段,空上下文清空,修订或截断无法可靠分辨时保守保留当前已发布全文。无动作、turn ID、控制标签、GT边界或未来词。是新的因果输入预处理实验,不能把结果标为“原完整转录不变”的协议,也不能据其直接晋升正式训练。 ++ ++单元验证前缀不变性、追加未来不影响过去、空上下文过期、修订/截断保守处理、非法时间拒绝及确定性恢复。使用禁止未来索引的序列测试真实Source.condition绑定,确认只访问已发布前缀且发布时间不变。真实输入核验先发现按需provider用SimpleNamespace构造Source、绕过__init__,已在启动前改为共享Source.condition接口;未启动无效训练。修正后961 decision276/432从“turns left Hmm... walks forward”变为“walks forward”,120为“Hmm...”;2327仍可能保留“stop jog forward left diagonally”,此规则不保证解决全部歧义,不预设改善。收据artifacts/recovery_v15_causal_text_checks.json、recovery_v15_actual_text_audit.json、recovery_v15_config_diff.json。只用train/validation,Test1024留出。 ++ ++HY2 GiB、Kimodo12 GiB预算重新实写/fsync并验证物理分配;原子checkpoint上限256 MiB/4 GiB、8小时墙钟、磁盘/inode/目录项/日志和受控停止保护保持。本地约1.54 GiB空闲、共享约34130空闲inode,探针写入通过,不把df当配额。两监督器567139/567143与子进程已核实启动,先初始验证和完整两更新保存/严格恢复;当前恢复检查未完成、效果未知。源码SHA和启动收据artifacts/recovery_v15_launch.json。 ++ ++清理已完成v14重复索引656.24 MiB,确认旧进程退出、无打开句柄和新副本SHA一致,正式/必要候选/完整恢复断点、全部轨迹和配置指标保留。清单artifacts/recovery_cleanup_hourly_20261109T23.json;完整分析recovery_hourly_20261109T23_analysis.json。下一检查点为两模型恢复收据与实际训练文本、0/125/250闭环逐例改善,并同时关注旧指令被过早丢弃的失败案例。无正式训练,心跳416e9f54继续;任一正式训练健康启动后立即删除心跳的条件保持。 ++ ++ ++### 小时检查与固定权重文本复核 v16(2026-11-10 00 UTC) ++ ++HY v15仍运行,00:05核验194/250步,进程567475正常推进,已记录loss/梯度有限、完整两更新恢复通过,无错误或资源停止。125步幅度0.258、接段2.068、姿态0.447 rad、根位置0.512 m,行走2327进度0.111/0.032、961为-0.052/-0.015,仍不合格;等待有界250步最终验证,不重复启动。 ++ ++Kimodo v15已完成250步、完整恢复通过、无错误,预留已释放。读取初始/125/250全部完整轨迹并确认有限。最终幅度1.193、接段2.259、姿态0.465 rad、根位置0.685 m、jerk501.95;2327进度0.679/0.537,961为-0.938/-1.307,明显反向行走。相比v12最终根位置0.441 m、961进度0.107/0.510退化。v15第0步961已为-1.171/-1.180(根位置均值0.739 m),125步进一步到-2.071/-2.205(根位置0.996 m);因此退化并非全由训练更新造成,截短上下文可能丢失方向信息,但尚不能确认唯一根因。不延长此配方、不进入正式训练;这些结果不是FID。 ++ ++Kimodo v16仅固定v15最终权重,比较adapted与full_context两臂,原验证6案例/两种子、20步CFG2、全部其他输入/推理/门槛保持。目录`/root/autodl-fs/pangxinlei/models/hfd-recovery-text-v16-20261110/kimodo`,入口scripts/recovery_text_v16/supervise_pilot.py。两臂依次写logs/diagnostic.jsonl,轨迹step-000000/000001是诊断臂而非训练步;0优化更新,进入优化器构造前返回。先复现v15最终适配文本,再恢复完整已发布因果文本,区分输入变换和训练后权重效应;不使用未来词或GT选指令,Test1024未使用。通过共享模块开关切换已绑定Source.condition,验证切换/还原与发布时间保持;避免运行入口与模型入口各自加载配置导致开关无效。源码/静态核验artifacts/recovery_v16_static_check.json,启动收据recovery_v16_launch.json。监督器568479/子进程568481已确认存活,诊断结果尚未完成。 ++ ++v16重新实写/fsync预留2GiB,4小时监督上限,磁盘/inode/日志/目录项保护保留。本地约1.539GiB可用,共享34052空闲inode,本地与共享1MiB写入/fsync探针通过;不把共享df总量当账号配额。清理已结束Kimodo v15重复索引和失败125快照,共1407.06MiB:确认训练/监督器已退出、无打开句柄、缓存与v16副本SHA一致、中间权重无配置输入引用。保留v15最终完整断点、全部配置/指标/轨迹及必要历史/正式权重,不触碰仍运行HY资源。清单artifacts/recovery_cleanup_hourly_20261110T00.json;完整轨迹/健康分析recovery_hourly_20261110T00_analysis.json。 ++ ++下一检查点为HY v15最终结果与Kimodo v16原适配组复现、完整文本恢复后的逐例方向/进度/幅度/接段/jerk。无正式训练,小时任务416e9f54继续;任一正式训练健康启动后立即删除心跳,不等待正式训练结束。 ++ ++ ++### 第十七轮检查与恢复实验(2026-11-10 01 UTC) ++ ++HY v15完成250更新,完整两更新恢复通过、无错误/资源停止,预留释放。初始/125/最终全部完整轨迹已读取且有限,记录loss/梯度有限。最终幅度0.247、接段2.003、姿态0.452 rad、根位置0.524 m、jerk279.65;行走2327进度0.073/-0.018、961为-0.038/0.005。相对原正式初始幅度0.291、姿态0.372、根位置0.449退化,低jerk不代表恢复。原门槛不变,不正式开训。 ++ ++Kimodo v16两臂正常完成,零训练更新、预留释放。已读取全部轨迹,适配文本组预测/参考逐元素复现v15最终。完整文本组幅度1.212、接段2.311、姿态0.454 rad、根位置0.523 m、jerk473.87;2327进度1.052/0.669,961为-0.227/0.253。相比适配组根位置0.685和961 -0.938/-1.307改善,但仍明显未通过原门槛,且不如v12根位置0.441/961 0.107/0.510。说明截短因果上下文确有影响,恢复全文不足以解决全部问题;不将该诊断称为FID。 ++ ++HY v17固定v15最终权重进行adapted/full_context两臂验证,目录`/root/autodl-fs/pangxinlei/models/hfd-recovery-text-v17-20261110/hymotion`,入口scripts/recovery_text_v17/supervise_pilot.py。同一6案例/两种子、10步BF16、CFG1与原门槛保持;沿用v16已验证共享模块开关切换Source.condition,原发布时间不变,无未来词或GT选指令。0更新、优化器构造前返回、4小时上限;先检验适配组复现再分析全文效果,不默认可恢复。 ++ ++Kimodo v17回到v12完整文本配方及原正式best初始化,唯一算法变量stage1_root_displacement权重0.5→1.0,检验加强相对最后观测帧的未来根位移监督能否减少v12的行走进度不足。不能据单一失败案例宣称必然有效;监督只用训练目标,推理不输入未来GT。其余源码与v12逐字节一致、训练/推理设置和原baseline-validation.json保持。目录`/root/autodl-fs/pangxinlei/models/hfd-recovery-pilot-v17-20261110/kimodo`,入口scripts/recovery_training_v17/supervise_pilot.py。250更新、0/125/250固定两种子验证、先两更新完整保存/严格恢复;不延长或续训v15失败配方,不使用Test1024。当前仅启动及数据准备,恢复检查尚未完成。 ++ ++两监督器569655/569696及真实子进程均已核实存活,未发现错误;源码/配置核验artifacts/recovery_v17_checks.json,启动及源码SHA见recovery_v17_hymotion_launch.json、recovery_v17_kimodo_launch.json。HY2GiB/Kimodo12GiB实写fsync预留通过,Kimodo保留4GiB原子checkpoint上限、8小时墙钟及磁盘/inode/目录项/日志保护。本地约1.539GiB、共享约34000空闲inode,本地和共享1MiB写入/fsync探针通过;共享df不是账号配额。 ++ ++清理已结束HY v15及Kimodo v16重复索引、HY v15已分析失败125快照,共786.78MiB;核验旧训练/监督器退出、无打开句柄、重复索引与新副本SHA一致及权重无配置引用。保留最终完整断点、正式权重、必要对照和所有轨迹/指标/配置。清单artifacts/recovery_cleanup_hourly_20261110T01.json;完整审计recovery_hourly_20261110T01_analysis.json。下一检查点为HY两臂结果与Kimodo恢复收据/训练推进及逐例效果。无正式训练,心跳416e9f54继续;任一正式训练健康启动后立即取消的条件保持。 ++ ++ ++### 第十八轮扩大验证(2026-11-10 02 UTC) ++ ++HY v17固定权重文本诊断正常完成,无错误,预留释放;全部轨迹有限,adapted组逐元素复现v15最终。full_context幅度0.272、接段2.003、姿态0.444 rad、根位置0.459 m、jerk290.00,行走2327为0.454/0.449、961为-0.018/-0.007;无坍缩但幅度仍不足,姿态超过原初始10%上限。截短组仍有1197/2026坍缩。恢复全文有改善却未解决恢复问题,不再重复截短配方;本轮先集中验证Kimodo的新增积极证据。 ++ ++Kimodo v17正常完成250更新,完整2步恢复通过,所有记录loss/梯度有限、全部完整轨迹已读取有限,无资源停止,预留释放。125步幅度1.182、接段2.229、姿态0.463 rad、根位置0.386 m;2327进度0.924/1.035、961为1.464/0.541,无坍缩,首次此轮通过原小验证全部门槛。250步幅度1.187、接段2.317、姿态0.456、根位置0.369,但961第二种子进度0.325,最终不合格;不机械延长,不丢掉中途候选。125步jerk787.14、250步562.89,相比原正式初始357.30偏高,必须排除抖动收益,不能以门槛通过直接正式开训。原正式基线姿态0.468/根位置0.442,各阶段对照和逐例证据见artifacts/recovery_hourly_20261110T02_analysis.json;这些不是FID。 ++ ++v18是冻结权重的扩大validation,零更新。目录`/root/autodl-fs/pangxinlei/models/hfd-recovery-broad-v18-20261110/{baseline,candidate125}`,入口scripts/recovery_broad_v18/supervise_pilot.py。baseline用原正式best,candidate125用v17 weights/step-000125.pt。每交互模式确定性选择4个6–20秒validation案例,共24个、包含原6个、固定种子12345/2026;仅依据模式/长度/可用性选样,不能依据预测分数选样,无Test1024。两组均保持完整因果文本、20步CFG2、同一初始状态及其他推理协议。选择器验证确定性、各模式4例、原6例保留;源码/检查recovery_v18_checks.json,源码SHA/启动recovery_v18_launch.json。两组监督器570911/570912、子进程570913/570914已核实存活,当前数据准备,结果未知。 ++ ++每组logs/diagnostic.jsonl及validation-traces/step-000000.npz为单个冻结权重验证,不是训练步。日志内门槛暂以原六例baseline计算,不能据此自动晋升;最终必须用同一扩展案例原正式baseline重新计算非退化门槛,检查逐例进度、幅度、接段/段内突变、速度误差、根jerk及失败轨迹。即使小验证125合格也不绕过更广验证和正式容量/恢复配置要求。 ++ ++两组各2GiB预算实写/fsync通过,4小时监督上限、日志轮转/目录项/资源停止保持,扩大轨迹单文件上限128MiB以覆盖48条轨迹且纳入预算。本地约1.539GiB、共享33893空闲inode,本地和共享1MiB写入/fsync通过,共享df不视为账号配额。清理已结束Kimodo v17重复索引328.12MiB,核验旧训练/监督器退出、无打开句柄、与新副本SHA一致。保留125候选、best、最终完整恢复断点及全部证据;HY诊断缓存暂留待后续。清单recovery_cleanup_hourly_20261110T02.json。无正式训练,心跳416e9f54保持,下一检查03 UTC;任一正式训练健康启动后立即取消,不等待训练结束。 ++ ++ ++### 扩大验证故障修复与候选重试(2026-11-10 03 UTC) ++ ++v18 baseline已完成24案例×2种子,零训练更新;读取48条完整轨迹均有限。幅度0.446、接段1.995、姿态0.453 rad、根位置0.632 m、jerk587.06;24个GT明显位移的案例×种子组合中17个进度未达0.5–2。扩大集不能直接与原六例均值比较,候选必须和这份同集基线比较。 ++ ++candidate125没有完成,因CUDA OOM失败,不能将缺失结果解释成模型效果差。根因是v18并行主进程使用不同pilot.physical_gpu,但文本编码器FrozenTextWorker通过config.gpu独立指定物理GPU,两份配置均沿用gpu=1,导致两个约14.75GiB文本工作进程挤占GPU1。已保留phase-error/supervisor-error/日志和原源码配置;两个旧主进程与监督器均已退出,GPU任务检查为空,预留已释放。baseline已经正常完成,保留不重复计算。 ++ ++仅重跑候选,目录`/root/autodl-fs/pangxinlei/models/hfd-recovery-broad-v18-20261110/candidate125-retry`,入口scripts/recovery_broad_v18r/supervise_pilot.py。新冻结源码在run_pilot启动前强制config.gpu==pilot.physical_gpu,明确主进程和文本工作进程都用GPU1;本轮只运行这一组。权重仍v17 step-000125.pt、24案例选择规则/两个种子/全文/20步CFG2及全部模型推理设置不变,未改训练或放宽门槛,无Test1024。原运行与源码不覆盖。监督器571978、子进程571982已存活并正常准备数据;当前尚未完成文本加载和全部验证,不虚称效果或恢复成功。 ++ ++新目录2GiB容量重新实写/fsync通过,4小时监督/磁盘/inode/日志轮转/目录项和轨迹128MiB上限保持;保留失败证据及原正式/候选/完整恢复断点。修复及预留收据artifacts/recovery_v18_retry.json,启动与源码SHA recovery_v18_retry_launch.json。清理两旧组重复索引656.24MiB,确认旧监督器/主进程退出、无打开句柄、与重试副本SHA一致;错误日志和完成基线全部轨迹保留。清单recovery_cleanup_hourly_20261110T03.json。本地约1.539GiB、共享33869空闲inode,本地和共享1MiB写入/fsync通过;不以共享df推断账号配额。 ++ ++HY无新训练/验证结果,保留v17失败诊断,不重复失败配方。完整本轮分析recovery_hourly_20261110T03_analysis.json。下一检查先核验候选重试完成和实际GPU分配,再用同集baseline重算非退化门槛及逐例进度/抖动;原六例门槛标记不能替代扩大集判断。无正式训练,小时心跳416e9f54继续,下一次04 UTC;任一正式训练健康启动后立即取消的条件不变。 ++ ++ ++### 第十九轮恢复实验(2026-11-10 04 UTC) ++ ++v18候选单独重试正常完成,零更新,无错误/资源停止、预留释放。原正式与候选24案例×2种子的完整轨迹均有限,case_keys及全部参考轨迹完全匹配,已按同集重新计算门槛。候选幅度1.021(基线0.446)、接段2.209(1.995)、姿态0.429 rad(0.453)、根位置0.598 m(0.632),无坍缩(基线1例)。明显位移24个案例×种子中,失败从17降至10;两个种子各5个失败,不能用平均改善掩盖失败。根jerk956.56 vs587.06,增约63%;1863两种子进度0.917/1.100虽合格,jerk7422/6808,相比基线4476/3907明显恶化。幅度/接段/姿态位置非退化门槛通过,逐例行走门槛不通过,且抖动增加,不批准正式训练。指标为SMPL-X诊断,不是FID。 ++ ++下一有界实验v19以v17为对照,原正式best初始化,不从125候选或最终断点延长;唯一训练变量decoded_root_acceleration监督权重0.001→0.01,保持根位移1.0、全部其他损失/数据/学习率/因果全文/推理参数不变。已有该损失约束解码实际根位置的二阶差分与相同C6目标,并非独立辅助速度通道;源码不改、加强权重检验抖动是否可控,不预设有效,须防止以静止换低jerk。仍250更新、0/125/250固定原六例两种子验证及原正式初始门槛,通过后还必须复查扩展集。无Test1024。HY无新证据,本轮不重复失败配方,保留诊断。 ++ ++目录`/root/autodl-fs/pangxinlei/models/hfd-recovery-pilot-v19-20261110/kimodo`,入口scripts/recovery_training_v19/supervise_pilot.py。增加启动前config.gpu==pilot.physical_gpu校验,两者均1,GPU空闲检查后单独启动,避免v18文本工作进程冲突。训练源码除运行器GPU断言/准备入口外与v17逐字节一致,原baseline-validation.json不变。收据artifacts/recovery_v19_checks.json,源码SHA/启动recovery_v19_launch.json。监督器572624及真实子进程已确认存活,无错误;当前初始验证/两步保存与严格恢复阶段,恢复检查尚未完成,不宣称训练改善。 ++ ++12GiB预算重新实写/fsync通过(实际预留约12.265GB),保留4GiB原子checkpoint上限、8小时墙钟、日志轮转、目录项/磁盘/inode保护。目录项32,本地约1.538GiB,共享33844空闲inode,本地/共享1MiB写入fsync探针通过;共享df不作账号配额。清理已完成候选重试的重复索引328.12MiB,核验旧监督器/子进程退出、无打开句柄、与v19新副本SHA一致;125候选、必要历史对照、正式权重、最终完整断点和全部轨迹/配置/指标保留。清单recovery_cleanup_hourly_20261110T04.json;配对验收与失败明细recovery_hourly_20261110T04_analysis.json。 ++ ++下一检查为v19恢复收据、训练推进、0/125/250逐例进度/幅度及jerk,不因平均数改善放宽门槛。尚无正式训练,心跳416e9f54继续,下次05 UTC;任一正式训练健康启动后立即取消的条件保持。 ++ ++ ++### 第二十轮候选扩大验证(2026-11-10 05 UTC) ++ ++Kimodo v19完成250更新,完整两步保存/optimizer/scheduler恢复通过,已记录loss和嵌套梯度均有限;全部0/125/250完整轨迹读取且有限,无错误/资源停止,预留释放。125步幅度1.121、接段2.187、姿态0.449 rad、根位置0.308 m、jerk795.43;行走2327进度1.044/0.910、961为1.329/0.904,无坍缩,原小验证全部门槛通过。对照v17的125步根位置0.386 m有所改善,但jerk787.14→795.43,并未证明加强加速度监督减少抖动。250步幅度0.974、接段2.436、姿态0.416、根位置0.407、jerk530.99,961进度0.338/0.349不合格,2327为0.817/0.635;无坍缩。继续保留中途候选,不机械延长最终失败配方,不以低jerk代替正确行走。上述诊断不是FID。 ++ ++v20冻结v19 weights/step-000125.pt,扩大到与v18完全相同的24个validation案例/两个固定种子,复用已完成原正式基线及v17候选对照,不重跑基线。目录`/root/autodl-fs/pangxinlei/models/hfd-recovery-broad-v20-20261110/candidate125`,入口scripts/recovery_broad_v20/supervise_pilot.py。除准备入口外所有源码与修复GPU绑定的v18r逐字节相同,选择规则、原六例包含、完整因果文本、20步CFG2及推理不变;仅输入权重换为v19的125步。0训练更新,无Test1024。输出logs/diagnostic.jsonl与validation-traces/step-000000.npz;日志原六例门槛仅为临时标记,必须基于同一扩大集baseline重新计算非退化、逐例进度和抖动后判断,不能自动晋升。 ++ ++监督器573160及真实子进程已核验存活、未发现错误,目前数据准备。启动/源码SHA artifacts/recovery_v20_launch.json,源码一致性recovery_v20_checks.json。2GiB容量重新实写/fsync通过,单独运行GPU1,config.gpu与pilot.physical_gpu一致断言保持;4小时墙钟、128MiB轨迹上限、目录项/日志轮转/磁盘inode保护保持。完整本轮分析recovery_hourly_20261110T05_analysis.json。 ++ ++清理已结束v19训练的重复索引328.12MiB,核验旧监督器/子进程退出、无打开句柄且与v20副本SHA一致。保留125候选、best、最终完整恢复断点、正式和必要历史对照及全部原始轨迹/指标/配置。清单recovery_cleanup_hourly_20261110T05.json。本地约1.538GiB可用、共享约33780空闲inode,本地和共享1MiB写入fsync通过;共享df不视为账号配额。HY暂无新结果,不重复失败实验。 ++ ++下一检查将v20候选与v18基线/上一候选逐例匹配,核实范围更广的行走改善及jerk而非只看均值。尚未正式训练,心跳416e9f54继续,下次06 UTC;任一正式训练健康启动后立即取消的条件保持。 ++ ++ ++### 第二十一轮接段与执行节奏诊断(2026-11-10 06 UTC) ++ ++v20候选扩大验证正常完成,零训练更新,无错误/资源停止,预留释放,进程退出。读取原正式基线、v17候选、v19候选三组全部完整轨迹,数值有限、case_keys及参考逐元素匹配。v19候选幅度0.991、接段2.213、姿态0.416 rad、根位置0.528 m、jerk906.78;上一候选对应1.021/2.209/0.429/0.598/956.56,原正式基线0.446/1.995/0.453/0.632/587.06。明显位移失败17→10→8(总24案例×种子),最新无坍缩、其他均值门槛通过,但8个行走失败与较基线偏高jerk仍不满足正式条件,不能仅因持续改善就晋升。不是正式FID。 ++ ++进一步读取逐帧三阶差分,将跨过每12帧重规划首个新预测帧的差分窗口计为接段附近:约24.72%样本占基线88.18%、v17候选86.32%、v19候选86.48%的根jerk平方能量。表明高频接段附近值得定位,不等于已证明唯一根因;仅继续加大损失权重缺乏充分依据。 ++ ++v21固定v19 step-000125权重,同一24validation案例、两种子,执行12帧与24帧两臂对照。目录`/root/autodl-fs/pangxinlei/models/hfd-recovery-cadence-v21-20261110/kimodo`,入口scripts/recovery_cadence_v21/supervise_pilot.py。原12帧组先跑,要求复现v20完整预测;24帧组只改重规划/执行节奏,每次使用截至该决策时刻已发布文本和模型历史,不输入未来GT、不改变权重/采样步数/CFG。通过共享runtime CONFIG切换,避免配置副本导致开关无效;逐帧覆盖/无重复遗漏静态验证通过。0优化更新、无Test1024。 ++ ++24帧组改变响应延迟、文本更新时机与随机抽样次数,是定位频繁重规划接缝的诊断,不能当作原协议训练改善或正式晋升依据。logs/diagnostic.jsonl arms为execute_12/execute_24,完整轨迹step-000000/000001为两诊断臂,非训练步。旧12帧边界指标与资格标记对24帧组不能直接使用,下一轮需按各自执行节奏重算接段指标,并同时查看行走/停步延迟,避免平滑却不响应任务。 ++ ++监督器573629及真实子进程已存活、无错误,当前数据准备。配置/覆盖检查artifacts/recovery_v21_checks.json,源码SHA/启动recovery_v21_launch.json。2GiB预算重新实写fsync验证,4小时墙钟、单独GPU1、主/文本GPU一致断言、日志轮转/目录项/磁盘inode与128MiB轨迹上限保持。本地约1.538GiB、共享约33750空闲inode,本地/共享1MiB写入fsync通过,共享df不视为账号配额。 ++ ++清理已结束v20重复索引328.12MiB,核对旧进程/监督器退出、无打开句柄及与新副本SHA一致;正式模型、完整恢复断点、必要候选及全部证据保留。清单recovery_cleanup_hourly_20261110T06.json;三组配对验收和接段能量分析recovery_hourly_20261110T06_analysis.json。HY暂无新结果,不重复失败训练。无正式训练,心跳416e9f54保持,下次07 UTC;正式训练健康启动后立即取消的规则不变。 ++ ++ ++### 第二十二轮根轨迹组件诊断(2026-11-10 07 UTC) ++ ++v21两臂正常完成,零训练更新,无错误/资源停止,预留释放、旧进程退出。全部完整轨迹已读取且有限,12帧组与v20预测/参考逐元素一致。24帧组jerk1431.62,比12帧906.78更差;根位置0.593 vs0.528 m,幅度0.847 vs0.991,行走失败均8/24,不能据旧12帧接段均值下降宣称改善,也不采用24帧协议。 ++ ++按真实执行节奏重算:以新计划首帧为接段、SO3逐关节角增量求均值后取接段/段内比,12帧均值2.329、24帧2.924(此定义与已有指标实现细节不同,单列诊断不覆盖旧指标);根jerk跨接段窗口占样本24.72%/12.31%,对应能量86.48%/88.05%。减少重规划次数并未消除接缝,可能每次接缝更大;响应延迟同时改变,尚未单独量化停止延迟,不能以此代替任务协议。不是FID。 ++ ++v22固定相同v19 step-000125权重,恢复原12帧执行及24validation案例/两种子,仅增加只读根轨迹组件采集。目录`/root/autodl-fs/pangxinlei/models/hfd-recovery-components-v22-20261110/kimodo`,入口scripts/recovery_components_v22/supervise_pilot.py。采集每次最终inverse的unnormalized原生channels0:8(含stage1根0:3、骨盆偏移5:8)与decoded root xyz,原生帧59:73;首行最后已知历史,随后12行执行,末行仅诊断的未执行预测。保存为同一validation-traces/step-000000.npz中的root_components_i,形状[decisions,14,11],不新增逐步文件。钩子返回原解码对象,不修改输入/输出、不新增随机采样,并在finally恢复inverse;完整轨迹必须先逐元素复现v20,之后组件证据才能用于判断根/骨盆抵消或突跳。组件是模型预测,未向条件加入未来GT,无Test1024。 ++ ++该实验零优化更新、原20步CFG2/完整因果文本/坐标及其他推理保持。通过语法、钩子还原、12帧保持检查,收据artifacts/recovery_v22_checks.json;源码SHA/启动recovery_v22_launch.json。监督器574134及真实子进程已核验存活、无错误,当前数据准备,尚无组件诊断结论。4小时监督上限、2GiB实写fsync预留、GPU1主/文本一致断言、目录项/磁盘inode/日志轮转与128MiB轨迹上限保留。 ++ ++清理已结束v21重复索引328.12MiB,核验旧进程/监督器退出、无打开句柄,与新副本SHA一致;全部诊断轨迹/配置/指标、正式/必要候选及完整恢复权重保留。清单recovery_cleanup_hourly_20261110T07.json,完整分析recovery_hourly_20261110T07_analysis.json。本地约1.537GiB、共享33711空闲inode,本地和共享1MiB写入fsync通过,共享df不作为账号配额。HY无新证据,不重复失败训练。 ++ ++下一检查先确认v22输出复现,再比较失败/正常案例的stage1根、骨盆偏移和解码根在接段前后变化,以证据决定训练或表征修复。无正式训练,心跳416e9f54继续,下次08 UTC;正式训练健康启动后立即取消规则保持。 ++ ++ ++### 用户要求双模型同步推进;HY v23启动(2026-11-10) ++ ++用户明确要求:两个模型必须一起推进,不能因为Kimodo出现改善就搁置HY;只有某模型确认改进、通过验证并正式训练成功启动后,才可聚焦另一个。该要求已写入artifacts/recovery_hourly_task.json的parallel_progress_requirement,后续每轮必须分别检查并推进两者。此前任一正式训练健康启动后立即取消小时任务的条件仍保留,不把pilot当正式训练。 ++ ++HY新一轮v23已在空闲GPU0启动,目录`/root/autodl-fs/pangxinlei/models/hfd-recovery-pilot-v23-20261110/hymotion`,入口scripts/recovery_training_v23/supervise_pilot.py。以完整因果文本的v7为对照,原正式best初始化、同一数据/种子/250更新/0、125、250验证/根运动监督与全部原门槛保持;唯一算法变化rotation_velocity_weight 0.05→0,检验额外旋转速度辅助目标是否限制动作幅度。只是待验证假设,不声称根因已确定;原flow损失和根导数目标保留,不靠降低验收标准。无Test1024,训练源码除准备入口与v7逐字节相同。检查recovery_v23_checks.json,启动与源码SHA recovery_v23_launch.json。 ++ ++HY监督器574500和真实子进程已确认存活,目前初始验证/两更新完整保存恢复阶段,尚不能宣称恢复检查通过或效果改善。Kimodo v22在GPU1的根组件诊断真实进程仍存活且无错误,继续推进,不中断源码/配置。两模型健康检查recovery_v23_health.json。 ++ ++HY2GiB容量重新实写/fsync预留通过,256MiB原子checkpoint上限、8小时墙钟、日志轮转/目录项/磁盘inode保护保持。本地与共享1MiB写入/fsync检查通过,本地约1.537GiB可用、共享约33660空闲inode。保留当前有效权重与证据,重复缓存继续按每轮已结束/无占用/哈希一致规则清理。下一小时检查须同时分析HY v23恢复与训练进度、Kimodo v22组件结果,再分别确定后续动作;不能再次默认只推进Kimodo。 ++ ++ ++### 双模型同步检查与Kimodo v24恢复(2026-11-10 08 UTC) ++ ++HY v23仍在GPU0正常训练,检查时151步,结束核验163/250步,真实进程存活、完整2步保存/optimizer/scheduler恢复通过、无错误/资源停止。已读取初始/125完整轨迹且有限,记录loss/梯度有限;125幅度0.270、接段2.156、姿态0.442 rad、根位置0.463 m,仍不合格。继续既定有界250步及最终验证,不重启、不覆盖正在运行源码配置。 ++ ++Kimodo v22组件诊断完成,全部完整预测/参考与v20逐元素完全相同,组件数值有限。高抖动1863两种子每次第一帧stage1根平均跳变0.230/0.207 m,骨盆偏移变化0.262/0.236 m,方向余弦约-0.987/-0.989;解码第一帧仅0.043/0.038 m,第二帧达0.186/0.160 m。810与277也出现类似抵消/后续突跳。这支持针对根轨迹第一、第二帧直接约束,尚不声称唯一根因。已查原生inverse:平滑根仅加到水平x/z,根高度来自骨盆位置;水平组合的最大残差2.38e-7,三维直接相加的残差不是解码bug。完整证据artifacts/recovery_hourly_20261110T08_analysis.json。 ++ ++Kimodo v24以v19为基线、原正式best初始化,唯一算法变量stage1_boundary_weight 3→10,加强已有stage1_root_velocity中前两次差分权重;其余根位移1.0、解码加速度0.01、训练数据/学习率/种子/完整因果文本/推理及原门槛不变。并非延长失败最终断点,仍250更新、0/125/250固定两种子验证、先2步完整保存严格恢复;不以静止或抑制幅度换低抖动,无Test1024。目录`/root/autodl-fs/pangxinlei/models/hfd-recovery-pilot-v24-20261110/kimodo`,入口scripts/recovery_training_v24/supervise_pilot.py。配置差异与基线核验recovery_v24_checks.json,源码SHA/启动recovery_v24_launch.json。 ++ ++Kimodo监督器575150及子进程已核验存活,GPU1启动前空闲,GPU0的HY持续推进。目前Kimodo初始验证/两步恢复阶段,恢复测试尚未完成、效果未知。12GiB预算重新实写fsync通过,保留4GiB原子checkpoint上限、8小时墙钟、主/文本GPU绑定校验、日志轮转、目录项/磁盘inode保护。HY当前预留不清理。本地约1.537GiB、共享33638空闲inode,本地/共享1MiB写入fsync通过;共享df不当账号配额。 ++ ++清理已完成Kimodo v22与HY v17诊断重复索引共656.24MiB,确认旧训练/监督器退出、无打开句柄,与新v24/v23保留副本SHA一致。保留正式权重、完整恢复断点、必要候选和全部原始证据。清单recovery_cleanup_hourly_20261110T08.json。下次同时分析HY最终结果和Kimodo恢复/推进及逐例结果;双模型必须持续推进的要求保持。无正式训练,心跳416e9f54继续,下次09 UTC;任一模型正式训练健康启动后立即取消规则不变。 ++ ++ ++### 第二十五轮双模型推进(2026-11-10 09 UTC) ++ ++HY v23与Kimodo v24均正常完成250更新、完整2步optimizer/scheduler恢复通过,无错误/资源停止、预留释放。已读取两模型初始/125/最终所有完整轨迹且有限,loss及嵌套梯度记录有限。HY最终幅度0.275、接段2.098、姿态0.451 rad、根位置0.474 m、jerk293.03;2327进度0.359/0.334,961为0.000/0.004,仍不合格。移除旋转速度辅助并未恢复幅度,不继续延长此配方。 ++ ++Kimodo v24的125步通过小验证门槛,jerk775.87,2327进度1.540/0.822、961为1.327/0.609;但250步961第二种子0.209失败(第一0.608),2327为0.793/0.553,最终幅度1.008、接段2.431、姿态0.421 rad、根位置0.390 m、jerk561.37。小验证通过不等于更广交互恢复,不直接正式训练,也不把这些诊断称为FID。 ++ ++HY v25目录`/root/autodl-fs/pangxinlei/models/hfd-recovery-pilot-v25-20261110/hymotion`,入口scripts/recovery_training_v25/supervise_pilot.py。在v23基础上唯一算法变量root_velocity_weight 0.1→0(旋转速度权重已为0),保留原flow损失、C6目标/导数一致构造、完整因果文本、原正式best初始化、数据/种子/学习率和原门槛,做flow-only对照检验辅助目标是否冲突,而非预设根因。250更新、0/125/250两种子验证、完整2步保存恢复检查;当前恢复检查尚未完成。GPU0监督器575697与真实子进程已存活,无错误。 ++ ++Kimodo v25目录`/root/autodl-fs/pangxinlei/models/hfd-recovery-broad-v25-20261110/candidate125`,入口scripts/recovery_broad_v25/supervise_pilot.py。冻结v24 step-000125.pt,重复同一24validation案例/两种子/12帧执行/20步CFG2及完整因果文本,复用v18原正式基线及v20候选对照。零优化更新,只有权重与输出目录变化;要求用匹配扩大集原基线重算门槛并检查抖动,不用日志六例临时标记晋升。GPU1监督器575698与真实子进程已存活,无错误。两任务都不使用Test1024;双模型同步推进要求保持。 ++ ++每组2GiB预算重新实写fsync通过;HY8小时/256MiB原子checkpoint上限,Kimodo4小时/128MiB轨迹上限,原磁盘/inode/目录项/日志轮转保护保持。主模型与文本GPU分配保持各自0/1,不共享Kimodo文本工作进程。配置差异artifacts/recovery_v25_checks.json,源码SHA/启动recovery_v25_launch.json。原始结果、恢复与实际写入检查recovery_hourly_20261110T09_analysis.json。 ++ ++清理两旧训练重复索引及HY已分析失败125快照,共786.78MiB;确认旧训练/监督器退出、无打开句柄、缓存SHA与保留副本一致、中间快照无配置输入引用。保留两模型最终完整断点、Kimodo125候选与best、正式/必要历史对照及全部轨迹/配置/指标。清单recovery_cleanup_hourly_20261110T09.json。本地约1.536GiB、共享约33560空闲inode,本地/共享1MiB写入fsync通过,不把共享df当配额。下次同时检查HY恢复/训练推进、Kimodo扩大验证结果。无正式训练,小时心跳416e9f54继续,下次10 UTC;任一正式训练健康启动后立即取消规则保持。 ++ ++ ++### 第二十六轮同步检查与历史诊断(2026-11-10 10 UTC) ++ ++HY v25持续GPU0训练,开始检查161步,结束169/250步;完整2步恢复通过,记录loss/梯度及初始/125完整轨迹有限,无错误/资源停止。125幅度0.272、接段2.299、姿态0.450 rad、根位置0.469 m,仍不合格。继续本轮有界预算,不重启,不碰正在使用的预留或断点。 ++ ++Kimodo v25扩大验证正常完成,无错误,预留释放。三组同24案例/两种子完整轨迹均有限、参考与case_keys匹配。v24的125步候选幅度0.966、接段2.234、姿态0.412 rad、根位置0.585 m、jerk971.45;明显位移失败11/24,相比v19候选8/24、根位置0.528、jerk906.78退化。加强stage1前两帧速度权重没有带来扩大集改善,不采用、不机械延长。不能根据小验证通过正式开训。 ++ ++Kimodo v26固定较好的v19 step-000125权重,比较predicted_history与expert_past两臂,目录`/root/autodl-fs/pangxinlei/models/hfd-recovery-history-v26-20261110/kimodo`,入口scripts/recovery_history_v26/supervise_pilot.py。保持原24validation案例/两种子、12帧执行、20步CFG2、因果全文与其余推理设置。第一臂必须逐元素复现v20;第二臂每次用截至当前decision帧的真实过去60帧替换生成历史,索引绝不超过decision、不读取未来GT条件。目标是区分反馈误差累积与单次条件预测不足,非训练配方或部署结果;真实过去在部署不可获得,重置也会制造拼接突变,不能把该组较低误差/jerk或资格标记当恢复成功。零更新、无Test1024,保留完整两臂轨迹。 ++ ++静态验证过去索引上界、共享runtime CONFIG切换和原组复现要求,收据artifacts/recovery_v26_checks.json;启动及源码SHA recovery_v26_launch.json。监督器576671及真实子进程已核验存活,无错误,GPU1启动前空闲;HY继续GPU0。2GiB实写fsync预留、4小时墙钟、主/文本GPU绑定校验、日志轮转/目录项/磁盘inode与128MiB轨迹上限保留。当前尚无诊断结论。 ++ ++清理已完成Kimodo扩大验证重复索引328.12MiB,核验旧训练/监督器退出、无打开句柄及与新副本SHA一致;正式/必要候选、完整断点和全部证据保留。清单recovery_cleanup_hourly_20261110T10.json;三组配对轨迹与HY健康分析recovery_hourly_20261110T10_analysis.json。本地约1.536GiB、共享33503空闲inode,本地/共享1MiB写入fsync通过,共享df不是账号配额。下次同时分析HY最终结果和Kimodo历史对照,不搁置任一模型。无正式训练,心跳416e9f54继续,下次11 UTC;正式训练健康启动后立即取消规则不变。 ++ ++### 第二十七轮双模型历史诊断(2026-11-10 11 UTC) ++ ++HY v25完成250步,完整恢复检查通过,loss/梯度记录和0/125/250全部轨迹有限;最终幅度0.283、接段2.224、姿态0.457 rad、根位置0.461 m,仍不合格。去掉旋转及根速度辅助损失没有恢复动作幅度,不机械延长。Kimodo v26两臂完成;生成历史组逐元素复现v20。真实过去组姿态0.176 rad、根位置0.093 m,但接段6.186、jerk5296.08,且每段重置使行走进度看似通过,不能作为恢复成功。按当前段真实起点计算,真实历史组段内根位移RMSE仍0.139 m;生成历史组按自己当前起点计算为0.078 m,二者锚点定义不同,不能据此单独下因果结论。完整证据见HFD_bench/artifacts/recovery_hourly_20261110T11_analysis.json;其中mean_root_jerk_rms是坐标分量RMS,和既有向量模jerk相差sqrt(3),不替换旧指标。 ++ ++两个模型同步启动v27,根目录/root/autodl-fs/pangxinlei/models/hfd-recovery-history-v27-20261110。HY固定v25最终权重,原6validation×2种子对照生成历史与截至decision的真实过去;首臂应复现v25最终轨迹。Kimodo固定v19 step125,在原24validation×2种子中分别只用真实历史根平移、只用真实历史旋转替换,分解条件误差来源。混合历史可能存在平移/姿态不一致,因此仅作诊断,不能视为可部署配方或正式资格。索引不超过decision,未来占位和原采样设置保持;零更新、无Test1024。入口scripts/recovery_history_v27_{hymotion,kimodo}/supervise_pilot.py,源码/启动及健康收据recovery_v27_launch_*.json、recovery_v27_health.json。下次检查必须核对HY原臂复现,并结合Kimodo两组的段内误差/接缝而非全局重置进度选择训练改动。 ++ ++GPU0/1分别确认真实子进程存活并占用显存,当前为加载及诊断阶段,不宣称结果已改善。每组2GiB预算实写fsync预留通过,4小时监督器、磁盘/inode/目录项/日志保护保留。本地约1.536GiB,磁盘本地/共享1MiB实写fsync通过,共享空闲inode超过33000;共享df不作为账号配额。确认旧进程已退出、无打开句柄及新副本SHA相同后,清理HYv25和Kimv26重复启动索引共656.24MiB;所有正式权重、完整断点和轨迹保留,清单recovery_cleanup_hourly_20261110T11.json。 ++ ++尚无正式训练,心跳416e9f54继续;下一次12 UTC。任一模型正式训练通过验收且真实健康启动后立即取消心跳,不关闭训练监督器。两模型继续同步推进。 ++ ++### 第二十八轮双模型恢复试验(2026-11-10 12 UTC) ++ ++v27两模型正常完成,无错误/资源停止,预留已释放;两臂所有完整轨迹有限。HY生成历史组逐元素复现v25最终结果,幅度0.283、接段2.224、姿态0.457 rad、根位置0.461 m、jerk334.62,2个幅度坍缩案例;4个明显位移组合均未严格达到0.5(其中一个0.499574,不能四舍五入算通过)。真实过去组幅度0.644、姿态0.146、根位置0.034,但接段12.068、jerk1878.39;重置产生的行走改善不计资格。Kimodo只替换历史平移:幅度1.235、接段2.397、姿态0.394、根位置0.100、jerk5448.41;只替换历史旋转:幅度1.217、接段6.417、姿态0.183、根位置0.383、jerk1256.41,行走仍失败7/24。真实历史及混合条件均非部署方法,不能用其较低误差正式开训。完整逐案例/轨迹核验在artifacts/recovery_hourly_20261110T12_analysis.json,不是FID。 ++ ++两边同步启动v28各250步有界恢复试验,根目录/root/autodl-fs/pangxinlei/models/hfd-recovery-pilot-v28-20261110,入口scripts/recovery_training_v28/supervise_pilot.py。HY以v25配方为对照,仅将生成历史训练比例0.25提高到1,检验真实历史和自主历史的分布差异;Kimodo以v19为对照,仅将监督标签过渡6帧延长至12帧,检验更缓的姿态/高度纠正能否减小接段冲击。后者不是直接改变XY轨迹或推理后处理,也不能预设有效。均从原正式best初始化,保留0/125/250同6案例×2种子及原始验收门槛;未达幅度/行走条件的平滑不算改善,若小验证通过仍须扩大验证。Test1024未使用。 ++ ++配置差异记录recovery_v28_config_diff.json;12帧过渡测试验证第12帧起与对齐标签完全一致、历史不被改写、数值有限,推理条件/部署实现不改。HY与Kimodo关键运行源码分别匹配v25/v19;启动先做2步optimizer/scheduler断点回读与恢复,随后监督器继续至250步,当前恢复检查尚未完成,不宣称已通过。两个真实子进程已核验,GPU分别0/1,不存在正式训练。HY2GiB/Kimodo12GiB实写fsync预留均通过;8小时墙钟、日志轮转、目录项/容量/inode检查、原子断点失败保留旧文件机制沿用。启动/源码SHA和健康证据recovery_v28_launch_*.json、recovery_v28_health.json。 ++ ++本轮清理已结束v27两模型重复启动索引656.24MiB,删除前确认旧进程/监督器退出、无打开句柄并与新副本SHA一致;保留全部权重、断点和原始轨迹,清单recovery_cleanup_hourly_20261110T12.json。本地约1.536GiB,本地/共享1MiB实写fsync通过,共享inode超过33000,共享df不作账号配额。小时任务416e9f54保持,下一次13 UTC重点核对两边恢复收据、loss/梯度和初始/中途轨迹;任一正式训练健康启动立即取消,不终止监督器。 ++ ++### 双模型检查及Kimodo扩大验证(2026-11-10 13 UTC) ++ ++HY v28仍正常训练,检查时153步、结束记录158/250步;完整2步optimizer/scheduler断点回读与恢复通过,无错误或资源停止。125步幅度0.480(未达0.5)、接段1.695、姿态0.418 rad(仍超过原初始1.1倍)、根位置0.430 m、jerk351.95。2327进度0.965/1.173改善,961仍-0.007/0.001,尚不合格;继续当前有界训练,不重启、不动其预留。full_predicted_history_fraction表示当前窗口已有60帧自身历史的比例,训练初始段不足60帧,因此不应把该日志低于1误读为配置没有启用100%预测链。 ++ ++Kimodo v28正常完成250步,恢复收据通过,预留已释放。125步通过原6案例两种子小验证:幅度0.859、接段2.352、姿态0.425 rad、根位置0.344 m、jerk652.17;2327进度0.770/0.947、961为1.388/0.929。250步退化:幅度0.734、根位置0.506 m、jerk277.79,但2327一个种子-0.006、961一个种子0.031,不能因jerk降低采用末步。保留125候选和完整最终断点,不续训退化配方。两模型已保存全部轨迹有限,case_keys/参考与对应配方对照一致,初始逐元素复现;训练loss/梯度记录有限,证据artifacts/recovery_hourly_20261110T13_analysis.json。这些诊断不是正式FID。 ++ ++立即启动Kimodo v29扩大验证:/root/autodl-fs/pangxinlei/models/hfd-recovery-broad-v29-20261110/candidate125,入口scripts/recovery_broad_v29/supervise_pilot.py,固定v28 step-000125.pt,沿用原24validation案例(各模式覆盖)、两种子、12帧执行与20步CFG2;零训练更新,无Test1024。必须与v18 expanded baseline、v20候选比较全部配对轨迹、行走/坍缩和jerk;日志中的原六案例门槛标记仍仅为临时诊断,不能据此正式开训。候选jerk仍高于原初始,扩大验证不能只看平均误差。GPU1启动前空闲,真实子进程580064和监督器580063已核验存活;HY继续GPU0。2GiB预算实写fsync预留通过,4小时墙钟/日志轮转/容量inode监督保留;源码SHA/启动/健康证据recovery_v29_launch_kimodo.json、recovery_v29_health.json。 ++ ++清理已结束Kimodo v28的重复启动索引328.12MiB,确认旧进程/监督器退出、无打开句柄、新副本SHA一致;所有权重和轨迹保留,HY运行资源未动。清单recovery_cleanup_hourly_20261110T13.json。本地约1.536GiB、共享inode超过37000,本地/共享实写fsync通过;共享df不作账号配额。尚无正式训练,小时任务416e9f54继续,下一次14 UTC同步分析HY完成结果与Kimodo扩大集;正式训练健康启动后立即取消规则保持。 ++ ++### 第三十轮双模型恢复试验(2026-11-10 14 UTC) ++ ++HY v28正常完成250步,2步恢复收据通过,全部训练loss/梯度及0/125/250轨迹有限,初始与原对照精确复现。最终幅度0.561、接段1.772、无低于0.1坍缩,姿态0.421 rad、根位置0.441 m、jerk359.06;相对原始姿态0.372仍超过10%恶化限,961两种子行走进度-0.057/-0.006,未达资格。100%生成历史使幅度改善,但未解决行走,不能将整体平均改善当正式开训依据。 ++ ++Kimodo v29扩大验证正常完成,无错误/资源停止、预留释放;原始扩大基线、v19step125和v28step125三组48条完整轨迹均有限、case_keys/GT完全匹配。v28候选幅度0.794、接段2.239、姿态0.408 rad、根位置0.568 m、jerk961.69,无坍缩、明显位移失败8/24;v19候选根位置0.528、jerk906.78、同样8/24失败,因此不采用延长标签过渡。完整配对指标及案例保存在artifacts/recovery_hourly_20261110T14_analysis.json。这些是原始SMPL-X诊断,不是正式FID。 ++ ++同步启动v30各250步,根目录/root/autodl-fs/pangxinlei/models/hfd-recovery-pilot-v30-20261110,入口scripts/recovery_training_v30/supervise_pilot.py。HY以v28为对照,仅将根速度辅助损失0改为0.1,保留全生成历史与旋转辅助0,检验在动作幅度恢复后明确根速度目标是否改善行走。Kimodo回到扩大集较好的v19配方(6帧标签过渡),仅统一把root/body学习率峰值和下限乘0.5,warmup/衰减时点及总250步保持;目的是检验后期退化与更新幅度的关系,不预设成功、不无限延长。两者仍从原正式best初始化,保持原始验收基线、0/125/250同案例两种子,Test1024未使用。 ++ ++配置差异和源码匹配检查recovery_v30_checks.json通过;没有新增监督表示/推理路径代码,恢复合同按独立源码/配置生成。HY2GiB/Kimodo12GiB预算实写fsync预留完成;GPU0/1启动前空闲,双模型真实子进程和监督器核验存活,目前执行两步恢复检查,尚未宣称该检查通过。启动SHA收据recovery_v30_launch_*.json,资源和PID见recovery_v30_health.json。保留8小时监督器、磁盘/inode/目录项检查、日志轮转和原子断点失败保留旧文件机制。 ++ ++清理已结束HYv28及Kimodo v29重复启动索引656.24MiB,确认旧进程/监督器退出、无打开句柄及新副本SHA一致,清单recovery_cleanup_hourly_20261110T14.json;正式权重、完整恢复断点、候选和全部轨迹保留。本地约1.535GiB,共享inode超过37000,本地/共享写入fsync通过;共享df不当账号配额。尚未正式开训,小时任务416e9f54继续,下次15 UTC检查两边恢复收据和训练/验证;任一正式训练健康启动立即取消规则保持。 ++ ++调度更新(2026-11-10):用户要求将检查频率改为每半小时。已通过paseo heartbeat update将现有任务416e9f54改为 `*/30 * * * *`(Etc/UTC),持久化记录核验通过;每小时整点及30分执行,下次2026-11-10 15:00 UTC。任务旧名称仍含“每小时”,实际频率以cron为准。继续同步推进两模型;任一正式训练健康启动立即取消的条件不变,不创建重复任务。 ++ ++### 半小时检查及Kimodo v31扩大验证(2026-11-10 15:00 UTC) ++ ++HY v30持续GPU0训练,检查157步,结束记录160/250;两步optimizer/scheduler断点回读及恢复通过,无错误/资源停止。125步幅度0.446、接段1.732、姿态0.427 rad、根位置0.414 m,仍不合格,继续当前有界预算。训练loss/梯度记录及已保存0/125完整轨迹有限,不重启不清理正在使用的资源。 ++ ++Kimodo v30完成250步,恢复检查通过,无错误/资源停止、预留释放。125步通过小验证:幅度0.730、接段2.124、姿态0.425 rad、根位置0.363 m、jerk628.92;2327进度0.751/0.761,961为1.129/0.507(后者接近下限,须关注扩大集稳定性)。250步幅度0.718、接段2.346、姿态0.414、根位置0.403、jerk384.19,但2327一个种子0.461、961一个种子0.365不合格。减半学习率没有消除后期退化,不续训失败末步。所有0/125/250完整轨迹有限、case_keys/参考与v19一致,初始精确复现;对照证据artifacts/recovery_hourly_20261110T1500_analysis.json,非正式FID。 ++ ++GPU1已启动v31:/root/autodl-fs/pangxinlei/models/hfd-recovery-broad-v31-20261110/candidate125,入口scripts/recovery_broad_v31/supervise_pilot.py,固定v30 step-000125.pt。同24validation案例、两个种子、原采样和12帧执行;零更新无Test1024。原6案例资格标记不能替代扩大集验收,须与v18原始扩大基线、v20及v29完整轨迹比较,检查行走失败、接段、幅度、jerk与姿态/位置误差。实际子进程582073及监督器582072已核验存活;2GiB实写fsync预留通过,4小时及容量/inode/日志保护沿用。启动和健康证据recovery_v31_launch_kimodo.json、recovery_v31_health.json。 ++ ++清理已结束Kimodo v30重复索引328.12MiB,确认旧进程/监督器退出、无打开句柄且新副本SHA一致;权重/完整断点/轨迹保留,HY预留和运行缓存未动。清单recovery_cleanup_hourly_20261110T1500.json。本地约1.535GiB,共享inode超过37000,本地/共享实写fsync检查通过;不将共享df作为账号配额。无正式训练,保持半小时cron */30,下一次15:30 UTC,优先分析任一已完成模型;正式训练健康启动后立即取消任务416e9f54的条件不变。 ++ ++### 半小时检查及v32采样覆盖对照(2026-11-10 15:30 UTC) ++ ++HY v30正常完成250步,两步恢复检查通过,无错误/资源停止;最终幅度0.512、接段1.707、姿态0.419 rad、根位置0.436 m、jerk302.88。961行走进度-0.048/-0.006,姿态仍超原初始10%限,未合格;增加根速度辅助降低jerk但没有恢复该行走案例。训练loss/梯度和0/125/250完整轨迹有限,参考和初始复现核验通过。 ++ ++Kimodo v31扩大验证正常完成,预留释放、无错误。v30半学习率候选幅度0.708、接段2.061、姿态0.407 rad、根位置0.575 m、jerk732.01,无坍缩但行走失败13/24;v19候选为8/24、根位置0.528、jerk906.78。降低学习率有平滑收益但伴随运动不足,不能正式晋级。与原扩大基线/v19全部48条轨迹有限且case_keys/GT匹配,原始及逐案例证据artifacts/recovery_hourly_20261110T1530_analysis.json;非FID。 ++ ++转查采样覆盖:代码确认预测历史链都从decision0开始,原链长50个12帧窗口最多覆盖开头20秒。v32两模型各只把链长50改25(10秒),在固定250更新内增加会话切换机会;HY对照v30,Kimodo回到扩大集较好的v19对照。它减少连续窗口相关性,但也缩短后半段覆盖,属于待证假设,不能预设修复;下一次需从训练locations检查独立会话数和实际时域覆盖并结合运动/行走/姿态。仍从原正式best初始化、全预测历史、初始单姿态启动,不引入GT历史替换。验证不缩短,保持原完整时长、原案例和两个种子;原门槛不变,无Test1024。 ++ ++新根目录/root/autodl-fs/pangxinlei/models/hfd-recovery-pilot-v32-20261110,入口scripts/recovery_training_v32/supervise_pilot.py。配置差异及生效字段检查recovery_v32_checks.json;HY的curriculum.chain_segments和pilot.chain_segments均25,Kimodo读取pilot.chain_segments。未改监督表示/推理代码。HY2GiB、Kimodo12GiB实写fsync预留通过,启动前GPU0/1空闲,真实进程/监督器已核验存活,当前执行两步恢复检查,尚不宣称完成。8小时监督/有限日志/目录项与磁盘inode检查/原子断点保护保留,启动及健康收据recovery_v32_launch_*.json、recovery_v32_health.json。 ++ ++清理已结束HYv30及Kimodo v31重复索引656.24MiB,旧进程/监督器退出、无打开句柄及新副本SHA一致后删除,清单recovery_cleanup_hourly_20261110T1530.json。所有权重、完整恢复断点及轨迹保留。本地约1.534GiB、共享inode超过36000,本地/共享实写fsync通过,共享df不作账号配额。尚无正式训练,半小时任务416e9f54继续,下一次16:00 UTC;任一正式训练健康启动立即取消规则不变。 ++ ++### v32运行中半小时检查(服务器时间2026-11-10 16:00 UTC) ++ ++本次核查以服务器date和运行收据为时间依据;对话环境日期标为11月11日,但实际服务器为11月10日16:00 UTC。最新仍是v32,没有发现正式训练启动。两模型真实进程和监督器存活,HY结束检查79/250步,Kimodo 202/250步;两步optimizer/scheduler恢复收据均通过,未出现错误或资源停止,训练loss/梯度记录和所有已保存完整轨迹有限,初始复现及参考配对通过。不重启、不抢占GPU。 ++ ++HY尚只有初始验证;Kimodo125幅度1.035、接段2.372、姿态0.455 rad、根位置0.454 m,未合格:2327种子12345进度0.430、961种子2026进度0.364低于0.5,其余两个1.013/0.956。等待预算内中途/最终轨迹,不仅看loss或幅度。链长改动实际增加了覆盖:相同79步HY训练独立会话8→16;相同200步Kimodo为18→33,所有比较更新的locations均有记录。最远起点588→288帧,验证了减少后段覆盖的代价;这些变化不等于质量改善,仍按原完整验证门槛判定。 ++ ++证据artifacts/recovery_hourly_20261110T1600_analysis.json。本地约1.534GiB、共享空闲inode36968,本地/共享1MiB实写fsync通过,当前每运行目录项低于500。两模型正在使用的预留、缓存、断点保持;本轮无新增已核验可删除项,清理检查写recovery_cleanup_hourly_20261110T1600.json(0字节删除)。不重复创建实验或扩大预算。文档和任务状态已更新,半小时调度下一次16:30 UTC,任一正式训练健康启动立即取消规则保持。 ++ ++### 半小时检查及Kimodo模块诊断(服务器2026-11-10 16:30 UTC) ++ ++HY v32继续GPU0,检查158步,结束记录177/250。恢复检查通过,无错误/资源停止,记录loss/梯度和已保存初始/125完整轨迹有限。125幅度0.419、接段1.962、姿态0.451 rad、根位置0.419 m,仍不合格,继续既定有界预算。 ++ ++Kimodo v32正常完成250步,恢复检查通过,预留释放,无错误。125和250均未合格;250幅度0.865、接段2.551、姿态0.426 rad、根位置0.477 m、jerk325.49;2327进度0.312/0.537、961为0.186/0.438,4组合3失败。增加会话覆盖未解决行走,不能因jerk降低晋级。完整0/125/250轨迹有限、参考匹配及初始复现通过,证据artifacts/recovery_hourly_20261110T1630_analysis.json。保持原正式基线与门槛,无Test1024,不是FID。 ++ ++下一步停止连续标量调参,启动冻结v33模块交换诊断:/root/autodl-fs/pangxinlei/models/hfd-recovery-modules-v33-20261110/kimodo,入口scripts/recovery_modules_v33/supervise_pilot.py。以v19 step125为候选,三组分别原候选、只换回原正式best根模块、只换回原正式best身体模块;原正式best不是官方原生权重。各组先严格重新加载完整候选,再覆盖一个模块,24validation案例两种子同协议,零训练更新、无未来GT条件。权重键/形状匹配通过,root/body各204键,检查recovery_v33_checks.json。第一组必须逐元素复现v20;模块交互可能导致分布不匹配,不能将混合权重改善直接当因果损坏证明或正式资格。结果用于定位下一训练改动。 ++ ++GPU1启动前空闲,新子进程583485和监督器583482存活,HY原进程持续。2GiB实写fsync预留、4小时墙钟、日志轮转及容量/inode检查保留;启动/源码SHA和健康收据recovery_v33_launch_kimodo.json、recovery_v33_health.json。清理已结束Kimodo v32重复索引328.12MiB,旧进程退出、无打开句柄、新副本SHA一致后删除,清单recovery_cleanup_hourly_20261110T1630.json;全部模型断点和轨迹保留,HY在用资源不动。本地约1.534GiB、共享inode36935,本地/共享实写fsync通过,目录项低于500。无正式训练,半小时任务继续,下次17:00 UTC;正式训练健康启动即取消条件不变。 ++ ++### 双模型运行中检查(服务器2026-11-10 17:00 UTC) ++ ++HY v32检查时249步,结束记录249/250;真实训练进程和监督器存活,尚无completed收据及最终验证结论,不将步数达到预算当成已完成。两步恢复检查通过,loss/梯度记录及已保存轨迹有限,初始/参考复现通过。保留当前预算,等待后台最终验证与保存,不重复启动。 ++ ++Kimodo v33仍运行,原候选组已完成:全部48条轨迹逐元素复现v20,幅度0.991、接段2.213、姿态0.416 rad、根位置0.528 m、jerk906.78,说明对照基线一致。模块交换组尚无完整结果,不预判根或身体模块的作用。GPU1真实主进程583485及文本进程运行,GPU0为HY;无错误/监督器停止。证据artifacts/recovery_hourly_20261110T1700_analysis.json。诊断不当正式FID或正式训练资格。 ++ ++本地约1.534GiB、共享inode36931,本地/共享1MiB实写fsync通过,两个运行目录项均低于500。当前预留/缓存/断点正在使用,本轮未发现新增已核验可清理项,0字节删除,清理检查recovery_cleanup_hourly_20261110T1700.json。两份文档与任务状态更新,未正式开训,半小时任务保持,下一次17:30 UTC;任一正式训练健康启动立即取消任务416e9f54。 ++ ++### 双模型采样诊断v34(服务器2026-11-10 17:30 UTC) ++ ++HY v32完成250步,无错误/资源停止,恢复检查通过;全部训练记录及初始/中途/最终轨迹有限。最终幅度0.476、接段2.025、姿态0.454 rad、根位置0.443 m、jerk318.87,2327进度1.015/1.698,961为-0.055/-0.005。缩短链后仍不合格,不延长该配方。Kimodo v33三臂完成,无错误,预留释放;48案例种子轨迹有限、参考配对一致,候选逐元素复现v20。候选行走失败8/24、根位置0.528、jerk906.78;原根模块替换失败10/24、根位置0.550、jerk864.14;原身体模块替换幅度0.453、失败17/24、根位置0.633、jerk590.09,接近原正式模型。说明恢复后的身体模块对幅度/行走改善很重要,简单回滚根或身体都不解决问题;模块交互影响仍在,不能断言单一模块损坏。证据artifacts/recovery_hourly_20261110T1730_analysis.json,非FID。 ++ ++同步启动v34固定权重采样步数诊断,目录/root/autodl-fs/pangxinlei/models/hfd-recovery-sampling-v34-20261110/{hymotion,kimodo},入口scripts/recovery_sampling_v34_{hymotion,kimodo}/supervise_pilot.py。HY固定较好幅度的v28最终权重,6validation×2种子比较10与50步;Kimodo固定v19 step125,24validation×2种子比较20与50步。先要求原采样组逐元素复现HYv28最终/Kimv20,再分析50步是否改善运动、行走和jerk及延迟代价。之前50步HY对照针对原正式权重,本轮针对恢复候选,非重复同一配方。未更改初始状态/执行窗口/因果文本/GT条件,不训练更新、不读取Test1024,验收基线不重置。更高步数即使改善也须记录计算成本,不能混同原等预算评测。 ++ ++两组各2GiB实写fsync预留完成,GPU0/1启动前空闲,真实子进程/监督器核验存活,启动/源码SHA及健康见recovery_v34_launch_*.json、recovery_v34_health.json。4小时墙钟、磁盘/inode/目录项限制、有限日志/缓存保留。清理已结束HYv32和Kimv33重复索引656.24MiB,删除前确认旧进程退出、无句柄占用、新副本SHA一致,清单recovery_cleanup_hourly_20261110T1730.json;所有权重、完整恢复断点及轨迹保留。本地约1.533GiB、共享inode超过36800,本地/共享实写fsync通过,共享df不是账号配额。无正式训练,半小时调度继续,下次18:00 UTC;任一正式训练健康启动立即取消任务416e9f54。 ++ ++### HY文本条件诊断及Kimodo运行检查(服务器2026-11-10 18:00 UTC) ++ ++HY v34两组正常完成,无错误/资源停止、预留释放。10步组逐元素复现v28最终;50步幅度0.887、接段1.671、姿态0.454 rad、根位置0.687 m、jerk368.75,较10步根位置0.441明显退化;2327进度2.055/3.613过冲,961为-0.283/-0.573反向,4组全不合格。采样更精细不等于恢复,不采用50步。两组完整轨迹有限、case_keys/GT相同。 ++ ++Kimodo v34仍在运行50步组;已完成20步组全部48条轨迹逐元素复现v20。尚无50步完整结果,不推断改善。真实进程584260及监督器584259存活,无错误/资源停止。完整核验和可写性证据artifacts/recovery_hourly_20261110T1800_analysis.json。 ++ ++HY继续推进冻结v35:/root/autodl-fs/pangxinlei/models/hfd-recovery-text-v35-20261110/hymotion,入口scripts/recovery_text_v35/supervise_pilot.py,固定v28最终权重与原10步采样,比较full_text和empty_text两臂、原6案例两种子。检查已恢复幅度候选的持续行走失败是否依赖文本条件;此前空文本实验使用别的权重,本轮不冒充相同对照。空文本经原编码器编码,不是零向量;只改变推理文本,轨迹/未来GT不输入条件。空文本偏离任务条件,仅用于机制诊断,不能作为正式评测或放宽资格。要求full_text精确复现v28最终。零权重更新,无Test1024。 ++ ++GPU0启动前空闲,HY真实子进程584643和监督器584642存活,GPU1继续Kimodo。2GiB实写fsync预留通过,4小时墙钟、日志轮转、容量/inode/目录项保护保留;启动与健康收据recovery_v35_launch_hymotion.json、recovery_v35_health.json。清理已结束HYv34重复启动索引328.12MiB,核验旧进程退出、无打开句柄、新副本SHA相同后删除;Kimodo在用预留不动,所有权重和轨迹保留,清单recovery_cleanup_hourly_20261110T1800.json。本地约1.533GiB、共享inode36828,本地/共享1MiB实写fsync通过。尚无正式训练,半小时任务继续,下次18:30 UTC,正式训练健康启动即取消规则保持。 ++ ++### HY v36执行窗口权重试验(服务器2026-11-10 18:30 UTC) ++ ++HY v35正常完成,预留释放,无错误。全文组完整轨迹逐元素复现v28最终;空文本组幅度0.544、接段1.874、姿态0.413 rad、根位置0.514 m、jerk339.82,2327进度0.082/0.011,961为-0.051/-0.010。去掉文本未恢复961且破坏2327行走,不能采用;仅为机制诊断。两臂轨迹有限、案例/参考相同,证据artifacts/recovery_hourly_20261110T1830_analysis.json。Kimodo v34仍正常运行50步组,原20步组精确复现已核验,目前没有50步完整结果,不重启或抢占GPU1。 ++ ++HY恢复训练v36,/root/autodl-fs/pangxinlei/models/hfd-recovery-pilot-v36-20261110/hymotion,入口scripts/recovery_training_v36/supervise_pilot.py。对照v28配方,仅把training.execution_weight从3改6,保持全生成历史、50窗口链、原正式best初始化及全部其他训练设置。代码核实该权重作用于未来60帧中实际执行的前12帧,并以加权总和归一化,因此并非简单放大总体loss。目的是检验近端执行目标的相对权重,不预设解决行走。原10步推理/正常因果全文/固定validation两种子/原验收基线不改,无Test1024,各250步,0/125/250比较。配置差异、入口源码一致核验recovery_v36_checks.json。 ++ ++GPU0启动前空闲,真实子进程585015及监督器585014存活,当前执行2步恢复检查,尚不宣称检查已完成;2GiB实写fsync预留通过,8小时墙钟、日志轮转/目录项/磁盘inode/原子断点保护保留。启动SHA和健康收据recovery_v36_launch_hymotion.json、recovery_v36_health.json。Kimodo主进程584260持续,监督器无错误或停止。 ++ ++清理已结束HYv35重复索引328.12MiB,确认旧进程退出、无句柄占用及新副本SHA相同;清单recovery_cleanup_hourly_20261110T1830.json,全部权重/断点/轨迹保留,Kimodo在用资源不动。本地约1.533GiB,共享inode36795,本地/共享实写fsync通过,目录项低于500。未正式开训,半小时任务保持,下次19:00 UTC;任一正式训练健康启动即取消规则不变。 ++ ++ ++### 半小时检查与Kimodo引导强度诊断(服务器2026-11-10 19:00 UTC) ++ ++HY v36仍在GPU0,真实进程585133、监督器585014,记录124/250步,正等待中途验证结果;控制台有数据读取进展,容量监督收据更新,无错误/资源停止。两步optimizer/scheduler恢复和读回通过,全部训练记录有限、已保存完整轨迹有限且初始/参考复现通过,不重复启动。当前尚不能判断本轮改进。 ++ ++Kimodo v34两组已完成且预留释放。全部48条/组完整轨迹有限、case_keys及参考一致,20步组精确复现v20。50步相比20步:幅度0.991→1.293、接段2.213→2.063、姿态0.416→0.447rad、根位置0.528→0.563m、根jerk906.78→1103.45,行走失败仍8/24;不能凭接段变小和幅度增加采用更高采样步数。证据artifacts/recovery_hourly_20261110T1900_analysis.json,非正式FID,无Test1024。 ++ ++Kimodo继续冻结v37引导强度单变量诊断:/root/autodl-fs/pangxinlei/models/hfd-recovery-guidance-v37-20261110/kimodo,入口scripts/recovery_guidance_v37/supervise_pilot.py。固定v19step125、20步采样、原24validation两种子、12帧执行及正常因果全文,只比较guidance2与1,排查引导放大是否贡献抖动及行走失败。guidance2必须精确复现v20;不预设低引导能恢复,不改变验收门槛。参数通过closed_loop的config.guidance实际传入采样器;零训练更新,不是正式训练。 ++ ++GPU1启动前空闲,真实子进程585526与监督器585525存活、初始化日志推进。2GiB实写fsync预留通过,4小时墙钟/容量inode/目录项/日志轮转保护保留,源码SHA与健康证据recovery_v37_launch_kimodo.json、recovery_v37_health.json。清理已结束Kimodo v34重复索引328.12MiB,旧进程退出、无打开句柄及新副本SHA一致后删除;所有权重和轨迹保留,HY在用资源未动,清单recovery_cleanup_hourly_20261110T1900.json。本地约1.532GiB,共享空闲inode36761,本地/共享1MiB实写fsync通过;共享df不作账号配额。半小时任务416e9f54继续,下次19:30 UTC;尚无正式训练,正式训练健康启动即取消规则保持。 ++ ++ ++### 双模型半小时运行检查(服务器2026-11-10 19:30 UTC) ++ ++HY v36实际推进到224/250,进程585133/监督器585014存活,两步恢复读回通过,全部loss/梯度记录及初始/125完整轨迹有限,参考/案例匹配、初始精确复现。125步幅度0.533、接段1.733、姿态0.420rad、根位置0.438m、jerk364.83;2327进度1.150/1.325,961为-0.030/-0.004。幅度达线但行走仍失败,姿态较原初始0.3723增约12.8%,超过10%门槛,尚不能正式开训。继续既定250步预算等待最终结果,不以平均数或loss作恢复结论。 ++ ++Kimodo v37运行中,真实进程585526/监督器585525存活,GPU1约17GiB。guidance2对照48条完整轨迹精确复现v20;guidance1组尚无完整结果,不预判改善或重复启动。HY继续GPU0。无错误/资源停止,证据artifacts/recovery_hourly_20261110T1930_analysis.json。当前两模型均为pilot或冻结诊断,没有正式训练。 ++ ++本地约1.532GiB,共享空闲inode36748,本地/共享1MiB实写fsync通过;共享df不是账号配额。两运行目录项低于500,现有预留与缓存正在使用,本轮无新增核验可删资源,0字节清理,记录recovery_cleanup_hourly_20261110T1930.json。全部断点和完整轨迹保留。半小时任务416e9f54保持,下次20:00 UTC;任一正式训练健康启动后立即取消规则不变。 ++ ++ ++### 双模型v38诊断(服务器2026-11-10 20:00 UTC) ++ ++HY v36完成250步,恢复读回通过,无错误,预留释放。全部训练记录及0/125/250完整轨迹有限,初始精确复现,参考/案例一致。最终幅度0.609、接段1.780、姿态0.41234rad(较原初始0.37227约+10.77%)、根位置0.452m、jerk355.50;2327进度1.489/1.855,961为-0.074/-0.071。仍不合格,不延长失败配方。Kimodo v37完成两臂,48条/组轨迹有限且参考一致,原引导2精确复现v20;引导1幅度0.980、接段2.258、姿态0.409、根位置0.614、jerk529.29(原906.78),行走失败8/24→16/24,不能以更平滑替代任务能力。证据artifacts/recovery_hourly_20261110T2000_analysis.json,无Test1024,非FID。 ++ ++两模型同步进入冻结v38,目录/root/autodl-fs/pangxinlei/models/hfd-recovery-guidance-v38-20261110/{hymotion,kimodo},入口scripts/recovery_guidance_v38_模型名/supervise_pilot.py。HY固定v36最终权重及10步,原6validation两种子比较guidance1与2;此前更强引导针对其他权重,本轮检查已恢复幅度候选。Kimodo固定v19step125、20步及24validation两种子,对照constant2与前10次去噪引导2/后10次引导1,检验强引导行走收益与抖动是否可按去噪阶段分开;不是根/身体模块分离。仅改变引导,无未来GT输入,原验收门槛不变,零训练更新。原臂必须精确复现HYv36最终/Kimv20。Kimodo本地采样副本撤销该切换后逐字节还原原实现,反向去噪迭代顺序验证通过,检查recovery_v38_checks.json;不修改上游或历史运行源码。 ++ ++两GPU启动前空闲,各2GiB实写fsync预留完成;HY真实进程586140/监督器586137、Kimodo586144/586142存活,初始化中,无错误。启动SHA及健康证据recovery_v38_launch_*.json、recovery_v38_health.json。4小时墙钟、容量/inode/目录项/日志保护保留。清理已结束HYv36和Kimv37重复启动索引656.24MiB,旧进程退出、无打开句柄及新副本SHA一致后删除;所有权重、恢复断点及轨迹保留,清单recovery_cleanup_hourly_20261110T2000.json。本地约1.532GiB,共享inode36697,本地/共享实写fsync通过,共享df不作账号配额。没有正式训练,半小时调度保持,下次20:30 UTC;任一正式训练健康启动即取消任务416e9f54。 ++ ++ ++### v38分析、故障修复及双模型v39(服务器2026-11-10 20:30 UTC) ++ ++HY v38完成、预留释放。两臂完整轨迹有限且参考/案例相同,引导1逐元素复现v36最终。引导2幅度0.645、接段1.893、姿态0.423rad、根位置0.729m、jerk560.34,2327进度2.581/1.443、961为0.079/0.011,3/4行走组合失败;原引导1根位置0.452、jerk355.50,仍不采用更强引导。证据artifacts/recovery_hourly_20261110T2030_analysis.json,非FID。 ++ ++Kimodo v38失败退出,原因是新增独立recovery_generate.py保留包内相对导入,make_window调用from .kimodo_causal_rep报ImportError。此前AST检查未覆盖实际导入路径,是本轮入口错误,未产生完整诊断结果,不归因于模型。监督器已释放预留,失败日志保留。v39独立复制源码只把两处包内相对导入改为hfd_bridge绝对导入;模块及依赖实际导入检查通过、AST确认无相对导入,recovery_v39_checks.json。保持v38的冻结权重、constant2/late1两臂和全部验证协议;原臂轨迹复现仍待完成,不把导入通过当效果验证。 ++ ++HY同步v39有界250步,从原正式best初始化,以v36全部配置为对照,仅根速度辅助权重0→0.1,保留execution_weight6、全生成历史50窗口链及10步推理。依据是v36改善幅度,而v30在execution_weight3下的小速度辅助曾减少jerk;本轮检验该作用能否在v36保留,不能预设改善行走,不机械延长或改门槛。固定原validation两种子,无Test1024。 ++ ++目录/root/autodl-fs/pangxinlei/models/hfd-recovery-v39-20261110/{hymotion,kimodo},入口scripts/recovery_v39_模型名/supervise_pilot.py。HY真实进程586697/监督器586694处于两步恢复检查初始化;Kimodo586701/586699处于诊断初始化,无新错误,未宣称恢复/诊断已通过。两GPU启动前空闲,各2GiB实写fsync预留通过;HY8小时/Kim4小时墙钟、日志/缓存/容量inode/断点保护保留,启动与健康见recovery_v39_launch_*.json、recovery_v39_health.json。 ++ ++清理已结束HYv38和已失败Kimv38重复索引656.24MiB,确认旧进程退出、无打开句柄、新副本SHA相同后删除;全部权重、恢复断点、轨迹及错误证据保留,清单recovery_cleanup_hourly_20261110T2030.json。本地约1.531GiB、共享inode36632,本地/共享实写fsync通过,共享df不作账号配额,目录项低于500。尚无正式训练,半小时任务继续,下次21:00 UTC;任一正式训练健康启动即取消规则保持。 ++ ++ ++### v39半小时运行检查(服务器2026-11-10 21:00 UTC) ++ ++HY v39推进108/250,真实训练进程586859/监督器586694存活;两步optimizer/scheduler恢复及断点读回通过,已记录loss/梯度全部有限,初始完整轨迹精确复现及参考配对通过。尚未到125步中途验证,不能判断新辅助项效果,不重复启动或扩大预算。 ++ ++Kimodo v39修复重试正常运行,进程586701/监督器586699存活,GPU1约17GiB。constant2对照完成,全部48条完整轨迹逐元素精确复现v20,有限性及案例/参考匹配通过,确认本次导入修复与采样副本没有改变对照输出。late1诊断仍运行,无完整结果,不预判改善;没有再次出现导入错误或资源停止。HY继续GPU0。证据artifacts/recovery_hourly_20261110T2100_analysis.json。 ++ ++本地约1.531GiB,共享空闲inode36605,本地/共享1MiB实写fsync通过,共享df不作账号配额。两个运行目录项低于500,预留/缓存/恢复断点正在使用,本轮没有新增已核验可清理项,0字节删除,清单recovery_cleanup_hourly_20261110T2100.json。尚无正式训练,保留半小时任务416e9f54,下一检查21:30 UTC;任一正式训练健康启动即取消规则不变。 ++ ++ ++### Kimodo v40与HY运行检查(服务器2026-11-10 21:30 UTC) ++ ++HY v39检查209步、结束更新210/250,真实进程586859及监督器586694存活。恢复检查通过、全部loss/梯度及初始/125完整轨迹有限,参考配对/初始复现通过。125幅度0.499、接段1.747、姿态0.423rad、根位置0.417m、jerk302.33;2327进度1.017/1.255,961为-0.047/-0.019,仍未达门槛,继续原250步预算。预留实写期间出现一次114.8秒训练步耗时,进程仍推进,无资源停止,不将此解释为模型性能变化。 ++ ++Kimodo v39两组完成,原臂精确复现v20、完整48条/组轨迹有限且参考一致。late1幅度0.978、接段2.351、姿态0.417、根位置0.542m、jerk892.22(对照906.78,仅降约1.6%),行走失败仍8/24,根位置误差增大,不采用该推理改动。证据artifacts/recovery_hourly_20261110T2130_analysis.json,无Test1024,非FID。 ++ ++Kimodo启动有界训练v40:/root/autodl-fs/pangxinlei/models/hfd-recovery-pilot-v40-20261110/kimodo,入口scripts/recovery_training_v40/supervise_pilot.py。从原正式best初始化,以v19为基线只把recovery.temporal_weights.decoded_root_acceleration从0.01提高到0.1,其余训练/验证/恢复配方保持,250步并保存0/125/250对照。该项跟踪解码后根位置二阶差分与参考加速度(包含fps及评分掩码),不是把所有加速度压零。此前引导改动无法兼顾行走与抖动,检验加强监督是否有帮助,不预设改善或放宽门槛。原入口/runtime源码一致、单变量配置核验recovery_v40_checks.json。 ++ ++GPU1启动前空闲,12GiB预算实写fsync预留通过(实际reserve约11.42GiB,非shared df配额推断);监督器587407与真实子进程存活,当前恢复检查初始化,尚未宣称通过。8小时墙钟、原子断点/日志/缓存/容量inode保护保留,启动SHA/健康recovery_v40_launch_kimodo.json、recovery_v40_health.json。HY GPU0继续运行。清理已结束Kimodo v39重复索引328.12MiB,确认旧进程退出、无打开句柄及新副本SHA相同;权重/恢复断点/完整轨迹保留,清单recovery_cleanup_hourly_20261110T2130.json。本地约1.531GiB,共享inode36585,本地/共享实写fsync通过,目录项低于500。未正式开训,半小时调度继续,下次22:00 UTC;正式训练健康启动即取消规则保持。 ++ ++ ++### HY扩展验证v41及Kimodo v40(服务器2026-11-10 22:00 UTC) ++ ++HY v39正常完成250步、恢复读回通过,预留释放;所有训练记录及0/125/250完整轨迹有限,原初始精确复现、参考/案例配对通过。最终幅度0.563、接段1.724、姿态0.418rad、根位置0.439m、jerk302.30;2327进度1.069/1.454,961为-0.069/-0.007,姿态及行走仍不达门槛。对比v36降低jerk但未恢复行走,不延长此配方。Kimodo v40检查196步、结束200/250,真实587571/监督器587407存活;完整3.394GB断点两步恢复通过、loss/梯度和初始/125轨迹有限,初始/参考匹配通过。125小验证幅度0.886、接段2.343、姿态0.442rad、根位置0.367m,现有小验证门槛通过;尚需最终结果及更大validation,不能据此正式开训。证据artifacts/recovery_hourly_20261110T2200_analysis.json,非FID。 ++ ++HY继续冻结扩展诊断v41:/root/autodl-fs/pangxinlei/models/hfd-recovery-broad-v41-20261110/hymotion,入口scripts/recovery_broad_v41_hymotion/supervise_pilot.py。同24validation案例、两种子比较原正式best、v36最终、v39最终三臂;6种交互模式各4例,选择不依赖模型结果并包含原6例,复用Kimodo扩展选样规则。先产生相同24例原正式基线,不沿用6例误差作比较门槛;逐臂使用已有asset/adapter键/shape校验完整重载权重,保持10步/引导1/12帧执行及因果文本,无训练更新/无Test1024。原6例子集需精确复现,各臂参考需一致。此轮为失败分布与覆盖诊断,不是已通过小验证后的正式晋级复核,不宣称候选合格。代码检查recovery_v41_checks.json。 ++ ++GPU0启动前空闲,监督器587842和子进程存活,仍在初始化;2GiB实写fsync预留通过,4小时墙钟、日志/缓存/容量inode保护保留,启动SHA和健康recovery_v41_launch_hymotion.json、recovery_v41_health.json。Kimodo GPU1继续训练,当前预留与恢复断点不动。清理已结束HYv39重复索引328.12MiB,旧进程退出、无打开句柄和新副本SHA一致后删除,所有权重/完整轨迹保留,清单recovery_cleanup_hourly_20261110T2200.json。本地约1.531GiB,共享inode36532,本地/共享实写fsync通过,目录项低于500,共享df不作账号配额。没有正式训练,半小时任务继续,下次22:30 UTC;任一正式训练健康启动即取消规则保持。 ++ ++ ++### Kimodo最终分析及扩大验证v42(服务器2026-11-10 22:30 UTC) ++ ++Kimodo v40正常完成250步,恢复读回通过、预留释放。全部训练loss/梯度以及0/125/250完整轨迹有限,初始精确复现、参考与案例匹配。125步幅度0.886、接段2.343、姿态0.442rad、根位置0.367m、jerk536.63;2327进度0.887/1.015、961为1.272/0.888,四组均达标。250步幅度0.665、接段2.671、姿态0.418、根位置0.493、jerk206.09,但2327进度0.004/0.788、961为0.045/0.162,三组行走失败且根位置超原基线10%。更低jerk伴随行走退化,不能晋级最终权重,不继续延长。证据artifacts/recovery_hourly_20261110T2230_analysis.json,非FID,无Test1024。 ++ ++保留125步候选,启动冻结扩大验证v42:/root/autodl-fs/pangxinlei/models/hfd-recovery-broad-v42-20261110/candidate125,入口scripts/recovery_broad_v42/supervise_pilot.py。同24validation×2种子、各模式4例,原20步guidance2/12帧执行不变。启动前baseline-validation.json替换为v18/baseline的48条原正式模型结果,避免旧6案例误差门槛与扩展样本错配;来源及复现操作写experiment.json和recovery_v42_checks.json,门槛未放宽。需核验原6子集精确复现v40step125、全量参考/案例一致和逐失败案例。小验证通过不能直接启动正式训练。 ++ ++GPU1启动前空闲,新监督器588139与子进程存活,初始化推进;2GiB实写fsync预留通过,4小时墙钟/日志/缓存/容量inode保护保留,recovery_v42_launch_kimodo.json、recovery_v42_health.json。HY v41进程587845/监督器587842存活,GPU0约18GiB,容量心跳更新,第一臂尚无完整输出,不声称已复现,不重启。原来较小HY验证耗时不能直接套到此次24例×三权重上。 ++ ++清理已完成Kimodo v40重复索引328.12MiB,确认旧进程退出、无打开句柄及新副本SHA相同后删除,所有权重/完整恢复断点/轨迹保留;recovery_cleanup_hourly_20261110T2230.json。HY在用资源不动。本地约1.531GiB,共享inode约36500,本地/共享实写fsync通过,共享df不作账号配额,目录项低于500。尚无正式训练,半小时任务保持,下次23:00 UTC;任一正式训练健康启动即取消规则保持。 ++ ++ ++### 扩展结果及Kimodo中点诊断v43(服务器2026-11-10 23:00 UTC) ++ ++HY v41原正式模型扩展基线已完成:幅度0.271、接段1.250、姿态0.373rad、根位置0.768m、jerk331.81,明显位移组合失败23/24。48条轨迹有限、原6例子集逐元素复现;恢复候选尚未完成,进程587845/监督器587842和容量心跳正常,不能只根据基线宣布改善。Kimodo v42完成,预留释放、无错误;完整48条轨迹有限、与扩展原基线参考/案例相同,原6例精确复现v40step125。扩展幅度0.741、接段2.356、姿态0.403rad、根位置0.558m、jerk725.25,行走失败9/24,未过原门槛。对比v19候选失败8/24、jerk906.78、姿态0.416,平滑/姿态有所改善但行走没有整体改善,不能正式开训。证据artifacts/recovery_hourly_20261110T2300_analysis.json,无Test1024,非FID。 ++ ++启动冻结v43诊断:/root/autodl-fs/pangxinlei/models/hfd-recovery-midpoint-v43-20261110/kimodo,入口scripts/recovery_midpoint_v43/supervise_pilot.py。两臂是原v40step125对照及v40/v19两个step125完整denoiser浮点张量等权中点,非浮点缓冲保持一致;两者同初始化/架构,仅加速度训练权重不同。固定0.5而非扫描参数,检验行走与平滑性折中是否存在,不假设权重平均必然改善。实际检查键/shape/dtype一致、中点数值有限、非浮点缓冲相同,recovery_v43_checks.json。使用相同24validation两种子/20步guidance2与原48例基线,原臂需精确复现v42,零训练更新,不导出新大权重;原门槛不变。 ++ ++GPU1启动前空闲,监督器588526和真实子进程存活、初始化中;2GiB实写fsync预留完成,4小时墙钟/日志/缓存/容量inode保护继续,启动及健康recovery_v43_launch_kimodo.json、recovery_v43_health.json。HY GPU0继续。清理已结束v42重复索引328.12MiB,确认旧进程退出、无打开句柄、新副本SHA一致,所有权重/恢复断点/轨迹保留;recovery_cleanup_hourly_20261110T2300.json。本地约1.530GiB,共享inode约36450,本地/共享实写fsync通过,共享df不是账号配额,目录项低于500。尚无正式训练,半小时任务保持,下次23:30 UTC;任一正式训练健康启动即取消规则不变。 ++ ++ ++### 双模型扩展诊断运行检查(服务器2026-11-10 23:30 UTC) ++ ++HY v41进程587845及监督器587842存活,累计约88分钟,GPU0约18GiB且有实际计算活动;原正式模型基线已核验,恢复候选仍未完成,没有新完整结果,不重复启动。Kimodo v43进程588528及监督器588526存活,GPU1约17GiB且计算活跃;v40对照臂48条完整轨迹逐元素精确复现v42,数值有限、原6例和全量参考匹配。中点臂仍运行,不预判权重平均收益。两组无错误或资源停止,证据artifacts/recovery_hourly_20261110T2330_analysis.json。均为冻结诊断,零训练更新,不是正式训练/FID。 ++ ++本地约1.530GiB、共享空闲inode36448,本地/共享1MiB实写fsync通过,共享df不作账号配额。目录项低于500,运行中预留与缓存保持,本轮无新增已核验可清理项,0字节删除,清单recovery_cleanup_hourly_20261110T2330.json。两份文档及任务状态已更新;半小时调度保持,下一次11月11日00:00 UTC,任一正式训练健康启动立即取消416e9f54规则不变。 ++ ++ ++### Kimodo根训练范围v44及HY扩展进展(服务器2026-11-11 00:00 UTC) ++ ++HY v41完成v36臂:幅度0.544、接段1.776、姿态0.402rad、根位置0.738m、jerk357.01,行走失败22/24(原基线23/24);幅度恢复不等于行走恢复,未过门槛。完整48轨迹有限、全量参考/案例匹配、原6例子集逐元素复现v36最终。v39仍运行,真实587845/监督器587842存活,无错误,不抢占GPU0。Kimodo v43两臂完成,预留释放、无错误,中点幅度0.798、接段2.325、姿态0.408、根位置0.564m、jerk732.78,行走失败10/24,较v40的9/24和v19的8/24更差,不采用。完整轨迹有限且参考匹配。证据artifacts/recovery_hourly_20261111T0000_analysis.json,非FID,无Test1024。 ++ ++Kimodo启动v44有界250步:/root/autodl-fs/pangxinlei/models/hfd-recovery-pilot-v44-20261111/kimodo,入口scripts/recovery_training_v44/supervise_pilot.py。以v19为对照、原正式best初始化,仅training.root_trainable由full改projections;根模块input_linear/output_linear训练,内部参数冻结,身体模块全量训练,全部学习率/损失(加速度仍0.01)/链/验证配置保持。基于身体适配必要且多次根全量训练后期行走退化,检验限制根更新自由度是否有帮助,不预设根更新就是原因。使用已有受支持模式,不修改上游。单变量配置及优化器排除冻结参数验证通过,recovery_v44_checks.json;真实模型训练计数/恢复检查待启动完成,不宣称已通过。 ++ ++GPU1启动前空闲,监督器589285和真实子进程存活,处于恢复检查初始化;12GiB预算实写fsync预留通过(实际reserve约11.42GiB),8小时墙钟、日志/缓存/目录项/容量inode/原子断点保护保留,recovery_v44_launch_kimodo.json、recovery_v44_health.json。HY继续扩展诊断。清理已结束v43重复索引328.12MiB,确认旧进程退出、无打开句柄和新副本SHA相同;所有正式权重/恢复断点/候选/轨迹保留,清单recovery_cleanup_hourly_20261111T0000.json。本地约1.530GiB,共享inode36432,本地/共享实写fsync通过,共享df不作账号配额,目录项低于500。无正式训练,半小时调度保持,下次00:30 UTC;正式训练健康启动立即取消规则不变。 ++ ++ ++### 双模型运行检查(服务器2026-11-11 00:30 UTC) ++ ++HY v41监督器587842、真实子进程587845存活,扩展v39臂继续;原基线及v36臂已完成,v36行走失败22/24,尚无新完整候选结果。Kimodo v44监督器589285、训练子进程589455存活,审计182/250步,loss/梯度及已保存完整轨迹有限,初始复现、参考匹配通过。两步恢复读回通过,optimizer/scheduler均为2,完整断点2268515405字节。实际根可训练565253、身体141529356参数,符合根projections模式。125步小验证幅度1.144、接段2.255、姿态0.462rad、根位置0.317m,既定小验证门槛通过;jerk773.16仍需关注。最终及扩展验证未完成,不据此宣布正式训练资格。证据artifacts/recovery_hourly_20261111T0030_analysis.json;非FID,无Test1024。 ++ ++两组无错误收据、容量心跳继续,HY约2小时28分钟,仍在4小时诊断上限内;Kimodo受8小时训练上限保护。本地约1.530GiB、共享空闲inode36397,本地/共享1MiB实写fsync通过,共享df不作账号配额。目录项均低于500;保留运行中预留、缓存、完整断点及轨迹,无新增可安全清理项,删除0字节,清单recovery_cleanup_hourly_20261111T0030.json。未重复启动。尚无正式训练,半小时调度继续,下一检查点01:00 UTC,检查HY扩展结果及Kimodo最终验证;任一正式训练健康启动即取消416e9f54规则保持。 ++ ++ ++### 完成结果及双模型v45(服务器2026-11-11 01:00 UTC) ++ ++HY v41三臂完成、无错误且预留已释放,所有完整轨迹有限、48例参考匹配,原6例子集精确复现各自候选。v39扩展幅度0.495、接段1.718、姿态0.400rad、根位置0.739m、jerk303.50,行走失败22/24;对比v36幅度0.544、jerk357.01、同样22/24失败,平滑收益没有解决广泛行走不足,未过原门槛。Kimodo v44正常完成250步、恢复读回通过、预留释放;0/125/250完整轨迹和训练数值有限,初始及参考匹配。最终幅度1.043、接段2.447、姿态0.439rad、根位置0.454m、jerk394.23,但961两种子进度0.352/0.298,不合格。125步通过小验证(幅度1.144、根位置0.317m、jerk773.16),保留该候选进行扩展复核,不延长退化终点。证据artifacts/recovery_hourly_20261111T0100_analysis.json,非FID,无Test1024。 ++ ++已启动共享根/root/autodl-fs/pangxinlei/models/hfd-recovery-v45-20261111:HY入口scripts/recovery_v45_hymotion/supervise_pilot.py,从原正式best重置进行250步有界实验,相对v36仅training.text_dropout由0.1改0;其余训练参数、执行权重6、因果预测历史及原门槛不变。依据扩展行走普遍不足、此前空文本消融损害行走,检验训练时完整保留文本条件的假设,尚不能归因或宣布改善。Kimodo入口scripts/recovery_v45_kimodo/supervise_pilot.py,冻结v44step125,固定24validation×2种子,沿用原正式48例基线,验证原6例子集精确复现及扩大覆盖后的行走/抖动表现。配置单变量、48例基线和语法检查通过,recovery_v45_checks.json。 ++ ++启动前GPU0/1均空闲;HY监督器590055/子进程590059、Kimodo监督器590058/子进程590060存活,无错误,处于初始化(HY恢复检查尚未完成,不宣称已通过)。各2GiB预算实写fsync预留通过,HY8小时/Kimodo4小时上限及原容量/inode/日志/原子断点保护保持,健康及源码SHA见recovery_v45_health.json和launch收据。清理已结束HYv41、Kimodov44重复启动索引共656.24MiB,确认进程退出、无打开句柄、新副本SHA相同;全部权重/恢复断点/诊断轨迹保留,清单recovery_cleanup_hourly_20261111T0100.json。本地约1.529GiB,共享空闲inode36348,实写fsync检查通过,目录项低于500,共享df不作账号配额。尚无正式训练,保留半小时调度,下一检查点01:30 UTC;任一正式训练健康启动立即取消416e9f54条件不变。 ++ ++ ++### HY v45继续、Kimodo位移监督v46(服务器2026-11-11 01:30 UTC) ++ ++HY v45真实训练590274/监督器590055存活,检查55步、轨迹审计时57/250,loss/梯度有限;原初始轨迹逐元素复现v36初始,2步完整恢复读回通过(optimizer/scheduler为2,136884553字节断点)。尚未到125验证,保留正在运行配方,不重复启动。Kimodo v45扩展验证完成,无错误、预留释放;48条完整轨迹有限、参考/case_keys与原扩展基线一致,原6例精确复现v44step125。幅度0.938、接段2.249、姿态0.426rad、根位置0.561m、jerk900.12,行走失败9/24,较v19的8/24及根位置0.528m未改善。限制根训练范围未解决泛化不足,不晋级正式训练。证据artifacts/recovery_hourly_20261111T0130_analysis.json,非FID,无Test1024。 ++ ++Kimodo实际启动v46:/root/autodl-fs/pangxinlei/models/hfd-recovery-pilot-v46-20261111/kimodo,入口scripts/recovery_training_v46/supervise_pilot.py。以v19原正式best初始化配方为对照,仅recovery.temporal_weights.stage1_root_displacement由1增至2,其余训练/损失/学习率/根全量训练/因果输入/推理不变。检验更强已有位移目标能否减少扩展行走不足,不预设有效,仍需防止抖动与误差退化。既有损失实现不变,只使用训练目标和有效帧掩码;单变量与原基线保持检查recovery_v46_checks.json。250步有界、固定0/125/250两种子验证、原门槛不变,不从退化断点延长。 ++ ++GPU1空闲后启动监督器590548和真实子进程,健康收据recovery_v46_health.json、源码SHA见launch收据;处于恢复检查初始化,不宣称新轮恢复或效果通过。12GiB预算实际写入fsync预留通过,8小时墙钟、容量/inode/目录项/日志/原子断点保护保持;HY GPU0继续。清理已结束Kimodo v45重复索引328.12MiB,确认旧进程退出、无打开句柄和新副本SHA一致;所有模型权重、完整断点及轨迹保留,清单recovery_cleanup_hourly_20261111T0130.json。本地约1.529GiB、共享空闲inode超过36000,本地/共享实写fsync通过,目录项低于500,共享df不作为账号配额。无正式训练,半小时调度继续,下次02:00 UTC;正式训练健康启动即取消416e9f54规则不变。 ++ ++ ++### 双模型中期健康检查(服务器2026-11-11 02:00 UTC) ++ ++HY v45真实进程590274/监督器590055存活,检查131步、审计135/250;Kimodo v46真实进程590704/监督器590548存活,检查196步、审计200/250。双方训练loss/梯度有限,初始/125完整轨迹有限、参考及案例匹配,原初始逐元素复现;两步恢复读回通过,optimizer/scheduler均为2,HY完整断点136884553字节、Kimodo3394249205字节。无错误或watchdog停止收据,GPU0/1分别约20.1/20.3GiB,监督器继续保护。 ++ ++HY125幅度0.644、接段1.729、姿态0.409rad、根位置0.504m(比原初始约恶化12.3%)、jerk383.25;2327两种子进度1.644/2.514,961为-0.267/-0.129,3/4行走组合失败,尚未合格。Kimodo125幅度1.194、接段2.224、姿态0.452rad、根位置0.424m、jerk766.11;2327进度0.750/0.876,961为1.434/0.524,现有小验证门槛通过,最终及扩展验证待完成,不能据此正式开训。证据artifacts/recovery_hourly_20261111T0200_analysis.json,非FID,无Test1024。保持原250步有界预算,不重复启动或修改运行源码。 ++ ++本地约1.529GiB、共享空闲inode36269,本地/共享1MiB实写fsync通过,共享df不作账号配额;两运行目录项低于500,预留及旧断点保护正常。没有新核验可删项,删除0字节,保留活动预留、缓存、权重及轨迹,清单recovery_cleanup_hourly_20261111T0200.json。无正式训练,半小时调度保持,下次02:30 UTC检查最终结果;正式训练健康启动即取消416e9f54规则不变。 ++ ++ ++### Kimodo v46完成及扩展v47(服务器2026-11-11 02:30 UTC) ++ ++HY v45进程590274/监督器590055存活,最新日志249/250,尚无完成收据或最终验证结果;训练数值及已保存初始/125完整轨迹有限,原初始复现和参考匹配通过,恢复检查通过。保留运行中预留及源码,未重复启动。Kimodo v46正常完成250步,预留释放、无错误;初始/125/250完整轨迹有限、参考/案例匹配、原初始精确复现,两步完整恢复通过。最终幅度1.219、接段2.329、姿态0.451rad、根位置0.437m、jerk478.29;2327进度0.864/0.585,961进度0.242/0.237,最终不合格。125小验证通过,但这不等于扩大覆盖合格。证据artifacts/recovery_hourly_20261111T0230_analysis.json,非FID,无Test1024。 ++ ++已在空闲GPU1启动冻结扩展v47:/root/autodl-fs/pangxinlei/models/hfd-recovery-broad-v47-20261111/candidate125,入口scripts/recovery_broad_v47/supervise_pilot.py。使用v46的step125候选,同24validation案例×两种子及原正式48例基线,20步/guidance2,零训练更新,保持原门槛。需核对完整参考及原6例精确复现后比较v19等扩展对照,不延长失败终点。2GiB预算实写fsync预留通过;监督器591195和真实子进程存活,初始化中,无错误,4小时墙钟/容量inode/日志缓存/原子保存保护保持,recovery_v47_checks.json、recovery_v47_health.json和launch收据。 ++ ++清理已结束Kimodo v46重复启动索引328.12MiB,旧进程退出、无打开句柄、新副本SHA相同;保留所有模型权重、完整恢复断点及轨迹,清单recovery_cleanup_hourly_20261111T0230.json。本地约1.529GiB、共享空闲inode超过36000,本地/共享实写fsync通过,目录项低于500,共享df不作账号配额。仍无正式训练,半小时调度保持,下次03:00 UTC检查HY最终结果和Kimodo扩展结果;正式训练健康启动即取消416e9f54规则不变。 ++ ++ ++### 完成分析及历史通道诊断v48(服务器2026-11-11 03:00 UTC) ++ ++HY v45正常完成250步,恢复检查通过、预留释放;训练及0/125/250完整轨迹有限、原初始精确复现和参考配对通过。最终幅度0.700、接段1.758、姿态0.406rad、根位置0.522m(较原初始恶化约16.2%)、jerk366.42;2327进度1.740/2.607,961为-0.326/-0.119,3/4行走组合失败,取消文本dropout未解决行为问题,不延长。Kimodo v47扩展完成、预留释放,无错误;48完整轨迹有限、原扩展参考匹配、原6例精确复现v46step125。幅度0.971、接段2.228、姿态0.421rad、根位置0.581m、jerk961.83,无低于0.1坍缩,但行走失败10/24,比v19的8/24、根位置0.528m及jerk906.78更差,不采用。证据artifacts/recovery_hourly_20261111T0300_analysis.json,非FID,无Test1024。 ++ ++两边实际启动冻结诊断v48,共享根/root/autodl-fs/pangxinlei/models/hfd-recovery-history-v48-20261111/{hymotion,kimodo},入口scripts/recovery_history_v48_{hymotion,kimodo}/supervise_pilot.py。HY固定v36最终候选、原6validation×两种子;Kimodo固定v19step125候选、扩展24validation×两种子。各含自主预测历史对照、只替换过去根平移、只替换过去旋转三臂。根平移是原始轨迹0:3通道,其余旋转3:;历史索引仅0..decision,未来占位仍由末历史帧构造。只作反馈误差定位,真实过去臂不可部署,位置重置造成的进度收益不算自主行走,所有oracle臂明确禁止用于正式晋级;自主对照需精确复现既有结果后才解释差异。不更新训练权重、不导出新大模型,保留原门槛。检查recovery_v48_checks.json,源码SHA见launch收据。 ++ ++GPU0/1启动前空闲;HY监督器591617/子进程591619、Kimodo监督器591621/子进程591622存活,初始化中,无错误,健康recovery_v48_health.json。各2GiB预算实写fsync预留通过,4小时墙钟/容量inode/目录项/日志缓存/原子保存保护保持。清理已结束HYv45、Kimodov47重复索引共656.24MiB,确认旧进程退出、无打开句柄、新副本SHA相同;全部模型权重、完整断点和轨迹保留,清单recovery_cleanup_hourly_20261111T0300.json。本地约1.528GiB、共享inode36185,本地/共享实写fsync通过,目录项低于500,共享df不作账号配额。尚无正式训练,半小时调度保持,下次03:30 UTC检查诊断对照复现与进度;正式训练健康启动即取消416e9f54条件不变。 ++ ++ ++### HY历史诊断细分v49、Kimodo v48继续(服务器2026-11-11 03:30 UTC) ++ ++HY v48三臂完成,无错误且预留释放,完整轨迹有限、同参考匹配,自主预测历史逐元素精确复现v36最终。只替换过去平移后根误差0.055m、行走进度接近1,但jerk2521.70,属于位置重置收益,不是自主行走改善;身体幅度0.607基本不变。只替换过去全部旋转后幅度0.832、姿态误差0.183rad、根位置0.279m,961两种子进度0.507/0.717;接段比10.516、jerk476.27,不可部署。表明旋转历史可能参与行走失败,但尚未区分根朝向和身体姿态,不能把oracle收益作为资格。Kimodo v48监督器591621/子进程591622存活,GPU1约17GiB、计算活动正常,尚在后续诊断;自主对照48条完整轨迹有限并逐元素精确复现v20,参考匹配,原有8/24行走失败不变。证据artifacts/recovery_hourly_20261111T0330_analysis.json,非FID,无Test1024。 ++ ++HY启动冻结v49:/root/autodl-fs/pangxinlei/models/hfd-recovery-history-v49-20261111/hymotion,入口scripts/recovery_history_v49_hymotion/supervise_pilot.py。相同v36最终候选、6validation×两种子,三臂为自主对照、仅替换过去根朝向3:6、仅替换过去身体关节旋转6:69;已核对原始3:69为22个轴角关节,保持根平移预测不重置,历史索引仅0..decision,不读未来GT。零训练更新,oracle臂禁止正式晋级;先要求自主对照精确复现,再据通道差异设计恢复实验。检查recovery_v49_checks.json。 ++ ++GPU0启动前空闲,监督器591942及真实子进程存活,无错误、初始化中;2GiB预算实写fsync预留通过,4小时墙钟/容量inode/日志缓存/原子保存保护保持,健康recovery_v49_health.json和启动SHA收据。Kimodo原运行继续、不重复启动。清理已结束HYv48重复索引328.12MiB,进程退出、无打开句柄、新副本SHA相同;所有权重/断点/轨迹保留,清单recovery_cleanup_hourly_20261111T0330.json。本地约1.528GiB、共享inode超过36000,实写fsync通过、目录项低于500,共享df不作账号配额。尚无正式训练,半小时调度继续,下次04:00 UTC;正式训练健康启动即取消416e9f54条件保持。 ++ ++ ++### 历史诊断完成及双模型混合历史v50(服务器2026-11-11 04:00 UTC) ++ ++HY v49和Kimodo v48均正常完成、预留释放;三臂完整轨迹有限、参考配对通过,自主对照逐元素精确复现HYv36最终及Kimodov20扩展。HY仅替换过去根朝向时961进度0.209/0.146仍不足,2327一seed超2;仅替换身体旋转时961仍反向-0.256/-0.108、接段10.516。与此前全部旋转替换相比,单通道修正不足,不能简单归因为某一独立通道。Kimodo替换平移使进度接近1但jerk5448.41,属于位置重置;替换旋转后姿态0.183rad、根位置0.383m,但接段6.417、jerk1256.41、行走失败7/24(自主对照8/24),仍不可部署。oracle臂不作为正式晋级依据。证据artifacts/recovery_hourly_20261111T0400_analysis.json,非FID,无Test1024。 ++ ++已启动双模型v50:/root/autodl-fs/pangxinlei/models/hfd-recovery-pilot-v50-20261111/{hymotion,kimodo},入口scripts/recovery_training_v50_{hymotion,kimodo}/supervise_pilot.py。分别以HYv36、Kimodov19为对照,从原正式best重置,仅pilot.predicted_probability从1降至0.75,使训练含25%真实历史窗口;其余训练参数、损失、学习率、链长、文本和0/125/250固定两种子验证均不变。诊断提示历史与姿态误差耦合,但并未证明混合历史训练有效,此轮检验是否有助于改善反馈误差。推理仍完全自主、无真实过去替换,无未来GT,无Test1024,原门槛保持;不延长失败断点。配置检查recovery_v50_checks.json。 ++ ++启动前GPU0/1空闲;两监督器及真实子进程存活、恢复检查初始化中,无错误,不宣称本轮恢复或效果已通过。HY2GiB、Kimodo12GiB预算实写fsync预留完成,8小时墙钟及容量/inode/日志缓存/原子断点保护保持,进程及源码SHA见recovery_v50_health.json和launch收据。清理已结束HYv49、Kimodov48重复索引共656.24MiB,旧进程退出、无打开句柄、新副本SHA一致后删除;正式权重、候选、完整断点及轨迹全部保留,清单recovery_cleanup_hourly_20261111T0400.json。本地约1.528GiB、共享inode超过36000,实写fsync检查通过、目录项低于500,共享df不作账号配额。无正式训练,半小时调度保持,下次04:30 UTC检查恢复和训练进度;正式训练健康启动即取消416e9f54条件不变。 ++ ++ ++### 双模型v50运行检查(服务器2026-11-11 04:30 UTC) ++ ++HY真实训练592427/监督器592244存活,审计98/250,尚未完成125验证;Kimodo真实训练592540/监督器592293存活,检查200步、审计205/250。双方两步完整断点恢复读回通过,optimizer/scheduler均为2,HY断点136888521字节、Kimodo3394253045字节;训练loss/梯度和已保存完整轨迹有限,初始逐元素复现原对照、参考/案例匹配。Kimodo日志已包含预测与真实历史链,混合训练路径生效;推理保持完全自主。无错误或watchdog停止,GPU0/1分别约20.1/20.3GiB,监督器正常。 ++ ++Kimodo125幅度0.764、接段2.262、姿态0.413rad、根位置0.520m(比原初始恶化约17.8%)、jerk590.96;2327两种子进度0.907/0.720,961为0.806/0.143,一组明显位移失败,未过门槛,不能把jerk下降视为已恢复质量。等待原250步有界预算的最终验证,不重复启动、不改动运行源码。证据artifacts/recovery_hourly_20261111T0430_analysis.json,非FID,无Test1024。 ++ ++本地约1.527GiB、共享空闲inode36041,本地/共享1MiB实写fsync通过,目录项低于500,共享df不作账号配额。活动预留、完整断点、缓存及轨迹保留,本轮无新核验可清理项,删除0字节,清单recovery_cleanup_hourly_20261111T0430.json。尚无正式训练,半小时调度继续,下次05:00 UTC检查中期/最终结果;正式训练健康启动即取消416e9f54规则不变。 ++ ++ ++### Kimodo v50完成与细分历史诊断v51(服务器2026-11-11 05:00 UTC) ++ ++HY v50进程592427/监督器592244存活,最新日志249/250、无错误、最终验证待返回。125幅度0.447低于0.5,姿态误差0.448rad较原初始恶化约20.3%,961进度-0.045/-0.001,尚不合格,不重复启动或覆盖运行源码。Kimodo v50正常完成250步、恢复通过、预留释放;双方训练及已保存初始/中期/最终完整轨迹有限,原初始精确复现、参考/案例匹配。Kimodo最终幅度0.658、接段2.562、姿态0.427rad、根位置0.409m、jerk372.26;2327进度0.373/1.001,961为0.771/0.221,2/4行走组合失败,混合历史未过门槛,不能以低jerk当恢复成功。证据artifacts/recovery_hourly_20261111T0500_analysis.json,非FID,无Test1024。 ++ ++Kimodo已启动冻结v51:/root/autodl-fs/pangxinlei/models/hfd-recovery-history-v51-20261111/kimodo,入口scripts/recovery_history_v51_kimodo/supervise_pilot.py。沿用v19step125、相同24validation×两种子及原正式48例基线,将v48全部旋转历史诊断细分为自主对照、仅替换过去根朝向3:6、仅替换过去身体旋转6:69;根平移保持预测,历史索引0..decision,未来占位不取GT。用于区分反馈误差来源,零训练更新,oracle臂不可部署且禁止正式晋级,自主对照必须精确复现。检查recovery_v51_checks.json。 ++ ++GPU1空闲后启动监督器593012及真实子进程,存活、初始化中、无错误;2GiB预算实写fsync预留通过,4小时墙钟/容量inode/日志缓存/原子保存保护保持,健康recovery_v51_health.json及launch源码SHA。HY GPU0继续。清理已结束Kimodo v50重复索引328.12MiB,旧进程退出、无打开句柄、新副本SHA一致后删除,所有模型权重、完整断点及轨迹保留,清单recovery_cleanup_hourly_20261111T0500.json。本地约1.527GiB、共享inode超过36000,本地/共享实写fsync通过、目录项低于500,共享df不作账号配额。尚无正式训练,半小时调度继续,下次05:30 UTC;正式训练健康启动即取消416e9f54规则不变。 ++ ++ ++### HY v50完成及标签过渡v52(服务器2026-11-11 05:30 UTC) ++ ++HY v50正常完成250步、无错误、预留释放、恢复通过;完整初始/125/250轨迹有限、参考配对、原初始精确复现。最终幅度0.557、接段1.887、姿态0.429rad(原初始+15.2%)、根位置0.450m、jerk358.13;2327进度1.316/1.838,961为-0.062/-0.006,姿态及行走失败,不延长混合历史配方。Kimodo v51真实593013/监督器593012存活,GPU1约17GiB并有计算活动,无错误;自主对照48条完整轨迹精确复现v20,参考一致且有限,后续诊断尚未完成,不预判oracle收益。证据artifacts/recovery_hourly_20261111T0530_analysis.json,非FID,无Test1024。 ++ ++HY启动v52有界250步:/root/autodl-fs/pangxinlei/models/hfd-recovery-pilot-v52-20261111/hymotion,入口scripts/recovery_training_v52/supervise_pilot.py。回到v36原正式best初始化对照,唯一语义变量标签SO3/高度过渡6帧缩至3帧,pilot和curriculum字段同步;其他训练参数、全预测历史、文本、损失、执行12帧及原验证保持。依据旋转历史误差耦合及混合历史无效,检验更快纠正监督姿态是否有效;存在接段恶化风险,原门槛不变,不预设改善。已有过渡实现不改,测试第三个未来帧起精确对齐目标、XY保持、有效掩码保持、输入不修改、真实历史恒等及数值有限通过,recovery_v52_checks.json。只修改训练标签,不增加未来GT输入,不改变自主推理,无Test1024。 ++ ++GPU0启动前空闲,监督器593503及真实子进程存活、恢复检查初始化中、无错误;2GiB预算实写fsync预留通过,8小时墙钟及容量/inode/日志缓存/原子断点保护保持,健康recovery_v52_health.json及launch源码SHA。不宣称新轮恢复通过。Kimodo原诊断继续。清理已结束HYv50重复索引328.12MiB,旧进程退出、无打开句柄和新副本SHA一致后删除,权重/断点/轨迹保留,清单recovery_cleanup_hourly_20261111T0530.json。本地约1.527GiB、共享inode超过35000,实写fsync通过、目录项低于500,共享df不作账号配额。无正式训练,半小时调度继续,下次06:00 UTC;正式训练健康启动立即取消416e9f54规则不变。 ++ ++ ++### HY v52继续及Kimodo标签过渡v53(服务器2026-11-11 06:00 UTC) ++ ++HY v52真实训练593626/监督器593503存活,124/250步、中期验证待返回;两步完整恢复读回通过,optimizer/scheduler均为2,断点136884553字节。训练数值及已存完整轨迹有限,原初始精确复现、参考配对,无错误。不重复启动或覆盖活动源码。Kimodo v51完成三臂、预留释放;48例完整轨迹有限、参考匹配、自主对照精确复现v20。自主行走失败8/24;仅替换过去根朝向失败10/24,幅度1.115、接段2.343、姿态0.387rad、根位置0.617m、jerk1000.12;仅替换身体旋转失败11/24,幅度1.251、接段6.582、姿态0.205rad、根位置0.549m、jerk1635.70。单通道oracle修正没有恢复自主行走,oracle不可晋级。证据artifacts/recovery_hourly_20261111T0600_analysis.json,非FID,无Test1024。 ++ ++Kimodo已启动v53:/root/autodl-fs/pangxinlei/models/hfd-recovery-pilot-v53-20261111/kimodo,入口scripts/recovery_training_v53/supervise_pilot.py。从原正式best初始化,以v19为对照,仅pilot.blend_frames从6改3,检验更快的标签姿态/高度纠正;此前12帧过渡和混合历史均未解决行走失败,因此不重复旧配方。250步有界、0/125/250固定两种子验证;训练、损失、学习率、推理、原门槛均不变。只改变标签不增加未来GT输入,存在接段恶化风险,不预设改善。准备时已移除误带的HY专属curriculum字段,未启动任何错误配置训练。三帧目标数值测试及单变量合同检查通过,recovery_v53_checks.json。 ++ ++GPU1空闲后监督器启动,真实子进程初始化中;12GiB预算实写fsync预留通过,8小时墙钟、磁盘/inode、日志缓存及原子断点保护保持。新轮恢复结果尚待返回,不宣称通过。健康及源码SHA见recovery_v53_health.json和launch收据。HY GPU0继续运行。仅清理已结束Kimodo v51重复索引328.12MiB,确认旧进程退出、无打开句柄、新副本SHA一致,权重/断点/完整轨迹保留,清单recovery_cleanup_hourly_20261111T0600.json。本地约1.526GiB,共享inode超过35000,实写fsync通过,目录项均低于500;共享df不作账号配额。无正式训练,半小时调度继续,下次06:30 UTC;任一正式训练健康启动即取消416e9f54的规则保持。 ++ ++ ++### 双模型标签过渡实验运行检查(服务器2026-11-11 06:30 UTC) ++ ++HY v52真实训练593626/监督器593503存活,审计190/250;Kimodo v53真实训练594424/监督器594263存活,125/250且中期验证尚待返回。双方两步完整断点恢复读回通过,optimizer/scheduler均为2,HY断点136884553字节、Kimodo3394249141字节。训练日志数值与已保存完整轨迹有限、参考和案例匹配、原初始精确复现,运行源码SHA与启动收据一致,无错误/监督器停止。GPU0/1占用约20.1/20.3GiB。HY日志full_predicted_history_fraction统计历史60帧是否全由模型生成,链启动阶段可低于1,不代表配置概率变为0.75;pilot.predicted_probability仍为1。 ++ ++HY125幅度0.538、接段2.124、姿态0.419675rad(原初始+12.7%)、根位置0.440253m、jerk368.13;2327两种子进度1.159/1.348,961为-0.031/0.001,姿态与行走不合格。Kimodo尚无中期结果,不能以初始结果判断本轮改善。双方继续已授权250步有界预算,不重复启动、不延长失败配方、不覆盖运行源码;下一次分析各自完成结果。证据artifacts/recovery_hourly_20261111T0630_analysis.json,非FID,无Test1024。 ++ ++本地约1.526GiB、共享空闲inode35910,本地及共享1MiB实写fsync检查通过,目录项均低于500,共享df不作账号配额。活动预留、缓存、完整断点及轨迹保留,无新增核验可删除项,本轮清理0字节,清单recovery_cleanup_hourly_20261111T0630.json。尚无正式训练,半小时调度保持,下次07:00 UTC;正式训练健康启动立即取消416e9f54的条件不变。 ++ ++ ++### 标签过渡实验完成及双模型v54(服务器2026-11-11 07:00 UTC) ++ ++HYv52与Kimodov53正常完成250步、旧进程退出、预留释放;恢复检查均通过,初始/125/250完整轨迹有限、参考配对、原初始精确复现。HY最终幅度0.609、接段2.452、姿态0.413428rad(原初始+11.1%)、根位置0.460978m、jerk363.53;2327进度1.552/1.887,961为-0.076/-0.070,仍不合格;较v36接段1.780恶化而姿态/行走未改善。Kimodo125幅度1.151、接段2.387、姿态0.44648rad、根位置0.43974m、jerk816.99,961一seed进度0.471低于0.5,不能放宽门槛;最终幅度1.210、接段2.652、姿态0.44417rad、根位置0.39969m、jerk514.15,2327进度0.814/0.455,961为0.794/0.195,两组失败。证据recovery_hourly_20261111T0700_analysis.json包括每例段内突变、根速度/jerk、坍缩及完整轨迹核对,非FID,无Test1024。 ++ ++启动双模型v54有界250步,根目录/root/autodl-fs/pangxinlei/models/hfd-recovery-pilot-v54-20261111,入口scripts/recovery_training_v54_{hymotion,kimodo}/supervise_pilot.py。HY回到v36原正式best初始化,仅旋转速度辅助权重0改0.05,原暖启动200步保持,检验姿态动态监督;不同于早期带根速度辅助/不同历史配方,不宣称该损失首次尝试。Kimodo回到v19原正式best,仅身体学习率曲线减半(1e-6/2e-7变5e-7/1e-7),根学习率2e-7/5e-8保持,检验身体更新过快与中后期退化的关系;区别于v30同时减半双模块。两者均非旧断点续训。其余配置、标签6帧、因果输入、推理、固定两种子验证及原验收保持;假设尚待验证。逐字段单变量比较及HY辅助损失有限/非零梯度检查通过,recovery_v54_checks.json。 ++ ++HY2GiB/Kimodo12GiB预算实写fsync预留通过后,在空闲GPU0/1分别启动监督器与真实子进程;新轮恢复检查初始化中,不宣称已通过。8小时墙钟、容量/inode、日志缓存、原子断点保护保持,健康和源码SHA见recovery_v54_health.json及launch收据。清理已结束HYv52/Kimodov53重复索引共656.24MiB,确认旧进程退出、无打开句柄、新副本SHA一致,权重/完整恢复断点/轨迹全部保留,清单recovery_cleanup_hourly_20261111T0700.json。本地约1.526GiB、共享inode超过35000,实写fsync检查通过,目录项低于500,共享df不作为账号配额。无正式训练,半小时调度继续,下次07:30 UTC;任一正式训练健康启动立即取消416e9f54规则不变。 ++ ++ ++### 双模型v54运行及Kimodo中期候选(服务器2026-11-11 07:30 UTC) ++ ++HY真实训练595667/监督器595433存活,审计90/250;Kimodo真实训练595769/监督器595479存活,检查184步、审计188/250。双方两步完整恢复读回通过,optimizer/scheduler均为2,HY断点136884553字节、Kimodo3394249205字节;训练数值有限,已保存完整轨迹有限、参考/案例匹配、原初始精确复现。源码SHA与启动收据一致,无错误/监督停止,GPU0/1占用约20.1/20.3GiB,日志持续更新。HY尚无125中期结果,不据初始结果推断本轮效果。 ++ ++Kimodo125在当前6案例×两种子上通过原门槛:幅度0.791598、接段2.058991、姿态0.423648rad、根位置0.314195m、jerk639.63;2327进度0.696/0.787,961为1.148/0.685,明显位移四组均在0.5–2。jerk仍高于原初始357.30,不能把小样本门槛通过等同于动作质量完全恢复。保留125候选,先完成当前250步有界预算;GPU1释放后优先对125进行相同24案例×两种子的扩展验证,与原正式48例基线及v19候选比较。未完成扩展验证前不启动正式训练,不重复占用GPU,不延长预算。证据recovery_hourly_20261111T0730_analysis.json包含完整轨迹核对及每例幅度、段内突变、速度、jerk、进度;非FID,无Test1024。 ++ ++本地约1.526GiB、共享空闲inode35821,本地及共享1MiB实写fsync检查通过、目录项低于500,共享df不作账号配额。活动预留、缓存、完整断点及候选保留,无新增核验可删除项,本轮清理0字节,清单recovery_cleanup_hourly_20261111T0730.json。尚无正式训练,半小时调度保持,下次08:00 UTC;正式训练健康启动立即取消416e9f54条件不变。 ++ ++ ++### 用户结束定时检查(2026-11-11T07:52:28.163857+00:00) ++ ++按用户明确指令,已执行 paseo heartbeat delete 416e9f54,CLI返回deleted,并核实调度记录不存在。任务状态已更新为cancelled_by_user,不再自动检查,不自动重建。此次操作仅取消心跳调度,不终止已授权训练,也不关闭监督器的容量/inode、断点保护及资源清理。 ++ ++ ++## 2026-11-11 三模型训练交付 ++ ++原训练工作区、代码差异、正式 latest/best、恢复候选、数据索引和特征缓存、环境锁、历史日志及学习曲线已上传并核验到以下私有仓库: ++ ++- [genmo](https://huggingface.co/starstone3/HFD-GENMO-SMPL-Training),commit `979e5130d6d2289582e8171830700d49dd0cdaf2`,清单 94 个文件。 ++- [hymotion](https://huggingface.co/starstone3/HFD-HY-Motion-SMPL-Training),commit `f858d43583e5897f850394fef886e7860f3a84f9`,清单 97 个文件。 ++- [kimodo](https://huggingface.co/starstone3/HFD-Kimodo-SMPLX-Training),commit `c9a71c47e7e16e9700779e986a607da6d920669e`,清单 107 个文件。 ++ ++三个模型均已实际短程续训并保存完整更新断点:HY 14000→14002(loss/LR/样本与原记录一致);Kimodo 98000→98002(loss/LR一致);GENMO 98000→98002(LR一致,loss略有差异,不能称为逐步精确一致)。GENMO 保存时触发容量受控停止,Kimodo 验收监督器异常和 HY 路径身份修复均已保留证据。原始数据和部分外部基础/文本/许可资产以固定来源与 SHA 清单交付,未声称已做全新环境重建。三个正式最新断点均到100K原预算,后续新阶段调度迁移尚未验证。 ++ ++Kimodo 原 best 仅权重;另交付与 best 参数逐位相同的同一步完整 last 对应文件。每仓库 README/acceptance-report/MANIFEST 记录具体命令、缺口和验证范围。心跳416e9f54已删除,不重建;既有训练未被终止。 ++ ++本地交付证据:`/root/autodl-fs/pangxinlei/handoffs/three-human-models-20261111/FINAL_DELIVERY.json`。核验时间:2026-11-11T08:35:41.057469+00:00。 ++ ++HY pilot54 已补交 step200 的最新完整断点与明确截止时间的日志快照;活动训练后续变化不包含在快照中。最终 MANIFEST 已从固定远端提交下载回读核验。最终核验时间:2026-11-11T08:39:04.787213+00:00。 ++ +diff --git a/SESSION_FORMAT.md b/SESSION_FORMAT.md +deleted file mode 100644 +index 03f5053..0000000 +--- a/SESSION_FORMAT.md ++++ /dev/null +@@ -1,82 +0,0 @@ +-# HFD session 格式与当前输入边界 +- +-更新:2026-10-21。正式默认为英文前景、`aligned_reference_context`、 +-`clock_policy: microphone`、`text_conditioning: none`:离线原文词对齐、最近30秒 +-已释放文本,不运行ASR或翻译。五模型的具体输入/输出、帧率与实现状态见 +-[当前协议](docs/HFD_HUMAN_BASELINE_INPUT_PROTOCOL.md),不再沿用旧中文翻译/ASR默认说明。 +- +-portable不携带原始tar、模型权重、缓存、视频或训练产物。HFD原始数据在 +-`/root/autodl-fs/pangxinlei/hfd_data`,按索引读取tar,保留分片,不全量解压。 +-原生训练session按配方动态组装并缓存在进程内;不是必须提前导出所有NPZ。 +-第三方读取已经组装的生产session。人体源动作也需要按计划组装,不能直接当完整session。 +- +-## 导出格式 +- +-`prepare_sessions.py`调用原生FullDuplexAudioCommandDataset;导出用于固定评测和短测, +-不改变任务、背景、监听/保持/过渡或事件时间。命令示例: +- +-```bash +-python prepare_sessions.py --config configs/sessions.yaml +-python prepare_sessions.py --config configs/sessions.yaml --set split=val +-python validate_sessions.py /data/hfd-sessions/train +-``` +- +-每个G1 session为NPZ及同名JSON: +- +-| 字段 | 内容 | +-| --- | --- | +-| `motion_q29_xyz_quat` | `[T,36]`:IsaacLab顺序29关节+根xyz+wxyz四元数;50Hz | +-| `audio` | 16kHz单声道混合PCM,保留原前景语音与背景 | +-| `control_labels` | 0=listen、1=motion、2=idle;监督/评测元数据,不作为输入 | +-| `turn_index_at_decision` | turn元数据,仅审计/展示,不作为输入 | +-| `valid_start_count` | 原生有效决策起点数;各模型再检查自己的窗口覆盖 | +-| `.json` | 源动作/音频、实际事件位置、关节/四元数约定、split、哈希 | +- +-`full_session`包含listen、motion、idle、过渡及事件前后的合法训练窗口。 +-`motion_only`仅保留作历史消融。源片段在session内部按配方组合,但训练窗口不跨session。 +-候选session数不是独立原始动作数,也不是已经缓存到磁盘的NPZ数量。 +- +-## 当前文本sidecar +- +-`data.asr_dir`是历史字段名;在当前协议下实际读取 +-`.reference.json`,不是`.asr.json`。动态训练则从同一离线对齐库和 +-当次session音频组件偏移直接生成相同记录,不需要在线ASR。 +- +-格式 `hfd-aligned-reference-v1` 包含NPZ/元数据哈希、源TTS片段、估计词末时间、 +-可用时间、首次释放时间、30秒上下文记录及异常fallback。已释放的完整句子可以输入; +-尚未释放的句尾、最终修正指令或整条GT全文不能提前提供。 +- +-动作时刻t对应麦克风t+0.2秒。词末加0.1秒guard,向上取整到麦克风 +-0.6/1.0/1.4…秒释放,模型只选择 `r_k <= t+0.2` 的记录。 +-0.6秒音频窗每0.4秒更新,是重叠窗口,不是每步新增0.6秒。 +-文本按释放点最近30秒滚动,过期可以为空;空文本不无限沿用旧命令。 +-无效词跨度在整个对应音频组件结束后释放。对齐误差和guard未完成逐词人工认证。 +- +-这是正确前景说话人的oracle转写实验:背景PCM保留,背景文字不输入。 +-旧 `causal_asr_context`、`stable_latest_sentence`、Whisper与中文→英文翻译仅用于 +-显式历史复现,不与当前结果混用。现有原始双语数据和历史产物不改写。 +- +-## 动作与坐标 +- +-原始NPZ不是MuJoCo/native关节顺序,不能直接切片当qpos。 +-`hfd_bridge/joints.py`按名称转换:TextOp原生23关节/57D,OMG原生29关节/125D。 +-当前TextOp未建模的腕关节固定在bootstrap姿态;指标只对其原生23关节计分。 +-根四元数在原始HFD文件为wxyz,进入各框架时显式重排。 +- +-GenMo读取配对人体trans、root_orient、pose_body、betas和sample_times, +-训练与推理不使用G1关节作为网络输入或目标。当前人体组装仍使用原生G1配方/ +-配对映射追踪源与时序;不能说整个数据构造程序与G1文件完全无关。 +-人体输入与retarget契约见[人体协议](docs/HFD_HUMAN_BASELINE_INPUT_PROTOCOL.md) +-及[评测转换说明](docs/HUMAN_RETARGET_EVALUATION.md)。 +- +-## 训练与连续评测 +- +-各模型使用自己的原生表示、归一化、损失和框架。OMG与GenMo训练提供GT历史, +-TextOp沿用原生分阶段历史课程。连续评测只初始化一次GT状态,之后反馈预测; +-真值历史诊断必须明确标注。GT事件仅用于选择/标注片段,不作为模型条件。 +- +-评测保留原混合音轨,将共同0.2秒microphone前置帧和bootstrap从动作评分中排除, +-不做DTW或事后根对齐。OMG/TextOp原生G1输出不做人体retarget;人体模型对完整轨迹 +-做一次retarget,并同时报告人体本身误差及人体GT重定向对照。 +-当前独立test未建立;val用于训练监测,不能称完全未见测试集。视频为运动学回放, +-不声称物理跟踪、端到端实时语音系统或语义任务成功率。 +diff --git a/UPSTREAM_LOCK.json b/UPSTREAM_LOCK.json +index 6c9873d..f30a85f 100644 +--- a/UPSTREAM_LOCK.json ++++ b/UPSTREAM_LOCK.json +@@ -2,11 +2,17 @@ + "external_repositories": { + "genmo": { + "url": "https://github.com/NVlabs/GENMO.git", +- "commit": "16bebf402d8893184249ee206d957b8248cd8310" ++ "commit": "16bebf402d8893184249ee206d957b8248cd8310", ++ "local_packaging_patch": "patches/genmo-numpy1264.patch", ++ "replay_numpy": "1.26.4" + }, + "kimodo": { + "url": "https://github.com/nv-tlabs/kimodo.git", + "commit": "e1c219c99b4b94ba48be1fb6a1cbc64ecee4c954" ++ }, ++ "hymotion": { ++ "url": "https://github.com/Tencent-Hunyuan/HY-Motion-1.0.git", ++ "commit": "4e426f5a1021cbcf7f375458c37b840ee7225229" + } + }, + "genmo_helper_assets": { +@@ -118,5 +124,16 @@ + "revision": "93ba4a7a9c66b3e1b7041e5f4ea038ef2418b948", + "metadata_manifest": "TRAINING_ASSETS.json", + "note": "Includes paired SMPL-X and all prior shards; do not bootstrap using the older pre-SMPL-X revision." ++ }, ++ "hfd_final_release": { ++ "data_repo": "Tinghong-Ye22/data_vqa_all_1025", ++ "data_revision": "bdca2fe7581c263e9ceb3a96204b314e199faddf", ++ "pipeline_repo": "Tinghong-Ye22/data_vqa_pipeline_1025", ++ "pipeline_revision": "9a2407adeed36a2645110390f1521304e36a2215", ++ "profile": "isolated_benchmark_1024_20261026", ++ "replay_arithmetic": "numpy2_float32_scalars", ++ "vendored_files": { ++ "upstream/hfd/src/psi/data/seed/third_party_dataset.py": "212fd08f50a6211caf3442d19772c8437b4de36ca0fb6de7f6865bfd621e84da" ++ } + } + } +diff --git a/configs/genmo.yaml b/configs/genmo.yaml +index 68fd000..a442cf6 100644 +--- a/configs/genmo.yaml ++++ b/configs/genmo.yaml +@@ -2,13 +2,15 @@ + model: genmo + genmo_root: /root/autodl-tmp/hfd-human-models/genmo + python: /root/autodl-tmp/envs/genmo/bin/python +-gpu: 1 +-output: /root/autodl-tmp/runs/hfd-genmo-english-v2 +-weights_output: /root/autodl-fs/pangxinlei/models/genmo/hfd-english-v2 ++gpu: 0 ++output: /root/autodl-tmp/runs/hfd-genmo-english-v2-split20261026-replay1264 ++weights_output: /root/autodl-fs/pangxinlei/models/genmo/hfd-english-v2-split20261026-replay1264 + init_checkpoint: /root/autodl-fs/pangxinlei/models/genmo/gem_smpl.ckpt + text_encoder: /root/autodl-fs/pangxinlei/models/genmo/t5-3b-local + body_models: /root/autodl-fs/pangxinlei/models/body_models + data: ++ replay_audit_status: /root/autodl-tmp/cache/hfd/genmo-replay1264/audit.json ++ require_replay_risk_gate: true + config: /root/autodl-tmp/hfd-baseline-portable/configs/sessions.yaml + reference_cache: /root/autodl-tmp/cache/hfd/reference-alignments.sqlite + language: en +@@ -43,6 +45,7 @@ strategy: + max_frames: 900 + seed: 12345 + training: ++ early_checkpoint_steps: [128, 256, 512, 1024] + # Total budget; review validation and closed-loop rollouts at 20K/40K/60K. + # Cosine T_max follows max_steps; resume must retain this original budget. + max_steps: 100000 +diff --git a/configs/genmo_legacy.yaml b/configs/genmo_legacy.yaml +index 57a7ea1..fa7f7aa 100644 +--- a/configs/genmo_legacy.yaml ++++ b/configs/genmo_legacy.yaml +@@ -3,8 +3,8 @@ model: genmo + genmo_root: /root/autodl-tmp/hfd-human-models/genmo + python: /root/autodl-tmp/envs/genmo/bin/python + gpu: 1 +-output: /root/autodl-tmp/runs/hfd-genmo-english-train +-weights_output: /root/autodl-fs/pangxinlei/models/genmo/hfd-english-train ++output: /root/autodl-tmp/runs/hfd-genmo-english-legacy-split20261026 ++weights_output: /root/autodl-fs/pangxinlei/models/genmo/hfd-english-legacy-split20261026 + init_checkpoint: /root/autodl-fs/pangxinlei/models/genmo/gem_smpl.ckpt + text_encoder: /root/autodl-fs/pangxinlei/models/genmo/t5-3b-local + body_models: /root/autodl-fs/pangxinlei/models/body_models +diff --git a/configs/kimodo.yaml b/configs/kimodo.yaml +index 5c27a61..4537176 100644 +--- a/configs/kimodo.yaml ++++ b/configs/kimodo.yaml +@@ -1,11 +1,11 @@ + # Full native English paired SMPL-X; text-only causal 30s context. + model: kimodo +-gpu: 0 ++gpu: 1 + model_dir: /root/autodl-fs/pangxinlei/models/kimodo/Kimodo-SMPLX-RP-v1 + session_config: /root/autodl-tmp/hfd-baseline-portable/configs/sessions.yaml + reference_cache: /root/autodl-tmp/cache/hfd/reference-alignments.sqlite +-output: /root/autodl-tmp/runs/hfd-kimodo-english-train +-weights_output: /root/autodl-tmp/models/kimodo/hfd-english-train ++output: /root/autodl-tmp/runs/hfd-kimodo-english-split20261026 ++weights_output: /root/autodl-fs/pangxinlei/models/kimodo/hfd-english-split20261026 + representation: + causal_root: true + future_frames: 60 +diff --git a/configs/omg.yaml b/configs/omg.yaml +index 9d973ce..c94cf3e 100644 +--- a/configs/omg.yaml ++++ b/configs/omg.yaml +@@ -1,4 +1,4 @@ +-# English foreground only; original indices/seed preserved. No ASR or translation. ++# English foreground only; isolated final split and recorded per-session seeds. + # Full-data English subset; long-run step budget still requires calibration. + schema_version: 1 + mode: train +@@ -7,9 +7,9 @@ framework: omg_lightning + paths: + data_disk: ${oc.env:HFD_DATA_DISK,/root/autodl-tmp} + upstream_root: upstream/omg +- session_root: ${paths.data_disk}/datasets/hfd-sessions ++ session_root: ${paths.data_disk}/datasets/hfd-sessions-split20261026 + cache_root: ${paths.data_disk}/cache +- output_dir: ${paths.data_disk}/runs/hfd-omg-english-train ++ output_dir: ${paths.data_disk}/runs/hfd-omg-english-split20261026 + runtime: + python: ${paths.data_disk}/envs/omg/bin/python + num_gpus: 1 +@@ -21,6 +21,8 @@ runtime: + distributed_timeout_seconds: 1800 + environment: + HFD_LOCAL_INDEX_CACHE: '1' ++ HFD_NATIVE_STARTUP_CACHE: '1' ++ HFD_STARTUP_INDEX_CACHE_DIR: ${paths.cache_root}/hfd/startup-indexes + data: + train_sessions: + - ${paths.session_root}/train/*.npz +@@ -28,7 +30,7 @@ data: + - ${paths.session_root}/val/*.npz + window_policy: full_session + max_windows: null +- cache_sessions: 512 ++ cache_sessions: 2048 + asr_dir: ${paths.session_root}/reference + asr_latency_seconds: 0.0 + audio_policy: hfd_past_2s +@@ -41,7 +43,7 @@ data: + prefetch_batches: 1 + native: + config: configs/sessions.yaml +- memory_mib: 2048 ++ memory_mib: 8192 + reference_cache: ${paths.cache_root}/hfd/reference-alignments.sqlite + reference_guard_seconds: 0.1 + refresh_batches: 256 +@@ -56,7 +58,7 @@ training: + precision: bf16-mixed + accumulate_grad_batches: 1 + gradient_clip_val: 1.0 +- save_every: 2000 ++ save_every: 256 + eval_every: 2000 + experiment: 50m + scheduler: +diff --git a/configs/omg_variants.yaml b/configs/omg_variants.yaml +index 87e9636..9d98296 100644 +--- a/configs/omg_variants.yaml ++++ b/configs/omg_variants.yaml +@@ -16,11 +16,11 @@ + experiment: 300m + checkpoint: checkpoints/paper/300m/sstep=055000.ckpt + sha256: df4ce9aeced84fd96a39583c7f6e2c5773b76f942cb43012c6c51dd9eb806806 +- batch_size: 32 +- accumulate_grad_batches: 16 ++ batch_size: 64 ++ accumulate_grad_batches: 8 + 500m: + experiment: 500m + checkpoint: checkpoints/paper/500m/sstep=050000.ckpt + sha256: f13bc58f07f104fcff67e35e79650def0fec5a6eb62a193fa1ae9f979fb911cf +- batch_size: 16 +- accumulate_grad_batches: 32 ++ batch_size: 32 ++ accumulate_grad_batches: 16 +diff --git a/configs/sessions.yaml b/configs/sessions.yaml +index c73045e..fa5652a 100644 +--- a/configs/sessions.yaml ++++ b/configs/sessions.yaml +@@ -2,7 +2,7 @@ + code_root: upstream/hfd + raw_root: /root/autodl-fs/pangxinlei/hfd_data + cache_root: ${oc.env:HFD_DATA_DISK,/root/autodl-tmp}/cache/hfd +-output_dir: ${oc.env:HFD_DATA_DISK,/root/autodl-tmp}/datasets/hfd-sessions/${split} ++output_dir: ${oc.env:HFD_DATA_DISK,/root/autodl-tmp}/datasets/hfd-sessions-split20261026/${split} + split: train + limit_sources: 0 + count: null +@@ -14,3 +14,11 @@ candidate_pool: false + recipe_overrides: + command_interruption_session_variants: 2 + include_third_party: true ++ ++# Final outer split selects sessions; upstream split is used only for replay. ++final_release: ++ profile: isolated_benchmark_1024_20261026 ++ revision: bdca2fe7581c263e9ceb3a96204b314e199faddf ++ directory: ${oc.env:HFD_DATA_DISK,/root/autodl-tmp}/cache/hfd/releases/bdca2fe7581c263e9ceb3a96204b314e199faddf/final_session_delivery ++ selection_index: ${oc.env:HFD_DATA_DISK,/root/autodl-tmp}/cache/hfd/releases/bdca2fe7581c263e9ceb3a96204b314e199faddf/motion-selection.sqlite ++ replay_support_root: ${oc.env:HFD_DATA_DISK,/root/autodl-tmp}/cache/hfd/releases/bdca2fe7581c263e9ceb3a96204b314e199faddf/runtime +diff --git a/configs/textop.yaml b/configs/textop.yaml +index 2235dab..49be7df 100644 +--- a/configs/textop.yaml ++++ b/configs/textop.yaml +@@ -1,4 +1,4 @@ +-# English foreground only; original indices/seed preserved. No ASR or translation. ++# English foreground only; isolated final split and recorded per-session seeds. + # Full-data English subset; long-run step budget still requires calibration. + schema_version: 1 + mode: train +@@ -7,9 +7,9 @@ framework: textop_native + paths: + data_disk: ${oc.env:HFD_DATA_DISK,/root/autodl-tmp} + upstream_root: upstream/textop +- session_root: ${paths.data_disk}/datasets/hfd-sessions ++ session_root: ${paths.data_disk}/datasets/hfd-sessions-split20261026 + cache_root: ${paths.data_disk}/cache +- output_dir: ${paths.data_disk}/runs/hfd-textop-english-train ++ output_dir: ${paths.data_disk}/runs/hfd-textop-english-split20261026 + runtime: + python: ${paths.data_disk}/envs/textop/bin/python + num_gpus: 1 +@@ -19,6 +19,8 @@ runtime: + offline: true + environment: + HFD_LOCAL_INDEX_CACHE: '1' ++ HFD_NATIVE_STARTUP_CACHE: '1' ++ HFD_STARTUP_INDEX_CACHE_DIR: ${paths.cache_root}/hfd/startup-indexes + data: + train_sessions: + - ${paths.session_root}/train/*.npz +@@ -26,7 +28,7 @@ data: + - ${paths.session_root}/val/*.npz + window_policy: full_session + max_windows: null +- cache_sessions: 512 ++ cache_sessions: 2048 + asr_dir: ${paths.session_root}/reference + asr_latency_seconds: 0.0 + audio_policy: released_asr +@@ -36,7 +38,7 @@ data: + backend: native_ondemand + native: + config: configs/sessions.yaml +- memory_mib: 2048 ++ memory_mib: 8192 + reference_cache: ${paths.cache_root}/hfd/reference-alignments.sqlite + reference_guard_seconds: 0.1 + refresh_batches: 64 +@@ -51,7 +53,7 @@ training: + precision: 32-true + accumulate_grad_batches: 1 + gradient_clip_val: 1.0 +- save_every: 2000 ++ save_every: 256 + eval_every: 2000 + num_primitive: 4 + stages: +diff --git a/docs/ASR_CONTEXT_OVERFIT.md b/docs/ASR_CONTEXT_OVERFIT.md +deleted file mode 100644 +index ce25532..0000000 +--- a/docs/ASR_CONTEXT_OVERFIT.md ++++ /dev/null +@@ -1,73 +0,0 @@ +-# OMG / TextOp 完整历史 ASR 过拟合测试 +- +-2026-10-18。结果目录:`/root/autodl-tmp/runs/archive/legacy-20261020/hfd-asr-context-overfit`。 +- +-本实验复用 `hfd-final/videos/omg_full_session_playable.mp4` 对应的原生双任务 session、混合音频和因果 Whisper sidecar。从相同官方初始权重重新训练,完整监督监听、执行、收尾动作;不把 GT 指令、任务阶段或控制标签作为输入。原来的 `hfd-final` 文件没有被覆盖。 +- +-## 输入变化 +- +-新增 `data.caption_policy: causal_asr_context`:每个释放点使用本次 ASR 对最近最多 30 秒已释放音频的**完整识别结果**;非空立即更新,空结果沿用上一份非空文本。不拼接历次重识别结果,不等两次一致,不用未来识别结果回填过去。这里保留的是 ASR 当前假设,最后一句可能尚未识别完整。 +- +-旧 `stable_latest_sentence` 只保留最近一句,并要求连续两次识别一致。本实验同时改变了文本覆盖范围和确认等待规则,因此不能把结果变化单独归因于历史句子的数量。本轮最初通过独立实验配置选择新策略;随后用户确认将其设为五个 baseline 的默认文本策略,见 [共同文本约定](BASELINE_TEXT_CONDITION.md)。旧 run 配置保留用于历史复现。 +- +-两个模型使用同一字符串,各自经过原生编码器。TextOp 保留 CLIP 的 77-token 限制和已有的超长文本最近后缀策略;OMG 保留 T5 的 50-token 限制。**本 session 的全部 37 个释放点,两个模型都没有截断。** 第二条指令出现后,输入可同时包含 `Taking an item from someone. Picking an item up from the floor.`,没有使用 metadata 修正 ASR。 +- +-`input_audit.json` 保存逐释放点的完整文本、旧稳定句、实际编码文本/令牌、截断标志;还核对 session 哈希、每个 ASR 音频切片哈希、释放边界与前缀不变性。ASR 仍是预计算的因果 sidecar,不计入实际识别计算延迟。 +- +-## 保持的训练与推理设置 +- +-| 项目 | OMG | TextOp | +-| --- | --- | --- | +-| 训练初始化 | 原官方 50M checkpoint | 原官方预训练 checkpoint | +-| 训练 seed | 42 | 42 | +-| 更新次数 | 1600 | 3200,阶段 `[800,1600,800]` | +-| batch / learning rate | 8 / 6e-5 | 4 / 1e-4 | +-| 额外条件 | 过去 2 秒已释放音频,10 帧动作历史 | 2 帧动作历史;不输入 PCM | +-| 动作预测 | 60 帧,30Hz | 8 帧,50Hz;训练链 4 个 primitive | +-| 推理 | DDIM 10 步、joint_one CFG、原生 async、24 帧余量、测得规划延迟 | CFG 5,每轮执行 8 帧 | +- +-保持原生网络、损失、优化器框架和动作表示。训练数据窗口与旧 run 相同:OMG 366 个窗口、TextOp 88 条连续链。只有开场使用 GT 初始化,随后完全反馈已执行的预测历史。TextOp 旧 run 保存的 `audio_policy: offline_target_aligned` 是未使用的历史字段,本轮按当前校验器写为 `released_asr`;两次都只使用因果 ASR。完整配置差异见 `experiment.json`。 +- +-OMG 的异步交接使用本次测量的规划耗时,旧新耗时可能不同,故误差差异包含实际调度影响,不是固定交接时刻的纯文本消融。主种子平均规划耗时旧版约 0.157 秒、新版约 0.175 秒。 +- +-## 结果 +- +-主视频沿用 seed=12345,没有挑选最优种子。关节 RMSE 排除开场初始化;OMG 统计 29 关节、TextOp 统计原生 23 关节,跨模型绝对值不能直接排名。 +- +-| 模型 | 旧稳定句关节 RMSE | 完整历史 ASR 关节 RMSE | 旧 / 新根坐标 RMSE | +-| --- | ---: | ---: | ---: | +-| OMG | 0.06191 rad | 0.07455 rad | 0.00844 / 0.01052 m | +-| TextOp | 0.29922 rad | 0.17019 rad | 0.04596 / 0.01603 m | +- +-| 模型 / 文本策略 | seed 12345 | seed 12346 | seed 12347 | 三种子均值 | +-| --- | ---: | ---: | ---: | ---: | +-| OMG 旧稳定句 | 0.06191 | 0.15385 | 0.05854 | 0.09143 | +-| OMG 完整历史 | 0.07455 | 0.13126 | 0.08651 | 0.09744 | +-| TextOp 旧稳定句 | 0.29922 | 0.36330 | 0.28990 | 0.31747 | +-| TextOp 完整历史 | 0.17019 | 0.30671 | 0.32361 | 0.26684 | +- +-TextOp 主种子的深弯腰峰值从旧版约 11.66 秒提前至 9.74 秒,接近 GT 的约 9.58 秒;但第一段动作仍有时序偏差,三个种子中仍有一个变差。其固定 GT 历史短窗口 RMSE 为 0.00976 rad,不能代替闭环结果。**完整历史文本对本条 session 的 TextOp 有帮助迹象,尚未实现稳定过拟合;OMG 没有整体改善。** +- +-三种子仅改变推理采样,每种文本策略只有一次训练、同一条训练 session;这不是独立训练重复或泛化评测。也不能据此认定模型已理解“旧指令被新指令替代”的对话语义。 +- +-## 产物与核验 +- +-- `videos/omg_full_session_playable.mp4`、`videos/textop_full_session_playable.mp4`:主种子视频,预测与原始 G1 GT 对照,保留混合音轨;均为运动学回放,没有物理 tracker。 +-- `comparison.png`、`video_contact_sheet.jpg`:旧新轨迹及视频帧对照。 +-- `summary.json`:全部采样种子、根误差、关节误差、跳变和规划耗时。 +-- `input_audit.json`、`experiment.json`、`verification.json`、`code_sha256.json`:输入、配置、检查与代码记录。 +-- 两个 `pipeline_check.json` 均通过有限 loss/梯度、参数更新、预期优化步数、原生 checkpoint 严格重载检查。28 项相关单元测试和两个数据合同检查通过。两个主视频完整解码及音频检查通过。 +-- OMG 冗余 checkpoint 先逐 tensor 比对最终权重与 global_step,确认一致后删除,保留最终 `last.ckpt` 与全部日志;记录在 `omg/checkpoint_cleanup.json`。 +- +-## 复现 +- +-以下训练输出目录应尚不存在;重跑时先覆盖 `paths.output_dir` 为新目录。 +- +-```bash +-/root/autodl-tmp/envs/omg/bin/python train.py --config configs/history/omg_asr_context_overfit.yaml +-/root/autodl-tmp/envs/textop/bin/python train.py --config configs/history/textop_asr_context_overfit.yaml +-/root/autodl-tmp/envs/omg/bin/python infer_overfit.py --run /root/autodl-tmp/runs/archive/legacy-20261020/hfd-asr-context-overfit/omg --continuous-only --skip-window-eval --task-seconds 0 14.8 +-/root/autodl-tmp/envs/textop/bin/python infer_overfit.py --run /root/autodl-tmp/runs/archive/legacy-20261020/hfd-asr-context-overfit/textop --continuous-only --task-seconds 0 14.8 +-/root/autodl-tmp/envs/omg/bin/python render_overfit.py --inference /root/autodl-tmp/runs/archive/legacy-20261020/hfd-asr-context-overfit/omg/inference --assets-root /root/autodl-tmp/models/OMG/OMG/assets --gpu 1 +-/root/autodl-tmp/envs/omg/bin/python render_overfit.py --inference /root/autodl-tmp/runs/archive/legacy-20261020/hfd-asr-context-overfit/textop/inference --assets-root /root/autodl-tmp/models/OMG/OMG/assets --gpu 1 +-``` +diff --git a/docs/BASELINE_TEXT_CONDITION.md b/docs/BASELINE_TEXT_CONDITION.md +deleted file mode 100644 +index 8d819f7..0000000 +--- a/docs/BASELINE_TEXT_CONDITION.md ++++ /dev/null +@@ -1,63 +0,0 @@ +-# 五个 baseline 的共同文本条件 +- +-2026-10-21 更新:按用户决定,当前训练仅选英文前景 session,使用 `aligned_reference_context`、`clock_policy: microphone`、`text_conditioning: none`。停止对中文样本做训练和在线翻译。原始中文数据及历史报告保留,中文文本缓存已清理。完整输入/输出与实现状态见[当前五模型协议](HFD_HUMAN_BASELINE_INPUT_PROTOCOL.md)。 +- +-模型接收截至当前释放点、最近 30 秒的英文前景文字。文字来自原始 TTS 文本的离线词级对齐,不再经过 ASR。词结束后加实验性的 0.1 秒 guard,再向上取整到 `0.6+0.4k` 秒释放;该 guard 不是已证明的对齐误差上界。无效词跨度整段结束后释放,禁止提前展示完整指令。背景 PCM 原样保留,背景文字不作为输入。这是 oracle 前景转写实验。 +- +-原生动作时刻 `t` 对应麦克风 `t+0.2s`;只允许使用已经到达的文字和音频。文字不跨 session 累积,不输入控制标签。编码器 token 预算仍有效:OMG T5 50 token、TextOp CLIP 77 token(75正文+两个边界)、GenMo T5-3B 50 token,超长时保留最近后缀。前端 30 秒不等于全部文字都能进入编码器。文本过期后允许为空,不沿用旧ASR“空结果保留最后非空”的行为。 +- +-| Baseline | 文本与额外输入 | 当前接入范围 | +-| --- | --- | --- | +-| OMG | 原生 T5-Base;过去2秒已释放混合音频+10帧动作历史@30Hz | 动态英文训练、连续验证与视频 | +-| TextOp | 原生 CLIP ViT-B/32;2帧57D动作历史@50Hz,不输入音频 | 动态英文300K训练、连续验证与视频 | +-| GenMo / GEM-SMPL | 原生 T5-3B;60帧/2秒人体历史@30Hz,当前音频关闭 | 动态原生英文人体训练、验证、完整resume、retarget视频;100K配置 | +-| HY-Motion | 原生 Qwen3 与 CLIP-L | 共同契约;完整 HFD 训练入口尚未实现 | +-| Kimodo | 冻结 LLM2Vec / Llama-3-8B,最多400正文token,mean pooling为1×4096 | 共同契约;固定英文配对子集训练/闭环/GMR入口已接通,见[Kimodo训练](KIMODO_HFD_TRAINING.md) | +- +-OMG/TextOp 的正式默认配置与 reference 配置均设置 `data.native.language: en`。原生索引、session ID 和 seed 不重排,仅将中文候选的采样权重置零;第三方已组装 session 同样按前景语言过滤。剩余 train **168,583**、val **5,109** 个候选 session。共享声明见 `configs/english_text_protocol.yaml`,详细时间戳与动态组装见 [动态 reference 训练说明](REFERENCE_DYNAMIC_TRAINING.md)。 +- +-GenMo只使用配对人体的原生英文子集:train **155,117**、val **4,719**,排除尚无人体真值的第三方生产session。全量session没有提前落盘;离线缓存的是源语音词对齐,原生session在训练时动态组装并做有上限的内存缓存。三模型的数据范围与音频可见性不同,不能仅按step或误差绝对值直接比较架构优劣。 +- +-GenMo 新缓存示例: +- +-```bash +-python cache_genmo_text.py --asr SESSION.reference.json --output NEW_TEXT_CACHE.pt --caption-policy aligned_reference_context --text-conditioning none +-``` +- +-旧 ASR 对照配置归档在 `configs/history/omg_bilingual_asr.yaml` 和 `configs/history/textop_bilingual_asr.yaml`。历史策略 `causal_asr_context` 使用混合音频的因果 ASR 及冻结 OPUS 中文→英文转换,曾有动作语义误译。旧配置、权重和日志不改写,但重跑中文历史实验需重建已删除的中文缓存。历史依据见 [ASR 过拟合报告](ASR_CONTEXT_OVERFIT.md) 和 [语义时钟修复](SEMANTICS_CLOCK_FIX.md)。 +- +-实际误差、吞吐和清理记录见 [纯英文短测报告](ENGLISH_ONLY_PROBE.md)。 +- +-## 历史双语盘点(清理前) +- +-### 本地中英文文本配对盘点(2026-10-19) +- +-原始 `audio/seed_tts/metadata.jsonl` 共 142,220 行,包含 5,220 个原始音频目录及 5,220 对不同的中英文完整指令;按字符串去重后为中文 5,060 条、英文 5,220 条。中英字段缺失均为 0。同一指令会对应多份动作和多个声音版本,142,220 行不能当作独立句子数。 +- +-加入停止、附和、犹豫、修正及拆分的指令片段后,已对齐的前景资产包含 9,313 种不同中文字符串和 9,140 种不同英文字符串,中文音频 61,860 个、英文音频 61,860 个。这不是动态 session 组合文本的数量,也不包含对背景说话者做完整中英配对的承诺。停止短语有 12 种中文及 12 种英文,附和短语各 15 种。 +- +-完整动作指令都有英文对应字段,但不能据此保证全部等义或一一对应: +- +-- 141 个不同中文字符串对应多个英文字符串,有些仅为措辞变体,也有“单手/双手”“轻/重”等实质区别。例如“左手持小发光物向左转身”同时对应英文的 one hand 和 two hands 描述。该计数不是已确认错误条数。 +-- 已知 `retro`(复古)被译为“后退”的问题涉及原始 594 行、30 个音频目录。原生规则会排除对应的 120 个中文声线版本;实际选定第三方 recipe 中也未发现这些错误中文资产。完整资产对齐库仍保留这些原始资源,不能将其存在误认为都进入训练。 +-- 中英指令独立拆分后,prefix/suffix 不一定语义对齐。例如“检查全身”对应 checking whole body,但中文 prefix 是“检查”,英文 prefix 是 checking whole;不能直接逐片替换。 +- +-现有完整配对英文可以作为复用与质量审查的起点,但在中文只说出一部分时不能提前释放完整英文。前面“转写正确”指文字与合成语音所用原文一致,不代表已完成全部中英等义人工校验。数据报告:`/root/autodl-tmp/runs/archive/legacy-20261020/reference-full-preparation-20261019/bilingual-inventory-audit.json`。 +- +-## 官方语言支持核查(2026-10-19) +- +-核查对象为 OMG(Omni-Modal Motion Generation)、TextOp、NVlabs GENMO/GEM、HY-Motion-1.0 和 Kimodo。区分动作模型的语言验证、文本编码器的语言说明、以及前端翻译/改写能力。“未找到中文声明”不等于证明中文绝对不可用,也不等于已证明只支持英文。 +- +-| 模型 | 官方可核查的说明 | 对中文输入的结论 | +-| --- | --- | --- | +-| OMG | [论文附录 B.3](https://arxiv.org/html/2606.10340v1) 明确要求 action 和 style 标注写成英文;C.3 使用冻结 T5-Base、最多 50 token。 | 查阅的论文和官方 README 未列中文支持或中文动作评测。英文标注规范与本地中文 token 坍缩相符。 | +-| TextOp | [论文 III-B/III-C](https://arxiv.org/html/2602.07439v1) 使用 BABEL 动作文本标注和 CLIP;[CLIP 官方模型卡](https://github.com/openai/CLIP/blob/main/model-card.md#out-of-scope-use-cases) 说明未专门训练或评估英文以外语言,限定英文用途。 | TextOp 未给完整语言清单,未找到中文动作评测或中文支持承诺。不能把中文可分词视为已支持中文语义。 | +-| GenMo / GEM-SMPL | [论文附录](https://arxiv.org/html/2505.01425v1) 使用 T5 文本编码;[官方配置](https://github.com/NVlabs/GENMO/blob/main/configs/exp/gem_smpl.yaml) 指定 `t5-3b`。 | 未找到动作模型语言清单、中文支持声明或中文动作评测。英文示例和 T5 选型是使用证据,不是所有其他语言被排除的证明。 | +-| HY-Motion-1.0 | [官方模型卡](https://huggingface.co/tencent/HY-Motion-1.0) 标注 English/Chinese,但其 Prompting Guide 要求使用英文,其他语言由 Text2MotionPrompter 改写。[论文第 3.3 节](https://arxiv.org/html/2512.23464v1) 的改写模块训练输入包含中英混合请求。 | 官方支持链路包含中文用户输入,但通过前置改写生成英文动作描述。未找到绕过改写后,核心动作模型直接读中文的明确质量保证或专门评测。 | +-| Kimodo | [技术报告第 4.2 节](https://research.nvidia.com/labs/sil/projects/kimodo/assets/kimodo_tech_report.pdf) 使用 LLM2Vec;[官方编码器配置](https://github.com/nv-tlabs/kimodo/blob/main/kimodo/scripts/run_text_encoder_server.py) 指向 Llama-3-8B-Instruct 的 LLM2Vec 适配器,其[模型卡](https://huggingface.co/McGill-NLP/LLM2Vec-Meta-Llama-3-8B-Instruct-mntp-supervised) 标注 English。 | Kimodo 技术报告和 SMPLX-RP-v1 模型卡未提供动作提示词语言清单或中文动作评测。Llama 能处理某些中文,不构成动作模型中文能力的验证。 | +- +-HY-Motion 的官方实现直接要求改写输出为英文,见 [prompt engineering 的系统提示](https://github.com/Tencent-Hunyuan/HY-Motion-1.0/blob/master/hymotion/prompt_engineering/model_constants.py)。因此模型卡的中英文标签与“动作生成推荐英文”应一起阅读,不能仅依据标签宣称核心网络原生中英双语,也不能把整套系统说成完全不能接收中文。 +- +-另需澄清:[T5-Base](https://huggingface.co/google-t5/t5-base) 和 [T5-3B](https://huggingface.co/google-t5/t5-3b) 的官方 NLP 语言列表为英语、法语、罗马尼亚语、德语,没有中文。这也不表示采用 T5 的动作模型已验证了法语、德语或罗马尼亚语动作生成;编码器能力与下游训练覆盖应分别报告。 +- +-对当前计划的含义:使用共享英文文本前端符合已查到的主要官方使用路径,尤其是 HY-Motion 已明确建议其他语言先改写。共享前端仍须验证翻译质量;直接中文训练或更换多语编码器应单列实验,不能以“官方已支持中文”跳过检查。本轮没有修改训练或推理配置。 +diff --git a/docs/ENGLISH_ONLY_PROBE.md b/docs/ENGLISH_ONLY_PROBE.md +deleted file mode 100644 +index 62771a8..0000000 +--- a/docs/ENGLISH_ONLY_PROBE.md ++++ /dev/null +@@ -1,90 +0,0 @@ +-# 纯英文训练切换与短测(2026-10-19) +- +-后续性能更新:已启用本地只读索引镜像并通过原生训练对照;旧正式训练已停止并清理,详见 [最新性能报告](LOCAL_INDEX_TRAINING.md)。下文耗时属于优化前记录。 +- +-清理说明:本文记录的一次性实验脚本已归档,复现前请按[脚本恢复说明](WORKSPACE_CLEANUP.md)取回;历史指标与日志保留。 +- +-已将当前 OMG/TextOp 正式默认、reference、smoke 和完整 session 配置切换为英文前景、离线词级对齐、最近 30 秒上下文、不翻译。GenMo 单 session 配置和新建文本缓存默认策略同步;HY-Motion/Kimodo 记录共同约定,尚无完整 HFD 训练适配。本轮实际训练及误差、吞吐验证仅覆盖 OMG/TextOp。没有启动 100K/300K 正式长训练。 +- +-英文路线可完成训练和严格重载,但本轮没有证据证明它一致改善动作质量。OMG 相对旧双语短测的关节误差下降 4.4%,TextOp 上升 1.8%;OMG 仍未优于官方初始模型。数据准备仍是主要效率瓶颈。 +- +-## 数据与清理 +- +-原生 train 155,117、val 4,719;选定第三方 train 13,466、val 390,合计 **168,583 / 5,109**。原始全量索引与 session ID、seed 不重排,中文候选权重置零;六种交互类型均保留。native train 覆盖 4,245 种英文指令和 99,099 个动作文件 ID,这两个计数不包含第三方的额外去重统计。 +- +-删除中文资产对齐条目 61,860 个、中文孤立对齐记录 27 个、非英文全文 recipe 13,695 个、翻译缓存 10,692 条、含中文文本的旧 ASR 缓存 109,323 条,以及 63 份中文 reference/ASR sidecar,约回收 0.8 GB。英文资产内容摘要删除前后相同,当前对齐库只含 61,860 个英文资产,完整性检查通过。旧英文 ASR 条目 100,509 个作为历史缓存保留,当前训练不使用。 +- +-原始 tar、原始双语 metadata、日志、历史报告、模型及英文缓存保留。recipe 的 ASCII 语言标签保留以过滤中文 ID,中文全文 recipe 已删除。背景 PCM 不变,过滤的是前景目标语言。清理记录:[cache-cleanup.json](/root/autodl-tmp/runs/archive/legacy-20261020/english-only-20261019/cache-cleanup.json)。 +- +-文字是原始 TTS 文本的 oracle 前景转写,按实际动态 session 拼接位置平移词时间戳;不需要在线 ASR。释放时钟为 `0.6+0.4k` 秒、词结束后加实验性 0.1 秒 guard,保留最近 30 秒。英文中 6,608/61,860(10.68%)个资产词边界无效,整段结束后释放;该 guard 未证明能覆盖所有对齐误差。OMG 输入过去 2 秒音频、已释放文本及动作历史;TextOp 输入已释放文本及动作历史。两者均未加载 ASR 或翻译模型。 +- +-## 误差对照 +- +-两次新训练均从官方权重初始化。OMG batch64、512 步、lr6e-5、bf16;TextOp batch64、768 步、lr1e-4、FP32,三阶段各 256 步。对照旧双语试验采用相同训练步数、batch、学习率和 seed42;采样范围按实验目的改变为英文。不是在旧双语 checkpoint 上续训。 +- +-在同一组三个英文 val session(普通指令、打断、急停)上,对官方初始、旧双语、纯英文三个 checkpoint 做连续闭环回放。各 checkpoint 使用完全相同的原始音频、oracle 文本、释放时钟、一次 GT bootstrap、推理参数和 seed12345。OMG 固定规划延迟 4 帧、原生异步规划;TextOp 原生 CFG5、预测历史。计分排除 bootstrap;OMG 计 29 关节,TextOp 按其原生覆盖关节计分,不能横向比较两模型绝对值。 +- +-以下为三个 session 各自 RMSE 的等权平均,单位 rad: +- +-| 模型 | 官方初始 | 旧双语 | 纯英文 | 纯英文相对双语 | +-| --- | ---: | ---: | ---: | ---: | +-| OMG | 0.4043 | 0.4257 | 0.4070 | -4.4% | +-| TEXTOP | 0.4848 | 0.4117 | 0.4193 | +1.8% | +- +-根节点 xyz RMSE 的相同等权平均,单位 m: +- +-| 模型 | 官方初始 | 旧双语 | 纯英文 | +-| --- | ---: | ---: | ---: | +-| OMG | 0.3002 | 1.5859 | 0.7177 | +-| TEXTOP | 0.7033 | 0.5317 | 0.6196 | +- +-逐 session 的关节 RMSE: +- +-| 模型 | 交互 | 官方初始 | 旧双语 | 纯英文 | +-| --- | --- | ---: | ---: | ---: | +-| OMG | turn-taking | 0.3616 | 0.3794 | 0.3654 | +-| OMG | interruption | 0.3404 | 0.3319 | 0.3097 | +-| OMG | emergency-stop | 0.5109 | 0.5657 | 0.5459 | +-| TEXTOP | turn-taking | 0.3750 | 0.3913 | 0.3873 | +-| TEXTOP | interruption | 0.3582 | 0.3401 | 0.3461 | +-| TEXTOP | emergency-stop | 0.7211 | 0.5037 | 0.5244 | +- +-OMG 新版比旧双语更接近 GT,但官方初始关节误差仍略低,根节点误差也更低。TextOp 新版比双语略差,虽然优于官方初始。这只是三个样本、单一随机种子的短测,没有统计显著性保证;单一 GT 的运动 RMSE 也不是指令执行成功率,不能据此宣称全双工能力改善。 +- +-OMG 英文固定验证损失为 0.05414 → 0.05118 → 0.04272 → 0.04288(128/256/384/512 步)。TextOp 为 0.13086、0.13235、0.13669、0.14283、0.13936、0.12966(每 128 步);阶段历史分布在 256/512 步改变,不能把整段曲线当作同一目标的收敛曲线。旧双语 native validation 含中文,新版仅英文,二者验证损失不能直接比较优劣。 +- +-## 训练效率 +- +-OMG 固定 GPU1,TextOp 固定 GPU0。另从官方初始权重做 batch512 测速:OMG64步、TextOp96步,分别与 batch64 对照处理相同的 32,768 / 49,152 个训练实例;优化器更新次数不同,batch512 权重不用于上面的质量结论。TextOp 的“实例”按每次更新的 primitive 样本计,OMG 为动作窗,不能拿两模型的样本/秒直接比架构效率。 +- +-“池内”是排除超过 2 秒的冷池/验证间隔后的连续步间隔中位数,包含热数据准备和训练。“训练阶段”包含组池、验证和周期保存;“总墙钟”还包括索引/模型初始化和结束检查。batch512 复用了前一阶段的 CPU 原始索引,因此它的总墙钟不能与含冷索引构建的 batch64 直接当作正式加速比。 +- +-| 模型 | Batch | 步数 | 池内秒/步 | 池内实例/秒 | 训练阶段秒 | 训练阶段实例/秒 | 总墙钟秒 | +-| --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: | +-| OMG | 64 | 512 | 0.118 | 543 | 547.4 | 59.9 | 729.1 | +-| OMG | 512 | 64 | 0.535 | 957 | 385.6 | 85.0 | 397.9 | +-| TEXTOP | 64 | 768 | 0.083 | 775 | 300.9 | 163.3 | 478.8 | +-| TEXTOP | 512 | 96 | 0.149 | 3430 | 399.6 | 123.0 | 410.3 | +- +-OMG batch512 的首池准备 341.6 秒,TextOp 为 374.2 秒。训练阶段分别约 88.6% / 93.6% 时间用于该首池准备。batch64 OMG 8 池共 431.5 秒,TextOp 3 池共 169.4 秒。实际观察到共享盘 tar 读取等待,GPU 在组池时基本空闲。两秒采样、排除各卡训练后的推理后,包含组池/验证的平均 GPU 利用率仅约 GPU1 3.1%、GPU0 3.9%;采样从索引初始化后开始,不能当成全程精确 GPU 工时。 +- +-batch512 峰值 PyTorch 分配显存 OMG **16.21 GiB**、TextOp **2.62 GiB**,均可在当前卡运行。TextOp 的 batch512 短测没有覆盖完整 256 更新的数据池刷新周期,只用一个池跑了96步;batch64 跑了三个64-session池,故新鲜 session 数不同。大 batch 的池内吞吐高,但首池成本和数据复用程度必须一起看。 +- +-OMG 两阶段启用逐步训练健康检查;TextOp matched 为匹配旧试验关闭额外健康 hook,batch512 开启,因此不能把全部吞吐差异归因于 batch。历史双语 OMG 还多准备了一个未使用的尾池,当前已修复;其 1063.8 秒历史 fit 时间与本轮 547.4 秒的差异包含此独立修复及缓存状态差异,不是单独“去中文”的严格加速比。 +- +-当前瓶颈仍在数据准备。后续效率工作应优先测后台组池、原始资产/动作特征缓存及共享盘读取,再决定正式步数和刷新频率;仅增大 batch 或把旧 100K/300K 步原样启动,不能保证合理训练时长或动作质量。 +- +-## 验证与产物 +- +-60 项相关单元测试、两套正式英文配置 dry-run 通过。OMG matched/batch512 分别记录到 512/64 次真实优化器更新,有限损失/梯度、参数变化和严格重载通过;TextOp batch512 记录96次真实更新并通过同样检查。TextOp matched 记录768步有限损失/梯度范数,108个 denoiser 状态张量变化且有限,实际回放完成严格加载;该阶段没有单独 hook 计数优化器。没有训练结束后的多余冷池。 +- +-- [完整指标与逐样本误差](/root/autodl-tmp/runs/archive/legacy-20261020/english-only-20261019/summary.json) +-- [英文验证曲线](/root/autodl-tmp/runs/archive/legacy-20261020/english-only-20261019/english-validation.png) +-- [GPU 利用率](/root/autodl-tmp/runs/archive/legacy-20261020/english-only-20261019/gpu-utilization.png) +-- [最终数据和前端检查](/root/autodl-tmp/runs/archive/legacy-20261020/english-only-20261019/final-audit.json) +-- [OMG 急停视频](/root/autodl-tmp/runs/archive/legacy-20261020/english-only-20261019/evaluation/omg/emergency-stop/english/inference/complete_task.mp4) +-- [TextOp 急停视频](/root/autodl-tmp/runs/archive/legacy-20261020/english-only-20261019/evaluation/textop/emergency-stop/english/inference/complete_task.mp4) +- +-视频保留原始混合音频,显示预测与 GT、麦克风时钟和实际释放的文本;两份均已完整解码检查。OMG/TextOp 直接输出原生 G1,无需 SMPL-X retarget。GenMo 等人体模型仍须遵循现有 SMPL-X→G1 retarget 与视频评测链路,本轮没有宣称已完成它们的新训练。 +- +-运行配置、日志及 checkpoint 位于 `/root/autodl-tmp/runs/archive/legacy-20261020/english-only-20261019/`。脚本为 `tools/run_english_training_probe.py`、`tools/compare_english_checkpoints.py`、`tools/summarize_english_probe.py`。已更新正式默认配置,但正式长训练未启动;100K/300K 仍属于待校准预算。 +diff --git a/docs/FORMAL_TRAINING_READINESS.md b/docs/FORMAL_TRAINING_READINESS.md +deleted file mode 100644 +index 426a887..0000000 +--- a/docs/FORMAL_TRAINING_READINESS.md ++++ /dev/null +@@ -1,117 +0,0 @@ +-# 正式训练准备检查 +- +-后续性能更新:已启用本地只读索引镜像并通过原生训练对照;旧正式训练已停止并清理,详见 [最新性能报告](LOCAL_INDEX_TRAINING.md)。下文耗时属于优化前记录。 +- +-2026-10-19 英文路线更新:正式默认与 reference 配置均仅采样英文前景 session,离线对齐、无 ASR/翻译;train 168,583、val 5,109。中文缓存已清理。当前有步数上限的双卡误差与吞吐短测见 [纯英文短测报告](ENGLISH_ONLY_PROBE.md),100K/300K 正式训练未启动,旧预算仍需校准。 +- +-2026-10-19 较早更新:离线前景转写对齐和动态 session 读取已接入独立 reference 配置,详见 [离线对齐与动态训练](REFERENCE_DYNAMIC_TRAINING.md)。当前是一卡 OMG、一卡 TextOp;下文的双卡 OMG、TextOp batch64 和 ASR 预算属于历史记录。中文翻译质量、TextOp 连续回放和新版训练预算仍需验证,不能直接沿用旧文中的启动结论。 +- +-2026-10-17:OMG 已完成一轮 100k 更新,训练循环 108.55 小时。实际结果、数据范围、性能瓶颈和新配置见 [正式训练复查](OMG_TRAINING_REVIEW.md)。当前增加 `data.batched_motion=true`、`training.fixed_validation=true` 和 `training.save_best=true`;早停默认关闭,可用 `training.early_stopping.enabled=true` 启用。新增最佳 checkpoint 约多占 1 GiB。下面 10 月 10 日的短测和资源规划为历史准备记录,不应当作此次完整训练的实测速度。 +- +-同日后续:OMG 默认 `data.prefetch_batches=1`。双卡 64 步同步/异步分段对照,排除边界两步后,池内平均每步由 1.07 秒降到 0.62 秒。该数值不包含下一池刷新和 ASR,也不是整轮加速承诺。主机内存预算需另留每卡一个 CPU batch;`num_workers=0` 保持不变。 +- +-2026-10-10 更新:419 个原始 tar 已全部下载且 SHA256 校验通过;正式配置改为 OMG 双卡、TextOP 单卡,每次只运行一个模型。全量原生 session 配方已通过索引构造,原有两模型 batch=2 的短测已通过。OMG 每卡 batch 512 的显存探针已通过;低吞吐正式运行已按用户要求终止并删除,当前仅进行有明确步数上限的性能短测。 +- +-## 已确认的配方修正 +- +-用户确认将 `configs/sessions.yaml` 中的 `command_interruption_session_variants` 从 1 调到 2。原值在当前数据版本只能产生 36,836 个打断候选,低于原生配比要求的 39,743;修改后有 73,672 个候选,原生生成、筛选和最长 session 剔除流程均通过。 +- +-不复制 session,不跳过最终配比,不改变原生任务类型目标。增加 variant 可能改变后续随机组装结果,应把这个 override 同步记录到论文 HFD 和 baseline 数据配方中。原始 recipe 源码保持不变,覆盖参数与源码哈希保存在检查报告中。 +- +-| 类型 | train | val | +-| --- | ---: | ---: | +-| turn-taking | 158,812 | 4,539 | +-| interruption | 39,703 | 1,134 | +-| emergency-stop | 26,468 | 756 | +-| backchannel | 39,703 | 1,134 | +-| hesitation | 34,071 | 1,016 | +-| pause | 23,486 | 693 | +-| 合计 | 322,243 | 9,272 | +- +-这些是原生筛选及每类最长 0.1% 剔除后的计数。interrupt/control/backchannel 的目标按基础任务 session 数定义;hesitation/pause 按窗口数加入,因此不能直接拿最终六类 session 百分比验证全部配比。train/val 沿用原生动作来源分组划分。原生长度估计对应约 3,304.3/94.6 小时,实际导出长度会随 idle、打断时刻等而变化。 +- +-## 全量数据准备的实际限制 +- +-- 检查开始时,两张 RTX 5090 32 GB 空闲;TextOP、OMG、CLIP、T5、Whisper 权重均已在本机。 +-- 原始分片共 419 个,完整 tar 约 49.8 GiB,另有索引和 metadata。2026-10-09 已有 419 个完整分片,大小和 SHA256 均通过清单校验,无需继续下载。原始目录保持在共享盘。后续使用 `tools/download_hfd_tars.py` 校验 SHA256、跳过完整文件并续传 `.partial`;命令及空间要求见 [迁移说明](MIGRATION.md)。实际 hesitation 样本仍读取 v1 分片中的资源,不能仅凭配方名为 v2 就忽略这些文件。 +-- 按原生估计长度,在每个交互类型中选一个中位附近 session,六个均完成导出、`validate_sessions.py` 检查和完整因果 ASR;这些仅用于容量与吞吐测量,没有进入训练。 +-- 六个样本的压缩 NPZ 约为每音频秒 37,765 字节。乘以原生总估计时长,train+val 约 **430 GiB**,还不含 JSON、ASR 和 checkpoint。此为旧完整导出方案的六样本推算,当前按需方案无需分配这部分空间;收尾时数据盘剩余约 44 GiB。 +-- 逐窗口 ASR 在五个热身后样本上约 0.048–0.091 秒/窗口。六个样本共 483 个因果窗口,逐条总耗时 57.9 秒,batch=32 总耗时 7.45 秒(约 7.8 倍);两个测量的热身条件并不完全一致,不能作为严格性能基准。16/483 个窗口的文本有差异,PCM 和截止时刻一致。保持混合音频,不用 GT 替代 ASR。 +-- 旧 NPZ 加载器只有 4-session 缓存且 OMG 保留所有周期 checkpoint。当前正式配置改用按需读取、OMG 512-session 特征缓存、每个 split 2 GiB 数组预算和最近 3 个周期 checkpoint,旧方案只保留用于已有小样本复现。 +- +-已选择按需组装方案:`configs/textop.yaml` 和 `configs/omg.yaml` 使用 `data.backend: native_ondemand`。每个 batch 槽位按合法动作窗口的长度加权抽 session,再从 session 内均匀取窗口;OMG 当前每卡 batch 为 512,正式值每 256 个 batch 刷新一次 session 池。相邻 batch 有相关性,但保持合法窗口的边际采样概率;每个 batch 仍会在已保留 session 内重新抽动作点。较长刷新间隔用于摊薄原生组装和 ASR 成本,若需更强 session 随机性可改回 64。OMG 缓存覆盖一个完整槽位池;批量音频特征计算保留逐窗口 35 维特征定义,独立对照容差为 rtol=atol=1e-6(存在浮点运算顺序差异)。原生配比定义的是候选 session 集合,不保证每个 batch 都满足该比例。增强使用固定 seed=42,缓存失效后重组同一 session 不改变其内容,也不每个 epoch 重新随机增强。ASR 结果以压缩 SQLite 缓存,Whisper 使用 batch=32 的因果窗口;批处理参数和实现哈希记录在 run manifest,实际识别结果位于缓存。它与逐条识别不是字节级等价,做论文复现时应一并保存匹配的缓存。 +- +-OMG 和 TextOP 的按需 smoke 都已通过(各 4 次更新、有限梯度/参数变化、验证读取、严格 checkpoint 加载)。六个样本的动作/PCM 与旧导出逐元素一致。少数窗口文字因 batch/半精度数值产生差异,这是已记录的预处理口径变化,不是未来音频或 GT 输入。 +- +-按需训练不生成数百 GiB 的派生 NPZ;当前数据盘只存 raw-range cache、ASR SQLite、日志和模型输出。建议先补齐原始 tar,再长时间训练;当前代码支持缺失文件的 range 回退,但受网络可用性与重试影响,不能按本地数据速度估算。`runtime.offline: true` 控制模型权重加载,不会关闭这个 raw tar 网络回退。 +- +-## 正式训练命令和资源预算 +- +-2026-10-09 配置改为**每次只训练一个模型**:OMG 使用 GPU 0、1 的 DDP;TextOP 沿用原生单卡训练器。OMG 的按需加载器在每个 rank 使用独立窗口采样随机流、对应 GPU 的 Whisper,以及分别命名的采样记录;session 构造种子保持一致。两个 rank 从完整原生集合独立有放回抽样,允许偶然抽到同一个 session,不是两卡复制同一个 batch。每个 rank 保留完整索引,模型梯度由 DDP 同步。 +- +-先在仓库目录检查配置(不启动训练): +- +-```bash +-cd /root/autodl-tmp/hfd-baseline-portable +-export HFD_DATA_DISK=/root/autodl-tmp +-/root/autodl-tmp/envs/omg/bin/python train.py --config configs/omg.yaml --dry-run +-/root/autodl-tmp/envs/textop/bin/python train.py --config configs/textop.yaml --dry-run +-``` +- +-在 screen 中运行 OMG: +- +-```bash +-screen -S hfd-omg +-cd /root/autodl-tmp/hfd-baseline-portable +-/root/autodl-tmp/envs/omg/bin/python -u train.py --config configs/omg.yaml +-``` +- +-OMG 结束后,另开一个 screen 运行 TextOP: +- +-```bash +-screen -S hfd-textop +-cd /root/autodl-tmp/hfd-baseline-portable +-/root/autodl-tmp/envs/textop/bin/python -u train.py --config configs/textop.yaml +-``` +- +-已经在 screen 内时只执行 `cd` 和 Python 命令,不嵌套创建 screen。`Ctrl+A` 然后 `D` 脱离,`screen -r hfd-omg` 或 `screen -r hfd-textop` 恢复。不要同时执行这两个训练命令。终端直接显示数据准备和训练进度,无需 `nohup`、`tee` 或输出重定向;run 中仍保留配置、checkpoint 和小型结构化指标,供复查实验。 +- +-| 项目 | TextOP | OMG | +-| --- | --- | --- | +-| GPU | 1 张 RTX 5090 32 GB,GPU 0 | 2 张 RTX 5090 32 GB,GPU 0、1 | +-| batch | 64 | 每卡 512,全局 1024 | +-| 精度 | FP32 | bf16 mixed | +-| 更新步数 | 300,000;三阶段各 100,000 步 | 100,000 | +-| 学习率 | 1e-4 | 6e-5 | +-| 数据读取 | 按需原生组装,完整 train/val | 同左,DDP 独立采样 | +-| ASR | Whisper tiny,batch 32,共用训练卡 | 每个 rank 在对应训练卡使用 Whisper tiny,batch 32 | +-| CPU | 4 线程,num_workers=0 | 每 rank 4 线程,num_workers=0;音频特征按 batch 向量化 | +-| 输出目录 | `/root/autodl-tmp/runs/hfd-textop-train` | `/root/autodl-tmp/runs/archive/legacy-20261020/hfd-omg-train` | +-| 保存策略 | 每 20,000 步,保留最新 3 份 | 每 2,000 步,保留最新 3 份 | +- +-输出目录必须尚不存在;启动新实验可加 `--set paths.output_dir=/root/autodl-tmp/runs/新的名称`。OMG 入口自动调用 torchrun,无需手工加启动器。默认沿用本机可用的 Gloo,DDP 超时 1,800 秒,给按需组装和 ASR 留余量。OMG 的 `training.batch_size` 是每卡 batch;当前每卡 512、总 batch 1024。真实 forward/backward 探针峰值约为 21.9 GiB/卡(另有 Whisper 常驻);没有把 640 作为正式值,因为 640 已达到约 27.3 GiB,给优化器和 DDP 临时张量的余量较小。改回 OMG 单卡时加 `--set runtime.num_gpus=1 'runtime.gpu_ids=[0]' training.batch_size=512`;保持学习率 6e-5 以便与 baseline 口径一致。 +- +-两模型都对完整 session 的动作窗口进行监督,保留 listen/idle 对应关节动作,不输入控制标签。OMG 使用过去 2 秒音频+因果 ASR;TextOP 使用因果 ASR。此次双卡改动不改变上述输入或数据配方。 +- +-训练是按优化更新步数停止的随机窗口采样,并不是逐个遍历全部 session 的 epoch:TextOP 每次取一个 batch 后依次对 4 个 primitive 做 4 次更新,共取约 75,000 组 batch;OMG 共取 100,000 组双卡 batch。OMG 槽位每 256 个 batch 刷新一次,TextOP 每 64 个 batch 刷新;全量集合可被抽样,但有限步数不保证每个 session 都被访问。论文应记录这一训练预算。 +- +-资源规划:OMG 两进程建议预留 8 CPU 核、32 GiB 主机 RAM;TextOP 建议至少 4 核、16 GiB RAM。OMG 每个 rank 的 train/val 各有 2 GiB 数组 LRU,另有 512-session 特征缓存和完整索引;2 GiB 不是整个进程内存上限。仅数据回放进程的 RSS 约 3.9 GiB;这不包含训练模型、Whisper、验证集与优化器,不能把它当作完整训练的 RAM 占用。ASR SQLite 共用 2 GiB 有效记录预算;原始 tar 齐全时无需网络 range 回退。无需生成约 430 GiB 的全量派生 NPZ。 +- +-数据盘在已有权重和环境之外建议预留 15–20 GiB。TextOP 输出约 1.2 GiB;OMG 最新 3 个周期 checkpoint 加最终 last.ckpt 约 4 GiB,多卡不会各保存一套 checkpoint。其余用于缓存、SQLite 页/WAL 和保存瞬间空间。当前约 44 GiB 空闲足够这份规划。原始 tar 约 49.8 GiB+索引保存在指定共享盘,已下载完成;共享盘 df 的整体空闲仍不能视作账号配额。 +- +-本轮已终止的 OMG batch=512 实跑暴露出旧配置的问题:64-session 缓存无法覆盖一个 512-session 槽位池,普通 batch 也会重复构造 session 及运动特征;每 64 个 batch 重新抽 session 时还需要处理 ASR,导致运行近 10.5 小时只完成约 500 步。配置已改为 512-session 缓存、每 256 个 batch 刷新一次。此前 session 回放基准(ASR 命中磁盘缓存)显示,缓存 512、数组预算 2 GiB 时首次准备约 50 秒;另一轮文件读取较冷的回放约 188 秒。后者包含不同缓存状态,不能把首次准备时间差全归因于代码优化。热缓存下 batch 准备约 1.7 秒;同样缓存容量、旧逐窗口音频实现的整个 batch 准备约 7.5 秒,改为 batch 向量化后降至约 1.7 秒。双卡真实 train mode 短测中,模型计算约 0.43 秒/步,缓存命中时完整 optimizer step 约 2.0 秒,首次装载一个 512-session 池约 72 秒;这些测试不包含完整 refresh 周期。按 100,000 步估算,纯缓存命中下约 56 小时,实际时间还取决于每 256 步的刷新、ASR 命中率和共享盘读取速度。正式重启前仍应做 3–5 个 optimizer step 的端到端短测,再估计总耗时。 +- +-OMG 每 2,000 步每 rank 抽取 16 个验证 batch(双卡共 32 个),TextOP 每 2,000 步运行原生 eval_steps=12(3 组四 primitive batch)。这是训练监测,不是全量独立测试集评测。统一入口目前仍未提供恢复优化器及步数的 resume;`weights.init_checkpoint` 仅表示初始化。断点恢复和按需 run 的论文评测入口是后续工程项,不能把重新加载初始化权重当作续训。 +- +-## 可复查的证据 +- +-小型 JSON 结果在 [formal_training_readiness.json](../reports/formal_training_readiness.json)。本机测量样本和日志位于 `/root/autodl-tmp/cache/hfd/readiness/`,不进入 Git,也不混入 `runs/hfd-final`。 +- +-重建索引检查(不导出 session、不启动训练): +- +-```bash +-/root/autodl-tmp/envs/omg/bin/python prepare_sessions.py --config configs/sessions.yaml --audit-only --report /root/autodl-tmp/cache/hfd/train_audit.json +-/root/autodl-tmp/envs/omg/bin/python prepare_sessions.py --config configs/sessions.yaml --audit-only --report /root/autodl-tmp/cache/hfd/val_audit.json --set split=val +-``` +- +-迁移时替换 Python、数据盘和原始 tar 路径;详见 [迁移说明](MIGRATION.md)。`--audit-only` 通过只说明索引构造与配比可行,不保证全部 tar、ASR 或训练就绪。全量 tar 已校验、独立 train/val 索引已构造;本轮正式 batch 的短测结果见本文开头及报告。长时吞吐和论文独立测试集评测须另行衡量。 +diff --git a/docs/GENMO_FORMAL_TRAINING.md b/docs/GENMO_FORMAL_TRAINING.md +deleted file mode 100644 +index d2b273f..0000000 +--- a/docs/GENMO_FORMAL_TRAINING.md ++++ /dev/null +@@ -1,150 +0,0 @@ +-> 本文保留旧正式配方的记录。当前默认训练入口采用 [GenMo v2](GENMO_V2_TRAINING.md),包括执行段损失、预测历史恢复、CPU预取及闭环选权重;旧配置见 `configs/genmo_legacy.yaml`。 +- +-# GenMo 原生英文 HFD 训练 +- +-2026-10-24:本轮100K正式训练已完成,耗时约85.8小时。best为92K,已对官方初始化、92K和最终100K按既有六类验证规则完成闭环、整段GMR及有声视频评测。当前六段结果以92K略好,后退/移动任务仍明显不足;完整数值与视频位置见[正式评测报告](HUMAN_FORMAL_VALIDATION_20261024.md)。以下命令和训练约定保留;新正式权重未修改或删除。 +- +-正式入口为 `train_genmo_full.py --config configs/genmo.yaml`。旧 `train_genmo.py` +-和 `run_genmo_overfit.py` 仍是单session检查入口,不用于全数据训练。 +-沿用锁定版本的原生 GEM `training_step`、regression→diffusion 顺序、AdamW、151D表示、 +-冻结预训练统计量及原生3D损失;适配层保留此前验证的坐标/位移逆变换、历史锁定和未来评分mask。 +- +-## 数据范围与输入 +- +-只使用原生英文 HFD:155,117条train候选session、4,719条val候选session。 +-第三方生产包目前只有G1时间线,没有经过验证的配对SMPL-X session,明确排除。 +-因此不能把这次训练与包含第三方数据的OMG/TextOp结果直接作为同数据量架构对比。 +- +-每次按相同索引、原生seed组装音频和G1轨迹,同时观察同次执行的源选择、裁剪、 +-持姿、对齐、过渡和拼接,依据权威配对表构造人体轨迹。源SMPL-X SHA256逐条核验, +-所有原生记录及stationary源的配对ID在首次使用时检查;未支持的操作或缺失源直接失败。 +-全双工人体session不是从G1反解出来的,也不要求预先把整套session解压存盘。 +- +-每条已组装人体session的30Hz重采样只计算一次并受CPU缓存容量限制。决策点同时位于 +-50Hz和30Hz网格;历史切片的插值支撑均不晚于决策时刻,末端无效目标仍保持原50Hz终帧。 +-优化前后实际batch32的前4步loss逐值一致;未来动作扰动和末端mask回归检查通过。 +- +-输入为已知/已执行的2秒人体动作历史,以及当前已释放的最近30秒英文GT前景文本。 +-词时间取现有离线对齐缓存,guard=.1秒,0.6+.4k秒释放,motion到microphone偏移.2秒。 +-冻结T5-3B以BF16编码,原生50token预算超长时保留最近token;空文本特征为零。 +-编码器常驻当前GPU,CPU特征缓存有256MiB上限。无需ASR、翻译或重新跑强制对齐。 +-本版本保留已采用的文本+动作历史条件,不混入独立的past-audio实验。 +- +-输出为未来2秒人体动作,30Hz,每次执行12帧(.4秒)再规划。不输入GT阶段、 +-listen/motion/idle标签或未来音频。训练窗口可有初始重复姿态和末端无效目标, +-已知历史不计入未来几何损失;未知边界位移保留监督。 +- +-## 采样、验证、续训 +- +-在原生session的合法.4秒决策窗口上带放回抽样,使用保守上界拒绝采样维持窗口边际。 +-一个槽位绑定一条session,默认16批后刷新槽位,每批重新抽该session中的决策点。 +-训练按step预算运行,epoch不表示无重复遍历全部session。 +- +-每2K步进行固定验证:原生GEM双模式目标、固定窗口和噪声,无梯度;网络保持eval, +-只切换diffusion调度器的训练目标标志,防止原生图像验证入口不兼容或BN统计更新。 +-该loss不是连续推理或任务成功率,应同时使用下面的retarget视频评测。 +-默认每次16批×16窗口,验证每批刷新session,使用固定的256个窗口;这仍是验证子集, +-不是遍历4,719条候选session。训练池的16批复用策略不用于验证。 +-GEM的时间步采样使用NumPy,故验证同时保存、固定并恢复Python/NumPy/Torch随机数, +-不能仅固定Torch高斯噪声。同一真实checkpoint重复验证两次loss均为7.5329627991, +-仅用于验证复现性,不代表正式训练效果。 +- +-checkpoint保存模型、AdamW、按step的cosine调度器、RNG和已消费的采样池状态。 +-先写临时文件再原子替换,保留最近两份和最佳验证权重。完整checkpoint单份约5.1GiB, +-默认放共享盘 `models/genmo/hfd-english-train`,日志放本地 `runs/hfd-genmo-english-train`。 +-常态约需3份权重空间,保存时还需要临时文件空间;共享盘写入错误会中止并保留原checkpoint。 +- +-续训核对输入配置、模型适配实现、原生配方、离线文本缓存、配对表和body model身份。 +-只允许在原总步数预算内续训;max_steps不是“再训练多少步”,不能直接修改总预算或学习率后冒充恢复。 +-可更换GPU,允许把同一checkpoint分支到新的空输出目录。 +- +-```bash +-cd /root/autodl-tmp/hfd-baseline-portable +-/root/autodl-tmp/envs/genmo/bin/python -u train_genmo_full.py --config configs/genmo.yaml +- +-# 中断后恢复完整状态 +-/root/autodl-tmp/envs/genmo/bin/python -u train_genmo_full.py --config configs/genmo.yaml --resume auto +-``` +- +-当前总预算100K步,20K/40K/60K为阶段检查点,不表示已验证必须训练到100K才收敛。 +-学习率2e-5起步,按全部100K步做cosine衰减,在100K降到1e-6;`T_max`直接读取 +-`training.max_steps`,20K不再是学习率衰减终点。BF16、GPU1、batch32,合计320万次窗口采样。 +-采样仍有放回、槽位每16批刷新,不保证每个session都出现,不把100K解释为固定epoch数。 +-每2K步固定验证和保存,保留最近两份及最佳验证checkpoint;阶段检查应及时运行连续评测, +-20K/40K/60K权重不会永久保留。评测需单独使用空闲GPU或在训练暂停后进行,避免抢占显存。 +-验证或rollout停止改善时可提前停止,并从最佳checkpoint评测;当前不自动按val/loss早停, +-因为它不是语义成功率。100K是总上限,不是达到20K后再额外训练100K。 +-若中断,用同一100K配置恢复;旧20K协议checkpoint不能通过修改YAML直接兼容续训。 +-本次预算调整时正式输出目录尚未创建、权重目录为空,无旧正式训练需要迁移。 +- +-实测含T5-3B的GPU峰值约23.7GiB。热批次数据准备约4.9秒降到0.05–0.07秒, +-计算约0.4秒;换session池仍有组装开销,不能把热批次加速当作整轮训练的加速倍数。 +-不会自动开始长训练。 +- +-## 连续推理、retarget与视频 +- +-```bash +-/root/autodl-tmp/envs/genmo/bin/python -u evaluate_genmo.py \ +- --run /root/autodl-tmp/runs/hfd-genmo-english-train \ +- --output /root/autodl-tmp/runs/genmo-english-validation \ +- --gpu 1 +-``` +- +-默认读取最佳完整checkpoint,六类原生英文val各选首条6–30秒样本,可用 `--checkpoint PATH` +-和 `--modes turn-taking interruption ...` 显式指定。样本选择在推理前,使用与训练相同的 +-配对provider与文本对齐;只初始化一次GT历史,随后反馈预测,不按窗口重置GT。 +-推理中的指令显示时间读取实际session元数据;GT文本出处传递到retarget与视频,避免误标ASR。 +- +-每条完整预测和完整人体GT各retarget一次,保留三种比较:预测→G1/原G1、人体GT→G1/原G1、 +-预测→G1/人体GT→G1。retarget误差是独立诊断,不能从模型误差直接相减。 +-视频保留原混合音轨与共同microphone前缀。默认生成三条视频,属于运动学回放,不是物理跟踪。 +-原生task源按train/val隔离;val用于训练监测,不是独立test,预训练数据重合未审计。 +- +-## 2026-10-20 训练链路复核 +- +-可复用检查入口: +- +-```bash +-/root/autodl-tmp/envs/genmo/bin/python -u tools/audit_genmo_training.py \ +- --config configs/genmo.yaml --gpu 1 \ +- --output /root/autodl-tmp/setup/genmo-training-audit-20261020/audit.json +-``` +- +-此次检查对六类交互各取首条合格train session,共核对483个决策点,其中465个非空文本。 +-逐点重建文本并核对词末时间、guard和离散释放时间,没有违反标注释放规则; +-相对强制对齐词末的披露延迟为0.110875–0.48秒,一处不合格对齐采用整段结束后释放。 +-强制对齐是事后估计,0.1秒guard不是人工验证的误差上界;这仍是oracle文本实验。 +-六条样本最长29个T5 token,未发生截断,不代表全量语料截断率为零。 +- +-实测改变未来GT位置和姿态,regression预测逐值不变;清空文本后未来预测发生变化 +-(标准化151D特征的平均绝对变化0.05681275;不是语义成功率)。 +-regression和diffusion分别反向传播,文本投影、动作输入投影、主干和相关输出头均有 +-有限非零梯度;固定历史的原始姿态输出梯度为零,未知边界位移保留梯度。 +-执行一次用于检查的diffusion优化器更新后,文本/动作输入投影权重及bias均变化; +-T5参数保持冻结且无梯度。检查不保存模型权重,不代表收敛或闭环质量。 +-配套44项回归检查通过;报告位于上述output路径,日志在同级目录的 +-`genmo-training-audit-20261020.log`。 +- +-此次修复:位移目标i表示i→i+1,原生编码会复制窗口末段位移。当前屏蔽不存在下一有效帧 +-的位移和静止标签,包括窗口末帧及session尾部填充边界,保留有效姿态和历史→未来边界 +-位移监督。该修改改变训练目标mask和实现身份,旧实验loss不可直接逐值对比,旧完整状态 +-不能伪装为新协议resume。此次修改前未启动GenMo正式训练。 +- +-评测新增retarget之前的人体根朝向、根速度误差,与已有身体关节旋转及根位置误差一起报告; +-`evaluate_human.py`的`human_metrics`及`evaluate_genmo.py`摘要都会保留这些结果。 +-它们不依赖原G1轨迹;retarget后三种比较仍用于单独诊断。误差均不做根对齐或DTW, +-排除GT bootstrap。速度RMSE按三个坐标分量求均方根,单位m/s。 +- +-本设计适合“已释放文本+动作历史”的条件生成对照,存在以下能力/衡量限制: +- +-- 文本释放前没有语音起点信息,不能要求模型准确判断倾听开始;语调、韵律和背景声也不可见。 +-- 最近30秒是文本候选范围,实际最多50个T5 token;没有显式词时间或当前说话状态输入。 +-- 训练历史为GT,闭环历史为预测;未来2秒内的新指令尚不可见,目标天然存在不确定性。 +-- 输出预测2秒,每0.4秒重新规划执行;原生位置损失使用GT朝向积分,实际推理使用预测朝向, +- 因而必须同时看根朝向和连续rollout,不能只依据val/loss判定部署误差。 +-- 原生loss沿用全窗口均值乘mask,尾部有效目标少时loss会偏小;固定验证的窗口/噪声可比, +- 该loss不能直接当成“每个有效未来帧”的误差或跨模型通用分数。 +-- 现有指标衡量运动学轨迹一致性,尚不衡量独立定义的指令成功率、停止延迟或物理跟踪稳定性。 +- +-当前结论是结构与梯度检查通过,可开展该文本条件baseline训练;输入信息不足造成的反应差异 +-不能简单归为训练未收敛。按既定设置保留文本条件,音频条件仍作为另行定义的实验。 +diff --git a/docs/GENMO_HFD_OVERFIT.md b/docs/GENMO_HFD_OVERFIT.md +deleted file mode 100644 +index 805e6b3..0000000 +--- a/docs/GENMO_HFD_OVERFIT.md ++++ /dev/null +@@ -1,105 +0,0 @@ +-# GenMo / GEM-SMPL 的 HFD 数据适配与过拟合实验 +- +-清理说明:本文记录的一次性实验脚本已归档,复现前请按[脚本恢复说明](WORKSPACE_CLEANUP.md)取回;历史指标与日志保留。 +- +-新增原生过去音频分支的匹配实验见 [GENMO_PAST_AUDIO_TEST.md](GENMO_PAST_AUDIO_TEST.md)。音频输入已接通,当前小样本表现未优于 ASR-only,未替换本文工作版本。 +- +-本轮使用官方 `NVlabs/GENMO` commit `16bebf402d8893184249ee206d957b8248cd8310`、`nvidia/GEM-X` 的 **`gem_smpl.ckpt`**。模型源码没有修改;适配代码位于 portable 仓库。官方训练入口是 `scripts/train.py exp=gem_smpl`,本地桥接入口 `train_genmo.py` 直接继承官方 `GEM.training_step`、`train_step`、AdamW 和 regression→diffusion 执行顺序,使用原生 Lightning Trainer、噪声过程、151D 编解码与 3D 损失。 +- +-## 当前验证范围 +- +-只验证一条 train split 的双指令 interruption session:接别人递来的物品,随后切换到捡地上物品。输入、音频与此前 OMG/TextOp 的 `train-002203-seed42` 相同。human session 为 740 帧 @50Hz;连续导出覆盖 444 帧 @30Hz,视频 14.8 秒。首帧为已知初始化,其余帧闭环生成。此实验不是全数据训练、未见样本评估或物理跟踪。 +- +-原 human 包仍是源动作。`prepare_genmo_session.py` 使用 `hfd_bridge/human_sessions.py` 观察同一次原生 HFD recipe 执行,记录 source ID、精确采样时间、随机裁剪、listen 持姿帧、过渡 envelope、截断/恢复和 append 计划。人体端从 SQLite/tar 索引读取 SMPL-X,不批量解压;根据权威配对表和 SHA256 对应源动作。平移按秒插值,旋转按 SO(3)/SLERP 处理,使用人体自身的根高度和 pelvis offset 重新对齐。idle 采用局部旋转残差,listen 采用所选人体持姿与连续过渡,整个 session 保持同一个 16D shape。 +- +-原生 G1 复现与旧结果的 PCM、控制标签和决策索引逐值一致。受环境浮点实现影响,G1 数组最大差异为 `2.3841858e-7`;输出保留旧 G1 数组作为参考。该误差已记录,不宣称 G1 数组逐字节一致。50Hz/源时间的末端舍入最多允许一个 50Hz 帧的边界保持,实际保持时间写入计划,超过即报错。其他任务的几何/事件组合尚未逐一验收;观察器遇到无法追踪的姿态会拒绝导出。 +- +-## 输入与因果边界 +- +-- 原生 GEM-SMPL 151D:21×6D body rotations、10D shape、camera/gravity 两组根旋转、3D 根局部逐帧位移。沿用 checkpoint 的统计量,不另拟合 HFD 归一化。 +-- 使用相同 Whisper 因果 sidecar 的 `stable_latest_sentence`,经原生 T5-3B 编码,最多 50 tokens。冻结编码器,以 BF16 计算并缓存 FP32 特征;空文本缓存为零。缓存核对原 session 哈希。 +-- ASR 释放仍为 `0.6+0.4*k` 秒,动作决策为 `0.4*k` 秒;保留既有 HFD 三时钟约定。每次固定预测未来 60 帧,执行 12 帧,再重规划。 +-- 对照输入历史为 12 帧(0.4s)或 60 帧(2s)。历史包含当前状态。闭环仅使用实际已生成的历史,不在每次规划读取真值;首帧不足的历史重复该已知姿态。 +-- 视觉、2D、相机、music、speech 分支全部标缺失。没有未来音频、GT 指令、listen/motion/idle 标签、未来 Stop、任务进度、GT turn ID 或 GT 总时长作为条件。 +-- 固定虚拟坐标映射为 AMASS Z-up→GEM Y-up,camera 和 gravity 使用同一固定坐标;平移转换包含 shape 对应的 SMPL-X pelvis offset,导出时返回 Z-up。 +-- 原生根“速度”使用前向帧差。历史单独重采样/编码,最后一个已知姿态的下一步位移保持未知;其 mask=0,作为未来边界监督。不能使用第一帧未来 GT 计算该历史特征。 +- +-沿用原生 observed-motion overwrite,不添加新的 mask embedding 或网络层。已知前缀在网络输入及输出均固定;原生根位移 loss 会积分整段窗口,因此仅锁输入、却让任意历史输出补偿未来位移是不一致的。输出固定通过无参数 forward hook 实现,并覆盖训练和 DDIM 采样。未知边界位移保留梯度。 +- +-## 训练损失与执行 +- +-保留原生 simple MSE、root-relative 3D joints/437 vertices、world translation、static-contact BCE;缺少真实图像时关闭相机平移和 2D/投影损失。simple loss 的 mask 排除复制的已知特征;几何/contact loss 只对未来有效帧评分。注意原生 masked mean 仍除以固定窗口尺寸,因此**不同历史长度的原始 loss 数值不宜直接比较**。比较应使用同样长度的闭环输出指标。 +- +-full model 共约 518M 参数,原始权重 709 个 state tensors 全部严格匹配。motion 模型全部微调,T5 冻结;batch=2,BF16,AdamW,gradient clip=0.5。本轮先做 3 步训练烟雾检查,再探索 300 步及低学习率继续训练。排查完 prefix 输出一致性后,两种历史长度都从同一官方权重、相同 seed=42、lr=2e-5 开始,分别训练 600 步。`locked-h12-fit600` / `locked-h60-fit600` 是用于比较的修正版本;此前 `pilot`、`fit300`、`fit1300`、`history60-fit600` 是排查记录。 +- +-## 位移表示往返检查 +- +-固定版本的 `get_local_transl_vel` 文档写 R^T,但实际代码为 R×world_delta;`rollout_local_transl_vel` 也乘 R,两者不互逆。真值完整 session 的原生往返误差约 0.01168m,正确逆变换 R^T 的误差约 6.4e-10m。桥接保留原生编码/预训练特征定义,修正解码与 world-translation loss 的积分方向;不修改上游 checkout。随机旋转的往返测试防止仅在近似单位旋转上“偶然正确”。 +- +-`final-h12` / `final-h60` 是输出前缀锁定后的匹配 600 步对照,保留了原生积分方向,用于记录历史长度实验。发现积分问题后,以 h60 的 600 步权重为初值,lr=5e-6 再训练 300 步(AdamW 状态重新初始化),保存为 `corrected-h60-fit900`;最终复现入口默认采用正确逆变换。`initial-corrected-h60` 是相同解码规则下的官方权重基线。修正这项确定性错误不等于消除了自由闭环的速度偏置/误差累积。 +- +-## 重要发现:短历史存在观测歧义 +- +-`input-ambiguity-h12.json` 找到两对完全相同的文本和历史姿态: +- +-| 动作时刻 | 当前文本 | 两份未来 body rotation 标签的 RMSE 差异 | +-| --- | --- | --- | +-| 2.8s / 3.2s | Taking an item from someone | 0.230 rad | +-| 7.6s / 8.0s | Picking an item up from the floor | 0.264 rad | +- +-同样的输入不能唯一确定两份不同的未来真值,不能把无法精确过拟合都归为优化问题。2s 历史在这条样例中消除了上述精确重复,但这不证明所有任务已经可观测,也不保证模型能记住更长的语义状态。若仍有事件响应问题,后续应单列原生 past-audio 或因果状态记忆实验,不能注入 GT 阶段计时器。 +- +-## Retarget 与视频 +- +-`infer_genmo.py` 导出标准 AMASS SMPL-X。GEM 不预测手指,预测端使用零残差手姿态;known shape 使用 session bootstrap 的 16D betas,GEM 原生特征取前 10 维。human GT 保留实际手残差。评测使用 `flat_hand_mean=False`、`use_pca=False`、`ext=npz`,与 human 包约定一致。 +- +-`evaluate_human.py` 将**完整连续轨迹**各 retarget 一次到 G1,每条轨迹使用独立、跨帧连续的 GMR IK 状态。GMR commit `bb1bbe40774794fceb2a7c579a3464a28e68c844`,solver=daqp,`offset_to_ground=False`。仅稀疏检出 GMR package 与 Unitree G1 资产,约 79MB;AutoDL 安装依赖失败后通过清空子进程代理的阿里云 HTTPS PyPI 国内镜像直连完成依赖安装。 +- +-同时保留三组比较和视频:prediction→G1 vs 原 G1、human GT→G1 vs 原 G1、prediction→G1 vs human GT→G1。人体真值转换与原 G1 的关节 RMSE 约 0.202 rad,是几何转换/表示的诊断,不能从模型误差中直接相减。视频沿用 `render_overfit.py` / OMG MuJoCo renderer,含原始混音声音,展示真实输入的因果 ASR 和初始化/预测历史标记。 +- +-## 文件和复现 +- +-实验根目录:`/root/autodl-tmp/runs/archive/legacy-20261020/genmo-overfit`。模型训练权重放 `/root/autodl-fs/pangxinlei/models/genmo/hfd-overfit`,数据盘只保留软链接。源数据仍保留 tar;原生 GenMo 和 GMR checkout 没有修改。 +- +-配置入口为 `configs/genmo_overfit.yaml`,一键运行: +- +-```bash +-cd /root/autodl-tmp/hfd-baseline-portable +-/root/autodl-tmp/envs/genmo/bin/python run_genmo_overfit.py \ +- --config configs/genmo_overfit.yaml \ +- --output /root/autodl-tmp/runs/genmo-new-overfit \ +- --prepared-session /root/autodl-tmp/runs/archive/legacy-20261020/genmo-overfit/session \ +- --text-cache /root/autodl-tmp/runs/archive/legacy-20261020/genmo-overfit/text.pt +-``` +- +-不提供后两个参数时,会重新准备同一条 paired session 和文本缓存。输出目录必须是新的。此 runner 调用的只是本轮已逐步验证的准备、训练、推理、retarget、渲染入口;组合 runner 自身不额外启动一轮重复训练。 +- +-契约检查:`tests/test_genmo_contract.py`、`tests/test_human_eval.py`。真实输入/梯度审计为 `tools/audit_genmo_inputs.py`:扰动全部未来 GT 后,observed motion、mask、f_cond/uncond/empty 和 T5 条件逐值不变;原生 regression/diffusion 两模式实际反向,532 个 denoiser 梯度 tensors 有限且非零。另检查 SO(3) 跨 ±π 插值、坐标/pelvis roundtrip、前缀损失掩码与未知边界位移梯度。 +- +-## 本轮结果(2026-10-18) +- +-正式比较使用同样的 2s 历史、固定坐标和修正后的位移逆变换,单次已知初始化后闭环生成,不使用真值重置: +- +-| 设置 | 对原始 G1 的 29 关节 RMSE (rad) | 根位置每坐标 RMSE (m) | +-| --- | ---: | ---: | +-| 官方权重 | 0.4786 | 0.3804 | +-| 微调 600+300 步,自身历史 | 0.1998 | 0.1285 | +-| 同一权重,真值历史诊断 | 0.1857 | 0.0125 | +-| human GT 经同一 retarget | 0.2016 | 0.0093 | +- +-最终模型与 **同样 retarget 的 human GT** 比较为关节 0.1384 rad、根位置 0.1236m;这更直接衡量相同转换后的生成误差。人体原始空间的闭环根位置每坐标 RMSE 为 0.1426m,真值历史诊断为 0.0070m。曲线显示根位置存在随时间积累的偏差,不能将本轮称作完整成功过拟合。动作阶段也仍有偏差。 +- +-前缀输出锁定、原生积分方向下的匹配 600 步对照:0.4s 历史的 G1 关节 RMSE=0.3420、2s 历史=0.1975 rad;两者根位置误差均约 0.34–0.35m。这组仅用于记录输入长度实验,与修正后 900 步结果分开。 +- +-结论:**数据适配和完整推理评测链路通过;当前小样本闭环过拟合未完全通过。** 后续建议从 2s 历史开始,分别验证闭环预测历史训练、原生 contact 后处理或显式因果状态记忆;每一项单独标注适配和对照,不加入 GT 任务阶段或手工真值位置校正。增加其他任务前,先冻结并验收对应人体/G1 session。 +- +-结果入口: +- +-- `/root/autodl-tmp/runs/archive/legacy-20261020/genmo-overfit/README.md`、`summary.json`。 +-- 主视频:`final/evaluation/rollout.mp4`;相同 retarget 比较:`final/evaluation/retargeted_human_comparison.mp4`;转换诊断:`final/evaluation/retarget_reference.mp4`。 +-- 最终三条视频均为 444 帧、30fps、14.8 秒,原始混音声音已解码核验;`videos.json` 保存详细检查。人工查看了 listen、第一动作、打断后的 listen、第二动作的代表帧。 +-- `trajectory_diagnostics.png` 展示关节误差和根漂移;视频是参考姿态播放,不代表物理控制成功。 +-- 最终权重:`/root/autodl-fs/pangxinlei/models/genmo/hfd-overfit/corrected-h60-fit900/fitted.ckpt`。 +-- 14 个相关契约测试通过;完整官方权重严格加载、真实 regression/diffusion 反向、未来 GT 扰动后的采样输出逐值不变、真实 GMR 和音视频集成已验证。GPU 训练峰值约 9.17GB。ASR/T5 为预计算缓存,采样耗时不是端到端在线延迟。 +- +-复现最终继续训练阶段:在 portable 目录执行 `GENMO_HISTORY_FRAMES=60 /root/autodl-tmp/envs/genmo/bin/python train_genmo.py --session /root/autodl-tmp/runs/archive/legacy-20261020/genmo-overfit/session --text-cache /root/autodl-tmp/runs/archive/legacy-20261020/genmo-overfit/text.pt --checkpoint /root/autodl-tmp/runs/archive/legacy-20261020/genmo-overfit/locked-h60-fit600/fitted.ckpt --output /root/autodl-tmp/runs/genmo-reproduce900 --steps 300 --batch-size 2 --lr 5e-6`。当前 `run_genmo_overfit.py` 默认是从官方权重直接运行修正后的 600 步实验配方,与本轮分阶段 900 步训练的优化器历史不同,应在报告中区分。 +diff --git a/docs/GENMO_LISTENING_DIAGNOSIS.md b/docs/GENMO_LISTENING_DIAGNOSIS.md +deleted file mode 100644 +index a7ab5de..0000000 +--- a/docs/GENMO_LISTENING_DIAGNOSIS.md ++++ /dev/null +@@ -1,25 +0,0 @@ +-2026-10-24:针对GenMo多个episode长期保持倾听姿态,使用当前92K权重做离线对照。正式权重未修改,训练损失、采样配方和推理规则均未改动。 +- +-已经确认:失败发生在人体生成阶段,早于retarget。按词释放的英文文本确实进入模型,模型也能产生变化的未来特征;当前证据不支持“没有读到文本”或“整个网络已经输出常数”。主要暴露的问题是:GT历史训练下能延续运动,反馈自身预测时却容易停留在倾听状态,动作启动不足。 +- +-| 后退样本val-backchannel-000057 | 人体姿态RMSE(rad) | 根位置RMSE(m) | 整段根路径(m) | +-|---|---:|---:|---:| +-| 正常预测历史闭环 | 0.536 | 1.252 | 0.136 | +-| 每0.4秒使用真实历史,仅作诊断 | 0.155 | 0.030 | 4.702 | +-| 人体GT | — | — | 3.652 | +- +-GT历史对照会不断重置条件及空间位置,因此位置误差天然受保护,其速度误差反而较大;它不能作为正式评测或有效修复。该对照说明运动历史对结果影响很大,并不证明模型能独立完成后退。 +- +-同一8秒决策点、相同英文caption和噪声,单窗口把真实运动历史改为固定倾听历史:CFG1时未来两秒的根路径由0.266m降至0.011m;CFG2.5时由0.206m降至0.093m。改变历史会明显影响运动启动。直接清空文本也会改变预测:t=999时未来151D特征平均绝对差约0.00963,真实文本特征范数约17.05;此时未来特征的帧间RMS约0.0271。因此不能把倾听姿态卡住等同于编码器完全失效。 +- +-曾怀疑原生CFG的无文本分支移除历史条件特征。实际检查当前`in_attr`不含该独立历史特征,历史主要由`motion_mask_3d/observed_motion_3d`约束;强制让两分支共用`f_cond`后,单窗口和完整闭环输出均保持相同。该候选原因已经排除,没有据此修改生产推理。 +- +-训练数据抽查来自真实`train_sampling`记录的最早、约三分之一、三分之二和末尾四个池,共128个窗口。按根速度>0.1m/s或平均身体角速度>0.3rad/s、活动帧占比至少25%分类:80个窗口的两秒未来有活动,63个窗口的实际12帧执行段有活动,49个窗口的历史较安静,其中13个执行段转入活动,122个窗口有非空文字。此抽查不是全语料统计,但排除了“训练全部读成静止样本”的解释。起动转换仅占该抽查约10.2%,不能据此推断全库比例。 +- +-当前训练完整地监督未来60帧,闭环只执行最前12帧;全部历史来自GT,没有预测历史混入。模型一旦在起动处继续预测倾听,下一次重规划又看到自己的倾听历史;caption不变时容易形成持续保持姿态的闭环。近真实未来加噪的去噪损失下降,不能保证从安静历史自主启动动作。这是目前最直接的失效机制,具体损失权重、历史课程、数据分布的独立贡献尚未完成受控消融。 +- +-后续训练策略应优先验证:提高已执行12帧及起动窗口的监督占比,加入真实解码运动的变化监督,并用train-only的预测历史训练恢复能力。输入仍保持原协议,不能用GT阶段、未来文本或评测时真实历史替代这些训练工作。本轮没有直接启用这些未验收配方,也没有重新训练100K。 +- +-本地诊断代码在被忽略的`test/diagnostics/`,数值与中间轨迹在本机cache;GitHub仅保留本文结论和复现当前正式训练所需的代码。当前训练源快照见`TRAINING_SNAPSHOT.json`及对应Git历史;本次迁移改动仅处理路径、资源定位和训练依赖拆分,没有改变训练算法。 +- +-后续已实现的训练适配改动见 [GenMo v2](GENMO_V2_TRAINING.md)。本文诊断数值仍对应旧正式权重,不能当作新策略全量训练后的效果。 +diff --git a/docs/GENMO_PAST_AUDIO_TEST.md b/docs/GENMO_PAST_AUDIO_TEST.md +deleted file mode 100644 +index 4d9c162..0000000 +--- a/docs/GENMO_PAST_AUDIO_TEST.md ++++ /dev/null +@@ -1,93 +0,0 @@ +-# GEM-ASR-past-audio:原生音频分支测试 +- +-2026-10-18。在已有 GenMo/HFD bridge 中启用官方 speech encoder,使用同一条冻结 interruption session、同一因果 ASR、同一人体真值和 GMR 链路。不是新增语音语义模型;ASR 文本仍经 T5-3B,音频由 checkpoint 自带的卷积前端编码。 +- +-## 输入构造 +- +-第 k 次决策的动作时刻是 `t_k=0.4*k`,按既有 HFD 协议,PCM 释放点是 `r_k=0.6+0.4*k`。取**释放点之前**的 `[r_k-2,r_k)` 两秒混合 PCM;不能把动作轴、PCM 轴和墙钟直接视为同一个零点。这里保留 OMG/TextOp/HFD 的释放规则。 +- +-1. 从原 16kHz session 混音按采样点截取,右端严格为 `9600+6400*k`。启动时不足两秒的左侧补静音,只有真实覆盖的历史位置 availability=1;首个决策实际覆盖 18 个音频 bin。 +-2. **先截断,再重采样和卷积。** SciPy polyphase FIR 按 9/8 重采样到 18kHz,保持波形幅度。未来样本没有进入 FIR 或卷积。 +-3. 原生六个 BasicBlock 将 `[B,1,36000]` 编码为 `[B,128,60]`;stride=600 samples,恰为 30Hz。卷积感受野为 18,635 samples(约 1.035s)。所有卷积支持都限制在本次已释放窗口及边界补零内;不是在完整 session 上先提特征再切片。 +-4. **编码后**补 60 帧未来特征,再用原生 `audio_embedder`、`cond_exists_embedder` 和 CFG 逻辑融合。120 帧条件中只有历史音频位置可用,未来 audio availability 恒为 0。未来的 f_cond 数值可以有“缺失条件”的学习偏置,但与波形内容逐值无关。 +-5. 特征按时间顺序放入 60 个动作历史槽位。卷积的中心时间为 `r_k-2+j/30`,最后一个中心比释放点早 1/30 秒;保留原生卷积的 phase,没有为对齐而读取释放点之后的样本。原生训练 BatchNorm 会在已截断的 batch/time 内统计;推理使用 eval/running statistics。 +- +-其它输入保持:60 帧已执行动作历史+T5 当前稳定 ASR;预测 60 帧,执行 12 帧,30Hz。训练使用真值历史,闭环使用模型自身历史;仅首帧初始化已知。不输入未来音频、GT 指令、状态/turn 标签或任务进度。music、视觉、2D 和相机条件仍缺失。 +- +-## 实现边界 +- +-- 新增 `hfd_bridge/genmo_audio.py` 负责释放窗口、重采样和卷积后 padding。 +-- `GENMO_PAST_AUDIO=1` 且 `GENMO_HISTORY_FRAMES=60` 启用音频;默认关闭,因此已有 ASR-only 入口保留。 +-- 在原生 `GEM.create_condition_mask` 外接入已截断的波形;encoder 的无参数 forward hook 仅在卷积之后补齐 future features。没有添加新的学习层、mask embedding 或改变 709 个 state tensor 的结构。上游 checkout 保持 clean。 +-- 保留原生音频条件 dropout=0.1、BatchNorm、卷积和音频投影的训练方式。模型主体及音频前端一起微调,T5 保持冻结。 +-- 之前修正的 prefix 输出锁定、未知边界位移和实际位移编码逆变换在两组实验中一致。 +- +-## 对照设计 +- +-两组均从 `/root/autodl-tmp/runs/archive/legacy-20261020/genmo-overfit/corrected-h60-fit900/fitted.ckpt` 初始化,再各训练 600 个 optimizer steps,batch=2、lr=5e-6、seed=42。motion 模型累计微调 1500 步;音频前端在本轮训练 600 步。优化器均重新初始化。两组的窗口、文本、监督、损失权重、DDIM 10 步、CFG 和采样种子相同;音频前端 dropout 会消耗额外随机数,因此不宣称两组训练噪声逐位一致。 +- +-| 训练目录 | 输入 | +-| --- | --- | +-| `asr-control-fit600` | ASR+2s 动作历史 | +-| `past-audio-fit600` | 同上+2s 已释放混合 PCM | +- +-另对音频模型使用 `infer_genmo.py --drop-audio` 做消融:清零波形并把 audio availability 全设 0,ASR、动作历史和权重不变。视频仍播放原 session 的真实声音,标题明确显示 `AUDIO OFF`;播放声音不等于模型收到了音频。 +- +-## 验证 +- +-- 完整模型严格加载仍为 709 个 state tensors。 +-- 扰动释放点之后的原始 PCM,重采样输入逐值相同;同时扰动未来 GT,原生条件及 DDIM 输出逐值相同。 +-- 过去 PCM 改变实际条件和未来采样结果,未来 audio 条件槽与波形内容无关;CFG unconditioned branch 对波形不变。 +-- 原生 regression/diffusion 两模式实际反向,音频 encoder 与 projection 的梯度有限且非零。 +-- 训练完成后的 checkpoint 检查:ASR-only 的音频 encoder/projection state 与初值完全相同;音频组有实际更新,BN batch counter 增加 1200,符合每步 regression/diffusion 各一次前端调用。 +-- 四个音频契约测试覆盖未来 PCM 隔离、启动补零与可用帧、16→18kHz 的频率保持、拒绝未覆盖的释放点。原有六个 GenMo 契约测试在音频开/关两种进程配置下通过。 +- +-测试根目录:`/root/autodl-tmp/runs/archive/legacy-20261020/genmo-audio-overfit`,真实前端审计为 `audio-audit.json`,权重更新审计为 `trained-audio-state-audit.json`。 +- +-## 配置与运行 +- +-配置:`configs/genmo_audio_overfit.yaml`;组合 runner 为 `run_genmo_overfit.py`,设置 `past_audio: true` 会在训练和推理子进程中显式启用对应环境变量。要做匹配的 ASR-only 对照,复制配置只把 `past_audio` 改为 false,保持初值、步数和学习率一致。 +- +-本轮直接调用的训练方式(两组用不同的新输出目录): +- +-```bash +-GENMO_HISTORY_FRAMES=60 GENMO_PAST_AUDIO=1 \ +-/root/autodl-tmp/envs/genmo/bin/python train_genmo.py \ +- --session /root/autodl-tmp/runs/archive/legacy-20261020/genmo-overfit/session \ +- --text-cache /root/autodl-tmp/runs/archive/legacy-20261020/genmo-overfit/text.pt \ +- --checkpoint /root/autodl-tmp/runs/archive/legacy-20261020/genmo-overfit/corrected-h60-fit900/fitted.ckpt \ +- --output /root/autodl-tmp/runs/genmo-new-audio-test \ +- --steps 600 --batch-size 2 --lr 5e-6 +-``` +- +-权重默认写 fs 的 `models/genmo/hfd-overfit//fitted.ckpt`,数据盘只保存软链接。推理时使用同样的两个环境变量。评测和视频继续调用 `evaluate_human.py`、`render_overfit.py`,对完整连续轨迹 retarget;不在每个规划窗口重置 IK。ASR、T5 为预计算缓存,报告的采样耗时不是端到端在线耗时。 +- +-## 实测结果 +- +-以原始 G1 为参考,全部排除首帧初始化;根位置为每坐标 RMSE,未做根对齐或 DTW: +- +-| 设置 | 采样 seed | 29 关节 RMSE (rad) | 根位置 RMSE (m) | +-| --- | ---: | ---: | ---: | +-| 匹配 ASR-only 对照 | 42 | 0.2061 | 0.0366 | +-| 原生过去音频,eval/running BN | 42 | 0.2893 | 0.1030 | +-| 匹配 ASR-only 对照 | 43 | 0.1992 | 0.0261 | +-| 原生过去音频,eval/running BN | 43 | 0.2773 | 0.1096 | +-| 音频模型,关闭 speech 条件 | 42 | 0.1992 | 0.1076 | +-| 音频模型,当前窗口 BN 诊断 | 42 | 0.2380 | 0.1085 | +- +-与同样 retarget 的 human GT 比较,seed 42 的 ASR-only 为关节 0.1395 rad、根位置 0.0350m,音频版为 0.2480 rad、0.0989m,方向一致。因此,退化不能仅归因于原始 G1 与 human GT 的转换差异。`summary.json` 另存 listen/motion/idle 三类 GT 阶段的评分;这些标签仅用于评估,未输入模型。 +- +-诊断中的 `--audio-normalization window` 只改变推理时音频 BatchNorm:其它模块保持 eval、Dropout 关闭,BN 使用**当前已截断的两秒窗口**统计,关闭 running-stat 更新,不使用未来 PCM 或整段 session 的统计。这是单独的归一化适配诊断,不是默认原生推理。姿态有所改善,但根漂移没有解决,尚不足以将归一化差异认定为唯一原因。未来 PCM 扰动的逐值不变性检查针对原生 eval 推理;训练保留原生跨 batch/time 的 BN 统计语义。 +- +-**结论:当前版本的因果性、维度、权重加载、梯度和完整评测链路通过,但小样本性能没有优于 ASR-only,音频版尚未成功过拟合。** 两个采样种子都显示相同方向;它们不是两次独立训练,更不是未见数据泛化验证。暂保留 ASR-only 为工作版本,音频模型作为实验结果保存。 +- +-可能的后续检查包括:固定音频前端/BN 的训练对照、原生逐帧音频预训练与仅历史条件之间的分布差异、闭环预测历史训练。当前结果不证明原生音频能力无效,也不能证明上述某一项已经解释了退化。 +- +-## 产物 +- +-- 汇总与曲线:`/root/autodl-tmp/runs/archive/legacy-20261020/genmo-audio-overfit/README.md`、`summary.json`、`comparison.png`。 +-- 音频版视频:`past-audio/evaluation/rollout.mp4`;ASR-only 对照:`asr-only/evaluation/rollout.mp4`。 +-- 两组均另有 `retarget_reference.mp4`(human GT→G1 vs 原 G1)和 `retargeted_human_comparison.mp4`(相同转换后的预测比较)。保留原始混音声音;视频为参考姿态播放,不是物理控制跟踪。 +-- 消融和归一化诊断保存在 `audio-disabled`、`audio-window-bn`,明确标注模型收到的输入;seed 43 复核保存推理数组与完整 GMR 数值,不重复渲染视频。 +-- 训练权重:fs 的 `models/genmo/hfd-overfit/asr-control-fit600/fitted.ckpt`、`past-audio-fit600/fitted.ckpt`。数据盘训练目录内为软链接。 +-- 未安装新依赖、下载新模型或复制主体资产;沿用已准备的 GenMo、T5、SMPL-X 和 GMR。 +diff --git a/docs/GENMO_V2_TRAINING.md b/docs/GENMO_V2_TRAINING.md +deleted file mode 100644 +index 0cc76f7..0000000 +--- a/docs/GENMO_V2_TRAINING.md ++++ /dev/null +@@ -1,52 +0,0 @@ +-# GenMo v2:执行段监督、预测历史恢复与 CPU 预取 +- +-本版本针对长期保持倾听姿态的失效机制改造训练适配。模型骨干、英文前景 GT 词时间释放、最近 30 秒上下文、2 秒动作历史、2 秒预测及每次执行 0.4 秒的输入输出协议不变。没有新增音频、未来文本或 GT 阶段条件。正式旧权重保留;短程链路检查不能替代新策略的全量质量评估。 +- +-## 训练目标 +- +-保留原生 regression + diffusion 两个训练分支及姿态几何损失;全局根平移监督改为按预测朝向解码的实际轨迹,避免用GT朝向掩盖推理时的方向误差。151D 特征误差中,实际执行的前 12 帧及历史末帧到第一未来帧的位移权重设为 3。逐样本、逐特征归一化权重,保持有效监督总权重与原配方一致,避免仅仅放大总 loss。无效尾帧及未知末帧位移仍不参与监督。 +- +-另加前 12 帧的解码根速度与身体关节旋转增量 SmoothL1 监督,默认权重分别为 1 和 0.05。根位移按原生编码器的实际逆变换 `Rᵀ @ local_delta` 解码,旋转来自预测本身;不以未来 GT 朝向代替预测朝向。旋转增量在 SO(3) 上计算,避免轴角跨分支的假速度。速度项不是鼓励一直运动,正常倾听/停止的 GT 速度仍是低速或零。 +- +-## 预测历史恢复 +- +-每个训练 batch 最多选 25% 的可用窗口,保持原始窗口采样分布。前 1,000 步不混入预测历史,接下来 4,000 步逐步增加,5,000 步起达到该上限。源窗口不足 0.4 秒历史时不参与恢复。 +- +-对时刻 `t` 的恢复样本,在 `t−0.4s` 决策点用当时已释放的文字及过去动作生成一次原生 10 步 DDIM 预测,只取实际执行的 12 帧,替换 `t` 输入历史的末 12 帧。生成所需的未知未来位置填成末个已知姿态,不读 GT 未来;预测过程关闭梯度及 dropout/BN 更新。未来GT关节姿态、根旋转、位移增量及时间不变;将未来根轨迹整体平移到实际历史的末位置,保持相同的局部运动。GEM的151D输入含根速度而不含绝对根位置,因此不能要求它纠正输入中不可观测的全局位置偏移。目标编码及速度/旋转增量的边界以这份一致的历史/未来为起点,避免监督矛盾。目标编码、解码和新增时序损失使用FP32几何计算。恢复样本仍需从偏离的自身预测回到正确轨迹。 +- +-这是一段预测历史的恢复课程,不等同于对无限长闭环的保证;必要时可另做多段恢复消融,不能把评测时使用 GT 历史当作修复。 +- +-## 验证与权重选择 +- +-继续记录固定噪声 `val/loss`。默认另选两个不同 interaction mode 的固定英文验证 session,各最多 900 帧。只用一个初始真实姿态启动,之后全部反馈自身预测,文字沿原时间轴释放。训练/验证 RNG 在检查后恢复。 +- +-默认 `best` 按 `val/closed_loop_score` 选择,定义为两个 session 的 `body geodesic RMSE(rad) + root coordinate RMSE(m)` 的平均值。另保存根速度误差与预测路径长度。这是训练选权重的代理指标,不是全测试集分数,也不替代人体到 G1 retarget 后的正式评测。`last` 仍按训练步保存完整恢复状态。 +- +-## 吞吐改动 +- +-- 单独的 CPU producer 负责 session 组装和窗口构造,最多预取 16 个 batch(一个样本池周期,深度上限 32);GPU 主进程负责冻结 T5 与训练。进程间只传配置、CPU batch 和采样状态,不共享 SQLite 连接或 CUDA 对象。 +-- checkpoint 记录已消费 batch 的池/RNG 状态,忽略预取但尚未消费的数据。producer 及主进程训练/验证 DataLoader 均使用独立随机生成器,不挤占模型的扩散随机流。 +-- GenMo 的 SO(3) 插值/混合批量化;人体几何 observer 只挂接实际需要的函数返回事件,避免全进程 `sys.setprofile` 的逐函数开销。原 HFD 文件及其校验哈希不修改。此优化需通过与旧 observer 的成对 session 对照,不能只看速度。 +-- 训练 session 内存预算为 4 GiB,验证预算为 8 GiB,文本特征 LRU 为 512 MiB。另有进程及索引开销;内存较少的机器可降低这些预算,采样规则不变。 +-- 启用内容校验的本地原生 session 索引缓存,缓存的是源组合索引,不是 GT session 全量导出或模型输入。原始 tar 继续留在共享盘,离线词时间缓存、索引缓存、代码和日志位于工作盘。 +-- 保持 `refresh_batches=16`,没有通过延长同一批 session 的重复使用来虚增吞吐。GT 几何在两个训练分支间复用一次计算。NumPy/OpenBLAS默认限制为1线程,避免小矩阵使用64线程的调度开销;正式配置关闭逐 batch 进度条,保留结构化指标输出。 +- +-耗时记录新增 `cpu_batch_seconds`、`text_encoding_seconds` 和样本池刷新标记。`data_wait_seconds` 是训练批次间的主进程等待,不能直接等同于磁盘 IO;验证耗时不再计入该字段。评估速度时应跨越多个样本池刷新,并单独报告初始化。 +- +-本机稳定阶段实测约快 1.4–3.1 倍,较保守复测为 1.44 倍;冷启动和共享存储波动仍明显。完整测量口径、正确性与小规模质量限制见 [GENMO_V2_VALIDATION.md](GENMO_V2_VALIDATION.md)。 +- +-## 命令与旧实验复现 +- +-现有服务器从仓库根目录正常前台运行: +- +-```bash +-/root/autodl-tmp/envs/genmo/bin/python -u train_genmo_full.py --config configs/genmo.yaml +-``` +- +-新输出为 `runs/hfd-genmo-english-v2`,权重为共享盘 `models/genmo/hfd-english-v2`;不覆盖旧的 `hfd-english-train`。跨机器先按 [PORTABLE_TRAINING.md](PORTABLE_TRAINING.md) 生成本机配置。 +- +-新 run 的完整断点使用同配置加 `--resume auto`。如要从旧正式权重初始化新策略,使用 `--init-checkpoint /path/to/old.ckpt`;这是权重 warm start,优化器、调度与步数重新开始,不能伪装成旧 run 的续训。默认仍从官方 `gem_smpl.ckpt` 开始。 +- +-`configs/genmo_legacy.yaml` 保留旧配方;配置生成器 `--genmo-recipe legacy` 可以复现该配方。若要审计旧训练的原实现,Git `0bcc49e` 保留了与旧 GenMo/Kimodo 正式训练源码记录匹配的快照,`fa5c1a5` 是本次策略修改前的便携版。恢复旧完整断点应使用对应原配置、代码和资产。跨版本/路径改动不承诺逐位相同。 +- +-测试、基准和过拟合脚本均在忽略的 `test/`,不随训练代码发布;实验日志、权重和视频不进入 Git。 +diff --git a/docs/GENMO_V2_VALIDATION.md b/docs/GENMO_V2_VALIDATION.md +deleted file mode 100644 +index ab834f4..0000000 +--- a/docs/GENMO_V2_VALIDATION.md ++++ /dev/null +@@ -1,45 +0,0 @@ +-# GenMo v2 本机验证记录(2026-10-24) +- +-本页记录训练实现和短程实测,不能视为新策略已经完成全量训练或测试集验收。正式配置与设计见 [GENMO_V2_TRAINING.md](GENMO_V2_TRAINING.md),跨机器准备见 [PORTABLE_TRAINING.md](PORTABLE_TRAINING.md)。诊断脚本、原始日志、测试权重和视频均不发布到 Git。 +- +-## 原耗时来自哪里 +- +-旧 100K run 耗时约 85.8 小时,记录的平均计算约 0.367 秒/步,平均批次等待约 2.718 秒/步;旧等待字段还混入验证间隔,不能直接当作 IO 时间。一次冷启动的 8-session CPU profile 中,实际读取约 625.6 MB 的累计耗时为 1.027 秒,而旋转插值累计约 6.644 秒。session 组装、旋转处理及主进程串行等待也是瓶颈,单纯搬动 tar 不能解释全部耗时。 +- +-本次保持原始 tar 在共享盘、缓存和派生索引在本地盘;采用向量化几何、定点 observer hook、本地索引缓存、单 CPU producer 和最多 16-batch 预取。保持每 16 batch 刷新样本池,未通过延长同一批 session 的重复使用换取吞吐。 +- +-## 吞吐实测 +- +-同机 RTX 5090 32 GB,batch 32,各执行 64 次真实优化更新,跨越 4 个样本池。基线与优化版的 session ID、决策帧和采样顺序一致;新策略开启恢复训练,GPU 计算负担比旧策略更高。吞吐检查关闭验证,不含正式闭环验证开销;总耗时栏是训练指标计时,另有进程初始化和退出开销。 +- +-| 测量 | 前 16 步(秒) | 64 步(秒) | 第 17–64 步平均(秒/步) | 同区间计算 / 等待(秒/步) | +-|---|---:|---:|---:|---:| +-| 旧策略、串行数据 | 51.61 | 284.00 | 4.84 | 0.35 / 4.49 | +-| 新策略、预取 16,较热缓存的一次 | 91.60 | 167.05 | 1.57 | 0.68 / 0.89 | +-| 修正监督后的新策略复测 | 111.50 | 272.33 | 3.35 | 0.68 / 2.67 | +- +-稳定阶段观察到约 **1.4–3.1 倍**吞吐提升;更保守的复测是约 **1.44 倍**。包含前期等待的 64 步复测只缩短约 4%,不能承诺全量训练固定快 3 倍。共享存储/主机负载与缓存状态使样本池生成耗时明显波动,预取仍未完全遮蔽 CPU 等待。新策略两次峰值显存约 20.7 GB(十进制),基线约 25.4 GB;这不是最大 batch 搜索结果。根轨迹监督最后增加了平移重锚,最终源码另通过参数更新与恢复检查,未再重复整套吞吐基准。 +- +-## 正确性检查 +- +-- 14 个成对 session 覆盖 6 种交互模式:音频、G1、文字、时间及参考元数据完全相同;人体数组最大绝对差约 `7.45e-9`。原始 HFD 文件及哈希不变,旧 profiler 状态保持。 +-- 7 项单元检查通过,覆盖 SO(3) 插值、执行段权重总量/有效 mask/梯度、姿态键、显式与隐式返回 hook、恢复样本条件、配置迁移与资源链接。 +-- 原生编码器检查通过:扰动历史后,保持未来局部运动并平移重锚的完美预测,其新增速度、根轨迹损失接近零;错误预测有有限且非零的梯度。未来 GT 关节姿态、根旋转和位移增量保持,源 GT batch 未改写。 +-- 训练恢复以预取队列中**已消费**的样本状态为准。最终源码连续 4 步与 2 步后中断恢复至 4 步的对照通过:709 个模型张量、Adam、调度器、采样状态及 Python/NumPy/Torch/CUDA RNG 均逐位一致。对照开启预取、预测历史恢复、固定噪声验证和闭环验证;这是短程恢复检查,不是长期质量保证。 +-- 外部资产清单中 141 个文件(约 53.6 GB)的 SHA256 全部核验通过。OMG 50M/100M/300M/500M 均严格加载并以配置中的微批次执行一次丢弃的真实参数更新;这不代表各尺寸已完成 HFD 正式训练。 +-- 从 Git 暂存内容导出的干净副本检查训练入口和模块依赖,不依赖忽略的 `test/` 或本机旧适配仓库。OMG/TextOp dry-run、GenMo 配置及新增模块导入检查通过;Kimodo 原生离线检查通过,真实参数更新成功,文本敏感性为 0.0632,根/身体梯度均有限且非零。 +- +-## 小规模闭环质量复核 +- +-使用旧 GenMo 正式 best 初始化,在两个**训练集** session 的全部合法窗口上训练 600 步,batch 8;恢复课程压缩为 warmup 40、ramp 160。此实验只检查新框架能否学习与闭环是否改善,不是默认 100K 配方的验证集比较,也不是与旧策略同预算的完整消融。 +- +-评估各自前 450 帧(15 秒):一个 GT 初始姿态,之后仅反馈预测历史,执行段为 12 帧;文字按原时间释放。无根对齐或时间扭曲。误差从第二帧开始统计,路径长度包含整个轨迹。 +- +-| 训练 session | 根坐标 RMSE,前→后(m) | 身体旋转 RMSE,前→后(rad) | 根路径,预测前→后 / GT(m) | +-|---|---:|---:|---:| +-| 53390 | 2.794 → 0.922 | 0.462 → 0.317 | 2.973 → 4.328 / 6.075 | +-| 28035 | 0.125 → 0.142 | 0.630 → 0.619 | 0.549 → 0.468 / 1.588 | +- +-首 20 步平均训练 loss 从 2.623 降至末 20 步的 1.499,身体、边界及未来速度输出均有非零梯度。一段恢复明显,另一段仍动作不足;此前其他短程窗口实验也有改善和退化并存的情况,不能声称倾听停滞已经彻底解决。 +- +-两段预测和人体 GT 分别对完整连续序列 retarget 一次到 G1,保存原 G1、人体 GT retarget 与预测的三种对照视频。视频本机目录为 `runs/genmo-v2-debug-20261024/videos/{53390,28035}/`;它们是训练集诊断,不用于报告测试集成绩。新正式 run 应在 20K/40K/60K 查看自主闭环与 retarget 视频后判断效果;默认 best 已改按固定自主闭环分数选择,旧正式权重保留。 +diff --git a/docs/HFD_HUMAN_BASELINE_INPUT_PROTOCOL.md b/docs/HFD_HUMAN_BASELINE_INPUT_PROTOCOL.md +deleted file mode 100644 +index 6579d20..0000000 +--- a/docs/HFD_HUMAN_BASELINE_INPUT_PROTOCOL.md ++++ /dev/null +@@ -1,149 +0,0 @@ +-# HFD 五个 baseline:当前输入、输出与训练约定 +- +-更新:2026-10-21(Asia/Taipei)。本文记录当前本地实现与约定;源码/运行配置优先。 +-原2026-10-17调研、上游commit、论文链接和待实现方案已完整保存在 +-[历史调研](history/HFD_HUMAN_BASELINE_INPUT_PROTOCOL_RESEARCH_20261017.md)。 +-历史ASR/中文翻译、12帧GenMo历史等配方不作为当前默认。 +- +-## 共同文本协议 +- +-五个baseline统一约定如下;OMG、TextOp、GenMo已实现,Kimodo已增加全量原生英文动态人体训练与闭环评测入口,HY-Motion仍待适配。Kimodo的小规模验收不代表全库正式训练已经完成。 +- +-```yaml +-foreground_language: en +-caption_policy: aligned_reference_context +-clock_policy: microphone +-text_conditioning: none +-context_seconds: 30.0 +-alignment_guard_seconds: 0.1 +-``` +- +-输入是已经释放的英文前景GT文本。原始TTS原文与对应源音频由 +-Qwen3-ForcedAligner-0.6B离线估计词时间,再按动态session内的真实音频拼接偏移平移。 +-原生指令、停止、附和、犹豫、修正片段均按各自的实际播放位置处理,不提前给出最终完整指令。 +-不运行ASR、不翻译、不用动作标签选择“当前任务”,也不做摘要或prompt rewrite。 +-背景混合PCM原样保留,但背景文字不输入。这是oracle foreground transcript实验, +-不代表已经解决部署时的ASR或目标说话人识别。 +- +-当前词级对齐库含61,860条英文前景资产;这是音频资产数,不是不同句子数或session数。 +-对齐跨度无效、越界、非单调或文本覆盖不完整时,在对应完整音频片段结束后释放整段文字。 +-0.1秒guard是实验缓冲,不是经过人工验证的对齐误差上界;事后强制对齐也不等同在线识别。 +-实现见 `hfd_bridge/reference_text.py`、`reference_runtime.py`,共享声明见 +-`configs/english_text_protocol.yaml`。 +- +-## 时钟与文本释放 +- +-动作源为50Hz,混合音频16kHz。动作时刻t对应麦克风时刻 `t+0.2s`。 +-文本释放点采用麦克风时间 `r_k=0.6+0.4k`,k从0开始。 +-词末估计为e时,先等到 `e+0.1s`,再向上取整到下一个r_k: +- +-```text +-r(word) = 0.6 + 0.4 * max(0, ceil((e + 0.1 - 0.6) / 0.4)) +-当前可见文本 = 最近已到达释放点r_k上的文本记录,要求r_k <= t + 0.2 +-``` +- +-因此麦克风0.6/1.0/1.4秒对应动作0.4/0.8/1.2秒;动作起点没有预先获得0.6秒文本。 +-这是当前microphone协议;旧 `legacy_hfd` 按动作组提前取记录的行为仅供历史复现。 +-旧材料“同一动作组直接读取0.6秒结果”不能继续套入当前正式实验。 +- +-每条记录只保留该释放点前30秒内、已经完成并获准释放的词;不等待句子完整, +-不拼接历次窗口字符串,过期后可以为空,不沿用任意久远的最后一句。session切换时清空。 +-例如词末0.85秒,加guard到0.95秒,在麦克风1.0秒释放,动作0.8秒决策才能看见。 +-0.6秒音频窗与0.4秒步长是重叠窗口:相邻只新增0.4秒,不能说每步新增0.6秒。 +- +-模型不接收未来PCM、未释放文字、GT turn ID、listen/motion/idle标签、未来事件时刻、 +-真实剩余动作时长、未来根路径或未来姿态约束。标签仅用于监督、样本选择、评测与视频标注。 +-扩散训练的带噪未来目标是学习变量;不能将干净未来GT作为条件。 +- +-## 实际输入与输出 +- +-共同的是信息释放规则,不要求模型拥有相同张量或动作历史长度。 +- +-| 模型 | 当前输入 | 当前输出及执行 | 本地接入状态 | +-| --- | --- | --- | --- | +-| OMG | 已释放英文文本→冻结T5-Base,最多50 token、保留最近后缀;过去2秒已释放混合PCM;10帧动作历史@30Hz | 原生125D动作特征,解码G1全29关节及根状态;计划60帧/2秒;当前原生异步缓冲剩24帧触发规划 | 正式英文训练及验证评测已接通 | +-| TextOp | 已释放英文文本→冻结CLIP ViT-B/32,77 token含两个边界token;2帧57D原生动作历史特征@50Hz;不输入音频 | 128D primitive latent,经冻结VAE解码8帧57D/0.16秒动作;G1原生23关节及根状态,未建模腕关节固定 | 正式300K训练完成;连续验证和视频见TextOp评测入口 | +-| GenMo/GEM-SMPL | 已释放英文文本→冻结T5-3B,最多50 token、保留最近后缀;60帧/2秒人体历史@30Hz;固定已知体型;当前past_audio=false | 原生151D表示,解码身体21关节旋转、根朝向及根位移;未来60帧/2秒,每次执行12帧/0.4秒后重规划 | 100K已完成,best92K与最终100K的闭环/GMR/视频见[正式评测](HUMAN_FORMAL_VALIDATION_20261024.md) | +-| HY-Motion | 共同英文文本计划接原生Qwen3 token与CLIP-L全局特征;流式历史条件需显式扩展 | 原生201D人体运动表示;连续输出、历史帧数与mask尚未验收 | 已有模型环境/调研;本地完整HFD训练尚未接通,不声称已按此协议训练 | +-| Kimodo-SMPLX | 已释放英文文本→冻结LLM2Vec,最近400正文token,mean pooling为1×4096;60帧/2秒人体历史@30Hz;203维历史姿态约束,无音频;当前配置采用因果root分解 | root→body两阶段,SMPLX22的273D表示;默认未来60帧/2秒,执行12帧/0.4秒;中性手部,保留源体型;12帧预测诊断未优于默认 | 修复版100K已完成,best即100K;冻结根骨干仅训练根投影/身体,见[训练修复](KIMODO_TRAINING_REPAIR.md)及[新评测](HUMAN_FORMAL_VALIDATION_20261024.md);未宣称闭环质量已达标 | +- +-30秒是候选文本范围,不等于30秒所有文字都进入编码器。TextOp的75个正文token +-加边界共77;T5为原生50 token。当前各自编码和缓存embedding,不共享不同模型的向量。 +-Kimodo的400正文token进入原生512总预算及模板;独立文本worker验证双向注意力和两份适配器权重。HY-Motion的token预算和前缀长度仍待实际adapter固定。 +- +-Kimodo使用归一化x0主损失,历史70维未观察辅助量在每个DDIM步保持训练噪声分布,避免未监督历史输出继续回灌。2026-10-23修复版增加原始根历史重建及实际解码运动的连续性监督;根/身体分别优化和裁剪,冻结根注意力骨干,未增加未来观测或推理修正。早期2K+1K、在线历史课程属于历史诊断。全量配置从完整英文索引按窗口均匀边缘抽样,不采用诊断中的70%活动窗口筛选;文本worker与动作模型同卡、独立进程,已编码特征在本地SQLite复用。共同文本释放规则及其余四模型输入未更改。 +- +-TextOp每8帧重新查询当前可见文字,文本本身只在0.4秒释放点改变;不同模型的执行周期 +-不必等于文本释放周期。其原生历史特征含相邻帧差分,2帧特征需读到当前决策状态, +-不能简单理解成两个完全独立的旧姿态。TextOp原生57D包含根姿态/位移、23个关节、 +-关节差分和接触;沿用官方统计与VAE,不直接拿G1原始36D训练。 +- +-OMG的60×35声学位置来自已经截取的过去PCM,不是未来2秒音频;文本和音频释放时钟一致。 +-初始不足两秒的音频左补零。TextOp和GenMo主实验没有声学分支,因此在文字出现前 +-没有显式语音起点、韵律或背景说话信息,不能要求它们仅靠文字精确判断倾听开始。 +- +-## 人体数据与session构造 +- +-源人体资产包含 `trans[T,3]`、`root_orient[T,3]`、`pose_body[T,63]`、 +-`pose_hand[T,90]`、`betas[16]`、`sample_times`。它们是源动作,不是完整全双工人体session。 +-以 `metadata/seed_g1_smplx_pairs.parquet` 的motion_id连接源,按原生配方的真实选取、 +-裁剪、持姿、朝向/位置衔接、过渡和事件时间构造人体时间线。 +-当前实现仍借用原G1 session组合程序及配对表追踪计划;不是从G1反解人体。 +-G1关节不是GenMo或Kimodo网络输入或训练损失目标。 +- +-GenMo训练实际使用trans、root_orient、21个身体关节及前10维betas; +-16维体型还用于SMPL-X pelvis/坐标转换。手指与面部不作为生成目标,推理导出中手部为中性残差。 +-151D由21×6D身体旋转、10D体型、两组6D根朝向和3D根局部相邻帧位移构成。 +-位移编码没有乘FPS,不能直接称为m/s;坐标转换和推理积分必须互为逆变换。 +- +-原生源动作约50Hz对齐成完整session后,按时间与SO(3)插值重采样到30Hz。 +-输入历史不能通过插值使用决策之后的GT;GenMo历史单独编码,最后一个历史帧指向 +-第一未来帧的位移未知,保留为生成目标。已知历史锁定,无下一有效帧的位移/静止标签不计损失。 +- +-## 数据范围与缓存 +- +-| 模型范围 | 英文train候选session | 英文val候选session | +-| --- | ---: | ---: | +-| OMG/TextOp:原生+已选第三方 | 168,583 | 5,109 | +-| GenMo:仅原生配对人体 | 155,117 | 4,719 | +- +-第三方生产包有G1时间线,但尚无经验证的配对人体session;GenMo排除该部分。 +-因此不能直接把OMG/TextOp与GenMo的结果差异完全归因于模型结构。 +-原生索引/seed保持,英文过滤不重排session ID;原生任务源train/val隔离。 +-当前没有独立test集,代码中test别名val不能当作未触碰测试集;预训练/共享stationary源重叠未排除。 +- +-离线完成的是源英文语音词对齐及索引,不是提前导出全部人体session。 +-OMG/TextOp/GenMo正式训练均按需组装原生session,组装结果放有上限的进程内缓存,退出即失效; +-第三方则读取已组装生产session。GenMo额外缓存30Hz重采样,train/val各512MiB, +-冻结文本特征共享256MiB CPU LRU;这不是全量持久session缓存。 +-原始tar在共享盘保留,本地只读索引镜像减轻共享盘随机读取。 +-采样有放回、槽位复用,step/epoch不等于逐条无重复遍历整个候选集。 +- +-## 训练与评测 +- +-正式配置:`configs/omg.yaml`、`configs/textop.yaml`、`configs/genmo.yaml`。 +-OMG预算100K、TextOp三阶段合计300K、GenMo100K;不按步数直接比较训练数据量。 +-OMG/GenMo训练使用GT历史,TextOp沿用原生分阶段GT/rollout/static-history课程。 +-正式连续评测只做一次初始GT bootstrap,之后使用预测历史;重复喂GT的结果必须标为oracle诊断。 +-TextOp当前根位置与朝向锚点取已知初始决策状态,不能每个primitive重新用GT根状态对齐。 +- +-评测入口:`evaluate_omg.py`、`evaluate_textop.py`、`evaluate_genmo.py`。 +-OMG/TextOp直接输出G1,不需要人体retarget。 +-人体模型先比较人体关节旋转、根位置、根朝向和根速度,再对整条预测及整条人体GT各做一次 +-SMPL-X→G1 retarget,不能在每个规划窗口重置IK。三种比较分别是预测→G1/原G1、 +-人体GT→G1/原G1、预测→G1/人体GT→G1;重定向误差不能直接从模型误差中相减。 +-统一保存原混合音轨、microphone前置帧、文本释放审计和bootstrap排除范围。 +- +-验证loss用于监测或选权重,但不等于连续生成质量或语义成功率。特别是GenMo原生位移loss +-使用GT朝向积分,闭环使用预测朝向,必须同时看朝向和长时漂移。 +-当前视频是运动学回放,不代表物理控制跟踪、独立任务成功率或真实端到端停止延迟。 +-GenMo的双分支梯度/因果性与44项检查见[正式训练说明](GENMO_FORMAL_TRAINING.md)。 +- +-更多入口:[共同文本条件](BASELINE_TEXT_CONDITION.md)、[session格式](../SESSION_FORMAT.md)、 +-[动态reference组装](REFERENCE_DYNAMIC_TRAINING.md)、[GenMo正式训练](GENMO_FORMAL_TRAINING.md)、 +-[人体retarget](HUMAN_RETARGET_EVALUATION.md)。历史研究及社区训练代码核查仍见归档, +-不把当年的待实现建议、旧ASR输入或旧缓存描述当作当前已经验收的实现。 +- +-## GenMo v2训练适配补充(2026-10-24) +- +-输入/输出与本协议相同。训练额外使用一段模型预测历史作为因果恢复条件,GT未来仅作监督,前0.4秒增加损失权重;这不增加模型推理时可见的信息。默认最佳权重用自主闭环验证选择,执行细节见 [GENMO_V2_TRAINING.md](GENMO_V2_TRAINING.md)。旧配方保留在 `configs/genmo_legacy.yaml`。 +diff --git a/docs/HFD_HUMAN_BASELINE_SOURCES.json b/docs/HFD_HUMAN_BASELINE_SOURCES.json +index 8752b9f..5cf7f2f 100644 +--- a/docs/HFD_HUMAN_BASELINE_SOURCES.json ++++ b/docs/HFD_HUMAN_BASELINE_SOURCES.json +@@ -175,10 +175,6 @@ + } + ], + "local_evidence": [ +- { +- "path": "/root/autodl-tmp/hfd-baseline-portable/SESSION_FORMAT.md", +- "sha256": "b478da021db41576b023de4ee057350e93539b89a9427fa680f16004abfb226b" +- }, + { + "path": "/root/autodl-tmp/hfd-baseline-portable/README.md", + "sha256": "f2b9adf6cf12bcbe7ebd3af72e74355d7a1910f0238c2744571a03a6193c7f40" +diff --git a/docs/HUMAN_FORMAL_VALIDATION_20261024.md b/docs/HUMAN_FORMAL_VALIDATION_20261024.md +deleted file mode 100644 +index 531d697..0000000 +--- a/docs/HUMAN_FORMAL_VALIDATION_20261024.md ++++ /dev/null +@@ -1,38 +0,0 @@ +-2026-10-24:GenMo、修复版 Kimodo 均已完成 100K 步;本轮按既有六类 native English validation 规则执行闭环生成、整段 GMR 和有声视频。没有训练、修改或删除正式权重。 +- +-输入为按既定词时间释放的最近30秒英文GT文本与60帧已执行人体历史,无音频输入。只使用第一帧GT初始化,随后完全反馈预测;规划60帧、执行12帧、30Hz。Kimodo采用DDIM20/CFG2,GenMo采用DDIM10/CFG2.5,均seed12345。人体GT和完整预测分别连续retarget一次,无轨迹对齐、DTW或位置修正。 +- +-GenMo依原入口每类选首个预计6–30秒的样本;Kimodo依原入口选6–20秒。打断、急停的样本ID不同,其余四类相同。所有样本均在生成前确定。下表用于比较各模型微调前后,不能据此直接排名架构;也不是全HFD验证集、独立test或物理跟踪/任务成功率评测。 +- +-| 模型/权重 | 人体姿态RMSE(rad) | 人体根位置RMSE(m) | G1关节RMSE(rad) | G1根位置RMSE(m) | G1根朝向RMSE(rad) | +-|---|---:|---:|---:|---:|---:| +-| GenMo official | 0.651 | 0.995 | 0.688 | 0.864 | 2.342 | +-| GenMo best92k | 0.407 | 0.701 | 0.373 | 0.604 | 0.373 | +-| GenMo last100k | 0.427 | 0.758 | 0.376 | 0.655 | 0.390 | +-| Kimodo official | 0.627 | 0.502 | 0.464 | 0.433 | 0.575 | +-| Kimodo last100k | 0.422 | 0.306 | 0.373 | 0.264 | 0.487 | +- +-六段分别评分后等权平均,排除GT初始化帧。人体姿态是21个身体关节的SO(3)角误差RMSE;根位置为三坐标逐元素RMSE。G1指标参考同流程retarget的human GT,原始G1仅作另一个参考;retarget误差不从模型误差中相减。 +- +-GenMo:优先保留92K的best作为当前评测候选。固定去噪验证loss在92K最低(2.409),100K为2.464;本轮六段连续生成的人体姿态、根位置及根速度误差也均为92K略低。训练耗时约85.8小时。批量评测复用模型但恢复原始初始化后的采样RNG;官方和92K各抽一个固定样本与未修改的infer_genmo.py命令复核,人体位置/根旋转/身体旋转均逐位相同。 +- +-Kimodo:best与last均为100K,训练耗时约18.7小时,固定验证loss为1.179→0.0669。根骨干200个参数张量仍与官方初始化逐位一致;同一后退诊断窗口t=0的末层时序RMS为0.0360、输入梯度范数0.00586,旧失效100K分别约1.08e-7和4.76e-12。本轮未重现此前根网络恒定输出失效。 +- +-与上一轮失败的Kimodo 100K历史报告相比,G1根位置误差从0.544降至0.264m,人体窗口首帧根速度误差从约0.839降至0.109m/s;但G1根朝向误差从0.266升至0.487rad。修复包含可训练参数范围、学习率、损失和注意力后端的组合修改,不能把这些收益归因于某个单独超参数。 +- +-两者相对官方初始化均有平均误差收益,但动作质量仍有明显限制:后退、长距离移动与部分指令跟随不足。backchannel样本的人体GT净后退位移约3.50m,Kimodo仅沿GT后退方向移动约0.37m;GenMo该样本的整段根路径只有约0.14m,而GT约3.65m,接近保持位置。低平均误差不能当成动作任务已经完成。 +- +-保持初始姿态的对照也保存在JSON中。Kimodo新100K的人体姿态平均误差0.422略低于静止对照0.431,但G1关节0.373仍略高于静止对照0.366。GenMo92K的人体姿态0.407低于静止对照0.429。静止对照无法执行指令,只用于识别静态区间对平均分的影响。 +- +-本轮导出42段有声MP4:Kimodo 18段、GenMo 24段。均核对帧数、30Hz、音轨和完整音视频解码;抽帧检查六类预测/参考、文本与时钟显示。runs只保留视频,工作数据和日志在cache。 +- +-- `/root/autodl-tmp/runs/kimodo-formal-validation-20261024/last100k/`:六类预测→G1 / human-GT→G1及六段retarget参考;`official/`含六段官方初始化对照。 +-- `/root/autodl-tmp/runs/genmo-formal-validation-20261024/best92k/`、`last100k/`、`official/`:各六段预测→G1 / human-GT→G1;last100k另含六段GT retarget参考。 +-- `retargeted_human_comparison.mp4`左侧预测,右侧同流程retarget的human GT;`retarget_reference.mp4`比较human-GT→G1与原始G1。 +- +-完整报告与曲线: +- +-- [Kimodo数值、逐例结果、根骨干检查和视频校验](../reports/kimodo_formal_validation_20261024.json) +-- [GenMo数值、逐例结果、命令一致性和视频校验](../reports/genmo_formal_validation_20261024.json) +-- [Kimodo验证曲线与后退轨迹](../reports/kimodo_formal_validation_20261024.png) +-- [GenMo验证曲线与后退轨迹](../reports/genmo_formal_validation_20261024.png) +diff --git a/docs/HUMAN_RETARGET_EVALUATION.md b/docs/HUMAN_RETARGET_EVALUATION.md +deleted file mode 100644 +index 3176e0b..0000000 +--- a/docs/HUMAN_RETARGET_EVALUATION.md ++++ /dev/null +@@ -1,98 +0,0 @@ +-# 人体模型的 retarget 评测入口 +- +-`evaluate_human.py` 共用于 Kimodo、GenMo、HY-Motion。输入为模型导出的 **完整连续 rollout**,先转换为标准 SMPL-X / AMASS,再通过 GMR 转到 G1。该入口不加载生成模型、不下载权重,也不改变现有 TextOP / OMG 推理逻辑。 +- +-2026-10-21更新:GenMo已接通动态原生英文人体训练及`evaluate_genmo.py`连续推理、完整轨迹GMR与带音轨视频;Kimodo已增加固定原生英文子集训练与`evaluate_kimodo.py`完整轨迹导出、GMR和视频。两者当前均输入已释放英文GT文本+60帧人体历史,不输入音频。Kimodo范围见[训练说明](KIMODO_HFD_TRAINING.md),HY仍待接入。共同输入协议见[当前约定](HFD_HUMAN_BASELINE_INPUT_PROTOCOL.md)。GMR位于 `/root/autodl-tmp/hfd-human-models/gmr-runtime`,主体资产在fs的`models/body_models`。 +- +-## 同时保存三组结果 +- +-| report.json 指标 | 左侧 | 右侧 | 用途 | +-| --- | --- | --- | --- | +-| `prediction_vs_original_g1` | 预测 SMPL-X→G1 | 原始 G1 真值 | 与机器人 baseline 相同目标上的轨迹误差 | +-| `human_gt_vs_original_g1` | human 真值→G1 | 原始 G1 真值 | 检查体型、session 几何渲染及 retarget 自身的偏差 | +-| `prediction_vs_retargeted_human_gt` | 预测 SMPL-X→G1 | human 真值→G1 | 相同转换后的预测误差 | +- +-第二项是转换链路的诊断,不能从第一项中直接相减,也不是数学意义的误差下界。所有指标包含 29 关节 RMSE(rad)、根位置每坐标 RMSE(m)、根朝向 SO(3) 测地角 RMSE(rad)。不做 DTW、逐帧根对齐、截短或补齐。`bootstrap_frames` 在三组评分中一致排除。teacher-forced 必须标为 oracle,不能作为闭环结果。 +- +-此外,`report.json.human_metrics`在retarget前直接比较人体身体旋转、根位置、根朝向及根速度,排除同一bootstrap。根速度按实际FPS计算,单位m/s;人体关节角采用SO(3)测地角。优先用人体自身指标判断生成误差,G1比较辅助诊断转换与机器人覆盖差异。 +- +-一个连续序列仅创建一次 GMR retargeter,内部状态跨帧保留;预测和 human GT 分别新建,不能共享 IK 状态,也不能每个 0.4s 规划块重置。保持旧 benchmark 的 `offset_to_ground=False` 和身高估计 `1.66 + 0.1 * beta[0]`;体型使用 16 个 betas、`use_pca=False`、`flat_hand_mean=False`。不根据未来 GT 调整预测的朝向或地面。该身高估计是否适用于当前体型应在真实集成测试中核对。 +- +-## 输入契约 +- +-路径可相对 manifest 所在目录。manifest 的三份动作文件必须携带同一个已冻结的 `session_plan_id`。**这个 ID 只用于发现误配,脚本不能证明 human/G1 已按同一个 plan 正确渲染**;生成端必须先完成该保证。不要仅凭相同随机种子或相同动作名配对。 +- +-预测和 human 真值两个 NPZ 都必须包含: +- +-| 字段 | 要求 | +-| --- | --- | +-| `trans` / `root_orient` / `pose_body` | `[T,3]` / `[T,3]` / `[T,63]`,轴角,有限值 | +-| `betas` | 恒定 `[10]` 或 `[16]`,10 维补零至 16;不从 GT 偷换预测体型 | +-| `pose_hand` | 可选 `[T,90]` 左右手轴角;相对 SMPL-X 手均值的残差;缺省零表示模型默认手姿态 | +-| `gender` | 可选 `neutral` / `male` / `female`,缺省 neutral | +-| `mocap_frame_rate` | 正标量;预测和 human GT 必须相同 | +-| `coordinate_system` | 字符串 `amass_z_up`;`trans` 是模型平移,不是 pelvis 世界坐标 | +-| `sample_times` | 可选;若存在必须等于 `arange(T)/fps`,时间不均匀时先在导出端显式重采样 | +-| `session_plan_id` | 冻结 plan 的标识,推荐使用内容哈希 | +-| `start_seconds` | 该 rollout 第 0 帧在 session 中的时间,与 manifest 一致 | +- +-Kimodo 的 Y-up joint positions / rotations、HY 的 201D 特征、GenMo 的 151D 特征均不能直接送入此入口,须通过各模型原生解码与坐标逆变换导出 AMASS。`trans` 不做隐式 pelvis 偏移修正。非零 `pose_jaw` / `pose_eye` 暂不支持。预测与 human GT 的体型/性别是否相同会记录在 provenance,差异需要在实验说明中解释。 +- +-原始 G1 session NPZ 必须包含: +- +-- `layout="xyz_wxyz_native29"` + `qpos_36[T,36]`,或 `layout="hfd_isaac29_xyz_wxyz"` + `motion[T,36]`。后者走现有按关节名排列的转换,不能直接切片当 MuJoCo qpos。 +-- `fps`、`session_plan_id`、`start_seconds=0`。 +-- `audio`(一维PCM)、`audio_sample_rate=16000`、`audio_start_seconds`。当前`microphone`协议下原PCM起点填0,动作时刻t显示在麦克风t+0.2秒;从动作0开始的回放给两侧共同增加0.2秒前置姿态。不能把当前PCM起点机械填成-0.2或再次偏移。历史协议需按原manifest显式复现。 +- +-G1 真值按真实时间线重采样到 human rollout 帧率:位置/关节线性插值,根四元数 SLERP。覆盖不足立即报错;音频也必须覆盖完整的 `T/fps` 播放区间。当前入口面向均匀 rollout,不接受拼接成连续序列的 disconnected windows。 +- +-manifest 示例(示意路径,不是现有数据): +- +-```json +-{ +- "schema_version": 1, +- "model": "genmo", +- "session_plan_id": "sha256-of-frozen-plan", +- "prediction": "prediction_amass.npz", +- "human_reference": "human_gt_amass.npz", +- "g1_reference": "g1_session.npz", +- "start_seconds": 0.0, +- "bootstrap_frames": 1, +- "history": "predicted", +- "clock_policy": "microphone", +- "caption_policy": "aligned_reference_context", +- "text_conditioning": "none", +- "text_source": "oracle_foreground_transcript", +- "inputs": {"text": "released English reference text <=30s", "history": "60-frame history; initial pose held at startup, then predictions"}, +- "audio_policy": "Aligned reference text <=30s; no raw audio conditioning", +- "use_asr_text": true, +- "command_seconds": [0.4, 5.2] +-} +-``` +- +-`command_seconds` 仅供叠字展示,不用于生成或 retarget。输入策略字段是生成端声明,评测脚本不能据此证明模型没有使用未来信息。 +- +-## 执行与渲染 +- +-在具有 NumPy / SciPy 的模型环境中,先验证导出数据: +- +-```bash +-python evaluate_human.py --manifest /path/to/manifest.json \ +- --output-dir /root/autodl-tmp/runs/kimodo-eval \ +- --gmr-root /path/to/GMR --body-models /path/to/body_models --preflight +-``` +- +-preflight 检查文件、布局、时间覆盖和 plan ID,只报告 GMR/资产目录是否存在;不会导入 GMR 或验证资产许可证、内容、依赖是否完整,也不写评测目录。具备真实依赖后去掉 `--preflight` 执行转换。输出目录已存在时拒绝覆盖。 +- +-输出 `report.json`、三组配对 NPZ,以及汇总 `retargeted.npz`(预测、human GT、原始 G1 GT 和绝对 timestamps),包含输入 SHA256、GMR commit/dirty 状态、路径、转换配置和体型一致性。真实资产存在时记录主体 NPZ 的 SHA256,并记录相关包版本;GMR 工作树仍需保持固定。 +- +-现有渲染环境中执行: +- +-```bash +-python render_overfit.py --inference /root/autodl-tmp/runs/kimodo-eval \ +- --assets-root /path/to/OMG/assets --gpu 0 +-``` +- +-三组视频均使用原 session 音频;标题区分预测、human GT retarget 和原始 G1 真值。视频为参考姿态播放,不是物理跟踪结果。human GT 转换对照不显示模型 ASR 输入或预测历史提示。 +- +-开发验证:`python -m unittest tests.test_human_eval tests.test_joint_mapping -v`。合成测试覆盖三组指标、bootstrap 排除、序列级后端调用、拒绝误配/不均匀时间/覆盖不足、四元数双覆盖和 SLERP,以及 HFD→G1 关节名映射。 +diff --git a/docs/KIMODO_CLOSED_LOOP_STABILITY.md b/docs/KIMODO_CLOSED_LOOP_STABILITY.md +deleted file mode 100644 +index ca51599..0000000 +--- a/docs/KIMODO_CLOSED_LOOP_STABILITY.md ++++ /dev/null +@@ -1,144 +0,0 @@ +-# Kimodo 训练策略与闭环稳定性实验 +- +-本轮从官方权重重新构建实验。此前及本轮的测试权重均已清理,复现需重训;报告、代码和视频保留,正式GenMo训练不受影响。 +- +-## 当前保留的诊断路线与最终状态 +- +-**训练与推理的关键一致性问题已修正,但六种模式混合闭环过拟合、少样本验证均未通过;不启动全量正式Kimodo训练。** +- +-目前较好的诊断路线是:官方初始化→归一化x0目标训练2,000步→只加载模型权重、重置Adam,再用GT历史训练1,000步。输入协议不变,不增加GT阶段、语音结束状态或动作时钟。两个公开配置为`configs/kimodo_overfit.yaml`与`configs/kimodo_refine_probe.yaml`,准备/复用输入与续训命令见[训练说明](KIMODO_HFD_TRAINING.md)。首阶段配置与实测`normalized-x0`除输出路径外完全相同。 +- +-| 修正采样后的实验 | 累计步数 | 训练身体RMSE(rad) | 训练根RMSE(m) | 训练根速度RMSE(m/s) | +-| --- | ---: | ---: | ---: | ---: | +-| 物理Huber,同一模型旧历史反馈规则 | 2,000 | 0.3703 | 0.2580 | 0.4307 | +-| 物理Huber,同一模型修正历史反馈规则 | 2,000 | 0.3747 | 0.2137 | 0.3638 | +-| 归一化x0,50%高噪声/70%运动窗口 | 2,000 | 0.3822 | 0.2055 | 0.4102 | +-| 上述模型再用GT历史训练1,000步 | 3,000 | **0.3105** | **0.0915** | **0.3183** | +-| 同一初始化,再用在线历史训练1,000步 | 3,000 | 0.4355 | 0.2826 | 0.5754 | +-| 归一化x0,均匀扩散步/均匀窗口 | 2,000 | 0.3898 | 0.5479 | 0.3337 | +- +-第一、二行已核对checkpoint SHA相同,唯一改动是推理历史辅助量规则:根位置/根速度误差分别降低约17%/16%,身体误差没有改善。第四、五行已核对初始化权重相同、各追加1,000步;在线课程更差,**不采用为默认**。归一化x0在2,000步时并未全面优于物理Huber,不能把第4行的改善全部归因于损失更换,训练预算和优化器重启也改变了。 +- +-GT历史3K模型的六条结果如下,均为完整预测历史rollout: +- +-| 模式 | 身体RMSE(rad) | 根RMSE(m) | 预测/GT身体角速度比 | +-| --- | ---: | ---: | ---: | +-| turn-taking | 0.4387 | 0.0398 | 1.747 | +-| interruption | 0.3192 | 0.0686 | 1.165 | +-| emergency-stop | 0.0564 | 0.0396 | 1.149 | +-| backchannel | 0.3220 | 0.0236 | 1.107 | +-| hesitation | 0.3703 | 0.1470 | 1.103 | +-| pause | 0.3564 | 0.2307 | 1.148 | +- +-逐条精确值以[机器报告](../reports/kimodo_closed_loop_stability_20261021.json)为准。行走/停止已能较准确复现,其他模式虽运动幅度不再明显不足,动作相位和多余运动仍存在;幅度正确不等于动作正确。六条验证session宏平均为身体0.5220rad、根0.8850m、根速度0.3673m/s,身体只有1/6好于冻结初始姿态,泛化没有通过。当前训练入口仍是固定小规模子集,不是全库动态训练器。 +- +-优先查看本轮较好模型的三组预测↔人体GT视频(双方各完整retarget到G1,均含音轨): +- +-- [行走/停止](/root/autodl-tmp/runs/kimodo-closed-loop-20261021/score-control/train-emergency-stop-000036/retargeted_human_comparison.mp4) +-- [动作切换,仍有相位偏差](/root/autodl-tmp/runs/kimodo-closed-loop-20261021/score-control/train-turn-taking-000013/retargeted_human_comparison.mp4) +-- [犹豫,仍未准确复现](/root/autodl-tmp/runs/kimodo-closed-loop-20261021/score-control/train-hesitation-000087/retargeted_human_comparison.mp4) +- +-## 修正历史辅助量的推理反馈 +- +-训练时,历史未观察的70维速度/接触辅助量始终是`epsilon * sigma_t`;它们不参与历史损失。旧推理却把模型对这些未监督历史字段的预测继续经DDIM反馈,低噪声阶段也保留其信号,与训练不一致。 +- +-当前`inference.prefix_aux_policy: training_noise`在每次去噪调用前恢复相应扩散步的历史纯噪声边缘分布,然后重新锁定已知203维姿态。噪声来自本轮初始epsilon,没有额外GT或随机数;未来采样变量不被这一步改写,也不对输出做平滑。训练与20步稀疏DDIM的sigma映射有独立一致性检查,旧行为仅通过`legacy`供对照复现。在线历史刷新也使用同一规则。 +- +-新增`normalized_x0`主损失直接在原生归一化特征上做Gaussian x0最小二乘;root/body两阶段分别平均,已知/填充帧和未定义末端速度排除。它是训练目标对照,不宣称复现未公开的官方训练配方。`physical_huber`保留供历史结果对照。 +- +-当前代码41项测试通过;最新完整模型续训重放含修正后的采样与在线缓存,在确定性模式下逐位一致。该完整重放使用physical_huber分支;normalized_x0共用持久化路径,并完成了实际全模型训练与梯度检查。本轮共16条训练路径(含共享初始化和后续阶段),不是16次独立随机实验。以下为历史实验明细;旧表均使用修正前的`legacy`历史辅助量反馈,重新运行需显式指定该规则。 +- +-## 视频审计修正 +- +-快速probe最初没有导出每轮`executed_frames`,导致显示层误报“文本未释放”。模型实际使用的caption、运动轨迹和误差均未受影响。现已补齐字段,核对执行帧总数等于预测总长减去bootstrap,并重渲染全部72个视频。现有retarget结果未重新计算;报告记录了仅显示元数据修正及前后manifest哈希。 +- +-## 改动与对照原则 +- +-输入协议不变:按时释放的最近30秒英文GT文本、60帧已知人体历史、预测60帧、执行12帧、无音频。推理仅用一帧GT初始化,后续全部接自己的输出,不加入GT阶段、未来指令、动作时钟或目标姿态,也不做平滑后处理。 +- +-`hfd_bridge/kimodo_closed_loop.py` 实现两项可独立开关的训练策略: +- +-1. `closed_loop.tail_weight`:实际执行的前12个未来帧权重为1,余下48帧为较低的正权重。无效填充帧依然排除,分母使用有效权重之和。这个实验检验远期监督是否稀释当前决策,不代表已经证明未来文字是退化的唯一原因。 +-2. `closed_loop.online_history`:按 `refresh_every` 用当前模型重新闭环生成训练session,混入原GT历史窗口。`max_probability` 在 `ramp_steps` 内逐渐增加。刷新使用与推理相同的DDIM步数、CFG、文字释放时钟和单帧初始化;只刷新训练集,不对验证集训练。预测历史只作为条件,GT未来保留原世界坐标,不再平移根位置或混合前6帧目标姿态。 +- +-历史刷新不反传穿过整个采样轨迹;随后训练步对网络两阶段正常反传。独立的历史选择随机数不会改变原窗口抽样顺序,刷新会恢复训练的全局随机数和train/eval模式。完整checkpoint包含历史缓存、缓存对应的policy step和历史选择随机数;恢复训练不重新生成另一份历史替代原缓存。 +- +-原 `prepare_kimodo_predicted_history.py` 为历史实验保留,不作为本轮在线策略入口。 +- +-可选 `reset_schedule` 只控制训练历史课程:例如 `[2,5,0]` 在前两阶段每2轮、5轮从GT过去重新启动短rollout,最后阶段从一个GT初始姿态生成完整session。每个窗口保存模型当时实际看到的60帧历史,不能从包含后续GT重启的拼接轨迹倒推历史。正式闭环评测入口没有这个开关,始终不重启、不补GT历史。`curriculum_stage_steps` 决定阶段长度,实际历史在下一次刷新时切换。 +- +-## 固定实验 +- +-数据保持六种模式各一条训练session与验证session,分别205与215个窗口。所有数据来自原生英文配对人体session。81种合法释放文本使用同一份冻结LLM2Vec缓存。 +- +-- 先按原方案训练2,000步,作为公共初始化。 +-- 三组分别从同一初始化再训练1,000步:原方案对照;`tail_weight=0.1`;`tail_weight=0.1`加在线历史(最大概率0.5,500步渐增,每250步刷新)。 +-- 三组保持batch8、LR1e-5→1e-6、随机种子、窗口抽样顺序、原七项损失、文本dropout0.1、DDIM20、CFG2一致。本轮不同时叠加高噪声抽样、活跃采样或额外速度损失。 +-- GT历史去噪loss只用于诊断,不能跨不同权重目标直接比较。主要比较六条训练session的完整闭环误差、静止基线、角速度幅度、根速度误差和重规划加速度误差;验证集另列,不作为训练过拟合的替代。 +- +-前三组之后增加一个同预算的课程恢复候选:短历史→长历史→完整历史,50%样本t=999,并在root输出进入body阶段前锁定已知root历史。它是组合配方,不把其结果归因于单一因素;仍不增加活跃抽样或额外速度损失。 +- +-快速多session评测入口为 `tools/probe_kimodo_closed_loop.py`,导出的manifest兼容 `evaluate_human.py` 和 `render_overfit.py`。`--teacher-forced`只能生成明确标记的诊断结果,不计入闭环成绩。每次评测必须使用新目录,避免误用其他权重的旧输出。 +- +-工作数据、临时日志与权重分别在 `cache/kimodo-stability-20261021`、`setup/kimodo-stability-*` 和 `/dev/shm/kimodo-stability-*`,不重新往 `runs` 堆积中间数据。`runs`只保存最终对照视频。 +- +-## 结果 +- +-**结论:混合闭环过拟合仍未通过,不推荐启动正式Kimodo训练。** 已落地表示一致性、导数标签与续训状态修正;在线历史、执行帧加权、动态损失和短预测时长没有形成可靠的混合训练配方。 +- +-下表为原始六条训练session完整预测历史rollout的等权宏平均,全部仅第一帧GT初始化、固定seed12345、DDIM20、CFG2,无后处理。累计步数包含公共初始化,表中不同累计预算不能解释为单因素同预算对照。 +- +-| 配方 | 累计更新 | 身体RMSE(rad) | 根RMSE(m) | 根速度RMSE(m/s) | +-| --- | ---: | ---: | ---: | ---: | +-| 原方案公共初始化 | 2000 | 0.4785 | 0.5938 | 0.2945 | +-| 原方案继续训练 | 3000 | 0.4186 | 0.5964 | 0.2944 | +-| 执行帧加权 | 3000 | 0.4353 | 0.5961 | 0.2946 | +-| 加权+整段在线历史 | 3000 | 0.3540 | 0.5789 | 1.2121 | +-| 高噪声+历史课程 | 3000 | 0.3673 | 0.2337 | 0.4767 | +-| 强动态监督/GT历史 | 5000 | 0.3583 | 0.2913 | 1.1430 | +-| 强动态监督/课程历史 | 5000 | 0.3470 | 0.3921 | 0.4033 | +-| 因果编码+强动态监督/GT历史 | 5000 | 0.4352 | 0.4061 | 0.9534 | +-| 因果编码+较轻动态监督/课程历史 | 6000 | 0.3231 | 0.4638 | 1.0686 | +-| 同预算追加/12帧预测 | 7000 | 0.3854 | 0.4401 | 1.3143 | +-| 同预算追加/60帧预测 | 7000 | 0.3387 | 0.3743 | 1.0379 | +- +-较轻课程组身体误差比3K原方案低约23%,但根速度误差从0.2944升至1.0686m/s,不能以身体误差下降宣称修复。其行走/停止样本根路径21.83m,GT5.29m,根速度误差4.7618m/s;backchannel运动速度仍只有GT约15%。强动态监督的在线版本相对稳定一些,但hesitation/pause仍偏静止、行走仍有额外速度。因果root编码解决了实际表示不一致,却没有独立带来混合动作质量提升。 +- +-12帧与60帧预测从同一个较轻课程组最终权重开始,均追加1,000步GT历史训练,LR1e-5→1e-6、batch8、运动抽样70%,仅预测长度不同。12帧组的三项宏平均均未优于60帧组,故默认保留60帧。该对照不是把不同预算或不同初始化混为预测时长的效果。 +- +-较轻课程组的六条验证session宏平均为身体0.4430rad、根0.4699m、根速度0.1859m/s,仅作小规模诊断,不是全库泛化或独立测试成绩。CFG1训练宏平均为0.3747rad/0.4654m/1.2224m/s,也没有改善本轮问题;验证结果与逐条指标见机器报告。 +- +-此处为旧采样规则下的历史对照;当前诊断配置和修正后的采样结果见文首。在线历史、执行帧加权及12帧预测未采纳为默认;其他四个baseline配置未改动。 +- +-[完整配置、逐条指标与检查记录](../reports/kimodo_closed_loop_stability_20261021.json)。旧采样规则阶段共11条训练路径(包含公共初始化),有共享初始化关系,不是11次独立随机实验。 +- +-视频均含原始音轨,分别先对完整预测和人体GT做一次GMR,再比较G1轨迹;不是逐窗口重置retarget。示例: +- +-- [原方案:行走/停止](/root/autodl-tmp/runs/kimodo-closed-loop-20261021/control/train-emergency-stop-000036/retargeted_human_comparison.mp4) +-- [较轻课程:行走/停止,仍有抖动](/root/autodl-tmp/runs/kimodo-closed-loop-20261021/causal-online-lite/train-emergency-stop-000036/retargeted_human_comparison.mp4) +-- [较轻课程:犹豫样本](/root/autodl-tmp/runs/kimodo-closed-loop-20261021/causal-online-lite/train-hesitation-000087/retargeted_human_comparison.mp4) +-- [12帧预测诊断:犹豫样本](/root/autodl-tmp/runs/kimodo-closed-loop-20261021/horizon12/train-hesitation-000087/retargeted_human_comparison.mp4) +- +-保留范围:报告和代码在仓库,工作缓存位于cache;runs中本轮只新增MP4。所有本轮过拟合权重已清理,正式OMG/TextOp权重及GenMo活跃输出保留。旧中间初始化权重不再可直接复用,复现需按报告中各阶段重建。 +- +-## 表示与速度监督的额外修正 +- +-复查原205个训练窗口,独立编码的历史末帧与整段监督中的同一帧,其虚拟平滑root平均相差0.0213m、最大0.1492m。这是特征分解差异,不是人体GT发生了15cm跳变;但把两份表示在边界拼接后,会影响root→body阶段计算的局部root速度。原先只检查“未来不能改变observed”,没有检查“known部分与target使用同一分解”。 +- +-新增 `representation.causal_root: true`:在原273D表示中使用单向、半径0.06m受限的根轨迹滤波(alpha0.2),并同步重算hips相对位移。FK、全局旋转、人体root物理位置及预训练统计保持一致。此滤波只用于特征分解,不对预测轨迹做事后平滑。前203维已知姿态在prefix与整段编码中一致,改变未来不改变输入,解码后保持原人体运动;已对全部420个train/val窗口检查。 +- +-`hfd_bridge/kimodo_causal_rep.py` 是可选编码实现;准备数据、在线历史、单session推理、多sessionprobe均读取相同配置。启用或切换编码后必须重新准备windows,不能直接把旧编码缓存标成新编码。严格训练合同会拒绝代码或数据已经变化的旧resume。 +- +-对于预测历史,位置目标仍为原绝对GT未来;另外的 `temporal_target` 只提供真实GT速度/加速度标签。构造这些标签时使用最后3个GT过去姿态,但它们不进入observed/mask或网络条件。这样不会把“从错误预测位置瞬间跳回GT”的位移当成期望速度。原GT历史样本的temporal target就是原target。新增边界root速度及实际解码root加速度监督,未配置的权重为0。 +- +-两组动态监督对照均从课程组第1,000步初始化,再训练2,000步、LR3e-5→3e-6:GT历史组与在线课程组使用相同强度的五项导数损失(1、0.5、0.1、2、0.05)。因果编码另做相同预算的GT历史对照。最后的较轻组合改为3,000步,权重0.2、0.2、0.05、0.2、0.005,70%抽样为运动窗口,75%历史混入上限,750步刷新,短2轮→短5轮→完整session;不降低远期监督权重。该组合改变了预算和多项训练策略,不能将其改善唯一归因于因果编码。 +- +-## 续训与确定性 +- +-包含因果编码、预测历史缓存和分开的导数标签的完整282,790,161参数模型,已在确定性CUDA/math attention模式下验证:第3步checkpoint两次恢复到第6步,模型、Adam、scheduler、全局随机状态、窗口采样、历史选择随机数、缓存及刷新元数据逐项完全相同。默认CUDA attention的重放出现过微小数值差异,但缓存和随机状态一致,不能把默认运行称为逐位可复现。确定性检查使用`CUBLAS_WORKSPACE_CONFIG=:4096:8`、`torch.use_deterministic_algorithms(True)`并关闭flash/memory-efficient SDPA;正常实验保留原性能配置。 +- +-## 输入时机的限制 +- +-训练中发现一个具体例子:`train-hesitation-000087` 的首段“performing ride the pony”在mic2.2s释放,最后一个词在mic5.4s释放;GT持续根运动起点约motion6.67s(按>0.15m/s持续0.2s检测)。一些候选在motion2.5s就开始运动。释放时钟没有错位,但文本模型看到可执行的前半句时,并没有显式的讲话/结束状态。这个例子提示输入信息可能限制时机复现,没有证明所有提前/静止问题均由它引起。 +- +-本轮没有加入GT阶段、语音结束标签、动作时钟或额外音频条件,也没有改动其他四个baseline的输入协议。后续若测试语音状态/文字稳定时长,需明确区分可实时计算的因果信号与GT阶段标签,并单独报告协议变化。 +diff --git a/docs/KIMODO_FORMAL_VALIDATION.md b/docs/KIMODO_FORMAL_VALIDATION.md +deleted file mode 100644 +index 0cf05ca..0000000 +--- a/docs/KIMODO_FORMAL_VALIDATION.md ++++ /dev/null +@@ -1,72 +0,0 @@ +-2026-10-24:修复后新100K的正式评测已完成,见[最新报告](HUMAN_FORMAL_VALIDATION_20261024.md)。本文记录旧失败训练的历史评测,不能把下文100K结果当成新权重的结果。 +- +-2026-10-23 更新:用户随后要求修复训练并删除旧失败目录;原run及其全部权重已删除,本文数值、分析缓存和视频保留。训练配置及修改见[训练修复](KIMODO_TRAINING_REPAIR.md)。当前同名run/权重目录已被新训练重新创建。 +- +-2026-10-23:`322190.textop` screen 实际运行的是 Kimodo,已正常完成 100,000 步,耗时约 19.6 小时。 +-本轮对六类未参与训练的 validation session 做了完整闭环生成、整段 GMR retarget 和有声视频评测。 +-100K 权重已有明确的人体姿态与朝向收益,但根位置、窗口衔接和部分交互仍较弱,暂不能认定全双工指标已经稳定可用。 +- +-同日进一步检查确认 100K 权重的根网络已退化为近似常数;因此撤回单凭平均姿态误差推荐 100K 整模的建议。诊断结论见[根网络退化检查](KIMODO_ROOT_COLLAPSE_DIAGNOSIS.md)。 +-`best.pt` 实际为 2K 权重:它的去噪验证 loss 最低;初轮六段连续生成的平均姿态误差逊于 100K,但根网络仍有动态响应。 +-初轮评测没有修改训练配置或重新训练;失败权重已在后续修复阶段按用户要求删除。 +- +-评测约定: +- +-- 每类取 native validation index 中第一个符合条件的英文、具有人体配对、预计 6–20 秒的 session。样本在生成前选定,没有根据效果挑选。 +-- 六类:turn-taking、interruption、emergency-stop、backchannel、hesitation、pause。源动作共 88.4 秒;排除每段第一帧后评分 2,646 帧。 +-- 输入为按既定时间因果释放的英文 GT 词、最近 30 秒上下文和已执行的人体历史。无音频输入、无未来文本或阶段标签。 +-- 仅以第一帧 GT 姿态初始化,之后完全使用预测历史。历史 60 帧,规划 60 帧,每次执行 12 帧,30 Hz。 +-- DDIM 20 步、CFG 2.0、seed 12345。官方、2K、100K 使用相同的适配、文本、随机种子和闭环流程。 +-- 每段完整预测及 human GT 各连续 retarget 一次到 G1;不在窗口之间重置 IK,不作位置对齐、DTW 或轨迹修正。 +-- 下表是六段分别计算后等权平均,单位为 rad 或 m。位置误差为三个坐标的逐元素 RMSE。 +- +-| 权重/对照 | 人体关节旋转 RMSE (rad) | G1 关节 RMSE (rad) | G1 根位置 RMSE (m) | G1 根朝向 RMSE (rad) | +-|---|---:|---:|---:|---:| +-| 官方初始化,采用相同 HFD 适配 | 0.628 | 0.464 | 0.433 | 0.600 | +-| best.pt,2K | 0.643 | 0.467 | 0.463 | 0.825 | +-| last.pt,100K | **0.454** | **0.390** | 0.544 | **0.266** | +-| 保持每段初始 GT 姿态 | 0.431 | 0.366 | 0.365 | — | +- +-G1 主指标的参考是同流程 retarget 的 human GT。100K 相比官方初始化,人体姿态误差降低 27.7%,G1 关节误差降低 15.9%,根朝向误差降低 55.7%,但根位置误差增加 25.7%。 +-人体原生空间的位置 RMSE 同样从 0.502 m 升至 0.629 m,所以位置问题并非只由 retarget 造成。 +-GT 活动帧的人体旋转误差从 0.633 降至 0.478 rad;保持初始姿态的相应误差约 0.447 rad。 +-动作幅度和视频均表明 100K 并非完全静止,但在本轮静止/倾听时间较多的样本上,平均误差仍未超过保持初始姿态的简单对照。 +-该对照不能完成动作任务;也不能把低姿态误差直接当成语义任务成功。 +- +-| 验证样本 | 100K 人体旋转 RMSE (rad) | 100K G1 关节 RMSE (rad) | 100K G1 根位置 RMSE (m) | +-|---|---:|---:|---:| +-| turn-taking:向后跳 | 0.321 | 0.247 | 0.385 | +-| interruption:摸头后切换姿态 | 0.533 | 0.478 | 0.226 | +-| emergency-stop:向左行走并收到停止文本 | 0.350 | 0.286 | 0.899 | +-| backchannel:拄拐后退并收到附和文本 | 0.541 | 0.469 | 1.311 | +-| hesitation:拿物体指令中插入犹豫 | 0.572 | 0.529 | 0.242 | +-| pause:拿物体指令中暂停 | 0.406 | 0.332 | 0.201 | +- +-主要闭环问题集中在窗口接缝。将每个 12 帧执行块的第一帧标为重规划边界,100K 的边界根速度误差为 0.839 m/s,块内部为 0.118 m/s。 +-各段边界帧占比平均 8.35%,却贡献平均 81.5% 的根速度平方误差。 +-官方初始化对应为边界 0.724、内部 0.169 m/s;100K 改善了块内运动,但边界仍有明显跳变。 +-曲线还可见部分样本约 2 秒一次的位置脉冲。这些现象在人体输出中已存在;目前证据只支持“生成窗口衔接未解决”,尚不能断定是某一行代码或单一训练因素。 +- +-训练去噪验证 loss:初始 0.371、2K 最低 0.198、100K 为 0.340。 +-这种 GT 历史下的短窗口去噪指标与执行预测历史的整段质量不一致,不能据此简单判断 2K 最好或 100K 只是过拟合。 +-后续应将固定闭环样本的姿态、位置和边界速度一起用于 checkpoint 选择,并优先排查历史/未来交界处的根表示与解码,再考虑连续性监督。 +-当前没有证据支持直接追加长训练即可解决这些问题。 +- +-视频保存在 `/root/autodl-tmp/runs/kimodo-formal-validation-20261023/`: +- +-- `last100k/`:六段预测→G1 与 human-GT→G1 对比,以及六段 human-GT→G1 与原始 G1 对比,共 12 段。 +-- `best2k/`、`official/`:各六段预测→G1 与 human-GT→G1 对比。 +-- `retargeted_human_comparison.mp4` 左侧为预测,右侧为 human GT 经相同 retarget 得到的参考。 +-- `retarget_reference.mp4` 用于查看 retarget 本身的差异。human-GT→G1 对原始 G1 的平均关节误差 0.226 rad、根位置误差 0.045 m;这不是可从模型误差中减去的常数。 +-- 共 24 个有声 MP4,帧数、帧率、音轨和完整音视频解码均通过检查。runs 中仅导出视频,工作数据在 cache 中。 +- +-这是一轮六段、单随机种子的离线轨迹检查,未运行物理跟踪、任务成功率或完整 HFD 验证集评测。 +-可以确认身体姿态训练有收益;进一步诊断已确认 100K 根阶段退化,需修复后再选择正式比较权重,当前不足以保证正式全双工指标“不难看”。 +- +-完整数据与曲线: +- +-- [数值和样本溯源](../reports/kimodo_formal_validation_20261023.json) +-- [训练验证曲线](../reports/kimodo_formal_validation_curve_20261023.png) +-- [六段闭环运动速度曲线](../reports/kimodo_formal_validation_motion_20261023.png) +-- [24 段视频的校验结果](../reports/kimodo_formal_validation_videos_20261023.json) +-- 本轮输入、原生预测、G1 retarget、完整文本释放记录在 `/root/autodl-tmp/cache/kimodo-formal-validation-20261023/`。 +diff --git a/docs/KIMODO_HFD_OVERFIT_RESULTS.md b/docs/KIMODO_HFD_OVERFIT_RESULTS.md +deleted file mode 100644 +index f9830e2..0000000 +--- a/docs/KIMODO_HFD_OVERFIT_RESULTS.md ++++ /dev/null +@@ -1,66 +0,0 @@ +-# Kimodo 英文人体过拟合验收(2026-10-21) +- +-复核结论:**本轮仅通过训练链路检查,闭环过拟合未通过。** 用户指出视频动作偏静止后,运动幅度审计确认三条训练样本的关节运动速度仅约GT的三分之一。训练急停样本保持初始姿态的关节/根误差为0.3607rad/2.0503m,微调闭环为0.3624rad/2.0289m,基本未超过静止基线。下文误差表作为原始记录保留,不应据其宏平均下降宣称动作复现成功。详见评测目录的 `motion_amplitude_audit.json` 和该样本 `motion_audit.json`。 +- +-训练框架、输入和复现命令见 [Kimodo训练说明](KIMODO_HFD_TRAINING.md)。运行目录为 `/root/autodl-tmp/runs/kimodo-english-overfit-20261021`;视频与逐样本指标入口为 [/root/autodl-tmp/runs/kimodo-english-evaluation-20261021/README.md](/root/autodl-tmp/runs/kimodo-english-evaluation-20261021/README.md)。 +- +-## 已完成的检查 +- +-- train/val 各六个原生英文配对人体 session,分别为 84.4/88.4 秒、205/215 个窗口,覆盖六种交互模式。选择规则在训练前固定;不是全库正式训练。 +-- 27 项单元/数据契约测试通过。实际 session 的坐标转换回环误差:平移不超过 6e-8m、旋转不超过 2.8e-6rad;扰动未来不改变输入历史或 heading。 +-- 文本基础模型无缺失权重;MNTP/supervised 两份适配器各 448 个张量与文件逐项一致。冻结编码得到 81 个文本特征,包含空文本的零向量;双向注意力探针通过。动作环境 `pip check` 通过,旧版文本依赖单独隔离。 +-- 完整 282,790,161 参数模型的 root/body 梯度分别非零且有限,一次 Adam 更新实际改变参数;移除文本会改变预测。 +-- 从第 4 步独立恢复两次至第 6 步,全部模型参数、采样器和两步 loss 完全一致;Adam step、scheduler step 均正确恢复。 +- +-## 1,000 步训练 +- +-使用 batch8、AdamW 1e-5→1e-6、bf16,7 项结构损失和 10% 文本 dropout。固定噪声评估每次抽取 64 个窗口,训练/验证独立抽样,不能理解为遍历全数据集。 +- +-| 指标 | 初始 | 1,000步 | +-| --- | ---: | ---: | +-| 固定训练窗口去噪损失 | 1.61369 | 0.24366 | +-| 固定验证窗口去噪损失 | 1.28607 | 0.60110 | +- +-最佳固定验证损失为第 800 步的 0.55017。主闭环对照明确使用最后第 1,000 步,避免把窗口损失最优当成闭环最优。每步中位耗时 0.0896s,峰值显存 5.33GiB;第 6→1,000 步续训进程共约 107 秒,含验证和保存。此吞吐只代表小型预制窗口与缓存文本,不代表全量动态数据加载和 8B 文本编码成本。 +- +-## 闭环对照 +- +-训练和验证各选 turn-taking、interruption、emergency-stop 三条 session;官方初始化/微调模型均使用相同 seed、DDIM20、CFG2、单帧 GT 初始化和后续预测历史。下表为各三条的逐 clip 宏平均,无根对齐或 DTW。 +- +-| 范围 | 关节旋转RMSE,初始→微调(rad) | 根位置RMSE,初始→微调(m) | +-| --- | ---: | ---: | +-| 训练三条 | 0.5711 → 0.3492 | 1.5246 → 0.7362 | +-| 验证三条 | 0.6580 → 0.4471 | 0.9332 → 0.4737 | +- +-验证宏平均关节误差下降约32%,根位置误差下降约49%。这是小规模探索性验证,三条样本不足以证明泛化性能。 +- +-| 样本 | 关节初始→微调(rad) | 根位置初始→微调(m) | +-| --- | ---: | ---: | +-| train turn-taking 13 | 0.5262 → 0.3260 | 0.8745 → 0.0323 | +-| train interruption 73 | 0.6752 → 0.3591 | 1.0668 → 0.1476 | +-| train emergency-stop 36 | 0.5118 → 0.3624 | 2.6323 → 2.0289 | +-| val turn-taking 3 | 0.5458 → 0.3664 | 1.0103 → 0.3108 | +-| val interruption 75 | 0.7042 → 0.4871 | 1.0318 → 0.0803 | +-| val emergency-stop 200 | 0.7240 → 0.4877 | 0.7574 → 1.0299 | +- +-每条完整预测与人体 GT 分别做 GMR,再与原始 G1 作三组比较;主对照共 36 条带音轨视频。人体指标先于 retarget 计算,GT→G1 的差异单独保留,不能把它当作可直接相减的误差下界。视频为姿态回放,不是物理跟踪或任务成功率评测。 +- +-加上两组历史/CFG诊断共42条视频,全部逐帧解码通过,共16,812帧,音轨也全部可解码。校验清单在评测目录 `video_checks.json`。这次20步DDIM的单轮规划平均约0.46–0.65秒,尚未满足每0.4秒执行窗口的严格实时要求;当前验收针对离线推理。 +- +-## 急停样本与历史诊断 +- +-训练急停样本的 GT 先行走约 5 米再停下,微调闭环预测大多停留在起点附近。它的问题主要是没有启动该段行走,不能简单解释成“听到停止后仍一直向前”。该样本原始动作描述为 `stop fast walk forward`,随后释放 `Pause for a moment.`;没有为改善结果改写文本。 +- +-| train emergency-stop 36诊断 | 关节RMSE(rad) | 根位置RMSE(m) | +-| --- | ---: | ---: | +-| 预测历史,CFG2,主结果 | 0.3624 | 2.0289 | +-| GT历史,CFG2,仅诊断 | 0.0896 | 0.0910 | +-| 预测历史,CFG1,仅诊断 | 0.4779 | 2.0421 | +- +-这说明模型在给定正确历史时能较好地拟合该轨迹,但闭环可能停留在静止状态;降低 CFG 未解决问题。GT历史结果不能替代主闭环指标,也不能独立证明唯一原因。验证急停样本的根误差还从 0.757m 升到 1.030m,因此不能宣称已经学会急停。 +- +-结论是微调框架与梯度路径可用,小规模训练已改善多数样本;当前仍有闭环启动、状态切换和根运动不足的问题。下一阶段应扩大训练覆盖,增加连续多轮训练/预测历史扰动,并单列启动与停止指标,再决定正式训练预算;只延长这六条 session 的训练不能保证解决上述问题。纯文本输入在词释放前也没有语音起点信息。 +- +-## 留存 +- +-运行目录保留配置、数据/文本审计、源码与环境快照、loss 曲线和完整 resume 检查。当前 `best.pt` / `last.pt` 在 `/dev/shm/kimodo-english-overfit-20261021`,属于临时测试权重,重启不保留;两份独立恢复测试的重复 checkpoint 已清理,日志仍保留。GenMo 的 GPU1 正式训练未停止。 +diff --git a/docs/KIMODO_HFD_TRAINING.md b/docs/KIMODO_HFD_TRAINING.md +deleted file mode 100644 +index 4810ba7..0000000 +--- a/docs/KIMODO_HFD_TRAINING.md ++++ /dev/null +@@ -1,112 +0,0 @@ +-# Kimodo HFD 全量微调与小规模验收 +- +-本地入口保留 Kimodo-SMPLX-RP-v1 的 root→body 两阶段网络与官方初始化,增加监督训练、完整断点恢复、连续推理、人体指标、整段 GMR 和视频评测。上游代码固定在 `e1c219c99b4b94ba48be1fb6a1cbc64ecee4c954`,不修改上游实现。 +- +-2026-10-24:修复版已从官方初始化完成新一轮100K训练;当前best与last均为100K,新正式权重保留。根骨干200个参数张量仍与官方初始化逐位一致,梯度和时序响应正常,未重现旧恒定输出失效。六段验证的G1根位置RMSE为0.264m,但后退与部分指令跟随仍不足。完整数值和有声视频见[本轮正式评测](HUMAN_FORMAL_VALIDATION_20261024.md)。 +- +-训练方式见[训练修复与验收](KIMODO_TRAINING_REPAIR.md):冻结根 Transformer 骨干,仅微调根输入/输出投影,身体保持训练;增加实际解码运动连续性监督,分阶段学习率及梯度裁剪。此前失败正式训练的run和权重已经删除,当前同名目录属于新训练。以下早期过拟合结果保留为历史对照。 +- +-历史混合闭环过拟合未通过。较好的早期诊断路线为归一化x0目标、2,000步初训加1,000步GT历史微调:行走样本达到约0.056rad/0.040m,但其他模式相位与泛化仍不足。历史实验见[静止退化复查](KIMODO_MOTION_RECOVERY.md),完整对照见[闭环稳定性报告](KIMODO_CLOSED_LOOP_STABILITY.md)。`kimodo_overfit.yaml`和`kimodo_refine_probe.yaml`复现这条诊断路线;`kimodo.yaml`现为全量正式训练配置。未默认启用在线历史混入。旧缓存需按当前编码重新准备。 +- +-## 输入、目标与因果边界 +- +-- 原生英文 human 配对样本,按实际 session 组合操作拼接人体动作;G1 只用于配对和评测参考,不进入 Kimodo 输入或损失。小规模验收固定 train/val 各六个 session;正式入口覆盖完整英文原生索引,按需组装配对人体 session。 +-- 文本沿用共同协议:对齐后的前景英文 GT 单词,词结束加 0.1s 后等待麦克风 `0.6+0.4k` 释放,保留最近 30s。动作时刻 `t` 对应麦克风 `t+0.2s`。无音频、翻译、未来文本、交互类型标签输入。 +-- 冻结 LLM2Vec/Llama-3-8B-Instruct,MNTP 适配器合并,再加载 supervised 适配器;逐项核对权重。原生 mean pooling 得到 `1×4096`,网络内部的 50 个文本 token 由其自身投影构造。文档预算 400 token,超长保留最近文本;模板总预算 512,编码内部 batch=1。 +-- 30Hz,历史60帧、默认预测60帧、执行12帧。12帧预测仅用于诊断,未采纳为默认。历史mask锁定203维姿态,70维速度/接触辅助量不作为已知约束。当前配置使用因果root特征分解,使历史在输入和监督中一致,不改变物理人体动作;首个已知姿态确定heading与平移坐标系。 +-- 输出原生 273D 特征,由原生 inverse 解码到 22 个 SMPL-X 关节和根轨迹,再逆变换到 AMASS;手部中性。保留 session 体型,显式处理其 pelvis 偏移。完整输出与人体 GT 分别 GMR 到 G1。 +- +-当前`training.objective: normalized_x0`在归一化273D特征上做x0均方误差,root与body分别取均值后相加;仅未来有效帧计分,末端无下一帧时速度/接触不计分。保留原生两阶段之间的detach,两阶段各有直接梯度。诊断配置使用50% t=999、其余均匀t,以及70%运动窗口抽样。原七项物理量SmoothL1配方通过`physical_huber`保留作对照,YAML的`training.weights`只在该模式生效;不宣称复现官方未公开训练配方。 +- +-推理从一个已知初始姿态启动,后续历史来自已执行预测;CFG两支保留相同历史。历史70维未观察辅助量在每个DDIM步恢复为训练时的纯噪声边缘分布,避免把未监督的历史预测继续反馈;`legacy`仅用于旧规则对照。默认DDIM20、CFG2,无轨迹后处理。`--teacher-forced`仅诊断使用。 +- +-## 环境 +- +-动作使用 `/root/autodl-tmp/envs/kimodo/bin/python`,保留上游 Transformers 5.1.0。原生 LLM2Vec 的双向 Llama 代码需要独立旧版依赖,在 `/root/autodl-tmp/envs/kimodo-text-compat` 保存 `requirements/kimodo-text.txt` 所列包;仅文本进程将其加入 Python 路径。可用 `KIMODO_TEXT_COMPAT` 指向其他安装目录。 +- +-重新安装文本 overlay 时,应先用用户要求的 `/root/.local/bin/turbo` 调用 pip,并用 `--no-deps --target <目录> -r requirements/kimodo-text.txt`,复用动作环境已有的 Torch 等依赖。不要直接将该 requirements 安装进动作环境。文本 worker 自检未来 token 能影响前面 hidden state,防止误用单向注意力;小规模离线缓存完成后动作训练不加载 8B 编码器。全量训练采用常驻独立文本进程,与动作训练共用所选 GPU;缓存命中时不重复编码。 +- +-## 全量正式训练(2026-10-23 修复版) +- +-```bash +-cd /root/autodl-tmp/hfd-baseline-portable +-/root/autodl-tmp/envs/kimodo/bin/python -u train_kimodo.py --config configs/kimodo.yaml +-``` +- +-中断恢复使用相同命令追加 `--resume`,读取完整 `last.pt`,恢复模型、Adam、余弦调度器、全局 RNG、抽样 RNG 和每个 batch 槽位的 session;100K 是总步数。配置/实现/数据身份变化会拒绝恢复。 +- +-- GPU 0;batch 8;100K 步;BF16。根仅训练输入/输出投影,AdamW 2e-6→2e-7,在20K衰减到底;身体AdamW 1e-5→1e-6,在100K衰减到底;500步warmup,两阶段分别裁剪到1。因果root、normalized x0、10%强制t=999(其余均匀t)、文本dropout 0.1;执行12帧权重1,剩余规划帧0.25。增加原始根历史重建及实际解码速度/加速度监督;根注意力采用数学后端,运行配置随checkpoint保存并恢复。全量所有合法窗口均匀边缘抽样,不启用在线历史课程。具体权重以`configs/kimodo.yaml`为准。 +-- train 原生英文 155,117 个 session;val 独立原生英文 4,719 个 session;没有6–20秒筛选、没有每模式数量上限、无第三方G1动作、无中文。初始化统计会打印真实覆盖数。 +-- 不预拼155K个session。每个 batch 槽位独立按保守长度上界拒绝抽样,再均匀选0.4s决策点;训练槽位复用16个batch,验证复用4个batch。所有session有被抽到的机会,预算内不保证每个session都被访问;100K不是一个epoch。 +-- paired SMPL-X 动态拼装与 GenMo 相同来源和时间轴;本地512MiB窗口LRU、每split512MiB原始sessionLRU;冻结LLM2Vec独立进程按需编码,每次内部batch1,SQLite缓存本地 `/root/autodl-tmp/cache/kimodo-text-full`,特征payload上限4GiB,清缓存不改变抽样顺序。编码器不反传;主模型正常反传。 +-- 文本仍是按单词释放的最近30s英文GT,60帧历史,60帧目标,12帧执行,无音频或未来语义输入。正式训练依然用GT历史;闭环推理用预测历史。小样本泛化差属于质量风险,继续通过全量独立验证和闭环视频观察。 +-- 首次启动要构建train/val原生索引、加载8B文本模型并编码首次抽样文本,启动阶段可能数分钟。`FULL_DATA`、`DATA`是正常准备进度。每10步打印`TRAIN`;`seconds`含数据等待和更新,`data_seconds`单独显示组装/文本等待。 +-- 每2K步执行固定种子16×8个验证窗口,并记录同种子训练诊断;这些窗口从完整val索引抽取,绝非全val遍历。噪声损失用于选`best.pt`,不能替代闭环动作评测。 +-- 日志/配置在 `/root/autodl-tmp/runs/hfd-kimodo-english-train`;权重在 `/root/autodl-tmp/models/kimodo/hfd-english-train`,保留完整last和最近2个完整快照,以及仅模型best。本地唯一临时文件完成flush/fsync后原子替换;写入失败最多重试3次,保留上一断点。快照用硬链接避免重复写入。 +- +-训练结束后,使用正式run选择六模式验证session,执行预测历史→整段GMR→带音频视频: +- +-```bash +-/root/autodl-tmp/envs/kimodo/bin/python -u evaluate_kimodo.py \ +- --run /root/autodl-tmp/runs/hfd-kimodo-english-train \ +- --output /root/autodl-tmp/runs/kimodo-full-validation \ +- --splits val --variants trained --gpu 0 +-``` +- +-全量评测按每模式第一个满足6–20s的原生英文session作固定可视化子集,动作G1仅评测导出,训练不加载。推理可额外使用`--checkpoint`选择其他完整快照。评测GPU需要容纳动作模型和独立文本模型;勿与该GPU上的训练并发。批量全val指标可后续扩大选择,当前命令只展示六例。 +- +-启动验收:44项单元测试通过;实际全量入口完成2步、保存完整断点后恢复到3步,核对下一批位置、session槽位、抽样RNG、Adam和调度器。运行统计身份误变问题已修复。临时测试权重已删除;这不是100K质量验收。报告见 [`kimodo_full_training_ready_20261021.json`](../reports/kimodo_full_training_ready_20261021.json)。 +- +-## 小规模诊断命令 +- +-```bash +-cd /root/autodl-tmp/hfd-baseline-portable +-PY=/root/autodl-tmp/envs/kimodo/bin/python +-$PY prepare_kimodo.py --config configs/kimodo_overfit.yaml +-$PY cache_kimodo_text.py --run /root/autodl-tmp/cache/kimodo-overfit --gpu 0 +-$PY train_kimodo.py --config configs/kimodo_overfit.yaml --offline-check +-$PY train_kimodo.py --config configs/kimodo_overfit.yaml +-# 中断后恢复,max_steps 是总预算: +-# $PY train_kimodo.py --config configs/kimodo_overfit.yaml --resume auto +-``` +- +-复现第二阶段时,复用首阶段的冻结输入,再从其权重warm-start(首阶段权重需暂时保留): +- +-```bash +-$PY - <<'PY' +-from pathlib import Path +-from omegaconf import OmegaConf +-base = OmegaConf.load('configs/kimodo_overfit.yaml') +-refine = OmegaConf.load('configs/kimodo_refine_probe.yaml') +-destination = Path(refine.output) +-destination.mkdir(parents=True, exist_ok=True) +-for name in ['data.json', 'text_features.pt', 'text_features.json', 'texts.json']: +- target = destination / name +- if not target.exists(): +- target.symlink_to(Path(base.output) / name) +-OmegaConf.save(refine, destination / 'resolved_config.yaml') +-PY +-$PY train_kimodo.py --config configs/kimodo_refine_probe.yaml +-``` +- +-完整checkpoint包含网络、AdamW、cosine scheduler、全局step、Torch/CUDA/Python/NumPy随机状态和窗口采样器;启用在线历史时还包含历史缓存、对应policy step与混入选择随机数。恢复核对数据、缓存文本、代码、官方权重/统计及配置。`--stop-after 4`可用于检查。`best.pt`按固定验证去噪损失选,`last.pt`保存最后状态;闭环质量另行评测。逐位复现需显式启用确定性CUDA配置,见最新报告。 +- +-```bash +-$PY evaluate_kimodo.py \ +- --run /root/autodl-tmp/cache/kimodo-overfit-refine \ +- --output /root/autodl-tmp/cache/kimodo-evaluation \ +- --checkpoint /dev/shm/hfd-kimodo-overfit-refine/last.pt --gpu 0 \ +- --modes turn-taking interruption emergency-stop backchannel hesitation pause +-``` +- +-默认分别报告训练/验证的 turn-taking、interruption、emergency-stop,每个对照官方初始化与微调模型。先计算人体根/旋转误差,再按每个完整轨迹做 GMR,保存 predicted human↔retargeted GT、predicted human↔原始 G1、retargeted GT↔原始 G1 三组比较。视频含原始音轨和文本释放记录。两种模型共用 seed、采样步数、CFG 与历史策略;验证集用于诊断,不作为独立测试集。 +- +-首阶段YAML为2,000步、batch8、AdamW 3e-5→3e-6;`kimodo_refine_probe.yaml`再从首阶段模型权重初始化,重置Adam,以1e-5→1e-6训练1,000步。第二阶段不是`--resume`,而是显式warm-start。工作数据和日志在cache,临时权重在/dev/shm,runs只存视频。正式训练仍需另行实现更大session范围并验证预算与泛化。 +- +-## 验收产物 +- +-cache中的工作目录记录数据、文本审计、完整模型梯度检查与训练指标;runs只保存筛选后的MP4。首轮失败诊断见[过拟合验收](KIMODO_HFD_OVERFIT_RESULTS.md),最新对照、视频与保留范围见[闭环稳定性](KIMODO_CLOSED_LOOP_STABILITY.md)。 +- +-2026-10-22 保存故障修复:旧训练在64K保存时无法在共享盘创建临时文件,62K完整断点及best已归档到 `/root/autodl-fs/pangxinlei/models/kimodo/archive/hfd-english-train-failed-20261022`。默认正式配置改用本地盘保存新权重,旧失败run的日志等已清理。标准启动命令从头训练;旧归档含匹配原断点契约的训练脚本副本,新代码不会自动绕过旧断点的实现校验。见[故障报告](../reports/kimodo_checkpoint_failure_20261022.json)。 +- +-2026-10-23 旧一轮正式训练完成 100K 步。六类 held-out session 的历史数值与 24 段有声视频见[正式训练验证](KIMODO_FORMAL_VALIDATION.md)。进一步确认 100K 根网络退化为近似常数;替换早期根权重也尚未恢复正确后退。随后加入的退化检测和自动停止已按用户要求全部撤回。本轮改为修复训练策略,旧失败run及其权重已全部删除,历史视频和分析数据保留;具体原因见[训练失效机制](KIMODO_ROOT_TRAINING_CAUSE_ANALYSIS.md)。 +diff --git a/docs/KIMODO_MOTION_RECOVERY.md b/docs/KIMODO_MOTION_RECOVERY.md +deleted file mode 100644 +index 6cc4edb..0000000 +--- a/docs/KIMODO_MOTION_RECOVERY.md ++++ /dev/null +@@ -1,78 +0,0 @@ +-# Kimodo 静止退化复查与恢复实验 +- +-本文保留历史恢复实验。后续在线历史课程、因果root表示修正和同预算预测时长对照见[最新闭环报告](KIMODO_CLOSED_LOOP_STABILITY.md);混合闭环验收仍未通过。 +- +-用户观看视频后指出预测几乎不动。复查确认:[原1,000步实验](KIMODO_HFD_OVERFIT_RESULTS.md)的闭环过拟合**没有通过**。宏平均误差下降不足以证明动作复现;先前报告对此解释不充分。 +- +-## 直接证据 +- +-三条原训练评测的预测关节运动速度仅为GT的32%–36%。`train-emergency-stop-000036` 的根路径为0.83m,GT为5.29m;一直保持初始姿态的关节/根误差为0.3607rad/2.0503m,模型为0.3624rad/2.0289m,几乎没有超过静止基线。 +- +-新增 `audit_kimodo_motion.py --inference <目录>` 同时报告原人体误差、静止初始姿态基线、根路径、关节角速度、GT活跃区间误差和持续运动启动时间。运动幅度不能单独证明正确性:抖动和错误动作也可能增加路径/角速度。启动检测仅为诊断,使用根速度>0.15m/s持续0.2s,不是通用任务成功判据。 +- +-## 训练改动与对照 +- +-输入协议保持不变:已释放英文文本、60帧历史、60帧未来、每轮执行12帧;无音频、GT阶段、未来姿态或剩余时长条件。文本缓存和预训练动作参数均沿用原实验。 +- +-`hfd_bridge/kimodo_recovery.py` 增加可选训练组合: +- +-- 在root阶段输出传给body阶段之前锁定已知历史root,保持两阶段条件一致。旧实现仅在两阶段完成后锁定最终输出。单独给旧权重加此约束,根误差仅2.0289→2.0265m,**它本身没有解决静止问题**。 +-- 原速度损失针对独立的velocity特征块,不能保证实际解码姿态在动。增加由解码root、FK关节和旋转增量计算的时序监督,权重为1、2、0.5;不是只拟合辅助速度输出。 +-- 每批50%样本使用t=999,增加从近纯噪声依靠文字/历史生成的训练;其余扩散步仍均匀采样。 +-- 训练按GT运动标签将70%抽样分给明显运动窗口,保留30%等待/停止窗口。六条原始session有83个活跃窗口、122个其余窗口。标签只用于训练抽样,不输入模型。 +- +-## 单样本、同预算对照 +- +-均从原六条训练的1,000步模型初始化,只训练同一条急停/行走session,新增1,000步、batch8、LR3e-5、相同cosine预算、text dropout=0、同seed。原配方对照保留七项损失与均匀窗口/扩散步;恢复配方同时启用上述改动。单条实验是定位问题,不是泛化评测。 +- +-| 配方(预测历史、CFG2) | 关节RMSE(rad) | 根RMSE(m) | +-| --- | ---: | ---: | +-| 原六条1,000步模型 | 0.3624 | 2.0289 | +-| 原配方单条新增1,000步 | 0.3164 | 0.5658 | +-| 恢复组合单条新增1,000步 | 0.0427 | 0.1332 | +- +-恢复组合开始真正行走,持续运动起点约3.17s,GT约3.33s;关节速度是GT的1.23倍。不过根路径7.31m仍高于GT5.29m,有抖动/多余位移。CFG1诊断为0.0437rad/0.0759m,根路径7.08m;不能只看较低的位置误差忽略多余路径。原配方更充分训练也改善根误差,且恢复实验同时改变多项因素,不能将改善完全归因于某一项损失。 +- +-[单条真实闭环视频](/root/autodl-tmp/runs/kimodo-motion-recovery-20261021/probe1000/evaluation/retargeted_human_comparison.mp4),仅第一帧GT初始化;[动作/位移曲线](/root/autodl-tmp/runs/kimodo-motion-recovery-20261021/single_motion_comparison.png)。 +- +-## 六条混合实验 +- +-`configs/kimodo_motion_recovery_six.yaml` 从原1,000步模型继续微调3,000步,恢复10%文本dropout,默认CFG2。六种训练模式均用模型自己历史做完整rollout。 +- +-| 训练模式 | 关节RMSE(rad) | 预测/GT关节角速度比 | +-| --- | ---: | ---: | +-| turn-taking | 0.3158 | 0.360 | +-| interruption | 0.3138 | 1.175 | +-| emergency-stop | 0.0912 | 1.354 | +-| backchannel | 0.3228 | 1.110 | +-| hesitation | 0.5201 | 0.269 | +-| pause | 0.4859 | 0.568 | +- +-因此单条恢复**没有自动推广到六条混合过拟合**。行走改善,部分舞蹈仍运动不足;幅度接近GT的样本也可能动作/相位错误。hesitation给GT历史时关节误差0.0630rad,而闭环为0.5201rad,说明历史状态分布差异仍然很大。GT历史只用于诊断,不计入验收结果。 +- +-## 预测历史训练 +- +-`prepare_kimodo_predicted_history.py` 仅对六条训练session增加一份由已完成模型闭环生成的历史视图,验证集保持原样。每个训练窗口的60帧输入来自缓存的真实预测历史;未来监督来自GT。专家相对位移从当前预测根位置出发,避免要求瞬移回已错过的GT位置;前6帧身体旋转用0.2s过渡接向专家姿态。此过渡属于训练目标构造,不是推理后处理。 +- +-原GT历史视图与预测历史视图各205个窗口,训练均使用最近已释放文字。数据记录输入预测的哈希与plan ID,并拒绝GT历史rollout;`augmentation`视图不当作额外独立评测session。新增 `configs/kimodo_predicted_history.yaml`,从六条3,000步模型初始化,再训练2,000步。不能把六条的两种历史视图称为12条独立训练session。 +- +-所有阶段均为固定小规模数据诊断,不代表全库训练已经准备就绪。后续应按闭环动作幅度、与静止基线的比较、动作一致性与启动/停止共同判断,不能以去噪loss或平均姿态误差单独宣布通过。 +- +-## 预测历史实验结果与结论 +- +-完成新增2,000步后,仍使用相同seed、DDIM20、CFG2、一个GT初始姿态和后续完整预测历史,评估原始六条训练session(不把增强视图当成新样本)。 +- +-| 训练模式 | 关节RMSE(rad) | 预测/GT关节角速度比 | +-| --- | ---: | ---: | +-| turn-taking | 0.4061 | 1.458 | +-| interruption | 0.3091 | 1.217 | +-| emergency-stop | 0.1933 | 1.910 | +-| backchannel | 0.4659 | 1.615 | +-| hesitation | 0.3127 | 1.197 | +-| pause | 0.3637 | 1.403 | +- +-角速度比的精确数值以各session的 `motion_audit.json` 和汇总 `motion_recovery_comparison.json` 为准。相比上阶段,hesitation/pause的静止退化缓解,但有些模式更不准确。急停根位置误差0.7804m、速度RMSE3.9659m/s,明显抖动;turn-taking虽然根位置误差只有0.0595m,关节误差仍0.4061rad。这再次说明移动更多或根误差更小,都不能代替动作复现。 +- +-**当前六条混合闭环过拟合仍未通过。** 不将这两个混合恢复版本替换为正式配方,不推荐直接扩大全量训练。单条真实动作已经可以学到,但混合数据的闭环启动、状态切换、动作相位和轨迹稳定性尚不可靠。实验支持历史分布差异的重要性,没有证明它是唯一原因;缓存一轮预测历史也没有充分解决问题。现阶段继续盲目增加正式训练步数,无法保证排除静止或错误动作。 +- +-本次30项回归测试通过;预测历史缓存的因果边界、输入预测哈希和已知物理历史一致性检查通过。链路正确性与动作验收分开报告。所有新实验保留数据盘日志和视频,测试权重仍暂存在`/dev/shm`;原GenMo GPU1训练未停止。 +diff --git a/docs/KIMODO_ROOT_COLLAPSE_DIAGNOSIS.md b/docs/KIMODO_ROOT_COLLAPSE_DIAGNOSIS.md +deleted file mode 100644 +index 6fa6855..0000000 +--- a/docs/KIMODO_ROOT_COLLAPSE_DIAGNOSIS.md ++++ /dev/null +@@ -1,55 +0,0 @@ +-2026-10-23:后退失败的直接原因已定位到 **100K 权重中的根轨迹网络退化为近似常数输出**。 +-窗口坐标转换与已知历史保留检查通过。单纯换回早期根网络仍未恢复正确后退,因此还需修复根网络训练稳定性与闭环动作续接,不能将本轮替换对照视为可用修复。 +- +-证据来自已有 validation 样本 `val-backchannel-000057`,未用该样本训练: +- +-- 输入正确:3.0 秒已释放完整 `walking backwards with crutches`,7.8 秒才追加 `Wow.`。去掉 `Wow.` 后仍反向漂移。 +-- GT 的人体坐标编码→解码→导出最大平移误差约 0.0000083 m;采样末尾的已知历史解码误差为 0。 +-- 在决策帧 0、132、240、396,使用预测历史或 GT 历史,100K 根网络都输出几乎相同的未来位置。 +-- 在几乎无噪声的输入(t=0,提供加噪后的 GT 未来)上也如此。t=526/999、FP32/BF16、train/eval 均复现,排除了仅由半精度或评估模式引起的异常。 +-- 官方初始化和 2K 网络在相同输入下输出随时间变化的根轨迹。 +- +-下面对八个 GT 存在明显位移的窗口进行额外检查,混合 t=0/526/999。运动 RMS 定义为有效未来帧相对其时间均值的三维位置偏差 RMS,使用根阶段的 smooth-root 通道,不让身体阶段的髋部残差掩盖根阶段退化。 +- +-| 网络 | 预测根轨迹运动 RMS (m) | GT RMS (m) | 近似常数窗口 | +-|---|---:|---:|---:| +-| 官方初始化 | 0.1594 | 0.1958 | 0 / 8 | +-| 2K | 0.2381 | 0.1958 | 0 / 8 | +-| 100K | **0.000057** | 0.1958 | **8 / 8** | +- +-现有窗口以已知历史第一帧确定原点/朝向,编码解码的逆变换本身成立。 +-但退化网络不断在该原点附近输出几乎相同的位置,而不是延续历史末尾位置。 +-当历史以 60 帧长度滚动时,原点跟随过去预测变化,从而出现约两秒一次的位置脉冲。 +-这是“近似常数预测 + 滚动窗口原点”共同产生的现象,不能简单认定为坐标轴翻转或平移加错一次。 +- +-本轮进行了仅替换根网络、保留 100K 身体网络的闭环消融,均保持原始因果文本、随机种子、窗口和采样配置,不添加轨迹修正或 GT 历史: +- +-| 组合 | 沿 GT 后退方向的净位移 (m) | 人体位置 RMSE (m) | 人体姿态 RMSE (rad) | +-|---|---:|---:|---:| +-| GT | +3.498 | — | — | +-| 100K 整模 | -0.780 | 1.513 | 0.541 | +-| 2K 根 + 100K 身体 | -0.075 | 1.300 | 0.570 | +-| 官方根 + 100K 身体 | +0.183 | 1.132 | 0.706 | +- +-替换缓解了漂移,但没有完成约 3.5 m 的后退;身体姿态误差还增大。 +-这表明常数根网络确实是问题之一,现有早期根权重也不足以解决完整 HFD 动作续接。 +-两个替换对照均已整段 retarget 并保存有声视频: +- +-- `/root/autodl-tmp/runs/kimodo-root-diagnosis-20261023/root2k-body100k/backchannel.mp4` +-- `/root/autodl-tmp/runs/kimodo-root-diagnosis-20261023/root-official-body100k/backchannel.mp4` +- +-训练日志提供了时间线索:1–6K 的平均 normalized-root loss 约 0.042,6.5–7K 升至约 0.149,7–8K 约 0.220,而身体 loss 仍在缓慢下降。 +-没有保留这个区间的逐步权重,无法确定精确退化时刻,也未分别消融学习率、高噪声比例或样本分布来确定触发因素。 +-可以确定训练后的根阶段失效;不能把“原因就是学习率过大”等猜测当作已验证结论。 +- +-用户随后明确要求分析原因,不加入退化检测。上一轮加入的检测模块、测试、配置和自动停止已全部撤回,训练脚本恢复为本次 100K 正式训练时的原版本。原因分析见[根网络训练失效机制](KIMODO_ROOT_TRAINING_CAUSE_ANALYSIS.md)。 +- +-正式模型权重与已有视频均保留。本轮没有追加训练,没有生成临时训练 checkpoint。 +-旧 100K 训练脚本快照保存在 `/root/autodl-tmp/models/kimodo/hfd-english-train/source-100k-20261023/train_kimodo.py`。 +-训练脚本已恢复到旧正式训练契约的相同内容;诊断推理仍可正常加载全部旧权重。 +- +-下一轮训练应从健康根权重恢复,分别控制根阶段与身体阶段的更新,加入实际解码位移/接缝监督,并用固定闭环验证选择权重。 +-100K 身体阶段的姿态收益可作为初始化候选;当前不能把 100K 整模或上述混合模型直接作为已经修复的正式 baseline。 +- +-[完整检查数据](../reports/kimodo_root_collapse_20261023.json)。 +-逐窗口特征、模式/精度对照与诊断推理保存在 `/root/autodl-tmp/cache/kimodo-formal-validation-20261023/root-probe/`。 +diff --git a/docs/KIMODO_ROOT_TRAINING_CAUSE_ANALYSIS.md b/docs/KIMODO_ROOT_TRAINING_CAUSE_ANALYSIS.md +deleted file mode 100644 +index aa08523..0000000 +--- a/docs/KIMODO_ROOT_TRAINING_CAUSE_ANALYSIS.md ++++ /dev/null +@@ -1,59 +0,0 @@ +-2026-10-23:本轮分析的核心结论是,微调后的根网络丢失了时序和条件信息,随后落入前层梯度接近零的常数回归状态。上一轮所说的“根网络退化”只是输出表现;这里补充网络内部如何失效、为何训练无法自行恢复,以及哪些训练设置最值得怀疑。 +- +-本文记录旧100K训练的原因分析。分析阶段按用户要求撤回全部退化检测及自动停止,恢复了旧训练代码和配置。用户随后授权修复训练框架并删除失败目录;当前代码和配置已按[训练修复](KIMODO_TRAINING_REPAIR.md)更新,失败run及其全部权重已删除,既有数值、分析缓存和评测视频保留。以下机制测量均来自删除前保存的诊断结果。 +- +-已确认的内部机制: +- +-1. 根 Transformer 的不同时间帧逐层变得相同。在同一 GT 历史、同一接近无噪声未来输入下,2K 网络最后一层的帧间特征 RMS 为 0.0489,100K 只有约 1.08e-7。100K 第 1 层仍有约 0.0601 的差异,到第 8 层约 0.000854,第 13 层约 3.56e-7。信息在 Transformer 内部被抹平,早于坐标解码。 +-2. 深层注意力同时趋于均匀。该输入共 172 个 token;均匀注意力的熵为 ln(172)=5.14749,100K 最后一层实测约 5.14749,2K 约 1.0776。不同帧的同质化与近似全局平均相互强化,最后输出不再区分时间或条件。 +-3. 前层几乎无法再获得有效梯度。以同一根 MSE 对照,2K 对输入的梯度范数为 0.0170,100K 为 4.76e-12,下降约 35.8 亿倍。100K 输入投影权重梯度约 9.50e-11,但输出层权重梯度仍约 8.14。因此“根分支总梯度非零”不代表整个根网络仍能学习;末端有更新,前层已经很难恢复时序信息。 +-4. 在这种状态下,当前 MSE 的最佳行为是调整末端,使固定输出接近训练目标的平均值。实际抽样验证了这一点,见下表。这是“进入近似常数状态后为什么会维持这个平均答案”的解释;不代表 MSE 的全局最优只能是常数。 +- +-网络内部的这些测量使用 FP32、普通数学注意力实现,避免把加速后端的额外差异混入梯度结论。之前的 train/eval、FP32/BF16 对照也都确认 100K 输出近似常数,因此最终失效本身并非单一评估模式的现象。 +- +-对原正式训练采样器以 seed 42 抽取前 1,024 个窗口,覆盖 64 个 session;维持 batch8、16 批刷新一次的原约定: +- +-| 项目 | 结果 | +-|---|---:| +-| 根轨迹运动 RMS <1 cm 的窗口 | 49.3% | +-| 根轨迹运动 RMS ≥3 cm 的窗口 | 39.3% | +-| 尚无已释放文本的窗口 | 6.7% | +-| 仅输出样本目标均值的最优常数根 MSE | 0.22739 | +-| 100K 模型固定输出在同一批目标上的根 MSE | 0.22990 | +- +-100K 的固定答案只比“完全忽略输入、始终回答平均值”的最优常数解差约 1.1%。这与均值回归状态相符。 +-同时约 39% 窗口有明显位移,2K 网络也能利用输入得到更低误差,不能把问题简单归结为“数据全是静止”或“输入没有动作信息”。这只是 64 个 session 的抽样,不是完整数据集比例普查。 +- +-当前训练实现让这种状态难以被其他目标纠正: +- +-- 根阶段和身体阶段是两个 Transformer。原生 `TwostageDenoiser.forward` 在训练时以 `torch.no_grad()` 将根预测转成身体条件,并显式 `detach()`。所以身体阶段即使发现行走不对,其损失也不能传回根网络。 +-- 正式目标是 `normalized_x0`:根阶段五个特征的 MSE,加身体阶段特征的 MSE。YAML 中 `weights.root_position`、`joint_velocity` 等物理损失权重属于另一个 objective,本次没有启用。 +-- 身体输出中的 `velocities` 是独立辅助通道。导出人体运动时实际位置由 smooth-root 和髋部位置残差解码,未通过积分这些速度通道生成。辅助速度通道学对,不等于实际导出的根位移正确。 +-- 本次未配置 `recovery.temporal_weights`,因此没有针对实际解码轨迹速度、接缝速度或加速度的额外监督。 +-- 已知历史在输出端会被覆盖回观测值,且不计入未来损失。因而“历史输出保留准确”并不能证明 Transformer 内部仍在利用历史:根网络内部可以已经忘记历史,输出检查仍通过。 +- +-这些结构事实解释了为什么身体 loss 可以继续下降、整体梯度也非零,而根网络的坏状态得不到修正。 +-它们并不单独证明某一条代码必然使模型退化;健康 2K 权重已经说明同一结构可以暂时正常学习。 +- +-最值得怀疑的触发设置,是较强的根网络全参数更新与高噪声目标组合: +- +-- 根和身体共享 AdamW、3e-5 初始学习率和按 100K 步拉长的 cosine 调度,没有根阶段独立的更新预算。 +-- 2K 时学习率仍为 2.997e-5,6.5K 为 2.972e-5,8K 为 2.958e-5,几乎仍处于初始强度。 +-- 日志中 1–6K 的平均根 loss 约 0.042;6.5–7K 约 0.149,7–8K 约 0.220。身体 loss 同期仍缓慢下降。这是优化恶化的信号,不是仅在未见过的验证集上变差。 +-- `high_noise_fraction=0.5` 将约 50.05% 样本设为最大噪声 t=999;均匀采样时这一比例只有 0.1%。这大幅增加了无法依赖未来 noisy motion、必须依赖历史和文本的样本权重。 +-- 最大噪声时历史和文本仍然存在,因此高噪声本身不要求模型输出常数。但一旦条件表征开始丢失,这部分目标对模型来说就变成高不确定性 MSE 回归,平均答案成为容易维持的方向。 +- +-最符合现有证据的解释是:在这套长程微调设置下,根网络的时序/条件表示被后续更新破坏,深层特征和注意力进入近似均匀状态;前层梯度随之消失;MSE 将尚能更新的输出端推向平均答案,而隔离的身体损失无法救回根网络。 +-随后滚动窗口以历史开头为原点,重复生成同一个非零局部位置偏差,把根网络的均值偏差转成约两秒一次的世界坐标漂移。 +- +-严格的因果边界:目前没有保存 6.5–8K 附近的连续权重,也没有在相同全量数据顺序和随机种子下,分别改变根学习率、最大噪声比例与实际轨迹监督重训。因此不能宣称“唯一原因就是学习率”或“唯一原因就是高噪声”。已实测的是内部信息同质化、梯度消失、均值回归状态及其不能被身体梯度纠正的机制。 +- +-补充数值排查:本环境 PyTorch 2.7.1+cu128、RTX 5090。FP32 的 efficient attention 在特定高噪声输入上出现明显异常的大梯度;普通数学后端没有这一现象。 +-但正式训练使用 BF16,profile 确认使用 efficient attention。对六个既有训练 session 重建的 32 批、每批八个窗口,按原噪声/dropout 策略对照 BF16 efficient 与 math,没有复现那种量级的梯度爆炸:全参数梯度范数比最大约 1.061。 +-这不足以覆盖整次训练或排除稀有数值问题,但目前不能把 FP32 的异常直接当作本次 BF16 训练退化的已证实触发原因。 +- +-参数替换对照中,仅恢复输入投影、文本投影或输出层都不能恢复时序变化。 +-仅将注意力参数恢复为 2K,而保留其他 100K 参数时,输入梯度从 4.75e-12 恢复到约 0.00407,根轨迹运动 RMS 从约 6e-7 m 提高到 0.00626 m。这是一个直接的内部干预,支持注意力参数更新与信息/梯度消失存在联系;但该组合仍不能生成完整正确轨迹,不能据此宣称已经修复。 +-这支持失效发生在 Transformer 的内部条件处理,而不是某一个坐标输出 bias 写错。本轮所有替换仅用于只读诊断,没有生成新的训练权重。 +- +-[完整原因分析数据](../reports/kimodo_root_training_cause_20261023.json)。原始逐层、梯度、参数替换、数据采样及后端记录均在 `/root/autodl-tmp/cache/kimodo-formal-validation-20261023/root-probe/`。 +diff --git a/docs/KIMODO_TRAINING_REPAIR.md b/docs/KIMODO_TRAINING_REPAIR.md +deleted file mode 100644 +index 013ccff..0000000 +--- a/docs/KIMODO_TRAINING_REPAIR.md ++++ /dev/null +@@ -1,50 +0,0 @@ +-2026-10-23:修复 Kimodo 的训练方式,针对根 Transformer 在此前长程微调中失去时序差异的问题。输入仍为按时间释放的英文 GT 文本、最近 30 秒上下文和已执行人体历史;没有新增未来文本、阶段标签、GT 推理历史或轨迹修正。 +- +-这次修改训练本身,未加入退化检测、阈值触发、自动停止或自动替换权重。 +- +-- 根阶段保留官方预训练的注意力、FFN、归一化、文本和时间嵌入,只训练输入/输出投影,共 565,253 个参数;身体阶段训练 141,529,356 个参数。冻结的根骨干仍参与前向和对输入投影的反向传播。 +-- 根和身体分组优化、分别裁剪梯度。根学习率 2e-6→2e-7,在 20K 步衰减到底;身体 1e-5→1e-6,在 100K 步衰减;均有 500 步 warmup。末端根学习率在 20K 后保持下限。 +-- 最大噪声强制采样比例从 0.5 降至 0.1,其余 timestep 仍均匀采样。 +-- 保留有效未来帧的 normalized-x0 监督,前 12 帧执行段权重为 1,剩余规划段为 0.25。 +-- 加入原始根网络对已知历史的重建损失,读取硬覆盖历史之前的网络输出。目标只来自已知过去,不使用未来作为观测。 +-- 对实际解码的根速度、关节速度、旋转变化、首帧接缝速度和根加速度监督;原来的独立速度辅助通道不能替代这些目标。 +-- 根注意力使用数学后端,身体保留原加速实现。运行配置保存进 checkpoint,推理加载时自动恢复。 +-- 完整 checkpoint 保留模型、分组 AdamW、调度器、随机状态及 native session 槽位,继续支持 `--resume`。 +- +-正式配置仍为全量 native 英文人体配对数据、batch8、100K 步、GPU0。缓存与官方预训练资产复用;新训练不继承失败的 100K 微调权重。 +- +-```bash +-cd /root/autodl-tmp/hfd-baseline-portable +-/root/autodl-tmp/envs/kimodo/bin/python -u train_kimodo.py --config configs/kimodo.yaml +-``` +- +-中断后在相同命令末尾追加 `--resume`。进程可在 screen 中前台运行。 +- +-用户要求删除的旧目录已经完整移除,包括其中的 best、last、98K/100K 快照、日志和脚本副本: +- +-- `/root/autodl-tmp/runs/hfd-kimodo-english-train` +-- `/root/autodl-tmp/models/kimodo/hfd-english-train` +- +-这两个路径后来由修复后的正式训练重新创建,当前包含新一轮100K产物。删除旧失败训练时,官方模型、原始数据、共用文本缓存、GenMo活跃训练及既有诊断视频均保留。 +- +-验收报告:[kimodo_training_repair_20261023.json](../reports/kimodo_training_repair_20261023.json)。 +- +-- 30 项单元测试通过,覆盖真实两阶段小网络的冻结/反传、原始历史监督、分阶段优化与恢复、因果表示、解码运动损失和 session 槽位恢复。全尺寸官方网络离线检查通过,根/身体均有有效梯度、参数实际更新、文本条件有效。 +-- 全量入口实际读取 155,117 个 train、4,719 个 val session,完成2步保存,再恢复到4步;AdamW和调度器均到第4步,8个native session槽位随断点恢复。短测试尚未触发2K周期验证,`best_validation`在正式JSON报告中记为null,表示尚未选best。 +-- 六段 train session 小样本训练到8K,1K时执行过完整恢复;六段val不参与更新。固定GT历史验证loss由0.577降至1K的0.222,8K为0.152。所有8,000步记录的loss有限。验证loss不是闭环动作成功率。 +-- 8K根骨干的200个参数张量与官方初始化逐位一致;四个根投影参数及204个身体参数张量均已更新。同一后退窗口、t=0时,根末层时序RMS为0.0364,输入梯度范数为0.00944;旧失败100K分别约1.08e-7、4.76e-12。本次检查未重现旧恒定输出失效,不能外推保证100K质量。 +-- 六类val和两段train完成单帧GT初始化后的全程预测历史生成,整段预测和人体GT分别GMR到G1;导出8段有声视频。帧数、30Hz、音轨及完整音视频解码全部通过,也抽帧查看了预测/参考和文本显示。 +- +-同六段val、同seed/DDIM20/CFG2的等权平均如下;1K只做人体空间诊断,未单独retarget: +- +-| 模型 | 人体姿态 RMSE(rad) | 人体根位置 RMSE(m) | 人体根速度 RMSE(m/s) | G1关节 RMSE(rad) | G1根位置 RMSE(m) | +-|---|---:|---:|---:|---:|---:| +-| 官方初始化+当前适配 | 0.627 | 0.502 | 0.269 | 0.464 | 0.433 | +-| 新训练1K | 0.411 | 0.476 | 0.150 | — | — | +-| 新训练8K | 0.418 | 0.602 | 0.328 | 0.367 | 0.519 | +- +-动作质量仍未通过完整闭环验收:8K姿态误差优于官方初始化,但根位置/速度相较1K恶化,行走与停止样本仍有轨迹漂移;此前后退任务也不能宣称已解决。该实验验证训练链路及旧根骨干失效是否重现,不把增加小样本训练步数视为质量保证。全量训练命令可执行,最终效果仍需独立闭环评测;这不是完整HFD benchmark。 +- +-视频目录:`/root/autodl-tmp/runs/kimodo-training-repair-20261023/`,每例只有`retargeted_human_comparison.mp4`,左侧预测、右侧同流程retarget的human GT。工作数据、曲线和校验日志位于`/root/autodl-tmp/cache/kimodo-training-repair-20261023/`。本次临时checkpoint验收后全部删除,未启动新的100K正式训练。 +- +-2026-10-24后续:用户已启动并完成修复版100K训练,新正式权重保留。根骨干及其时序/梯度复查通过;六类闭环、整段retarget及18段有声视频见[新100K评测](HUMAN_FORMAL_VALIDATION_20261024.md)。前面的8K数值属于修复时的小样本验收。 +diff --git a/docs/LOCAL_INDEX_TRAINING.md b/docs/LOCAL_INDEX_TRAINING.md +deleted file mode 100644 +index ba1d40e..0000000 +--- a/docs/LOCAL_INDEX_TRAINING.md ++++ /dev/null +@@ -1,71 +0,0 @@ +-# 本地索引镜像与训练性能(2026-10-19) +- +-优化已完成,两个旧训练进程已按用户授权停止,原输出目录已删除;screen 保留。旧日志和配置保存在本报告证据目录,新正式训练尚未自动启动。当前正常训练命令见文末。 +- +-## 原因与改动 +- +-64 次实际合法窗口采样的 CPU 剖析耗时 50.64 秒,其中 tar 索引查询累计 33.52 秒,SQLite execute/fetch 占 33.80 秒,实际 tar 字节读取仅约 2.88 秒。主要瓶颈是对共享盘 `index.sqlite` 的大量小查询,而不是显存不足或大文件吞吐。 +- +-新增 `hfd_bridge/index_cache.py`:把约 **433.59 MiB** 的原始只读索引镜像到数据盘 `cache/hfd/index-mirrors/<身份>/index.sqlite`。按源路径、大小、修改时间和 SQLite 头部身份区分版本,使用进程文件锁和原子替换;复制时检查源未变化,并核对源复制流和本地文件的 SHA256。明确指定 `HFD_INDEX_PATH` 时仍尊重该覆盖路径。 +- +-共享 tar 读取器和第三方 session provider 都使用同一个索引入口。原始 tar 保留共享盘;这是元数据索引的本地副本,不是把全量 session 或原始数据复制到内存/本地盘。它可以供后续使用相同读取层的 baseline 复用。 +- +-正式 OMG/TextOp 及 reference YAML 开启 `runtime.environment.HFD_LOCAL_INDEX_CACHE: '1'`;设为 `'0'` 可做旧路径对照。正式 batch512、刷新频率、语言/文本协议、学习率和100K/300K步计划保持原配置。当前优化没有更改原生组装算法、训练循环或数据采样。 +- +-## 对照证据 +- +-CPU ABBA 对照(远端、本地、本地、远端)同64个样本依次为 **48.81、3.95、3.61、10.46 秒**。两次远端时长差异说明文件/页缓存有明显影响;即便采用已经变热的最后一轮远端,本地仍更快。四轮采样ID、动作、音频、控制标签及 reference segment 内容哈希一致。 +- +-随后使用两个真实原生训练循环,batch128,每种路径经历3次数据池准备。OMG48步,TextOp96步;同模型两种路径使用相同初始化、seed、更新预算、采样ID与窗口帧。以下为训练段耗时,包含组池、更新和有界验证,排除索引/模型初始化: +- +-| 模型 | 远端索引秒 | 本地索引秒 | 加速比 | +-| --- | ---: | ---: | ---: | +-| OMG | 251.6 | 33.9 | 7.42× | +-| TEXTOP | 219.9 | 35.6 | 6.18× | +- +-OMG两条loss曲线逐项相同,TextOp最大绝对差为2.98e-8。两种路径的采样ID与窗口帧完全一致。第二种路径复用了原始CPU索引,故不能把包含冷启动的总墙钟差异全算成优化收益。此轮与旧正式训练并行运行,存在共享资源干扰;这些加速比只描述该短测,不能直接当成100K/300K完整训练加速承诺。 +- +-## 正式 batch512 复核 +- +-确认上述加速和135项回归通过后,先停止并清理旧正式训练,再独立运行两张卡各512次更新、每模型两个完整数据池刷新周期。采用当前正式batch512与刷新频率,短学习率/阶段预算和有界验证用于测速,权重不作为质量结论。 +- +-| 模型 | 训练段秒 | 两次组池秒 | 池内秒/步中位数 | 训练段实例/秒 | +-| --- | ---: | --- | ---: | ---: | +-| OMG | 283.8 | 38.0, 53.5 | 0.271 | 924 | +-| TEXTOP | 183.9 | 41.4, 40.3 | 0.143 | 1426 | +- +-训练段包含组池与最终验证/周期保存,初始化墙钟另记在JSON中。TextOp的实例按每次更新的primitive样本计,OMG为动作窗;两模型绝对实例/秒不能横向解释为架构优劣。池内中位数也不包含完整的冷组池时间,长程耗时应使用整体训练段与后续正式验证成本估算。 +- +-两模型各记录512次更新、有限训练loss,保存的Adam状态计数均为512,且恰好准备2个训练池,无多余尾池。临时测速checkpoint已清理,仅保留配置、采样、指标和日志。性能变好不代表动作质量变好;英文前景oracle文本及其已有局限保持原协议,仍需正式验证曲线与连续回放判断质量。 +- +-## 停止、清理与重启 +- +-仅停止已核验命令行和工作目录的旧OMG PID322001、TextOp PID322362。OMG收到SIGTERM后未及时退出,再使用SIGKILL;TextOp由SIGTERM结束。以下两个目录已删除: +- +-- `/root/autodl-tmp/runs/hfd-omg-english-train` +-- `/root/autodl-tmp/runs/hfd-textop-english-train` +- +-小体积日志、配置、数据manifest和采样记录保存在 [old-training](/root/autodl-tmp/runs/archive/legacy-20261020/data-pipeline-optimization-20261019/old-training),checkpoint未保留;原始数据、官方权重、其他历史run和screen未清理。测试用索引副本已去重,保留一份当前有效的本地镜像。 +- +-在两个screen中分别从头启动,不加 `--resume`: +- +-```bash +-cd /root/autodl-tmp/hfd-baseline-portable +-/root/autodl-tmp/envs/omg/bin/python -u train.py --config configs/omg.yaml +-``` +- +-```bash +-cd /root/autodl-tmp/hfd-baseline-portable +-/root/autodl-tmp/envs/textop/bin/python -u train.py --config configs/textop.yaml +-``` +- +-## 验证文件 +- +-- [加速与输入/loss对照](/root/autodl-tmp/runs/archive/legacy-20261020/data-pipeline-optimization-20261019/speed-gate.json) +-- [CPU ABBA与输入哈希](/root/autodl-tmp/runs/archive/legacy-20261020/data-pipeline-optimization-20261019/index-abba.json) +-- [CPU剖析](/root/autodl-tmp/runs/archive/legacy-20261020/data-pipeline-optimization-20261019/provider-profile.txt) +-- [OMG batch512](/root/autodl-tmp/runs/archive/legacy-20261020/data-pipeline-optimization-20261019/omg-full512.json) +-- [TextOp batch512](/root/autodl-tmp/runs/archive/legacy-20261020/data-pipeline-optimization-20261019/textop-full512.json) +-- [旧训练停止与目录删除记录](/root/autodl-tmp/runs/archive/legacy-20261020/data-pipeline-optimization-20261019/old-training-cleanup.json) +- +-分环境135项回归测试与两套正式配置dry-run通过。性能探针保存在仓库外的本报告证据目录,没有重新堆入日常 `tools/`。 +diff --git a/docs/MIGRATION.md b/docs/MIGRATION.md +deleted file mode 100644 +index d2d72fe..0000000 +--- a/docs/MIGRATION.md ++++ /dev/null +@@ -1,5 +0,0 @@ +-# 迁移入口 +- +-当前完整操作见[训练复现指南](PORTABLE_TRAINING.md):四模型环境、外部资源、原始tar、英文词对齐、路径生成和正式训练命令。 +- +-本版本发布训练代码、配置、资源版本元数据和文档。本机测试/评测在被忽略的`test/`。历史ASR、双语和过拟合配置不作为当前默认。旧迁移说明的本地副本在`test/docs/`,未随Git发布。 +diff --git a/docs/OMG_ENGLISH_VALIDATION.md b/docs/OMG_ENGLISH_VALIDATION.md +deleted file mode 100644 +index 24fe908..0000000 +--- a/docs/OMG_ENGLISH_VALIDATION.md ++++ /dev/null +@@ -1,56 +0,0 @@ +-# OMG 英文对齐文本训练后的评测 +- +-`evaluate_omg.py` 从正式训练的 `resolved_config.yaml`、`data_manifest.json` 和 +-`val_native_manifest.json` 恢复验证协议,调用相同的原生及第三方数据 provider。 +-它只支持当前英文 `aligned_reference_context` / `microphone` OMG 配置。 +- +-默认选六类原生场景各一条6–30秒英文 session,另选三条第三方英文验证 session。 +-原生按索引取首个符合条件的样本,第三方在合格索引中等距取样;选择发生在生成之前。 +-这不是按类别比例采样的全量验证估计。`test` 是 `val` 的别名,验证数据用于过训练监测。 +- +-```bash +-cd /root/autodl-tmp/hfd-baseline-portable +-/root/autodl-tmp/envs/omg/bin/python -u evaluate_omg.py \ +- --trained-run /root/autodl-tmp/runs/hfd-omg-english-train \ +- --output /root/autodl-tmp/runs/omg-english-validation-20261020 \ +- --gpu 1 +-``` +- +-默认对比官方初始化与 `best/best.ckpt`。可用 `--checkpoint PATH` 显式选择新权重, +-`--per-mode N` / `--third-party N` 控制样本数,`--stage prepare|run|summarize` +-分阶段运行。相同参数重跑会复用完整结果;不完整推理目录须检查后另选输出目录或移走, +-不会静默覆盖。不会启动训练、识别、翻译、强制对齐或下载。 +- +-## 与训练一致的输入 +- +-- 相同英文过滤、配方、原生种子、验证有序索引及第三方 selection。 +-- 原始混合 PCM;过去2秒音频。词级离线对齐结果来自训练用缓存。 +-- 截至当前释放点的最近30秒 GT 前景文本,0.1秒 guard,0.6+0.4k 秒更新。 +-- 原生 T5-Base 保留最近50 token;实际 token ID、截断方向、PCM截止和文字释放时间记录在规划日志。 +-- 动作时间到麦克风时间偏移0.2秒,音视频对预测和GT共同前置0.2秒显示。 +- +-评测冻结时逐值校验原生导出和 compact provider 的音频、动作、标签及有效起点数, +-逐个50Hz状态帧比较导出 sidecar 和训练 provider 的文字条件。第三方直接导出训练 provider 数组。 +-标签、完整原文和未来动作不传入规划器;标签仅用于计分分层,事件元数据只用于视频标注。 +- +-## 连续推理与指标 +- +-两份权重用相同噪声seed12345、DDIM10、joint_one、剩余24帧触发的原生异步缓冲。 +-公平对照固定规划延迟为4个30Hz帧(约133ms),同时单独记录实际规划耗时。 +-这不是包含前端和部署调度的实测端到端延迟。仅开头13帧动作作GT初始化,之后反馈已执行预测, +-不按窗口重置GT。初始化及共同显示前缀不计分。 +- +-统计29关节RMSE、根xyz分量RMSE、根朝向误差、FK链接位置误差,以及listen/motion/idle分层误差。 +-标签索引先减去视频共同前缀,再转换到50Hz原始动作时钟,避免使用未来标签。 +-没有DTW或事后根位置对齐;这些误差衡量参考跟随,不等于语义成功率。 +- +-OMG直接输出G1的29关节qpos,不需要人体retarget。渲染为运动学预测/GT对照,带原始混合音轨, +-不是物理tracker。所有视频完整解码,音轨核验;两份权重的GT和音频必须相同。 +- +-输出包含 `README.md` 视频索引、`metrics.csv`、`summary.json`、`verification.json`、 +-权重SHA256、冻结样本与对齐sidecar、逐次推理/渲染日志。 +-官方预训练重合、第三方与原生动作源重合未审计;小样本单seed不能替代完整评测。 +- +-视频封装保留完整H.264视频流,不再用 `-shortest`:实测相同长度PCM与含B帧的视频流封装时, +-该参数可能裁掉末尾视频包(508帧样本曾变为505帧)。新增实际编码/全帧解码回归测试; +-渲染中间视频只在最终MP4帧数、帧率和音轨通过检查后删除。 +diff --git a/docs/OMG_TRAINING_REVIEW.md b/docs/OMG_TRAINING_REVIEW.md +deleted file mode 100644 +index 9002c7c..0000000 +--- a/docs/OMG_TRAINING_REVIEW.md ++++ /dev/null +@@ -1,105 +0,0 @@ +-# OMG 正式训练复查与吞吐优化(2026-10-17) +- +-清理说明:本文记录的一次性实验脚本已归档,复现前请按[脚本恢复说明](WORKSPACE_CLEANUP.md)取回;历史指标与日志保留。 +- +-本轮输出:`/root/autodl-tmp/runs/archive/legacy-20261020/hfd-omg-train-20261010-v2`。已正常完成 **100,000 次优化器更新**,不是被 OOM 或异常中断。screen 的 `Trainer.fit stopped: max_steps=100000 reached` 和 checkpoint 的 `global_step=100000` 一致。训练循环耗时 108 小时 33 分 7 秒,整个命令约 108 小时 45 分,约 4.53 天。完成时间为 2026-10-15 01:30(Asia/Taipei)。 +- +-## 本轮结果 +- +-| 指标 | 结果 | +-| --- | --- | +-| 模型 | OMG 原生 50m 配方;denoiser 参数约 54.14M,另有冻结的 T5 等模块 | +-| 训练 | RTX 5090 × 2;每卡 batch 512,总 batch 1024;bf16 mixed | +-| 训练 loss,前 1,000 步平均 | 0.04943 | +-| 训练 loss,最后 10,000 步平均 | 0.02315 | +-| 首次验证,2,000 步 | 0.0365847 | +-| 最低日志验证 loss,90,000 步 | 0.0220401 | +-| 最后验证,100,000 步 | 0.0234458 | +-| 最终权重检查 | 可加载,394 个 state tensors 均无 NaN/Inf,含 optimizer/scheduler 状态 | +- +-![训练与验证曲线](../reports/omg_training_20261017.png) +- +-训练 loss 是 rank 0 的原生扩散损失;验证 loss 由两卡同步归约。它们不是厘米级动作误差,也不是任务成功率。`val/motion_loss=0` 来自原生 `diffusion_only` 配方,不代表动作没有监督;动作特征是扩散损失的预测目标。这里只确认训练收敛趋势和权重健康,尚未完成论文独立测试集或连续 rollout 评测。 +- +-最终文件是 `last.ckpt`。保留的周期权重是 `checkpoints/step-00096000.ckpt`、`step-00098000.ckpt` 和 `step-00100000.ckpt`;`checkpoints/last.ckpt` 是周期保存的链接。原策略按步数保留最近三份,**90,000 步的最低 loss 权重已被删除,不能恢复为 best**。现存三份记录的验证 loss 分别是 0.0222620、0.0223724、0.0234458;要据此选模型,应在同一固定验证样本上重新比较。 +- +-旧验证并不固定:每次验证每卡 16 个 batch,池每累计 256 个验证 batch 换一次。因此 2k、34k、66k、98k 步分别开始使用不同的验证 session 池;池内窗口和扩散噪声也不断变化。最后两个 loss 不能直接与 96k 比较后认定过拟合。 +- +-本轮 manifest 记录 train=322,243、val=9,272 个原生候选 session,`command_interruption_session_variants=2`。两卡记录合计访问 207,173 个不同训练 session,约占候选的 64.3%;有放回随机窗口训练不等于完整遍历一个 epoch。**本轮不包含后来加入的 third-party 包**:当前 `configs/sessions.yaml` 已有 `include_third_party: true`,不能用今天的配置文件反向描述旧实验。复查性能时使用关闭 third-party 的临时配方,重放旧 run 的 session 索引;当前正式数据配置未被改回旧值。 +- +-## 是否使用 OMG 原生脚本 +- +-入口是 portable 的 `train.py::omg`,没有直接调用上游 `omg.cli.generation.train.main`。入口组合上游 `exp=50m`,加载官方模型定义及初始化权重,再构造 `pytorch_lightning.Trainer`,训练、验证、优化器和 scheduler 仍调用上游 `MotionGenerator` 的实现。替换的是 HFD 数据适配器和训练管理 callback。准确表述是“使用 OMG 原生模型及训练步骤,通过 HFD 适配入口训练”。 +- +-本轮保持原生 100k 更新、2k warmup、6e-5 学习率及 cosine schedule。数据输入是过去两秒已释放音频 + 因果 ASR 当前指令,完整 session 动作监督;未改变动作映射、音频截止时间或模型目标。 +- +-## 慢在哪里 +- +-1. 每卡 batch=512 的 session 池每 256 个 batch 刷新。本轮 391 次刷新,取每次两个 rank 中较慢的准备时间累加约 **59.22 小时**,约占训练循环 54.6%。这包含共享 tar 读取、原生 session 组装、ASR、运动预处理和首个 batch,不能全算成磁盘 I/O。rank 0 的刷新中位数 493.6 秒。 +-2. 四次验证池刷新另需约 1.10 小时。完整训练平均每步 3.91 秒,旧日志没有每步数据/计算细分计时。 +-3. ASR SQLite 有 209,704 条记录、约 1.18 GiB 压缩有效内容。旧实现每插入一条都 `SUM(length(data))` 扫全表;复查时同一只读扫描缓存热后仍约 0.20–0.29 秒。这是可去除的元数据开销,不是模型计算。 +-4. 旧普通 batch 逐个调用 512 次 motion codec,许多小 CPU tensor 运算串行执行。音频特征此前已做批处理,动作特征仍有批处理空间。 +-5. tar `Member.readinto` 对小读请求仍取一个 256 KiB 块,且本地 tar 路径没有块复用。`BufferedReader`/NPZ 的小读和回读可反复读取同一块。 +- +-`num_workers=63/127` 是 Lightning 通用提示。当前数据对象持有 SQLite、状态采样器以及同卡 Whisper,且明确禁止 DataLoader worker;仅修改 `num_workers` 会被配置校验拒绝。若做并行预取,需要把 CPU 读盘/组装与 GPU ASR 分开,并管理 rank/worker 的采样状态,不能直接照提示填值。 +- +-## 已落地的改动 +- +-- 动作预处理:用同一个 OMG 原生 codec 一次计算整批 10 帧历史与 60 帧目标。`data.batched_motion=true`,可设为 false 回退逐样本实现。测试比较全部 batch 字段,容差 `rtol=atol=1e-6`,保持动作定义与时间对齐。 +-- ASR SQLite:增加由数据库 trigger 原子维护的容量计数,首次迁移扫描一次,此后不再每次插入扫全表。兼容已有 ASR 内容与 key,两卡连接共同维护;访问时间最多每分钟写一次,减少热缓存写锁。原有 2 GiB 有效内容上限仍生效;该上限不等于 SQLite 文件/WAL 的物理上限。 +-- tar I/O:每个打开的成员保留最多四个 256 KiB 块,减少小读放大;关闭时释放。原始 tar 留在共享盘,不解压,不复制完整数据集。 +-- 训练管理:固定验证窗口与扩散随机数,验证结束恢复训练 RNG;单独保存 `best/best.ckpt`,不受最近三份周期权重清理影响;记录每步等待/计算时间及最终训练摘要。时间字段是主机墙钟,`data_wait_seconds` 可能包含周期 checkpoint 等 callback 开销,不等于纯磁盘 I/O。 +-- 可选早停:原生 Lightning `EarlyStopping`,监测 `val/loss`,从 30k 步开始计数,`patience=10`,`min_delta=0.0002`。每 2k 步验证,因此对应连续 20k 步没有足够改进。patience 计算的是验证次数,见 [Lightning 官方说明](https://lightning.ai/docs/pytorch/latest/optional-api/api/lightning.pytorch.callbacks.EarlyStopping)。 +- +-默认 `training.fixed_validation=true`、`training.save_best=true`,但 `training.early_stopping.enabled=false`,保留原生固定 100k 预算。要选择早停训练,在原训练命令上加 `--set training.early_stopping.enabled=true paths.output_dir=/root/autodl-tmp/runs/新的目录`。这会成为不同的训练预算,论文中须记录。 +- +-固定验证复用同一组每卡 16 个 batch(双卡 16,384 次窗口抽样,窗口允许重复),仍只是监测子集。不能按旧随机验证日志选择最有利的早停阈值:事后机械套 patience=10 会在 88k 停,随后 90k 还有改善,说明不能保证早停明显省时且完全不损失质量。优化每步吞吐比激进缩短训练更有依据。 +- +-## 验证和性能测量 +- +-单元测试覆盖原生动作 codec 对照、固定验证样本、验证 RNG 恢复、真实 Lightning 早停/最佳权重保存、旧 SQLite 迁移与双连接容量计数、tar 小读/seek/EOF 等。性能测量仅重放旧训练池、运行限定步数,不保存更新后的训练权重,也不改写已完成 run。 +- +-本机测量记录集中在 `/root/autodl-tmp/cache/hfd/review-20261017/`,不提交原始数据或大文件。47 项相关测试与 12 个 subtests 通过,配置 dry-run 和 `git diff --check` 通过。 +- +-同一旧 run 的 512-session 池,4 CPU 线程,比较结果: +- +-| 测量口径 | 优化前 | 优化后 | +-| --- | ---: | ---: | +-| session 已缓存、仍有新音频窗口的 batch 数据准备 | 1.68 秒 | 0.79–0.80 秒 | +-| 同一进程、同一批窗口且音频已缓存,逐样本/批量动作 codec 对照 | 0.91–0.92 秒 | 0.048–0.078 秒 | +-| 209,704 条 ASR 记录的容量查询 | 约 0.16 秒/次全表扫描 | 约 2–44 微秒/次读计数 | +- +-真实 batch 对照比较了所有 tensor、caption 和索引,`rtol=atol=1e-6` 通过。热 session、新音频窗口的数据准备约快 **2.1 倍**;音频全命中的对照展示了消除逐样本动作计算后的收益,不能把该倍数当成整轮训练加速。 +- +-两次首次装入同一池分别耗时约 417 和 209 秒,数组读取/原生组装分别占约 407 和 200 秒。它们共享文件系统缓存且执行有重叠,不是受控冷盘对照,不能归因于代码修改。新 tar 块缓存有读取一致性与读放大量测试,但尚无隔离的整轮冷读测速。 +- +-双卡端到端短测已完成:加载本轮最终 checkpoint,在真实 HFD 数据上每卡 batch=512,各完成 6 次前向、反向与优化器更新;所有记录的 loss 有限,denoiser 确认为训练模式。第 2–6 步完整迭代为 **1.25–1.32 秒/步,均值约 1.28 秒**,模型计算约 0.42–0.50 秒/步。每卡 PyTorch 峰值 allocated 显存约 16.42 GiB;本次 ASR 已命中缓存,没有同时运行缺失文本的 Whisper 识别,不能据此提高正式 batch。 +- +-首次准备在 rank 0/1 分别是 68.9/371.2 秒;rank 0 第一轮又等待约 304 秒,而后两卡进入上述稳定吞吐。这个对照直接展示了共享盘读取/原生组装的不均衡及 DDP 等待。短测没有跨过 256-step 池刷新边界,不能用 1.28 秒直接乘 100k 预测完整训练时间。检查已结束,两张 GPU 已释放;没有启动新的正式训练,也没有覆盖本轮 checkpoint。 +- +-下一阶段若刷新仍占主导,优先将因果 ASR 从主训练循环拆成可复用的准备阶段,再考虑有界异步预取。单纯把 `refresh_batches` 从 256 加到 1024 虽然摊薄刷新,却减少相同步数下接触的新 session、增加相邻 batch 相关性,本次没有采用。当前数据盘不足以额外复制约 49.8 GiB 的全部原始 tar,因此也没有把“全部搬到本地”当作立即可行的方案。 +- +-## 异步预取与并行 I/O(2026-10-17 后续) +- +-增加 `data.prefetch_batches`(0 或 1)。开启后,OMG 在 GPU 训练当前 batch 时,用一个后台线程准备下一个已缓存的 CPU batch,包括动作编码和音频特征。最多额外保留一个 CPU batch;validation 仍在主线程执行。 +- +-后台线程不构造新 session、不访问 SQLite、不运行 Whisper,也不执行 GPU 工作。`WindowPool` 保留当前槽位的合法时刻范围,避免热 batch 为查询长度重新组装数组;缓存未就绪或到达刷新边界时,自动回到主线程。固定 seed 下,测试跨多个刷新边界逐一比较动作、音频、caption、采样时刻,并验证训练 RNG 不变。后台异常传回主线程,关闭 iterator 会结束任务。实现使用标准库 [ThreadPoolExecutor](https://docs.python.org/3.10/library/concurrent.futures.html)。 +- +-另外增加本地 tar 成员的有界并行预读:`runtime.environment.HFD_IO_READAHEAD_BLOCKS=4`,`runtime.environment.HFD_IO_WORKERS=4`。线程仅执行本地文件读取,不共享 SQLite 连接,不进行网络下载。大型/连续读取才触发预读,seek 会取消旧预读,每个成员的未来块数有界;原有四块缓存仍保留。它作用于原生 `ArchivePath/Member` 读取路径,third-party 已组装 WAV/NPZ 的直接大块读取路径没有替换。 +- +-8 个实际音频成员的同步→异步→异步→同步对照,全部 SHA256 相同,时间为 0.704、0.094、0.071、0.052 秒。第一次同步读取受冷缓存影响;最后一次同步比异步更快,因此不能把第一次与第二次相除宣称并行 I/O 提升。目前并行 tar 预读**默认关闭**,保留为不同文件系统上的可测选项。 +- +-CPU batch 异步预取则使用双卡 64 步的同步→异步→异步→同步分段对照;每段排除开头两步,以减少流水线切换影响。该测试仍重放旧训练池,不包含 third-party,也不执行新的正式训练。它测的是池内准备与 GPU 计算的重叠,不包含下一池的后台组装/ASR,不能视为解决全部刷新等待。 +- +-最终采用 **OMG 一个热 CPU batch 预取,tar 并行预读默认关闭**。两卡各 64 步均完成,所有 loss 有限,denoiser 确认为 train mode。每段排除两步后,按每一步较慢的 rank 汇总: +- +-| 模式 | 平均每步 | 分段结果(rank 0) | +-| --- | ---: | --- | +-| 同步 | 1.072 秒 | 第一段 1.244 秒;末段 0.897 秒 | +-| CPU batch 异步预取 | 0.617 秒 | 中间两段 0.656 / 0.572 秒 | +- +-该对照约提高池内吞吐 1.74 倍,单步时间减少约 42%。音频缓存会逐步变热,ABBA 仅减轻该影响,不能彻底消除;异步中段也比更晚、更热的同步末段快。CPU/GPU 并发时模型阶段本身略慢,但整体迭代更快。每卡 PyTorch allocated 峰值约 16.42 GiB。首次装载两卡分别约 227/223 秒,此部分未被热 batch 预取隐藏。 +- +-当前默认 `configs/omg.yaml` 中 `data.prefetch_batches: 1`,同步回退为 0;并行 I/O 的 `HFD_IO_READAHEAD_BLOCKS` 默认仍为 0。52 项相关测试及 12 个 subtests 通过,另有真实双卡 64 步检查。这里只运行了有限步数的性能探针,不保存更新权重;GPU 已释放,旧正式训练权重未改写。原始测量在 `ddp-async-abba/{rank0,rank1,summary}.json` 与 `tar-async-abba.json`(位于本机 review 缓存目录)。 +- +-复查命令可使用 `tools/benchmark_omg_training.py --steps 64 --compare-prefetch`(双卡需 torchrun,配合实际配置、`--replay-dir` 和 `--report-dir`);单独 tar 对照使用 `tools/benchmark_omg_io.py --members 8 --report /数据盘/新的测量.json`。这些命令是测试工具,不是正式训练或完整训练时长评估。 +diff --git a/docs/OMG_VALIDATION_DIAGNOSIS.md b/docs/OMG_VALIDATION_DIAGNOSIS.md +deleted file mode 100644 +index 9898751..0000000 +--- a/docs/OMG_VALIDATION_DIAGNOSIS.md ++++ /dev/null +@@ -1,77 +0,0 @@ +-# OMG 验证视频:提前动作与听完不执行的诊断 +- +-清理说明:本文记录的一次性实验脚本已归档,复现前请按[脚本恢复说明](WORKSPACE_CLEANUP.md)取回;历史指标与日志保留。 +- +-本次仅检查冻结的 100k checkpoint,没有训练、改权重、替换模型输入或覆盖原视频。诊断产物位于 `/root/autodl-tmp/runs/archive/legacy-20261020/omg-100k-validation/diagnostics/`。 +- +-## 1. 已确认的中文文本语义丢失 +- +-在实际使用的 `/root/autodl-tmp/models/OMG/t5-base-local` tokenizer 上,以下实际规划文本的 token IDs 完全相同: +- +-| 实际文本 | token IDs | token | +-| --- | --- | --- | +-| 慢跑向前 | `[3, 2, 1]` | `▁, , ` | +-| 小禮母 | `[3, 2, 1]` | 同上 | +-| 慢跑向前跳越中身手架 | `[3, 2, 1]` | 同上 | +-| 慢跑向前跳越中身手构想后跳 | `[3, 2, 1]` | 同上 | +- +-原生 `FrozenT5TextEncoder` 只根据这些 token IDs 和 mask 计算冻结 T5 特征,因此这些输入的文本条件相同,无法区分任务。打断样本第二、第三条中文指令到来后,ASR 字符串虽然变化,模型文本分支却收不到对应的语义变化。35D 声学分支包含频带能量等声学特征,不能假定它会自动补回丢失的指令语义。 +- +-这不只是 Whisper 的中文识别不准;即使正确识别“慢跑向前”,当前文本入口仍会丢失含义。之前只验证长度和截断,没有验证未知词覆盖,是适配验收的遗漏。保留完整历史 ASR 也不能修复该 tokenizer 问题;继续增加同配方训练步数不会恢复已经被映射成同一个 token 的指令信息。 +- +-`clock-and-token-audit.json` 保存六个视频每次规划实际文本及 token。用户指出的 `interruption-stable_latest_sentence.mp4` 对应 `val-000064-seed42`: +- +-- 视频任务语音起点为 1.6、9.2、14.8 秒。 +-- 进入模型的后续文本含“慢跑向前跳越中身手架”等 ASR 错误,且中文段均发生上述 token 合并。 +-- 预测根位置相对开场的最大水平位移仅 0.0242 m,GT 为 4.0543 m;模型没有跟随参考里的明显移动。 +- +-这些证据确认输入适配有问题,不能直接把该样本的失败归因为训练步数不足。 +- +-## 2. 提前动作包含生成模型自身的提前起步 +- +-六条视频的首次明显姿态变化几乎都在 1.6–1.7 秒,尽管语音起点不同。这里的“明显变化”定义为相对开场姿态、29 关节 RMS 偏移首次超过 0.05 rad,不是一个训练出的倾听分类器。 +- +-尤其是英文紧急停止样本 `val-000111-seed42`: +- +-- 视频指令语音到 2.4 秒才开始;GT 姿态超过阈值在约 3.47 秒。 +-- 预测在约 1.63 秒已经超过阈值,呈视频中的倾听/准备样式。 +-- 产生这些姿态的首轮规划只读到 PCM 1.0 秒,实际任务语音到 PCM 2.6 秒才开始;首轮 ASR 为背景上的幻觉 `you`。 +- +-对同一 checkpoint、同一开场历史、同一 seed,重新生成首个 60 帧计划,实际条件输出与保存视频的已执行部分匹配。分别去文本、去音频、两者都去掉,仍会产生相近幅度的动作: +- +-| 首轮条件 | 最大关节姿态偏移 RMS | +-| --- | ---: | +-| 实际音频+文本 | 0.5723 rad | +-| 去文本 | 0.5752 rad | +-| 去音频 | 0.5722 rad | +-| 两者都去掉 | 0.5756 rad | +- +-这是首轮计划诊断,不是完整 session 消融,也不表示屏蔽条件时原生扩散模型本应静止。它表明这些动作无需当前语音指令触发即可生成。当前模型只预测姿态,没有显式“检测到指令后才能进入倾听/执行”的门控;HFD 所需的等待和触发行为尚未被可靠学到。不能将其解释为模型提前听懂了后面的指令。 +- +-数值见 `silent-first-plan-audit.json`,轨迹见 `silent-first-plans.npz` 和 `early-motion.png`。 +- +-## 3. 现有动作轴与音频释放轴还会影响视频观感 +- +-现有评测沿用训练的 HFD 时钟,动作组 `k` 使用截止 `0.6+0.4*k` 秒的 PCM;视频按动作时间显示,播放的音轨却取原 PCM 的 `t+0.2` 秒位置。这不是统一麦克风墙钟的真实在线系统。 +- +-按实际计划交接记录换算后,打断视频有新计划开始执行时、输入音频内容比视频当时播放位置领先约 0.167 秒的区间;停止视频最大约 0.233 秒。它会影响“提前/延迟”的直观判断,当前视频不应直接用于真实响应延迟评估。 +- +-这个小幅度时钟领先不能解释停止样本中约 0.77 秒的提前起步:首轮规划根本未包含任务语音,且无条件首轮仍产生动作。两件事应分开处理。修正评测墙钟后需要重新核对动作标签和训练时的释放映射,不能只把音轨随意平移来美化视频。 +- +-## 4. 仍有生成/闭环能力问题,但尚未完全归因 +- +-英文跳跃样本文本能正常编码;预测根高度范围仅约 0.0633 m,GT 为约 0.3907 m。因此中文入口并非全部问题。现有模型、训练窗口和闭环规划组合仍有动作幅度与阶段推进不足。 +- +-异步规划执行旧缓冲、替换未执行的计划后半段,模型是否反复把动作推迟到未来,以及真实历史与预测历史的差异,尚需固定条件、固定交接时刻和历史模式对照才能分解。不能仅凭当前视频断言优化器错误、模型架构无能力或某个单一训练参数是根因。 +- +-后续优先级:先修复中文语义入口并审计五个模型的实际 tokenizer 覆盖;再统一评测墙钟;最后区分模型等待/触发学习和闭环规划问题,再决定重训配方。完整历史 ASR 仍是用户确认的共同策略,但不能把“完整文本存在”当作“模型获得了文本语义”。 +- +-复现主诊断: +- +-```bash +-# diagnostics 目录必须尚不存在;保留原诊断时改用新的评测根目录副本。 +-/root/autodl-tmp/envs/omg/bin/python tools/diagnose_omg_validation.py --root /root/autodl-tmp/runs/archive/legacy-20261020/omg-100k-validation +-``` +- +- +-后续输入与时钟修复已实现,旧诊断结果保持不变,见 [修复记录](SEMANTICS_CLOCK_FIX.md)。 +diff --git a/docs/OMG_VALIDATION_ROLLOUT.md b/docs/OMG_VALIDATION_ROLLOUT.md +deleted file mode 100644 +index a7cff48..0000000 +--- a/docs/OMG_VALIDATION_ROLLOUT.md ++++ /dev/null +@@ -1,67 +0,0 @@ +-# 已完成训练的 OMG:六类验证 session 连续推理 +- +-清理说明:本文记录的一次性实验脚本已归档,复现前请按[脚本恢复说明](WORKSPACE_CLEANUP.md)取回;历史指标与日志保留。 +- +-结果目录:`/root/autodl-tmp/runs/archive/legacy-20261020/omg-100k-validation/`。 +- +-后续诊断确认了 **T5 中文输入映射为相同 `` token、无任务语音时提前起步,以及视频/释放时钟差异**。详见 [提前动作与不执行诊断](OMG_VALIDATION_DIAGNOSIS.md)。本轮指标包含这些输入适配问题,不能单独用于判断动作模型能力。 +- +-## 数据和权重 +- +-使用 `/root/autodl-tmp/runs/archive/legacy-20261020/hfd-omg-train-20261010-v2/last.ckpt`,确认 `global_step=100000`。没有重新训练或改写权重,评测目录仅用软链接引用原 checkpoint。`evaluation.json` 保存 SHA256;每次推理使用原生模型的严格权重加载。 +- +-仓库有 HFD 原生 train/val 构造入口,数据来自本机 `/root/autodl-fs/pangxinlei/hfd_data` 的 tar 分片。旧训练对应 322243 条 train、9272 条 val session。任务动作源按哈希隔离;验证集曾用于训练监测。**源码中的 `test` 是 val 的别名,没有第三份独立、未使用过的 test 划分。** 公共 idle/listen 素材可能共用;官方 OMG 预训练数据与这些动作的重合没有审计,不能声称模型整个训练过程从未见过它们。 +- +-本次重建的 val 索引指纹 `dc6cb1533a00d2fd983d220628a999cad60023339fbf5abba106cd704d2421dd` 与训练保存的 manifest 一致,任务源哈希划分验证通过。每类选择第一条估计时长 6–30 秒的 session,按类别交替优先选择英文和中文;选择完成后才运行模型,没有根据输出挑样本。具体索引和动作身份见 `selection.json`。 +- +-| 场景 | session index | 语言 | 长度 | +-| --- | ---: | --- | ---: | +-| 轮流指令 | 3 | 英文 | 10.8 s | +-| 任务打断 | 64 | 中文 | 21.6 s | +-| 紧急停止 | 111 | 英文 | 16.8 s | +-| 附和 | 60 | 中文 | 15.6 s | +-| 犹豫 | 28 | 英文 | 16.8 s | +-| 停顿 | 41 | 中文 | 27.2 s | +- +-## 推理设置 +- +-旧权重采用 `stable_latest_sentence` 训练,因此主评测保持原策略;另以同一 checkpoint 测试当前默认 `causal_asr_context`。后者属于输入分布变化对照,不代表已用完整历史重训。 +- +-两组共用相同的因果 ASR sidecar、音频、GT 和 seed=12345。ASR 使用本地 Whisper-tiny、auto 语言、最多 30 秒历史、batch 32;每条输入切片的哈希与释放边界均核验。OMG 输入过去两秒已释放音频及文本,保持原生 10 帧动作历史、60 帧未来、30Hz、DDIM 10 步、joint_one CFG、剩余 24 帧触发的 async 缓冲。异步规划耗时实测,旧新策略的耗时和交接时刻可能不同;ASR 识别计算耗时不计入。 +- +-开场 13 帧 GT bootstrap 不计分,之后始终反馈已执行预测,没有逐窗口 GT 重置。listen/motion/idle 标签仅用于评测分层,不进入推理。输出直接是 G1 qpos,无需人体 retarget;视频仅为运动学回放,没有物理 tracker。 +- +-## 结果 +- +-关节 RMSE 为原生 29 关节,根坐标 RMSE 为 xyz 分量 RMSE;它不是逐帧根位置欧氏距离的均值。所有误差按共同时间轴直接计算,没有 DTW 或事后根位置对齐。 +- +-| 场景 | 原稳定句关节 RMSE | 完整历史关节 RMSE | 原稳定句根坐标 RMSE | +-| --- | ---: | ---: | ---: | +-| 轮流指令 | 0.2501 rad | 0.2264 rad | 0.0915 m | +-| 任务打断 | 0.3253 rad | 0.3244 rad | 1.6924 m | +-| 紧急停止 | 0.4356 rad | 0.4614 rad | 0.6104 m | +-| 附和 | 0.2860 rad | 0.2625 rad | 0.0133 m | +-| 犹豫 | 0.4297 rad | 0.4278 rad | 0.0437 m | +-| 停顿 | 0.2919 rad | 0.2695 rad | 0.5386 m | +-| 六条等权均值 | **0.3364 rad** | **0.3287 rad** | **0.4983 m** | +- +-按计分帧合并的关节 RMSE 为 0.3436 / 0.3393 rad;两种汇总的含义不同,完整记录在 `summary.json`。另保存根朝向误差、FK 链接位置误差、listen/motion/idle 分层误差及实际规划文本非空比例。 +- +-这些验证动作的参考跟随偏差明显高于之前单条 session 的过拟合结果。完整历史 ASR 的平均改善很小,紧急停止样本的关节误差反而上升。部分中文转写有明显识别偏差,例如打断样本的末次文本为“曼跑向前跳躍中身手架我想后跳”;保留实际转写,不用 GT 修正。这是 ASR+OMG 链路的结果,不能把所有偏差归因于动作模型。 +- +-这是六条验证样本、一个采样 seed 的初步检查,不是全量测试或任务成功率评测。不同动作也可能满足同一句话,因此 GT RMSE 不能单独判断语义是否正确。未提供停止响应延迟、任务成功率或物理稳定性结论。 +- +-## 查看和核验 +- +-打开结果目录的 `README.md` 可逐类访问两种策略的 12 条视频,均为左侧预测、右侧 GT,带原始混合音轨;`comparison.png`、`metrics.csv` 和 `summary.json` 分别为图表、表格和完整数值。`video_contact_sheet.jpg` 是中段视频帧概览。部分中文视频字幕受现有字体支持限制显示为问号,实际 Unicode 输入保存在各条 `report.json` 和 ASR sidecar 中。 +- +-12 条视频共 6528 帧完整解码,音轨检查通过;两种策略的 GT 数组逐值一致,ASR PCM 切片哈希通过。13 项相关缓冲、关节映射、文本因果性测试通过,记录在 `verification.json`。此次只读取本地 tar,没有下载数据。 +- +-## 复现 +- +-```bash +-# output 必须为新目录;从原训练记录恢复验证配方 +-/root/autodl-tmp/envs/omg/bin/python tools/prepare_omg_validation.py --trained-run /root/autodl-tmp/runs/archive/legacy-20261020/hfd-omg-train-20261010-v2 --output /root/autodl-tmp/runs/archive/legacy-20261020/omg-100k-validation +-/root/autodl-tmp/envs/omg/bin/python tools/run_omg_validation.py --root /root/autodl-tmp/runs/archive/legacy-20261020/omg-100k-validation +-/root/autodl-tmp/envs/omg/bin/python tools/summarize_omg_validation.py --root /root/autodl-tmp/runs/archive/legacy-20261020/omg-100k-validation +-``` +diff --git a/docs/PORTABLE_TRAINING.md b/docs/PORTABLE_TRAINING.md +deleted file mode 100644 +index c73bd69..0000000 +--- a/docs/PORTABLE_TRAINING.md ++++ /dev/null +@@ -1,167 +0,0 @@ +-# 在另一台机器复现训练 +- +-当前四个正式入口及配方见README。HFD/OMG/TextOp原生源码在`upstream/`;GenMo和Kimodo使用`UPSTREAM_LOCK.json`指定的外部源码。模型、原始数据和离线词时间缓存单独准备。本版本不发布测试/评测脚本、正式权重或日志。 +- +-以下以`/data/hfd_work`为工作盘、`/data/hfd_assets`为外部资产盘。已验证环境是Linux、Python3.10、PyTorch2.7.1+cu128、RTX5090 32GB。GenMo训练峰值约25GB;Kimodo还需要同卡常驻的8B冻结文本worker。已有资源可由路径参数复用,不必搬迁。 +- +-## 代码及环境 +- +-AutoDL下载/安装命令先加`/root/.local/bin/turbo`,不要全局source代理。下面下载命令展示该前缀;其他机器使用自身已授权的网络方式。HF失败可使用已配置的`hfmirror`,不能自动转用个人代理。 +- +-```bash +-/root/.local/bin/turbo git clone https://github.com/starstone3/Dual-hfd-portable.git /workspace/hfd-baseline-portable +-cd /workspace/hfd-baseline-portable +-mkdir -p /data/hfd_work/envs /data/hfd_work/hfd-human-models +-``` +- +-创建独立Python3.10环境。完整冻结依赖在`requirements/`;GenMo、Kimodo和forced-align的Torch版本另列于`runtime-versions.json`: +- +-```bash +-python3.10 -m venv /data/hfd_work/envs/genmo +-/root/.local/bin/turbo /data/hfd_work/envs/genmo/bin/python -m pip install \ +- --index-url https://download.pytorch.org/whl/cu128 torch==2.7.1 torchvision==0.22.1 +-/root/.local/bin/turbo /data/hfd_work/envs/genmo/bin/python -m pip install -r requirements/genmo.txt +- +-python3.10 -m venv /data/hfd_work/envs/kimodo +-/root/.local/bin/turbo /data/hfd_work/envs/kimodo/bin/python -m pip install \ +- --index-url https://download.pytorch.org/whl/cu128 torch==2.7.1 torchvision==0.22.1 +-/root/.local/bin/turbo /data/hfd_work/envs/kimodo/bin/python -m pip install -r requirements/kimodo.txt +-/root/.local/bin/turbo /data/hfd_work/envs/kimodo/bin/python -m pip install --no-deps \ +- --target /data/hfd_work/envs/kimodo-text-compat -r requirements/kimodo-text.txt +- +-python3.10 -m venv /data/hfd_work/envs/omg +-/root/.local/bin/turbo /data/hfd_work/envs/omg/bin/python -m pip install \ +- --extra-index-url https://download.pytorch.org/whl/cu128 -r requirements/omg.txt +-/root/.local/bin/turbo /data/hfd_work/envs/omg/bin/python -m pip install --no-deps -e upstream/omg +- +-python3.10 -m venv /data/hfd_work/envs/textop +-/root/.local/bin/turbo /data/hfd_work/envs/textop/bin/python -m pip install \ +- --extra-index-url https://download.pytorch.org/whl/cu128 -r requirements/textop.txt +-/root/.local/bin/turbo /data/hfd_work/envs/textop/bin/python -m pip install --no-deps \ +- -e upstream/textop/deps/CLIP -e upstream/textop/deps/isaac_utils -e upstream/textop/TextOpRobotMDAR +-``` +- +-固定外部代码revision,不换成最新分支: +- +-```bash +-/root/.local/bin/turbo git clone https://github.com/NVlabs/GENMO.git /data/hfd_work/hfd-human-models/genmo +-git -C /data/hfd_work/hfd-human-models/genmo checkout 16bebf402d8893184249ee206d957b8248cd8310 +-/root/.local/bin/turbo git clone https://github.com/nv-tlabs/kimodo.git /data/hfd_work/hfd-human-models/kimodo +-git -C /data/hfd_work/hfd-human-models/kimodo checkout e1c219c99b4b94ba48be1fb6a1cbc64ecee4c954 +-/root/.local/bin/turbo /data/hfd_work/envs/genmo/bin/python -m pip install --no-deps -e /data/hfd_work/hfd-human-models/genmo +-env SKIP_MOTION_CORRECTION_IN_SETUP=1 /root/.local/bin/turbo /data/hfd_work/envs/kimodo/bin/python -m pip install \ +- --no-deps --no-build-isolation -e /data/hfd_work/hfd-human-models/kimodo +-``` +- +-## 外部资源 +- +-完整身份清单在`TRAINING_ASSETS.json`。HFD索引与人体分片固定为`93ba4a7a9c66b3e1b7041e5f4ea038ef2418b948`,不能再用不含SMPL-X的旧版本。先准备已授权的Llama base及SMPL-X主体;其余缺失资产可按清单下载,不覆盖已有文件: +- +-```bash +-/root/.local/bin/turbo /data/hfd_work/envs/omg/bin/python tools/training_assets.py \ +- --models-root /data/hfd_assets/models --raw-root /data/hfd_assets/hfd_data \ +- --models omg textop genmo kimodo --omg-sizes 50m 100m 300m 500m --download +-/root/.local/bin/turbo /data/hfd_work/envs/omg/bin/python tools/download_hfd_tars.py \ +- --root /data/hfd_assets/hfd_data +-``` +- +-已有资产去掉`--download`进行SHA256核验;`--quick`只检查大小,`--plan`只列出所需文件及大小。HF源可使用授权的`hfmirror`重试;CLIP的Azure下载不能借此自动转用个人代理。只训练某一模型可缩小`--models`/`--omg-sizes`。不要在缺少原始tar时直接启动训练,训练缺分片会报错,不再隐式联网补齐。若复用既有完整资产,直接核验即可,不必重新下载。 +- +-准备下面的相对布局;权重和数据不在Git中: +- +-```text +-/data/hfd_assets/ +- hfd_data/ +- index.sqlite, MANIFEST.json, PATH_MAP.json +- metadata/seed_g1_smplx_pairs.parquet +- 原始tar分片(保留,不批量解压) +- models/ +- body_models/smplx/SMPLX_NEUTRAL.npz +- genmo/gem_smpl.ckpt +- genmo/t5-3b-local/ # 完整tokenizer/config/权重 +- kimodo/Kimodo-SMPLX-RP-v1/ # model.safetensors、config.yaml及完整stats/ +- kimodo/Meta-Llama-3-8B-Instruct/ # 完整base及分片 +- kimodo/LLM2Vec-Meta-Llama-3-8B-Instruct-mntp/ +- kimodo/LLM2Vec-Meta-Llama-3-8B-Instruct-mntp-supervised/ +- OMG/official/checkpoints/paper/50m/sstep=090000.ckpt +- OMG/t5-base-local/ +- TextOp-Data/TextOpRobotMDAR/logs/pretrained/checkpoint/{ckpt_200000.pth,vae.pth} +- TextOp-Data/TextOpRobotMDAR/dataset/PRIVATE-DATA/meanstd.pkl +- TextOp-Data/TextOpRobotMDAR/description/robots/g1/ +- TextOp-Data/clip/ViT-B-32.pt +- asr/Qwen3-ForcedAligner-0.6B/ # 重建对齐时需要 +-``` +- +-GenMo权重来自`nvidia/GEM-X` revision `5b9a69d5ed4713ac4ba2a63b335e66db87279a3e`的`gem_smpl.ckpt`。Kimodo权重来自`nvidia/Kimodo-SMPLX-RP-v1` revision `1419ba56b734c48bbafb41fefa84088ca94583b5`;两个LLM2Vec适配器revision和SHA256在`UPSTREAM_LOCK.json`。Llama本机来源为ModelScope的`LLM-Research/Meta-Llama-3-8B-Instruct`,也可按已有授权获取对应HF base,不能用量化版或不同base替代。SMPL-X主体需自行从已授权来源获取;本机neutral-2020文件SHA256为`bdf06146e27d92022fe5dadad3b9203373f6879eca8e4d8235359ee3ec6a5a74`。 +- +-GenMo需要四个额外身体回归/顶点表,URL、大小、revision和hash已固定: +- +-```bash +-/root/.local/bin/turbo /data/hfd_work/envs/genmo/bin/python tools/fetch_genmo_helpers.py \ +- --genmo-root /data/hfd_work/hfd-human-models/genmo +-/data/hfd_work/envs/genmo/bin/python tools/prepare_human_runtime.py \ +- --models-root /data/hfd_assets/models --genmo-root /data/hfd_work/hfd-human-models/genmo +-``` +- +-Kimodo安装时跳过setup中的MotionCorrection自动构建,避免其另行联网拉取构建依赖;随后由准备器使用已安装的Eigen和pybind11构建,并关闭CMake联网。准备器也建立本地base/adapter链接,不复制权重。构建需要C++17编译器和CMake>=3.15: +- +-```bash +-/data/hfd_work/envs/kimodo/bin/python tools/prepare_human_runtime.py \ +- --models-root /data/hfd_assets/models --kimodo-root /data/hfd_work/hfd-human-models/kimodo \ +- --build-correction --build-dir /data/hfd_work/cache/kimodo-motion-correction-build +-``` +- +-## 目标机器配置 +- +-```bash +-/data/hfd_work/envs/omg/bin/python tools/configure_training.py \ +- --work-root /data/hfd_work --assets-root /data/hfd_assets \ +- --raw-root /data/hfd_assets/hfd_data \ +- --external-code-root /data/hfd_work/hfd-human-models \ +- --env-root /data/hfd_work/envs --gpu 0 +-source /data/hfd_work/training-configs/training.env +-``` +- +-生成器按所选配方保留batch、预算、损失、调度、随机种子及数据规则,改成本机路径和GPU。默认四模型均使用GPU0,分别运行。已有OMG/TextOp资源可用`--robot-models-root /existing/models`复用;`--output`可改配置目录。它还生成`omg-{50m,100m,300m,500m}.yaml`和`genmo-{v2,legacy}.yaml`;`--omg-size`和`--genmo-recipe`选择默认项。默认OMG仍为50M,GenMo改为v2;更大OMG的微批次/累积与各原生损失见[复现矩阵](TRAINING_REPRODUCTION_MATRIX.md)。 +- +-## 英文词时间缓存 +- +-优先把现有`cache/hfd/reference-alignments.sqlite`迁移到新工作盘。它是源语音词对齐及第三方recipe索引,session仍动态组装。复制应在写入进程结束后保留完整SQLite状态。 +- +-没有缓存时创建独立Python3.10环境`forced-align`,按`runtime-versions.json`安装Torch,再安装`requirements/forced-align.txt`。准备`Qwen/Qwen3-ForcedAligner-0.6B` revision `c7cbfc2048c462b0d63a45797104fc9db3ad62b7`后运行: +- +-```bash +-/data/hfd_work/envs/forced-align/bin/python tools/align_reference_corpus.py \ +- --root /data/hfd_work/cache/reference-preparation \ +- --cache /data/hfd_work/cache/hfd/reference-alignments.sqlite \ +- --model /data/hfd_assets/models/asr/Qwen3-ForcedAligner-0.6B \ +- --language en --gpus 0 --batch-size 16 +-/data/hfd_work/envs/forced-align/bin/python tools/index_reference_recipes.py \ +- --cache /data/hfd_work/cache/hfd/reference-alignments.sqlite --language en \ +- --report /data/hfd_work/cache/reference-recipes.json +-``` +- +-双GPU可用`--gpus 0 1`。完成后应有61,860条英文源资产、缺失0;OMG/TextOp另用选定第三方英文recipe,GenMo/Kimodo仅用具有人体配对的原生英文session。重新对齐可能有数值差异;恢复原完整断点时应保留原缓存及代码/配置身份。原始tar、索引、PATH_MAP和配对metadata需完整;`tools/download_hfd_tars.py`可按已有manifest补齐分片。 +- +-## 预检、训练、恢复 +- +-从仓库根目录执行: +- +-```bash +-/data/hfd_work/envs/omg/bin/python train.py --config /data/hfd_work/training-configs/omg.yaml --dry-run +-/data/hfd_work/envs/textop/bin/python train.py --config /data/hfd_work/training-configs/textop.yaml --dry-run +-/data/hfd_work/envs/genmo/bin/python train_genmo_full.py --config /data/hfd_work/training-configs/genmo.yaml --check-only +-``` +- +-上述预检不更新参数。Kimodo的`--offline-check`会执行一次丢弃的梯度更新,加载全量索引及冻结文本worker。正式命令分别为: +- +-```bash +-/data/hfd_work/envs/omg/bin/python -u train.py --config /data/hfd_work/training-configs/omg.yaml +-/data/hfd_work/envs/textop/bin/python -u train.py --config /data/hfd_work/training-configs/textop.yaml +-/data/hfd_work/envs/genmo/bin/python -u train_genmo_full.py --config /data/hfd_work/training-configs/genmo.yaml +-/data/hfd_work/envs/kimodo/bin/python -u train_kimodo.py --config /data/hfd_work/training-configs/kimodo.yaml +-``` +- +-可在screen中前台执行。恢复时加`--resume auto`,不改变总预算;版本、路径或数据身份变化会拒绝伪装成同一次训练。`TRAINING_SNAPSHOT.json`与Git历史保留迁移前实现。跨机器从头复现使用生成配置;逐位复现还取决于缓存、完整随机状态、CUDA确定性及硬件,不能只凭seed保证。 +- +-GenMo旧权重的倾听姿态问题见[诊断说明](GENMO_LISTENING_DIAGNOSIS.md)。当前默认采用[GenMo v2训练策略](GENMO_V2_TRAINING.md),旧配方和原代码快照仍保留,未覆盖旧正式权重。新策略完整恢复使用`--resume auto`;旧权重迁入新策略使用`--init-checkpoint`,这是新run的权重初始化。 +diff --git a/docs/REFERENCE_DYNAMIC_TRAINING.md b/docs/REFERENCE_DYNAMIC_TRAINING.md +deleted file mode 100644 +index 99edc89..0000000 +--- a/docs/REFERENCE_DYNAMIC_TRAINING.md ++++ /dev/null +@@ -1,84 +0,0 @@ +-# 离线对齐与动态 session 训练 +- +-清理说明:本文记录的一次性实验脚本已归档,复现前请按[脚本恢复说明](WORKSPACE_CLEANUP.md)取回;历史指标与日志保留。 +- +-2026-10-19。当前实现保留原生 HFD 动态组装流程:缓存原始语音资产的词时间戳,训练时按实际组装位置平移时间戳。无需把全部 session 导出为 NPZ,也无需在线运行 ASR。 +- +-## 已落实的数据链路 +- +-1. `tools/align_reference_corpus.py` 从本地 tar 索引枚举原始任务、停止、附和、犹豫与修正语音,按唯一资产对齐。最初完成 123,720 个双语资产;按英文训练决策,当前库仅保留 61,860 个英文资产,英文缺失 0。没有批量解压或复制原始音频。 +-2. `/root/autodl-tmp/cache/hfd/reference-alignments.sqlite` 保存资产路径、原始音频 SHA256、16 kHz 长度、原文与词时间戳。训练时只读查表;不是把全部 session 常驻内存。进程内仍有受容量限制的 session/特征缓存。 +-3. 原生组装器在实际音频拼接时记录源资产和起止采样点,包括犹豫、停顿、修正、停止及附和。该元数据不改变动作、PCM、控制监督或随机组装流程。 +-4. 第三方已组装 session 按本地 `session_recipes.jsonl` 的前景目标语音位置映射,检查原始音频哈希与长度。背景音频保留,背景说话者的文本不作为条件。索引含 27,584 个 recipe、82,590 个目标片段;实际 train/val 仍以原选择清单的 26,737/783 个 session 为准。 +-5. 词结束后加实验性的 0.1 秒 guard,再向上取整到 `0.6, 1.0, 1.4, ...` 秒释放。模型只看到当前时刻已释放、过去 30 秒内的文字。原生动作时钟与麦克风时钟的 0.2 秒偏移按 `clock_policy: microphone` 统一处理。 +- +-24 个双语冻结样例的动态 PCM、动作与控制监督逐元素相同;动态生成的完整文本 records 也与先前导出的 `.reference.json` 完全相同。该检查覆盖 train/val、中英及六种交互类型,属于代表性回归检查,不表示已逐一实例化所有原生 session。 +- +-这属于 **oracle 前景转写条件**,不能作为真实 ASR 识别准确率实验。强制对齐也是时间估计,0.1 秒不是已证明的误差上界。历史双语库 8,342/123,720(6.74%)个资产存在无效跨度;当前保留的英文资产中为 6,608/61,860(10.68%)。这些资产保守地整段结束后释放文字;不猜测缺失词边界,不丢弃文本。 +- +-## 模型实际输入 +- +-| 模型 | 条件 | 监督与输出 | +-| --- | --- | --- | +-| OMG | 当前可用的过去 2 秒音频、已释放的最近 30 秒文字、原生动作历史 | 原生 G1 动作表示与原生训练损失 | +-| TextOp | 已释放的最近 30 秒文字、原生动作历史 | 原生 G1 动作表示、四 primitive 的原生三阶段 DAR 训练 | +- +-不输入控制标签、未来音频或未来原文。当前只采样英文前景 session,`text_conditioning: none`,不加载 ASR 或翻译模型。背景音频仍原样保留。中文翻译实验属于历史对照。 +- +-“保留 30 秒文字”是前端上下文约定,不代表编码器具有无限 token 容量。TextOp 的 CLIP 保留预算内的最近文本后缀;OMG 原生 T5 为 50 token,新 reference 配置通过适配工厂设置 `truncation_side=left`,超长时保留最近文字。架构、长度、参数与严格 checkpoint 加载不变;旧 run 不追溯修改。 +- +-## 配置和复查 +- +-新配置为 `configs/omg_reference_train.yaml`、`configs/textop_reference_train.yaml`,使用 `native_ondemand` 和 `aligned_reference_context`。`omg.yaml`、`textop.yaml` 也已同步为英文 reference 默认;旧 ASR 对照移至 `configs/history/`。OMG 使用物理 GPU 1,TextOp 使用物理 GPU 0。 +- +-```bash +-/root/autodl-tmp/envs/omg/bin/python train.py --config configs/omg_reference_train.yaml --dry-run +-/root/autodl-tmp/envs/textop/bin/python train.py --config configs/textop_reference_train.yaml --dry-run +-``` +- +-`--dry-run` 检查完整对齐索引、第三方 recipe 表和权重路径;完整读取与训练仍需短测。构建索引约需数分钟 CPU,首次组池和共享盘读取不能算作纯 GPU 更新吞吐。当前英文路线不做翻译。 +- +-本轮证据目录:`/root/autodl-tmp/runs/archive/legacy-20261020/reference-full-preparation-20261019/`。其中 `coverage.json`、`alignment-quality.json`、`dynamic-verification.json`、`provider-verification.json`、`translation-*.json` 分别记录覆盖、回退、动态文本、实际 provider 和翻译诊断。配置中的 100K/300K 步仍是旧预算;batch 改变后不能直接当作已验证的推荐预算。 +- +-TextOp 旧单 session 过拟合的 teacher-forced 短窗误差下降,但连续闭环误差变差。当前同 seed 引导系数 1 的补充检查仍是初始 0.3004 → 训练后 0.3121 rad;单纯把系数从 5 改成 1 没有解决。该证据不支持宣称全双工或长程动作能力已经验证。 +- +-全量正式训练尚未启动。当前英文短测报告将保存在 [纯英文短测报告](ENGLISH_ONLY_PROBE.md);历史双语结果不能直接与英文验证损失比较,必须使用同一英文回放集合。 +- +-## 历史双语短测结果 +- +-两模型都从官方权重初始化,batch=64,使用完整原生加第三方集合的随机窗口采样;不是逐一训练完所有 session。 +- +-| 检查 | OMG | TextOp | +-| --- | --- | --- | +-| 更新预算 | 512 步 | 768 步,三阶段各 256 步 | +-| 实际训练 session | 512 个槽位、510 个不同 ID | 192 个槽位、192 个不同 ID;每组四 primitive | +-| 原生验证损失 | 0.07212 → 0.04290,固定窗口和噪声 | 0.14213 → 0.14530,原生随机验证 | +-| 权重检查 | 512 次优化更新、梯度和损失有限、参数变化、严格重载通过 | 记录的梯度范数和损失有限,108 个状态张量变化,严格重载通过;未另挂 hook 独立计数 optimizer.step | +- +-OMG 池内单步中位耗时 0.115 秒,而包含冷池、验证和保存的训练阶段共约 1,064 秒;未包含此前索引与模型启动,不能用池内速度直接外推。OMG 每个 64-session 池准备中位约 103 秒;TextOp 三个池约 105/97/125 秒。当前耗时主要仍在共享盘读取、session 组装、首次翻译及特征准备,尚未证明正式 batch512 的端到端效率。 +- +-OMG 最后一轮因框架 lookahead 多准备了一个未训练的池,约 94 秒。这项开销已通过训练/验证 batch 上限修正;真实 CPU Lightning/DataLoader 的四步边界检查确认只请求四个源 batch,未触发下一冷池。57 项相关测试通过。本轮真实模型曲线和耗时来自修复前,未人为扣除这段耗时,也未重跑整个短测。 +- +-孤立批量翻译探针:64 条真实中文释放文本,batch16 与逐条结果 64/64 完全相同,纯模型翻译约 0.49 秒对 2.41 秒。这是后续优化依据,尚未改训练默认翻译策略,也没有解决译错动作的问题。 +- +-证据:运行目录下 `summary.json`、`pilot-curves.png`、`gpu-utilization.png`、`textop-checkpoint-verification.json`、`lightning-batch-bound-verification.json`;原生 run 的配置、采样记录、日志、验证曲线及 checkpoint 均保留。短测固定验证损失下降不等于连续闭环动作质量通过;旧 100K/300K 预算仍待校准。 +- +-## 正确中文直接输入的检查 +- +-2026-10-19 补充:文本转写正确与编码器支持中文是两项独立条件。原生冻结编码器实测如下: +- +-- OMG 的 T5-Base 将“向前走”“向后走”“向左转”“向右转”“停止动作”“继续动作”等 16 条中文对照指令全部编码为 `[3, 2, 1]`,其中 2 为 unknown token。相同 token 与 mask 的文本条件相同,后续仅训练动作网络无法恢复丢失的指令区别;音频及动作历史仍可能驱动输出。直接解冻这个编码器也不能区分已经被 tokenizer 合并的输入。 +-- TextOp 的 CLIP 字节 BPE 能保留不同中文字符,不能因此断言直接中文训练不可能。但当前冻结 CLIP 在这组手写 16 类中英文本匹配上仅 1/16 top1 正确。这是编码空间的小型诊断,不是中文动作准确率。 +- +-用上述短测 checkpoint,在同样两条中文验证 session(打断、急停)上保持种子、PCM、初始动作历史及文本释放时间相同,各回放中文原文、OPUS 英文和空文本,共 12 次。两条 session 的等权平均关节 RMSE: +- +-| 模型 | 直接中文 | 翻译英文 | 空文本 | +-| --- | ---: | ---: | ---: | +-| OMG | 0.4406 | 0.5121 | 0.4069 | +-| TextOp | 0.4355 | 0.3804 | 0.4307 | +- +-单位 rad,单种子 12345。OMG 的中文误差确实低于译文,但空文本更低;结合 token 坍缩,不能把这一结果解释成读懂中文。TextOp 译文在两个样例中均比中文误差低,但闭环关节误差仍不等于执行指令语义的准确率。 +- +-这只是已按英文条件训练的 checkpoint 的推理消融,尚未做直接中文与翻译英文的等预算训练比较。OMG 如需原生读懂中文,必须调整 tokenizer/encoder;TextOp 可单独做直接中文训练实验。更换中文/多语编码器还需要对齐原动作条件空间,并应标为独立模型变体。正式配置本轮未改变。 +- +-复现工具 `tools/probe_direct_chinese.py`;证据目录 `/root/autodl-tmp/runs/archive/legacy-20261020/direct-chinese-probe-20261019/`,其中 `encoders.json`、`omg-rollout.json`、`textop-rollout.json` 和 `summary.json` 保留完整结果。 +diff --git a/docs/ROLLOUT_DEBUG.md b/docs/ROLLOUT_DEBUG.md +deleted file mode 100644 +index fc94a97..0000000 +--- a/docs/ROLLOUT_DEBUG.md ++++ /dev/null +@@ -1,82 +0,0 @@ +-# 视频问题调试:输入、监督范围与执行节奏 +- +-清理说明:本文记录的一次性实验脚本已归档,复现前请按[脚本恢复说明](WORKSPACE_CLEANUP.md)取回;历史指标与日志保留。 +- +-本轮使用同一个原生双任务 session:接取别人递来的物品尚未完成时,输入捡起地上物品的新指令。数据未手工拼接,混合录音、背景音、ASR 与关节映射未改变。所有输出位于 `runs/hfd-final/`;主视频目录只展示两条当前结果,所有对照集中在 `debug/`。 +- +-## 结论与已采用修改 +- +-1. **监听姿态缺失首先是监督范围问题。** 之前把“不输入 listen 等控制标签”扩大成了“不训练这些帧的关节动作”。两个模型没有控制分类头,但都能拟合这些关节姿态。经用户确认改成 `full_session`:完整动作进入原生损失,控制标签仍不作为模型输入。原来 TextOP 仅 35 条四段连续训练链,OMG 仅 84 个两秒窗口;补回后分别为 88 条与 366 个窗口。监听、起步和收尾不再被过滤。 +-2. **OMG 重复动作与缺少收尾监督有关;卡顿还与异步计划交接有关。** 原先深弯腰在约 2.07、7.93、10.30、12.77 秒出现四次,当前主结果只有 9.57 秒一次,与真值相符。异步策略计算时执行旧计划,完成后跳过新计划前缀;两个计划在交接时的姿态不保证一致。最大跳变原来正好落在交接帧,并非播放器掉帧。该策略参照原生 `upstream/omg/src/omg/cli/pipeline/modes.py` 的 bridge/skip 流程,直接参考姿态回放没有物理跟踪器去缓和突变。 +-3. **TextOP 连续生成仍不合格,不能靠拉长文本窗口修好。** 它执行全部原生 8 帧(50 Hz 下 0.16 秒),且 2 帧动作历史是预测反馈,没有每个窗口重新初始化。固定真值历史时误差很小,但用自己的历史后逐渐停滞或延迟。完整监督后能出现监听姿态,却长时间停留在那里;捡取弯腰峰值约 11.66 秒,真值为 9.58 秒。延长训练反馈链也没有可靠解决。 +- +-当前 OMG 采用完整动作监督、过去两秒音频+持续因果 ASR、原生 DDIM 10 步、`joint_one` CFG、60 帧计划、剩余 24 帧时异步预取,使用实际测得规划耗时。修改的是原生采样参数,没有换网络、损失或添加手工平滑。TextOP 保留完整监督和原生 0.16 秒执行节奏,当前结果明确标记尚未收敛。 +- +-## 同 checkpoint 的对照 +- +-所有主对照固定 seed=12345。RMSE 为整段连续关节误差,排除开场真值初始化。TextOP 统计其 23 个关节,OMG 统计 29 个,不能把二者绝对值当作直接模型排名。 +- +-| 旧 motion-only checkpoint 的推理变化 | 连续 RMSE (rad) | 说明 | +-| --- | ---: | --- | +-| OMG 原推理 | 0.3294 | 50 步、测量延迟、异步 | +-| OMG 假设零延迟 | 0.3510 | 仅诊断;最大跳变从 0.4288 降到 0.1854 | +-| OMG 每 0.4 秒截断重规划 | 0.3708 | 不模拟计算耗时,仍差 | +-| OMG 执行完整 2 秒计划 | 0.3226 | 不模拟计算耗时,末尾仍重复 | +-| OMG 音频改为 0.6 秒 | 0.3204 | 零延迟,与训练 2 秒输入不匹配,仅诊断 | +-| OMG 音频改为完整前缀 | 0.3389 | 零延迟,前缀压成 60 个声学位置;仍无改善 | +-| OMG 去掉 ASR | 0.3867 | 零延迟,保持音频;文本条件有作用 | +-| OMG 去掉音频 | 0.3615 | 零延迟,保持 ASR;不能替代联合输入 | +-| TextOP 原推理 | 0.2434 | 原生 8 帧输出 | +-| TextOP 文本每 1.92 秒刷新 | 0.2441 | 仍执行原生 8 帧,改善几乎为零 | +- +-音频窗口和条件屏蔽是同 checkpoint 的敏感性检查;由于改变了训练时的输入分布,不能据此决定新训练的最佳窗口,更不能将诊断零延迟作为实时结果。 +- +-## 从官方权重重训,仅改为完整监督 +- +-TextOP 3200 次更新,阶段 `[800,1600,800]`,每次原生训练链 4 段;OMG 1600 次更新。其余参数与旧对应 run 保持一致。两者都通过有限 loss/梯度、参数更新、原生 checkpoint 严格加载检查。 +- +-| 完整监督后的设置 | 连续 RMSE (rad) | 最大逐帧跳变 RMS (rad) | +-| --- | ---: | ---: | +-| OMG 50 步+实际延迟 | 0.1865 | 0.3166 | +-| OMG 50 步+假设零延迟 | 0.0886 | 0.1242 | +-| OMG 10 步+实际延迟(当前) | 0.0619 | 0.1160 | +-| OMG 20 步+实际延迟 | 0.0622 | 0.1431 | +-| OMG 50 步+固定 5 帧延迟 | 0.0795 | 0.0976 | +-| OMG 10 步+固定 15 帧延迟 | 0.0753 | 0.2395 | +-| OMG 每次完整执行 2 秒,不模拟耗时 | 0.2293 | 0.0839 | +-| OMG 每 0.4 秒截断,不模拟耗时 | 0.2995 | 0.2400 | +-| TextOP 原生连续反馈(当前) | 0.2992 | 0.0197 | +-| TextOP 每块提供真值历史(oracle) | 0.0125 | 不作为连续性能 | +-| TextOP guidance=1 | 0.3247 | — | +-| TextOP 原生默认站姿初始化 | 0.4385 | — | +-| TextOP 去掉 ASR | 0.3260 | — | +-| TextOP 连续训练链从 4 段加到 16 段 | 0.2735 | 0.3357 | +- +-固定延迟交叉对照表明,采样轨迹与交接延迟都会影响结果,不能把全部误差改善只归因于速度。10 步采样平均规划约 0.157 秒,20 步约 0.230 秒,先前 50 步约 0.5 秒。没有手工插值、滤波或利用真值重置来美化当前视频。 +- +-OMG 10 步在 seed=12345/12346/12347 下 RMSE 分别是 **0.0619 / 0.1538 / 0.0585 rad**。主视频沿用原先 seed=12345,没有替换成最优种子。仍有随机性和残余接缝;真值最大逐帧跳变是 0.0355 rad,当前 OMG 的 0.1160 仍大于真值。 +- +-## 为什么 TextOP 的窗口很准,完整视频却不对 +- +-同一 ASR、同一个种子下,使用任务中的真值历史,8 帧误差约 0.0037–0.0054 rad;使用开场 HFD 站姿,预测的短期动作幅度约 0.003 rad。该旧 checkpoint 从原生默认站姿可以产生更大动作,但完整回放并未因此变好,说明不是简单换初始化就能解决。 +- +-完整监督的 TextOP 在每块真值历史诊断中误差 0.0125 rad,预测历史闭环却为 0.2992 rad,支持“训练短片段可拟合,闭环历史分布与任务阶段跟踪不稳定”的判断。缺少 KV cache 不是这里已经证实的唯一原因。TextOP 只有 ASR 条件:同一稳定句子可跨监听、执行、完成多个阶段;ASR 本身不显式提供动作是否已完成。保留完整标签用于动作监督不会自动增加阶段记忆。 +- +-16 段连续训练(2.56 秒)比 4 段(0.64 秒)误差略低,却引入大跳变,因此未采用。当前 TextOP 仍需更系统的原生训练与闭环验证;此次不将扩充模型状态或增添控制头偷偷并入 baseline。 +- +-## 复现与查看 +- +-当前配置:`configs/omg_full_session.yaml`、`configs/textop_full_session.yaml`。重训时改 `paths.output_dir` 为新目录。正式模板也已改为 `full_session`。当前 OMG 配置保存 `inference.omg_sampling_steps: 10`;TextOP smoke 保留此前诊断的 `static_prob=0.05`,正式模板仍是上游设置,应在论文里独立记录这一超参数。 +- +-```bash +-# 在对应模型环境运行;输出目录需尚不存在 +-python infer_overfit.py --run /root/autodl-tmp/runs/archive/legacy-20261020/hfd-final/omg --output-dir /data/runs/omg-replay --continuous-only --skip-window-eval --task-seconds 0 14.8 +-python infer_overfit.py --run /root/autodl-tmp/runs/archive/legacy-20261020/hfd-final/textop --output-dir /data/runs/textop-replay --continuous-only --task-seconds 0 14.8 +-# 仅作反事实诊断:--omg-latency-frames 0、--textop-history-mode teacher_forced +-# 其他对照:--omg-sampling-steps 50、--omg-execute-frames 60 --omg-replan-mode receding +-python tools/summarize_rollout_debug.py --root /root/autodl-tmp/runs/archive/legacy-20261020/hfd-final +-``` +- +-主视频:`videos/omg_full_session_playable.mp4`、`videos/textop_full_session_playable.mp4`。曲线:`debug/trajectory_comparison.png`。全部数值:`debug/summary.json`;旧结果在 `debug/previous_motion_only/`。这些均为有原始音频的关节参考姿态回放,没有物理跟踪器。 +- +-本轮沿用 HFD 名义动作时钟和音频释放映射,ASR 为预计算的因果 sidecar,未加入实际 ASR 计算延迟。OMG 仅计入本次声学特征、文本编码和模型规划耗时,不等同于完整麦克风到机器人端到端实时系统。所有测试来自同一个训练 session,不能作为独立测试集或论文泛化结论。 +diff --git a/docs/SEMANTICS_CLOCK_FIX.md b/docs/SEMANTICS_CLOCK_FIX.md +deleted file mode 100644 +index 82611b7..0000000 +--- a/docs/SEMANTICS_CLOCK_FIX.md ++++ /dev/null +@@ -1,59 +0,0 @@ +-# 中文语义与麦克风时钟修复 +- +-清理说明:本文记录的一次性实验脚本已归档,复现前请按[脚本恢复说明](WORKSPACE_CLEANUP.md)取回;历史指标与日志保留。 +- +-2026-10-18。修复已接入新 OMG/TextOp 训练配置、共享按需/third-party 数据入口、GenMo 文本缓存与窗口、连续推理和视频评测。未重训模型;历史 run、权重、ASR 和旧视频保持不变。HY-Motion/Kimodo 继续以同一输入契约为后续适配目标,其 HFD 训练器尚未实现。 +- +-## 语义条件 +- +-保留 Whisper-tiny `task=transcribe` 的原始 ASR,先按 `causal_asr_context` 获取已释放的完整假设,再通过冻结的 [Helsinki-NLP/opus-mt-zh-en](https://huggingface.co/Helsinki-NLP/opus-mt-zh-en) 转成英文。纯英文和空文本原样通过,仍由模型各自的原生文本编码器编码。没有使用 GT 指令或动作标签修正文本。 +- +-`慢跑向前` → `Jogging forward.`;`向后跳` → `Jump back.`。原先这些中文在 OMG T5-Base 中容易合并为 `[3,2,1]`,其中 `2` 是 UNK。六条指令的回归测试验证了 T5 token 可区分、无 UNK,保留左右手、否定和次序。实际三个样本共 42 次规划,翻译后文本均无 UNK,最长 19 token,未触发原生 50 token 截断。 +- +-实现:`hfd_bridge/semantics.py`。模型 eval、无梯度、beam=4、无采样,逐个 ASR 假设独立转换;长输入按 128 源 token 分块,输出遇到长度限制显式失败,不静默丢弃历史。初始化保护训练/采样 RNG。SQLite 缓存按模型身份、解码参数、规范化器版本和源字符串区分;缓存有容量限制,不跨释放点积累翻译记忆。 +- +-模型唯一存放在 `/root/autodl-fs/pangxinlei/models/translation/opus-mt-zh-en`,权重 312,087,009 字节。HF metadata 经 AutoDL 获取,权重在 AutoDL 超时、HF 镜像限流后从 ModelScope 国内直连下载(`trust_env=False`),未回退个人代理;本地 `download_identity.json` 记录版本和 SHA256。OMG/TextOp 依赖锁新增所需 tokenizer/规范化依赖;GenMo 环境已安装 `sacremoses==0.2.0`,其余依赖沿用现有环境。可用 `HFD_TRANSLATOR_ROOT` 指定相同身份的本地模型。 +- +-试验过 Whisper 直接 `task=translate`,但本例翻译误差明显,已放弃。输出目录下 `asr-en/` 是被拒绝的候选,三个修复视频均使用原始 `omg-100k-validation/asr` 的 transcribe sidecar,再做上述独立文本翻译。 +- +-## 信息可用时刻与视频 +- +-HFD 原始动作时刻 `t_motion` 对应麦克风时刻 `t_mic=t_motion+0.2s`。原始音频窗口 `[0.4k,0.4k+0.6]` 在 `r_k=0.6+0.4k` 到达。新 `microphone` 策略只读取满足 `r_k<=t_mic` 的最近完整窗口和 ASR。首个窗口还未到达时输入为空文本及静音。音频先裁剪再做归一化、重采样和卷积特征,未来 PCM 不参与这些运算。 +- +-例如原生动作 0.4s 请求对应麦克风 0.6s,只能用到 PCM 0.6s;历史 `legacy_hfd` 会选择到 1.0s。共享实现为 `hfd_bridge/clocks.py`,30Hz/50Hz 数据与推理共用;原始 PCM 和 sidecar 不需要重做。 +- +-完整视频从麦克风 0s 播放,预测与 GT 同时前置 0.2s 首帧(OMG/GenMo 6 帧、TextOp 10 帧)。这是显示时间变换,原生参考轨迹和计分范围保持不变;不是单独平移预测去匹配 GT。GenMo 仍只对完整预测和完整人体 GT 各 retarget 一次,原生 `retargeted.npz` 保持原始时间,显示前缀只写入视频比较文件。 +- +-新配置显式记录: +- +-```yaml +-caption_policy: causal_asr_context +-clock_policy: microphone +-text_conditioning: asr_to_english +-``` +- +-OMG/TextOp 位于 `data` 下,GenMo 位于顶层。无新字段的旧 run 保留 `legacy_hfd`、不翻译;新 GenMo 缓存记录翻译策略并包含启动所需空文本 embedding,不允许将旧缓存误当成新输入。 +- +-**延迟口径:** ASR/翻译为逐释放点因果离线缓存,计算耗时没有加入回放时间;OMG 规划仍使用实测生成耗时。信息因果性已验证,但不代表测得了包含前端的端到端实时延迟。 +- +-## 验证与同权重回放 +- +-输出:`/root/autodl-tmp/runs/archive/legacy-20261020/omg-semantics-clock-fix`。使用原始 100k OMG 权重、seed12345、DDIM10、native async、过去 2s PCM;对照旧 `causal_asr_context` 视频,同时修复语义与时钟,不能用这组三样本区分两项修复各自的效果。 +- +-| 样本 | 旧关节 RMSE(rad) | 新关节 RMSE(rad) | 旧根位置 RMSE(m) | 新根位置 RMSE(m) | +-| --- | ---: | ---: | ---: | ---: | +-| interruption | 0.3244 | 0.3539 | 1.6918 | 1.5484 | +-| backchannel | 0.2625 | 0.2776 | 0.0100 | 0.0112 | +-| emergency-stop | 0.4614 | 0.3864 | 0.4912 | 0.1547 | +- +-检查结果: +- +-- 70 项共享回归测试通过,另有 11 项 GenMo 几何/音频测试通过。 +-- OMG/TextOp 的真实数据契约检查通过:未来动作、未来 ASR/标签、未来 PCM 不影响当前条件,标量/批量音频特征一致。 +-- 三个视频共 1,638 帧完整解码且有音轨;NPZ 音频严格等于原始 PCM 从 0 开始的相应片段。 +-- 新视频去掉共同显示前缀后,GT 与旧评测逐元素完全相同;三个样本均链接同一原始 checkpoint。 +-- 42 次规划的 PCM 截止和 ASR 释放时刻均不晚于请求的麦克风时刻;保存每次原始 ASR、英文条件及实际 T5 token。 +- +-可复查 `verification.json`、`tests.log`、`genmo-tests.log`、`omg-contract.log`、`textop-contract.log`。逐次验证脚本为 `tools/verify_semantics_clock.py`。 +- +-**当前限制:** 输入已能区分中文,但 Whisper 错字、静音幻觉和翻译误差仍在;完整 ASR 也不保证模型理解纠错。旧权重曾在中文 UNK/旧时钟下训练,此次没有重训,动作仍有明显偏差,指标并非全面改善。不能据此宣称提前倾听、不执行指令等模型行为已经修好。 +diff --git a/docs/TEXTOP_ENGLISH_VALIDATION.md b/docs/TEXTOP_ENGLISH_VALIDATION.md +deleted file mode 100644 +index ecfd00c..0000000 +--- a/docs/TEXTOP_ENGLISH_VALIDATION.md ++++ /dev/null +@@ -1,83 +0,0 @@ +-# TextOp 300K 英文验证评测(2026-10-21) +- +-训练目录:`/root/autodl-tmp/runs/hfd-textop-english-train`,使用完整训练后的 +-`ckpt_300000.pth`,与官方初始化checkpoint做同条件对照。没有继续训练或改动已有权重。 +- +-结果根目录:[textop-english-validation-20261021](/root/autodl-tmp/runs/textop-english-validation-20261021/README.md)。 +-包含18条有原混合音轨的视频、原始qpos、文本/token审计、指标CSV、汇总JSON、比较图及验证记录。 +- +-## 样本与协议 +- +-- 六类原生英文验证各一条:普通指令、打断、急停、附和、犹豫、暂停;另取三条第三方验证。 +-- 使用训练时相同的val provider、配方、seed42、原生与第三方选择摘要和离线文本库; +- 导出动作、音频及逐帧文字与动态provider逐值相同,原生任务source hash划分核对通过。 +-- 样本在生成前选定:各类首条估计6–30秒的合格英文session,第三方按合格ID等距选取。 +- 不是按结果挑选。当前没有独立test,val曾用于训练监测。 +-- 两组生成seed均12345;相同一次初始GT历史/根锚点,之后只反馈预测。 +- 原生2帧57D历史、8帧primitive、50Hz、CFG5,使用训练配置中的5步原生扩散。 +- 覆盖所选session内的完整primitive,末尾不足8帧不额外补齐或重置GT。 +-- 文本为`aligned_reference_context`:已释放最近30秒英文GT前景文本,CLIP77最近后缀; +- microphone时钟、0.1秒guard,不运行ASR/翻译,不输入音频或控制标签。 +-- TextOp直接生成G1,不做人体系retarget;未建模腕关节固定,主关节指标只算原生23个关节。 +-- 视频左预测右GT,保留0.2秒共同microphone前置帧;评分排除前置及GT bootstrap, +- 不做DTW或事后根对齐。同步运动学回放,不把实测规划耗时模拟为控制器延迟。 +- +-## 结果 +- +-以下按session等权平均,关节单位rad,根xyz分量RMSE单位m: +- +-| 范围 | 初始关节RMSE | 300K关节RMSE | 初始根RMSE | 300K根RMSE | +-| --- | ---: | ---: | ---: | ---: | +-| 全部9条 | 0.4380 | 0.4152 | 1.2243 | 1.4877 | +-| 原生6条 | 0.4481 | 0.4569 | 0.7812 | 0.9121 | +-| 第三方3条 | 0.4178 | 0.3318 | 2.1105 | 2.6390 | +- +-整体关节RMSE降低约5.2%,根RMSE增加约21.5%;原生六类未整体改善。 +-附和、暂停样本改善,急停和打断仍有明显问题,不能以局部或平均关节误差下降宣称任务能力提升。 +- +-| 原生场景 | 初始关节RMSE | 300K关节RMSE | 初始根RMSE | 300K根RMSE | +-| --- | ---: | ---: | ---: | ---: | +-| 普通指令 | 0.4346 | 0.4541 | 0.2770 | 0.1805 | +-| 打断 | 0.4084 | 0.3946 | 1.3258 | 2.6228 | +-| 急停 | 0.5908 | 0.6995 | 1.4652 | 2.1198 | +-| 附和 | 0.3017 | 0.2429 | 1.5385 | 0.5089 | +-| 犹豫 | 0.4440 | 0.5042 | 0.0399 | 0.0328 | +-| 暂停 | 0.5091 | 0.4462 | 0.0410 | 0.0079 | +- +-优先查看: +- +-- [急停300K](/root/autodl-tmp/runs/textop-english-validation-20261021/videos/emergency-stop-trained.mp4) +-- [打断300K](/root/autodl-tmp/runs/textop-english-validation-20261021/videos/interruption-trained.mp4) +-- [附和300K](/root/autodl-tmp/runs/textop-english-validation-20261021/videos/backchannel-trained.mp4) +-- [全部对照与指标](/root/autodl-tmp/runs/textop-english-validation-20261021/README.md) +- +-急停样本`val-000111-seed42`中,`Stop.`在麦克风10.6秒已经释放且实际送入CLIP; +-12–16.88秒区间,预测平均根速度约1.008m/s、GT约0.014m/s,模型仍持续运动。 +-这是单条样例的停止后运动诊断,不是经过统一定义的停止延迟指标;可以排除该例是因 +-停止文字未释放、被token截断或ASR漏识别造成。不能仅凭此确定训练目标、历史依赖或文本理解 +-中哪一项是主要原因。详见结果目录的`emergency-stop-diagnostic.json`。 +- +-训练日志最终`eval_total=0.10740`;最后100K阶段最低约0.10626(250K)。 +-训练验证样本/阶段历史分布与连续闭环不同,不能把该loss解释为上述视频误差或用全程曲线 +-单独证明过拟合。完整抽取的150个验证点见结果目录`training-curve.json`。 +- +-## 复现与验证 +- +-```bash +-cd /root/autodl-tmp/hfd-baseline-portable +-/root/autodl-tmp/envs/textop/bin/python -u evaluate_textop.py \ +- --trained-run /root/autodl-tmp/runs/hfd-textop-english-train \ +- --output /root/autodl-tmp/runs/textop-english-validation-NEW \ +- --gpu 0 +-``` +- +-入口支持`--checkpoint`、`--stage prepare/run/summarize`。原生模型、冻结VAE、CLIP、统计量 +-均复用本地资产;不重新下载,不把验证样本写入训练数据。 +-推理入口已支持显式TextOp checkpoint,并可跳过无关训练窗口评测;每个primitive记录原文、 +-实际编码后缀、CLIP token ID、条件帧、释放时刻和规划耗时。 +-视频长文本按列宽换行,过长仅显示最近150字符;完整模型输入仍保留在report,不改变条件。 +- +-18条视频共16,504帧,完整解码、原音频一致、两权重GT一致、计划时钟/文本/token一致、 +-CLIP77与释放因果边界检查通过。配套25项数据/时钟/文本回归检查通过。 +-这些验证说明评测链路一致,不证明语义成功或统计显著改善。 +-共享stationary池、第三方与原生的源重叠及上游预训练重叠未排除,不能称完全未见动作泛化。 +diff --git a/docs/TRAINING_REPRODUCTION_MATRIX.md b/docs/TRAINING_REPRODUCTION_MATRIX.md +deleted file mode 100644 +index e420551..0000000 +--- a/docs/TRAINING_REPRODUCTION_MATRIX.md ++++ /dev/null +@@ -1,44 +0,0 @@ +-# 训练复现覆盖范围 +- +-当前发布四种 baseline 的训练入口,HY-Motion 不在已完成训练框架范围内。源码、正式配置、依赖版本、外部源码 revision 和必要资产身份均在仓库;数据与模型文件按单独授权准备,不上传权重/日志/测试脚本。 +- +-## 现有配方和 OMG 多个初始化权重 +- +-| 配方 | 入口/生成配置 | 训练预算 | 微批次 × 累积 | 说明 | +-|---|---|---:|---:|---| +-| OMG 50M paper | `train.py` / `omg-50m.yaml` | 100K optimizer steps | 512 × 1 | 已有 HFD 正式实验使用此初始化 | +-| OMG 100M paper | `train.py` / `omg-100m.yaml` | 100K | 64 × 8 | 独立输出目录;有效 batch 512 | +-| OMG 300M paper | `train.py` / `omg-300m.yaml` | 100K | 32 × 16 | 独立输出目录;有效 batch 512 | +-| OMG 500M paper | `train.py` / `omg-500m.yaml` | 100K | 16 × 32 | 独立输出目录;有效 batch 512 | +-| TextOp | `train.py` / `textop.yaml` | 300K,三阶段各 100K | 512 × 1 | DAR/VAE/CLIP/57D 统计及机器人资源配套 | +-| GenMo legacy | `train_genmo_full.py` / `genmo-legacy.yaml` | 100K | 32 × 1 | 保留原正式配方及旧代码快照 | +-| GenMo v2 | `train_genmo_full.py` / `genmo-v2.yaml` | 100K | 32 × 1 | 新恢复课程、执行段损失、闭环选权重 | +-| Kimodo | `train_kimodo.py` / `kimodo.yaml` | 100K | 8 × 1 | 当前因果动作表示、根锚点约束、预测历史恢复及冻结文本 worker | +- +-其他 OMG 尺寸是可复现的初始化与微调配置,不代表已经各自完成 HFD 全量训练。各尺寸的实际文件名和 SHA256 见 `configs/omg_variants.yaml`,下载来源固定为 `THU-MARS/OMG` revision `06a88419da9f9125d0b5480cc10966741a189358`。100M/500M 需显式打开与权重相符的音频/人体分支结构;人体条件的可用性mask仍为false,HFD适配不输入未来人体GT。500M 仅移除已知的 52 个废弃 `visual_film` 参数,其他缺失/多余参数仍拒绝加载。保留各原生 experiment 的损失配置,100M/500M 的 full loss 与 50M/300M 的 diffusion-only loss 不能混称完全相同的训练目标。 +- +-上述 OMG 微批次按单卡 32GB 保守设置,均保持有效 batch 512;实际峰值还受文本长度与数据影响,不能把单次检查当作最大 batch 调优。50M/100M/300M/500M 均已用对应微批次完成严格权重匹配和一次丢弃的参数更新。历史 HFD run 中的 `best`、98K、100K 等是同一次训练的不同 checkpoint,均由原生验证与保留策略产生;它们不是新的模型尺寸,也不在 Git 中。 +- +-## 生成与运行 +- +-按便携指南准备资产后: +- +-```bash +-/data/hfd_work/envs/omg/bin/python tools/configure_training.py \ +- --work-root /data/hfd_work --assets-root /data/hfd_assets \ +- --raw-root /data/hfd_assets/hfd_data --omg-size 300m --genmo-recipe v2 --gpu 0 +-source /data/hfd_work/training-configs/training.env +-/data/hfd_work/envs/omg/bin/python -u train.py --config /data/hfd_work/training-configs/omg-300m.yaml +-``` +- +-生成器始终输出四种 OMG 配置及两种 GenMo 配方;`omg.yaml` 和 `genmo.yaml` 保存所选默认项。所有配置默认单卡执行,不能把多条命令同时塞到同一块卡。 +- +-## 外部资产、数据与环境 +- +-`TRAINING_ASSETS.json` 记录必要模型、统计、机器人资源、SMPL-X 主体、原始索引和配对表的相对路径、大小、SHA256 以及已确认的下载来源。`tools/training_assets.py` 默认只验证;显式 `--download` 才会下载缺失文件,不覆盖已有文件。Llama 需按记录哈希获取同一份已授权 base,SMPL-X 主体仍需自行获取。派生 runtime adapter 路径与本机原生扩展由准备器重建。 +- +-HFD 原始数据版本必须使用 `93ba4a7a9c66b3e1b7041e5f4ea038ef2418b948`;旧 `7745110...` 缺少新增人体元数据/分片。先下载同一版本的索引、PATH_MAP、manifest 和人体配对表,再按 manifest 补齐 tar,保留 tar 不解压。训练不自动联网补缺失文件。 +- +-源文本时间戳使用 `reference-alignments.sqlite`。迁移已完成的数据库并校验身份可保持既有释放时间;也可按固定 ForcedAligner、依赖和离线对齐命令重建。重建可能有数值差异,不应声称与旧缓存逐位相同。 +- +-原始数据、外部模型和离线缓存齐备后,训练不再依赖当前服务器的 `setup/`、其他适配仓库、测试目录或运行日志。四模型的完整断点均包含优化器/调度/采样或课程所需状态。严格恢复仍要求相同代码、配方和资产;Git 历史保留原训练快照。完整操作见 [PORTABLE_TRAINING.md](PORTABLE_TRAINING.md),GenMo v2 见 [GENMO_V2_TRAINING.md](GENMO_V2_TRAINING.md)。 +diff --git a/docs/TRAINING_RESUME.md b/docs/TRAINING_RESUME.md +deleted file mode 100644 +index ba9c43b..0000000 +--- a/docs/TRAINING_RESUME.md ++++ /dev/null +@@ -1,43 +0,0 @@ +-# OMG / TextOp 训练续训 +- +-2026-10-19:两个原生训练入口已接入完整 checkpoint 恢复,适用于当前一卡 OMG、一卡 TextOp。正式配置 `training.resume: null` 表示新训练;已有输出目录在没有显式 resume 时仍会拒绝覆盖。 +- +-从仓库根目录运行,自动选择原输出目录的 checkpoint: +- +-```bash +-/root/autodl-tmp/envs/omg/bin/python train.py --config configs/omg.yaml --resume auto +-/root/autodl-tmp/envs/textop/bin/python train.py --config configs/textop.yaml --resume auto +-``` +- +-`--resume` 不带参数等价于 `--resume auto`。找不到文件会报错,不会退回从头训练。OMG 在根目录 `last.ckpt`、`checkpoints/last.ckpt` 和周期 checkpoint 中按保存的步数选择最新文件,避免旧的根目录 last 覆盖较新的周期保存点;TextOp 同样按保存步数选择。忽略原子保存中的 `.partial` 文件。 +- +-显式指定恢复点: +- +-```bash +-/root/autodl-tmp/envs/omg/bin/python train.py --config configs/omg.yaml --resume /path/to/step-00002000.ckpt +-/root/autodl-tmp/envs/textop/bin/python train.py --config configs/textop.yaml --resume /path/to/ckpt_2000.pth +-``` +- +-也可在 YAML 设置 `training.resume: auto` 或 checkpoint 绝对路径。显式命令行参数优先。可以追加 `--dry-run` 先检查 checkpoint、配置与输出目录。 +- +-## 恢复内容与边界 +- +-- OMG 由 Lightning 恢复模型、优化器、scheduler、训练步数、precision 和原生 callback 状态。 +-- TextOp 恢复 denoiser、AdamW、全局步数及启用时的 EMA;学习率退火和三阶段位置由原生逻辑按已恢复步数继续计算。 +-- 新 checkpoint 额外保存 Python、NumPy、Torch/CUDA 随机状态、动态池的 session 索引、批次位置和采样 RNG。OMG 保存实际已消费 batch 的池快照,不使用可能已被预取推进的现场状态。 +-- 不保存大型音频/动作内存缓存。恢复时重建当前池所需内容,首次读取可能仍慢。 +-- TextOp 必须在完整 primitive 链边界恢复;配置检查要求保存间隔能被 `num_primitive` 整除。 +- +-`max_steps` 是原训练计划的总目标,不是恢复后再增加的步数。续训会检查模型、数据语言、文本时钟、batch、学习率、总步数、scheduler/阶段等配置一致;允许更换单张 GPU、调整保存/验证频率,或指定新的空输出目录。当前没有接入多 GPU 的完整 HFD 采样恢复。 +- +-官方预训练权重使用 `weights.init_checkpoint` 初始化,不能当作此次 HFD 续训点。历史 HFD checkpoint 如缺少新增随机/数据池状态,仍可恢复其完整优化器和步数,但会明确提示采样不能精确接续。新 checkpoint 的训练恢复也不承诺 GPU 逐位一致;原生验证的执行与随机数消耗会影响后续数值轨迹。 +- +-只能恢复到最近成功保存的位置,断电前尚未保存的更新不会恢复。两模型正式配置现在均每 **2000 步**保存,保留最近 **3** 份;OMG 另存最佳验证 checkpoint。100K/300K 总预算未改动,正式长训练未启动。 +- +-## 日志与检查 +- +-恢复后标量日志追加写入;原始 `resolved_config.yaml` 保留。每次恢复在 `resume_history//` 保存本次配置、checkpoint 来源、代码摘要、数据清单及恢复前的原生配置/训练摘要。显式回退到较早 checkpoint 时,标量日志可能出现重复步数,可用恢复记录区分。 +- +-已通过两个真实 GPU 原生训练入口的 **4 步停止 → 恢复到 8 步**检查:恢复后各完成4次更新,Adam状态计数从4到8,OMG scheduler接续,TextOp跨阶段接续,有限损失/梯度与严格重载通过。该 smoke 使用已导出的英文 session;动态池与预取的一致性另用 CPU Lightning 回归覆盖,不代表正式全数据训练已完成。 +- +-分环境共 **132 项回归测试**通过,包括预取启用下的消费位置、优化器与 scheduler 连续性、TextOp 原生 manager 跨阶段、RNG 和采样池恢复。证据见 [验证记录](/root/autodl-tmp/runs/archive/legacy-20261020/resume-validation-20261019/verification.json)。本次验证生成的临时 checkpoint 已清理,配置、日志和结果保留。 +diff --git a/docs/WORKSPACE_CLEANUP.md b/docs/WORKSPACE_CLEANUP.md +deleted file mode 100644 +index 1551b42..0000000 +--- a/docs/WORKSPACE_CLEANUP.md ++++ /dev/null +@@ -1,78 +0,0 @@ +-# 工作区清理(2026-10-19) +- +-检查了 `git diff --stat`、逐文件 diff、`git status --short` 的未跟踪文件,以及运行时代码/测试的引用。仅用 `git diff` 会漏掉大量未跟踪脚本。新增的 bridge、动态数据读取、对齐、GenMo 和 retarget 模块仍是实际功能,不能按“未跟踪”直接删除。 +- +-已将 **43 个一次性实验脚本**移出工作区;包括旧 ASR/翻译探针、不同阶段的预算和吞吐 benchmark、中英 pilot、已完成过拟合与验证的专用汇总脚本。旧的 `omg_asr_context_overfit.yaml`、`textop_asr_context_overfit.yaml`、`omg_budget60k.yaml` 三份配置移入 `configs/history/`,引用已更新。未重置或提交用户已有改动。 +- +-以下保留在日常工作区: +- +-- 正式训练入口、当前配置与 resume 实现。 +-- 动态 session/缓存适配、文本时钟、音频和动作特征、映射等运行时模块。 +-- OMG/TextOp/GenMo 推理、retarget、视频评测入口及必要回归测试。 +-- 原始数据、模型、已有实验权重、日志和报告。 +-- 七个数据准备/审查工具:`align_reference_corpus.py`、`audit_reference_text.py`、`download_hfd_tars.py`、`index_reference_recipes.py`、`prepare_reference_sessions.py`、`verify_dynamic_reference.py`、`verify_reference_release.py`。 +- +-## 归档与恢复 +- +-脚本移出前已打包,并逐文件验证归档内容 SHA256,清单保存原路径、大小和哈希。归档在服务器工作区外,若迁移后还需复现实验,应同时复制该归档。 +- +-- [归档包](/root/autodl-tmp/hfd-code-archive/20261019-resume-cleanup/historical-experiments.tar.gz) +-- [逐文件清单](/root/autodl-tmp/hfd-code-archive/20261019-resume-cleanup/cleanup.json) +-- [清理前 Git 状态](/root/autodl-tmp/hfd-code-archive/20261019-resume-cleanup/git-status-before.txt) +-- [清理前已跟踪文件 diff](/root/autodl-tmp/hfd-code-archive/20261019-resume-cleanup/git-diff-before.patch) +- +-按需恢复某个脚本,例如: +- +-```bash +-tar -xzf /root/autodl-tmp/hfd-code-archive/20261019-resume-cleanup/historical-experiments.tar.gz -C /root/autodl-tmp/hfd-baseline-portable tools/run_english_training_probe.py +-``` +- +-历史报告中的脚本路径属于当时的实验记录,已添加归档提示;当前报告、指标和视频链接保留。归档 SHA256:`1fc40fe878610f4c37610d77a56aee9a5db70bfb32ed9bf7c30d25ae71fba06c`。 +- +-## 移出工作区的脚本 +- +-| 原路径 | 字节 | +-| --- | ---: | +-| `tools/analyze_omg_budget.py` | 5110 | +-| `tools/audit_asr_context.py` | 3829 | +-| `tools/audit_genmo_audio.py` | 4258 | +-| `tools/audit_genmo_inputs.py` | 3353 | +-| `tools/audit_reference_translation.py` | 3706 | +-| `tools/benchmark_omg_data.py` | 6149 | +-| `tools/benchmark_omg_io.py` | 2526 | +-| `tools/benchmark_omg_resident.py` | 4317 | +-| `tools/benchmark_omg_training.py` | 8946 | +-| `tools/benchmark_textop_batch.py` | 7413 | +-| `tools/benchmark_translation_batch.py` | 2800 | +-| `tools/check_genmo_ambiguity.py` | 1781 | +-| `tools/compare_asr_models.py` | 7332 | +-| `tools/compare_english_checkpoints.py` | 6201 | +-| `tools/diagnose_omg_validation.py` | 6358 | +-| `tools/eval_bilingual_pilot.py` | 3519 | +-| `tools/eval_reference_overfit.py` | 3580 | +-| `tools/initial_bilingual_validation.py` | 2183 | +-| `tools/prepare_bilingual_pilot.py` | 3325 | +-| `tools/prepare_omg_validation.py` | 4446 | +-| `tools/probe_asr_latency.py` | 2183 | +-| `tools/probe_bilingual_condition.py` | 4904 | +-| `tools/probe_direct_chinese.py` | 7190 | +-| `tools/probe_reference_alignment.py` | 15248 | +-| `tools/prune_chinese_caches.py` | 6225 | +-| `tools/render_bilingual_pilot.py` | 2202 | +-| `tools/run_bilingual_pilot.py` | 4289 | +-| `tools/run_english_training_probe.py` | 1537 | +-| `tools/run_omg_validation.py` | 5170 | +-| `tools/summarize_asr_context.py` | 5454 | +-| `tools/summarize_asr_throughput.py` | 6969 | +-| `tools/summarize_bilingual_pilot.py` | 8655 | +-| `tools/summarize_english_probe.py` | 7373 | +-| `tools/summarize_genmo_audio.py` | 5254 | +-| `tools/summarize_genmo_overfit.py` | 4932 | +-| `tools/summarize_omg_validation.py` | 9952 | +-| `tools/summarize_reference_overfit.py` | 9336 | +-| `tools/summarize_reference_preparation.py` | 6832 | +-| `tools/summarize_rollout_debug.py` | 4354 | +-| `tools/summarize_textop_batch.py` | 5570 | +-| `tools/verify_bilingual_pilot.py` | 3282 | +-| `tools/verify_native_batch_limit.py` | 2224 | +-| `tools/verify_semantics_clock.py` | 4106 | +diff --git a/docs/history/BASELINE_METHOD_BEFORE_SEMANTIC_FIX.md b/docs/history/BASELINE_METHOD_BEFORE_SEMANTIC_FIX.md +deleted file mode 100644 +index 0ea8331..0000000 +--- a/docs/history/BASELINE_METHOD_BEFORE_SEMANTIC_FIX.md ++++ /dev/null +@@ -1,95 +0,0 @@ +-# 最终采用的 baseline 输入与推理方式 +- +-本轮完成了用户提出的累计音频、原生推理节奏对照。两模型仍使用原生网络、损失、训练循环和官方初始化权重;没有新增 KV 模块、控制分类头、任务进度真值或动作平滑。测试样本为原生双任务 interruption session `train-002203-seed42`:接取物品未完成时输入捡起地上物品的新指令。原始声音、背景音和动作时间线保留。 +- +-## 最终选择 +- +-| 模型 | 输入 | 动作生成与衔接 | +-| --- | --- | --- | +-| OMG | 保存原始录音;模型每次编码截至 HFD 当前释放时刻的最近两秒 PCM。开头左侧补零,保留原生 60 个声学条件位置 | 每次预测原生 60 帧 / 两秒;模拟原生异步参考缓冲,剩余 24 个 30 Hz 参考帧(0.8 s)时请求新计划。计算期间继续执行旧参考,新计划跳过已耗时对应帧后接入 | +-| TextOP | 由已释放音频生成的因果 ASR;当前实验保持最新稳定 ASR 句子,文字内容连续两次一致才更新,标点变化不算新命令;继续使用原生 CLIP token 上限 | 原生 8 帧 / 0.16 s primitive,约 6.25 Hz;持续反馈预测历史,换文本时不重置姿态;原生 guidance=5 与三阶段训练 | +- +-音频到达频率、语义条件变化、动作生成频率是三个独立事件。音频仍按 HFD 0.6 秒窗口 / 0.4 秒释放间隔提供可用边界;模型可以重新编码更长的已听历史,但不能读取边界之后的录音。OMG 不强制每 0.4 秒抛弃未来计划,也不等整个任务做完才处理新指令。 +- +-OMG 缓冲参数对应原生 tracker 50 Hz 下剩余 40 帧的 0.8 秒时间,当前在 30 Hz 的参考轨迹上实现为 24 帧。两秒计划减去 0.8 秒缓冲,对应大约 1.2 秒的规划请求间隔;新计划真正接入还受计算时间影响。 +- +-## 为什么没有采用“全历史直接输入” +- +-累计录音已经保留。另一个 OMG 模型从官方初始化,用 `hfd_full_prefix` 在同一 session 上训练 1600 步:每个动作点读 `[0,当前释放截止)` 的全部 PCM,再把原生声学特征线性插值到固定 60 个条件位置。训练和推理策略一致,未来录音不会影响当前特征。 +- +-这个适配没有带来更好的结果,且随着历史增长降低时间分辨率、增加重新编码成本。OMG 原生音频编码为 35D 声学特征,不是语言语义编码器。此结果只能否定本次“全历史压缩到固定位置”的适配效果,不能概括为完整音频历史没有价值。 +- +-最终采用两秒历史,是为了保留原生特征的时间排列和一致的训练/推理接口。超过两秒的指令仍可能被截断,这是明确的 baseline 上下文限制。改成可变长语义编码器将涉及网络改造,本轮没有这样做。 +- +-TextOP 输入的是文字,不是原始 PCM。此 14.8 秒样本短于 ASR 的 30 秒历史上限,ASR 本来就在每个截止时刻重新识别从开始到当时的全部录音。真正改变的是供给 CLIP 的指令组织和稳定保持方式。`stable_latest_sentence` 是词法预处理,不使用 GT turn、事件标签或真值任务文本;它也不是可靠的语音任务边界检测器。多句复合指令、复杂犹豫或 ASR 幻觉可能使末句不代表完整任务;可用 `rolling_transcript` 做保留完整转写的对照,正式训练需在既定验证集核对这一限制。 +- +-## 结果 +- +-所有误差来自该训练 session 的连续生成,开头真值初始化帧被排除;不代表独立测试集或完整全双工评测。TextOP 指标为 23 个建模关节,OMG 为 29 个,不能直接横向排名。 +- +-| 实验 | 连续关节 RMSE(rad) | root RMSE(m) | +-| --- | --- | --- | +-| OMG 两秒历史,0.4 s 重规划(前轮) | 0.25026 | 0.04641 | +-| OMG 两秒历史,同步两秒块执行(前轮) | 0.18776 | 0.03093 | +-| OMG 两秒历史,模拟原生异步缓冲,本轮最终测时 | 0.09436 | 0.01764 | +-| OMG 全历史压缩,模拟原生异步缓冲 | 0.22073 | 0.03660 | +-| TextOP 滚动转写、原生周期、guidance=5(前轮) | 0.40908 | 0.03739 | +-| TextOP 稳定最新句、原生周期、guidance=5 | 0.30987 | 0.02933 | +- +-OMG 两秒历史模型已在前轮完成同预算 1600 步训练,本轮复用其唯一最终权重测试缓冲;累计历史模型和新 TextOP 均重新从官方初始化训练 1600 步。没有为不同执行周期挑选不同检查点。 +- +-TextOP 真实历史窗口 RMSE 为 0.00700 rad,但连续生成仍未充分拟合。相同静止历史下切换两条指令,8 帧输出差异约 0.00066 rad;相同文本下换成真实动作历史,输出差异约 0.22 / 0.47 rad。这说明此次小样本拟合仍主要依赖姿态续接,不能把稳定文本和原生周期描述为已经解决任务启动问题。 +- +-OMG 固定时延敏感性对照:异步缓冲用零计算时延时为 0.10596 rad,固定 0.2 秒为 0.08031 rad,最终实测时延为 0.09436 rad。人工固定时延仅用于诊断,最终没有选择最有利的 0.2 秒去改善指标。 +- +-## 推理耗时与时钟边界 +- +-本机 RTX 5090、FP32 原生 PyTorch,模型加载/预热后: +- +-| 输入 / 模型 | 每次规划平均耗时 | P95 | +-| --- | --- | --- | +-| OMG 两秒历史 | 0.479 s | 0.489 s | +-| OMG 累计历史 | 0.523 s | 0.555 s | +-| TextOP primitive | 0.066 s | 0.112 s | +- +-OMG 计时包含本轮声学特征、历史 FK/编码和原生生成,并在 CPU 取回预测后结束计时。TextOP 包含当前 CLIP 条件读取/编码及原生生成,不包含之前生成 ASR sidecar 的 Whisper 计算。渲染与真实机器人控制器不计入。 +- +-在当前未优化后端,OMG 本身已超过 0.4 秒预算,不能宣称 0.4 秒内完成完整推理和执行。异步缓冲允许旧参考在规划期间继续输出,最终实测无缓冲耗尽。模型输出回到缓冲时,按实测时间向上取整到 30 Hz 帧数,跳过新计划中已过去的帧;旧参考提供这段 bridge,而不是把未执行的新预测当作实际历史。 +- +-这是离线源时间参考轨迹上的执行模拟,使用真实测得的规划耗时;视频仍是原始音频内容对齐的姿态回放。它不是实时麦克风、ASR、机器人 tracker 的完整墙钟部署,不从视频直接推导论文中的硬件反应时延。HFD 音频释放截止、源动作时间、真实计算耗时分别记录在报告中。 +- +-## 配置、复现与迁移 +- +-采用的短训练配置:`configs/history/omg_adopted.yaml`、`configs/history/textop_adopted.yaml`。正式训练模板 `configs/omg.yaml`、`configs/textop.yaml` 也记录相同输入和推理策略,但独立数据划分、正式训练和论文评测尚未完成。按根目录 README 安装两个环境并准备外部权重、ASR、机器人 assets;更新 YAML 中路径、GPU 和 batch size 后训练到新的 output_dir。 +- +-```bash +-python train.py --config configs/history/omg_adopted.yaml +-python train.py --config configs/history/textop_adopted.yaml +-python infer_overfit.py --run NEW_OMG_RUN --continuous-only --task-seconds 0 14.8 +-python infer_overfit.py --run NEW_TEXTOP_RUN --continuous-only --task-seconds 0 14.8 +-``` +- +-当前已验证权重与输出: +- +-- OMG 权重:`/root/autodl-tmp/runs/hfd-omg-context-debug/omg-past2-train/last.ckpt`;推理输出 `hfd-native-context-final/omg-past2-adopted`。 +-- TextOP 权重:`/root/autodl-tmp/runs/hfd-native-context-final/textop-stable-command/ckpt_1600.pth`;推理输出该 run 的 `inference`。 +- +-复现最终 OMG 用 `--omg-replan-mode native_async --output-dir NEW_OUTPUT`,因为该保留权重的旧 resolved_config 仍记录前轮重规划设置。新 adopted 配置把这些推理设置放在 `inference` 中,无需 CLI 覆盖。 +- +-重要配置项:`data.audio_policy`、`data.caption_policy`、`inference.omg_replan_mode`、`inference.omg_buffer_remaining_frames`、`inference.textop_guidance_scale`。`--omg-latency-frames` 是诊断覆盖,正式采用实测时间,不在配置中固定有利的延迟。 +- +-全历史对照用 `configs/history/omg_native_context.yaml`,其中 `hfd_full_prefix` 不作为最终默认。缓存仅用于训练时重复窗口的确定性声学特征,单 source 最多 128 组;推理不复用未接收的未来条件。缓存与直接提取逐项一致。 +- +-## 原生依据与核验 +- +-- [TextOP 原生部署论文](https://arxiv.org/html/2602.07439v1#S3.SS5):生成器 6.25 Hz,tracker 50 Hz。 +-- [OMG 原生执行文档](https://github.com/Tsinghua-MARS-Lab/OMG/blob/main/docs/generation.md):同步与异步参考缓冲。 +-- 本地固定版 `upstream/omg/src/omg/cli/pipeline/modes.py`:旧参考 bridge、新参考跳过 bridge 时段;原生 tracker 参数剩余 40 帧 / 50 Hz。 +- +-新两个训练的优化器更新、参数变化、有限 loss/梯度和严格检查点重载均通过。数据合同检查和 37 项回归测试通过,覆盖累计音频未来不可见、缓存一致性、ASR 稳定策略的因果性、缓冲 bridge/跳过位置与 underrun 检测。全历史模型的三个等价重复检查点已清理,唯一最终权重与对照记录保留。 +- +-## 当前语义保持方案(2026-10-07) +- +-OMG 的正式 smoke 配置使用 `motion_only` 监督,并把原生 ASR 作为因果文本条件接入。每个推理时刻仍使用最近 2 秒 PCM;ASR 的最后一个已稳定句子在后续背景音阶段持续保留,因此任务描述结束后不会因为音频窗口滑走而丢失目标。ASR 文本只使用当时已经释放的转写,不能看到未来文本或真值动作。 +- +-OMG 代码保留两种 guidance:`joint_one` 是当前可复现实验默认值(单次采样,适合异步缓冲);`native` 启用原生文本/音频分支引导。后者在本机一次规划约需 1.37 秒,在 0.8 秒异步余量下会缓冲欠载,已记录为诊断结果,不能宣称实时可用。TextOP 的 3200-step 起步诊断连续误差为 0.2434 rad,窗口误差为 0.0094 rad;这说明训练链路有效,但整段闭环仍未达到理想过拟合。 +diff --git a/docs/history/DUPLEX_TASK.md b/docs/history/DUPLEX_TASK.md +deleted file mode 100644 +index 8ba0556..0000000 +--- a/docs/history/DUPLEX_TASK.md ++++ /dev/null +@@ -1,79 +0,0 @@ +-> 历史实验记录;当前入口见 [README](../../README.md)。旧运行目录已移动到 `/root/autodl-tmp/hfd-runs-archive-20261007/`。 +- +-# 两个正常任务的完整全双工 session +- +-本例选用原生 HFD `train-002203-seed42`:任务一是接取别人递来的物品(`neutral_item_take_R_001__A545`),任务二是弯腰捡起地上的物品(`neutral_item_pick_up_floor_R_002__A544`)。英文女声 `female_02_en`;保留原始录音与空调背景音。两项任务都包含实质动作,原有 listen/idle 过渡仍保留在时间线上,标签没有作为模型的额外控制输入。 +- +-这是原生 interruption session,第二条指令到达时第一项动作尚未完成。第一项动作会被中断,随后完成第二项动作;不是声称两个动作都完整执行结束。两任务之间的姿态过渡由原生 HFD 构造器处理,没有人为把两个独立视频拼在一起,也没有对预测做平滑。 +- +-## 时间线 +- +-运动时间线长 14.8 秒,50 Hz 共 740 帧。 +- +-| 时间 | 原生内容 | +-| --- | --- | +-| 0–1.2 s | 开场 idle;第一条录音实际在运动时间 0.4–2.16 s 可听见 | +-| 1.2–3.2 s | 第一条指令的 listen/姿态过渡 | +-| 3.2–6.0 s | 第一项动作;第二条录音在运动时间 5.2 s 到达,动作仍在执行 | +-| 6.0–8.0 s | 第二条指令的 listen/姿态过渡 | +-| 8.0–12.0 s | 第二项实质动作:捡起地上的物品 | +-| 12.0–14.8 s | 收尾 idle | +- +-`turn_speech_start_frames=[20,260]`。音频数组相对运动有原生 0.2 秒前置偏移:未裁剪音频中的实际发声时间为 0.6–2.36 s、5.4–7.16 s。导出的视频按此偏移对齐,所以视频中是 0.4–2.16 s、5.2–6.96 s。原生中断记录显示第一项动作剩余 86 帧,即 1.72 秒;不是由视频外观猜测。 +- +-## 训练与视频口径 +- +-两个模型从官方初始化权重开始,用各自原生训练框架在此 session 上训练 1600 步。训练目标窗口覆盖该 session 的时间范围 `[0,730)`,只跳过尾部 0.2 秒 idle;`max_windows: null`,不局限于此前八个稀疏窗口。TextOP batch size 4、原生四 primitive 链;OMG batch size 8、原生 60 帧未来动作窗口。OMG 使用完整目标区间音频,属于之前确认的离线口径;TextOP 使用原始录音的因果 ASR,不使用真值指令文本。少样本选择跳过正式训练的大规模交互配比,不改变原生构造算法。 +- +-视频覆盖整个 session,从开场连续推理到收尾。只在开头用一次真值历史初始化,随后使用预测历史,没有在第二条指令到达时重置姿态。TextOP 以原生 8 帧 primitive 为单位,可能略少于最后一个不完整 primitive;OMG 保留整段 30 Hz 时间线。前缀真值初始化单独标注,并从连续推理误差中排除。左右分别为预测和真值,视频只是关节姿态播放,不包含物理控制和实际物体交互。 +- +-视频统一为 H.264 + MP3 的 MP4,使用 `_playable.mp4` 新文件名。主入口为 `/root/autodl-tmp/runs/hfd-overfit-videos-jointfixed/README.md`。结果指标和音频验证写入 `duplex_task_summary.json`;窗口指标使用真值历史,连续整段指标使用预测历史,二者不能混为模型完全学会全双工任务。 +- +-## 迁移后复现 +- +-先按根目录 README 准备两个环境、官方权重、外部机器人 assets 与 HFD 原始 tar 分片路径,再更新 YAML 中绝对路径。依次执行(工作目录是代码文件夹): +- +-```bash +-/root/autodl-tmp/envs/omg/bin/python prepare_duplex_task.py --output /root/autodl-tmp/runs/hfd-duplex-task-overfit/sessions +-/root/autodl-tmp/envs/omg/bin/python transcribe_sessions.py /root/autodl-tmp/runs/hfd-duplex-task-overfit/sessions/train-002203-seed42.npz --device cuda:0 --language auto --output-dir /root/autodl-tmp/runs/hfd-duplex-task-overfit/asr +-/root/autodl-tmp/envs/textop/bin/python train.py --config configs/history/textop_duplex_task.yaml +-/root/autodl-tmp/envs/omg/bin/python train.py --config configs/history/omg_duplex_task.yaml +-/root/autodl-tmp/envs/textop/bin/python infer_overfit.py --run /root/autodl-tmp/runs/hfd-duplex-task-overfit/textop --continuous-only --task-seconds 0 14.8 --textop-guidance-scale 1 +-/root/autodl-tmp/envs/omg/bin/python infer_overfit.py --run /root/autodl-tmp/runs/hfd-duplex-task-overfit/omg --continuous-only --task-seconds 0 14.8 +-``` +- +-渲染每个 `RUN/inference`: +- +-```bash +-/root/autodl-tmp/envs/omg/bin/python render_overfit.py --inference RUN/inference --assets-root /root/autodl-tmp/datasets/eval_textop_omg/OMG_code_20260927/OMG/assets --gpu 0 +-``` +- +-选择脚本默认构建 512 条原始 source 的候选池,筛选指定两动作的原生 interruption session,保存 `selection.json` 和原生 metadata。输出目录必须不存在,防止混用旧样本和检查点。 +- +-## 清理检查点 +- +-`cleanup_checkpoints.py` 只清理 OMG 同一训练步、每一个 state_dict 张量完全相同的重复保存。正式初始化权重、唯一训练结果、TextOP 所需 VAE 都保留。先检查,然后执行: +- +-```bash +-/root/autodl-tmp/envs/omg/bin/python cleanup_checkpoints.py RUN +-/root/autodl-tmp/envs/omg/bin/python cleanup_checkpoints.py RUN --apply +-``` +- +-每个 run 保留 `last.ckpt`,删除等价 `pipeline.ckpt` 和 `checkpoints/*.ckpt`,记录 `checkpoint_cleanup.json` 并更新训练检查报告。不同参数或不同 global_step 会拒绝删除。可复现代码目录不包含这些大文件。 +- +-## 本次检查结果 +- +-两模型的 1600 次优化更新、参数变化、有限 loss/梯度和严格检查点重新加载均通过。23 项回归测试通过。 +- +-| 模型 | 训练窗口初始化关节 RMSE | 训练窗口训练后关节 RMSE | 整段连续关节 RMSE | 整段 root RMSE | +-| --- | --- | --- | --- | --- | +-| TextOP | 0.03169 rad | 0.00516 rad | 0.32428 rad | 0.02954 m | +-| OMG | 0.30846 rad | 0.10780 rad | 0.10088 rad | 0.01108 m | +- +-TextOP 的短窗口明显拟合,但整段预测仍未跟随两项动作。OMG 在整段中能看到接取和弯腰动作,但仍有姿态误差。这符合训练链路测试的用途,不作为两个 baseline 充分拟合或正式评测的结论。TextOP 的指标只计其建模的 23 关节,OMG 计 29 关节,不能直接据此排名。视频初始化分别占 2 帧 / 13 帧,已从整段指标中排除。 +- +-视频全部帧和音轨均完整解码通过,配音在编码前与原生 session 对应音频切片逐样本完全相等。两条视频音轨均非静音。TextOP 为 738 帧、14.76 s;OMG 为 444 帧、14.8 s。 +- +-本次共移除三个 OMG run 中的 9 份等价重复检查点,释放 9,535,716,429 bytes(约 8.88 GiB)。每个 run 的唯一最终 `last.ckpt` 都保留。清理记录汇总位于 `/root/autodl-tmp/runs/diagnostics/checkpoint_cleanup_summary.json`。 +- +-本页旧 TextOP 数值和视频使用 guidance=1。当前默认推理已补齐原生 guidance=5;复现旧结果须显式指定 `--textop-guidance-scale 1`。当前窗口输入与新检查见 [HFD_WINDOW_PROTOCOL.md](HFD_WINDOW_PROTOCOL.md)。 +diff --git a/docs/history/HFD_HUMAN_BASELINE_INPUT_PROTOCOL_RESEARCH_20261017.md b/docs/history/HFD_HUMAN_BASELINE_INPUT_PROTOCOL_RESEARCH_20261017.md +deleted file mode 100644 +index 4c560a7..0000000 +--- a/docs/history/HFD_HUMAN_BASELINE_INPUT_PROTOCOL_RESEARCH_20261017.md ++++ /dev/null +@@ -1,318 +0,0 @@ +-> 历史调研快照,归档于2026-10-21,包含2026-10-17起调研及后续历史补充。当前约定见[最新输入协议](../HFD_HUMAN_BASELINE_INPUT_PROTOCOL.md),以下旧默认与待实现状态不用于当前训练。 +- +-# HFD 人体 baseline:模型结构、因果输入与训练约定 +- +-当前输入修复:新训练配置使用 `clock_policy: microphone` 和 `text_conditioning: asr_to_english`,保留 `causal_asr_context`。按 `r_k <= t_motion+0.2s` 选择已释放 PCM/ASR,中文 ASR 经冻结翻译后交给原生文本编码器;视频完整保留麦克风开头。无新字段的历史 run 仍保留旧行为。详见 [中文语义与时钟修复](../SEMANTICS_CLOCK_FIX.md)。 +- +-调研日期:2026-10-17(Asia/Taipei)。状态:**代码与论文调研完成;以下新模型配方是待实现、待验证的建议,不是已经跑通的训练链路。** +- +-2026-10-18 实现进展:GenMo 的单条成对 session、原生训练桥接、连续推理、GMR 与视频链路已实现;短/长历史实验及限制见 [GENMO_HFD_OVERFIT.md](../GENMO_HFD_OVERFIT.md)。下文保留调研时的原始建议,不代表全部任务已验收。 +- +-2026-10-18 用户确认:五个 baseline 默认文本策略统一为 **`causal_asr_context`**;OMG/TextOp 已完成同 session 过拟合对照。此决定覆盖下文保留的历史建议。当前共同规则见 [五模型文本条件约定](../BASELINE_TEXT_CONDITION.md)。旧训练权重和结果仍按其原始 manifest 标识。 +- +-## 1. 推荐结论 +- +-先冻结同一份全双工 session 计划,生成经过验收的人体动作时间轴;再做各模型的表示转换。**统一信息可用时间,不强求模型接收相同形状的张量。** 0.6 秒是 HFD 每次新增的音频窗口,不是所有 baseline 的语义记忆长度,也不是动作预测长度。 +- +-第一版建议保留各模型的文本编码器,以与现有 TextOP/OMG 一致的因果 ASR 提供语义;用当前已知动作提供连续性。原生支持掩码条件的模型沿用该接口;没有该接口的模型必须明确标注为适配版本。不要把三者改造成同一个 Qwen 音频前端模型后仍称作原生 baseline。 +- +-| 模型 | 核实到的原生结构与条件 | 建议的 HFD 输入 | 适配性质 | +-| --- | --- | --- | --- | +-| 当前 HFD | Qwen 流式语音上下文;动作端有状态/历史和 RTC | 每次 0.6s PCM、0.4s 更新,session 内积累语义历史 | 参照系统;不等于独立短窗分类 | +-| 当前 TextOP | 文本条件的动作 primitive 生成 | 最近 30s 完整因果 ASR+原生 2 帧动作历史 | 本地已实现;50Hz,8 帧未来 | +-| 当前 OMG | 文本/声学条件的动作生成 | 同一因果 ASR+过去 2s 已释放 PCM+原生 10 帧历史 | 本地已实现;30Hz,60 帧未来 | +-| GenMo / GEM-SMPL | Transformer 扩散;T5 文本、逐帧音频/音乐/视觉条件;有 3D motion overwrite 接口 | 主实验:ASR 文本+已知动作前缀,视觉/音乐/语音分支标为缺失;另做 ASR+过去音频实验 | 有原生训练入口,但前缀训练分布需适配验证 | +-| HY-Motion 1.0 | 双流/单流 MMDiT,flow matching;Qwen3 token 特征+CLIP-L 全局特征 | ASR 文本经两个原生编码器;流式版另加显式动作前缀条件 | 没有公开的原生历史动作入口;流式版是方法扩展 | +-| Kimodo | root→body 两阶段 Transformer 扩散;LLM2Vec 文本+observed motion+mask | ASR 文本+历史姿态约束及 mask;未来无真值约束 | 原生约束接口可用;连续前缀微调和训练循环需实现 | +- +-建议先实现 **Kimodo-SMPLX 的输入转换与短样本闭环**,其次 GEM,最后 HY-Motion 的流式扩展。Kimodo 的 SMPLX 路线转换最直接,但官方提示该版本有 retargeting artifacts;这里只把它作为接入起点,不预先认定质量最好。若后续采用 SOMA checkpoint,需要单独验证人体到 SOMA 的转换,不能把当前 SMPL-X 数组换名当 SOMA。 +- +-## 2. 本次依据与版本 +- +-用户引用的两个线程已从本机保存记录核对: +- +-- `codex://threads/01a13a26-2519-7861-a188-c7adaa9667c1`:人体源动作交接与下载。 +-- `codex://threads/01a101b8-f059-72d1-90b4-9749bf524531`:OMG/TextOP 训练、third-party、I/O。 +- +-本地现行约定以 [SESSION_FORMAT.md](../../SESSION_FORMAT.md)、[README.md](../../README.md)、[SMPLX_HANDOFF.md](/root/autodl-fs/pangxinlei/hfd_data/SMPLX_HANDOFF.md) 为准,不从历史实验文档恢复旧策略。 +- +-官方源码通过 AutoDL 下载到 `/root/autodl-tmp/downloads/hfd-human-research-20261017/`,跳过 LFS 内容和子模块;未安装环境、下载模型权重或启动训练。 +- +-| 上游 | 本次固定 commit | 论文/官方说明 | +-| --- | --- | --- | +-| [NVlabs/GENMO](https://github.com/NVlabs/GENMO) | `16bebf402d8893184249ee206d957b8248cd8310` | [GENMO,ICCV 2025](https://arxiv.org/abs/2505.01425);仓库已改称 GEM | +-| [Tencent-Hunyuan/HY-Motion-1.0](https://github.com/Tencent-Hunyuan/HY-Motion-1.0) | `4e426f5a1021cbcf7f375458c37b840ee7225229` | [HY-Motion 1.0](https://arxiv.org/abs/2512.23464) | +-| [nv-tlabs/kimodo](https://github.com/nv-tlabs/kimodo) | `e1c219c99b4b94ba48be1fb6a1cbc64ecee4c954` | [Kimodo](https://arxiv.org/abs/2603.15546)、[官方文档](https://research.nvidia.com/labs/sil/projects/kimodo/docs/) | +- +-逐文件 SHA256 和源码定位另见 [调研来源清单](../HFD_HUMAN_BASELINE_SOURCES.json)。本次是静态核对,不声称已验证 checkpoint 兼容性、显存或推理速度。 +- +-## 3. 先构造正确的人体 session +- +-### 3.1 当前 human 包的边界 +- +-交接包提供 142,220 条成对 SMPL-X 源动作,而非已经验证的完整全双工人体训练集。本地看到 128 个 `seed_smplx/*.tar`,本次未重新做全量分片哈希审计。 +- +-按动作身份连接:本机实际读取的 `metadata/seed_g1_smplx_pairs.parquet` 使用 `motion_id` 列,对应 G1 metadata 的 `filename`;人体路径直接取表中的 `smplx_path`。保留原 split、八音色、背景、Stop、backchannel、hesitation、pause 等音频原料。禁止按行号、G1 帧下标或文件名字串猜对应关系。 +- +-人体源含 `trans[T,3]`、`root_orient[T,3]`、`pose_body[T,63]`、手/脸、`betas[16]` 和 `sample_times`。`sample_times` 是时间依据;源帧率并非严格 30Hz。`trans` 是 AMASS 模型平移,不能直接解释成世界 pelvis 位置。 +- +-### 3.2 两种身体共用计划,各自合成几何 +- +-应拆成 `SessionPlan → render_g1(plan) / render_human(plan)`。计划至少固定:split、motion/source ID、每段源时间映射、1–4 轮顺序、音色、背景偏移与增益、语音事件、动作开始/截断/保持/恢复、idle/listen 素材、过渡长度,以及所有随机选择。仅同 seed 或同 source index 不足以保证两版成对。 +- +-人体 renderer 采用共同 50Hz session 时间轴;根平移按秒插值,旋转用 SO(3)/SLERP。根据当前人体末位置和朝向拼接下一段,使用一致体型;不能复制机器人根高度、关节限制或保持姿态数组。idle/listen 和 Stop 过渡也必须重新构造人体几何。 +- +-验收后再把**已合成的整条人体 session**转到模型的 30Hz 时间网格。派生输入的最后一个历史采样点不得通过插值使用决策边界后的 GT 帧:在已知前缀上单独插值/边界保持;未来标签可按完整 GT 重采样。记录两者的取样规则。 +- +-### 3.3 third-party 不能自动宣称已有人体版 +- +-当前 portable 已接入 `hfd_third_party_multi_v5`,但其读取器加载的是生产包的 `[T,36] states`,即 G1 动作。新 SMPL-X 源动作不自动补出这些生产 session 的人体几何。必须恢复其完整源片段/时间映射,或从冻结计划同时渲染两版;不能从 G1 36D 直接构造 SMPL-X 真值。 +- +-若暂时无法恢复,应在所有参加该轮对比的模型上使用共同的、已验证的人体 session 子集,注明不含该部分 third-party。不要用新模型的较小数据集与既有 OMG 训练结果直接归因模型优劣。旁人说话的 PCM 应保留;数据层不能用旁人 GT 标签替模型过滤输入。 +- +-## 4. 因果边界与共同数据契约 +- +-### 4.1 三个时钟分开保存 +- +-当前 portable 对第 `k` 个语义组规定: +- +-```text +-新增音频窗口:W_k = audio[6400*k : 6400*k + 9600] # 16kHz +-音频释放时刻:r_k = 0.6 + 0.4*k # 音频轴,秒 +-原生动作起点:i = 20*k + j,j=0..19 # 50Hz 动作索引 +-该组内条件:仅使用截至 r_k 释放的音频/ASR +-``` +- +-这来自 `hfd_bridge/sessions.py::SessionWindows`。`i/50`、`r_k`、实际墙钟不是可随意替换的同一时间。接入人体系时必须保留这张映射;不要直接以 `audio[:i/50]` 替代现行约定,也不要因为 `r_k` 数值大于动作名义时间就擅自移位动作监督。原始语音与动作内容的约 0.2 秒偏移,也不能当作这里的释放延迟使用。 +- +-如果另建“麦克风物理时间零点”的真实在线评测,需明确初始预缓冲和动作轴到墙钟的映射,所有模型共用;它与复现现行 portable 是不同的时间协议。 +- +-### 4.2 共同可用信息 +- +-在一次决策处,允许输入:已释放的混合 PCM、由这些 PCM 得到的 ASR、已知当前姿态及过去已执行姿态、固定的身体/坐标定义,以及事先指定的计算窗口长度。 +- +-不允许输入:未来音频、整句 GT 指令、最终纠正后的文本、GT turn ID、listen/motion/idle 标签、未来 Stop 时刻、未来动作长度/进度、未来轨迹/落脚点、从未来 GT 动作算出的历史特征。GT 标签可用于监督、分层抽样和评估,不可变成推理时不可得的条件。 +- +-噪声化的未来动作是扩散/flow matching 的训练变量;这属于目标学习过程。问题在于把干净未来特征送入条件,或在评估时提供它们。未来生成 token 相互注意本身不违反在线输入因果性,因此不应把所有模型强制改成逐帧 causal attention。 +- +-### 4.3 建议缓存字段 +- +-```text +-session_id, split, plan_hash, recipe_hash, pcm_hash +-decision_index_50hz, semantic_group, audio_release_sample +-motion_sample_times, known_motion_end, native_fps +-asr_cache_hash, asr_policy, asr_text, asr_audio_end_sample +-body_model_id, shape_id, coordinate_transform, representation_id +-history_features, history_feature_mask, valid_history_mask +-future_target_features, valid_future_mask +-``` +- +-训练专用的目标和控制标签与输入字段隔开;ASR 缓存绑定 PCM 和识别配置,不绑定某种模型的姿态。G1/human 共用 PCM 后可以复用 ASR;人体特征缓存必须独立。 +- +-## 5. 文本条件、语义记忆与全双工限制 +- +-### 5.1 五个 baseline 默认采用完整历史 ASR +- +-复用现有 Whisper-tiny、`language=auto`,每个语义组仅识别截至 `r_k` 的最多 30 秒历史混合音频。原始 sidecar 保留中英文和背景,不翻译、不从 metadata 修正;当前模型条件另经冻结中文→英文转换。离线预计算必须逐释放点截断后识别;先识别完整 session 再按词时间戳切分会利用未来声学上下文。 +- +-用户已确认五个 baseline 统一采用 `causal_asr_context`:每次使用截至当前释放点、最近最多 30 秒音频的完整识别结果;非空立即更新,空结果沿用上一份非空文本,session 结束清空。不要把历次重识别结果不断拼接,也不要用未来识别结果回填过去。最后一句允许是不完整的当前假设,不再等待连续两次一致。 +- +-五个模型共用同一字符串,再各自通过原生 tokenizer/encoder:OMG 的 T5、TextOp 的 CLIP、GenMo 的 T5-3B、HY-Motion 的 Qwen3/CLIP、Kimodo 的 LLM2Vec。embedding 不共享,长度预算按原生配置,缓存并报告截断;不得静默换成摘要、GT 指令或“当前任务”。共同实现为 `hfd_bridge.textop_condition.caption_contexts`。 +- +-### 5.2 对照结果与历史兼容 +- +-OMG/TextOp 的相同 session 对照见 [完整历史 ASR 过拟合测试](../ASR_CONTEXT_OVERFIT.md)。TextOp 有改善但仍不稳定,OMG 没有整体改善。两种策略同时改变了历史覆盖范围和两次确认延迟,不能只归因于上下文长度;保留多个句子也不保证理解否定、指令替换或第三方话语。 +- +-`stable_latest_sentence` 和 `rolling_transcript` 只供显式历史复现/消融使用,不能混称新默认。已保存的旧权重、结果和缓存不改写。GenMo 新缓存记录策略;当前 runner 复用缓存时必须匹配请求策略,不匹配则要求重新生成。HY-Motion 和 Kimodo 的 HFD 训练适配仍待实现,但文本入口必须遵循此默认约定。 +- +-不建议第一版新增 GT 驱动的“当前任务选择器”,也不把所有 PCM 统一变成 Qwen hidden 后接新头。需要意图/说话人解析时,另立适配系统实验,单独报告其训练数据、错误和延迟。 +- +-## 6. GenMo / GEM:原生结构决定如何接入 +- +-2026-10-18:ASR+过去 2s 原生 PCM 分支已实现并完成小样本匹配实验,当前未优于 ASR-only;实际因果边界、维度、BN 诊断和结果见 [GENMO_PAST_AUDIO_TEST.md](../GENMO_PAST_AUDIO_TEST.md)。 +- +-### 6.1 已核实的能力 +- +-本次针对官方 **GEM-SMPL**,不是视频生成模型 Mochi。原生 `gem_smpl` 配方有 regression/diffusion 双模式,16 层、1024 隐藏宽度的 RoPE Transformer,文本通过 T5-3B 的 1024D token 特征做 cross attention。检查点和实际配置优先于论文中的概念表示。 +- +-公开编码器的动作输入/预测核心是 151D:身体 21×6D 旋转、10D betas、camera-view 根旋转 6D、gravity-view 根旋转 6D、根局部位移 3D。末三维虽叫 velocity,代码计算的是相邻帧位移,**没有乘 FPS**;不能直接套米/秒速度。见 [原生 encoder](https://github.com/NVlabs/GENMO/blob/16bebf402d8893184249ee206d957b8248cd8310/gem/network/endecoder.py#L122)、[位移计算](https://github.com/NVlabs/GENMO/blob/16bebf402d8893184249ee206d957b8248cd8310/gem/utils/motion_utils.py#L25)。 +- +-PCM 分支是多层卷积,输出逐帧 128D 特征,再投影进动作条件;BEAT2 loader 把音频重采样至 **18kHz**、动作 30Hz。它是逐帧对齐语音条件,不能把 HFD 16kHz 波形原样塞入并假定时长不变,也不能把 music 的 35D 特征当语音语义。见 [音频编码](https://github.com/NVlabs/GENMO/blob/16bebf402d8893184249ee206d957b8248cd8310/gem/gem.py#L175)、[BEAT2 loader](https://github.com/NVlabs/GENMO/blob/16bebf402d8893184249ee206d957b8248cd8310/gem/datasets/beat2/beat2.py)。 +- +-denoiser 有 `observed_motion_3d`、`motion_mask_3d`,可把已知动作写入 noisy motion。但这不等于已具备 HFD 的 KV/RTC。默认 `dual_mode` 的 `in_attr` 没有启用 observed-motion embedder;代码有相关分支引用,也不能仅凭该引用就认定实际 checkpoint 含对应参数。见 [overwrite 接口](https://github.com/NVlabs/GENMO/blob/16bebf402d8893184249ee206d957b8248cd8310/gem/network/gem_denoiser.py#L214)、[默认条件配置](https://github.com/NVlabs/GENMO/blob/16bebf402d8893184249ee206d957b8248cd8310/configs/pipeline/dual_mode.yaml)。 +- +-### 6.2 推荐训练构造 +- +-主实验命名 `GEM-ASR-prefix`。30Hz 下建议起始设定为 12 帧已知历史+60 帧未来;12 帧包含当前状态,所有 future 条件只使用当前决策的文本。未使用的视觉、2D、camera-motion、music、speech 条件必须通过原生 availability mask 标缺失,不从 GT 人体投影出“视觉输入”。保留固定虚拟相机/重力坐标以完成表示转换,不根据未来 GT 轨迹设置相机。 +- +-历史经原生 encoder 转成 151D,送入现成 overwrite 接口;未知未来 observed 值清零、mask=0。首版不为缺失的 observed-motion embedder 虚构权重,也不自动扩展 `in_attr`。先严格加载原始 checkpoint,再检查前缀分支和梯度;如必须增加 mask embedding,明确命名为额外结构扩展。 +- +-保留原生训练入口和双模式执行关系。其 `forward_train(mode='diffusion')` 要求 `regression_outputs`,不能直接删除 regression 调用而声称仍用原训练流程。HFD 有完整 3D 监督,无需伪造 2D-only 样本或启用无来源的相机重投影监督。采用原生噪声过程与 3D loss;将已知前缀/补齐帧从未来预测指标中排除,并审计每个损失的有效 mask,避免已知帧复制降低表观 loss。见 [训练路径](https://github.com/NVlabs/GENMO/blob/16bebf402d8893184249ee206d957b8248cd8310/gem/network/gem_diffusion.py#L73)。 +- +-### 6.3 语音增强实验 +- +-命名 `GEM-ASR-past-audio`,与主实验分开。只截过去 2 秒已释放 PCM,再重采样至 18kHz;编码前就截断,防止卷积跨边界。将过去音频的逐帧特征放在窗口的历史位置,未知未来位置设 availability=0,不能把未来动作对应的真实音频给模型。 +- +-如果仅保留 0.4 秒动作历史而希望使用整段 2 秒音频,现有逐帧对齐接口需要额外设计。第一版增强实验建议使用 **60 帧过去上下文+60 帧未来**,已知动作范围同步扩为两秒;另做相同两秒动作历史但关闭音频的对照。否则不能把收益单独归因音频。音频有效长度、卷积输出帧数、填充边界必须用真实 forward 验证,不能假设输出恰好逐点对齐。 +- +-## 7. HY-Motion:文本条件可以复用,流式历史需要扩展 +- +-### 7.1 原生能力与公开程度 +- +-原生动作表示是 SMPL-H 身体 22 关节的 **201D**:根平移 3、根旋转 6、21 个局部旋转 126、22 个局部位置 66。它不等于 HumanML3D 263D。网络结合双流/单流 Transformer,语义条件为 Qwen3-8B token 特征和 CLIP-L 全局向量,flow matching 预测速度。论文中的局部 attention 是双向局部性,不是在线语义记忆。见 [论文模型章节](https://arxiv.org/html/2512.23464v1#S3)、[文本编码器](https://github.com/Tencent-Hunyuan/HY-Motion-1.0/blob/4e426f5a1021cbcf7f375458c37b840ee7225229/hymotion/network/text_encoders/text_encoder.py)。 +- +-核查的 `generate()` 接口接收文本、seed、duration;`hidden_state_dict` 是预编码文本特征,并非历史动作状态。`x_mask_temporal` 表示有效序列位置,也不是 known-history mask。源码有 `mask_mode='causal'` 分支同样不等于已训练的流式预测器。公开仓库以推理为主,本次未发现可直接套用的完整官方训练循环。见 [生成接口](https://github.com/Tencent-Hunyuan/HY-Motion-1.0/blob/4e426f5a1021cbcf7f375458c37b840ee7225229/hymotion/pipeline/motion_diffusion.py#L477)、[forward](https://github.com/Tencent-Hunyuan/HY-Motion-1.0/blob/4e426f5a1021cbcf7f375458c37b840ee7225229/hymotion/network/hymotion_mmdit.py#L409)。 +- +-### 7.2 两种结果必须分开命名 +- +-`HY-native-window`:每次使用当前 ASR 文本、固定两秒计算窗口做文本到动作。保持原生网络,用于判断文本/动作表示是否接通;它没有历史姿态条件,不能冒充连续全双工 baseline,也不能靠 GT 对齐每个预测片段掩盖跳变。 +- +-`HY-ASR-prefix`:推荐的流式适配。采用 12 帧历史+60 帧未来、30Hz;增加显式 known-mask 和历史条件通路,例如零初始化的 mask 投影,保留原有 201D 输入投影权重。未来真值只用于 flow target。新增参数及加载差异必须写入 manifest;明确这是我们实现的扩展。 +- +-一种可实现的训练约定是:已知前缀始终为干净历史;未知部分按 `x_s=(1-s)ε+s*x_GT` 插值;网络得到 known mask 和当前 ASR,预测未知部分的 `x_GT-ε`,只在未知有效帧上计算 flow loss。采样每个 ODE 子步锁定历史,未知位置从噪声积分,结束后只执行未来段。训练/采样必须采用同一种前缀参数化。它改变了原始无前缀训练分布,需通过微调和闭环验证,不能只改推理器硬覆盖就宣称完成适配。 +- +-禁用 duration estimation 和 prompt rewrite,计算时域预先固定;不能用 GT 总动作长度或下一事件时间决定输出长度。保留双编码器,不把 Qwen 单独替换成 CLIP。可先冻结文本编码器、只训练 motion 模型及新增投影;显存/LoRA 策略在真实加载后决定。 +- +-原生解码默认包含旋转/平移平滑和按整段最低顶点落地。流式版若对整个“历史+预测”执行这些操作,会修改历史和世界根高度。需锁定已执行前缀、固定 session 地面/世界原点;后处理只作用于未执行计划,另外报告开启/关闭结果。利用自己预测的未来做后处理可以是因果规划,但不能回写已执行帧或利用 GT 未来。 +- +-## 8. Kimodo:优先复用原生约束机制 +- +-### 8.1 结构与版本选择 +- +-Kimodo 先预测 root,再把 root 转为局部表示供 body denoiser 使用。`motion_mask_mode='concat'` 时,已知值覆盖 noisy motion,特征 mask 与动作拼接进入网络;这是真正的约束接口。见 [两阶段 denoiser](https://github.com/nv-tlabs/kimodo/blob/e1c219c99b4b94ba48be1fb6a1cbc64ecee4c954/kimodo/model/twostage_denoiser.py#L73)。 +- +-文本默认是 LLM2Vec/Llama-3-8B 系列特征,不是 CLIP/T5;本次看到的配置可加载本地编码器或编码服务。训练实验应固定权重版本并本地缓存已释放文本的 embedding,服务不可静默换模型。 +- +-官方分别提供 SOMA、G1、SMPLX checkpoint。SOMA 内部 30 关节、公共输出可为 77 关节;SMPLX 为身体 22 关节。本批数据先对接固定的 `nvidia/Kimodo-SMPLX-RP-v1`,不使用会随版本变化的默认短名。官方说明 SMPLX 版本可能有较明显重定向伪影。见 [骨架说明](https://research.nvidia.com/labs/sil/projects/kimodo/docs/key_concepts/skeleton.html)。 +- +-源码的 AMASS→Kimodo 转换包含 pelvis offset、Y-up/Z-up 与额外朝向旋转;优先使用 `amass_arrays_to_kimodo_motion()` 并做往返检验,不能直接照搬 GEM 的坐标矩阵。源 16D betas 与 Kimodo 固定骨架体型也需核对;必要的形体标准化必须训练/验证一致,不能丢掉 betas 后声称几何完全相同。见 [官方转换](https://github.com/nv-tlabs/kimodo/blob/e1c219c99b4b94ba48be1fb6a1cbc64ecee4c954/kimodo/exports/smplx.py#L74)。 +- +-### 8.2 推荐输入与训练 +- +-命名 `Kimodo-SMPLX-ASR-prefix`。起始采用 12 帧历史+60 帧未来、30Hz;历史来自当前可知的姿态。优先按原生 full-body/末端约束格式构造历史,未来所有约束 mask=0;用 `first_heading_angle` 和可知根位置承接世界状态。不能输入未来全身关键帧、真实根路径、未来接触标签,即使原生 API 支持这些条件。 +- +-原生 multi-prompt 代码已经用上一段生成结果构造 transition constraints,说明无需另起一套条件网络;但原 UI/批处理可以看到全部 prompt,在线 wrapper 必须在每次释放后单步调用,不一次传入整条未来 prompt 时间线。原实现还会重新生成拼接重叠区,在线版不能修改已经执行的部分。见 [transition 构造](https://github.com/nv-tlabs/kimodo/blob/e1c219c99b4b94ba48be1fb6a1cbc64ecee4c954/kimodo/model/kimodo_model.py#L140)。 +- +-沿用 checkpoint 的扩散 schedule、两阶段网络和运动表示,新增 HFD 数据训练循环。论文目标是 clean-motion 分量的加权 smooth-L1 与 FK 一致性,不能用简单的统一 MSE 替换后称原生损失复现。连续历史前缀也不是原稀疏关键帧分布,需在微调中覆盖。历史姿态约束下的未约束辅助量仍可能需要损失;具体 mask 按语义/维度确定,不笼统把前缀全部当成未知目标。见 [论文训练目标](https://arxiv.org/html/2603.15546v1#S4.SS3)。 +- +-公开代码提供 denoiser、噪声过程与采样,不等于已提供完整训练器;优化器、EMA、loss 权重和约束采样需按论文/配置实现并标明复现范围。本轮不承诺两卡 batch 或训练时长。 +- +-### 8.3 平滑根与速度的未来泄漏 +- +-Kimodo 原生表示包含平滑根、heading、局部位置、全局旋转、速度和接触。SMPLX22 的该表示按源码组成是 `3+2+22*3+22*6+22*3+4=273D`,实际仍须以 checkpoint 的 `motion_rep_dim` 断言,不能套用 SOMA 的维数。见 [运动表示](https://github.com/nv-tlabs/kimodo/blob/e1c219c99b4b94ba48be1fb6a1cbc64ecee4c954/kimodo/motion_rep/reps/kimodo_motionrep.py)。 +- +-`get_smooth_root_pos()` 是整段轨迹平滑;速度是前向差分。**禁止对“GT 历史+GT 未来”整体编码后,把前 12 帧当输入。** 原生平滑甚至会影响历史段中更早的位置,单独屏蔽最后一帧速度不足以消除泄漏。见 [平滑根](https://github.com/nv-tlabs/kimodo/blob/e1c219c99b4b94ba48be1fb6a1cbc64ecee4c954/kimodo/motion_rep/smooth_root.py#L202)、[速度计算](https://github.com/nv-tlabs/kimodo/blob/e1c219c99b4b94ba48be1fb6a1cbc64ecee4c954/kimodo/motion_rep/feature_utils.py#L36)。 +- +-建议从已知前缀独立构造 full-body/end-effector constraints,而不是锁定整段 GT 编码的所有 273D 特征。若使用完整历史特征,只能在过去前缀上编码,边界速度用最近已知位移延拓;推理可以保留前次已生成计划的特征,但必须记录来源为模型预测。未来目标可以离线整体编码;输入中的平滑根与目标平滑根可能不同,需通过可观测姿态约束而非强求两者逐值相等来处理。 +- +-## 9. 新模型共同的窗口与 rollout 建议 +- +-以下是起始实验参数,不是三篇论文的原始训练配置: +- +-| 项目 | 第一版建议 | +-| --- | --- | +-| 新人体模型动作 FPS | 30Hz,严格按时间重采样 | +-| 动作历史 | 0.4s / 12 帧,包含当前状态;不足时重复首姿态并记录有效范围 | +-| 未来计划 | 2s / 60 帧;不依赖 GT 剩余动作长度 | +-| 名义条件更新/执行 | 0.4s / 12 帧后可重规划;每次仅用当时已释放条件 | +-| 语义条件 | 五者共享 `causal_asr_context`;旧稳定句仅显式历史对照 | +-| 训练覆盖 | `full_session`,包含 listen、idle、保持、过渡及事件前后 | +-| 外部控制 | 第一版不添加 GT stop gate、事件识别头或手工姿态重置 | +- +-预测两秒不表示必须播放完两秒。新语义组到来后可重规划,避免 Stop 要等完整计划结束。理想零计算延迟回放和实测延迟回放分别报告;如果规划/ASR 耗时超过 0.4 秒,不能声称达到该响应频率。 +- +-训练窗口在 session 内按合法起点取样,不能跨 session 连接;未来段即使覆盖下一条指令,也只提供当前可用文本。未来事件尚不可知时,监督天然存在多解,应保留并报告这种不确定性,不以未来 ASR 消除它。可增加只监督近期执行段的消融,但它改变训练损失,不能混为原生方法。 +- +-第一阶段使用 GT 过去姿态做 teacher-forcing,未来按模型原目标训练;正式评估必须用自身实际执行结果作为下一步历史,只允许 session 初始姿态来自 GT,不能每组重新喂 GT。后续可加入 rollout/noisy-history 微调,单独标注训练配方,不把它当作已经实现的能力。 +- +-统计归一化先沿用官方 checkpoint 的对应统计以保留初始化语义;人体分布若明显失配,再设计重估和权重/输入变换的成套方案。统计只能来自 train,不能把 SMPL-X 原始轴角、G1 36D、TextOP/OMG 的统计互相套用。 +- +-## 10. 实现前后的验收与比较口径 +- +-1. **人体 session**:逐任务覆盖 1–4 轮;与 G1 版比较计划、PCM 哈希、事件边界、split。人体单独检查旋转合法、无 NaN、拼接连续、固定体型与带声音视频。当前尚未完成。 +-2. **表示往返**:已知姿态→原生特征→解码,检查左右手、根朝向、位移单位、pelvis offset、帧率与固定形体;预先记录转换误差。所有模型按共同身体 22 关节评估,手指不纳入主体指标。 +-3. **因果反事实**:固定当前 PCM/历史,仅替换释放点后的音频,当前输入必须不变;仅替换当前姿态后的 GT 动作,历史特征/文本/约束必须不变;修改未来事件元数据也不得改变当前输入。 +-4. **原生模型接通**:严格加载官方 checkpoint,核对特征维数与实际前向;检查 loss、有效 mask、有限梯度、参数更新与 checkpoint 恢复。HY 新增参数单独列出;Kimodo 新训练器说明 loss 复现范围。 +-5. **短样本闭环**:包含普通指令、Stop、backchannel、hesitation、pause、third-party 的已验证人体样例;比较 teacher-forcing 与自主 rollout,检查“听完才动”、停止延迟、被打断后保持和下一指令切换。 +-6. **正式比较**:固定 session manifest、共同数据范围、ASR 版本、初始状态和随机种子;报告源动作/事件覆盖、训练更新数及有效目标帧,不强求不同架构拥有相同 batch/epoch。固定验证集,测试集不参与调参。 +-7. **结果**:同时报告动作质量、事件响应、语义错误、边界跳变、长时漂移、ASR 稳定等待、模型规划延迟和执行缓冲。文本输入不能完全覆盖声学/说话人信息,ASR 系统与原始音频 HFD 属于不同输入管线,表格应标明。 +- +-上游预训练数据与 HFD 源动作也要核查重叠。Kimodo 当前 README 明确说明部分旧 SOMA checkpoint 存在 BONES-SEED benchmark overlap,v1.2 修正;这不自动证明 SMPLX-v1 对 HFD 测试完全独立。无可核实的预训练划分时,写明“下游 session/source 隔离,预训练重叠未排除”,不能宣称完全未见动作泛化。 +- +-本说明不改变现行 OMG/TextOP 配置。下一步需要实现的是 session 计划冻结与 human renderer,再是三个 adapter;不是直接把新源动作交给既有训练命令。 +- +-## 11. 2026-10-18:官方与社区训练脚本综合核查 +- +-本节回答“是否有可用训练脚本”,不代表已经接通 HFD 训练。核查了官方当前源码树、官方训练代码 issue、定向搜索得到的社区仓库及实际训练代码;没有穷举所有 fork。固定版本、文件哈希与小规模 probe 的记录见 `reports/human_training_scripts_20261018.json`。选取的上游文本源码快照保存在 `/root/autodl-fs/pangxinlei/downloads/hfd-training-script-research-20261018/`,未替换三个原有模型仓库。 +- +-### 11.1 结论与选择 +- +-| 模型/候选 | 实际公开内容 | 能否用于当前 HFD 逐帧监督微调 | 验证程度 | +-| --- | --- | --- | --- | +-| GenMo/GEM 官方 | Lightning/Hydra 训练器、数据模块、双模式 diffusion/regression、损失、优化器与 checkpoint | **优先复用官方框架**,补 HFD 数据/条件 bridge 和专用配置 | 源码核对;现有环境及小型原生网络反向已通过;未跑真实 HFD 官方训练步 | +-| HY-Motion 官方 | 201D 模型、CLIP/Qwen 文本编码、采样、评测 | **没有官方训练入口**,监督 flow-matching 循环仍需补 | 官方源码与作者 issue 回复核对 | +-| HY-Motion:MotionRFT | 对 Full/Lite 的 LoRA、可微采样、奖励模型与微调循环 | **可参考训练组件,但不是逐帧监督训练的直接替代品** | 静态核对训练/梯度代码;未运行其完整奖励训练 | +-| Kimodo:Calix-L 后训练项目 | 特征数据集、文本缓存、训练步、优化器、梯度累积、checkpoint、七项结构化损失 | **适合作为轻量监督训练起点**;默认 SOMA 配置和离线约束采样必须替换 | 七项损失及其原样 train_step 已对 tiny 官方 SMPLX22 两阶段网络做反向 probe | +-| Kimodo:windydauy 的 scq 分支 | 完整 G1 微调目录、DDP/AMP、EMA、AdamAtan2、七项损失;另有蒸馏工作流 | **可复用训练设施**,G1 数据适配与损失须重审;不直接采用改变层数的蒸馏版本 | 损失模块在当前 SMPLX22 表示上反向通过;完整训练循环仅静态核对 | +- +-“可复用”指可以基于现有代码补齐当前任务,**不是修改数据路径即可开训**。共同缺口仍是第 4 节的冻结 SessionPlan 与 human renderer,以及因果 ASR/姿态历史、闭环推理和第 10 节验收。 +- +-### 11.2 GenMo/GEM:官方训练框架最完整 +- +-来源:[NVlabs/GENMO](https://github.com/NVlabs/GENMO),本地与远端核对版本均为 `16bebf402d8893184249ee206d957b8248cd8310`。这里是 NVIDIA 人体动作项目,不能误用同名 Genmo/Mochi 视频生成训练器,也不把 GEM-X 的另一套网络混入当前 baseline。 +- +-主要入口: +- +-- `scripts/train.py`:实例化模型和 datamodule,调用 Lightning `Trainer.fit`,支持断点恢复、日志和验证。 +-- `gem/gem.py`:`training_step`、`configure_optimizers`,原生 regression/diffusion 的执行次序。 +-- `gem/pipeline/gem_pipeline.py`、`gem/network/gem_diffusion.py`:扩散训练及几何损失。 +-- `configs/exp/gem_smpl.yaml`:包含文本/音频生成的完整模型配方;`gem_smpl_regression` 是回归配方,不能拿它代替文生动作训练。 +- +-官方命令 `python scripts/train.py exp=gem_smpl` 是框架入口示例;原配方会读取多套原始训练/评测集,并非 HFD 命令。默认 exp 引用缺失的 mixed,需显式指定 exp。模型构造还依赖主体模型、归一化与辅助资产,CLI --help 成功不证明这些运行资产已完整准备。 +- +-接 HFD 时应保留官方模型、噪声过程和合适的 3D 损失,接入 HFD datamodule 与原生条件 availability/overwrite。不要给模型伪造未来视觉观测;世界/重力/虚拟相机变换、body shape、语音特征与文本缓存仍须按第 6 节核对。diffusion 路径依赖 regression 输出,首轮不直接裁掉 regression 调用。输入历史的速度边界和未来 mask 需通过反事实检查。 +- +-### 11.3 HY-Motion:已有社区训练,但主要是奖励微调 +- +-官方 [issue #2](https://github.com/Tencent-Hunyuan/HY-Motion-1.0/issues/2) 中,协作者于 2025-12-31 明确回复没有开源该项目训练代码的计划。当前官方版本 `4e426f5a1021cbcf7f375458c37b840ee7225229` 的源码树也未找到训练入口。README 的预训练→高质量微调→RL 是方法说明,不能当成已发布训练器。网上部分教程给出 `train.py`/`configs/finetune.yaml` 却没有对应仓库文件,不作为可执行证据。 +- +-找到的主要可用社区实现是 [Xiaofeng-Tan/MotionRFT](https://github.com/Xiaofeng-Tan/MotionRFT),固定版本 `84543efec6bf8c0d819bfa75665f5c820c35daae`: +- +-- `RFT_HY/motionrft_hy.py` 包含 `LoRALayer`、`apply_lora_to_model`、`DifferentiableFlowMatchingSamplerV2`、训练器和 checkpoint;支持 Full/Lite 配置。 +-- 它先无梯度采样到随机时间,再通过一个可微末步产生动作,使用 `loss = -combined_reward` 更新 LoRA;不是标准配对动作的 flow-matching 速度场监督。 +-- 模型采样仍使用原生 **201D**。奖励/匹配模型使用的 **135D** 是取根平移与旋转部分的另一条数据通路,不可把 Mean_135/Std_135 或 HumanML263 统计替代 HY 的 201D 模型统计。 +-- 依赖 HumanML3D 风格数据、SPM/Sentence-T5 与可选 MotionReward checkpoint。代码即使选择 MotionReward 也会构造 SPM;不能假定仅准备 HY 权重即可启动。改为 HFD ASR/全 session 还需修改数据、目标和时间条件。 +- +-建议把它作为 LoRA 注入、可微网络调用和保存恢复的参考;第一版 HFD 过拟合仍采用原生 201D 与配对 GT 的监督 flow-matching 目标,不额外引入奖励模型作为第一阶段前提。若采用 MotionRFT 的奖励目标,应独立标为后训练实验,并评估事件响应是否被全片段奖励掩盖。原生 HY 接口没有经过训练的姿态历史约束,MotionRFT 也没有直接补齐我们的因果前缀接口。 +- +-另有 [FlowHMR](https://github.com/flowhmr/flowhmr/blob/main/docs/training.md) 复用 HY MMDiT backbone,提供 `train_v2m.py` 的监督训练与 GRPO;但它是视频到动作任务,数据要求相机、bbox 和 SAM 3D Body 特征,不能将其训练器/权重原样称为 HY 文本动作微调。另查了 5 个 HY mirror/streaming 候选,未发现额外可直接用的监督训练器;具体结果见快照 `hy-other-candidates.json`。 +- +-### 11.4 Kimodo:社区监督训练确实可以复用,但配方有差别 +- +-官方当前版本 `e1c219c99b4b94ba48be1fb6a1cbc64ecee4c954` 仍主要提供推理、约束与评测。[训练代码请求 #25](https://github.com/nv-tlabs/kimodo/issues/25) 的社区回复只描述尝试微调,没有给出官方训练发行版。 +- +-**候选 A:Calix-L/Kimodo-SOMA-RP-final-showcase**,固定版本 `2ece5366f8cee2a54999763f372c19bf3e20f5e2`。 +- +-- [train/train.py](https://github.com/Calix-L/Kimodo-SOMA-RP-final-showcase/blob/main/train/train.py) 包含完整的加噪→原生 denoiser→loss→backward→optimizer→保存循环,不只是展示网页。 +-- `train/kimodo_seed_repro/{data,kimodo_io,train_utils}.py` 提供特征读取、缓存文本、加载与恢复、AdamW/AdamAtan2、masked MSE/结构化 Smooth-L1;还有 `scripts/train_with_val.py`。 +-- 2026-10-18 本地实测:使用当前 Kimodo **SMPLX22 / 273D 官方统计**、tiny 随机两阶段 denoiser、synthetic cached text,原样调用其 `train_step`,打开关键帧条件与七项结构化损失;root 和 body 两分支均得到有限、非零梯度。另单独验证其 loss 反向。 +-- 该 probe 只验证接口和基本梯度流,不加载 full-size 预训练权重,不使用真实 HFD,不验证收敛、混合精度、checkpoint 恢复或官方论文复现。 +- +-必须修改/复核的点:现有 YAML 默认 SOMA、150 帧离线裁剪;关键帧可从完整 GT 有效区间随机抽取,并包含整体编码的 smooth root,不能直接用于 HFD 因果历史。`first_heading_angle` 固定零需与规范化策略一致。缓存文本路径下,`text_dropout` 仅修改文本字符串,已有缓存向量不会因此置空,使用缓存 CFG 时必须单独修正。基本 train.py 未见 EMA 更新;若声称沿用论文配方需要补齐。FK loss 当前比较 `FK(pred)` 与 `FK(target)`,没有直接使用目标的 position feature;应与论文的 FK→目标位置定义核对,不能只因七个权重相同就称一致。 +- +-**候选 B:[windydauy/kimodo_my,scq 分支](https://github.com/windydauy/kimodo_my/tree/scq/kimodo/training)**,固定版本 `59e12423f4a313aba8339f065bba7b326e7e7db2`。 +- +-其 `kimodo/training/train.py`、`dataset.py`、`loss.py`、`optimizers.py`、`ema.py` 已形成训练工作流,包含 DDP/AMP、断点、CFG dropout、EMA、AdamAtan2 和关键帧课程;比 Calix 的最小循环具备更多训练设施。七项 loss 在当前 SMPLX22 表示上的 CPU 合成反向 probe 通过,但完整循环还未执行。 +- +-它默认是 G1 CSV/机器人配置,不能直接读取 SMPL-X;七项 loss 实现使用 **普通 L1**,与当前约定的 Smooth-L1 不完全一致。默认 loss 权重与 YAML 指定权重也不同,运行必须显式锁定。其 `distillation` 目录另有 16→8 层、100→20 步等学生训练,改变了架构/目标;基础 human baseline 不直接采用。对应官方 [PR #44](https://github.com/nv-tlabs/kimodo/pull/44) 已关闭且未合入,不能标注官方训练代码。 +- +-其他容易误判的线索:`LavaTap/qwen-kimodo-train-fork` 的 `train_text_projection.py` 训练的是 Qwen→LLM2Vec 投影,不是动作模型。Blender/ComfyUI 接入器主要是推理或 retarget。近期 [Controllable-Motion-Exaggeration](https://github.com/AHHHZ975/Controllable-Motion-Exaggeration) 论文有训练适配,但当前仓库仅 README、标记 coming soon,暂没有可拿来运行的训练器。 +- +-### 11.5 对当前任务的执行建议 +- +-1. **GenMo**:保留官方 Lightning/Hydra,写 HFD bridge 与单独训练配方;不另造通用训练框架。 +-2. **Kimodo**:按用户既定的优先顺序,先用 Calix 的已验证训练步作为参考,沿用官方两阶段网络/扩散/统计,吸收另一项目的 EMA/DDP 设施;先统一损失、缓存 CFG、初始朝向与仅过去约束。所有复用的第三方代码保留来源与许可,不覆盖官方原文件。 +-3. **HY-Motion**:复用官方网络和编码器,参考 MotionRFT 的可训练调用/LoRA,补最小监督 flow-matching 循环;不要把奖励微调当过拟合测试,不能声称用了不存在的官方训练框架。 +-4. 三者接同一个冻结的 HFD session plan,分别编码为原生表示;训练和评测独立保留 teacher-forced / predicted-history。评测统一走已有 SMPL-X→G1 链路和 human GT retarget 对照。 +- +-本轮只完成调研、源码快照与上述合成 probe,没有启动正式训练或修改三者的实际训练配方。 +diff --git a/docs/history/HFD_WINDOW_PROTOCOL.md b/docs/history/HFD_WINDOW_PROTOCOL.md +deleted file mode 100644 +index 61916a7..0000000 +--- a/docs/history/HFD_WINDOW_PROTOCOL.md ++++ /dev/null +@@ -1,68 +0,0 @@ +-> 历史实验记录;当前入口见 [README](../../README.md)。旧运行目录已移动到 `/root/autodl-tmp/hfd-runs-archive-20261007/`。 +- +-# HFD 当前音频窗口口径与 TextOP 连续推理排查 +- +-旧 `omg_duplex_session_playable.mp4` 使用此前确认的 `offline_target_aligned`:动作目标未来 2 秒与这 2 秒对应音频一起输入模型。它可用于离线训练链路检查,不能据此宣称实时全双工能力;连续关节 RMSE 为 0.10088 rad,并非完美拟合。旧结果保留为离线对照。 +- +-## 新 OMG 输入时序 +- +-`data.audio_policy: hfd_current_window`。HFD 原始动作时钟为 50 Hz,每 20 个动作点使用一个 0.6 秒音频窗口;窗口每 0.4 秒更新: +- +-- 原始动作帧 `u` 的语义组 `k = u // 20`。 +-- 当前 PCM 范围 `[0.4*k, 0.4*k+0.6)` 秒,16 kHz 时为 `[6400*k, 6400*k+9600)`。 +-- 音频可用截止为 `0.6+0.4*k`。这沿用 HFD 释放时钟,与运动/音频内容偏移 0.2 秒不同,也不表示零延迟在线系统。 +-- OMG 保留原生 30 Hz 动作格式;状态帧 `t` 对应组 `k=floor(t/12)`。训练在每个可用 30 Hz 动作点取样,同组复用同一窗口。 +- +-OMG 当前音频接口需要每个预测位置一个 35D 声学特征,不接收原始 PCM,也没有 Qwen 的语言语义 KV 缓存。转换器只对当前 9600 个样本调用原生 `current35` 提取器,再将得到的特征序列线性插值到原生 60 个条件位置。**这 60 个位置都编码同一个当前窗口,不能解释为读取了未来 2 秒音频。** 包括峰值归一化在内的全部特征运算仅访问该窗口。没有额外的语义缓存前向,原生声学 embedding 和扩散采样仍然执行。 +- +-训练仍用原生 Lightning、125D 动作 codec、60 帧未来目标和 10 帧状态历史。未来动作是真值监督,未来音频不是输入。修改当前可用窗口以外的任何 PCM 不会改变输入;测试与数据合同检查覆盖此约束。 +- +-## 连续推理 +- +-只在开头使用一次真值历史。每轮用当前 0.6 秒音频窗口预测原生 60 帧(2 秒),执行前 12 帧(0.4 秒),再用已执行的预测姿态更新历史,读取新窗口并重新规划。未执行的预测丢弃,不能拿未执行的未来预测初始化下一轮。 +- +-不能把一整段 session 的各个未来音频窗口一次性传给 `generate(num_frames=整段)`:这会让当前规划提前接触之后才可用的条件。新推理代码逐轮调用原生生成器,并在报告中保存每轮 PCM 起止位置、状态帧和执行帧数。 +- +-这是沿用 HFD 音频释放规则的 OMG 适配,保留 OMG 原生预测长度;不是宣称 OMG 拥有 HFD 的语义记忆。当前窗口可能只包含长指令的一部分,缺少 KV 历史确实可能降低表现。音频在正式训练中也保留中文和背景音,不根据真值指令/任务边界提取音频。 +- +-## TextOP 为什么几乎不动 +- +-旧小样本配置 `stages: [1600]` 只进入原生 stage 0。虽然 `use_rollout: true`,`DARManager.should_rollout()` 在 stage 0 始终返回 False,所以训练全程使用每个 primitive 的真实历史。整段视频则只在开头给一次真实历史,随后反馈模型预测,训练和这种验证存在历史分布差异。 +- +-旧 TextOP 训练窗口关节 RMSE 为 0.00516 rad,整段连续 RMSE 为 0.32428 rad。两者不矛盾:从真实动作姿态继续生成 8 帧比从静止姿态独立启动并持续任务容易。文本在 listen/执行/收尾阶段也可能保持相同,而 TextOP 没有额外的真值控制标签或任务进度输入,容易停留在接近静止的历史状态。这个现象不能仅凭一条 loss 断定是映射问题;映射沿用已独立验证的关节名称转换。 +- +-本次保留原生训练循环与损失,用 `stages: [400,800,400]` 进入原生反馈预测历史的训练阶段。原生正式训练模板本来就有三阶段;此次是修正小样本实验只运行首阶段的局限。新配置为 `configs/history/textop_duplex_causal.yaml`,新 OMG 配置为 `configs/history/omg_duplex_causal.yaml`。两者均从官方初始化重训,没有复用旧离线 OMG 权重。 +- +-## 复现 +- +-已有原生 `train-002203-seed42` 与完整因果 ASR 后,先更新配置中的环境、session、权重和输出路径,再执行: +- +-```bash +-/root/autodl-tmp/envs/textop/bin/python train.py --config configs/history/textop_duplex_causal.yaml +-/root/autodl-tmp/envs/omg/bin/python train.py --config configs/history/omg_duplex_causal.yaml +-/root/autodl-tmp/envs/textop/bin/python infer_overfit.py --run /root/autodl-tmp/runs/hfd-duplex-causal/textop --continuous-only --task-seconds 0 14.8 +-/root/autodl-tmp/envs/omg/bin/python infer_overfit.py --run /root/autodl-tmp/runs/hfd-duplex-causal/omg --continuous-only --task-seconds 0 14.8 +-``` +- +-用原有 `render_overfit.py` 渲染两目录的 inference,默认 H.264/MP3 MP4。本页对应的旧 OMG 实验使用 `hfd_current_window`;后续正式模板已统一为 `hfd_past_2s`,详见 [OMG_CONTEXT_DEBUG.md](OMG_CONTEXT_DEBUG.md);旧离线 smoke/duplex 配置保留用于复现旧结果。 +- +-## 推理设置核验与本次结果 +- +-进一步对照原生 TextOP 可视化发现,旧导出使用普通条件采样(相当于 CFG guidance=1),原生 `robotmdar/config/vis_dar.yaml` 默认 guidance=5。当前 `infer_overfit.py` 已使用原生 `ClassifierFreeWrapper` 并默认 guidance=5;可用配置 `inference.textop_guidance_scale` 或 CLI `--textop-guidance-scale` 指定。旧视频/诊断的 guidance=1 设置明确保留为对照,不与新设置的误差直接混用。 +- +-同一静止历史、同一噪声种子,替换实际已释放的两条 ASR 条件:guidance=1 的输出差异约 0.00014 rad;guidance=5 约 0.00069 rad。guidance=5 下,保持文本不变而换成真实动作历史时,输出差异约 0.22 / 0.47 rad,真实历史窗口的误差约 0.00575 / 0.01026 rad。说明此次小样本拟合主要依赖动作历史续接,文本对从静止姿态启动的作用仍很弱。 +- +-三阶段加原生 guidance=5 后,TextOP 整段仍未拟合,不能将补齐这些设置描述为已解决几乎不动。旧单阶段配置是实验局限,guidance=1 是与原生默认可视化不同的设置;二者都不是静止现象的充分解释。诊断只针对本次两个任务和小样本预算,不断言整个 TextOP 架构没有能力。 +- +-OMG 当前窗口训练后窗口 RMSE 为 0.12468 rad,整段预测为 0.30954 rad;旧离线整段为 0.10088 rad。新旧同时改变音频范围、窗口取样密度与重新规划方式,这一对比不能单独量化未来音频的因果贡献。新口径不会向当前预测提供之后才释放的 PCM,而缺少语义缓存仍可能限制长指令处理。 +- +-两个 1600 步原生训练链路、严格权重重载与两个数据合同检查通过,27 项回归测试通过。新视频均保留原始音轨,音频和全部视频帧完整解码通过。OMG 推理每轮的窗口边界和执行历史记录在 `hfd_window_summary.json`。新 OMG run 的三份等价重复检查点已经清理,释放约 2.96 GiB,最终 `last.ckpt` 保留。 +- +-新视频入口: +- +-- `/root/autodl-tmp/runs/hfd-overfit-videos-jointfixed/textop_duplex_rollout_playable.mp4`:三阶段、原生 guidance=5、已释放 ASR。 +-- `/root/autodl-tmp/runs/hfd-overfit-videos-jointfixed/omg_duplex_hfd_window_playable.mp4`:当前 0.6 秒音频、0.4 秒重新规划。 +- +-诊断可复现:`python diagnose_textop.py --run RUN --frames 242 490 --guidance-scale 5`。这只比较同种子下原生 8 帧生成结果,不使用真值任务文本;真值姿态仅用于明确标注的历史对照。 +- +-最终原生 guidance=5 的 TextOP 训练窗口 RMSE 为 0.00687 rad,整段为 0.40908 rad(root 0.03739 m)。这再次表明短窗口拟合不能证明整段任务能力;新 OMG 整段为 0.30954 rad(root 0.06111 m)。两模型的关节维度不同,仍不能据此排名。 +diff --git a/docs/history/NORMAL_TASK.md b/docs/history/NORMAL_TASK.md +deleted file mode 100644 +index 90a97d6..0000000 +--- a/docs/history/NORMAL_TASK.md ++++ /dev/null +@@ -1,73 +0,0 @@ +-> 历史实验记录;当前入口见 [README](../../README.md)。旧运行目录已移动到 `/root/autodl-tmp/hfd-runs-archive-20261007/`。 +- +-# 正常单任务的小样本检查 +- +-推荐观看当前视频目录中的 `textop_complete_task_playable.mp4` 和 `omg_complete_task_playable.mp4`。这是连续播放的完整单任务,不是 `*_fitted_windows_playable.mp4` 的窗口拼接。 +- +-## 样本与任务 +- +-原始样本为 `sessions/train-000094-seed42.npz`,由 HFD 原生 session 构造器生成。元数据确认 `interaction_mode=turn-taking`、`turn_count=1`,没有打断、停止或后续任务。动作名是 `grab_walk_ff_180_002__A552_M`,即向前走并作抓取动作。英文女声为 `female_02_en`,ASR 识别为 “Character walks forward while grabbing something.”。没有额外场景物体或抓取物理模拟,播放的是机器人参考动作。 +- +-完整 session 长 17.6 秒: +- +-| 动作时间 | 内容 | +-| --- | --- | +-| 0.00–1.20 s | 初始待机;指令开始 | +-| 1.20–4.40 s | 原生听指令阶段与过渡 | +-| 4.40–16.00 s | 连续向前走并抓取,共 11.6 秒 | +-| 16.00–17.60 s | 动作结束后的待机 | +- +-原始麦克风语音约在 0.60–3.32 秒;与动作时钟相差 0.2 秒,按适配器既有时钟关系播放。数据中的听指令姿态也保留,但两个模型不接收 `listen/motion/idle` 控制标签。初始真实历史用于模型启动并在画面中标注;之后自由反馈预测。初始化画面不计入连续推理误差。 +- +-## 训练与导出 +- +-两模型从各自官方初始化权重重新训练 800 步,使用原生 DAR / Lightning 框架。配置选择整个 session 的 `[0,870)` 原始 50 Hz 帧,并保留这一连续区间内所有合法窗口,`max_windows=null`;不再均匀抽八个分散窗口。实际视频结尾根据模型帧网格取整到约 17.3–17.4 秒,只省略末尾少量待机。 +- +-TextOP 使用这个 session 新生成的逐释放窗口 ASR;OMG 使用完整离线目标段音频。音频、原始轨迹和关节映射保持不变,没有平滑、改速度、补帧或拼接其他任务来改善观感。 +- +-`textop-single/` 和 `omg-single/` 保存有效检查点、配置、指标与推理结果;`sessions/` 保存原始单轮任务及选择依据;`asr/` 保存新生成的完整转写;`summary.json` 保存本次训练和推理指标。视频和任务说明在 `/root/autodl-tmp/runs/hfd-overfit-videos-jointfixed`。 +- +-窗口拟合和自由连续推理分别报告。800 步只是训练链路的小样本检查,不保证完整任务的自由推理已拟合;模型偏离真值时不能当作视频拼接问题。右侧真值用来检查正常任务的实际动作连续性。 +- +-## 之前的视频为什么跳跃 +- +-旧 `textop_fitted_windows_playable.mp4` 拼接八段,每段仅 0.64 秒,总长 5.12 秒。八段中约 82% 的帧标为 motion、约 18% 为 idle,没有 listen 帧。但采样覆盖很偏:第一段是初始待机,其余几乎来自另一条长 session,没有覆盖原先向前走抓取任务的完整执行。训练窗口从不同时间点直接连接,必然出现时间与姿态跳变,不能当作完整任务视频。 +- +-新样本选择、训练和导出可以复现: +- +-```bash +-# 所有 output 必须指向未使用的新目录;选择仍使用原生构造,仅跳过小样本不适用的大规模最终配比。 +-python prepare_normal_task.py --output /data/normal-task/sessions --motion grab_walk_ff --limit-sources 128 +-# 转写选中的 NPZ,并按新路径修改 configs/*_normal_task.yaml。 +-python transcribe_sessions.py /data/normal-task/sessions/sample.npz --device cuda:0 --output-dir /data/normal-task/asr +-python train.py --config configs/history/textop_normal_task.yaml +-python train.py --config configs/history/omg_normal_task.yaml +-python infer_overfit.py --run /data/normal-task/textop --continuous-only --task-seconds 0 17.4 --textop-guidance-scale 1 +-python infer_overfit.py --run /data/normal-task/omg --continuous-only --task-seconds 0 17.4 +-python render_overfit.py --inference /data/normal-task/textop/inference --assets-root /data/OMG/assets --gpu 0 +-python render_overfit.py --inference /data/normal-task/omg/inference --assets-root /data/OMG/assets --gpu 1 +-``` +- +-`data.frame_range` 是原始 50 Hz 的 `[start,end)` 采样边界,仅允许 smoke 模式、单个 session,用于明确小样本检查范围;正式训练仍走完整数据配置。相机跟随使用机器人身体尺寸计算距离,避免走得远时机器人越画越小;这只影响视角,不改变原始世界坐标和误差。 +- +-本次 800 步结果(rad / m): +- +-| 模型 | 训练窗口关节 RMSE:初始 → 训练后 | 完整自由推理关节 RMSE | 完整自由推理根位置 RMSE | +-| --- | --- | --- | --- | +-| TEXTOP | 0.0448 → 0.0089 | 0.3164 | 2.5040 | +-| OMG | 0.2926 → 0.2252 | 0.2522 | 0.8357 | +- +-两模型训练检查均通过,但这些连续误差说明完整任务自由推理尚未充分拟合,不能作为已成功复现任务的结论。 +- +-## 当前视频格式与 VS Code 播放 +- +-当前交付统一为 MP4 容器、H.264 视频、48 kHz 双声道 MP3 音轨;文件名以 `_playable.mp4` 结尾。根据官方文档选择 H.264、MP4 与 MP3,避开不支持的 AAC。此前 VP8 WebM 在用户当前预览中仍无法打开,因此已清理。 +- +-官方支持清单:https://github.com/microsoft/vscode/issues/180780 +-官方 Webview 说明:https://code.visualstudio.com/api/extension-guides/webview#supported-media-formats +- +-所有视频检查过完整声画解码、非静音音轨、帧数和帧率;H.264 码流与原始视频的 SHA256 一致,转换只更换音频编码和封装,没有改动作。检查记录见视频目录的 `playable_mp4_check.json`。远程服务器上的解码检查不等于本地 VS Code 实测,首次打开请使用新文件名。 +- +-后续 `render_overfit.py` 与 `remux_video_audio.py` 默认导出 H.264/MP3 MP4,无需 WebM 或额外的 `--vscode-preview` 参数。 +- +-本页旧 TextOP 数值和视频使用 guidance=1。当前默认推理已补齐原生 guidance=5;复现旧结果须显式指定 `--textop-guidance-scale 1`。当前窗口输入与新检查见 [HFD_WINDOW_PROTOCOL.md](HFD_WINDOW_PROTOCOL.md)。 +diff --git a/docs/history/OMG_CONTEXT_DEBUG.md b/docs/history/OMG_CONTEXT_DEBUG.md +deleted file mode 100644 +index 92bcc43..0000000 +--- a/docs/history/OMG_CONTEXT_DEBUG.md ++++ /dev/null +@@ -1,79 +0,0 @@ +-> 历史实验记录;当前入口见 [README](../../README.md)。旧运行目录已移动到 `/root/autodl-tmp/hfd-runs-archive-20261007/`。 +- +-# OMG 两秒历史上下文与动作启动调试 +- +-样本仍是原生 `train-002203-seed42`:接取物品尚未完成时,输入捡起地上物品的新指令。完整原生音轨、中文/英文与背景音处理策略没有改变。此样本的两条录音在视频时间 0.4–2.16 s、5.2–6.96 s,视频到 14.8 s。最后一句话结束后已有约 7.8 s,所以不能直接归因于视频结束太早。 +- +-## 输入策略 +- +-新策略 `data.audio_policy: hfd_past_2s`:以 HFD 当前组的释放截止 `hi=9600+6400*k` 为准,读取 `[max(0,hi-32000),hi)` 原生 PCM。不足两秒时左侧补零。原生 OMG 仍保持 30 Hz、125D 动作、10 帧历史、60 帧预测与 `current35` 音频特征。对完整两秒 PCM 提取原生逐帧特征,取前 60 行,不做旧 0.6 秒到 60 位置的插值。 +- +-训练和推理都采用同一策略。没有额外 KV 预填充,每次重新编码已听过的最近两秒;未来音频不输入。两秒是有限历史上下文,不保证所有长指令都完整保留,也不是完整的 ASR/语言理解编码器。开头补零和窗口内峰值归一化在训练/推理中一致。 +- +-新增测试检查:截止时刻保持 HFD 时钟、左侧补零长度、原生特征顺序、修改未释放音频不影响当前特征,以及之前的录音确实能影响两秒条件。数据合同检查覆盖动作/音频因果边界。 +- +-## 对照设计 +- +-所有实验使用同一个 native session、相同初始随机种子与一次真实历史初始化。训练时动作点和未来目标保持一致;新模型从官方 50m 初始化重新训练 1600 步,没有从旧离线模型续训。 +- +-| 实验 | 训练音频 | 推理音频 | 每轮执行 | +-| --- | --- | --- | --- | +-| A | 当前 0.6 s | 当前 0.6 s | 0.4 s | +-| B | 当前 0.6 s | 最近两秒 | 0.4 s | +-| B2 | 当前 0.6 s | 最近两秒 | 2 s | +-| C | 最近两秒 | 最近两秒 | 0.4 s | +-| C2 | 最近两秒 | 最近两秒 | 2 s | +- +-B 就是用户提出的“训练 0.6 秒,推理两秒”,保留为诊断对照。B/C 的音频时间都不超出当前 HFD 释放边界。B2/C2 检查原生两秒预测能否在后半段启动;它们两秒才更新一次规划,不能与 0.4 秒重新规划混为同一个实时反应口径。 +- +-每轮报告记录两秒预测前 0.4 秒与之后 1.6 秒相对上一已执行姿态的关节变化。不能只看执行片段而忽略计划后半段:反复重规划可能把稍后的启动动作丢弃。执行后的状态才反馈到下一轮,没有使用未执行未来姿态。 +- +-## 配置与复现 +- +-新配置 `configs/history/omg_duplex_past2s.yaml` 包含原生框架、官方初始化权重、dataset、GPU、batch size、训练步数、音频策略与 `inference.omg_execute_frames: 12`。迁移时按根目录 README 准备环境和外部权重,修改 YAML 路径后运行: +- +-```bash +-python train.py --config configs/history/omg_duplex_past2s.yaml +-python infer_overfit.py --run RUN --continuous-only --task-seconds 0 14.8 +-python infer_overfit.py --run RUN --continuous-only --task-seconds 0 14.8 --omg-execute-frames 60 --output-dir RUN/inference-execute2 +-``` +- +-复现 B:将 RUN 指向已有当前 0.6 秒训练结果,再加 `--omg-audio-policy hfd_past_2s --output-dir NEW_OUTPUT`。推理报告分别记录训练和推理策略,防止把输入不匹配的结果误认为匹配训练。 +- +-`--gpu-id` 只改变推理用的物理 GPU,不改训练配置或权重。输出目录必须不存在。渲染仍使用原有 `render_overfit.py`,交付为 H.264/MP3 MP4,保留对应原生音轨。 +- +-## 本次结果 +- +-| 实验 | 整段关节 RMSE(rad) | root RMSE(m) | +-| --- | --- | --- | +-| A:0.6 s 训练 / 0.6 s 推理 / 执行 0.4 s | 0.30954 | 0.06111 | +-| B:0.6 s 训练 / 2 s 推理 / 执行 0.4 s | 0.30978 | 0.04566 | +-| B2:0.6 s 训练 / 2 s 推理 / 执行 2 s | 0.26777 | 0.04292 | +-| C:2 s 训练 / 2 s 推理 / 执行 0.4 s | 0.25026 | 0.04641 | +-| C2:2 s 训练 / 2 s 推理 / 执行 2 s | 0.18776 | 0.03093 | +- +-新两秒上下文模型的真实历史训练窗口 RMSE 从初始化 0.31303 rad 降到 0.09066 rad。两种执行长度使用相同最终权重,不是分别训练或按较好视频选择检查点。 +- +-只把旧 0.6 秒训练模型的推理输入延长到两秒,没有解决几乎不动。匹配训练与推理上下文后,第一项动作能启动;0.4 秒重规划下第二项仍较弱。以原生两秒块执行时,两段实质动作都可见,但捡取动作晚于真值,仍未充分拟合。 +- +-计划检查表明,某些静止历史下模型将明显动作安排在预测后半段。例如 B 在 source 2.4 s 的计划:前 0.4 s 相对上一姿态变化约 0.00518 rad,后 1.6 s 约 0.42096 rad。频繁重新规划会丢掉这部分计划,是此样本启动不足的一个因素。这是当前样本的诊断,不把两秒执行视为更好的实时策略;两秒执行降低指令更新频率,正式论文必须明确执行周期与反应延迟。 +- +-新视频: +- +-- `omg_duplex_past2_exec04_playable.mp4`:训练/推理统一两秒历史,0.4 秒重规划,14.8 s。 +-- `omg_duplex_past2_exec2_playable.mp4`:相同权重、相同两秒历史,原生两秒块执行,14.8 s。 +-- `omg_duplex_past2_exec2_tail4_playable.mp4`:相同两秒块推理追加四秒静音,18.8 s。 +- +-视频都在 `/root/autodl-tmp/runs/hfd-overfit-videos-jointfixed`。左侧预测、右侧原生真值,保留原始任务录音和背景音,H.264/MP3 MP4。第三条 14.8 s 之后右侧仅保持最终姿态,不存在真实真值,画面有明确标注。 +- +-延长尾段时,原始 14.8 s 的预测与原音轨逐项完全不变;新增部分全为诊断静音,并从误差计算中排除。因此延长版仍为 0.18776 rad,而不能借延长真值待机降低指标。新增四秒主要是收尾/待机,没有补出新的任务动作。原始模型的几乎不动不能归因于没有多给四秒等待时间。 +- +-本次 1600 步原生训练、严格最终权重重载、数据合同和 29 项回归检查通过;三条新视频完整声画解码通过。新 run 三份等价重复检查点已清理,释放约 2.96 GiB,唯一 `last.ckpt` 保留。完整数字、每轮窗口边界和执行片段记录见 `omg_context_debug_summary.json`。 +- +-延长尾段只用于诊断,必须跟在完整原生 session 后;复现命令: +- +-```bash +-python infer_overfit.py --run RUN --continuous-only --task-seconds 0 14.8 --omg-execute-frames 60 --tail-seconds 4 --output-dir RUN/inference-tail4 +-``` +- +-当前训练/推理两秒上下文模板仍默认执行 12 帧(0.4 s);两秒块对照用 CLI `--omg-execute-frames 60`。窗口长度、预测长度、执行长度是三个不同参数,不能把扩大上下文和延长等待混为同一件事。 +diff --git a/docs/history/README.md b/docs/history/README.md +deleted file mode 100644 +index c6d32cb..0000000 +--- a/docs/history/README.md ++++ /dev/null +@@ -1,9 +0,0 @@ +-> 最新调试改为 full_session 动作监督;OMG 10 步采样,TextOP 连续生成仍未解决。以根目录 README 和 docs/ROLLOUT_DEBUG.md 为准。 +- +-# 历史记录 +- +-本目录和 `configs/history/`、`reports/history/` 保留过去实验的文档、配置与检查记录。当前方案以根目录 README 为准。 +- +-这些文件没有删除,也没有移出 Git 仓库。引用的旧运行路径属于当时证据;实际产物现归档到 `/root/autodl-tmp/hfd-runs-archive-20261007/`。 +- +-旧运行产物后来经用户确认永久删除;历史配置和报告仍保留,但旧产物路径不再可用。当前样本与模型保存在 `runs/hfd-final/`。 +diff --git a/docs/history/VIDEO_SAMPLES.md b/docs/history/VIDEO_SAMPLES.md +deleted file mode 100644 +index 5a47fd5..0000000 +--- a/docs/history/VIDEO_SAMPLES.md ++++ /dev/null +@@ -1,92 +0,0 @@ +-> 历史实验记录;当前入口见 [README](../../README.md)。旧运行目录已移动到 `/root/autodl-tmp/hfd-runs-archive-20261007/`。 +- +-推荐查看新的完整单任务视频 `textop_complete_task_playable.mp4`、`omg_complete_task_playable.mp4`,任务与训练说明见 [NORMAL_TASK.md](NORMAL_TASK.md)。下文是此前八窗口及事件短片的记录。 +- +-# 修正关节映射后的推理视频与任务说明 +- +-七个 MP4 均已重新导出声音,使用 H.264 画面、48 kHz 双声道 MP3 音轨;逐个完整解码,检查非零音量、音频时长和视频帧数。音轨来自原始 session 的混合录音,保留语音、背景音和静音间隔。这批视频只对播放声音施加固定增益(不削波),没有改变模型输入、预测动作或训练结果。逐文件增益、音量和哈希见 `audio_verification.json`。 +- +-除 `gt_before_after_playable.mp4` 外,左侧是模型预测,右侧是真值。真值对照视频左侧是旧的错误关节解释,右侧是正确解释;两侧均为同一段真值,没有模型预测。 +- +-## 原生 session 中的任务 +- +-任务描述依据 session 的原始动作名与事件元数据,便于理解动作。下面引用的普通任务英文为已保存 ASR 的识别结果,可能有识别误差;`Stop now.` 是原始控制事件中记录的口令。GT 描述只用于本文解释,没有作为额外模型输入。 +- +-| session | 任务与交互 | 声音 | +-| --- | --- | --- | +-| `train-000002-seed42` | 第一轮从站立进入吸烟动作(`idle_to_smoke_idle_R_001__A552`);第一任务尚未结束时,第二轮要求向前走并抓取物体(`grab_walk_ff_180_001__A549`)。新任务打断旧任务。 | 英文男声;ASR 为 “from standing to standing and smoking.”,随后 “Character walks forward while grabbing something.”;保留空调背景音。 | +-| `train-000014-seed42` | 四轮分别为转身、站立喝水后扔瓶、再次转身、向前走并抓取物体。第二轮喝水/扔瓶任务未完成时插入 `Stop now.`,然后再进行后续任务。 | 英文女声;第二轮 ASR 为 “Drinks from bottle while standing, then throws bottle to the left.”;停止口令来自原始元数据;保留空调背景音。 | +- +-原生数据支持中文和背景音,本次选中的这两个 session 恰好是英文,未把中文样本转成英文。动作名和 ASR 的左右方向可能不一致,因此描述不据 ASR 推断精确转向;真值动作以原始轨迹为准。 +- +-## 连续事件视频 +- +-以下时间范围均为原始 session 的动作时间。原生适配器的麦克风时间比动作时间多 0.2 秒;音频片段按这一关系裁切,与事件时间对齐。 +- +-| 文件 | 动作时间范围 | 关键指令在视频中的时间 | 要观察的任务 | +-| --- | --- | --- | --- | +-| `textop_interruption_playable.mp4` | 7.72–13.64 s | 第二轮开始于约 1.08 s | 吸烟任务进行中输入向前走、抓取的新任务;听到的是第二轮指令,第一轮完整指令在更早处。 | +-| `omg_interruption_playable.mp4` | 7.63–13.63 s | 第二轮开始于约 1.17 s | 同上。 | +-| `textop_emergency-stop_playable.mp4` | 81.00–86.92 s | `Stop now.` 开始于约 1.00 s | 喝水/扔瓶任务中的急停;第二轮完整任务指令位于录音 73.8–77.0 s,不包含在这一短片中。 | +-| `omg_emergency-stop_playable.mp4` | 80.83–86.83 s | `Stop now.` 开始于约 1.17 s | 同上。 | +-| `gt_before_after_playable.mp4` | 7.00–13.00 s | 第二轮开始于约 1.80 s | 同一段吸烟被新任务打断的真值,仅用于对照错误/正确的关节映射。现在也包含对应原始音频。 | +- +-指令前后短暂只有背景音是 session 原本的内容。连续推理开头使用一次真实历史,之后反馈模型预测。不是所有事件窗口都在八窗口训练子集中,因此这些片段不是完整的全双工能力评测。 +- +-## 训练窗口拼接视频 +- +-`*_fitted_windows_playable.mp4` 拼接八个用于训练检查的窗口,并非完整连续任务。TextOP 每段 0.64 秒,总长 5.12 秒;OMG 每段 2 秒,总长 16 秒。窗口常落在动作执行或待机阶段,可能只含背景音,也可能截到半句话;不添加额外解说或补播片段外的指令。 +- +-### TEXTOP:`textop_fitted_windows_playable.mp4` +- +-| 样本 | 视频时间 | 原始 session | 原始动作时间 | +-| --- | --- | --- | --- | +-| 0 | 0.00–0.64 s | `train-000002-seed42.npz` | 0.04–0.68 s | +-| 1 | 0.64–1.28 s | `train-000014-seed42.npz` | 4.20–4.84 s | +-| 2 | 1.28–1.92 s | `train-000014-seed42.npz` | 33.32–33.96 s | +-| 3 | 1.92–2.56 s | `train-000014-seed42.npz` | 62.44–63.08 s | +-| 4 | 2.56–3.20 s | `train-000014-seed42.npz` | 91.72–92.36 s | +-| 5 | 3.20–3.84 s | `train-000014-seed42.npz` | 120.84–121.48 s | +-| 6 | 3.84–4.48 s | `train-000014-seed42.npz` | 149.96–150.60 s | +-| 7 | 4.48–5.12 s | `train-000014-seed42.npz` | 179.24–179.88 s | +- +-### OMG:`omg_fitted_windows_playable.mp4` +- +-| 样本 | 视频时间 | 原始 session | 原始动作时间 | +-| --- | --- | --- | --- | +-| 0 | 0.00–2.00 s | `train-000002-seed42.npz` | 0.43–2.43 s | +-| 1 | 2.00–4.00 s | `train-000014-seed42.npz` | 5.63–7.63 s | +-| 2 | 4.00–6.00 s | `train-000014-seed42.npz` | 34.43–36.43 s | +-| 3 | 6.00–8.00 s | `train-000014-seed42.npz` | 62.83–64.83 s | +-| 4 | 8.00–10.00 s | `train-000014-seed42.npz` | 91.63–93.63 s | +-| 5 | 10.00–12.00 s | `train-000014-seed42.npz` | 120.03–122.03 s | +-| 6 | 12.00–14.00 s | `train-000014-seed42.npz` | 148.83–150.83 s | +-| 7 | 14.00–16.00 s | `train-000014-seed42.npz` | 177.63–179.63 s | +- +-## 结果与重现 +- +-两个模型从官方初始化权重重新训练 800 步;修复前关节映射错误的检查点、视频和动作指标作废。TextOP 预测 23 个关节,其余腕关节固定在初始化姿态;OMG 预测全部 29 个关节。OMG 使用离线目标段音频,TextOP 使用已释放窗口的 ASR。视频是参考姿态播放,不是物理跟踪执行。 +- +-训练、误差与映射核验位于数据盘的视频目录 `/root/autodl-tmp/runs/hfd-overfit-videos-jointfixed`,见 `summary.json`、`textop_report.json`、`omg_report.json`、`mapping_audit.json`;新的声音核验见 `audio_verification.json`。 +- +-后续导出使用代码仓库的 `render_overfit.py`,其默认生成 48 kHz 双声道 MP3 并实测解码音量。真值对照使用 `check_joint_mapping.py`,现在也导出原始音频。已有画面可用 `remux_video_audio.py` 重建音轨,例如: +- +-```bash +-python remux_video_audio.py \ +- --video /data/runs/textop/inference/interruption_playable.mp4 \ +- --audio-npz /data/runs/textop/inference/interruption.npz \ +- --output /data/runs/textop/inference/interruption_with_audio.mp4 +-``` +- +-若从完整 session NPZ 裁切音频,需传 `--audio-start-seconds`(动作起点 + 0.2 秒)。默认播放增益为 1;`--gain` 仅影响导出的声音,不影响模型。脚本在增益会造成削波时拒绝导出。 +- +-## 当前视频格式与 VS Code 播放 +- +-当前交付统一为 MP4 容器、H.264 视频、48 kHz 双声道 MP3 音轨;文件名以 `_playable.mp4` 结尾。根据官方文档选择 H.264、MP4 与 MP3,避开不支持的 AAC。此前 VP8 WebM 在用户当前预览中仍无法打开,因此已清理。 +- +-官方支持清单:https://github.com/microsoft/vscode/issues/180780 +-官方 Webview 说明:https://code.visualstudio.com/api/extension-guides/webview#supported-media-formats +- +-所有视频检查过完整声画解码、非静音音轨、帧数和帧率;H.264 码流与原始视频的 SHA256 一致,转换只更换音频编码和封装,没有改动作。检查记录见视频目录的 `playable_mp4_check.json`。远程服务器上的解码检查不等于本地 VS Code 实测,首次打开请使用新文件名。 +- +-后续 `render_overfit.py` 与 `remux_video_audio.py` 默认导出 H.264/MP3 MP4,无需 WebM 或额外的 `--vscode-preview` 参数。 +diff --git a/hfd_bridge/genmo_prefetch.py b/hfd_bridge/genmo_prefetch.py +index e9445d1..b37742f 100644 +--- a/hfd_bridge/genmo_prefetch.py ++++ b/hfd_bridge/genmo_prefetch.py +@@ -15,6 +15,8 @@ class CPUBatches(IterableDataset): + + def __iter__(self): + os.environ['CUDA_VISIBLE_DEVICES']='' ++ from .genmo_runtime import require_replay_runtime ++ require_replay_runtime() + torch.set_num_threads(2) + from .genmo_data import PairedNativeProvider,GenmoBatches + provider=PairedNativeProvider(**self.options) +diff --git a/hfd_bridge/genmo_training.py b/hfd_bridge/genmo_training.py +index 3f094ed..eba28ac 100644 +--- a/hfd_bridge/genmo_training.py ++++ b/hfd_bridge/genmo_training.py +@@ -13,6 +13,19 @@ def evaluation_mode(model): + for module,state in states.items():module.training=state + + ++@contextmanager ++def sampling_autocast(device_type): ++ """Do not cache no-grad weight casts inside the outer training autocast. ++ ++ Reusing those casts in the subsequent loss forward silently removes the ++ denoiser's parameter gradients. Preserve the caller's AMP dtype and state. ++ """ ++ with torch.autocast(device_type=device_type, ++ enabled=torch.is_autocast_enabled(device_type), ++ dtype=torch.get_autocast_dtype(device_type),cache_enabled=False): ++ yield ++ ++ + def execution_mask(mask,history,execute,weight): + """Preserve each feature's total loss mass while emphasizing executed frames.""" + result=mask.float().clone() +@@ -61,7 +74,7 @@ def consistent_translation_loss(inputs,outputs,pipeline): + def generate_parameters(model,batch): + """Generate from past parameters only. Future placeholders contain no GT.""" + from .genmo import HISTORY,FUTURE +- with evaluation_mode(model): ++ with sampling_autocast(batch['smpl_params_w']['transl'].device.type),evaluation_mode(model): + model.prepare_batch(batch,'regression') + model.create_condition_mask(batch,{},'diffusion',train=False) + result=model.pipeline.denoiser3d.forward_test(batch) +diff --git a/hfd_bridge/human_sessions.py b/hfd_bridge/human_sessions.py +index 52fb0c1..89175d9 100644 +--- a/hfd_bridge/human_sessions.py ++++ b/hfd_bridge/human_sessions.py +@@ -65,11 +65,11 @@ class NativeHumanObserver: + def __init__(self, ds, module, root): + # Share the bridge's verified local metadata mirror. Raw SMPL-X assets + # remain indexed in their original shared-storage tar shards. +- from .storage import ArchivePath ++ from .storage import ArchivePath,store + class IndexedReader: + def read(self, path):return ArchivePath(path).read_bytes() + self.store = IndexedReader() +- pairs = pd.read_parquet(Path(root)/'metadata/seed_g1_smplx_pairs.parquet') ++ pairs = pd.read_parquet(store().metadata_path('metadata/seed_g1_smplx_pairs.parquet',root)) + self.pairs = {r.motion_id:r for r in pairs.itertuples()} + self.ds, self.module = ds, module + self.rows = {} +diff --git a/hfd_bridge/kimodo.py b/hfd_bridge/kimodo.py +index 67cb18d..e9742a6 100644 +--- a/hfd_bridge/kimodo.py ++++ b/hfd_bridge/kimodo.py +@@ -25,6 +25,8 @@ def setup(gpu=0): + HFD_BODY_MODELS_ROOT=os.environ.get('HFD_BODY_MODELS_ROOT','/root/autodl-fs/pangxinlei/models/body_models'), + TEXT_ENCODERS_DIR=os.environ.get('TEXT_ENCODERS_DIR',str(Path(os.environ.get('KIMODO_MODELS_ROOT','/root/autodl-fs/pangxinlei/models/kimodo'))/'runtime-text-encoders')), + TEXT_ENCODER_DEVICE='cuda:0',OMP_NUM_THREADS='4',MKL_NUM_THREADS='4') ++ os.environ.setdefault('HFD_NATIVE_STARTUP_CACHE','1') ++ os.environ.setdefault('HFD_STARTUP_INDEX_CACHE_DIR',str(Path(os.environ.get('HFD_DATA_DISK','/root/autodl-tmp'))/'cache/hfd/startup-indexes')) + torch.set_num_threads(4) + + +diff --git a/hfd_bridge/loader.py b/hfd_bridge/loader.py +index 31321d0..124b42d 100644 +--- a/hfd_bridge/loader.py ++++ b/hfd_bridge/loader.py +@@ -30,7 +30,7 @@ def _instrument_returns(text,filename): + tree=Returns().visit(ast.parse(text,filename)) + return compile(ast.fix_missing_locations(tree),filename,'exec') + +-def load_hfd(): ++def load_hfd(replay_numpy2=False): + root=Path(os.environ['HFD_CODE_ROOT']);sys.path.insert(0,str(root/'src')) + source=root/'src/psi/data/seed/full_duplex_audio_command_dataset.py' + text=source.read_text() +@@ -74,5 +74,9 @@ def load_hfd(): + module._hfd_return_hooks=hooked;module._hfd_current_observer=None + module._hfd_observer_return=_return_hook + exec(_instrument_returns(text,str(source)) if hooked else compile(text,str(source),'exec'),module.__dict__) ++ if replay_numpy2: ++ from .replay_numerics import quat_to_yaw_wxyz,wrap_angle ++ module.quat_to_yaw_wxyz=quat_to_yaw_wxyz ++ module.wrap_angle=wrap_angle + module.SOURCE_SHA256=hashlib.sha256(source.read_bytes()).hexdigest() + return module +diff --git a/hfd_bridge/omg_checks.py b/hfd_bridge/omg_checks.py +index 17e6574..7784fa9 100644 +--- a/hfd_bridge/omg_checks.py ++++ b/hfd_bridge/omg_checks.py +@@ -37,8 +37,9 @@ class WarmupEarlyStopping(EarlyStopping): + + + class OMGTrainingObserver(pl.Callback): +- def __init__(self,output,check_pipeline=False,audio_policy='hfd_past_2s',use_asr_text=False,input_text_source='asr'): ++ def __init__(self,output,check_pipeline=False,audio_policy='hfd_past_2s',use_asr_text=False,input_text_source='asr',checkpoint_root=None): + self.output=Path(output);self.check_pipeline=check_pipeline;self.optimizer_updates=0;self.validation_batches=0 ++ self.checkpoint_root=Path(checkpoint_root) if checkpoint_root else self.output + self.audio_policy=audio_policy + self.use_asr_text=use_asr_text + self.input_text_source=input_text_source +@@ -52,6 +53,7 @@ class OMGTrainingObserver(pl.Callback): + self.optimizer_updates+=1 + def on_train_start(self,trainer,module): + self.initial_step=trainer.global_step ++ self.logged_step=trainer.global_step;self.micro_losses=[];self.micro_compute=[];self.micro_wait=[] + def on_before_optimizer_step(self,trainer,module,optimizer): + if self.check_pipeline: + grads=[p.grad for p in module.parameters() if p.grad is not None] +@@ -60,12 +62,18 @@ class OMGTrainingObserver(pl.Callback): + def on_train_batch_end(self,trainer,module,outputs,batch,batch_idx): + loss=outputs['loss'] if isinstance(outputs,dict) else outputs + if not torch.isfinite(loss).all():raise RuntimeError('Nonfinite OMG training loss') +- if trainer.is_global_zero: ++ self.micro_losses.append(float(loss.detach())*trainer.accumulate_grad_batches) ++ self.micro_compute.append(time.monotonic()-self.batch_started) ++ if self.data_wait_seconds is not None:self.micro_wait.append(self.data_wait_seconds) ++ if trainer.is_global_zero and trainer.global_step>self.logged_step: + with (self.output/'metrics.jsonl').open('a') as f: +- f.write(json.dumps(dict(step=trainer.global_step,loss=float(loss.detach()), ++ f.write(json.dumps(dict(step=trainer.global_step,loss=sum(self.micro_losses)/len(self.micro_losses), ++ micro_batches=len(self.micro_losses),learning_rate=trainer.optimizers[0].param_groups[0]['lr'], + elapsed_seconds=time.monotonic()-self.started, +- data_wait_seconds=self.data_wait_seconds, +- compute_seconds=time.monotonic()-self.batch_started))+'\n') ++ data_wait_seconds=sum(self.micro_wait),compute_seconds=sum(self.micro_compute), ++ peak_allocated_mib=torch.cuda.max_memory_allocated(module.device)/2**20 if module.device.type=='cuda' else 0))+'\n') ++ if trainer.global_step>self.logged_step: ++ self.logged_step=trainer.global_step;self.micro_losses=[];self.micro_compute=[];self.micro_wait=[] + self.previous_batch_end=time.monotonic() + def on_train_batch_start(self,trainer,module,batch,batch_idx): + self.batch_started=time.monotonic() +@@ -91,7 +99,7 @@ class OMGTrainingObserver(pl.Callback): + elapsed_seconds=time.monotonic()-self.started,best_checkpoints=best,early_stopping=early),indent=2)+'\n') + if not self.check_pipeline:return + changed=any(not torch.equal(self.initial[n],p.detach().cpu()) for n,p in module.named_parameters() if n in self.initial) +- file=self.output/'pipeline.ckpt' ++ file=self.checkpoint_root/'pipeline.ckpt' + trainer.save_checkpoint(str(file)) + trainer.strategy.barrier() + ckpt=torch.load(file,map_location='cpu',weights_only=False) +diff --git a/hfd_bridge/ondemand.py b/hfd_bridge/ondemand.py +index 0fd7144..4d00440 100644 +--- a/hfd_bridge/ondemand.py ++++ b/hfd_bridge/ondemand.py +@@ -35,8 +35,15 @@ def native_dataset(config,split): + os.environ[env]=str(resolve(c[field])) + os.environ['HFD_STARTUP_INDEX_CACHE']=os.environ.get('HFD_NATIVE_STARTUP_CACHE','0') + os.environ['HFD_RANGE_CACHE_MAX_BYTES']=str(1024*2**20) +- ds=build_dataset(c.limit_sources,split,OmegaConf.to_container(c.recipe_overrides),candidate_pool=c.candidate_pool) ++ selection=None ++ if c.get('final_release'): ++ from .final_split import FinalSelection ++ selection=FinalSelection(OmegaConf.to_container(c.final_release,resolve=True),split) ++ selection.identity['third_party_replay_source_sha256']=sha256(resolve(c.code_root)/'src/psi/data/seed/third_party_dataset.py') ++ upstream_split=selection.upstream_split if selection is not None else split ++ ds=build_dataset(c.limit_sources,upstream_split,OmegaConf.to_container(c.recipe_overrides),candidate_pool=c.candidate_pool,replay_numpy2=selection is not None) + ds.bridge_compact_output=True ++ if selection is not None:ds.bridge_final_selection=selection + _datasets[key]=ds + return _datasets[key],c + +@@ -113,6 +120,24 @@ def legal_times(arrays,model,num_primitive=4): + return range(12,max(12,stop)) + + ++def final_window_times(session,prefix_frames,model,num_primitive): ++ """Use published clocks for rejection tests, without replaying a candidate. ++ ++ Native dense starts stop prefix_frames+1 before the final motion frame. ++ The pinned third-party replay keeps that same complete native timeline. ++ This supplies the exact legal grid before loading any motion/audio arrays. ++ """ ++ count=int(session['dense_steps']);groups=int(session['semantic_windows']) ++ geometry=dict(valid_start_count=count, ++ motion_q29_xyz_quat=range(count+int(prefix_frames)+1), ++ audio=range((groups-1)*6400+9600)) ++ return legal_times(geometry,model,num_primitive) ++ ++ ++def final_window_count(session,prefix_frames,model,num_primitive): ++ return len(final_window_times(session,prefix_frames,model,num_primitive)) ++ ++ + class NativeSessionProvider: + def __init__(self,config,split,model,seed=42,num_primitive=4,memory_mib=512, + asr_cache=None,asr_cache_mib=2048,asr_model=None,asr_device='cuda:0',asr_batch_size=32, +@@ -121,6 +146,8 @@ class NativeSessionProvider: + if language not in ('all','en','zh'):raise ValueError('Unknown foreground language filter') + self.language=language + self.ds,c=native_dataset(config,split) ++ self.split=split ++ self.final_selection=getattr(self.ds,'bridge_final_selection',None) + if c.candidate_pool or c.limit_sources or c.count is not None: + raise ValueError('On-demand formal sampling requires the full balanced native recipe') + self.model=model;self.seed=int(seed);self.num_primitive=num_primitive +@@ -152,6 +179,10 @@ class NativeSessionProvider: + upper=np.asarray([self.ds._session_max_stream_steps(s)*20+self.ds.prefix_frames+1 for s in self.ds.sessions],dtype=np.int64) + self.bounds=(upper+7)//8 if model=='textop' else (upper*3+4)//5 + self.eligible=np.array([language=='all' or s[0].get('_session_voice_name','').endswith('_'+language) for s in self.ds.sessions],dtype=bool) ++ if self.final_selection is not None: ++ allowed=self.final_selection.native ++ if allowed and max(allowed)>=len(self.ds.sessions):raise ValueError('Final native index exceeds the pinned source inventory') ++ self.eligible &= np.array([i in allowed for i in range(len(self.ds.sessions))],dtype=bool) + if not self.eligible.any():raise ValueError('No native sessions in requested foreground language') + self.bounds[~self.eligible]=0 + self.rank=int(os.environ.get('RANK','0'));self.world_size=int(os.environ.get('WORLD_SIZE','1')) +@@ -174,27 +205,39 @@ class NativeSessionProvider: + for session in self.ds.sessions: + digest.update(json.dumps([(int(r['row_id']),r.get('_session_interaction_mode'),r.get('_session_voice_name'),r.get('_session_voice_variant')) for r in session],sort_keys=True).encode()) + self.audit['ordered_session_selection_sha256']=digest.hexdigest() ++ if self.final_selection is not None: ++ self.audit['final_release']=self.final_selection.identity ++ self.audit['ordered_session_selection_sha256']=self.final_selection.ordered_identity(language) + def arrays(self,index): + if hasattr(self,'eligible') and not self.eligible[int(index)]:raise ValueError('Session excluded by foreground language filter') + cached=self.cache.get(index) + if cached is not None:return cached + raw=fixed_session(self.ds,int(index),self.seed);arrays=raw['arrays'] + raw['times']=legal_times(arrays,self.model,self.num_primitive) ++ if getattr(self,'final_selection',None) is not None and self.model in ('textop','omg'): ++ if raw['times']!=self.window_times(index):raise ValueError('Final native window grid changed') + if len(raw['times'])>self.bounds[index]:raise RuntimeError('Native conservative length bound violated') + raw['size']=sum(v.nbytes for v in arrays.values()) + self.cache.put(index,raw,raw['size']);return raw ++ def window_times(self,index): ++ if getattr(self,'final_selection',None) is not None and self.model in ('textop','omg'): ++ return final_window_times(self.final_selection.native_record(index)['session'], ++ self.ds.prefix_frames,self.model,self.num_primitive) ++ return self.arrays(index)['times'] ++ def window_count(self,index): ++ return len(self.window_times(index)) + def choose(self): + for _ in range(10000): + slot=int(self.rng.integers(int(self.cumulative[-1]))) + index=int(np.searchsorted(self.cumulative,slot,side='right')) +- raw=self.arrays(index);n=len(raw['times']) ++ n=self.window_count(index) + if self.rng.random()*self.bounds[index]self.bounds[index]:raise RuntimeError('Replayed third-party length bound violated') ++ raw_item=dict(arrays=arrays,times=times,size=sum(x.nbytes for x in arrays.values()),source_id='third-party::'+sid) ++ self.cache.put(cache_key,raw_item,raw_item['size']);return raw_item + import io, wave + with np.load(io.BytesIO(self._asset('timelines',sid)),allow_pickle=False) as z: + raw=np.asarray(z['states'],dtype=np.float32) +@@ -347,7 +418,7 @@ class ThirdPartySessionProvider: + for _ in range(10000): + slot=int(self.rng.integers(int(self.cumulative[-1]))) + index=int(np.searchsorted(self.cumulative,slot,side='right')) +- n=len(self.arrays(index)['times']) ++ n=self.window_count(index) + if self.rng.random()*self.bounds[index]self.capacity:self.files.popitem(last=False) ++ data=os.pread(handle.fileno(),count,start) ++ while len(data)1 and not ondemand, +- val_check_interval=c.training.eval_every,check_val_every_n_epoch=None) ++ val_check_interval=validation_interval(c.training),check_val_every_n_epoch=None) ++ from hydra.utils import instantiate ++ trainer_options['plugins']=[instantiate(cfg.checkpoint_io)] + if ondemand: +- trainer_options['limit_val_batches']=int(c.training.validation_batches) ++ trainer_options['limit_val_batches']=len(val_loader.dataset) + elif not val: + trainer_options['limit_val_batches']=0 + if os.environ.get('LOCAL_RANK','0')=='0': +@@ -316,7 +333,7 @@ def omg(c, train, val, output): + model.train() + if model.text_encoder is not None:model.text_encoder.encoder.eval() + trainer.fit(model,train_dataloaders=train_loader,val_dataloaders=val_loader,ckpt_path=c.training.get('resume')) +- trainer.save_checkpoint(str(output/'last.ckpt')) ++ trainer.save_checkpoint(str(weights_output/'last.ckpt')) + + + def native_options(c): +diff --git a/train_genmo_full.py b/train_genmo_full.py +index cd3a7cc..43bb8f7 100644 +--- a/train_genmo_full.py ++++ b/train_genmo_full.py +@@ -37,6 +37,20 @@ def main(): + for name in ['init_checkpoint','text_encoder','body_models']: + if not Path(c[name]).exists():raise FileNotFoundError(c[name]) + out=Path(c.output).resolve();weights=Path(c.weights_output).resolve() ++ from hfd_bridge.genmo_runtime import require_replay_runtime ++ runtime=require_replay_runtime() ++ from hfd_bridge.genmo_replay_audit import read_status ++ audit_path=c.data.get('replay_audit_status') ++ if audit_path: ++ audit_status=read_status(audit_path,runtime, ++ require_risk=bool(c.data.get('require_replay_risk_gate',True))) ++ from hfd_bridge.asr import sha256 as config_sha256 ++ if audit_status['identity']['config_sha256']!=config_sha256(Path(c.data.config)): ++ raise ValueError('Replay audit uses another session configuration') ++ for name,digest in audit_status['identity']['implementation'].items(): ++ if config_sha256(Path(__file__).parent/name)!=digest: ++ raise ValueError('Replay audit implementation changed: '+name) ++ if audit_status.get('failed'):raise RuntimeError('Replay audit failed: '+str(audit_status.get('failure'))) + if a.check_only: + print(OmegaConf.to_yaml(c));return + os.environ.update(CUDA_VISIBLE_DEVICES=str(c.gpu),GENMO_ROOT=str(c.genmo_root), +@@ -69,9 +83,11 @@ def main(): + protocol=OmegaConf.to_container(c,resolve=True) + for key in ['output','weights_output','gpu','python']:protocol.pop(key,None) + protocol['training'].pop('resume',None) ++ protocol['replay_runtime']=runtime + protocol['data_implementation']={name:sha256(Path(__file__).parent/name) for name in + ['train_genmo_full.py','hfd_bridge/genmo.py','hfd_bridge/genmo_data.py','hfd_bridge/genmo_validation.py', + 'hfd_bridge/genmo_training.py','hfd_bridge/genmo_geometry.py','hfd_bridge/genmo_prefetch.py', ++ 'hfd_bridge/genmo_runtime.py','hfd_bridge/genmo_replay_audit.py','hfd_bridge/genmo_safety.py', + 'hfd_bridge/ondemand.py','hfd_bridge/training.py','hfd_bridge/textop_condition.py', + 'hfd_bridge/human_sessions.py','hfd_bridge/reference_text.py','hfd_bridge/reference_runtime.py', + 'hfd_bridge/loader.py','hfd_bridge/sessions.py','hfd_bridge/storage.py','hfd_bridge/clocks.py']} +@@ -149,11 +165,17 @@ def main(): + finally:Path(temporary).unlink(missing_ok=True) + + class Audit(Callback): +- def __init__(self):self.start=time.monotonic();self.grad_checks=0;self.last_end=None ++ def __init__(self):self.start=time.monotonic();self.grad_checks=0;self.denoiser_grad_checks=0;self.last_end=None + def on_train_batch_start(self,trainer,module,batch,batch_idx): + self.started=time.monotonic();self.wait=None if self.last_end is None else self.started-self.last_end + self.data_profile=batch.pop('_hfd_data_profile',{}) + def on_before_optimizer_step(self,trainer,module,optimizer): ++ # Recovery sampling must never silence the trainable Transformer. ++ core=module.pipeline.denoiser3d.denoiser.blocks ++ missing=[name for name,param in core.named_parameters() if param.requires_grad and param.grad is None] ++ if missing: ++ raise FloatingPointError(f'GenMo Transformer gradients missing for {len(missing)} tensors: {missing[:3]}') ++ self.denoiser_grad_checks+=1 + if trainer.global_step>=4 and trainer.global_step%100 and self.grad_checks>=4:return + gradients=[v.grad for v in module.parameters() if v.requires_grad and v.grad is not None] + if not gradients or not all(torch.isfinite(g).all() for g in gradients):raise FloatingPointError('Missing/nonfinite gradients') +@@ -180,11 +202,13 @@ def main(): + def on_save_checkpoint(self,trainer,module,state):state['hfd_contract']=protocol + + audit=Audit() ++ from hfd_bridge.genmo_safety import GenmoReplaySafety ++ replay_safety=GenmoReplaySafety(weights,c.training.get('early_checkpoint_steps',[]),audit_path,runtime) + from hfd_bridge.genmo_validation import GenmoClosedLoopValidation + closed=c.get('strategy',{}).get('closed_loop_validation',{}) + monitor='val/closed_loop_score' if closed.get('enabled',False) else 'val/loss' + callbacks=[HFDResumeState(train,val),GenmoValidationNoise(int(c.training.seed)+1000003), +- GenmoClosedLoopValidation(val.provider,text,closed,out),audit, ++ GenmoClosedLoopValidation(val.provider,text,closed,out),audit,replay_safety, + ModelCheckpoint(dirpath=str(weights/'checkpoints'),every_n_train_steps=int(c.training.save_every), + save_top_k=int(c.training.keep_last_checkpoints),monitor='step',mode='max',save_last='link', + filename='step-{step:08d}',auto_insert_metric_name=False,enable_version_counter=False), +@@ -204,9 +228,12 @@ def main(): + DataLoader(val,batch_size=None,num_workers=0,generator=val_generator), + ckpt_path=str(checkpoint) if checkpoint else None) + summary=dict(global_step=trainer.global_step,elapsed_seconds=time.monotonic()-audit.start, +- checked_gradient_updates=audit.grad_checks,weights_output=str(weights),gpu_peak_bytes=torch.cuda.max_memory_allocated()) ++ checked_gradient_updates=audit.grad_checks,checked_denoiser_updates=audit.denoiser_grad_checks, ++ weights_output=str(weights),gpu_peak_bytes=torch.cuda.max_memory_allocated(), ++ replay_runtime=runtime,replay_audit_failure=replay_safety.failure) + (out/'training_summary.json').write_text(json.dumps(summary,indent=2)+'\n') + print('COMPLETE',json.dumps(summary),flush=True) ++ if replay_safety.failure:raise RuntimeError('Stopped at a complete update after replay audit failure') + + + if __name__=='__main__':main() +diff --git a/train_kimodo.py b/train_kimodo.py +index 75219f0..aba64ab 100644 +--- a/train_kimodo.py ++++ b/train_kimodo.py +@@ -51,6 +51,8 @@ def main(): + p.add_argument('--resume',nargs='?',const='auto');p.add_argument('--stop-after',type=int) + p.add_argument('--offline-check',action='store_true',help='Full pretrained gradient/text/causality checks only') + a=p.parse_args();c=OmegaConf.load(a.config);out=Path(c.output);weights=Path(c.weights_output) ++ # Model imports can initialise CUDA; select the physical GPU before them. ++ os.environ['CUDA_VISIBLE_DEVICES']=str(int(c.gpu)) + if c.get('kimodo_root'):os.environ['KIMODO_ROOT']=str(c.kimodo_root) + if c.get('body_models'):os.environ['HFD_BODY_MODELS_ROOT']=str(c.body_models) + if c.get('text_models_root'):os.environ['KIMODO_MODELS_ROOT']=str(c.text_models_root) diff --git a/handoff-v1/inventory/textop-git-status.txt b/handoff-v1/inventory/textop-git-status.txt new file mode 100644 index 0000000000000000000000000000000000000000..e02f3ef5e8126ca9153363f0d1393f8349abf734 --- /dev/null +++ b/handoff-v1/inventory/textop-git-status.txt @@ -0,0 +1,109 @@ + D FINAL_BASELINE_METHOD.md + D JOINT_MAPPING.md + M MANIFEST.json + M README.md + D SESSION_FORMAT.md + M UPSTREAM_LOCK.json + M configs/genmo.yaml + M configs/genmo_legacy.yaml + M configs/kimodo.yaml + M configs/omg.yaml + M configs/omg_variants.yaml + M configs/sessions.yaml + M configs/textop.yaml + D docs/ASR_CONTEXT_OVERFIT.md + D docs/BASELINE_TEXT_CONDITION.md + D docs/ENGLISH_ONLY_PROBE.md + D docs/FORMAL_TRAINING_READINESS.md + D docs/GENMO_FORMAL_TRAINING.md + D docs/GENMO_HFD_OVERFIT.md + D docs/GENMO_LISTENING_DIAGNOSIS.md + D docs/GENMO_PAST_AUDIO_TEST.md + D docs/GENMO_V2_TRAINING.md + D docs/GENMO_V2_VALIDATION.md + D docs/HFD_HUMAN_BASELINE_INPUT_PROTOCOL.md + M docs/HFD_HUMAN_BASELINE_SOURCES.json + D docs/HUMAN_FORMAL_VALIDATION_20261024.md + D docs/HUMAN_RETARGET_EVALUATION.md + D docs/KIMODO_CLOSED_LOOP_STABILITY.md + D docs/KIMODO_FORMAL_VALIDATION.md + D docs/KIMODO_HFD_OVERFIT_RESULTS.md + D docs/KIMODO_HFD_TRAINING.md + D docs/KIMODO_MOTION_RECOVERY.md + D docs/KIMODO_ROOT_COLLAPSE_DIAGNOSIS.md + D docs/KIMODO_ROOT_TRAINING_CAUSE_ANALYSIS.md + D docs/KIMODO_TRAINING_REPAIR.md + D docs/LOCAL_INDEX_TRAINING.md + D docs/MIGRATION.md + D docs/OMG_ENGLISH_VALIDATION.md + D docs/OMG_TRAINING_REVIEW.md + D docs/OMG_VALIDATION_DIAGNOSIS.md + D docs/OMG_VALIDATION_ROLLOUT.md + D docs/PORTABLE_TRAINING.md + D docs/REFERENCE_DYNAMIC_TRAINING.md + D docs/ROLLOUT_DEBUG.md + D docs/SEMANTICS_CLOCK_FIX.md + D docs/TEXTOP_ENGLISH_VALIDATION.md + D docs/TRAINING_REPRODUCTION_MATRIX.md + D docs/TRAINING_RESUME.md + D docs/WORKSPACE_CLEANUP.md + D docs/history/BASELINE_METHOD_BEFORE_SEMANTIC_FIX.md + D docs/history/DUPLEX_TASK.md + D docs/history/HFD_HUMAN_BASELINE_INPUT_PROTOCOL_RESEARCH_20261017.md + D docs/history/HFD_WINDOW_PROTOCOL.md + D docs/history/NORMAL_TASK.md + D docs/history/OMG_CONTEXT_DEBUG.md + D docs/history/README.md + D docs/history/VIDEO_SAMPLES.md + M hfd_bridge/genmo_prefetch.py + M hfd_bridge/genmo_training.py + M hfd_bridge/human_sessions.py + M hfd_bridge/kimodo.py + M hfd_bridge/loader.py + M hfd_bridge/omg_checks.py + M hfd_bridge/ondemand.py + M hfd_bridge/ondemand_data.py + M hfd_bridge/resume.py + M hfd_bridge/sessions.py + M hfd_bridge/storage.py + M prepare_sessions.py + M requirements/genmo.txt + M requirements/runtime-versions.json + M tools/configure_training.py + M train.py + M train_genmo_full.py + M train_kimodo.py +?? configs/hymotion.yaml +?? configs/hymotion_c24.yaml +?? configs/omg_l.yaml +?? configs/omg_xl.yaml +?? docs/CURRENT_RUNS.md +?? docs/HYMOTION_ASSETS.json +?? hfd_bridge/final_split.py +?? hfd_bridge/final_third_party.py +?? hfd_bridge/genmo_replay_audit.py +?? hfd_bridge/genmo_runtime.py +?? hfd_bridge/genmo_safety.py +?? hfd_bridge/hot_assets.py +?? hfd_bridge/hymotion_autoregressive.py +?? hfd_bridge/hymotion_data.py +?? hfd_bridge/hymotion_evaluation.py +?? hfd_bridge/hymotion_metrics.py +?? hfd_bridge/hymotion_model.py +?? hfd_bridge/hymotion_rep.py +?? hfd_bridge/hymotion_rollout.py +?? hfd_bridge/hymotion_text.py +?? hfd_bridge/hymotion_training.py +?? hfd_bridge/hymotion_validation.py +?? hfd_bridge/omg_scaling.py +?? hfd_bridge/replay_numerics.py +?? hfd_bridge/textop_text_cache.py +?? patches/ +?? requirements/hymotion.txt +?? tools/audit_genmo_replay.py +?? tools/download_hymotion_assets.py +?? tools/index_hymotion_text_cache.py +?? tools/prepare_hot_assets.py +?? train_hymotion.py +?? train_hymotion_autoregressive.py +?? upstream/hfd/src/psi/data/seed/third_party_dataset.py diff --git a/handoff-v1/inventory/textop-git-submodules.txt b/handoff-v1/inventory/textop-git-submodules.txt new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/handoff-v1/inventory/textop-historical-source.json b/handoff-v1/inventory/textop-historical-source.json new file mode 100644 index 0000000000000000000000000000000000000000..2b5b24db4711d1599ff8119fa0ee5395bd3f4851 --- /dev/null +++ b/handoff-v1/inventory/textop-historical-source.json @@ -0,0 +1,67 @@ +{ + "hfd_bridge/omg_checks.py": { + "expected": "1c80361412892538400cc225ab28aa03d5357ccab7a53e3a6d46deb5da9d32e8", + "recovered_commit": "ea269e3716e289e64fa3cec17f1366b57e3fab9b", + "status": "exact_historical_bytes" + }, + "hfd_bridge/ondemand.py": { + "expected": "ddf6c76a6b3d736de9dbf5e27c9d8d6e9384156cee68fc96bba0ce08237aa7c5", + "recovered_commit": null, + "status": "historical_bytes_missing_current_snapshot_provided" + }, + "hfd_bridge/resume.py": { + "expected": "e8f9c06323415cbdf011272efa4504f12e0c44079d4648a92277f27a5e5d0cc3", + "recovered_commit": "ea269e3716e289e64fa3cec17f1366b57e3fab9b", + "status": "exact_historical_bytes" + }, + "hfd_bridge/genmo_training.py": { + "expected": "8276d4f266b44e2c112e14021eb29c442753f2de07abf29e9adb321226dc87d8", + "recovered_commit": "ea269e3716e289e64fa3cec17f1366b57e3fab9b", + "status": "exact_historical_bytes" + }, + "hfd_bridge/genmo_prefetch.py": { + "expected": "3375352f7668dfd5c76ff629853aa8e4d0bcd2e2294a6e96932280f2eafa68e4", + "recovered_commit": "ea269e3716e289e64fa3cec17f1366b57e3fab9b", + "status": "exact_historical_bytes" + }, + "hfd_bridge/hymotion_model.py": { + "expected": "db0968384b87929c6495b87647a37e92e13729a94a34f6a5347f4d23ed05be5c", + "recovered_commit": null, + "status": "historical_bytes_missing_current_snapshot_provided" + }, + "hfd_bridge/hymotion_data.py": { + "expected": "06170526a18b60fa486f084864e2e65f2608274ed967ef367b29de269f8c9147", + "recovered_commit": null, + "status": "historical_bytes_missing_current_snapshot_provided" + }, + "hfd_bridge/hymotion_rollout.py": { + "expected": "a7fefe8d72f9e3614103632d68d1d63ce3ee072650739683131f5315d0a41181", + "recovered_commit": null, + "status": "historical_bytes_missing_current_snapshot_provided" + }, + "hfd_bridge/hymotion_validation.py": { + "expected": "cf6fd4b14fdbafc65a8a65abaade620b43d5b1f95e7a7a6194a21247f2ad4127", + "recovered_commit": null, + "status": "historical_bytes_missing_current_snapshot_provided" + }, + "hfd_bridge/hymotion_training.py": { + "expected": "67f40e79cc43950f6876f8c0c2c67fe1f606f1b67e692a7dd239457fa4b96a2c", + "recovered_commit": null, + "status": "historical_bytes_missing_current_snapshot_provided" + }, + "hfd_bridge/hymotion_clock.py": { + "expected": "1eaf969e9198f6902759457488ff8c80fc20c973b72388443b3ef0af7c8282a2", + "recovered_commit": null, + "status": "historical_bytes_missing_current_snapshot_provided" + }, + "hfd_bridge/hymotion_recovery.py": { + "expected": "e47b841f85dfc19b37a646a6cf6e865dd2642718073a3419a97195f16024c379", + "recovered_commit": null, + "status": "historical_bytes_missing_current_snapshot_provided" + }, + "train.py": { + "expected": "0724b3db6d374259db9f6378e9b663ef8438e5b969093f3b4d1e46ee56456ddc", + "recovered_commit": "ea269e3716e289e64fa3cec17f1366b57e3fab9b", + "status": "exact_historical_bytes" + } +} \ No newline at end of file diff --git a/handoff-v1/inventory/textop-latest-resume-source-audit.json b/handoff-v1/inventory/textop-latest-resume-source-audit.json new file mode 100644 index 0000000000000000000000000000000000000000..2a2994afc45edd795bf25b48357b45dc7c0ff2d7 --- /dev/null +++ b/handoff-v1/inventory/textop-latest-resume-source-audit.json @@ -0,0 +1,521 @@ +{ + "source": "/root/autodl-tmp/runs/hfd-textop-english-split20261026/resume_history/20261028T112850179940Z/code_sha256.json", + "files": { + "hfd_bridge/storage.py": { + "expected": "02c4fe4eed5f6d17bf140568af365b0abc8e5843df018d65553828890058a1fa", + "actual": "02c4fe4eed5f6d17bf140568af365b0abc8e5843df018d65553828890058a1fa" + }, + "hfd_bridge/__init__.py": { + "expected": "885528eb35015189a0bcd2dc62b60c61878efcf959caef68ff69e93501cafd5c", + "actual": "885528eb35015189a0bcd2dc62b60c61878efcf959caef68ff69e93501cafd5c" + }, + "hfd_bridge/loader.py": { + "expected": "13772137dc5cbb036bd5b5e4ee6ffe8cb45ba8d033ad580887b4dd63b253ac9b", + "actual": "13772137dc5cbb036bd5b5e4ee6ffe8cb45ba8d033ad580887b4dd63b253ac9b" + }, + "hfd_bridge/sessions.py": { + "expected": "c271aecf69e52acaee178b1381e465f22af82732f606d4e0bd5f891e2bb86076", + "actual": "c271aecf69e52acaee178b1381e465f22af82732f606d4e0bd5f891e2bb86076" + }, + "hfd_bridge/asr.py": { + "expected": "4baaf4faecfdce2054844a41326760220af301eb11d6fad892bed8c7380dceb0", + "actual": "4baaf4faecfdce2054844a41326760220af301eb11d6fad892bed8c7380dceb0" + }, + "hfd_bridge/training.py": { + "expected": "b513a632bcf010c83c23a35bcfd09e189149ff8c0d7d667d8966ec09e6e55970", + "actual": "b513a632bcf010c83c23a35bcfd09e189149ff8c0d7d667d8966ec09e6e55970" + }, + "hfd_bridge/native.py": { + "expected": "72ca9f675c2dfc7a2e304191da7b4d4016cb6bc226ccc22f9e7d0aeb2889c9f4", + "actual": "72ca9f675c2dfc7a2e304191da7b4d4016cb6bc226ccc22f9e7d0aeb2889c9f4" + }, + "hfd_bridge/paths.py": { + "expected": "2aaee82ecad571dcc1186c275e270af6c854c670db4ae354ee984d23a50b8bf3", + "actual": "2aaee82ecad571dcc1186c275e270af6c854c670db4ae354ee984d23a50b8bf3" + }, + "hfd_bridge/configured_data.py": { + "expected": "f916e0f9d4cdd2a4120fe94a9f1dfe59bbd24606d71ce1941023f04f97f181cd", + "actual": "f916e0f9d4cdd2a4120fe94a9f1dfe59bbd24606d71ce1941023f04f97f181cd" + }, + "hfd_bridge/textop_checks.py": { + "expected": "00662346781a638298cfa344b9ce8a58b6e8bba11261cd5783fdca956f4df7c8", + "actual": "00662346781a638298cfa344b9ce8a58b6e8bba11261cd5783fdca956f4df7c8" + }, + "hfd_bridge/omg_checks.py": { + "expected": "1c80361412892538400cc225ab28aa03d5357ccab7a53e3a6d46deb5da9d32e8", + "actual": "094d997043d4e0d868261b7f25b9452fca832dfa48572aeac4a3f2ba9f42d172" + }, + "hfd_bridge/textop_condition.py": { + "expected": "4fca0f4a02df3b843a908b28b64766877585aa3300977359d21fdb3adb5a3c28", + "actual": "4fca0f4a02df3b843a908b28b64766877585aa3300977359d21fdb3adb5a3c28" + }, + "hfd_bridge/joints.py": { + "expected": "7827c82ca3463a026db6007c41a66528e1294485b4acb081f27f09449da2d154", + "actual": "7827c82ca3463a026db6007c41a66528e1294485b4acb081f27f09449da2d154" + }, + "hfd_bridge/audio_condition.py": { + "expected": "cab6795bdddc36aa63d4f54c3c8e8aaee7958c86882abe35ae76f67a79b4b945", + "actual": "cab6795bdddc36aa63d4f54c3c8e8aaee7958c86882abe35ae76f67a79b4b945" + }, + "hfd_bridge/batch_asr.py": { + "expected": "9d4f0ba1d382f45c7a0adf562bc008485fce7cd00e86e29607528676f503f8f1", + "actual": "9d4f0ba1d382f45c7a0adf562bc008485fce7cd00e86e29607528676f503f8f1" + }, + "hfd_bridge/ondemand.py": { + "expected": "940339aa16a6fa25a27ace23a956017f7325813b0cd2a49f76e6f1fb1d014cc0", + "actual": "940339aa16a6fa25a27ace23a956017f7325813b0cd2a49f76e6f1fb1d014cc0" + }, + "hfd_bridge/ondemand_data.py": { + "expected": "24c61e36623eaca0bdc2f7bff2c8ce307946c4a5afef18f0427b96c06c9ed57e", + "actual": "24c61e36623eaca0bdc2f7bff2c8ce307946c4a5afef18f0427b96c06c9ed57e" + }, + "hfd_bridge/prefetch.py": { + "expected": "b227251cc6470f980231999bbecbeb1fc583c754eda56e70d4f3c84ed1845d64", + "actual": "b227251cc6470f980231999bbecbeb1fc583c754eda56e70d4f3c84ed1845d64" + }, + "hfd_bridge/human_sessions.py": { + "expected": "0461ea66b1c324ec38e9e833ee8eb19b95a24c3f478b094fb600dcdc3c53cdea", + "actual": "0461ea66b1c324ec38e9e833ee8eb19b95a24c3f478b094fb600dcdc3c53cdea" + }, + "hfd_bridge/genmo.py": { + "expected": "3c898d2828267f38f8ba9f41aae62c4a52cbae87cf63dd2cef2b6634f7160060", + "actual": "3c898d2828267f38f8ba9f41aae62c4a52cbae87cf63dd2cef2b6634f7160060" + }, + "hfd_bridge/genmo_audio.py": { + "expected": "1cdfa6ebc19a75af7540b2280f5dc9c3ecc584ed1c386c434de32e3777020c6f", + "actual": "1cdfa6ebc19a75af7540b2280f5dc9c3ecc584ed1c386c434de32e3777020c6f" + }, + "hfd_bridge/clocks.py": { + "expected": "27054c0b2d3e7d4087eb8fc3c59c6bcf5a581ca2e3022341e306732dcbe92637", + "actual": "27054c0b2d3e7d4087eb8fc3c59c6bcf5a581ca2e3022341e306732dcbe92637" + }, + "hfd_bridge/semantics.py": { + "expected": "42ed59d58609095b8053d0598a9c575d52af9b8508e699162743cc11599f03b4", + "actual": "42ed59d58609095b8053d0598a9c575d52af9b8508e699162743cc11599f03b4" + }, + "hfd_bridge/reference_text.py": { + "expected": "50da93fed3be2bad05035bf978c9be41ba7b48aaf3ce7c43a879519e7386a121", + "actual": "50da93fed3be2bad05035bf978c9be41ba7b48aaf3ce7c43a879519e7386a121" + }, + "hfd_bridge/reference_runtime.py": { + "expected": "2b7fe129152d0f7ea3a4a4e3f5fc9ae799e805138a0d76487e9f59065fe7d975", + "actual": "2b7fe129152d0f7ea3a4a4e3f5fc9ae799e805138a0d76487e9f59065fe7d975" + }, + "hfd_bridge/resume.py": { + "expected": "be1b665778f3c3bc137f4242b0ac4d1f587903357954725ed509e712dac0ab8f", + "actual": "e4e373802e0dd21e39fa2d899eff524080ce1735ca86eefbf4ef2080f170ffbf" + }, + "hfd_bridge/resume_lightning.py": { + "expected": "69e4c6bf11a199f3a979227eb551870c44e371fcae0af924ec64d2d58d3ce7a6", + "actual": "69e4c6bf11a199f3a979227eb551870c44e371fcae0af924ec64d2d58d3ce7a6" + }, + "hfd_bridge/index_cache.py": { + "expected": "d75cbcb0105448c61c841e525a1f17e63c55c83ad9201104e8118e22f7728b17", + "actual": "d75cbcb0105448c61c841e525a1f17e63c55c83ad9201104e8118e22f7728b17" + }, + "hfd_bridge/genmo_data.py": { + "expected": "4b102d1eecb0af9757a76d9600ede345c3f21f92166276a22cbf8e144fa9a38c", + "actual": "4b102d1eecb0af9757a76d9600ede345c3f21f92166276a22cbf8e144fa9a38c" + }, + "hfd_bridge/genmo_validation.py": { + "expected": "fa265cf9df51a4b36db977be945547dd78bdf105db9aed8ccba940b1a977f5e4", + "actual": "fa265cf9df51a4b36db977be945547dd78bdf105db9aed8ccba940b1a977f5e4" + }, + "hfd_bridge/kimodo_data.py": { + "expected": "35b3b7d8fc8f36fe65eb311e6e65e667fe1fdbcdfc23e58f37fb75c1c9e9feae", + "actual": "35b3b7d8fc8f36fe65eb311e6e65e667fe1fdbcdfc23e58f37fb75c1c9e9feae" + }, + "hfd_bridge/kimodo.py": { + "expected": "081a1047b853a2ee143aad2a5699f33cbecd0b4f93a07ad085760e16665f20ce", + "actual": "081a1047b853a2ee143aad2a5699f33cbecd0b4f93a07ad085760e16665f20ce" + }, + "hfd_bridge/kimodo_recovery.py": { + "expected": "b9177ff13b55b753430dd14d46778b723d2a03ce8927313c246c5d7ce780809b", + "actual": "b9177ff13b55b753430dd14d46778b723d2a03ce8927313c246c5d7ce780809b" + }, + "hfd_bridge/kimodo_closed_loop.py": { + "expected": "38d7e022b1450a715b35542cd23a70b1491a6dbd307dbb4be730b9f562100f91", + "actual": "38d7e022b1450a715b35542cd23a70b1491a6dbd307dbb4be730b9f562100f91" + }, + "hfd_bridge/kimodo_causal_rep.py": { + "expected": "c0c40003a3d36ec35e8490fa1165ac88723178b15a0e118d2d1d4c9f7ed7c537", + "actual": "c0c40003a3d36ec35e8490fa1165ac88723178b15a0e118d2d1d4c9f7ed7c537" + }, + "hfd_bridge/kimodo_diffusion_training.py": { + "expected": "078fe8cf1ee39d910a6ca5ef94d67f20c4398cfe505895a089acb80c05336539", + "actual": "078fe8cf1ee39d910a6ca5ef94d67f20c4398cfe505895a089acb80c05336539" + }, + "hfd_bridge/kimodo_text_worker.py": { + "expected": "a275c121a9364ae85b6233a67f3acb63ceddb2c9da490c2ea7d476f4cc710751", + "actual": "a275c121a9364ae85b6233a67f3acb63ceddb2c9da490c2ea7d476f4cc710751" + }, + "hfd_bridge/kimodo_full_data.py": { + "expected": "c62b4bb2673f1875792edb3403347f76bb52fa2360d92a9c1091eae10d43f905", + "actual": "c62b4bb2673f1875792edb3403347f76bb52fa2360d92a9c1091eae10d43f905" + }, + "hfd_bridge/kimodo_checkpoint.py": { + "expected": "b9cd24df225c91a6212c354983e3f51acb495373747f303c6096000af37f7d59", + "actual": "b9cd24df225c91a6212c354983e3f51acb495373747f303c6096000af37f7d59" + }, + "hfd_bridge/kimodo_finetuning.py": { + "expected": "a2c54f04af2b516c14d00025c8b695093ff09da6f04b0b73eb3269817bc0a544", + "actual": "a2c54f04af2b516c14d00025c8b695093ff09da6f04b0b73eb3269817bc0a544" + }, + "hfd_bridge/reference_assets.py": { + "expected": "e977a1d804807629fe925d511acdd25b488a11acc3767c0234020af8f699a43c", + "actual": "e977a1d804807629fe925d511acdd25b488a11acc3767c0234020af8f699a43c" + }, + "hfd_bridge/omg_model_loader.py": { + "expected": "f972080dd6140e1574124be615c9351001b51a73c5774b63bd1c063671bd56aa", + "actual": "f972080dd6140e1574124be615c9351001b51a73c5774b63bd1c063671bd56aa" + }, + "hfd_bridge/genmo_geometry.py": { + "expected": "4d3e912a858b8d3d4ddf73941c3120d5304ed70187c02fc969730816e155b2cb", + "actual": "4d3e912a858b8d3d4ddf73941c3120d5304ed70187c02fc969730816e155b2cb" + }, + "hfd_bridge/genmo_training.py": { + "expected": "8276d4f266b44e2c112e14021eb29c442753f2de07abf29e9adb321226dc87d8", + "actual": "556100608e466ca801abd623d24af30eebe7f9e3ac598c6709f813f07d5d45e9" + }, + "hfd_bridge/genmo_prefetch.py": { + "expected": "3375352f7668dfd5c76ff629853aa8e4d0bcd2e2294a6e96932280f2eafa68e4", + "actual": "0f00c9c5c951f53a0b55cad0c71d470c9312381e75ec31f756f606cf20309141" + }, + "hfd_bridge/hymotion_rep.py": { + "expected": "ee2f8282720d311b98d3ceb8cc72f63219622b5f61a138af6eb3221181a89f7a", + "actual": "ee2f8282720d311b98d3ceb8cc72f63219622b5f61a138af6eb3221181a89f7a" + }, + "hfd_bridge/hymotion_model.py": { + "expected": "db0968384b87929c6495b87647a37e92e13729a94a34f6a5347f4d23ed05be5c", + "actual": "ed29d2d03b247ed681a09dc230697885c5ab8fa8b30f82c4a827aa8fe0b41597" + }, + "hfd_bridge/hymotion_data.py": { + "expected": "06170526a18b60fa486f084864e2e65f2608274ed967ef367b29de269f8c9147", + "actual": "79503602a9a4de4f67931fa944a198d6f438b531a363ffbe545ecc4dd54f97e7" + }, + "hfd_bridge/hymotion_text.py": { + "expected": "f0238ffc3656e2a90a26360c5e9bc8c6de3d2d6be0af6e87787c83237914e98b", + "actual": "f0238ffc3656e2a90a26360c5e9bc8c6de3d2d6be0af6e87787c83237914e98b" + }, + "hfd_bridge/hymotion_rollout.py": { + "expected": "a7fefe8d72f9e3614103632d68d1d63ce3ee072650739683131f5315d0a41181", + "actual": "6e27c4e24bd42a211df7be2a0271337af4e658d66e5d303633912bbd40313517" + }, + "hfd_bridge/hymotion_validation.py": { + "expected": "cf6fd4b14fdbafc65a8a65abaade620b43d5b1f95e7a7a6194a21247f2ad4127", + "actual": "6043090f19315c9fde677168fed06dca1fb390b559a7b9668e0ba46799074cd2" + }, + "hfd_bridge/hymotion_training.py": { + "expected": "67f40e79cc43950f6876f8c0c2c67fe1f606f1b67e692a7dd239457fa4b96a2c", + "actual": "792b76cfc35f58b0c7c94b40a660f00c4971fff1d978ffd0c923b9f26a5fa733" + }, + "hfd_bridge/hymotion_clock.py": { + "expected": "1eaf969e9198f6902759457488ff8c80fc20c973b72388443b3ef0af7c8282a2", + "actual": null + }, + "hfd_bridge/hymotion_recovery.py": { + "expected": "e47b841f85dfc19b37a646a6cf6e865dd2642718073a3419a97195f16024c379", + "actual": null + }, + "hfd_bridge/human_eval.py": { + "expected": "c943c332315f16f41cb53beee8303f2d94a4b522e678cfd346e8a6c9ec203b71", + "actual": "c943c332315f16f41cb53beee8303f2d94a4b522e678cfd346e8a6c9ec203b71" + }, + "hfd_bridge/omg_rollout.py": { + "expected": "24ecdec4a2c698bd8cdc2b4ac6e1c0ad960c252b5a9f9a15d57a7716545fc4d7", + "actual": "24ecdec4a2c698bd8cdc2b4ac6e1c0ad960c252b5a9f9a15d57a7716545fc4d7" + }, + "hfd_bridge/final_split.py": { + "expected": "d0f0adef291d17d3b73477258b5de74bb8ed477ca1952c9b83ddb115efd34305", + "actual": "d0f0adef291d17d3b73477258b5de74bb8ed477ca1952c9b83ddb115efd34305" + }, + "hfd_bridge/final_third_party.py": { + "expected": "3926a0f72d419bb5006d8a62136a8f5ffd5ddca439c39dd7768f7cf03422d716", + "actual": "3926a0f72d419bb5006d8a62136a8f5ffd5ddca439c39dd7768f7cf03422d716" + }, + "hfd_bridge/replay_numerics.py": { + "expected": "827c70aa9b2308ac33b6fccee2c4dc88df58d6b89d902d8c215adfdf918ab231", + "actual": "827c70aa9b2308ac33b6fccee2c4dc88df58d6b89d902d8c215adfdf918ab231" + }, + "hfd_bridge/textop_text_cache.py": { + "expected": "3d5c8a187e461144aa97cc539661e6e90b475800379c0c096b243bcd337a0e38", + "actual": "3d5c8a187e461144aa97cc539661e6e90b475800379c0c096b243bcd337a0e38" + }, + "hfd_bridge/hot_assets.py": { + "expected": "43830324054a551ef4ed364ad45d9b2dfd8830b7dde2661f4404b1ffc8cabe10", + "actual": "43830324054a551ef4ed364ad45d9b2dfd8830b7dde2661f4404b1ffc8cabe10" + }, + "upstream/textop/TextOpRobotMDAR/robotmdar/__init__.py": { + "expected": "4e96c939ef598e8744521fedfda14febb8078e49c67fa89f620e287d7fa0cedc", + "actual": "4e96c939ef598e8744521fedfda14febb8078e49c67fa89f620e287d7fa0cedc" + }, + "upstream/textop/TextOpRobotMDAR/robotmdar/cli.py": { + "expected": "0c3f7017f3056b1a7895c1a3c3f6816e1fd90598c71caa42cd25f0d154e8c8ab", + "actual": "0c3f7017f3056b1a7895c1a3c3f6816e1fd90598c71caa42cd25f0d154e8c8ab" + }, + "upstream/textop/TextOpRobotMDAR/robotmdar/config/__init__.py": { + "expected": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", + "actual": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855" + }, + "upstream/textop/TextOpRobotMDAR/robotmdar/dataloader/__init__.py": { + "expected": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", + "actual": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855" + }, + "upstream/textop/TextOpRobotMDAR/robotmdar/dataloader/data.py": { + "expected": "4107f913690a857f175b807dc43d2991d6264f9b5b4956d890da7cb18c31cceb", + "actual": "4107f913690a857f175b807dc43d2991d6264f9b5b4956d890da7cb18c31cceb" + }, + "upstream/textop/TextOpRobotMDAR/robotmdar/diffusion/__init__.py": { + "expected": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", + "actual": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855" + }, + "upstream/textop/TextOpRobotMDAR/robotmdar/diffusion/fp16_util.py": { + "expected": "5ca8ca970755b1f12783da366290228945795216e9ec15ca7276174f7f6910fc", + "actual": "5ca8ca970755b1f12783da366290228945795216e9ec15ca7276174f7f6910fc" + }, + "upstream/textop/TextOpRobotMDAR/robotmdar/diffusion/gaussian_diffusion.py": { + "expected": "b8b426f0fb7bcff02d3fcbf5b665dd2265d9be1a7a0fe6634ac0ad3cd8864f9b", + "actual": "b8b426f0fb7bcff02d3fcbf5b665dd2265d9be1a7a0fe6634ac0ad3cd8864f9b" + }, + "upstream/textop/TextOpRobotMDAR/robotmdar/diffusion/logger.py": { + "expected": "615d9cb63dd2cfe017ac424bcc64340257c83eb9cb8e6b6913df4bd55831c1c0", + "actual": "615d9cb63dd2cfe017ac424bcc64340257c83eb9cb8e6b6913df4bd55831c1c0" + }, + "upstream/textop/TextOpRobotMDAR/robotmdar/diffusion/losses.py": { + "expected": "86a33188b9419c248e6f338fcb49ef45f2d59a1b9fb108966035f35d8c81b874", + "actual": "86a33188b9419c248e6f338fcb49ef45f2d59a1b9fb108966035f35d8c81b874" + }, + "upstream/textop/TextOpRobotMDAR/robotmdar/diffusion/nn.py": { + "expected": "0b742e6fa65c1bbd752467951b4f026ff1e8210f558601e653d3c65b5f83b517", + "actual": "0b742e6fa65c1bbd752467951b4f026ff1e8210f558601e653d3c65b5f83b517" + }, + "upstream/textop/TextOpRobotMDAR/robotmdar/diffusion/resample.py": { + "expected": "13a06f67e655dc145833896389576b0f8e68b2f677dc9eed60ed8839aa6a9c62", + "actual": "13a06f67e655dc145833896389576b0f8e68b2f677dc9eed60ed8839aa6a9c62" + }, + "upstream/textop/TextOpRobotMDAR/robotmdar/diffusion/respace.py": { + "expected": "778abccf75205dd186e13f1a422d2ab6a477aaeeef203846e2770ac8e3699a33", + "actual": "778abccf75205dd186e13f1a422d2ab6a477aaeeef203846e2770ac8e3699a33" + }, + "upstream/textop/TextOpRobotMDAR/robotmdar/dtype/__init__.py": { + "expected": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", + "actual": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855" + }, + "upstream/textop/TextOpRobotMDAR/robotmdar/dtype/abc.py": { + "expected": "3b292a44163cc5c382db1b19d15e240ca3ab3536a4fdede4b6d8e07c0aeba5ab", + "actual": "3b292a44163cc5c382db1b19d15e240ca3ab3536a4fdede4b6d8e07c0aeba5ab" + }, + "upstream/textop/TextOpRobotMDAR/robotmdar/dtype/debug.py": { + "expected": "44e4f52d6ecc92189f1bb444f882ecdef358729836a0cab625d358a11a3e7a4f", + "actual": "44e4f52d6ecc92189f1bb444f882ecdef358729836a0cab625d358a11a3e7a4f" + }, + "upstream/textop/TextOpRobotMDAR/robotmdar/dtype/device.py": { + "expected": "ef3cbab9467c5471c02ab886403ac7c7180a37adb36110652f77c7580ed05085", + "actual": "ef3cbab9467c5471c02ab886403ac7c7180a37adb36110652f77c7580ed05085" + }, + "upstream/textop/TextOpRobotMDAR/robotmdar/dtype/enum.py": { + "expected": "39225955eb12aa25c12594a6e40698d2c1787443e5421282f4f2900b4dc3ae56", + "actual": "39225955eb12aa25c12594a6e40698d2c1787443e5421282f4f2900b4dc3ae56" + }, + "upstream/textop/TextOpRobotMDAR/robotmdar/dtype/logger.py": { + "expected": "d72471467e1b1741b5b263ee335af9b02661d59485c3ef35c91f249d97345865", + "actual": "d72471467e1b1741b5b263ee335af9b02661d59485c3ef35c91f249d97345865" + }, + "upstream/textop/TextOpRobotMDAR/robotmdar/dtype/motion.py": { + "expected": "3dbe3832c309ebe1128c7137b51d444606b3f4f17d1e57f393141c0db8b3e168", + "actual": "3dbe3832c309ebe1128c7137b51d444606b3f4f17d1e57f393141c0db8b3e168" + }, + "upstream/textop/TextOpRobotMDAR/robotmdar/dtype/rotation.py": { + "expected": "733b5914adfe92efebc0f261d1124985f7221943dab7cce0799c712af47e77ba", + "actual": "733b5914adfe92efebc0f261d1124985f7221943dab7cce0799c712af47e77ba" + }, + "upstream/textop/TextOpRobotMDAR/robotmdar/dtype/seed.py": { + "expected": "1de944913a0d814a42e2224300c9ba716a02326eb473a3cf663b209414850675", + "actual": "1de944913a0d814a42e2224300c9ba716a02326eb473a3cf663b209414850675" + }, + "upstream/textop/TextOpRobotMDAR/robotmdar/dtype/vis_mjc.py": { + "expected": "5e4b7a56c45a424a09114c6e05d971f0e3e19135f1ec7b8bf420fd780e711c77", + "actual": "5e4b7a56c45a424a09114c6e05d971f0e3e19135f1ec7b8bf420fd780e711c77" + }, + "upstream/textop/TextOpRobotMDAR/robotmdar/eval/__init__.py": { + "expected": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", + "actual": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855" + }, + "upstream/textop/TextOpRobotMDAR/robotmdar/eval/freq_dar.py": { + "expected": "600d48e189e104da09a5764e32c3b1651dc130ef1505f5e45196b544e5c1b258", + "actual": "600d48e189e104da09a5764e32c3b1651dc130ef1505f5e45196b544e5c1b258" + }, + "upstream/textop/TextOpRobotMDAR/robotmdar/eval/generate_dar.py": { + "expected": "adcea83c85b55f38bc6e22ffd0430db246b22d42e0a27e37b2adff8f6f03da73", + "actual": "adcea83c85b55f38bc6e22ffd0430db246b22d42e0a27e37b2adff8f6f03da73" + }, + "upstream/textop/TextOpRobotMDAR/robotmdar/eval/loop_dar.py": { + "expected": "9c1ee5784432bc40408c5f8dcd6a1fd10ad31bafeed05373d51e506ca4a4c1b4", + "actual": "9c1ee5784432bc40408c5f8dcd6a1fd10ad31bafeed05373d51e506ca4a4c1b4" + }, + "upstream/textop/TextOpRobotMDAR/robotmdar/eval/vis_dar.py": { + "expected": "c4b2794a4e4a16e8eddc7efef68e50d10a981a565b5c8f6405486eb79bef7d88", + "actual": "c4b2794a4e4a16e8eddc7efef68e50d10a981a565b5c8f6405486eb79bef7d88" + }, + "upstream/textop/TextOpRobotMDAR/robotmdar/eval/vis_mvae.py": { + "expected": "78dde051a7250a33c93a3eba541b6074e3609ef4cd18c1a476d970802f431a32", + "actual": "78dde051a7250a33c93a3eba541b6074e3609ef4cd18c1a476d970802f431a32" + }, + "upstream/textop/TextOpRobotMDAR/robotmdar/model/__init__.py": { + "expected": "6a3222ab321dcc09cca255712c3a4a943a9347216e89459b6e2dc15c34fe9e22", + "actual": "6a3222ab321dcc09cca255712c3a4a943a9347216e89459b6e2dc15c34fe9e22" + }, + "upstream/textop/TextOpRobotMDAR/robotmdar/model/clip.py": { + "expected": "e9ef0eca2db8d8873737495ac2e9cd3f0f1b71caba63a732bf31b3cb135fac65", + "actual": "e9ef0eca2db8d8873737495ac2e9cd3f0f1b71caba63a732bf31b3cb135fac65" + }, + "upstream/textop/TextOpRobotMDAR/robotmdar/model/mld_denoiser.py": { + "expected": "a7b7d72347c125b5cf413e9c25c90723b1ffc4da2389d1dfc7ae4633ab74c687", + "actual": "a7b7d72347c125b5cf413e9c25c90723b1ffc4da2389d1dfc7ae4633ab74c687" + }, + "upstream/textop/TextOpRobotMDAR/robotmdar/model/mld_vae.py": { + "expected": "6be6b305c74d7cd7a8abacaf01260893412ad46df8f4e00f1ab48413e2c4f508", + "actual": "6be6b305c74d7cd7a8abacaf01260893412ad46df8f4e00f1ab48413e2c4f508" + }, + "upstream/textop/TextOpRobotMDAR/robotmdar/model/modules.py": { + "expected": "62ae3ae9a630f7401baa70b09b0b8931cba74f8f67a33131378658e8bc30e34c", + "actual": "62ae3ae9a630f7401baa70b09b0b8931cba74f8f67a33131378658e8bc30e34c" + }, + "upstream/textop/TextOpRobotMDAR/robotmdar/model/operator/__init__.py": { + "expected": "d5d517bc233a1c243330fd46af85e230b365235635e1f60876ca964cbca766dc", + "actual": "d5d517bc233a1c243330fd46af85e230b365235635e1f60876ca964cbca766dc" + }, + "upstream/textop/TextOpRobotMDAR/robotmdar/model/operator/adain.py": { + "expected": "3f7c3a64a877c2c978b400f66eac1d60b769c68d0f611318bf305558f88e2de4", + "actual": "3f7c3a64a877c2c978b400f66eac1d60b769c68d0f611318bf305558f88e2de4" + }, + "upstream/textop/TextOpRobotMDAR/robotmdar/model/operator/blocks.py": { + "expected": "931c1af282a5362a57ce9ba66ea0342e901fca71c7f2a784d2f7b7283dc0b128", + "actual": "931c1af282a5362a57ce9ba66ea0342e901fca71c7f2a784d2f7b7283dc0b128" + }, + "upstream/textop/TextOpRobotMDAR/robotmdar/model/operator/conv2d_gradfix.py": { + "expected": "ecf394964eb5cf04c5503a481c2534e6ebfe6908a9a196a2e1e3b72057269e66", + "actual": "ecf394964eb5cf04c5503a481c2534e6ebfe6908a9a196a2e1e3b72057269e66" + }, + "upstream/textop/TextOpRobotMDAR/robotmdar/model/operator/cross_attention.py": { + "expected": "179b2f043871f54ad66965154c71f9aa3384366b4a90fe58326875ed21fdfb3e", + "actual": "179b2f043871f54ad66965154c71f9aa3384366b4a90fe58326875ed21fdfb3e" + }, + "upstream/textop/TextOpRobotMDAR/robotmdar/model/operator/position_encoding.py": { + "expected": "914306fefb39cb05d3c0f7bb397e2a5a8dcd55daf681b47a37201ed1a748726b", + "actual": "914306fefb39cb05d3c0f7bb397e2a5a8dcd55daf681b47a37201ed1a748726b" + }, + "upstream/textop/TextOpRobotMDAR/robotmdar/model/operator/position_encoding_layer.py": { + "expected": "c0516d4da7f20cda5bfe25edfdb3887a61f8c2296f4eeba0cb9b8cc26b860b5c", + "actual": "c0516d4da7f20cda5bfe25edfdb3887a61f8c2296f4eeba0cb9b8cc26b860b5c" + }, + "upstream/textop/TextOpRobotMDAR/robotmdar/skeleton/__init__.py": { + "expected": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", + "actual": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855" + }, + "upstream/textop/TextOpRobotMDAR/robotmdar/skeleton/forward_kinematics.py": { + "expected": "f27c1e790862db65f7e901f64f4ad7c103976ce05e9db06c9f2a4008d9c8d9e3", + "actual": "f27c1e790862db65f7e901f64f4ad7c103976ce05e9db06c9f2a4008d9c8d9e3" + }, + "upstream/textop/TextOpRobotMDAR/robotmdar/skeleton/robot.py": { + "expected": "b26226ac8c9fca0e59fd8888caa907c88fac109d3e50df63b699f807a1967768", + "actual": "b26226ac8c9fca0e59fd8888caa907c88fac109d3e50df63b699f807a1967768" + }, + "upstream/textop/TextOpRobotMDAR/robotmdar/train/__init__.py": { + "expected": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", + "actual": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855" + }, + "upstream/textop/TextOpRobotMDAR/robotmdar/train/manager.py": { + "expected": "06572c88fa0c584e56c44205828f6efbcdf7e48328b6b94840a3661d898b08eb", + "actual": "06572c88fa0c584e56c44205828f6efbcdf7e48328b6b94840a3661d898b08eb" + }, + "upstream/textop/TextOpRobotMDAR/robotmdar/train/train_dar.py": { + "expected": "0ca27402bc4a0f7732e9077756d66e74900fccc784730b9bf078de5a58ad929f", + "actual": "0ca27402bc4a0f7732e9077756d66e74900fccc784730b9bf078de5a58ad929f" + }, + "upstream/textop/TextOpRobotMDAR/robotmdar/train/train_mvae.py": { + "expected": "471fb1559b02d5278b8fe148809a46855949c3c18f58afc6adeefa2c52070215", + "actual": "471fb1559b02d5278b8fe148809a46855949c3c18f58afc6adeefa2c52070215" + }, + "upstream/textop/TextOpRobotMDAR/robotmdar/train/train_platforms.py": { + "expected": "edfdf9d9984618b6fb1c7583c8c5eb3d74835e83bfd1f47b6e4b404d36e99a3f", + "actual": "edfdf9d9984618b6fb1c7583c8c5eb3d74835e83bfd1f47b6e4b404d36e99a3f" + }, + "upstream/textop/TextOpRobotMDAR/robotmdar/wrapper/__init__.py": { + "expected": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", + "actual": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855" + }, + "upstream/textop/TextOpRobotMDAR/robotmdar/wrapper/vae_decode.py": { + "expected": "e45b7ebffeb9600f0a31549eb87e7c2b018fb59fc196782ae9b65045d9dedaca", + "actual": "e45b7ebffeb9600f0a31549eb87e7c2b018fb59fc196782ae9b65045d9dedaca" + }, + "upstream/textop/dataset/cal_action_statistics.py": { + "expected": "866aa29fae28f1d5c63ea3865afbdd96594423b187d62bf4e1f404c5083b1b5e", + "actual": "866aa29fae28f1d5c63ea3865afbdd96594423b187d62bf4e1f404c5083b1b5e" + }, + "upstream/textop/dataset/motion_package.py": { + "expected": "9eefabf13b0edaa4838a558a055d219e6128be95d0c2a2b89c335308476b5a0f", + "actual": "9eefabf13b0edaa4838a558a055d219e6128be95d0c2a2b89c335308476b5a0f" + }, + "upstream/textop/dataset/pack_dataset.py": { + "expected": "54a039ff872358f50044844186957122f52bce00b6a3561f1d5ea67992b7a59e", + "actual": "54a039ff872358f50044844186957122f52bce00b6a3561f1d5ea67992b7a59e" + }, + "upstream/textop/dataset/process_retarget_data.py": { + "expected": "ba8bb84049a237149985a87ca241e3f9c8a67fa079ff03c5f4da3e6f0a6b07c4", + "actual": "ba8bb84049a237149985a87ca241e3f9c8a67fa079ff03c5f4da3e6f0a6b07c4" + }, + "upstream/textop/dataset/smplx_to_robot_dataset.py": { + "expected": "48c29987fc8e0673ad0253825d23a5bb90a6e576e348a4897b84703d93759415", + "actual": "48c29987fc8e0673ad0253825d23a5bb90a6e576e348a4897b84703d93759415" + }, + "upstream/textop/deps/isaac_utils/setup.py": { + "expected": "4022149687d6dd29680f0b00442e3d455c284e6001fce44535f07a9a1cbdb18f", + "actual": "4022149687d6dd29680f0b00442e3d455c284e6001fce44535f07a9a1cbdb18f" + }, + "upstream/textop/deps/isaac_utils/isaac_utils/__init__.py": { + "expected": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", + "actual": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855" + }, + "upstream/textop/deps/isaac_utils/isaac_utils/maths.py": { + "expected": "f44b9edded4433995d54fc3445b3d9613a018416437e2a3e5aadf779d5f9a8fb", + "actual": "f44b9edded4433995d54fc3445b3d9613a018416437e2a3e5aadf779d5f9a8fb" + }, + "upstream/textop/deps/isaac_utils/isaac_utils/rotations.py": { + "expected": "8233c8c7ea271ca9d4fba0ebfe3ad0f3138ee0e483b15397bbeb720f0f0c3a02", + "actual": "8233c8c7ea271ca9d4fba0ebfe3ad0f3138ee0e483b15397bbeb720f0f0c3a02" + }, + "upstream/textop/deps/CLIP/hubconf.py": { + "expected": "0641946610c17a743ebcc3f6e85cffff1400337bb9a99766740c40b01d9a3e14", + "actual": "0641946610c17a743ebcc3f6e85cffff1400337bb9a99766740c40b01d9a3e14" + }, + "upstream/textop/deps/CLIP/setup.py": { + "expected": "6a913890284a988293f7b0b3ca017de29c65ce2bdce9fdd09cb481fc6bd5a636", + "actual": "6a913890284a988293f7b0b3ca017de29c65ce2bdce9fdd09cb481fc6bd5a636" + }, + "upstream/textop/deps/CLIP/clip/__init__.py": { + "expected": "d7930eb290ebed432116795c91473200d098451a94c0aedd4ae0f934f97955e8", + "actual": "d7930eb290ebed432116795c91473200d098451a94c0aedd4ae0f934f97955e8" + }, + "upstream/textop/deps/CLIP/clip/clip.py": { + "expected": "9540f200fbf8145479fa655382a56dab048d238cc698b9cbd8df3b6d86d3f1b6", + "actual": "9540f200fbf8145479fa655382a56dab048d238cc698b9cbd8df3b6d86d3f1b6" + }, + "upstream/textop/deps/CLIP/clip/model.py": { + "expected": "9902cbe5ee90a1da2aa3e6f043e8a23dc1f8831193b963785c9af03d5c7bef2c", + "actual": "9902cbe5ee90a1da2aa3e6f043e8a23dc1f8831193b963785c9af03d5c7bef2c" + }, + "upstream/textop/deps/CLIP/clip/simple_tokenizer.py": { + "expected": "d1b09f10ed3d1e2c343619cb98cbfacb92363dd8607526a43aa7fe2d60d15bad", + "actual": "d1b09f10ed3d1e2c343619cb98cbfacb92363dd8607526a43aa7fe2d60d15bad" + }, + "upstream/textop/deps/CLIP/tests/test_consistency.py": { + "expected": "342f84e4814143e1117dbc14baa6cf2a2ec01f61542e3952edc91e68f5eed206", + "actual": "342f84e4814143e1117dbc14baa6cf2a2ec01f61542e3952edc91e68f5eed206" + }, + "train.py": { + "expected": "0724b3db6d374259db9f6378e9b663ef8438e5b969093f3b4d1e46ee56456ddc", + "actual": "61f49739e14275a671dc3343afcf2d6b225da547a952f963037d545afe680dd4" + } + } +} \ No newline at end of file diff --git a/handoff-v1/inventory/textop-runtime.txt b/handoff-v1/inventory/textop-runtime.txt new file mode 100644 index 0000000000000000000000000000000000000000..5401cdf377df1c55867a02d9f666e448d8b4746d --- /dev/null +++ b/handoff-v1/inventory/textop-runtime.txt @@ -0,0 +1 @@ +{"python": "3.10.21 (main, Aug 27 2026, 14:42:07) [GCC 14.3.0]", "torch": "2.7.1+cu128", "cuda": "12.8", "cudnn": 90701} diff --git a/handoff-v1/inventory/textop-source-audit.json b/handoff-v1/inventory/textop-source-audit.json new file mode 100644 index 0000000000000000000000000000000000000000..1d37d1692660091754767865c86d4d49c13dc10e --- /dev/null +++ b/handoff-v1/inventory/textop-source-audit.json @@ -0,0 +1,518 @@ +{ + "hfd_bridge/storage.py": { + "expected": "02c4fe4eed5f6d17bf140568af365b0abc8e5843df018d65553828890058a1fa", + "actual": "02c4fe4eed5f6d17bf140568af365b0abc8e5843df018d65553828890058a1fa" + }, + "hfd_bridge/__init__.py": { + "expected": "885528eb35015189a0bcd2dc62b60c61878efcf959caef68ff69e93501cafd5c", + "actual": "885528eb35015189a0bcd2dc62b60c61878efcf959caef68ff69e93501cafd5c" + }, + "hfd_bridge/loader.py": { + "expected": "13772137dc5cbb036bd5b5e4ee6ffe8cb45ba8d033ad580887b4dd63b253ac9b", + "actual": "13772137dc5cbb036bd5b5e4ee6ffe8cb45ba8d033ad580887b4dd63b253ac9b" + }, + "hfd_bridge/sessions.py": { + "expected": "c271aecf69e52acaee178b1381e465f22af82732f606d4e0bd5f891e2bb86076", + "actual": "c271aecf69e52acaee178b1381e465f22af82732f606d4e0bd5f891e2bb86076" + }, + "hfd_bridge/asr.py": { + "expected": "4baaf4faecfdce2054844a41326760220af301eb11d6fad892bed8c7380dceb0", + "actual": "4baaf4faecfdce2054844a41326760220af301eb11d6fad892bed8c7380dceb0" + }, + "hfd_bridge/training.py": { + "expected": "b513a632bcf010c83c23a35bcfd09e189149ff8c0d7d667d8966ec09e6e55970", + "actual": "b513a632bcf010c83c23a35bcfd09e189149ff8c0d7d667d8966ec09e6e55970" + }, + "hfd_bridge/native.py": { + "expected": "72ca9f675c2dfc7a2e304191da7b4d4016cb6bc226ccc22f9e7d0aeb2889c9f4", + "actual": "72ca9f675c2dfc7a2e304191da7b4d4016cb6bc226ccc22f9e7d0aeb2889c9f4" + }, + "hfd_bridge/paths.py": { + "expected": "2aaee82ecad571dcc1186c275e270af6c854c670db4ae354ee984d23a50b8bf3", + "actual": "2aaee82ecad571dcc1186c275e270af6c854c670db4ae354ee984d23a50b8bf3" + }, + "hfd_bridge/configured_data.py": { + "expected": "f916e0f9d4cdd2a4120fe94a9f1dfe59bbd24606d71ce1941023f04f97f181cd", + "actual": "f916e0f9d4cdd2a4120fe94a9f1dfe59bbd24606d71ce1941023f04f97f181cd" + }, + "hfd_bridge/textop_checks.py": { + "expected": "00662346781a638298cfa344b9ce8a58b6e8bba11261cd5783fdca956f4df7c8", + "actual": "00662346781a638298cfa344b9ce8a58b6e8bba11261cd5783fdca956f4df7c8" + }, + "hfd_bridge/omg_checks.py": { + "expected": "1c80361412892538400cc225ab28aa03d5357ccab7a53e3a6d46deb5da9d32e8", + "actual": "094d997043d4e0d868261b7f25b9452fca832dfa48572aeac4a3f2ba9f42d172" + }, + "hfd_bridge/textop_condition.py": { + "expected": "4fca0f4a02df3b843a908b28b64766877585aa3300977359d21fdb3adb5a3c28", + "actual": "4fca0f4a02df3b843a908b28b64766877585aa3300977359d21fdb3adb5a3c28" + }, + "hfd_bridge/joints.py": { + "expected": "7827c82ca3463a026db6007c41a66528e1294485b4acb081f27f09449da2d154", + "actual": "7827c82ca3463a026db6007c41a66528e1294485b4acb081f27f09449da2d154" + }, + "hfd_bridge/audio_condition.py": { + "expected": "cab6795bdddc36aa63d4f54c3c8e8aaee7958c86882abe35ae76f67a79b4b945", + "actual": "cab6795bdddc36aa63d4f54c3c8e8aaee7958c86882abe35ae76f67a79b4b945" + }, + "hfd_bridge/batch_asr.py": { + "expected": "9d4f0ba1d382f45c7a0adf562bc008485fce7cd00e86e29607528676f503f8f1", + "actual": "9d4f0ba1d382f45c7a0adf562bc008485fce7cd00e86e29607528676f503f8f1" + }, + "hfd_bridge/ondemand.py": { + "expected": "ddf6c76a6b3d736de9dbf5e27c9d8d6e9384156cee68fc96bba0ce08237aa7c5", + "actual": "940339aa16a6fa25a27ace23a956017f7325813b0cd2a49f76e6f1fb1d014cc0" + }, + "hfd_bridge/ondemand_data.py": { + "expected": "24c61e36623eaca0bdc2f7bff2c8ce307946c4a5afef18f0427b96c06c9ed57e", + "actual": "24c61e36623eaca0bdc2f7bff2c8ce307946c4a5afef18f0427b96c06c9ed57e" + }, + "hfd_bridge/prefetch.py": { + "expected": "b227251cc6470f980231999bbecbeb1fc583c754eda56e70d4f3c84ed1845d64", + "actual": "b227251cc6470f980231999bbecbeb1fc583c754eda56e70d4f3c84ed1845d64" + }, + "hfd_bridge/human_sessions.py": { + "expected": "0461ea66b1c324ec38e9e833ee8eb19b95a24c3f478b094fb600dcdc3c53cdea", + "actual": "0461ea66b1c324ec38e9e833ee8eb19b95a24c3f478b094fb600dcdc3c53cdea" + }, + "hfd_bridge/genmo.py": { + "expected": "3c898d2828267f38f8ba9f41aae62c4a52cbae87cf63dd2cef2b6634f7160060", + "actual": "3c898d2828267f38f8ba9f41aae62c4a52cbae87cf63dd2cef2b6634f7160060" + }, + "hfd_bridge/genmo_audio.py": { + "expected": "1cdfa6ebc19a75af7540b2280f5dc9c3ecc584ed1c386c434de32e3777020c6f", + "actual": "1cdfa6ebc19a75af7540b2280f5dc9c3ecc584ed1c386c434de32e3777020c6f" + }, + "hfd_bridge/clocks.py": { + "expected": "27054c0b2d3e7d4087eb8fc3c59c6bcf5a581ca2e3022341e306732dcbe92637", + "actual": "27054c0b2d3e7d4087eb8fc3c59c6bcf5a581ca2e3022341e306732dcbe92637" + }, + "hfd_bridge/semantics.py": { + "expected": "42ed59d58609095b8053d0598a9c575d52af9b8508e699162743cc11599f03b4", + "actual": "42ed59d58609095b8053d0598a9c575d52af9b8508e699162743cc11599f03b4" + }, + "hfd_bridge/reference_text.py": { + "expected": "50da93fed3be2bad05035bf978c9be41ba7b48aaf3ce7c43a879519e7386a121", + "actual": "50da93fed3be2bad05035bf978c9be41ba7b48aaf3ce7c43a879519e7386a121" + }, + "hfd_bridge/reference_runtime.py": { + "expected": "2b7fe129152d0f7ea3a4a4e3f5fc9ae799e805138a0d76487e9f59065fe7d975", + "actual": "2b7fe129152d0f7ea3a4a4e3f5fc9ae799e805138a0d76487e9f59065fe7d975" + }, + "hfd_bridge/resume.py": { + "expected": "e8f9c06323415cbdf011272efa4504f12e0c44079d4648a92277f27a5e5d0cc3", + "actual": "e4e373802e0dd21e39fa2d899eff524080ce1735ca86eefbf4ef2080f170ffbf" + }, + "hfd_bridge/resume_lightning.py": { + "expected": "69e4c6bf11a199f3a979227eb551870c44e371fcae0af924ec64d2d58d3ce7a6", + "actual": "69e4c6bf11a199f3a979227eb551870c44e371fcae0af924ec64d2d58d3ce7a6" + }, + "hfd_bridge/index_cache.py": { + "expected": "d75cbcb0105448c61c841e525a1f17e63c55c83ad9201104e8118e22f7728b17", + "actual": "d75cbcb0105448c61c841e525a1f17e63c55c83ad9201104e8118e22f7728b17" + }, + "hfd_bridge/genmo_data.py": { + "expected": "4b102d1eecb0af9757a76d9600ede345c3f21f92166276a22cbf8e144fa9a38c", + "actual": "4b102d1eecb0af9757a76d9600ede345c3f21f92166276a22cbf8e144fa9a38c" + }, + "hfd_bridge/genmo_validation.py": { + "expected": "fa265cf9df51a4b36db977be945547dd78bdf105db9aed8ccba940b1a977f5e4", + "actual": "fa265cf9df51a4b36db977be945547dd78bdf105db9aed8ccba940b1a977f5e4" + }, + "hfd_bridge/kimodo_data.py": { + "expected": "35b3b7d8fc8f36fe65eb311e6e65e667fe1fdbcdfc23e58f37fb75c1c9e9feae", + "actual": "35b3b7d8fc8f36fe65eb311e6e65e667fe1fdbcdfc23e58f37fb75c1c9e9feae" + }, + "hfd_bridge/kimodo.py": { + "expected": "081a1047b853a2ee143aad2a5699f33cbecd0b4f93a07ad085760e16665f20ce", + "actual": "081a1047b853a2ee143aad2a5699f33cbecd0b4f93a07ad085760e16665f20ce" + }, + "hfd_bridge/kimodo_recovery.py": { + "expected": "b9177ff13b55b753430dd14d46778b723d2a03ce8927313c246c5d7ce780809b", + "actual": "b9177ff13b55b753430dd14d46778b723d2a03ce8927313c246c5d7ce780809b" + }, + "hfd_bridge/kimodo_closed_loop.py": { + "expected": "38d7e022b1450a715b35542cd23a70b1491a6dbd307dbb4be730b9f562100f91", + "actual": "38d7e022b1450a715b35542cd23a70b1491a6dbd307dbb4be730b9f562100f91" + }, + "hfd_bridge/kimodo_causal_rep.py": { + "expected": "c0c40003a3d36ec35e8490fa1165ac88723178b15a0e118d2d1d4c9f7ed7c537", + "actual": "c0c40003a3d36ec35e8490fa1165ac88723178b15a0e118d2d1d4c9f7ed7c537" + }, + "hfd_bridge/kimodo_diffusion_training.py": { + "expected": "078fe8cf1ee39d910a6ca5ef94d67f20c4398cfe505895a089acb80c05336539", + "actual": "078fe8cf1ee39d910a6ca5ef94d67f20c4398cfe505895a089acb80c05336539" + }, + "hfd_bridge/kimodo_text_worker.py": { + "expected": "a275c121a9364ae85b6233a67f3acb63ceddb2c9da490c2ea7d476f4cc710751", + "actual": "a275c121a9364ae85b6233a67f3acb63ceddb2c9da490c2ea7d476f4cc710751" + }, + "hfd_bridge/kimodo_full_data.py": { + "expected": "c62b4bb2673f1875792edb3403347f76bb52fa2360d92a9c1091eae10d43f905", + "actual": "c62b4bb2673f1875792edb3403347f76bb52fa2360d92a9c1091eae10d43f905" + }, + "hfd_bridge/kimodo_checkpoint.py": { + "expected": "b9cd24df225c91a6212c354983e3f51acb495373747f303c6096000af37f7d59", + "actual": "b9cd24df225c91a6212c354983e3f51acb495373747f303c6096000af37f7d59" + }, + "hfd_bridge/kimodo_finetuning.py": { + "expected": "a2c54f04af2b516c14d00025c8b695093ff09da6f04b0b73eb3269817bc0a544", + "actual": "a2c54f04af2b516c14d00025c8b695093ff09da6f04b0b73eb3269817bc0a544" + }, + "hfd_bridge/reference_assets.py": { + "expected": "e977a1d804807629fe925d511acdd25b488a11acc3767c0234020af8f699a43c", + "actual": "e977a1d804807629fe925d511acdd25b488a11acc3767c0234020af8f699a43c" + }, + "hfd_bridge/omg_model_loader.py": { + "expected": "f972080dd6140e1574124be615c9351001b51a73c5774b63bd1c063671bd56aa", + "actual": "f972080dd6140e1574124be615c9351001b51a73c5774b63bd1c063671bd56aa" + }, + "hfd_bridge/genmo_geometry.py": { + "expected": "4d3e912a858b8d3d4ddf73941c3120d5304ed70187c02fc969730816e155b2cb", + "actual": "4d3e912a858b8d3d4ddf73941c3120d5304ed70187c02fc969730816e155b2cb" + }, + "hfd_bridge/genmo_training.py": { + "expected": "8276d4f266b44e2c112e14021eb29c442753f2de07abf29e9adb321226dc87d8", + "actual": "556100608e466ca801abd623d24af30eebe7f9e3ac598c6709f813f07d5d45e9" + }, + "hfd_bridge/genmo_prefetch.py": { + "expected": "3375352f7668dfd5c76ff629853aa8e4d0bcd2e2294a6e96932280f2eafa68e4", + "actual": "0f00c9c5c951f53a0b55cad0c71d470c9312381e75ec31f756f606cf20309141" + }, + "hfd_bridge/hymotion_rep.py": { + "expected": "ee2f8282720d311b98d3ceb8cc72f63219622b5f61a138af6eb3221181a89f7a", + "actual": "ee2f8282720d311b98d3ceb8cc72f63219622b5f61a138af6eb3221181a89f7a" + }, + "hfd_bridge/hymotion_model.py": { + "expected": "db0968384b87929c6495b87647a37e92e13729a94a34f6a5347f4d23ed05be5c", + "actual": "ed29d2d03b247ed681a09dc230697885c5ab8fa8b30f82c4a827aa8fe0b41597" + }, + "hfd_bridge/hymotion_data.py": { + "expected": "06170526a18b60fa486f084864e2e65f2608274ed967ef367b29de269f8c9147", + "actual": "79503602a9a4de4f67931fa944a198d6f438b531a363ffbe545ecc4dd54f97e7" + }, + "hfd_bridge/hymotion_text.py": { + "expected": "f0238ffc3656e2a90a26360c5e9bc8c6de3d2d6be0af6e87787c83237914e98b", + "actual": "f0238ffc3656e2a90a26360c5e9bc8c6de3d2d6be0af6e87787c83237914e98b" + }, + "hfd_bridge/hymotion_rollout.py": { + "expected": "a7fefe8d72f9e3614103632d68d1d63ce3ee072650739683131f5315d0a41181", + "actual": "6e27c4e24bd42a211df7be2a0271337af4e658d66e5d303633912bbd40313517" + }, + "hfd_bridge/hymotion_validation.py": { + "expected": "cf6fd4b14fdbafc65a8a65abaade620b43d5b1f95e7a7a6194a21247f2ad4127", + "actual": "6043090f19315c9fde677168fed06dca1fb390b559a7b9668e0ba46799074cd2" + }, + "hfd_bridge/hymotion_training.py": { + "expected": "67f40e79cc43950f6876f8c0c2c67fe1f606f1b67e692a7dd239457fa4b96a2c", + "actual": "792b76cfc35f58b0c7c94b40a660f00c4971fff1d978ffd0c923b9f26a5fa733" + }, + "hfd_bridge/hymotion_clock.py": { + "expected": "1eaf969e9198f6902759457488ff8c80fc20c973b72388443b3ef0af7c8282a2", + "actual": null + }, + "hfd_bridge/hymotion_recovery.py": { + "expected": "e47b841f85dfc19b37a646a6cf6e865dd2642718073a3419a97195f16024c379", + "actual": null + }, + "hfd_bridge/human_eval.py": { + "expected": "c943c332315f16f41cb53beee8303f2d94a4b522e678cfd346e8a6c9ec203b71", + "actual": "c943c332315f16f41cb53beee8303f2d94a4b522e678cfd346e8a6c9ec203b71" + }, + "hfd_bridge/omg_rollout.py": { + "expected": "24ecdec4a2c698bd8cdc2b4ac6e1c0ad960c252b5a9f9a15d57a7716545fc4d7", + "actual": "24ecdec4a2c698bd8cdc2b4ac6e1c0ad960c252b5a9f9a15d57a7716545fc4d7" + }, + "hfd_bridge/final_split.py": { + "expected": "d0f0adef291d17d3b73477258b5de74bb8ed477ca1952c9b83ddb115efd34305", + "actual": "d0f0adef291d17d3b73477258b5de74bb8ed477ca1952c9b83ddb115efd34305" + }, + "hfd_bridge/final_third_party.py": { + "expected": "3926a0f72d419bb5006d8a62136a8f5ffd5ddca439c39dd7768f7cf03422d716", + "actual": "3926a0f72d419bb5006d8a62136a8f5ffd5ddca439c39dd7768f7cf03422d716" + }, + "hfd_bridge/replay_numerics.py": { + "expected": "827c70aa9b2308ac33b6fccee2c4dc88df58d6b89d902d8c215adfdf918ab231", + "actual": "827c70aa9b2308ac33b6fccee2c4dc88df58d6b89d902d8c215adfdf918ab231" + }, + "hfd_bridge/textop_text_cache.py": { + "expected": "3d5c8a187e461144aa97cc539661e6e90b475800379c0c096b243bcd337a0e38", + "actual": "3d5c8a187e461144aa97cc539661e6e90b475800379c0c096b243bcd337a0e38" + }, + "hfd_bridge/hot_assets.py": { + "expected": "43830324054a551ef4ed364ad45d9b2dfd8830b7dde2661f4404b1ffc8cabe10", + "actual": "43830324054a551ef4ed364ad45d9b2dfd8830b7dde2661f4404b1ffc8cabe10" + }, + "upstream/textop/TextOpRobotMDAR/robotmdar/__init__.py": { + "expected": "4e96c939ef598e8744521fedfda14febb8078e49c67fa89f620e287d7fa0cedc", + "actual": "4e96c939ef598e8744521fedfda14febb8078e49c67fa89f620e287d7fa0cedc" + }, + "upstream/textop/TextOpRobotMDAR/robotmdar/cli.py": { + "expected": "0c3f7017f3056b1a7895c1a3c3f6816e1fd90598c71caa42cd25f0d154e8c8ab", + "actual": "0c3f7017f3056b1a7895c1a3c3f6816e1fd90598c71caa42cd25f0d154e8c8ab" + }, + "upstream/textop/TextOpRobotMDAR/robotmdar/config/__init__.py": { + "expected": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", + "actual": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855" + }, + "upstream/textop/TextOpRobotMDAR/robotmdar/dataloader/__init__.py": { + "expected": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", + "actual": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855" + }, + "upstream/textop/TextOpRobotMDAR/robotmdar/dataloader/data.py": { + "expected": "4107f913690a857f175b807dc43d2991d6264f9b5b4956d890da7cb18c31cceb", + "actual": "4107f913690a857f175b807dc43d2991d6264f9b5b4956d890da7cb18c31cceb" + }, + "upstream/textop/TextOpRobotMDAR/robotmdar/diffusion/__init__.py": { + "expected": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", + "actual": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855" + }, + "upstream/textop/TextOpRobotMDAR/robotmdar/diffusion/fp16_util.py": { + "expected": "5ca8ca970755b1f12783da366290228945795216e9ec15ca7276174f7f6910fc", + "actual": "5ca8ca970755b1f12783da366290228945795216e9ec15ca7276174f7f6910fc" + }, + "upstream/textop/TextOpRobotMDAR/robotmdar/diffusion/gaussian_diffusion.py": { + "expected": "b8b426f0fb7bcff02d3fcbf5b665dd2265d9be1a7a0fe6634ac0ad3cd8864f9b", + "actual": "b8b426f0fb7bcff02d3fcbf5b665dd2265d9be1a7a0fe6634ac0ad3cd8864f9b" + }, + "upstream/textop/TextOpRobotMDAR/robotmdar/diffusion/logger.py": { + "expected": "615d9cb63dd2cfe017ac424bcc64340257c83eb9cb8e6b6913df4bd55831c1c0", + "actual": "615d9cb63dd2cfe017ac424bcc64340257c83eb9cb8e6b6913df4bd55831c1c0" + }, + "upstream/textop/TextOpRobotMDAR/robotmdar/diffusion/losses.py": { + "expected": "86a33188b9419c248e6f338fcb49ef45f2d59a1b9fb108966035f35d8c81b874", + "actual": "86a33188b9419c248e6f338fcb49ef45f2d59a1b9fb108966035f35d8c81b874" + }, + "upstream/textop/TextOpRobotMDAR/robotmdar/diffusion/nn.py": { + "expected": "0b742e6fa65c1bbd752467951b4f026ff1e8210f558601e653d3c65b5f83b517", + "actual": "0b742e6fa65c1bbd752467951b4f026ff1e8210f558601e653d3c65b5f83b517" + }, + "upstream/textop/TextOpRobotMDAR/robotmdar/diffusion/resample.py": { + "expected": "13a06f67e655dc145833896389576b0f8e68b2f677dc9eed60ed8839aa6a9c62", + "actual": "13a06f67e655dc145833896389576b0f8e68b2f677dc9eed60ed8839aa6a9c62" + }, + "upstream/textop/TextOpRobotMDAR/robotmdar/diffusion/respace.py": { + "expected": "778abccf75205dd186e13f1a422d2ab6a477aaeeef203846e2770ac8e3699a33", + "actual": "778abccf75205dd186e13f1a422d2ab6a477aaeeef203846e2770ac8e3699a33" + }, + "upstream/textop/TextOpRobotMDAR/robotmdar/dtype/__init__.py": { + "expected": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", + "actual": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855" + }, + "upstream/textop/TextOpRobotMDAR/robotmdar/dtype/abc.py": { + "expected": "3b292a44163cc5c382db1b19d15e240ca3ab3536a4fdede4b6d8e07c0aeba5ab", + "actual": "3b292a44163cc5c382db1b19d15e240ca3ab3536a4fdede4b6d8e07c0aeba5ab" + }, + "upstream/textop/TextOpRobotMDAR/robotmdar/dtype/debug.py": { + "expected": "44e4f52d6ecc92189f1bb444f882ecdef358729836a0cab625d358a11a3e7a4f", + "actual": "44e4f52d6ecc92189f1bb444f882ecdef358729836a0cab625d358a11a3e7a4f" + }, + "upstream/textop/TextOpRobotMDAR/robotmdar/dtype/device.py": { + "expected": "ef3cbab9467c5471c02ab886403ac7c7180a37adb36110652f77c7580ed05085", + "actual": "ef3cbab9467c5471c02ab886403ac7c7180a37adb36110652f77c7580ed05085" + }, + "upstream/textop/TextOpRobotMDAR/robotmdar/dtype/enum.py": { + "expected": "39225955eb12aa25c12594a6e40698d2c1787443e5421282f4f2900b4dc3ae56", + "actual": "39225955eb12aa25c12594a6e40698d2c1787443e5421282f4f2900b4dc3ae56" + }, + "upstream/textop/TextOpRobotMDAR/robotmdar/dtype/logger.py": { + "expected": "d72471467e1b1741b5b263ee335af9b02661d59485c3ef35c91f249d97345865", + "actual": "d72471467e1b1741b5b263ee335af9b02661d59485c3ef35c91f249d97345865" + }, + "upstream/textop/TextOpRobotMDAR/robotmdar/dtype/motion.py": { + "expected": "3dbe3832c309ebe1128c7137b51d444606b3f4f17d1e57f393141c0db8b3e168", + "actual": "3dbe3832c309ebe1128c7137b51d444606b3f4f17d1e57f393141c0db8b3e168" + }, + "upstream/textop/TextOpRobotMDAR/robotmdar/dtype/rotation.py": { + "expected": "733b5914adfe92efebc0f261d1124985f7221943dab7cce0799c712af47e77ba", + "actual": "733b5914adfe92efebc0f261d1124985f7221943dab7cce0799c712af47e77ba" + }, + "upstream/textop/TextOpRobotMDAR/robotmdar/dtype/seed.py": { + "expected": "1de944913a0d814a42e2224300c9ba716a02326eb473a3cf663b209414850675", + "actual": "1de944913a0d814a42e2224300c9ba716a02326eb473a3cf663b209414850675" + }, + "upstream/textop/TextOpRobotMDAR/robotmdar/dtype/vis_mjc.py": { + "expected": "5e4b7a56c45a424a09114c6e05d971f0e3e19135f1ec7b8bf420fd780e711c77", + "actual": "5e4b7a56c45a424a09114c6e05d971f0e3e19135f1ec7b8bf420fd780e711c77" + }, + "upstream/textop/TextOpRobotMDAR/robotmdar/eval/__init__.py": { + "expected": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", + "actual": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855" + }, + "upstream/textop/TextOpRobotMDAR/robotmdar/eval/freq_dar.py": { + "expected": "600d48e189e104da09a5764e32c3b1651dc130ef1505f5e45196b544e5c1b258", + "actual": "600d48e189e104da09a5764e32c3b1651dc130ef1505f5e45196b544e5c1b258" + }, + "upstream/textop/TextOpRobotMDAR/robotmdar/eval/generate_dar.py": { + "expected": "adcea83c85b55f38bc6e22ffd0430db246b22d42e0a27e37b2adff8f6f03da73", + "actual": "adcea83c85b55f38bc6e22ffd0430db246b22d42e0a27e37b2adff8f6f03da73" + }, + "upstream/textop/TextOpRobotMDAR/robotmdar/eval/loop_dar.py": { + "expected": "9c1ee5784432bc40408c5f8dcd6a1fd10ad31bafeed05373d51e506ca4a4c1b4", + "actual": "9c1ee5784432bc40408c5f8dcd6a1fd10ad31bafeed05373d51e506ca4a4c1b4" + }, + "upstream/textop/TextOpRobotMDAR/robotmdar/eval/vis_dar.py": { + "expected": "c4b2794a4e4a16e8eddc7efef68e50d10a981a565b5c8f6405486eb79bef7d88", + "actual": "c4b2794a4e4a16e8eddc7efef68e50d10a981a565b5c8f6405486eb79bef7d88" + }, + "upstream/textop/TextOpRobotMDAR/robotmdar/eval/vis_mvae.py": { + "expected": "78dde051a7250a33c93a3eba541b6074e3609ef4cd18c1a476d970802f431a32", + "actual": "78dde051a7250a33c93a3eba541b6074e3609ef4cd18c1a476d970802f431a32" + }, + "upstream/textop/TextOpRobotMDAR/robotmdar/model/__init__.py": { + "expected": "6a3222ab321dcc09cca255712c3a4a943a9347216e89459b6e2dc15c34fe9e22", + "actual": "6a3222ab321dcc09cca255712c3a4a943a9347216e89459b6e2dc15c34fe9e22" + }, + "upstream/textop/TextOpRobotMDAR/robotmdar/model/clip.py": { + "expected": "e9ef0eca2db8d8873737495ac2e9cd3f0f1b71caba63a732bf31b3cb135fac65", + "actual": "e9ef0eca2db8d8873737495ac2e9cd3f0f1b71caba63a732bf31b3cb135fac65" + }, + "upstream/textop/TextOpRobotMDAR/robotmdar/model/mld_denoiser.py": { + "expected": "a7b7d72347c125b5cf413e9c25c90723b1ffc4da2389d1dfc7ae4633ab74c687", + "actual": "a7b7d72347c125b5cf413e9c25c90723b1ffc4da2389d1dfc7ae4633ab74c687" + }, + "upstream/textop/TextOpRobotMDAR/robotmdar/model/mld_vae.py": { + "expected": "6be6b305c74d7cd7a8abacaf01260893412ad46df8f4e00f1ab48413e2c4f508", + "actual": "6be6b305c74d7cd7a8abacaf01260893412ad46df8f4e00f1ab48413e2c4f508" + }, + "upstream/textop/TextOpRobotMDAR/robotmdar/model/modules.py": { + "expected": "62ae3ae9a630f7401baa70b09b0b8931cba74f8f67a33131378658e8bc30e34c", + "actual": "62ae3ae9a630f7401baa70b09b0b8931cba74f8f67a33131378658e8bc30e34c" + }, + "upstream/textop/TextOpRobotMDAR/robotmdar/model/operator/__init__.py": { + "expected": "d5d517bc233a1c243330fd46af85e230b365235635e1f60876ca964cbca766dc", + "actual": "d5d517bc233a1c243330fd46af85e230b365235635e1f60876ca964cbca766dc" + }, + "upstream/textop/TextOpRobotMDAR/robotmdar/model/operator/adain.py": { + "expected": "3f7c3a64a877c2c978b400f66eac1d60b769c68d0f611318bf305558f88e2de4", + "actual": "3f7c3a64a877c2c978b400f66eac1d60b769c68d0f611318bf305558f88e2de4" + }, + "upstream/textop/TextOpRobotMDAR/robotmdar/model/operator/blocks.py": { + "expected": "931c1af282a5362a57ce9ba66ea0342e901fca71c7f2a784d2f7b7283dc0b128", + "actual": "931c1af282a5362a57ce9ba66ea0342e901fca71c7f2a784d2f7b7283dc0b128" + }, + "upstream/textop/TextOpRobotMDAR/robotmdar/model/operator/conv2d_gradfix.py": { + "expected": "ecf394964eb5cf04c5503a481c2534e6ebfe6908a9a196a2e1e3b72057269e66", + "actual": "ecf394964eb5cf04c5503a481c2534e6ebfe6908a9a196a2e1e3b72057269e66" + }, + "upstream/textop/TextOpRobotMDAR/robotmdar/model/operator/cross_attention.py": { + "expected": "179b2f043871f54ad66965154c71f9aa3384366b4a90fe58326875ed21fdfb3e", + "actual": "179b2f043871f54ad66965154c71f9aa3384366b4a90fe58326875ed21fdfb3e" + }, + "upstream/textop/TextOpRobotMDAR/robotmdar/model/operator/position_encoding.py": { + "expected": "914306fefb39cb05d3c0f7bb397e2a5a8dcd55daf681b47a37201ed1a748726b", + "actual": "914306fefb39cb05d3c0f7bb397e2a5a8dcd55daf681b47a37201ed1a748726b" + }, + "upstream/textop/TextOpRobotMDAR/robotmdar/model/operator/position_encoding_layer.py": { + "expected": "c0516d4da7f20cda5bfe25edfdb3887a61f8c2296f4eeba0cb9b8cc26b860b5c", + "actual": "c0516d4da7f20cda5bfe25edfdb3887a61f8c2296f4eeba0cb9b8cc26b860b5c" + }, + "upstream/textop/TextOpRobotMDAR/robotmdar/skeleton/__init__.py": { + "expected": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", + "actual": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855" + }, + "upstream/textop/TextOpRobotMDAR/robotmdar/skeleton/forward_kinematics.py": { + "expected": "f27c1e790862db65f7e901f64f4ad7c103976ce05e9db06c9f2a4008d9c8d9e3", + "actual": "f27c1e790862db65f7e901f64f4ad7c103976ce05e9db06c9f2a4008d9c8d9e3" + }, + "upstream/textop/TextOpRobotMDAR/robotmdar/skeleton/robot.py": { + "expected": "b26226ac8c9fca0e59fd8888caa907c88fac109d3e50df63b699f807a1967768", + "actual": "b26226ac8c9fca0e59fd8888caa907c88fac109d3e50df63b699f807a1967768" + }, + "upstream/textop/TextOpRobotMDAR/robotmdar/train/__init__.py": { + "expected": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", + "actual": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855" + }, + "upstream/textop/TextOpRobotMDAR/robotmdar/train/manager.py": { + "expected": "06572c88fa0c584e56c44205828f6efbcdf7e48328b6b94840a3661d898b08eb", + "actual": "06572c88fa0c584e56c44205828f6efbcdf7e48328b6b94840a3661d898b08eb" + }, + "upstream/textop/TextOpRobotMDAR/robotmdar/train/train_dar.py": { + "expected": "0ca27402bc4a0f7732e9077756d66e74900fccc784730b9bf078de5a58ad929f", + "actual": "0ca27402bc4a0f7732e9077756d66e74900fccc784730b9bf078de5a58ad929f" + }, + "upstream/textop/TextOpRobotMDAR/robotmdar/train/train_mvae.py": { + "expected": "471fb1559b02d5278b8fe148809a46855949c3c18f58afc6adeefa2c52070215", + "actual": "471fb1559b02d5278b8fe148809a46855949c3c18f58afc6adeefa2c52070215" + }, + "upstream/textop/TextOpRobotMDAR/robotmdar/train/train_platforms.py": { + "expected": "edfdf9d9984618b6fb1c7583c8c5eb3d74835e83bfd1f47b6e4b404d36e99a3f", + "actual": "edfdf9d9984618b6fb1c7583c8c5eb3d74835e83bfd1f47b6e4b404d36e99a3f" + }, + "upstream/textop/TextOpRobotMDAR/robotmdar/wrapper/__init__.py": { + "expected": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", + "actual": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855" + }, + "upstream/textop/TextOpRobotMDAR/robotmdar/wrapper/vae_decode.py": { + "expected": "e45b7ebffeb9600f0a31549eb87e7c2b018fb59fc196782ae9b65045d9dedaca", + "actual": "e45b7ebffeb9600f0a31549eb87e7c2b018fb59fc196782ae9b65045d9dedaca" + }, + "upstream/textop/dataset/cal_action_statistics.py": { + "expected": "866aa29fae28f1d5c63ea3865afbdd96594423b187d62bf4e1f404c5083b1b5e", + "actual": "866aa29fae28f1d5c63ea3865afbdd96594423b187d62bf4e1f404c5083b1b5e" + }, + "upstream/textop/dataset/motion_package.py": { + "expected": "9eefabf13b0edaa4838a558a055d219e6128be95d0c2a2b89c335308476b5a0f", + "actual": "9eefabf13b0edaa4838a558a055d219e6128be95d0c2a2b89c335308476b5a0f" + }, + "upstream/textop/dataset/pack_dataset.py": { + "expected": "54a039ff872358f50044844186957122f52bce00b6a3561f1d5ea67992b7a59e", + "actual": "54a039ff872358f50044844186957122f52bce00b6a3561f1d5ea67992b7a59e" + }, + "upstream/textop/dataset/process_retarget_data.py": { + "expected": "ba8bb84049a237149985a87ca241e3f9c8a67fa079ff03c5f4da3e6f0a6b07c4", + "actual": "ba8bb84049a237149985a87ca241e3f9c8a67fa079ff03c5f4da3e6f0a6b07c4" + }, + "upstream/textop/dataset/smplx_to_robot_dataset.py": { + "expected": "48c29987fc8e0673ad0253825d23a5bb90a6e576e348a4897b84703d93759415", + "actual": "48c29987fc8e0673ad0253825d23a5bb90a6e576e348a4897b84703d93759415" + }, + "upstream/textop/deps/isaac_utils/setup.py": { + "expected": "4022149687d6dd29680f0b00442e3d455c284e6001fce44535f07a9a1cbdb18f", + "actual": "4022149687d6dd29680f0b00442e3d455c284e6001fce44535f07a9a1cbdb18f" + }, + "upstream/textop/deps/isaac_utils/isaac_utils/__init__.py": { + "expected": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", + "actual": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855" + }, + "upstream/textop/deps/isaac_utils/isaac_utils/maths.py": { + "expected": "f44b9edded4433995d54fc3445b3d9613a018416437e2a3e5aadf779d5f9a8fb", + "actual": "f44b9edded4433995d54fc3445b3d9613a018416437e2a3e5aadf779d5f9a8fb" + }, + "upstream/textop/deps/isaac_utils/isaac_utils/rotations.py": { + "expected": "8233c8c7ea271ca9d4fba0ebfe3ad0f3138ee0e483b15397bbeb720f0f0c3a02", + "actual": "8233c8c7ea271ca9d4fba0ebfe3ad0f3138ee0e483b15397bbeb720f0f0c3a02" + }, + "upstream/textop/deps/CLIP/hubconf.py": { + "expected": "0641946610c17a743ebcc3f6e85cffff1400337bb9a99766740c40b01d9a3e14", + "actual": "0641946610c17a743ebcc3f6e85cffff1400337bb9a99766740c40b01d9a3e14" + }, + "upstream/textop/deps/CLIP/setup.py": { + "expected": "6a913890284a988293f7b0b3ca017de29c65ce2bdce9fdd09cb481fc6bd5a636", + "actual": "6a913890284a988293f7b0b3ca017de29c65ce2bdce9fdd09cb481fc6bd5a636" + }, + "upstream/textop/deps/CLIP/clip/__init__.py": { + "expected": "d7930eb290ebed432116795c91473200d098451a94c0aedd4ae0f934f97955e8", + "actual": "d7930eb290ebed432116795c91473200d098451a94c0aedd4ae0f934f97955e8" + }, + "upstream/textop/deps/CLIP/clip/clip.py": { + "expected": "9540f200fbf8145479fa655382a56dab048d238cc698b9cbd8df3b6d86d3f1b6", + "actual": "9540f200fbf8145479fa655382a56dab048d238cc698b9cbd8df3b6d86d3f1b6" + }, + "upstream/textop/deps/CLIP/clip/model.py": { + "expected": "9902cbe5ee90a1da2aa3e6f043e8a23dc1f8831193b963785c9af03d5c7bef2c", + "actual": "9902cbe5ee90a1da2aa3e6f043e8a23dc1f8831193b963785c9af03d5c7bef2c" + }, + "upstream/textop/deps/CLIP/clip/simple_tokenizer.py": { + "expected": "d1b09f10ed3d1e2c343619cb98cbfacb92363dd8607526a43aa7fe2d60d15bad", + "actual": "d1b09f10ed3d1e2c343619cb98cbfacb92363dd8607526a43aa7fe2d60d15bad" + }, + "upstream/textop/deps/CLIP/tests/test_consistency.py": { + "expected": "342f84e4814143e1117dbc14baa6cf2a2ec01f61542e3952edc91e68f5eed206", + "actual": "342f84e4814143e1117dbc14baa6cf2a2ec01f61542e3952edc91e68f5eed206" + }, + "train.py": { + "expected": "0724b3db6d374259db9f6378e9b663ef8438e5b969093f3b4d1e46ee56456ddc", + "actual": "61f49739e14275a671dc3343afcf2d6b225da547a952f963037d545afe680dd4" + } +} \ No newline at end of file diff --git a/handoff-v1/inventory/upload-retry.json b/handoff-v1/inventory/upload-retry.json new file mode 100644 index 0000000000000000000000000000000000000000..0ba9f0d1a636e5cfce2dcf9c204f80b2117cf6c7 --- /dev/null +++ b/handoff-v1/inventory/upload-retry.json @@ -0,0 +1,9 @@ +{ + "model": "hymotion", + "operation": "final pilot250 upload", + "first_client": "hymotion environment legacy HF upload path", + "failure": "HTTPS connection retry exhaustion during LFS upload", + "retry_client": "omg environment already successful for other handoff uploads", + "network_route": "AutoDL turbo for both attempts; no personal proxy fallback", + "signed_urls": "excluded from packages; failed local log sanitized" +} \ No newline at end of file diff --git a/handoff-v1/inventory/vendor-provenance.json b/handoff-v1/inventory/vendor-provenance.json new file mode 100644 index 0000000000000000000000000000000000000000..23ed9602b77f32f88bafd6f5a445cbfda50273f2 --- /dev/null +++ b/handoff-v1/inventory/vendor-provenance.json @@ -0,0 +1,322 @@ +{ + "omg": { + "project": "OMG", + "source_directory": "/cpfs01/thscc/sharestorage/AIGC-06-01/Hithink_VideoGen_Group/user_workspace/buwendong/user/code/OMG", + "created_utc": "2026-09-27T13:41:26.483781+00:00", + "git_head": "61e196010b332acbace223b2c449e25f454ea0a3", + "snapshot": "Working files including local changes; no Git history", + "excluded_roots_and_files": [ + { + "path": ".git", + "reason": "cache, environment, build, logs, outputs or VCS metadata" + }, + { + "path": ".pytest_cache", + "reason": "cache, environment, build, logs, outputs or VCS metadata" + }, + { + "path": "__pycache__", + "reason": "cache, environment, build, logs, outputs or VCS metadata" + }, + { + "path": "models", + "reason": "dataset, weight, external asset or generated artifact" + }, + { + "path": "outputs", + "reason": "cache, environment, build, logs, outputs or VCS metadata" + }, + { + "path": "assets/assets", + "reason": "symlink omitted: /cpfs01/thscc/sharestorage/AIGC-06-01/Hithink_VideoGen_Group/user_workspace/buwendong/user/code/OMG/assets" + }, + { + "path": "assets/benchmarks", + "reason": "dataset, weight, external asset or generated artifact" + }, + { + "path": "configs/configs", + "reason": "symlink omitted: /cpfs01/thscc/sharestorage/AIGC-06-01/Hithink_VideoGen_Group/user_workspace/buwendong/user/code/OMG/configs" + }, + { + "path": "scripts/__pycache__", + "reason": "cache, environment, build, logs, outputs or VCS metadata" + }, + { + "path": "src/omg.egg-info", + "reason": "cache, environment, build, logs, outputs or VCS metadata" + }, + { + "path": "src/omg/__pycache__", + "reason": "cache, environment, build, logs, outputs or VCS metadata" + }, + { + "path": "src/omg/benchmarks/__pycache__", + "reason": "cache, environment, build, logs, outputs or VCS metadata" + }, + { + "path": "src/omg/benchmarks/evaluator/__pycache__", + "reason": "cache, environment, build, logs, outputs or VCS metadata" + }, + { + "path": "src/omg/benchmarks/metrics/__pycache__", + "reason": "cache, environment, build, logs, outputs or VCS metadata" + }, + { + "path": "src/omg/benchmarks/runners/__pycache__", + "reason": "cache, environment, build, logs, outputs or VCS metadata" + }, + { + "path": "src/omg/callbacks/__pycache__", + "reason": "cache, environment, build, logs, outputs or VCS metadata" + }, + { + "path": "src/omg/cli/__pycache__", + "reason": "cache, environment, build, logs, outputs or VCS metadata" + }, + { + "path": "src/omg/cli/data/__pycache__", + "reason": "cache, environment, build, logs, outputs or VCS metadata" + }, + { + "path": "src/omg/cli/evaluation/__pycache__", + "reason": "cache, environment, build, logs, outputs or VCS metadata" + }, + { + "path": "src/omg/cli/generation/__pycache__", + "reason": "cache, environment, build, logs, outputs or VCS metadata" + }, + { + "path": "src/omg/cli/pipeline/__pycache__", + "reason": "cache, environment, build, logs, outputs or VCS metadata" + }, + { + "path": "src/omg/cli/realtime/__pycache__", + "reason": "cache, environment, build, logs, outputs or VCS metadata" + }, + { + "path": "src/omg/cli/tracking/__pycache__", + "reason": "cache, environment, build, logs, outputs or VCS metadata" + }, + { + "path": "src/omg/core/__pycache__", + "reason": "cache, environment, build, logs, outputs or VCS metadata" + }, + { + "path": "src/omg/data/__pycache__", + "reason": "cache, environment, build, logs, outputs or VCS metadata" + }, + { + "path": "src/omg/generation/__pycache__", + "reason": "cache, environment, build, logs, outputs or VCS metadata" + }, + { + "path": "src/omg/generation/conditions/__pycache__", + "reason": "cache, environment, build, logs, outputs or VCS metadata" + }, + { + "path": "src/omg/generation/denoisers/__pycache__", + "reason": "cache, environment, build, logs, outputs or VCS metadata" + }, + { + "path": "src/omg/generation/diffusion/__pycache__", + "reason": "cache, environment, build, logs, outputs or VCS metadata" + }, + { + "path": "src/omg/generation/export/__pycache__", + "reason": "cache, environment, build, logs, outputs or VCS metadata" + }, + { + "path": "src/omg/generation/losses/__pycache__", + "reason": "cache, environment, build, logs, outputs or VCS metadata" + }, + { + "path": "src/omg/generation/metrics/__pycache__", + "reason": "cache, environment, build, logs, outputs or VCS metadata" + }, + { + "path": "src/omg/generation/models/__pycache__", + "reason": "cache, environment, build, logs, outputs or VCS metadata" + }, + { + "path": "src/omg/motion/__pycache__", + "reason": "cache, environment, build, logs, outputs or VCS metadata" + }, + { + "path": "src/omg/pipeline/__pycache__", + "reason": "cache, environment, build, logs, outputs or VCS metadata" + }, + { + "path": "src/omg/realtime/__pycache__", + "reason": "cache, environment, build, logs, outputs or VCS metadata" + }, + { + "path": "src/omg/render/__pycache__", + "reason": "cache, environment, build, logs, outputs or VCS metadata" + }, + { + "path": "src/omg/robots/__pycache__", + "reason": "cache, environment, build, logs, outputs or VCS metadata" + }, + { + "path": "src/omg/robots/g1/__pycache__", + "reason": "cache, environment, build, logs, outputs or VCS metadata" + }, + { + "path": "src/omg/runtime/__pycache__", + "reason": "cache, environment, build, logs, outputs or VCS metadata" + }, + { + "path": "src/omg/tracking/__pycache__", + "reason": "cache, environment, build, logs, outputs or VCS metadata" + }, + { + "path": "src/omg/tracking/holomotion/__pycache__", + "reason": "cache, environment, build, logs, outputs or VCS metadata" + }, + { + "path": "src/omg/utils/__pycache__", + "reason": "cache, environment, build, logs, outputs or VCS metadata" + }, + { + "path": "tests/benchmarks/__pycache__", + "reason": "cache, environment, build, logs, outputs or VCS metadata" + }, + { + "path": "tests/callbacks/__pycache__", + "reason": "cache, environment, build, logs, outputs or VCS metadata" + }, + { + "path": "tests/data/__pycache__", + "reason": "cache, environment, build, logs, outputs or VCS metadata" + }, + { + "path": "tests/generation/__pycache__", + "reason": "cache, environment, build, logs, outputs or VCS metadata" + }, + { + "path": "tests/pipeline/__pycache__", + "reason": "cache, environment, build, logs, outputs or VCS metadata" + }, + { + "path": "tests/realtime/__pycache__", + "reason": "cache, environment, build, logs, outputs or VCS metadata" + }, + { + "path": "tests/robots/__pycache__", + "reason": "cache, environment, build, logs, outputs or VCS metadata" + }, + { + "path": "tests/tracking/__pycache__", + "reason": "cache, environment, build, logs, outputs or VCS metadata" + } + ], + "notes": [ + "Excluded directories were pruned without walking large artifacts.", + "No symlinks included.", + "README_YTH and ENVIRONMENT snapshots added separately.", + "For bench only, data/README.md is explicitly restored." + ], + "current_snapshot": "Original workspace archive includes all local changes; upstream directory git rev-parse reports enclosing workspace commit, not original vendor commit." + }, + "textop": { + "project": "TextOp", + "source_directory": "/cpfs01/thscc/sharestorage/AIGC-06-01/Hithink_VideoGen_Group/user_workspace/linyuze/baselines/repos/TextOp", + "created_utc": "2026-09-27T13:41:21.110906+00:00", + "git_head": "ef6555fb174c9b5c44945a62c7ffc77b5ddbbf22", + "snapshot": "Working files including local changes; no Git history", + "excluded_roots_and_files": [ + { + "path": ".git", + "reason": "cache, environment, build, logs, outputs or VCS metadata" + }, + { + "path": "TextOpDeploy/src/cnpy/.git", + "reason": "cache, environment, build, logs, outputs or VCS metadata" + }, + { + "path": "TextOpDeploy/src/textop_ctrl/models", + "reason": "dataset, weight, external asset or generated artifact" + }, + { + "path": "TextOpDeploy/src/unitree_ros2/.git", + "reason": "cache, environment, build, logs, outputs or VCS metadata" + }, + { + "path": "TextOpDeploy/src/unitree_ros2/example/src/src/g1/audio_client/test.wav", + "reason": "weight, binary data, media or compiled artifact" + }, + { + "path": "TextOpDeploy/src/unitree_ros2/example/src/src/g1/lowlevel/behavior_lib/motion.seq", + "reason": "weight, binary data, media or compiled artifact" + }, + { + "path": "TextOpRobotMDAR/dataset", + "reason": "dataset, weight, external asset or generated artifact" + }, + { + "path": "TextOpRobotMDAR/description", + "reason": "symlink omitted: /cpfs01/thscc/sharestorage/AIGC-06-01/Hithink_VideoGen_Group/user_workspace/linyuze/baselines/models/TextOp-Data/TextOpRobotMDAR/description" + }, + { + "path": "TextOpRobotMDAR/logs", + "reason": "cache, environment, build, logs, outputs or VCS metadata" + }, + { + "path": "TextOpRobotMDAR/robotmdar.egg-info", + "reason": "cache, environment, build, logs, outputs or VCS metadata" + }, + { + "path": "TextOpRobotMDAR/robotmdar/__pycache__", + "reason": "cache, environment, build, logs, outputs or VCS metadata" + }, + { + "path": "TextOpRobotMDAR/robotmdar/dataloader/__pycache__", + "reason": "cache, environment, build, logs, outputs or VCS metadata" + }, + { + "path": "TextOpRobotMDAR/robotmdar/diffusion/__pycache__", + "reason": "cache, environment, build, logs, outputs or VCS metadata" + }, + { + "path": "TextOpRobotMDAR/robotmdar/dtype/__pycache__", + "reason": "cache, environment, build, logs, outputs or VCS metadata" + }, + { + "path": "TextOpRobotMDAR/robotmdar/eval/__pycache__", + "reason": "cache, environment, build, logs, outputs or VCS metadata" + }, + { + "path": "TextOpRobotMDAR/robotmdar/model/__pycache__", + "reason": "cache, environment, build, logs, outputs or VCS metadata" + }, + { + "path": "TextOpRobotMDAR/robotmdar/model/operator/__pycache__", + "reason": "cache, environment, build, logs, outputs or VCS metadata" + }, + { + "path": "TextOpRobotMDAR/robotmdar/skeleton/__pycache__", + "reason": "cache, environment, build, logs, outputs or VCS metadata" + }, + { + "path": "TextOpRobotMDAR/robotmdar/train/__pycache__", + "reason": "cache, environment, build, logs, outputs or VCS metadata" + }, + { + "path": "deps/isaac_utils/isaac_utils.egg-info", + "reason": "cache, environment, build, logs, outputs or VCS metadata" + }, + { + "path": "deps/isaac_utils/isaac_utils/__pycache__", + "reason": "cache, environment, build, logs, outputs or VCS metadata" + } + ], + "notes": [ + "Excluded directories were pruned without walking large artifacts.", + "No symlinks included.", + "README_YTH and ENVIRONMENT snapshots added separately.", + "For bench only, data/README.md is explicitly restored." + ], + "current_snapshot": "Original workspace archive includes all local changes; upstream directory git rev-parse reports enclosing workspace commit, not original vendor commit." + } +} \ No newline at end of file diff --git a/handoff-v1/inventory/workspace-commit.txt b/handoff-v1/inventory/workspace-commit.txt new file mode 100644 index 0000000000000000000000000000000000000000..b497fa5066ab5a016e8f4eb188e2c2e2fbe8ce67 --- /dev/null +++ b/handoff-v1/inventory/workspace-commit.txt @@ -0,0 +1 @@ +ea269e3716e289e64fa3cec17f1366b57e3fab9b diff --git a/handoff-v1/inventory/workspace-diff.txt b/handoff-v1/inventory/workspace-diff.txt new file mode 100644 index 0000000000000000000000000000000000000000..c9c545369cebbcf6d2a05f027ead2f705f40b6ea --- /dev/null +++ b/handoff-v1/inventory/workspace-diff.txt @@ -0,0 +1,6411 @@ +diff --git a/FINAL_BASELINE_METHOD.md b/FINAL_BASELINE_METHOD.md +deleted file mode 100644 +index f388a06..0000000 +--- a/FINAL_BASELINE_METHOD.md ++++ /dev/null +@@ -1,16 +0,0 @@ +-# 当前保留的方案 +- +-本文为历史完整 session 方法记录。完整调试记录见 [视频问题调试](docs/ROLLOUT_DEBUG.md)。对应产物已归档到 `/root/autodl-tmp/runs/archive/legacy-20261020/hfd-final/`;该目录不随 portable 代码迁移。当前正式训练与评测见 [runs索引](/root/autodl-tmp/runs/README.md)。 +- +-| 模型 | 配置 | 当前连续 joint RMSE | 状态 | +-| --- | --- | ---: | --- | +-| OMG | `configs/omg_full_session.yaml` | 0.0619 rad | 监听与收尾恢复,重复弯腰消失,残余接缝仍在 | +-| TextOP | `configs/textop_full_session.yaml` | 0.2992 rad | 监听姿态出现,任务切换延迟,尚未完整拟合 | +- +-共同改为 `full_session` 关节监督,不向模型输入 listen/idle 等控制标签。训练仍使用原生框架和官方初始化。OMG 保持过去两秒音频+因果 ASR,采用原生 DDIM 10 步、CFG=1、60 帧规划和剩余 24 帧的异步预取;TextOP 保持原生 8 帧输出、CFG=5。 +- +-OMG 三个固定种子的误差为 0.0619、0.1538、0.0585 rad,当前视频沿用原 seed=12345。它仍未证明泛化或完美实时执行。TextOP 真值历史诊断为 0.0125 rad,但这不是连续推理性能;长训练反馈链对照尚未可靠修复闭环。 +- +-视频位于 `videos/omg_full_session_playable.mp4` 和 `videos/textop_full_session_playable.mp4`,均保留原始混合音频。完整报告在各模型的 `inference/`;动作曲线和同 checkpoint 的对照集中在 `debug/`。论文独立测试集评测仍待完成。 +- +-重训前修改配置的 `paths.output_dir`。完整迁移步骤见 [迁移说明](docs/MIGRATION.md);`configs/*_full_session.yaml` 中的 `paths.data_disk` 可通过 `HFD_DATA_DISK` 覆盖。 +diff --git a/JOINT_MAPPING.md b/JOINT_MAPPING.md +deleted file mode 100644 +index 8ebd645..0000000 +--- a/JOINT_MAPPING.md ++++ /dev/null +@@ -1,113 +0,0 @@ +-# HFD 到 TextOP / OMG 的动作映射与历史 bug +- +-## 原始数据和目标接口 +- +-session 的 `motion_q29_xyz_quat` 是 `[T, 36]` 数组,列布局为: +- +-| 列(从 0 开始) | 内容 | 顺序 | +-| --- | --- | --- | +-| `0:29` | G1 的 29 个关节角(rad) | HFD / IsaacLab 顺序 | +-| `29:32` | 根位置 | xyz | +-| `32:36` | 根四元数 | wxyz | +- +-IsaacLab 关节顺序交错排列左右侧和躯干。OMG 的 G1 qpos 使用机器人树顺序:左腿 6 个、右腿 6 个、腰部 3 个、左臂 7 个、右臂 7 个。两边都是 29 个角度,但同一列不一定对应同一个关节。这里不改变机器人关节定义,也不拟合角度变换;按相同关节名重新排列列即可。 +- +-session 保留原始顺序,不在导出阶段预先重排。新导出的 JSON 明确标记 `joint_order=isaaclab`、`quaternion_order=wxyz`;历史 v1 session 缺少这两个字段时按同一原始契约读取。适配器会拒绝显式声明其他顺序的数据,避免二次转换。 +- +-## 按名称转换 +- +-实现集中在 [`hfd_bridge/joints.py`](hfd_bridge/joints.py): +- +-```python +-HFD_TO_NATIVE = [HFD_JOINT_NAMES.index(name) for name in NATIVE_JOINT_NAMES] +-native_joints = raw[:, :29][:, HFD_TO_NATIVE] +-qpos = np.concatenate((raw[:, 29:36], native_joints), axis=1) +-``` +- +-`HFD_TO_NATIVE[i]` 表示“目标第 i 个关节应从原始哪一列取值”,所有下标从 0 开始。得到的 36D qpos 布局是 `[xyz3, quat_wxyz4, native_joints29]`。函数 `hfd_motion_to_qpos` 不改变根位置和四元数。 +- +-实际排列为: +- +-```text +-[0, 3, 6, 9, 13, 17, 1, 4, 7, 10, 14, 18, 2, 5, 8, 11, 15, 19, 21, 23, 25, 27, 12, 16, 20, 22, 24, 26, 28] +-``` +- +-| 目标关节下标 | 关节名 | HFD 原始列 | +-| --- | --- | --- | +-| 0 | `left_hip_pitch_joint` | 0 | +-| 1 | `left_hip_roll_joint` | 3 | +-| 2 | `left_hip_yaw_joint` | 6 | +-| 3 | `left_knee_joint` | 9 | +-| 4 | `left_ankle_pitch_joint` | 13 | +-| 5 | `left_ankle_roll_joint` | 17 | +-| 6 | `right_hip_pitch_joint` | 1 | +-| 7 | `right_hip_roll_joint` | 4 | +-| 8 | `right_hip_yaw_joint` | 7 | +-| 9 | `right_knee_joint` | 10 | +-| 10 | `right_ankle_pitch_joint` | 14 | +-| 11 | `right_ankle_roll_joint` | 18 | +-| 12 | `waist_yaw_joint` | 2 | +-| 13 | `waist_roll_joint` | 5 | +-| 14 | `waist_pitch_joint` | 8 | +-| 15 | `left_shoulder_pitch_joint` | 11 | +-| 16 | `left_shoulder_roll_joint` | 15 | +-| 17 | `left_shoulder_yaw_joint` | 19 | +-| 18 | `left_elbow_joint` | 21 | +-| 19 | `left_wrist_roll_joint` | 23 | +-| 20 | `left_wrist_pitch_joint` | 25 | +-| 21 | `left_wrist_yaw_joint` | 27 | +-| 22 | `right_shoulder_pitch_joint` | 12 | +-| 23 | `right_shoulder_roll_joint` | 16 | +-| 24 | `right_shoulder_yaw_joint` | 20 | +-| 25 | `right_elbow_joint` | 22 | +-| 26 | `right_wrist_roll_joint` | 24 | +-| 27 | `right_wrist_pitch_joint` | 26 | +-| 28 | `right_wrist_yaw_joint` | 28 | +- +-HFD 部署代码 [`upstream/hfd/scripts/deploy/run_sonic_sim_record_online.py`](upstream/hfd/scripts/deploy/run_sonic_sim_record_online.py) 使用相反方向的表,把 MuJoCo 关节值写成 IsaacLab 状态: +- +-```text +-[0, 6, 12, 1, 7, 13, 2, 8, 14, 3, 9, 15, 22, 4, 10, +- 16, 23, 5, 11, 17, 24, 18, 25, 19, 26, 20, 27, 21, 28] +-``` +- +-转换回原生顺序要取这个排列的逆排列,不能直接沿用同一张表。当前代码按关节名生成转换索引,并用部署表独立核对。 +- +-## 两个模型分别如何使用 +- +-**TextOP:** 原生接口只建模 23 个关节。适配器按模型的关节名称从 HFD 29 列中取值,保持 50 Hz,再交给原生 FeatureVersion=3 编码得到 57D 动作特征。选取的原生 29 关节下标为 `0:19` 和 `22:26`,即双腿、腰、双臂的肩和肘;六个腕关节不属于模型输出。实现可以直接使用 `TEXTOP_HFD_JOINTS` 从 raw 取 23 列,等价于先重排 29 列再选取 23 列。TextOP 根旋转接口使用 xyzw,所以输入 wxyz 会归一化并改为 `[x, y, z, w]`;根位置保留。视频中的六个腕关节使用初始化时的值,模型没有预测它们,也没有填入未来真值。 +- +-**OMG:** 原生接口使用全部 29 个关节。适配器将 HFD 的 50 Hz 动作重采样到 30 Hz:关节角和位置做线性插值,四元数做 SLERP;随后把关节列重排为原生顺序。qpos 根四元数保持 wxyz。再通过原生 G1 FK 和 codec 生成 125D 特征。按用户指定的离线测试口径,音频与目标动作窗口对齐。 +- +-渲染统一使用原生 29 关节的 qpos。真值也要经过同一份正确的 HFD 到原生转换,不能把 raw 的前 29 列直接写给渲染器。实现见 [`hfd_bridge/training.py`](hfd_bridge/training.py)、[`infer_overfit.py`](infer_overfit.py) 和 [`render_overfit.py`](render_overfit.py)。 +- +-## 之前的 bug 及影响 +- +-旧适配器误认为 HFD 的前 29 列已经是原生机器人树顺序,只调整了 root 字段的位置,没有重排关节列;TextOP 也按原生数字下标直接选择 HFD 列,因而取错关节。 +- +-例如原始第 1 列(从 0 开始)是 `right_hip_pitch_joint`,而原生第 1 个关节是 `left_hip_roll_joint`。错误代码把右髋 pitch 角写给左髋 roll。多个关节同时错位,会造成姿态明显异常。真值视频也采用了错误解释,因此即使完全没有模型预测,真值机器人仍会表现奇怪。 +- +-之前的往返测试用同一套错误排列做编码和解码,能够还原输入数字,只证明转换自洽,不能证明数字对应了正确的命名关节。错误动作空间内的训练损失下降,也不能证明真实机器人动作拟合正确。 +- +-以下结果不能用于论文中的动作质量或过拟合结论: +- +-- 修复前由旧适配器生成的动作特征、训练检查点、动作误差和预测视频。 +-- `hfd-video-overfit-textop-20261005`、`hfd-video-overfit-omg-20261005` 和 `hfd-overfit-videos-20261005` 中的旧结果,以及更早依赖同一错误排列的动作实验。 +- +-原始 HFD session、混合音频和 ASR 转写可以继续使用;错误动作特征必须重新生成。两个 baseline 已从官方初始化权重重新进行 800 步小样本训练,正确结果目录带 `jointfixed`,摘要见 [`reports/history/overfit_videos.json`](reports/history/overfit_videos.json)。旧视频、废弃适配脚本及旧实验检查点已清理;历史小型配置、日志和检查摘要集中保存在 `runs/diagnostics/historical/`,不能作为修复后模型的评测证据。 +- +-## 如何验证映射 +- +-```bash +-# 在 TextOP 环境运行已有回归测试 +-python -m unittest discover -s tests -t . +- +-# 在 OMG 环境核对独立 FK,并导出相同真值的错误/正确解释对照 +-# --output 必须是新的目录;外部 assets 需要包含 G1 URDF 和 mesh +-python check_joint_mapping.py \ +- --session /data/sessions/sample.npz \ +- --assets-root /data/OMG/assets \ +- --output /data/runs/joint-mapping-audit +-``` +- +-核验包括:部署表的逆排列与按名称转换一致;29 个关节逐个脉冲时仅到达同名电机;根位置和 wxyz 保持不变;50/30 Hz 公共时刻的姿态一致;OMG FK 与独立 MuJoCo 身体位置一致。已有样本核验的最大身体位置误差约 `8.0e-7 m`,见 [`reports/joint_mapping_audit.json`](reports/joint_mapping_audit.json)。这些检查确认数据与机器人接口对应正确;小样本训练确认训练链路可用,尚不构成完整全双工评测结论。 +diff --git a/MANIFEST.json b/MANIFEST.json +index 645a2a1..187efd9 100644 +--- a/MANIFEST.json ++++ b/MANIFEST.json +@@ -5,21 +5,9 @@ + "bytes": 4810, + "sha256": "df3bcbf6da74ff04850db5f301b6b9f84bfbea2ed68798787ad4fbb3277ec63f" + }, +- "FINAL_BASELINE_METHOD.md": { +- "bytes": 1728, +- "sha256": "014307754a7ae7d347ca5b64b6d30cb4b3831639e010ace4b6ca0f2f0b644200" +- }, +- "JOINT_MAPPING.md": { +- "bytes": 7338, +- "sha256": "a5047dc18dd36107ce385f6391460c7782fee6578fc565e7b627792c5096405f" +- }, + "README.md": { +- "bytes": 3805, +- "sha256": "8c55fba302e73f73e43ee5ad0096b852a775c23fd7ba01ac6fc068dedd4a3dd0" +- }, +- "SESSION_FORMAT.md": { +- "bytes": 5054, +- "sha256": "35c27ea0f7fe575cfbe236846ed540c3c95df432c0af8c6148091739a6c92869" ++ "bytes": 9402, ++ "sha256": "a85b9ee971e5a855b347a2b097ae5e4437b38369f1fb250faf6b135664d7e040" + }, + "TRAINING_ASSETS.json": { + "bytes": 87123, +@@ -81,181 +69,9 @@ + "bytes": 16378, + "sha256": "a4245020980dcbe6f8ba1527e96317549937ecaf128d43b8b2f24fef6405b365" + }, +- "docs/ASR_CONTEXT_OVERFIT.md": { +- "bytes": 6819, +- "sha256": "9838a5173dde9bfcffc3bab190282cd04e932b0933b52f1dbb6b7558078610ba" +- }, +- "docs/BASELINE_TEXT_CONDITION.md": { +- "bytes": 9685, +- "sha256": "1d291382f90539d20bc37e491d7c684ee888531fc2241fe53a688b7443363b40" +- }, +- "docs/ENGLISH_ONLY_PROBE.md": { +- "bytes": 9948, +- "sha256": "4ef5a6967a1e9edec4739e9882d97acc18d119a6e23342e599c0f9168d6077c2" +- }, +- "docs/FORMAL_TRAINING_READINESS.md": { +- "bytes": 14837, +- "sha256": "3676c5395bb74d0092327d2db1f8db8ba11ee4327785277b9769f8d196f88621" +- }, +- "docs/GENMO_FORMAL_TRAINING.md": { +- "bytes": 11336, +- "sha256": "eb3761334b649a5806fa5dc1e5029356e4b55e85524621b1e01642bf67a488f6" +- }, +- "docs/GENMO_HFD_OVERFIT.md": { +- "bytes": 12848, +- "sha256": "ca24702e3cfb4f359bfba677b24a013711cd0471f4255d2f8b619ab9ca327e93" +- }, +- "docs/GENMO_LISTENING_DIAGNOSIS.md": { +- "bytes": 3858, +- "sha256": "d54bdc5141d0bde62f1d76258f4ab353cbfe2d7fc713bffc6788326d322f58ee" +- }, +- "docs/GENMO_PAST_AUDIO_TEST.md": { +- "bytes": 9364, +- "sha256": "2a95909082c8ad998c542a205f1c2d3b4600bac04b9334753e5a16428356dbef" +- }, +- "docs/GENMO_V2_TRAINING.md": { +- "bytes": 6611, +- "sha256": "5c02cd209ccd1f9cf87f4f60b2d5fc54f00b412bde97ffa261856a3d96f4a397" +- }, +- "docs/GENMO_V2_VALIDATION.md": { +- "bytes": 5740, +- "sha256": "40d7cf46d333e194205dedf310954111d9fdf5ec014c4ca86dea4e3417824285" +- }, +- "docs/HFD_HUMAN_BASELINE_INPUT_PROTOCOL.md": { +- "bytes": 12323, +- "sha256": "dec2e8df1522e6740ef53ee1147e8e65cd318408c7835f65c507d56b5d28edd5" +- }, + "docs/HFD_HUMAN_BASELINE_SOURCES.json": { +- "bytes": 10598, +- "sha256": "14dae98d032442b7c86cd5f4375ff8aac2b8a99a8e8fa6840d6266ecf93f93a8" +- }, +- "docs/HUMAN_FORMAL_VALIDATION_20261024.md": { +- "bytes": 4630, +- "sha256": "286eec4ad74f26c93aa5ca5c91acbfb8b0fc0412c4ce743269a67bb074081ec1" +- }, +- "docs/HUMAN_RETARGET_EVALUATION.md": { +- "bytes": 7664, +- "sha256": "9efd14122d7f8461ccc9cf2db2d5843e02aac9b1b3b392b728b13517d3775a87" +- }, +- "docs/KIMODO_CLOSED_LOOP_STABILITY.md": { +- "bytes": 17001, +- "sha256": "7c29773be52bd16548208c4365cf4b31e4a9fa5b0b87efa9ee0eba6400d9f3f5" +- }, +- "docs/KIMODO_FORMAL_VALIDATION.md": { +- "bytes": 6733, +- "sha256": "d89622850760e9889549495e1b93dd30d4bc4f099ff9faf67f3384089527a2cf" +- }, +- "docs/KIMODO_HFD_OVERFIT_RESULTS.md": { +- "bytes": 6166, +- "sha256": "19dc1cc2dd3a0c6dba023b7b6ca979f0478b82c427e582b6386a9d94b512c091" +- }, +- "docs/KIMODO_HFD_TRAINING.md": { +- "bytes": 13291, +- "sha256": "4b7033db2727ded4d64a015c43d12b6fcd547a40a2e38662b2a33e2e3c9f0037" +- }, +- "docs/KIMODO_MOTION_RECOVERY.md": { +- "bytes": 7245, +- "sha256": "dc4f6192b533301946b218a51c0ab442572dd96996544e27b100f5594d7d2f22" +- }, +- "docs/KIMODO_ROOT_COLLAPSE_DIAGNOSIS.md": { +- "bytes": 4618, +- "sha256": "42a7a07344f523abb3bfb4d22214655141e4d4b38f7ca4d081f688ab57dc695d" +- }, +- "docs/KIMODO_ROOT_TRAINING_CAUSE_ANALYSIS.md": { +- "bytes": 8012, +- "sha256": "e2c119d0cc7e0cc65bf3c741142ad88c7844c41b5d33228189288848caac29a2" +- }, +- "docs/KIMODO_TRAINING_REPAIR.md": { +- "bytes": 5601, +- "sha256": "880ba0e4477735500471f7ac36fd075d63d874b4578b4ba8284045530670915e" +- }, +- "docs/LOCAL_INDEX_TRAINING.md": { +- "bytes": 6004, +- "sha256": "ac2dc3538b1214bf348bb672a79a6d6e1ea0b4d74cee907fa2416fe0f66405bd" +- }, +- "docs/MIGRATION.md": { +- "bytes": 426, +- "sha256": "1bbfc85ca07c183bdb6f25143d08db7dc1e526e9fb29993225a3232e74c689af" +- }, +- "docs/OMG_ENGLISH_VALIDATION.md": { +- "bytes": 3791, +- "sha256": "86a2a91471737d81fe9ef741aac5e224d3f727db9f6c08161c3bb98b78dee89b" +- }, +- "docs/OMG_TRAINING_REVIEW.md": { +- "bytes": 14254, +- "sha256": "d38b0b5049701c620935e0bf51bf0da77aa3cb2de675c33ba348b972039609d8" +- }, +- "docs/OMG_VALIDATION_DIAGNOSIS.md": { +- "bytes": 6133, +- "sha256": "5269c34a70f40977a9848106a59e05abbe9ee8486870ca4682e808cd0ac4259b" +- }, +- "docs/OMG_VALIDATION_ROLLOUT.md": { +- "bytes": 6132, +- "sha256": "ff8c7c8f249c3c89607f080ff281e15acc3d8fccc710c0f8563e417ecfa7d82e" +- }, +- "docs/PORTABLE_TRAINING.md": { +- "bytes": 11281, +- "sha256": "4b47034c1d5916537605e96153450ba5d8f3e89cb5e1fa4796e4a18b960ba973" +- }, +- "docs/REFERENCE_DYNAMIC_TRAINING.md": { +- "bytes": 9587, +- "sha256": "e55ead5d171134b3ef2a9238e4f28bd5315ace74733481f5033f6befe8b175b7" +- }, +- "docs/ROLLOUT_DEBUG.md": { +- "bytes": 8444, +- "sha256": "c45aefa40fe66dd1504465ec8254d4438aa4325b2f9687afda871e7c312e971d" +- }, +- "docs/SEMANTICS_CLOCK_FIX.md": { +- "bytes": 5914, +- "sha256": "66d1d57f18e69176b5ad91bf980a8f180350137b9c2b02d97ae1a71fa5d627cc" +- }, +- "docs/TEXTOP_ENGLISH_VALIDATION.md": { +- "bytes": 5350, +- "sha256": "c653a371cb97f690349d96c9750366cb570bd53cedb64695b283be9f67371374" +- }, +- "docs/TRAINING_REPRODUCTION_MATRIX.md": { +- "bytes": 4693, +- "sha256": "0c977387862ae82c48c2667e5d3f4520a9c522e6394a8d4e68cfa9db2a9919ad" +- }, +- "docs/TRAINING_RESUME.md": { +- "bytes": 4215, +- "sha256": "3caa601fdaef09d30ab79c195299fc5dfe52cd0b32be4daaf97e6bc7d7484253" +- }, +- "docs/WORKSPACE_CLEANUP.md": { +- "bytes": 4422, +- "sha256": "d10560c0c62a61f6cbee040c90c5db4fbe7167aecb286c2d8d637d161c72e90b" +- }, +- "docs/history/BASELINE_METHOD_BEFORE_SEMANTIC_FIX.md": { +- "bytes": 10085, +- "sha256": "60cfdeb3ee30d306e40af466723eaa98d73c2efa0a806b8b219a9b79f8013847" +- }, +- "docs/history/DUPLEX_TASK.md": { +- "bytes": 7084, +- "sha256": "75c02d717b23b000d95b35fbda0799326dece83f0e6ec893a368fe5e07ba420c" +- }, +- "docs/history/HFD_HUMAN_BASELINE_INPUT_PROTOCOL_RESEARCH_20261017.md": { +- "bytes": 40979, +- "sha256": "ad30933320fc00b863d50f7f102ab5c546c5a82ce4919ab4956af6c93997f702" +- }, +- "docs/history/HFD_WINDOW_PROTOCOL.md": { +- "bytes": 7986, +- "sha256": "4a5a25129aa0b5cd1173af7d2c9d365cdcb118a56856cb197f547f1e5e05c3d4" +- }, +- "docs/history/NORMAL_TASK.md": { +- "bytes": 6244, +- "sha256": "9a5600eeb00436d77603501666c910ca3ed536a93c696015ed706b5f5bff9c55" +- }, +- "docs/history/OMG_CONTEXT_DEBUG.md": { +- "bytes": 6920, +- "sha256": "34d50e904a8cfc721656e7939f6aff3c07021726d9a46b818ac853945d90134e" +- }, +- "docs/history/README.md": { +- "bytes": 670, +- "sha256": "09b277b367872ab7203e6ee1e6e48d31c75269e651982a1286f074b343b2ef43" +- }, +- "docs/history/VIDEO_SAMPLES.md": { +- "bytes": 7997, +- "sha256": "ff3489e07a6233176669aa5f2567284dbd880cb36e3f6089dbc60f3d9516295b" ++ "bytes": 10428, ++ "sha256": "2cc65718bc6bf9d0d089ab341db266f9bfc18fee1d6cdec4f80978ddf74f1493" + }, + "env.sh": { + "bytes": 631, +@@ -478,8 +294,8 @@ + "sha256": "1dc280ec6f2c008924b1a69079ee50d8e8281c0c2da2d291fc5c9f9c91a36769" + }, + "tools/configure_training.py": { +- "bytes": 6935, +- "sha256": "1519a086061cda3bf93aa1f7188bc0746c95efe2084cafb7feefc09405855308" ++ "bytes": 7506, ++ "sha256": "88ebaa442e230394b7929a85f26a696a309c5315fc4ce2f0b3307af730b7f7bc" + }, + "tools/download_hfd_tars.py": { + "bytes": 6701, +@@ -2348,6 +2164,42 @@ + "validate_sessions.py": { + "bytes": 1431, + "sha256": "8f7c045d7e9f75deeffc5777eefe31760c893c929fb25d2dffda8d8e520bd880" ++ }, ++ "docs/CURRENT_RUNS.md": { ++ "bytes": 15135, ++ "sha256": "a65e3eb72a8da86e705d134a9c43aa97be17aa51f53f4d3d1bf4c858faf7c2eb" ++ }, ++ "train_hymotion.py": { ++ "bytes": 3070, ++ "sha256": "c93a486c828963e568750da26ff81256804633fd34d28d9415d3535168c7c0b0" ++ }, ++ "train_hymotion_autoregressive.py": { ++ "bytes": 15136, ++ "sha256": "1438ed061573814dfa5fe9afe6914fb59df1e70d0215958dcbd9e1b114be839f" ++ }, ++ "hfd_bridge/hymotion_model.py": { ++ "bytes": 7571, ++ "sha256": "ed29d2d03b247ed681a09dc230697885c5ab8fa8b30f82c4a827aa8fe0b41597" ++ }, ++ "hfd_bridge/hymotion_data.py": { ++ "bytes": 2203, ++ "sha256": "79503602a9a4de4f67931fa944a198d6f438b531a363ffbe545ecc4dd54f97e7" ++ }, ++ "hfd_bridge/hymotion_training.py": { ++ "bytes": 3382, ++ "sha256": "792b76cfc35f58b0c7c94b40a660f00c4971fff1d978ffd0c923b9f26a5fa733" ++ }, ++ "hfd_bridge/hymotion_rollout.py": { ++ "bytes": 2186, ++ "sha256": "6e27c4e24bd42a211df7be2a0271337af4e658d66e5d303633912bbd40313517" ++ }, ++ "hfd_bridge/hymotion_evaluation.py": { ++ "bytes": 9348, ++ "sha256": "fe72667281f04a55bc8c88375b7ca3cee1f1b6de2c23fcd7d45f1e73c1530a99" ++ }, ++ "configs/hymotion_c24.yaml": { ++ "bytes": 1123, ++ "sha256": "258c03cd4ed31c41575c4aafe0e3ddd5758a0654a5f75a903833be6ff0359bca" + } + } + } +diff --git a/README.md b/README.md +index be11660..aac9867 100644 +--- a/README.md ++++ b/README.md +@@ -1,43 +1,861 @@ +-# HFD baseline training ++# HFD 五模型训练约定 + +-OMG、TextOp、GenMo和Kimodo在全双工HFD上的训练代码、配置及复现文档。仓库包含动态session组装、因果文本释放、原生训练、验证loss和完整断点恢复。HY-Motion尚无完整训练入口。 ++本文件维护当前训练约定;运行进度、有效权重、评测结论和待办见 [CURRENT_RUNS.md](docs/CURRENT_RUNS.md)。除非用户明确要求,不另写方案、诊断、历史报告等文档。 + +-| 模型 | 入口 | 正式配置 | 预算 | +-|---|---|---|---| +-| OMG | `train.py` | `configs/omg.yaml`,另支持50M/100M/300M/500M生成配置 | 100K,有效batch512 | +-| TextOp | `train.py` | `configs/textop.yaml` | 300K,batch512,三阶段各100K | +-| GenMo/GEM-SMPL v2 | `train_genmo_full.py` | `configs/genmo.yaml`,旧配方`configs/genmo_legacy.yaml` | 100K,batch32 | +-| Kimodo-SMPLX | `train_kimodo.py` | `configs/kimodo.yaml` | 100K,batch8 | ++## 数据与信息释放 + +-输入采用按词时间释放的英文前景GT文本、最近30秒上下文及各自的动作历史。OMG另输入已释放的过去2秒音频。session在训练时动态组装,不预导出全量session,不运行ASR或中文翻译。详见[输入协议](docs/HFD_HUMAN_BASELINE_INPUT_PROTOCOL.md)。 ++- 共用 [configs/sessions.yaml](configs/sessions.yaml) 的 final_release:profile=isolated_benchmark_1024_20261026;索引 Tinghong-Ye22/data_vqa_all_1025@bdca2fe7581c263e9ceb3a96204b314e199faddf,原料 Tinghong-Ye22/hfd_data@93ba4a7a9c66b3e1b7041e5f4ea038ef2418b948。 ++- 完整发布 train/val/test=258661/512/1024,含中文。当前仅英文:OMG/TextOp=126950/302/436(原生+第三方);GenMo/Kimodo/HY=117290/297/419(有配对SMPL-X的原生样本)。英文子集不等于完整1024 benchmark。 ++- 只回放动作session和原始配方,不下载或传入VQA问答、视频、问答配音;语音—动作与视频—音频完全解耦。保留严格session身份、动作长度和音频时钟检查。先按最终split/语言过滤,再按记录的upstream_split、dataset_index、seed回放,不重新随机划分。 ++- 英文文本用 aligned_reference_context、text_conditioning=none;读取本地预建GT词时间映射 /root/autodl-tmp/cache/hfd/reference-alignments.sqlite,不在线ASR、翻译或重做对齐。6608条严格对齐失败的源资产沿用已建立的整句结束后保守释放映射。 ++- 麦克风窗长0.6秒、步长0.4秒,词末加0.1秒guard后在下一个释放点可见;动作时刻t只访问麦克风时刻≤t+0.2的已释放文本,保留最近30秒,编码器另有token上限。session切换时清空。 ++- 不输入未来PCM、未释放文字、GT turn/listen/idle标签、未来时刻、根路径或姿态。带噪未来训练目标不是额外观测条件。人体标签从配对SMPL-X按原生事件配方构造,不以G1关节角代替。 + +-GenMo默认启用执行段监督、预测历史恢复与CPU预取,使用新的输出目录;旧正式权重保留。训练链路和短程检查不代表已完成新策略的全量质量验收,详见[GenMo v2](docs/GENMO_V2_TRAINING.md)。四模型及OMG多个初始化权重的覆盖范围见[复现矩阵](docs/TRAINING_REPRODUCTION_MATRIX.md)。 ++## 五模型默认训练 + +-本次速度、断点恢复及过拟合检查的实际结果见[GenMo v2验证记录](docs/GENMO_V2_VALIDATION.md),包括仍存在的动作不足问题。 ++分别使用 /root/autodl-tmp/envs/模型名/bin/python,共同seed=42;精确超参数以配置和代码为准。 + +-从[训练复现指南](docs/PORTABLE_TRAINING.md)开始,准备对应环境和外部资产后生成本机配置: ++| 模型/配置 | 条件与输出 | 正式训练预算 | ++|---|---|---| ++| OMG-B:[omg.yaml](configs/omg.yaml) | 冻结T5-Base,50 token;过去2秒已释放PCM;10帧历史@30Hz。原生125D、G1全29关节,计划60帧,异步缓冲剩24帧重规划 | 50M,100K,batch512,BF16;LR6e-5,2000 warmup后cosine至1e-6 | ++| OMG-L:[omg_l.yaml](configs/omg_l.yaml) | 与B相同的HFD数据、因果条件、125D表示和diffusion-only微调损失;官方300M/55K初始化 | GPU0,100K,microbatch64×累积8=512;BF16,LR6e-5,warmup2000+cosine | ++| OMG-XL:[omg_xl.yaml](configs/omg_xl.yaml) | 与B相同的HFD协议;官方500M/50K初始化;沿用严格加载器,仅排除官方旧版52个未使用visual_film张量 | GPU1,100K,microbatch32×累积16=512;BF16,LR6e-5,warmup2000+cosine | ++| TextOp:[textop.yaml](configs/textop.yaml) | 冻结CLIP ViT-B/32,77 token含边界;2帧57D历史@50Hz,无音频。128D latent经冻结VAE解码8帧,G1原生23关节,未建模腕关节固定 | 300K,三个100K阶段,batch512,FP32,LR1e-4;保留原生分阶段历史恢复 | ++| GenMo:[genmo.yaml](configs/genmo.yaml) | GEM-SMPL全量训练,冻结T5-3B,50 token;60帧人体历史,past_audio=false。151D,未来60/执行12帧@30Hz | 100K,batch32,BF16,LR2e-5→1e-6;regression+diffusion、执行段权重3、根速度1.0、旋转增量0.05;预测历史最高25%,1000 warmup+4000 ramp,DDIM10 | ++| Kimodo:[kimodo.yaml](configs/kimodo.yaml) | 冻结LLM2Vec,最近400正文token,独立文本worker;60帧人体历史,无音频。SMPLX22原生273D,因果root分解、root→body两阶段,未来60/执行12帧 | 100K,batch8,BF16;normalized_x0+解码连续性/根历史重建;根只训投影,冻结根注意力骨干;根/身体分开优化裁剪,LR上限2e-6/1e-5 | ++| HY-Motion:[hymotion_c24.yaml](configs/hymotion_c24.yaml) | 官方Lite+rank16/alpha32 LoRA;冻结Qwen3-8B 128正文token+CLIP-L,无音频。201D,历史60/未来60/执行12帧;10段预测反馈,Euler10/CFG1,反馈detach | C24完整训练100K,batch4,BF16,LR2e-5 cosine至1e-6;24帧连续恢复,根速度1.0/旋转增量0.05,前200更新渐启,接段权重3 | + +-```bash +-python tools/configure_training.py \ +- --work-root /data/hfd_work --assets-root /data/hfd_assets \ +- --raw-root /data/hfd_assets/hfd_data --gpu 0 +-source /data/hfd_work/training-configs/training.env ++TextOp每条训练短链有4个8帧primitive,共0.64秒;链首始终使用GT历史,后续按课程选择预测历史,每个新batch重新初始化短链。第三阶段不等于整条session完全自主展开。 + +-/data/hfd_work/envs/omg/bin/python train.py --config /data/hfd_work/training-configs/omg.yaml +-/data/hfd_work/envs/textop/bin/python train.py --config /data/hfd_work/training-configs/textop.yaml +-/data/hfd_work/envs/genmo/bin/python train_genmo_full.py --config /data/hfd_work/training-configs/genmo.yaml +-/data/hfd_work/envs/kimodo/bin/python train_kimodo.py --config /data/hfd_work/training-configs/kimodo.yaml +-``` ++关节顺序及TextOp关节选择以 [joints.py](hfd_bridge/joints.py) 为准;沿用原生统计、冻结编码器/VAE、体型和表示变换。依赖与资产由 [requirements](requirements)、[UPSTREAM_LOCK.json](UPSTREAM_LOCK.json)、[TRAINING_ASSETS.json](TRAINING_ASSETS.json)、[HYMOTION_ASSETS.json](docs/HYMOTION_ASSETS.json) 记录。 + +-命令分别运行;同一GPU顺序执行。中断后原命令加`--resume auto`,Kimodo也支持`--resume`。恢复要求配置、代码、数据身份一致,预算是总步数。 ++## 缓存、存储与恢复 + +-- `hfd_bridge/`:训练数据、动作表示、适配、文本worker、验证与断点管理。 +-- `upstream/`:按训练依赖裁剪的HFD、OMG、TextOp原生代码,保留许可证;独立评测与部署代码归入本地`test/`。 +-- `tools/`:数据下载/对齐、配置生成、外部资源连接与构建。 +-- `requirements/`、`UPSTREAM_LOCK.json`、`TRAINING_ASSETS.json`:冻结环境、外部源码revision及必要外部文件的大小、SHA256和来源。 +-- `TRAINING_SNAPSHOT.json`:迁移前当前训练源码hash;Git历史保留对应实现。 +-- `docs/`:训练协议、复现操作和诊断说明,历史实验页不作为当前默认命令。 ++- 原始tar在 /root/autodl-fs/pangxinlei/hfd_data,保留分片、不批量解压;换索引无需删除原料。新下载默认共享盘models/datasets/downloads,下载/装依赖先用 /root/.local/bin/turbo,回退严格遵守 /root/AGENTS.md。 ++- 热索引、词映射、文本embedding、派生session放tmp。热资产 /root/autodl-tmp/cache/hfd/hot-assets 已约16.6 GiB有效内容,不用默认4 GiB准备命令缩小;长尾人体原料仍按需读共享tar。 ++- OMG/TextOp原生session内存缓存8 GiB,当前分别每256/64 batch刷新。GenMo train/val缓存4/8 GiB、文本512 MiB、预取16、训练每16 batch刷新、CPU4/BLAS1。GPU瞬时空闲不代表适合加batch或并发任务。 ++- 日志用 /root/autodl-tmp/runs;OMG/TextOp权重同目录,人体模型正式权重用 /root/autodl-fs/pangxinlei/models/模型名/。tmp给checkpoint和输出留空间;共享df不代表账号配额。 ++- 完整续训须恢复模型、优化器、scheduler、已消费采样池和原总预算,并校验代码/数据/资产合同;仅加载权重须明确重置优化器与步数。训练运行期间不改其合同约束源码/配置。 ++- OMG-B/TextOp每256步完整断点、每2000步验证;GenMo/Kimodo每2000步保存并验证。GenMo另在128/256/512/1024步滚动保存一份完整早期断点,正常2000步保存成功后清理早期断点,后续保留最近2份及best。 ++- OMG-L/XL先采样完整512窗口逻辑批,再切成GPU小批;保留B的512个session槽、每256逻辑批刷新及相同随机采样顺序。每2000次优化器更新验证8192窗口(16×512)并保存完整断点,保留最近2份及验证best;日志loss按完整更新聚合。`val_check_interval`已换算为小批次数,避免累积梯度时过频验证。 ++- OMG-L/XL日志与运行合同分别在`/root/autodl-tmp/runs/hfd-omg-{l,xl}-english-split20261026`,完整权重/优化器在共享盘`/root/autodl-fs/pangxinlei/models/OMG/hfd-{l,xl}-english-split20261026`,本地checkpoints/best/last路径为兼容链接。使用原生DirectTorchCheckpointIO流式临时文件+原子替换,恢复同时检查配置、原运行归属、RNG和已消费的逻辑批状态。 ++- GenMo环境固定NumPy1.26.4、SciPy1.15.3、Pandas2.2.3,主进程/CPU生产进程检查版本并写入合同及权重;保留无梯度AMP采样cache_enabled=False修复、主干梯度检查及严格回放。CPU审计失败时保存最近完成更新并停止。 + +-本地单元测试、过拟合/探针、评测/渲染代码和原始报告集中到`test/`,由`.gitignore`忽略。旧入口的兼容链接也被忽略。GitHub不包含这些代码、模型权重、数据、日志、环境、视频或缓存;CLIP自带的压缩分词词表是必要训练资源。原生validation及断点完整性检查属于训练实现。 ++以下仅用于任务停止后续训;不要对运行中的目录重复启动: + +-进一步阅读:[迁移操作](docs/PORTABLE_TRAINING.md)、[GenMo训练](docs/GENMO_FORMAL_TRAINING.md)、[Kimodo训练](docs/KIMODO_HFD_TRAINING.md)、[文本时间轴](docs/REFERENCE_DYNAMIC_TRAINING.md)、[resume](docs/TRAINING_RESUME.md)、[GenMo倾听姿态诊断](docs/GENMO_LISTENING_DIAGNOSIS.md)。 ++~~~bash ++cd /root/autodl-tmp/hfd-baseline-portable ++/root/autodl-tmp/envs/omg/bin/python -u train.py --config configs/omg.yaml --resume auto --set 'runtime.gpu_ids=[1]' ++/root/autodl-tmp/envs/omg/bin/python -u train.py --config configs/omg_l.yaml --resume auto ++/root/autodl-tmp/envs/omg/bin/python -u train.py --config configs/omg_xl.yaml --resume auto ++/root/autodl-tmp/envs/textop/bin/python -u train.py --config configs/textop.yaml --resume auto --set 'runtime.gpu_ids=[0]' ++/root/autodl-tmp/envs/genmo/bin/python -u train_genmo_full.py --config configs/genmo.yaml --resume auto ++/root/autodl-tmp/envs/kimodo/bin/python -u train_kimodo.py --config configs/kimodo.yaml --resume auto ++CUDA_VISIBLE_DEVICES=1 /root/autodl-tmp/envs/hymotion/bin/python -u train_hymotion_autoregressive.py --config configs/hymotion_c24.yaml --set gpu=0 --set text.device=cuda:0 --resume auto ++~~~ ++ ++HY训练入口为 train_hymotion_autoregressive.py,当前唯一正式配方为configs/hymotion_c24.yaml。train_hymotion.py只提供模型/合同公共函数,不再包含旧训练循环。 ++ ++## HY C24完整训练约定 ++ ++- 用户于2026-10-30授权完整训练及HY专项清理。原始训练数据、split、英文筛选、样本选择、采样顺序和时间映射不变;不加入静止历史增强。恢复监督在训练目标构造中计算,不回写GT。 ++- 从C24验证集选出的best仅加载权重。该best与上一轮C的2000步模型完全一致,已独立复制到共享models/hymotion/hfd-c24-full-split20261026/initialization.pt;初始化来源、SHA256和选择理由保存在新运行目录initialization.json。优化器/步数重置,新增100K预算,不声称完整续接旧优化器。 ++- 保留原XY/yaw对齐;姿态SO(3)差和高度差在24帧内按w=1−3s²+2s³消除。保留flow,根速度权重1.0、旋转增量0.05,前200更新渐启;执行段接段第一帧权重3,FP32几何、无效尾帧屏蔽。 ++- 预测历史策略仍以每链GT历史启动,之后反馈预测;每链10段共4秒,不能描述成跨整个session完全自主训练。训练数据和实际采样约定不因代码清理改变。 ++- 初始和每2000步用6条未参与旧38K选优的validation、两个固定噪声seed、最长20秒自主闭环。best按含运动误差的score选择,并要求平均接段比≤3;best只是当前最优,不等于行为验收合格。 ++- last每256步、验证结束及最终步原子保存;保留initial/best/last,不累计中间断点。启动先完成17步完整断点,再自动恢复原100K预算,检查恢复链路。 ++- 旧split完整训练runs/hfd-hymotion-english及共享models/hymotion/hfd-english保留;近期HY实验目录和中间权重清理后,正式训练不依赖它们。结果摘要见CURRENT_RUNS,必需初始化和基线指标保存在当前运行目录。 ++- 保留数据/连续性/恢复/模型回归检查;清除旧时钟输入、静止/单窗恢复实验分支和一次性对照脚本。当前测试评测入口为test/evaluation/evaluate_hymotion.py,可通过--run/--weights指定运行;仅在用户要求评测时启动GPU评测。 ++ ++## 评测约定 ++ ++自主闭环只开头给GT姿态,之后用预测历史,不逐窗重置GT;不做视频平滑、rootfit或DTW。人体模型先生成完整原始序列,再做整段GMR。训练loss或短smoke通过不等于行为、连续性和根运动验收;验证集用于选权重/调参,测试集只报告固定模型。 ++ ++## 动作评测脚本入口 ++ ++- 移植来源为 Tinghong-Ye22/HFD-motion-evaluation-handoff-20261101@7901e042c9cf06e0081d7303a4b2c229136ec70b;下载原件在 /root/autodl-fs/pangxinlei/downloads/HFD-motion-evaluation-handoff-20261101,工作目录在 /root/autodl-tmp/hfd-motion-evaluation/HFD_bench。 ++- 命令入口 /root/autodl-tmp/hfd-motion-evaluation/bench,使用 configs/local_motion.json;只读看板入口为同目录 review --port 8770。增量环境 /root/autodl-tmp/envs/hfd-motion-eval 复用 OMG 基础环境,不修改原训练环境;实际依赖见工作目录 requirements-local.lock。 ++- 冻结识别器5c04cec、阈值、GT+0.7秒计分参照和原始结果均保留;模型预测不平移。交互衡量发布的motion reference,非实机执行。UNKNOWN、未适用和未响应删失分别保留。 ++- 质量入口 motion-score 调用指定官方源码与本机已有 evaluator/T5-3B/HoloMotion,包含原生29DOF到MuJoCo的置换。旧113587的14项质量分数仍因 invalid_input_joint_order 失效;测试样例通过不代表新模型完整质量评测完成。大型参考缓存、视频和971502训练权重未全部下载,不直接启动旧调度/恢复队列。 ++ ++## 五模型全英文1024条动作测试集 ++ ++- 2026-11-01用户指定用于OMG、TextOp、GenMo、Kimodo、HY-Motion。独立派生集在 `/root/autodl-tmp/datasets/hfd-test-english1024-v1`:保留原1024个test场景及1937个动作轮,共2229个motion事件;原436条英文前景保留,588条中文前景重建为对应英文。动作场景、原始session身份和训练/验证划分不变,没有从其他划分补动作样本。 ++- 文件入口为 `manifest.jsonl`、`events.jsonl`、`membership.jsonl`、`evaluation_contract.json`;每例 `sessions/test-en-NNNN.npz` 包含50Hz IsaacLab G1与配对159D SMPL-X、16维betas、16kHz麦克风和评分标签。原文件保留,新版本不覆盖原始test合同或旧分数。 ++- 生成器与读取接口在 `/root/autodl-tmp/hfd-motion-evaluation/english1024`。使用OMG环境Python;`dataset_loader.English1024.input_at(index, motion_frame, model)` 只给初始已知姿态及已释放文本,后续历史由模型预测;`scoring_targets` 单独返回GT,`paired_human_30hz` 复用原人体几何变换。五模型各用相同1024例,不沿用旧人体provider的419条筛选。 ++- 英文均复用已有审计语音。更换语言后重算麦克风、控制事件和动作时间轴,沿用已发布的30帧终端时钟。2例改用其他合格英文音色,37例因英文提示语更长而最小延长停顿;20例中文第三方干扰换为上游held-out英文干扰,其中7例原前景已为英文。保留干扰类型、事件数及同音色aside关系;重叠背景按不相交聚合片段移动一次,aside避开前景指令。 ++- 文本协议仍为离线前景reference alignment,30秒上下文、0.1秒guard、microphone clock;它是oracle参考转写,不是新ASR结果,guard不代表经认证的时间误差上限。原训练/validation配置和已运行模型均未修改。 ++- 全量核验见数据集 `verification.json`:2002个实际英文语音资产SHA匹配;全部1024例G1/人体/时钟/语言/事件通过;419条原生英文的音频、动作和标签经旧加载器重放逐项一致;436条原英文动作轮标注一致,429例完整motion标注一致(另7例替换了中文背景,浮点时钟按1ns容差比较);35组任务×模型输入检查及2项混音/aside回归通过。 ++- 用户随后确认保留流式输入与listen/idle动作,已启动下述完整流式质量评测。交互识别/评分尚未重跑;本次质量使用预先固定的GT执行窗口,不需要推断预测首尾,也不复用旧971502分数。 ++ ++## 完整流式动作质量评测 ++ ++- 入口 `/root/autodl-tmp/hfd-motion-evaluation/stream-quality`;`status` 查看进度,`run --gpus 0 1` 启动或恢复(运行中会拒绝重复启动),`collect` 汇总已有评分。2026-11-01已启动,脚本在 `/root/autodl-tmp/hfd-motion-evaluation/stream_quality`;模型和训练代码未改。 ++- 协议 `full_stream_fixed_gt_windows_v1`:五模型逐包/逐次读取现有因果输入,单个已知姿态启动,之后只有执行过的预测历史。完整保留listen/idle/执行输出,不按GT任务边界重启,不把完整目标文本提前交给生成器。OMG保留过去2秒PCM及native async缓冲;其余模型保留既有因果reference文本协议,非在线ASR。 ++- 每模型生成全部1024例。1937个GT动作窗口在模型生成前按event ID排序,官方batch32取1920个,其余17个明确列入excluded;五模型成员一致。按固定绝对时间切出预测,响应慢、未响应或仍为listen的输出照常保留;不寻找最像GT的片段,不平移、不DTW、不平滑。模型没有收到这些评分窗口。 ++- 使用现有验证集best:OMG100000、GenMo72000、Kimodo100000、HY14000;TextOp使用最终300000。精确路径/SHA/配置见评测输出run_config.json、checkpoint_metadata.json、preflight.json。人体完整流先整段GMR,再切评分窗口,GMR只在session开头重置。 ++- 大文件在 `/root/autodl-fs/pangxinlei/evaluations/hfd-english1024-stream-quality-v1`;每模型 `models//sessions/` 保存完整预测及因果审计,`phase_diagnostics.json` 单独保存listen/motion/idle的关节、根误差和根速度,`packed/` 保存官方输入,`scores/motion_quality.json` 保存14项结果。汇总自动写summary.json/csv,不编造一个总分;阶段诊断不冒充官方14项。 ++- 日志与进度在 `/root/autodl-tmp/runs/hfd-five-stream-quality-english1024-v1`;两张GPU动态排队,CPU并行整段GMR,模型失败会保留错误并阻止该模型缺样本评分。GPU评分调用指定原版OmniHM/HoloMotion、tracker-mode=run与原有分块FK执行器。 ++- 恢复时校验并复用已完成的1024条生成结果与packed输入。评分经独立`stream_quality/score.py`调用交接包同一`OmniHM.evaluate`接口,将输出写入用户指定共享盘;仍执行冻结配置、官方源码及输入SHA校验。GMR读取进度文件遇到短暂缺失时等待下次读取,其他错误照常报告;summary同步显示运行/失败阶段。 ++- 启动前已通过五模型8秒流式生成、三模型整段GMR、4项因果/固定时窗回归,29个互异关节值经过baseline及指定官方转换后逐项一致。短检查不是完整1024评分,最终状态及分数以实际产物为准。 ++- 2026-11-02已完成五模型全部评分:各1024条完整生成、各1920固定窗口、14项均有值,完整结果见上述summary.json/csv与docs/CURRENT_RUNS.md。跟踪MPJPE等比较HoloMotion执行与生成目标;Diversity的GT参考约1.17965。显示FID已×100、C-Slide已×10,引用时保留该缩放。 ++ ++ ++## 隔离恢复训练 pilot(2026-11-09) ++ ++首轮 HY 与 Kimodo 的独立 500 更新恢复试验已结束,均未通过闭环验收;首轮入口 `/root/autodl-tmp/hfd-motion-evaluation/HFD_bench/scripts/recovery_training/supervise_pilot.py --config /config.yaml`,分别使用对应原训练环境。运行目录 `` 为共享盘 `models/hfd-recovery-pilot-20261109/{hymotion,kimodo}`。必须先有实际容量预留收据;supervisor 先做 2 更新保存与完整恢复检查,再继续固定预算。新试验加入单姿态启动的预测历史链和真实历史回放,HY 去掉 C24 混合标签,Kimodo 保留独立 expert 速度标签;从旧 best 显式权重初始化,保持原推理 steps/CFG,旧正式训练及已有成绩不替换。 ++ ++权重、缓存、临时文件和日志都有独立预算与持续磁盘/inode 检查。验证新增动作幅度、进度和误差非退化门槛;只有合格闭环 checkpoint 可成为新 best。500 更新结束不自动延长,不把 loss 降低等同于 FID 提升。具体预算、选优门槛、实时状态路径见 docs/CURRENT_RUNS.md。若受控停训,先排查对应错误收据并重新验证可写空间,再严格恢复,不能用删除合同检查的方式续跑。 ++ ++### 当前第二轮恢复试验 ++ ++第二轮已启动,入口改为 `HFD_bench/scripts/recovery_training_v2/supervise_pilot.py --config /config.yaml`,运行目录为共享盘 `models/hfd-recovery-pilot-v2-20261109/{hymotion,kimodo}`。从各自原正式 best 重新初始化,各 250 更新,0/125/250 固定验证;首轮输出保留,不采用其退化 last 替换正式模型。历史采样、推理 steps/CFG 和验收条件保持。 ++ ++监督改为 XY/yaw 对齐、六帧 SO(3)/高度过渡,第六帧起回到对齐 GT;位置与导数 loss 使用同一目标。HY 连续性辅助项恢复,warmup 200 更新,root velocity 权重 1、rotation velocity 0.05、边界权重 3。仅修改训练标签,推理不加入 GT 或平滑。增加段内突变、根 jerk 诊断及三次完整验证轨迹;125 更新额外保留一份权重快照(Kimodo 为 model-only,完整恢复使用 last)。只有原定门槛全部通过才写 best,尚无 FID 改善结论。 ++ ++实际新增容量预留 HY 2 GiB、Kimodo 8 GiB 已通过;checkpoint 单次写入上限分别为 256/2560 MiB,其他资源阈值和断点保护沿用首轮。第二轮进度命令为对应 Python 执行 `HFD_bench/scripts/status_recovery_training.py --root /root/autodl-fs/pangxinlei/models/hfd-recovery-pilot-v2-20261109`。证据与当前状态见 docs/CURRENT_RUNS.md。 ++ ++迭代资源维护:每轮核查已失效临时文件和可再生的重复缓存,确认进程结束、无打开句柄和保留副本有效后及时清理;当前运行的容量预留、有效断点和诊断证据保留。日志轮转、缓存限额、受控结束释放剩余预留继续执行。 ++ ++用户已授权每小时自主检查 HY/Kimodo 恢复训练,完成后分析并持续迭代,及时清理无用中间产物;验证效果有明确希望且空间预算通过时,可直接展开分阶段正式训练,无需例行再次确认。当前会话心跳为每小时整点(Etc/UTC),复用原 auto-review 权限;具体任务、状态与操作范围见 HFD_bench/artifacts/recovery_hourly_task.json 及 recovery_hourly_task_request.json。只用 train/validation 选参,保留既定验收,更多 validation 复核后再升级正式训练,最终测试协议保持。 ++ ++ ++### 第三轮恢复试验(2026-11-09 11:14 UTC) ++ ++第二轮两模型均已完成 250 更新,均未通过门槛。HY 最终活跃动作幅度比 0.358、姿态误差 0.443 rad(初始 0.372,恶化约 19%),行走案例 961 仍未响应;Kimodo 幅度比 0.842、接段比 2.355,但根位置误差 0.527 m(初始 0.442,恶化约 19%)且行走结果不稳定。不能升级正式训练,也没有新的正式 FID。 ++ ++当前最新目录为 `/root/autodl-fs/pangxinlei/models/hfd-recovery-pilot-v3-20261109/{hymotion,kimodo}`,入口 `HFD_bench/scripts/recovery_training_v3/supervise_pilot.py`。各从原正式 best 初始化,仍为 250 更新及固定 0/125/250 验证。每模型只改变一个训练变量:HY 预测历史链概率 0.75→0.25,增加真实历史回放;Kimodo 根网络从仅投影训练改为完整解冻,学习率不变。保留 C6 标签、因果输入、两随机种子与全部验收门槛。监督器已启动,先执行初始验证及两更新保存/严格恢复检查,未通过不会继续剩余预算。 ++ ++实际写入/fsync 预算 HY 2 GiB、Kimodo 12 GiB 已通过,Kimodo 完整断点预计 3.2 GiB、单次原子写入上限 4 GiB。第二轮重复启动索引已清理 656.24 MiB,正式权重、恢复断点、125 快照和完整轨迹保留。逐例结果、配置差异、启动与清理清单见 HFD_bench/artifacts/recovery_hourly_20261109T11_analysis.json、recovery_v3_config_diff.json、recovery_v3_launch.json、recovery_cleanup_hourly_20261109T11.json。每小时任务的 latest_run_root 已更新为 v3。 ++ ++ ++用户最新要求(2026-11-09):任一模型正式训练成功启动并确认进程、日志、容量及恢复检查正常后,立即删除每小时心跳任务 `416e9f54`,不等待训练或最终评测完成,不自动重建。当前有界恢复 pilot 仍继续每小时检查;取消心跳不停止训练,也不关闭训练监督器的资源与断点保护。此要求覆盖此前“最终评测完成后取消”的旧条件。 ++ ++ ++### 每小时检查与第四轮 Kimodo(2026-11-09 12:04 UTC) ++ ++当前两模型轮次不同:HY 仍在 `/root/autodl-fs/pangxinlei/models/hfd-recovery-pilot-v3-20261109/hymotion`,Kimodo 最新为 `/root/autodl-fs/pangxinlei/models/hfd-recovery-pilot-v4-20261109/kimodo`,以 `artifacts/recovery_hourly_task.json` 的 latest_models 分别定位。尚未启动正式训练,小时任务保持;任一正式训练健康启动后立即取消任务的要求不变。 ++ ++HY v3 在本次检查期间推进至 168/250,loss/梯度有限、无错误,完整两更新恢复检查通过(断点 136,897,993 bytes)。125 验证幅度比 0.261、接段比 1.672、姿态误差 0.423 rad、根位置误差 0.487 m,未通过;维持原有 250 更新上限,不延长或重启。 ++ ++Kimodo v3 已完成 250,两更新完整断点 3,394,252,981 bytes、恢复校验通过,无数值或资源故障。最终幅度比 0.942、接段比 2.299、姿态误差 0.422 rad,但根位置误差从 0.442 增至 0.675 m(约 +53%),根 jerk RMS 从 357.30 增至 700.46 m/s³。行走案例 961 两种子进度比 -0.466/-0.055;案例 2327 最大根速度仍有 10.98/18.81 m/s,尖峰仍落在每段第二帧。125 同样不合格。已读取 0/125/250 全部逐例完整轨迹,不能将动作幅度达标当成恢复成功,也无新的正式 FID。 ++ ++Kimodo v4 从相同原正式 best 重新开始,相对 v3 只新增 stage1_root_velocity 权重 0.5:对反归一化第一阶段根位置求物理速度,与相同 C6 因果目标比较,包含历史/未来边界并屏蔽无效帧。目标是阻止第一阶段错误被身体骨盆偏移短暂抵消。保持完整根网络解冻、学习率、250 更新、固定 0/125/250 验证、两个种子和原验收门槛;不改变推理或加入 GT 条件。新损失通过真实表征检查:精确目标为零、补偿式跳变仍被识别、梯度有限且仅作用于根分量、无效尾帧不计分,见 artifacts/recovery_v4_objective_checks.json。源码入口 scripts/recovery_training_v4/supervise_pilot.py;准备与运行独立于 HY v3。 ++ ++第四轮 12 GiB 预算已实际写入/fsync,通过物理分配校验;4 GiB 原子写入上限、8 小时墙钟及原存储阈值不变。监督器已启动,先进行初始验证和两更新完整恢复检查,该检查尚待本轮执行完成。当前本地约 1.54 GiB 空闲、共享约 34,900 空闲 inode,小文件写入/fsync 通过;没有把共享 df 总空闲当账号配额。 ++ ++清理 v3 Kimodo 的 16 个重复启动索引及已分析且不再使用的 125 步落败快照,共 1,407.06 MiB。已核对进程结束、无打开句柄、无配置引用、缓存保留副本 SHA 相同;保留最终完整断点、原正式 best、必要旧对照及全部验证原始轨迹。清单 artifacts/recovery_cleanup_hourly_20261109T12.json。分析、配置差异与启动收据分别为 recovery_hourly_20261109T12_analysis.json、recovery_v4_config_diff.json、recovery_v4_launch.json。下一检查点为 HY 最终验证及 Kimodo v4 恢复/125 验证,不因平均指标改善放宽逐例行走门槛。 ++ ++ ++### 第五轮会话覆盖对照(2026-11-09 13:05 UTC) ++ ++最新两模型统一位于 `/root/autodl-fs/pangxinlei/models/hfd-recovery-pilot-v5-20261109/{hymotion,kimodo}`,入口 `HFD_bench/scripts/recovery_training_v5/supervise_pilot.py`。HY v3 和 Kimodo v4 均已完成 250 更新,完整恢复检查通过、训练数值有限、无资源错误,但未通过闭环门槛。尚未正式开训,小时检查继续;任一正式训练健康启动后立即取消心跳。 ++ ++HY v3 最终幅度比 0.263、接段比 1.686、姿态误差 0.427 rad、根位置误差 0.467 m、根 jerk 279.24 m/s³;两行走案例四种子组合进度比 0.336/-0.037/0.352/-0.010,均低于下限。Kimodo v4 最终幅度比 0.856、接段比 2.336、姿态误差 0.408 rad、根位置误差 0.564 m、根 jerk 376.07 m/s³。相对 v3 根 jerk 改善,但行走仍不可靠:案例 2327 两种子进度比 0.046/0.790,961 为 -0.168/0.653。v4 的 125 步姿态/位置非退化门槛通过,但两例进度未通过,故保留该快照诊断,不升级正式训练。已读取各轮初始/125/250 全部逐例完整轨迹;这些不是正式 FID。 ++ ++采样链路发现:WindowPool.refresh 按链起点批次计数,不按优化器更新计数。Kimodo v4 最终完整断点的 pool.step=6、refresh=16、slots=[32781,251404,112021,179604],证明 250 更新一直使用最初四个会话;HY v3 pool.step=7、refresh=2,日志实际覆盖 16 个不同会话。长链重复同一池显著限制本轮样本覆盖,但不能据此认定它是质量差的唯一原因。第五轮每模型只将 data.refresh_batches 改为 1(Kimodo 16→1,HY 2→1),每批新链重新抽取会话;仍保留最多 50 段闭环链、原概率、各自上一轮损失和学习率、同一原正式 best 初始化及全部验收门槛,各 250 更新,0/125/250 验证。没有读取 Test1024 选参。Kimodo 日志新增实际采样位置和池计数,便于后续核查覆盖。 ++ ++真实 WindowPool 的受控测试验证六批起点在 refresh16/2/1 时分别覆盖 4/12/24 个模拟会话,保存并恢复后下一批完全一致;实际覆盖仍由运行日志判断。检查和差异收据为 artifacts/recovery_v5_sampling_checks.json、recovery_v5_config_diff.json,启动源码 SHA 与进程在 recovery_v5_launch.json。原因果标签及推理实现未改;两更新完整断点/严格恢复仍由监督器先验证,第五轮目前处于该启动阶段,尚无训练后改善结论。 ++ ++HY 2 GiB、Kimodo 12 GiB 预算已重新实际写入/fsync 并确认物理分配;本地约 1.54 GiB 空闲,共享约 34,800 空闲 inode。保持原子替换预算、8 小时墙钟、目录项/缓存/日志上限及资源受控停止,不将共享 df 视为账号配额。清理结束轮次的重复启动索引及 HY v3 125 落败快照 786.77 MiB,已检查无进程占用、无必要配置引用、缓存 SHA 一致,保留当前完整断点、Kimodo v4 125 候选及全部轨迹。清单 recovery_cleanup_hourly_20261109T13.json,完整分析 recovery_hourly_20261109T13_analysis.json。状态脚本已修复缺少某模型目录时的报错,返回 present=false;没有修改任何旧训练合同源码。 ++ ++下一检查点为第五轮两模型恢复收据、实际会话覆盖以及 125/250 固定验证。失败时继续依据证据调整,不能仅延长当前失败配方或降低验收条件。 ++ ++ ++### 小时检查与 Kimodo 第六轮(2026-11-09 14:07 UTC) ++ ++当前 HY 仍为 `models/hfd-recovery-pilot-v5-20261109/hymotion`,Kimodo 最新为 `models/hfd-recovery-pilot-v6-20261109/kimodo`,根目录均在 `/root/autodl-fs/pangxinlei/`,以任务 JSON 的 latest_models 为准。没有正式训练启动;达到正式启动并核实健康条件后,仍应立即删除心跳 416e9f54。 ++ ++HY 本次检查从 138 推进到约 154/250;恢复检查通过(136,897,929 bytes)、训练数值有限、无错误,实际已记录 17 个训练会话。125 验证幅度比 0.252、接段比 1.661、姿态误差 0.424 rad、根位置误差 0.478 m,未通过。保持当前有界训练与原定最终验证,不抢占 GPU0。 ++ ++Kimodo v5 完成 250 更新,严格恢复通过(3,394,253,109 bytes),实际训练覆盖 25 个会话。最终幅度比 0.919、接段比 2.365、姿态误差 0.409 rad、根位置误差 0.371 m;相比原初始 0.468 rad/0.442 m,两种误差均改善。但行走案例 2327 两种子进度 0.496/0.519、961 为 0.547/0.227,仍有低于 0.5 的案例,不能四舍五入放宽门槛。根 jerk RMS 411.10 m/s³,高于初始 357.30,段首尖峰仍存在。125 的根误差 0.592 m、jerk 606.01,亦未通过。已读取 0/125/250 全部逐例原始轨迹;平均数改善不等于正式训练资格,也不是 FID 改善。 ++ ++完成同一 v5 final 权重下四个 validation 历史窗口、两种子、预测过去与真实过去的 16 次配对诊断,输出 `/root/autodl-fs/pangxinlei/models/hfd-recovery-history-diagnostics-v5-20261109/`。输入仅为截至 decision 的历史,无未来 GT;真实历史是诊断对照,不能作为正式验证或替代闭环成绩。案例 2327 decision396 的第二帧最大根速度从预测历史约 5.96/5.97 m/s 降至真实过去 0.12/0.32,但 decision504 真实过去仍约 2.85/2.63;961 decision276 真实过去仍有 2.32/4.12 m/s。说明历史累积参与部分异常,段首问题也能在真实过去下出现。低速度窗口可能只是静止,不能认作正确完成行走。 ++ ++第六轮从同一原正式 best 重新开始,保持第五轮采样刷新1、原链长和概率、完整根解冻、学习率、stage1 速度权重0.5、250 更新及两种子固定验证,只新增 stage1_boundary_weight=3:第一阶段根速度的历史末帧→首未来帧及首→第二未来帧两条差分加权3倍。推理无平滑、无 GT,验收不变。真实表征检查确认边界/内部误差比3、精确目标零、无效尾帧忽略、有限且仅作用于根分量的梯度;收据 recovery_v6_boundary_checks.json,配置差异 recovery_v6_config_diff.json,入口 scripts/recovery_training_v6/supervise_pilot.py。监督器已启动,初始验证和两更新完整恢复检查尚待完成,不声称本轮已改善。 ++ ++新增 12 GiB 预算已实际写入/fsync 并验证物理分配,保留4 GiB原子 checkpoint 上限、8小时墙钟及原资源门槛。本地约1.54 GiB空闲、共享约34,750空闲inode,小文件写入/fsync通过。已清理 v5 Kimodo 125 落败快照及结束诊断后的重复索引共1,407.06 MiB;无进程句柄和必要配置引用,缓存保留副本 SHA 一致。保留 v5 最终完整断点、v4 125候选、所有轨迹与活跃 HY 预留。诊断预留已释放。清单 recovery_cleanup_hourly_20261109T14.json;完整分析 recovery_hourly_20261109T14_analysis.json;启动收据 recovery_v6_launch.json。下一检查点是 HY 最终结果、Kimodo v6恢复收据与固定验证,而非直接延长未通过配方。 ++ ++ ++### 第七轮监督权重对照(2026-11-09 15:06 UTC) ++ ++当前两模型均进入 `/root/autodl-fs/pangxinlei/models/hfd-recovery-pilot-v7-20261109/{hymotion,kimodo}`,入口 `HFD_bench/scripts/recovery_training_v7/supervise_pilot.py`。上一轮 HY v5、Kimodo v6 均正常完成250更新,恢复检查通过、训练数值有限,未通过闭环验收;无正式训练,继续小时调度。任一正式训练成功启动并验证健康后立即取消心跳的要求保持。 ++ ++HY v5 最终幅度比0.264、接段比1.666、姿态误差0.429 rad、根位置误差0.487 m、根jerk285.94 m/s³。行走案例2327两种子进度0.279/0.289、961为-0.008/0.004,明显不合格。Kimodo v6 最终幅度比1.018、接段比2.283、姿态误差0.412 rad、根位置误差0.362 m、根jerk330.86 m/s³;虽然jerk低于原初始357.30,进度仍不可靠:2327为0.114/0.510,961为0.885/0.285。两个模型都已读取初始/125/250全部逐例完整轨迹,保存位置/姿态/速度误差、幅度、接段/段内突变、jerk与路径;中途候选也未通过,不用平均改善掩盖逐例失败,未计算新的正式FID。 ++ ++第七轮每模型仅改一个参数,保持原正式best初始化、固定训练种子、refresh1会话采样、原链长/概率、学习率、C6标签、推理及原验收不变,仍250更新、0/125/250固定验证。HY根速度辅助权重1→0.1,旋转速度权重0.05和warmup200保持;依据是v5后期根速度项数值均值约0.102,大于flow约0.071,同时幅度和姿态退化,这只是权重失衡假设,不能将损失数值比较当成梯度主导的证明。Kimodo closed_loop.tail_weight由0.25→1,完整监督原有60帧预测,不再将后48帧降权;检验局部接段改善却仍走不够是否与短时域拟合有关。执行仍只用前12帧,stage1_boundary_weight=3保留,不加推理平滑或未来GT。 ++ ++真实函数检查确认尾段权重1时原有效性掩码、历史条件和未定义末帧导数掩码完全保留、输入不被修改;HY调度只改变根速度系数。检查、配置差异、源码SHA与启动收据为 artifacts/recovery_v7_weighting_checks.json、recovery_v7_config_diff.json、recovery_v7_launch.json。两监督器真实进程已启动,先初始验证和两更新完整保存/严格恢复,本轮恢复检查尚未结束,不能声称训练效果已改善。本轮选参只用train/validation。 ++ ++已重新实际写入/fsync HY2 GiB与Kimodo12 GiB预算并验证物理分配,原子替换上限256 MiB/4 GiB、8小时墙钟、目录项/缓存/日志及空间inode护栏均保持。本地约1.54 GiB空闲,共享约34,690空闲inode,小文件实写/fsync通过,未把共享挂载总空闲视为账号配额。 ++ ++清理约4.98 GiB:已结束轮次的重复启动索引、HY v5及Kimodo v6的125步落败快照,以及不再需要的Kimodo v3最终完整断点。删除前核对运行结束、无打开句柄和配置输入依赖,缓存保留副本哈希一致。v3的配置/源码/完整原始轨迹/指标仍保留;保留当前HY v5、Kimodo v6完整断点、较好Kimodo v5完整候选、v4的125候选和监督审计所需旧断点,不无限累积落败权重。清单 recovery_cleanup_hourly_20261109T15.json,完整分析 recovery_hourly_20261109T15_analysis.json。下次先检查v7恢复与资源健康,任一完成即分析逐例,不机械延长失败配方。 ++ ++ ++### 小时检查与 Kimodo 原生初始化对照(2026-11-09 16 UTC) ++ ++HY v7仍运行,检查时194/250更新,数值有限、两步完整恢复检查通过、无错误。125验证幅度比0.269、姿态误差0.443 rad、根位置误差0.464 m,不合格,继续原定有界运行。 ++ ++Kimodo v7正常完成250更新、完整恢复通过,已读取初始/125/250全部逐例轨迹。最终幅度比0.770、接段比2.394、姿态误差0.422 rad、根位置误差0.376 m;原正式初始分别为幅度0.405、姿态0.468 rad、根位置0.442 m。行走案例2327两种子进度0.513/0.896、961为0.911/0.346;最后一个未达到0.5。根jerk511.39 m/s³高于初始357.30和v6最终330.86;段内姿态变化p95均值0.07275 rad,GT为0.05927。125根误差0.496 m、jerk643.45亦不合格。不能因平均误差改善晋升正式训练,也未计算新FID。 ++ ++第八轮仅Kimodo:`/root/autodl-fs/pangxinlei/models/hfd-recovery-pilot-v8-20261110/kimodo`,入口`HFD_bench/scripts/recovery_training_v8/supervise_pilot.py`。目录日期采用当前会话日期,检查收据时间来自机器UTC。只把initial_checkpoint从旧HFD正式best改为None,由上游配置加载原生Kimodo-SMPLX-RP-v1/model.safetensors。CPU逐张量精确核验408个张量、282790161个元素与原文件完全一致,非随机初始化。其余学习率、完整根训练、250更新、采样、C6标签、因果输入、推理及0/125/250两种子验证不变。复制原HFD初始baseline-validation.json且逐字节核对一致,误差门槛仍以原正式初始为准,不能因原生初始化改变验收。 ++ ++新监督器558793、子进程558794已核实存活,处于初始验证/两更新恢复检查阶段;本轮恢复尚未通过,效果未知。收据artifacts/recovery_v8_initialization_check.json、recovery_v8_config_diff.json、recovery_v8_launch.json。新预算12 GiB已实写/fsync并验证物理分配,4 GiB原子保存上限及原空间/inode/目录项/日志/8小时保护保持。本地约1.54 GiB可用、共享约34600空闲inode,本地及共享小文件实写通过;不将共享df当账号配额。 ++ ++已清理v7 Kimodo重复索引和125落败快照1407.06 MiB,核实结束进程、无打开句柄、快照无配置输入引用,索引与v8保留副本SHA一致。保留v7最终完整恢复断点、必要旧候选、配置与全部轨迹;活跃HY预留不动。清单artifacts/recovery_cleanup_hourly_20261109T16.json;逐例分析recovery_hourly_20261109T16_analysis.json。 ++ ++无正式训练启动,小时心跳416e9f54继续;一旦正式训练成功启动并验证健康,立即删除心跳,不停止训练监督器。下次先检查HY最终结果和Kimodo v8完整恢复、初始及训练后闭环验证。不把本次原生初始化假设当成已证实改善。 ++ ++ ++### 第九轮有界对照(2026-11-09 17 UTC) ++ ++HY v7、Kimodo v8均完成250更新,完整两更新保存/恢复通过,训练及已读取0/125/250全部逐例轨迹数值有限,无错误或资源受控停止。没有正式训练;小时心跳416e9f54保持,正式训练成功启动并核实健康后立即删除。 ++ ++HY v7降低根速度辅助权重后,最终幅度比0.272、接段比2.044、身体姿态误差0.450 rad、根位置误差0.478 m、jerk291.97 m/s³。原初始姿态0.372 rad,最终退化超过10%,且有1个幅度坍缩案例;行走2327两种子进度0.350/0.315,961约0/0.005。中途125亦未过关,不能将较低jerk解释为动作恢复。 ++ ++Kimodo v8原生初始化在当前因果适配接口下,初始幅度1.258、姿态0.603 rad、根位置0.608 m、jerk723.84;最终幅度1.545、接段比2.948、姿态0.601 rad、根位置0.569 m、jerk758.54。段内姿态变化p95均值0.10262 rad(GT0.05927);行走2327进度0.014/-0.019,961为-0.029/-0.036。125也不合格。原生权重增加运动幅度但未恢复正确进度,不支持“换回原生权重即可解决”,也不能据此否定原生模型在其原协议下的能力;当前结果包含因果表征、历史输入及任务域的适配影响。未计算新FID。 ++ ++下一轮根目录`/root/autodl-fs/pangxinlei/models/hfd-recovery-pilot-v9-20261110/{hymotion,kimodo}`,入口`HFD_bench/scripts/recovery_training_v9/supervise_pilot.py`。HY以v7为对照,仅initial_checkpoint改为None,build_model仍加载官方latest.ckpt。CPU核对302个backbone张量及null文本参数与官方文件完全一致,新增观察列为0、LoRA B为0,非随机网络;其余损失和配置不变。Kimodo以更好的v7为对照(不是从v8续训),保持原正式best初始化,仅新增stage1_root_displacement权重0.5:对有效未来帧的第一阶段物理根位置相对最后历史帧的位移,用beta0.1的SmoothL1监督并按score掩码归一。检验短时速度监督下仍走不够的问题;不改变推理、标签、历史构造或执行长度。测试确认精确目标损失0、无效尾段与空掩码忽略、整体平移不变、有限且仅根通道梯度。此项是假设检验,不预设改善。 ++ ++两模型仍250更新、0/125/250固定validation与两个种子,train/validation选参,Test1024未使用。原HFD初始baseline-validation.json逐字节保留,误差门槛不随新初始化放宽。配置差异仅指定变量与路径/元数据,收据artifacts/recovery_v9_config_diff.json;测试recovery_v9_displacement_check.json、recovery_v9_hy_initialization_check.json;启动源码SHA和进程分别recovery_v9_hymotion_launch.json、recovery_v9_kimodo_launch.json。两监督器及子进程已启动,目前初始验证/两更新恢复检查尚未结束,不声称本轮恢复或效果已通过。 ++ ++HY2 GiB、Kimodo12 GiB预算重新实写/fsync并验证物理分配,256 MiB/4 GiB原子checkpoint上限、8小时墙钟、空间/inode/日志/文件数限制保持。本地约1.54 GiB可用、共享约34500空闲inode,另做本地/共享小文件实写;共享df不代表账号配额。清理HY v7与Kimodo v8重复启动索引及125落败快照1865.72 MiB,核实进程结束、无打开句柄、缓存与v9副本SHA一致、快照无配置输入引用。保留最终完整断点、原正式权重、少量既有对照、全部轨迹。清单artifacts/recovery_cleanup_hourly_20261109T17.json,完整分析recovery_hourly_20261109T17_analysis.json。下次先检查v9恢复收据和资源健康,完成后按原门槛分析逐例。 ++ ++ ++### 小时检查与 Kimodo 第十轮(2026-11-09 18 UTC) ++ ++HY v9仍运行,检查时195→198/250更新,真实进程560419存活,GPU0约20.6 GiB,loss和已记录梯度有限,完整两步恢复通过,无错误或资源停止。125验证幅度1.952、接段比5.406、姿态误差0.925 rad、根位置0.717 m,明显未通过。保留当前有界实验等待最终验证,不重复启动或占用其GPU。 ++ ++Kimodo v9完成250,完整恢复检查通过,无非有限值,实际训练覆盖25会话,记录中预测历史槽位占比0.835(设定抽样概率0.75;长链会影响实际槽位占比)。已读取初始/125/250全部逐例完整轨迹及指标。最终幅度0.730、接段2.450、姿态0.404 rad、根位置0.396 m、jerk353.50 m/s³;与v7的0.422 rad、0.376 m、511.39相比,姿态与jerk改善,根位置略差但仍优于原初始0.442 m。段内姿态p95均值0.06497 rad,GT0.05927。没有幅度坍缩,但2327两种子进度0.332/1.090、961为0.517/0.261,仍有两个低于0.5。125根误差0.497 m、jerk642.08、961种子2026进度0.272,亦不合格。不能据此正式开训,不将诊断指标当FID。 ++ ++第十轮仅Kimodo,目录`/root/autodl-fs/pangxinlei/models/hfd-recovery-pilot-v10-20261110/kimodo`,入口`HFD_bench/scripts/recovery_training_v10/supervise_pilot.py`。以v9为对照,仅pilot.predicted_probability由0.75降至0.25;假设长预测历史反馈误差限制正确运动监督,增加真实过去训练占比进行检验。保持原正式best初始化、位移权重0.5、所有损失、学习率、采样刷新、链长、C6标签和250更新。运行源码除准备器外逐字节一致;推理完全闭环、无未来GT,不把真实过去诊断当正式成绩。原初始baseline-validation.json逐字节保留,验收与0/125/250两种子协议不变,Test1024未使用。配置核验artifacts/recovery_v10_config_diff.json,启动进程和源码SHA在recovery_v10_launch.json。 ++ ++12 GiB预算已重新实写/fsync且物理分配通过,保留4 GiB原子checkpoint上限、8小时墙钟、空间/inode/文件数/日志保护。监督器561515及子进程已核实启动,处于初始验证和两步恢复检查,尚无本轮改善结论。本地约1.54 GiB可用、共享约34480空闲inode,本地与共享小文件实写通过;不把共享df当账号配额。活跃HY预留不动。 ++ ++清理4.535 GiB:v9 Kimodo重复启动索引、125落败快照,以及已分析且无后续用途的v8 Kimodo原生初始化最终权重。删除前核实完成收据、进程结束、无打开句柄、无配置输入引用,缓存与v10副本SHA一致。保留v9当前完整断点、较优既有对照、原正式权重,以及v8配置/全部轨迹/指标/原生权重核验;不再无限积累落败完整权重。清单artifacts/recovery_cleanup_hourly_20261109T18.json,完整分析recovery_hourly_20261109T18_analysis.json。 ++ ++当前无正式训练,心跳416e9f54继续。下次先检查HY最终结果与Kimodo v10恢复收据、资源和逐例验证;任何正式训练健康启动后立即取消心跳,不终止监督器。 ++ ++ ++### 小时检查与固定权重推理诊断 v11(2026-11-09 19 UTC) ++ ++HY v9及Kimodo v10均正常完成250更新,完整两步恢复通过,训练loss/已记录梯度与0/125/250全部读取的完整轨迹有限,无错误、无资源停止、预留已释放。均未通过固定门槛,没有正式训练,小时心跳416e9f54保持。 ++ ++HY v9原生初始化最终幅度1.888、接段5.423、姿态0.882 rad、根位置0.710 m、jerk2306.62 m/s³;原生初始分别为幅度1.694、接段5.710、姿态1.061 rad、根位置0.811 m、jerk2498.46。训练有局部改善但仍远差于原HFD正式初始姿态0.372/根位置0.449,验收不随初始化放宽。行走2327进度0.024/0.116、961为-0.173/-0.014;无幅度坍缩不代表正确运动。段内姿态p95均值0.11329 rad,GT0.05927。125也不合格。 ++ ++Kimodo v10降低预测历史抽样后,最终幅度0.723、接段2.020、姿态0.460 rad、根位置0.429 m、jerk408.32;125分别0.616、2.144、0.466 rad、0.431 m、329.05。最终行走2327进度1.097/0.882,但961为-0.015/0.285;较v9的961进度0.517/0.261,第一种子退化。因此不接受“增加真实历史一定改善”的假设,也不凭更平滑中途结果正式开训。没有新的正式FID。 ++ ++连续单变量训练配方未带来稳定逐例通过,本轮先做有界推理诊断,暂不再追加新的损失权重或延长失败训练。目录`/root/autodl-fs/pangxinlei/models/hfd-recovery-inference-v11-20261110/{hymotion,kimodo}`,入口`HFD_bench/scripts/recovery_inference_v11/supervise_pilot.py`。HY固定原HFD正式best,比较10/50采样步;Kimodo固定v9最终完整权重,比较20/100步。各自CFG、相同6个validation案例、两个种子、因果历史、坐标转换、执行长度和原验收不变;不使用Test1024。先跑原步数对照用于复现已有结果,然后跑更细采样。先前Kimodo单窗口20/100步诊断不能替代这次完整闭环比较。 ++ ++诊断分支在优化器创建前返回,零训练更新、不保存训练断点,不计作正式训练或恢复续训;无需伪造两步恢复检查。复用原监督器空间/inode/目录项保护、日志轮转和受控终止,单次墙钟上限4小时,完成后释放预留。静态检查artifacts/recovery_v11_static_check.json,固定输入路径和SHA及基线一致性recovery_v11_inputs.json,启动源码SHA/进程recovery_v11_launch.json。输出为logs/diagnostic.jsonl及validation-traces/step-000010或000050等(此处编号是采样步数,不是训练更新);完成收据标记training_updates=0。状态脚本last_validation可能为空,应读取diagnostic.jsonl。 ++ ++两监督器562275/562278及子进程562279/562280已核实存活,无启动错误;结果尚待完成。每模型2 GiB预算已实写/fsync且验证物理分配,因无新训练checkpoint只预算有界缓存、日志和两套轨迹。本地约1.54 GiB可用、共享约34400空闲inode,本地/共享小文件实写通过,不将共享df当配额。 ++ ++清理1.822 GiB:HY v9和Kimodo v10已结束轮次的重复启动索引、125不合格快照;删除前核验进程结束、无打开句柄、快照无配置引用、缓存与诊断副本SHA一致。保留最终完整断点、原正式权重、诊断所用Kimodo v9完整权重、其他必要对照及全部原始轨迹。清单artifacts/recovery_cleanup_hourly_20261109T19.json;完整分析recovery_hourly_20261109T19_analysis.json。下次检查诊断完成/错误、原步数复现及增加步数的逐例影响,再据此决定训练改动;不将诊断直接当正式训练资格。任一正式训练健康启动后立即取消心跳的条件保持。 ++ ++ ++### 小时检查、HY精度修复与Kimodo v12(2026-11-09 20 UTC) ++ ++v11两模型诊断均正常完成,零训练更新,无错误/资源停止,预留释放。读取全部两套完整逐例轨迹,数值有限,输入权重SHA与启动前完全一致。没有正式训练,心跳416e9f54继续。 ++ ++Kimodo20步与v9最终全部预测及参考轨迹逐元素相同,确认复现。100步幅度1.004、接段2.255、姿态0.451 rad、根位置0.434 m、jerk542.36 m/s³;20步分别0.730、2.450、0.404 rad、0.396 m、353.50。100步行走2327进度0.670/0.863,961为-0.003/0.487,未达标且jerk升高,不能用更多采样步替代训练修复,不采用100步为新默认。 ++ ++HY v11发现诊断入口缺少原训练验证的BF16 autocast上下文。10步原协议复现失败,预测轨迹最大元素绝对差0.5620(混合位置/旋转通道,不作为单一物理量解释);其jerk255.71与原基线335.19不同。v11的10/50步结果仍保留作FP32诊断证据,但不能用于判断原BF16协议增加步数的收益,也不能与既有BF16指标混合统计。原因是新增诊断分支只包torch.no_grad,原validate同时包torch.autocast;旧训练结果未受该新增分支影响。 ++ ++HY v12复制独立源码到scripts/recovery_inference_v12,添加与原validate一致的BF16 autocast和validation seed;目录`/root/autodl-fs/pangxinlei/models/hfd-recovery-inference-v12-20261110/hymotion`。仍固定原正式best、10/50步、同一6案例与两个种子,零更新,4小时监督上限,原误差基线不变。监督器563273、子进程563276已核实存活。下一轮必须先验证10步复现;若仍不一致,继续排查文本缓存/随机上下文,不先宣布改善。诊断输出logs/diagnostic.jsonl、validation-traces,编号是采样步数。 ++ ++Kimodo v12以v9为基线,仅预测历史概率0.75→1.0。v10增加真实历史后961第一种子进度从v9的0.517退到-0.015,因此测试增加闭环暴露,而非延长失败预算。目录`/root/autodl-fs/pangxinlei/models/hfd-recovery-pilot-v12-20261110/kimodo`,入口scripts/recovery_training_v12/supervise_pilot.py。保持原正式best初始化、位移监督0.5、所有其他损失/学习率/采样刷新/链长/推理、250更新和0/125/250固定两种子验证;不把v10结果作为初始化,不改验收,不用Test1024。除准备器外训练与推理源码和v9逐字节一致,原baseline-validation.json逐字节一致。监督器563316及训练子进程已启动,目前初始验证/两更新完整恢复检查阶段,恢复检查尚未完成、效果未知。 ++ ++HY2 GiB、Kimodo12 GiB预算重新实写/fsync并确认物理分配,保留原资源保护;本地约1.54 GiB可用、共享约34340空闲inode,本地与共享写入探针通过,不把df总量当账号配额。清理v11已完成诊断的重复索引656.24 MiB,核对进程结束、无打开句柄和v12保留副本SHA一致,保留所有诊断轨迹/指标/配置与正式及必要候选权重。清单artifacts/recovery_cleanup_hourly_20261109T20.json;完整分析recovery_hourly_20261109T20_analysis.json;配置/静态核验recovery_v12_config_diff.json、recovery_v12_static_check.json;两启动收据recovery_v12_hymotion_launch.json、recovery_v12_kimodo_launch.json。任一正式训练健康启动后立即删除心跳的要求保持。 ++ ++ ++### 小时检查与文本条件诊断 v13(2026-11-09 21 UTC) ++ ++HY v12固定权重诊断和Kimodo v12训练均正常完成,预留已释放,无错误/资源停止。已读取所有完整轨迹且有限;Kimodo训练loss/记录梯度有限、完整两更新恢复通过。无正式训练,心跳416e9f54继续。 ++ ++HY修复BF16上下文后,10步所有预测/参考轨迹与原正式初始逐元素完全一致,确认复现问题解决。50步幅度0.489、接段1.324、姿态0.404 rad、根位置0.498 m、jerk304.67;10步分别0.291、1.290、0.372 rad、0.449 m、335.19。50步幅度仍低于0.5,根位置超过原初始10%上限,行走2327进度1.188/1.869、961为-0.222/-0.283,仍失败。不放宽门槛、不采用50步作为已通过方案。 ++ ++Kimodo v12全预测历史250更新最终幅度1.100、接段2.386、姿态0.439 rad、根位置0.441 m、jerk405.63。行走2327进度0.827/0.678,961为0.107/0.510,仍有一个失败;125时961为1.108/0.390,jerk890.09。中途和最终对同一案例不同种子不稳定,不能以均值改善晋升正式训练,也不机械延长250预算。所有诊断都不是新FID。 ++ ++读取既有validation历史诊断的实际因果文本:961在decision276/432是“turns left Hmm... walks forward”;2327在decision396/504是“interchanging an item between hands stop jog forward left diagonally”。Source.condition按已发布时钟取完整30秒参考转录上下文,未用GT turn或控制标签选新指令。文本中旧/新动作混合是可检验假设,尚不能断言根因或擅自丢弃旧指令。 ++ ++v13固定权重原文本/空文本对照已启动,目录`/root/autodl-fs/pangxinlei/models/hfd-recovery-text-v13-20261110/{hymotion,kimodo}`,入口scripts/recovery_text_v13/supervise_pilot.py。HY用原正式best、10步、BF16;Kimodo用v12最终完整权重、20步;同一6个validation案例、两种子、CFG、因果历史、原始验收和其余推理保持。两臂只在text.encode入口把已发布文本保留或替换为空字符串;空文本是编码空转录,并非假定Kimodo得到严格CFG无条件向量。没有未来文本、GT动作/标签选择,也不改变训练或评测门槛。原文本先跑以检验复现。该诊断测量文本敏感性,空文本即使分数变好也不能当任务成功或正式训练资格。 ++ ++零训练更新、无优化器更新、无新训练checkpoint,4小时监督上限;输出logs/diagnostic.jsonl的arm字段为original/empty,validation-traces/step-000000与step-000001对应两臂,不是训练更新。静态检查artifacts/recovery_v13_static_check.json,输入权重SHA/步数recovery_v13_inputs.json,源码SHA/进程recovery_v13_launch.json。监督器564572/564576及子进程已核实存活,尚无结果结论。 ++ ++每模型2 GiB预算实写/fsync验证通过,保持磁盘/inode/目录项/日志保护;本地约1.54 GiB可用、共享约34240空闲inode,写入探针通过,不将df当账号配额。清理4.856 GiB:HY v12诊断与Kimodo v12训练的重复索引、Kimodo v12不合格125快照,以及无后续用途的Kimodo v10最终权重。核对结束收据、进程已退出、无打开句柄、无配置输入引用,缓存与新副本哈希一致。保留当前完整断点、必要对照/正式权重及全部轨迹/配置/指标。清单artifacts/recovery_cleanup_hourly_20261109T21.json,完整分析recovery_hourly_20261109T21_analysis.json。 ++ ++下一轮先检验原文本组复现,再读取两臂完整逐例轨迹判断文本敏感性,结合行走进度/幅度/接段/jerk而非单一平均误差确定后续训练。未达到正式条件,不提前取消心跳;任一正式训练健康启动后立即取消的要求保持。 ++ ++ ++### 小时检查与引导强度诊断 v14(2026-11-09 22 UTC) ++ ++v13两模型文本对照正常完成,零更新,无错误/资源停止,预留释放。已读取全部两臂完整逐例轨迹并确认有限;original组与既有对应权重结果逐元素完全一致。原文本对照可复现,空文本效果不是入口精度差异导致。无正式训练,心跳416e9f54保持。 ++ ++HY原文本幅度0.291、姿态0.372 rad、根位置0.449 m、jerk335.19;空文本为0.269、0.387 rad、0.530 m、334.10。行走2327从0.483/0.680降至0.173/0.056;961仍近零。Kimodo原文本幅度1.100、姿态0.439 rad、根位置0.441 m、jerk405.63;空文本为1.176、0.439 rad、0.526 m、56.93。空文本行走2327仅0.067/0.035,961为-0.022/-0.030,原文本分别0.827/0.678和0.107/0.510。空文本低jerk来自缺少正确行走,不能作为改善或替代输入。两臂都未通过门槛;不能据此断言历史文本冲突是唯一根因,也不能宣称模型忽略文本。未测新FID。 ++ ++下一有界诊断v14只改变引导强度,目录`/root/autodl-fs/pangxinlei/models/hfd-recovery-guidance-v14-20261110/{hymotion,kimodo}`,入口scripts/recovery_guidance_v14/supervise_pilot.py。保持v13相同固定权重、完整因果文本、同一6个validation案例/两个种子、HY10步BF16与Kimodo20步、历史/坐标/执行长度及原门槛。HY比较guidance1→2,检验增强条件响应;Kimodo比较2→1,检验文本引导是否放大根运动抖动。假设未证实,不以分数改善预设正确;仍须检查逐例进度与幅度。无未来文本/GT选指令,Test1024未使用。 ++ ++两臂original/adjusted_guidance写入logs/diagnostic.jsonl,guidance字段记实际数值;validation-traces/step-000000和000001分别对应两臂,非训练步。零优化更新、不保存训练断点,4小时监督上限与资源受控停止保留。输入权重SHA与步数在artifacts/recovery_v14_inputs.json;源码SHA与进程recovery_v14_launch.json;静态检查recovery_v14_static_check.json。监督器565596/565599及子进程已核实存活,尚无本轮结果。 ++ ++各2 GiB预算重新实写/fsync并验证物理分配,本地约1.54 GiB空闲、共享约34170空闲inode,探针写入通过;不将df当账号配额。清理已完成v13的重复索引656.24 MiB,确认旧进程结束、无打开句柄、与v14副本SHA一致,保留全部轨迹/指标/源码/配置及正式与必要候选权重。清单artifacts/recovery_cleanup_hourly_20261109T22.json;完整分析recovery_hourly_20261109T22_analysis.json。下次先检验original组复现,再分析调整引导强度的逐例效果,为后续恢复训练选有证据的设置;没有通过全部门槛不正式开训。正式训练健康启动后立即取消小时任务的条件保持。 ++ ++ ++### 第十五轮因果文本适配恢复(2026-11-09 23 UTC) ++ ++v14两模型引导诊断正常完成,无错误/资源停止,预留释放;全部两臂完整轨迹已读取且有限,original组逐元素复现既有结果。HY guidance2幅度0.341、接段1.530、姿态0.379 rad、根位置0.575 m、jerk529.91;行走2327进度2.108/2.568过冲,961仍为-0.024/-0.008。Kimodo guidance1幅度1.095、接段2.412、姿态0.429 rad、根位置0.475 m、jerk143.44,但2327进度0.120/0.366、961为0.075/0.206。两者均未通过,低jerk不能代替任务运动。不采用这些推理设置,无正式训练。 ++ ++恢复实验v15目录`/root/autodl-fs/pangxinlei/models/hfd-recovery-pilot-v15-20261110/{hymotion,kimodo}`,入口scripts/recovery_training_v15/supervise_pilot.py。HY以v7、Kimodo以v12为对照,均从原正式best重新初始化,不从诊断输出续训;保持各自损失/学习率/预测历史概率/250更新/0、125、250两种子验证/原推理设置,仅加入训练和推理共享的因果文本适配。原baseline-validation.json逐字节保持,原误差门槛不因输入变换放宽。 ++ ++适配器causal_text.py只读取截至当前发布时刻的文本前缀和发布时间。新文本为旧文本词前缀的延续时,距上次新增词不足1.2秒则继续当前片段,否则从新增加的词开始;不变文本保持当前片段,空上下文清空,修订或截断无法可靠分辨时保守保留当前已发布全文。无动作、turn ID、控制标签、GT边界或未来词。是新的因果输入预处理实验,不能把结果标为“原完整转录不变”的协议,也不能据其直接晋升正式训练。 ++ ++单元验证前缀不变性、追加未来不影响过去、空上下文过期、修订/截断保守处理、非法时间拒绝及确定性恢复。使用禁止未来索引的序列测试真实Source.condition绑定,确认只访问已发布前缀且发布时间不变。真实输入核验先发现按需provider用SimpleNamespace构造Source、绕过__init__,已在启动前改为共享Source.condition接口;未启动无效训练。修正后961 decision276/432从“turns left Hmm... walks forward”变为“walks forward”,120为“Hmm...”;2327仍可能保留“stop jog forward left diagonally”,此规则不保证解决全部歧义,不预设改善。收据artifacts/recovery_v15_causal_text_checks.json、recovery_v15_actual_text_audit.json、recovery_v15_config_diff.json。只用train/validation,Test1024留出。 ++ ++HY2 GiB、Kimodo12 GiB预算重新实写/fsync并验证物理分配;原子checkpoint上限256 MiB/4 GiB、8小时墙钟、磁盘/inode/目录项/日志和受控停止保护保持。本地约1.54 GiB空闲、共享约34130空闲inode,探针写入通过,不把df当配额。两监督器567139/567143与子进程已核实启动,先初始验证和完整两更新保存/严格恢复;当前恢复检查未完成、效果未知。源码SHA和启动收据artifacts/recovery_v15_launch.json。 ++ ++清理已完成v14重复索引656.24 MiB,确认旧进程退出、无打开句柄和新副本SHA一致,正式/必要候选/完整恢复断点、全部轨迹和配置指标保留。清单artifacts/recovery_cleanup_hourly_20261109T23.json;完整分析recovery_hourly_20261109T23_analysis.json。下一检查点为两模型恢复收据与实际训练文本、0/125/250闭环逐例改善,并同时关注旧指令被过早丢弃的失败案例。无正式训练,心跳416e9f54继续;任一正式训练健康启动后立即删除心跳的条件保持。 ++ ++ ++### 小时检查与固定权重文本复核 v16(2026-11-10 00 UTC) ++ ++HY v15仍运行,00:05核验194/250步,进程567475正常推进,已记录loss/梯度有限、完整两更新恢复通过,无错误或资源停止。125步幅度0.258、接段2.068、姿态0.447 rad、根位置0.512 m,行走2327进度0.111/0.032、961为-0.052/-0.015,仍不合格;等待有界250步最终验证,不重复启动。 ++ ++Kimodo v15已完成250步、完整恢复通过、无错误,预留已释放。读取初始/125/250全部完整轨迹并确认有限。最终幅度1.193、接段2.259、姿态0.465 rad、根位置0.685 m、jerk501.95;2327进度0.679/0.537,961为-0.938/-1.307,明显反向行走。相比v12最终根位置0.441 m、961进度0.107/0.510退化。v15第0步961已为-1.171/-1.180(根位置均值0.739 m),125步进一步到-2.071/-2.205(根位置0.996 m);因此退化并非全由训练更新造成,截短上下文可能丢失方向信息,但尚不能确认唯一根因。不延长此配方、不进入正式训练;这些结果不是FID。 ++ ++Kimodo v16仅固定v15最终权重,比较adapted与full_context两臂,原验证6案例/两种子、20步CFG2、全部其他输入/推理/门槛保持。目录`/root/autodl-fs/pangxinlei/models/hfd-recovery-text-v16-20261110/kimodo`,入口scripts/recovery_text_v16/supervise_pilot.py。两臂依次写logs/diagnostic.jsonl,轨迹step-000000/000001是诊断臂而非训练步;0优化更新,进入优化器构造前返回。先复现v15最终适配文本,再恢复完整已发布因果文本,区分输入变换和训练后权重效应;不使用未来词或GT选指令,Test1024未使用。通过共享模块开关切换已绑定Source.condition,验证切换/还原与发布时间保持;避免运行入口与模型入口各自加载配置导致开关无效。源码/静态核验artifacts/recovery_v16_static_check.json,启动收据recovery_v16_launch.json。监督器568479/子进程568481已确认存活,诊断结果尚未完成。 ++ ++v16重新实写/fsync预留2GiB,4小时监督上限,磁盘/inode/日志/目录项保护保留。本地约1.539GiB可用,共享34052空闲inode,本地与共享1MiB写入/fsync探针通过;不把共享df总量当账号配额。清理已结束Kimodo v15重复索引和失败125快照,共1407.06MiB:确认训练/监督器已退出、无打开句柄、缓存与v16副本SHA一致、中间权重无配置输入引用。保留v15最终完整断点、全部配置/指标/轨迹及必要历史/正式权重,不触碰仍运行HY资源。清单artifacts/recovery_cleanup_hourly_20261110T00.json;完整轨迹/健康分析recovery_hourly_20261110T00_analysis.json。 ++ ++下一检查点为HY v15最终结果与Kimodo v16原适配组复现、完整文本恢复后的逐例方向/进度/幅度/接段/jerk。无正式训练,小时任务416e9f54继续;任一正式训练健康启动后立即删除心跳,不等待正式训练结束。 ++ ++ ++### 第十七轮检查与恢复实验(2026-11-10 01 UTC) ++ ++HY v15完成250更新,完整两更新恢复通过、无错误/资源停止,预留释放。初始/125/最终全部完整轨迹已读取且有限,记录loss/梯度有限。最终幅度0.247、接段2.003、姿态0.452 rad、根位置0.524 m、jerk279.65;行走2327进度0.073/-0.018、961为-0.038/0.005。相对原正式初始幅度0.291、姿态0.372、根位置0.449退化,低jerk不代表恢复。原门槛不变,不正式开训。 ++ ++Kimodo v16两臂正常完成,零训练更新、预留释放。已读取全部轨迹,适配文本组预测/参考逐元素复现v15最终。完整文本组幅度1.212、接段2.311、姿态0.454 rad、根位置0.523 m、jerk473.87;2327进度1.052/0.669,961为-0.227/0.253。相比适配组根位置0.685和961 -0.938/-1.307改善,但仍明显未通过原门槛,且不如v12根位置0.441/961 0.107/0.510。说明截短因果上下文确有影响,恢复全文不足以解决全部问题;不将该诊断称为FID。 ++ ++HY v17固定v15最终权重进行adapted/full_context两臂验证,目录`/root/autodl-fs/pangxinlei/models/hfd-recovery-text-v17-20261110/hymotion`,入口scripts/recovery_text_v17/supervise_pilot.py。同一6案例/两种子、10步BF16、CFG1与原门槛保持;沿用v16已验证共享模块开关切换Source.condition,原发布时间不变,无未来词或GT选指令。0更新、优化器构造前返回、4小时上限;先检验适配组复现再分析全文效果,不默认可恢复。 ++ ++Kimodo v17回到v12完整文本配方及原正式best初始化,唯一算法变量stage1_root_displacement权重0.5→1.0,检验加强相对最后观测帧的未来根位移监督能否减少v12的行走进度不足。不能据单一失败案例宣称必然有效;监督只用训练目标,推理不输入未来GT。其余源码与v12逐字节一致、训练/推理设置和原baseline-validation.json保持。目录`/root/autodl-fs/pangxinlei/models/hfd-recovery-pilot-v17-20261110/kimodo`,入口scripts/recovery_training_v17/supervise_pilot.py。250更新、0/125/250固定两种子验证、先两更新完整保存/严格恢复;不延长或续训v15失败配方,不使用Test1024。当前仅启动及数据准备,恢复检查尚未完成。 ++ ++两监督器569655/569696及真实子进程均已核实存活,未发现错误;源码/配置核验artifacts/recovery_v17_checks.json,启动及源码SHA见recovery_v17_hymotion_launch.json、recovery_v17_kimodo_launch.json。HY2GiB/Kimodo12GiB实写fsync预留通过,Kimodo保留4GiB原子checkpoint上限、8小时墙钟及磁盘/inode/目录项/日志保护。本地约1.539GiB、共享约34000空闲inode,本地和共享1MiB写入/fsync探针通过;共享df不是账号配额。 ++ ++清理已结束HY v15及Kimodo v16重复索引、HY v15已分析失败125快照,共786.78MiB;核验旧训练/监督器退出、无打开句柄、重复索引与新副本SHA一致及权重无配置引用。保留最终完整断点、正式权重、必要对照和所有轨迹/指标/配置。清单artifacts/recovery_cleanup_hourly_20261110T01.json;完整审计recovery_hourly_20261110T01_analysis.json。下一检查点为HY两臂结果与Kimodo恢复收据/训练推进及逐例效果。无正式训练,心跳416e9f54继续;任一正式训练健康启动后立即取消的条件保持。 ++ ++ ++### 第十八轮扩大验证(2026-11-10 02 UTC) ++ ++HY v17固定权重文本诊断正常完成,无错误,预留释放;全部轨迹有限,adapted组逐元素复现v15最终。full_context幅度0.272、接段2.003、姿态0.444 rad、根位置0.459 m、jerk290.00,行走2327为0.454/0.449、961为-0.018/-0.007;无坍缩但幅度仍不足,姿态超过原初始10%上限。截短组仍有1197/2026坍缩。恢复全文有改善却未解决恢复问题,不再重复截短配方;本轮先集中验证Kimodo的新增积极证据。 ++ ++Kimodo v17正常完成250更新,完整2步恢复通过,所有记录loss/梯度有限、全部完整轨迹已读取有限,无资源停止,预留释放。125步幅度1.182、接段2.229、姿态0.463 rad、根位置0.386 m;2327进度0.924/1.035、961为1.464/0.541,无坍缩,首次此轮通过原小验证全部门槛。250步幅度1.187、接段2.317、姿态0.456、根位置0.369,但961第二种子进度0.325,最终不合格;不机械延长,不丢掉中途候选。125步jerk787.14、250步562.89,相比原正式初始357.30偏高,必须排除抖动收益,不能以门槛通过直接正式开训。原正式基线姿态0.468/根位置0.442,各阶段对照和逐例证据见artifacts/recovery_hourly_20261110T02_analysis.json;这些不是FID。 ++ ++v18是冻结权重的扩大validation,零更新。目录`/root/autodl-fs/pangxinlei/models/hfd-recovery-broad-v18-20261110/{baseline,candidate125}`,入口scripts/recovery_broad_v18/supervise_pilot.py。baseline用原正式best,candidate125用v17 weights/step-000125.pt。每交互模式确定性选择4个6–20秒validation案例,共24个、包含原6个、固定种子12345/2026;仅依据模式/长度/可用性选样,不能依据预测分数选样,无Test1024。两组均保持完整因果文本、20步CFG2、同一初始状态及其他推理协议。选择器验证确定性、各模式4例、原6例保留;源码/检查recovery_v18_checks.json,源码SHA/启动recovery_v18_launch.json。两组监督器570911/570912、子进程570913/570914已核实存活,当前数据准备,结果未知。 ++ ++每组logs/diagnostic.jsonl及validation-traces/step-000000.npz为单个冻结权重验证,不是训练步。日志内门槛暂以原六例baseline计算,不能据此自动晋升;最终必须用同一扩展案例原正式baseline重新计算非退化门槛,检查逐例进度、幅度、接段/段内突变、速度误差、根jerk及失败轨迹。即使小验证125合格也不绕过更广验证和正式容量/恢复配置要求。 ++ ++两组各2GiB预算实写/fsync通过,4小时监督上限、日志轮转/目录项/资源停止保持,扩大轨迹单文件上限128MiB以覆盖48条轨迹且纳入预算。本地约1.539GiB、共享33893空闲inode,本地和共享1MiB写入/fsync通过,共享df不视为账号配额。清理已结束Kimodo v17重复索引328.12MiB,核验旧训练/监督器退出、无打开句柄、与新副本SHA一致。保留125候选、best、最终完整恢复断点及全部证据;HY诊断缓存暂留待后续。清单recovery_cleanup_hourly_20261110T02.json。无正式训练,心跳416e9f54保持,下一检查03 UTC;任一正式训练健康启动后立即取消,不等待训练结束。 ++ ++ ++### 扩大验证故障修复与候选重试(2026-11-10 03 UTC) ++ ++v18 baseline已完成24案例×2种子,零训练更新;读取48条完整轨迹均有限。幅度0.446、接段1.995、姿态0.453 rad、根位置0.632 m、jerk587.06;24个GT明显位移的案例×种子组合中17个进度未达0.5–2。扩大集不能直接与原六例均值比较,候选必须和这份同集基线比较。 ++ ++candidate125没有完成,因CUDA OOM失败,不能将缺失结果解释成模型效果差。根因是v18并行主进程使用不同pilot.physical_gpu,但文本编码器FrozenTextWorker通过config.gpu独立指定物理GPU,两份配置均沿用gpu=1,导致两个约14.75GiB文本工作进程挤占GPU1。已保留phase-error/supervisor-error/日志和原源码配置;两个旧主进程与监督器均已退出,GPU任务检查为空,预留已释放。baseline已经正常完成,保留不重复计算。 ++ ++仅重跑候选,目录`/root/autodl-fs/pangxinlei/models/hfd-recovery-broad-v18-20261110/candidate125-retry`,入口scripts/recovery_broad_v18r/supervise_pilot.py。新冻结源码在run_pilot启动前强制config.gpu==pilot.physical_gpu,明确主进程和文本工作进程都用GPU1;本轮只运行这一组。权重仍v17 step-000125.pt、24案例选择规则/两个种子/全文/20步CFG2及全部模型推理设置不变,未改训练或放宽门槛,无Test1024。原运行与源码不覆盖。监督器571978、子进程571982已存活并正常准备数据;当前尚未完成文本加载和全部验证,不虚称效果或恢复成功。 ++ ++新目录2GiB容量重新实写/fsync通过,4小时监督/磁盘/inode/日志轮转/目录项和轨迹128MiB上限保持;保留失败证据及原正式/候选/完整恢复断点。修复及预留收据artifacts/recovery_v18_retry.json,启动与源码SHA recovery_v18_retry_launch.json。清理两旧组重复索引656.24MiB,确认旧监督器/主进程退出、无打开句柄、与重试副本SHA一致;错误日志和完成基线全部轨迹保留。清单recovery_cleanup_hourly_20261110T03.json。本地约1.539GiB、共享33869空闲inode,本地和共享1MiB写入/fsync通过;不以共享df推断账号配额。 ++ ++HY无新训练/验证结果,保留v17失败诊断,不重复失败配方。完整本轮分析recovery_hourly_20261110T03_analysis.json。下一检查先核验候选重试完成和实际GPU分配,再用同集baseline重算非退化门槛及逐例进度/抖动;原六例门槛标记不能替代扩大集判断。无正式训练,小时心跳416e9f54继续,下一次04 UTC;任一正式训练健康启动后立即取消的条件不变。 ++ ++ ++### 第十九轮恢复实验(2026-11-10 04 UTC) ++ ++v18候选单独重试正常完成,零更新,无错误/资源停止、预留释放。原正式与候选24案例×2种子的完整轨迹均有限,case_keys及全部参考轨迹完全匹配,已按同集重新计算门槛。候选幅度1.021(基线0.446)、接段2.209(1.995)、姿态0.429 rad(0.453)、根位置0.598 m(0.632),无坍缩(基线1例)。明显位移24个案例×种子中,失败从17降至10;两个种子各5个失败,不能用平均改善掩盖失败。根jerk956.56 vs587.06,增约63%;1863两种子进度0.917/1.100虽合格,jerk7422/6808,相比基线4476/3907明显恶化。幅度/接段/姿态位置非退化门槛通过,逐例行走门槛不通过,且抖动增加,不批准正式训练。指标为SMPL-X诊断,不是FID。 ++ ++下一有界实验v19以v17为对照,原正式best初始化,不从125候选或最终断点延长;唯一训练变量decoded_root_acceleration监督权重0.001→0.01,保持根位移1.0、全部其他损失/数据/学习率/因果全文/推理参数不变。已有该损失约束解码实际根位置的二阶差分与相同C6目标,并非独立辅助速度通道;源码不改、加强权重检验抖动是否可控,不预设有效,须防止以静止换低jerk。仍250更新、0/125/250固定原六例两种子验证及原正式初始门槛,通过后还必须复查扩展集。无Test1024。HY无新证据,本轮不重复失败配方,保留诊断。 ++ ++目录`/root/autodl-fs/pangxinlei/models/hfd-recovery-pilot-v19-20261110/kimodo`,入口scripts/recovery_training_v19/supervise_pilot.py。增加启动前config.gpu==pilot.physical_gpu校验,两者均1,GPU空闲检查后单独启动,避免v18文本工作进程冲突。训练源码除运行器GPU断言/准备入口外与v17逐字节一致,原baseline-validation.json不变。收据artifacts/recovery_v19_checks.json,源码SHA/启动recovery_v19_launch.json。监督器572624及真实子进程已确认存活,无错误;当前初始验证/两步保存与严格恢复阶段,恢复检查尚未完成,不宣称训练改善。 ++ ++12GiB预算重新实写/fsync通过(实际预留约12.265GB),保留4GiB原子checkpoint上限、8小时墙钟、日志轮转、目录项/磁盘/inode保护。目录项32,本地约1.538GiB,共享33844空闲inode,本地/共享1MiB写入fsync探针通过;共享df不作账号配额。清理已完成候选重试的重复索引328.12MiB,核验旧监督器/子进程退出、无打开句柄、与v19新副本SHA一致;125候选、必要历史对照、正式权重、最终完整断点和全部轨迹/配置/指标保留。清单recovery_cleanup_hourly_20261110T04.json;配对验收与失败明细recovery_hourly_20261110T04_analysis.json。 ++ ++下一检查为v19恢复收据、训练推进、0/125/250逐例进度/幅度及jerk,不因平均数改善放宽门槛。尚无正式训练,心跳416e9f54继续,下次05 UTC;任一正式训练健康启动后立即取消的条件保持。 ++ ++ ++### 第二十轮候选扩大验证(2026-11-10 05 UTC) ++ ++Kimodo v19完成250更新,完整两步保存/optimizer/scheduler恢复通过,已记录loss和嵌套梯度均有限;全部0/125/250完整轨迹读取且有限,无错误/资源停止,预留释放。125步幅度1.121、接段2.187、姿态0.449 rad、根位置0.308 m、jerk795.43;行走2327进度1.044/0.910、961为1.329/0.904,无坍缩,原小验证全部门槛通过。对照v17的125步根位置0.386 m有所改善,但jerk787.14→795.43,并未证明加强加速度监督减少抖动。250步幅度0.974、接段2.436、姿态0.416、根位置0.407、jerk530.99,961进度0.338/0.349不合格,2327为0.817/0.635;无坍缩。继续保留中途候选,不机械延长最终失败配方,不以低jerk代替正确行走。上述诊断不是FID。 ++ ++v20冻结v19 weights/step-000125.pt,扩大到与v18完全相同的24个validation案例/两个固定种子,复用已完成原正式基线及v17候选对照,不重跑基线。目录`/root/autodl-fs/pangxinlei/models/hfd-recovery-broad-v20-20261110/candidate125`,入口scripts/recovery_broad_v20/supervise_pilot.py。除准备入口外所有源码与修复GPU绑定的v18r逐字节相同,选择规则、原六例包含、完整因果文本、20步CFG2及推理不变;仅输入权重换为v19的125步。0训练更新,无Test1024。输出logs/diagnostic.jsonl与validation-traces/step-000000.npz;日志原六例门槛仅为临时标记,必须基于同一扩大集baseline重新计算非退化、逐例进度和抖动后判断,不能自动晋升。 ++ ++监督器573160及真实子进程已核验存活、未发现错误,目前数据准备。启动/源码SHA artifacts/recovery_v20_launch.json,源码一致性recovery_v20_checks.json。2GiB容量重新实写/fsync通过,单独运行GPU1,config.gpu与pilot.physical_gpu一致断言保持;4小时墙钟、128MiB轨迹上限、目录项/日志轮转/磁盘inode保护保持。完整本轮分析recovery_hourly_20261110T05_analysis.json。 ++ ++清理已结束v19训练的重复索引328.12MiB,核验旧监督器/子进程退出、无打开句柄且与v20副本SHA一致。保留125候选、best、最终完整恢复断点、正式和必要历史对照及全部原始轨迹/指标/配置。清单recovery_cleanup_hourly_20261110T05.json。本地约1.538GiB可用、共享约33780空闲inode,本地和共享1MiB写入fsync通过;共享df不视为账号配额。HY暂无新结果,不重复失败实验。 ++ ++下一检查将v20候选与v18基线/上一候选逐例匹配,核实范围更广的行走改善及jerk而非只看均值。尚未正式训练,心跳416e9f54继续,下次06 UTC;任一正式训练健康启动后立即取消的条件保持。 ++ ++ ++### 第二十一轮接段与执行节奏诊断(2026-11-10 06 UTC) ++ ++v20候选扩大验证正常完成,零训练更新,无错误/资源停止,预留释放,进程退出。读取原正式基线、v17候选、v19候选三组全部完整轨迹,数值有限、case_keys及参考逐元素匹配。v19候选幅度0.991、接段2.213、姿态0.416 rad、根位置0.528 m、jerk906.78;上一候选对应1.021/2.209/0.429/0.598/956.56,原正式基线0.446/1.995/0.453/0.632/587.06。明显位移失败17→10→8(总24案例×种子),最新无坍缩、其他均值门槛通过,但8个行走失败与较基线偏高jerk仍不满足正式条件,不能仅因持续改善就晋升。不是正式FID。 ++ ++进一步读取逐帧三阶差分,将跨过每12帧重规划首个新预测帧的差分窗口计为接段附近:约24.72%样本占基线88.18%、v17候选86.32%、v19候选86.48%的根jerk平方能量。表明高频接段附近值得定位,不等于已证明唯一根因;仅继续加大损失权重缺乏充分依据。 ++ ++v21固定v19 step-000125权重,同一24validation案例、两种子,执行12帧与24帧两臂对照。目录`/root/autodl-fs/pangxinlei/models/hfd-recovery-cadence-v21-20261110/kimodo`,入口scripts/recovery_cadence_v21/supervise_pilot.py。原12帧组先跑,要求复现v20完整预测;24帧组只改重规划/执行节奏,每次使用截至该决策时刻已发布文本和模型历史,不输入未来GT、不改变权重/采样步数/CFG。通过共享runtime CONFIG切换,避免配置副本导致开关无效;逐帧覆盖/无重复遗漏静态验证通过。0优化更新、无Test1024。 ++ ++24帧组改变响应延迟、文本更新时机与随机抽样次数,是定位频繁重规划接缝的诊断,不能当作原协议训练改善或正式晋升依据。logs/diagnostic.jsonl arms为execute_12/execute_24,完整轨迹step-000000/000001为两诊断臂,非训练步。旧12帧边界指标与资格标记对24帧组不能直接使用,下一轮需按各自执行节奏重算接段指标,并同时查看行走/停步延迟,避免平滑却不响应任务。 ++ ++监督器573629及真实子进程已存活、无错误,当前数据准备。配置/覆盖检查artifacts/recovery_v21_checks.json,源码SHA/启动recovery_v21_launch.json。2GiB预算重新实写fsync验证,4小时墙钟、单独GPU1、主/文本GPU一致断言、日志轮转/目录项/磁盘inode与128MiB轨迹上限保持。本地约1.538GiB、共享约33750空闲inode,本地/共享1MiB写入fsync通过,共享df不视为账号配额。 ++ ++清理已结束v20重复索引328.12MiB,核对旧进程/监督器退出、无打开句柄及与新副本SHA一致;正式模型、完整恢复断点、必要候选及全部证据保留。清单recovery_cleanup_hourly_20261110T06.json;三组配对验收和接段能量分析recovery_hourly_20261110T06_analysis.json。HY暂无新结果,不重复失败训练。无正式训练,心跳416e9f54保持,下次07 UTC;正式训练健康启动后立即取消的规则不变。 ++ ++ ++### 第二十二轮根轨迹组件诊断(2026-11-10 07 UTC) ++ ++v21两臂正常完成,零训练更新,无错误/资源停止,预留释放、旧进程退出。全部完整轨迹已读取且有限,12帧组与v20预测/参考逐元素一致。24帧组jerk1431.62,比12帧906.78更差;根位置0.593 vs0.528 m,幅度0.847 vs0.991,行走失败均8/24,不能据旧12帧接段均值下降宣称改善,也不采用24帧协议。 ++ ++按真实执行节奏重算:以新计划首帧为接段、SO3逐关节角增量求均值后取接段/段内比,12帧均值2.329、24帧2.924(此定义与已有指标实现细节不同,单列诊断不覆盖旧指标);根jerk跨接段窗口占样本24.72%/12.31%,对应能量86.48%/88.05%。减少重规划次数并未消除接缝,可能每次接缝更大;响应延迟同时改变,尚未单独量化停止延迟,不能以此代替任务协议。不是FID。 ++ ++v22固定相同v19 step-000125权重,恢复原12帧执行及24validation案例/两种子,仅增加只读根轨迹组件采集。目录`/root/autodl-fs/pangxinlei/models/hfd-recovery-components-v22-20261110/kimodo`,入口scripts/recovery_components_v22/supervise_pilot.py。采集每次最终inverse的unnormalized原生channels0:8(含stage1根0:3、骨盆偏移5:8)与decoded root xyz,原生帧59:73;首行最后已知历史,随后12行执行,末行仅诊断的未执行预测。保存为同一validation-traces/step-000000.npz中的root_components_i,形状[decisions,14,11],不新增逐步文件。钩子返回原解码对象,不修改输入/输出、不新增随机采样,并在finally恢复inverse;完整轨迹必须先逐元素复现v20,之后组件证据才能用于判断根/骨盆抵消或突跳。组件是模型预测,未向条件加入未来GT,无Test1024。 ++ ++该实验零优化更新、原20步CFG2/完整因果文本/坐标及其他推理保持。通过语法、钩子还原、12帧保持检查,收据artifacts/recovery_v22_checks.json;源码SHA/启动recovery_v22_launch.json。监督器574134及真实子进程已核验存活、无错误,当前数据准备,尚无组件诊断结论。4小时监督上限、2GiB实写fsync预留、GPU1主/文本一致断言、目录项/磁盘inode/日志轮转与128MiB轨迹上限保留。 ++ ++清理已结束v21重复索引328.12MiB,核验旧进程/监督器退出、无打开句柄,与新副本SHA一致;全部诊断轨迹/配置/指标、正式/必要候选及完整恢复权重保留。清单recovery_cleanup_hourly_20261110T07.json,完整分析recovery_hourly_20261110T07_analysis.json。本地约1.537GiB、共享33711空闲inode,本地和共享1MiB写入fsync通过,共享df不作为账号配额。HY无新证据,不重复失败训练。 ++ ++下一检查先确认v22输出复现,再比较失败/正常案例的stage1根、骨盆偏移和解码根在接段前后变化,以证据决定训练或表征修复。无正式训练,心跳416e9f54继续,下次08 UTC;正式训练健康启动后立即取消规则保持。 ++ ++ ++### 用户要求双模型同步推进;HY v23启动(2026-11-10) ++ ++用户明确要求:两个模型必须一起推进,不能因为Kimodo出现改善就搁置HY;只有某模型确认改进、通过验证并正式训练成功启动后,才可聚焦另一个。该要求已写入artifacts/recovery_hourly_task.json的parallel_progress_requirement,后续每轮必须分别检查并推进两者。此前任一正式训练健康启动后立即取消小时任务的条件仍保留,不把pilot当正式训练。 ++ ++HY新一轮v23已在空闲GPU0启动,目录`/root/autodl-fs/pangxinlei/models/hfd-recovery-pilot-v23-20261110/hymotion`,入口scripts/recovery_training_v23/supervise_pilot.py。以完整因果文本的v7为对照,原正式best初始化、同一数据/种子/250更新/0、125、250验证/根运动监督与全部原门槛保持;唯一算法变化rotation_velocity_weight 0.05→0,检验额外旋转速度辅助目标是否限制动作幅度。只是待验证假设,不声称根因已确定;原flow损失和根导数目标保留,不靠降低验收标准。无Test1024,训练源码除准备入口与v7逐字节相同。检查recovery_v23_checks.json,启动与源码SHA recovery_v23_launch.json。 ++ ++HY监督器574500和真实子进程已确认存活,目前初始验证/两更新完整保存恢复阶段,尚不能宣称恢复检查通过或效果改善。Kimodo v22在GPU1的根组件诊断真实进程仍存活且无错误,继续推进,不中断源码/配置。两模型健康检查recovery_v23_health.json。 ++ ++HY2GiB容量重新实写/fsync预留通过,256MiB原子checkpoint上限、8小时墙钟、日志轮转/目录项/磁盘inode保护保持。本地与共享1MiB写入/fsync检查通过,本地约1.537GiB可用、共享约33660空闲inode。保留当前有效权重与证据,重复缓存继续按每轮已结束/无占用/哈希一致规则清理。下一小时检查须同时分析HY v23恢复与训练进度、Kimodo v22组件结果,再分别确定后续动作;不能再次默认只推进Kimodo。 ++ ++ ++### 双模型同步检查与Kimodo v24恢复(2026-11-10 08 UTC) ++ ++HY v23仍在GPU0正常训练,检查时151步,结束核验163/250步,真实进程存活、完整2步保存/optimizer/scheduler恢复通过、无错误/资源停止。已读取初始/125完整轨迹且有限,记录loss/梯度有限;125幅度0.270、接段2.156、姿态0.442 rad、根位置0.463 m,仍不合格。继续既定有界250步及最终验证,不重启、不覆盖正在运行源码配置。 ++ ++Kimodo v22组件诊断完成,全部完整预测/参考与v20逐元素完全相同,组件数值有限。高抖动1863两种子每次第一帧stage1根平均跳变0.230/0.207 m,骨盆偏移变化0.262/0.236 m,方向余弦约-0.987/-0.989;解码第一帧仅0.043/0.038 m,第二帧达0.186/0.160 m。810与277也出现类似抵消/后续突跳。这支持针对根轨迹第一、第二帧直接约束,尚不声称唯一根因。已查原生inverse:平滑根仅加到水平x/z,根高度来自骨盆位置;水平组合的最大残差2.38e-7,三维直接相加的残差不是解码bug。完整证据artifacts/recovery_hourly_20261110T08_analysis.json。 ++ ++Kimodo v24以v19为基线、原正式best初始化,唯一算法变量stage1_boundary_weight 3→10,加强已有stage1_root_velocity中前两次差分权重;其余根位移1.0、解码加速度0.01、训练数据/学习率/种子/完整因果文本/推理及原门槛不变。并非延长失败最终断点,仍250更新、0/125/250固定两种子验证、先2步完整保存严格恢复;不以静止或抑制幅度换低抖动,无Test1024。目录`/root/autodl-fs/pangxinlei/models/hfd-recovery-pilot-v24-20261110/kimodo`,入口scripts/recovery_training_v24/supervise_pilot.py。配置差异与基线核验recovery_v24_checks.json,源码SHA/启动recovery_v24_launch.json。 ++ ++Kimodo监督器575150及子进程已核验存活,GPU1启动前空闲,GPU0的HY持续推进。目前Kimodo初始验证/两步恢复阶段,恢复测试尚未完成、效果未知。12GiB预算重新实写fsync通过,保留4GiB原子checkpoint上限、8小时墙钟、主/文本GPU绑定校验、日志轮转、目录项/磁盘inode保护。HY当前预留不清理。本地约1.537GiB、共享33638空闲inode,本地/共享1MiB写入fsync通过;共享df不当账号配额。 ++ ++清理已完成Kimodo v22与HY v17诊断重复索引共656.24MiB,确认旧训练/监督器退出、无打开句柄,与新v24/v23保留副本SHA一致。保留正式权重、完整恢复断点、必要候选和全部原始证据。清单recovery_cleanup_hourly_20261110T08.json。下次同时分析HY最终结果和Kimodo恢复/推进及逐例结果;双模型必须持续推进的要求保持。无正式训练,心跳416e9f54继续,下次09 UTC;任一模型正式训练健康启动后立即取消规则不变。 ++ ++ ++### 第二十五轮双模型推进(2026-11-10 09 UTC) ++ ++HY v23与Kimodo v24均正常完成250更新、完整2步optimizer/scheduler恢复通过,无错误/资源停止、预留释放。已读取两模型初始/125/最终所有完整轨迹且有限,loss及嵌套梯度记录有限。HY最终幅度0.275、接段2.098、姿态0.451 rad、根位置0.474 m、jerk293.03;2327进度0.359/0.334,961为0.000/0.004,仍不合格。移除旋转速度辅助并未恢复幅度,不继续延长此配方。 ++ ++Kimodo v24的125步通过小验证门槛,jerk775.87,2327进度1.540/0.822、961为1.327/0.609;但250步961第二种子0.209失败(第一0.608),2327为0.793/0.553,最终幅度1.008、接段2.431、姿态0.421 rad、根位置0.390 m、jerk561.37。小验证通过不等于更广交互恢复,不直接正式训练,也不把这些诊断称为FID。 ++ ++HY v25目录`/root/autodl-fs/pangxinlei/models/hfd-recovery-pilot-v25-20261110/hymotion`,入口scripts/recovery_training_v25/supervise_pilot.py。在v23基础上唯一算法变量root_velocity_weight 0.1→0(旋转速度权重已为0),保留原flow损失、C6目标/导数一致构造、完整因果文本、原正式best初始化、数据/种子/学习率和原门槛,做flow-only对照检验辅助目标是否冲突,而非预设根因。250更新、0/125/250两种子验证、完整2步保存恢复检查;当前恢复检查尚未完成。GPU0监督器575697与真实子进程已存活,无错误。 ++ ++Kimodo v25目录`/root/autodl-fs/pangxinlei/models/hfd-recovery-broad-v25-20261110/candidate125`,入口scripts/recovery_broad_v25/supervise_pilot.py。冻结v24 step-000125.pt,重复同一24validation案例/两种子/12帧执行/20步CFG2及完整因果文本,复用v18原正式基线及v20候选对照。零优化更新,只有权重与输出目录变化;要求用匹配扩大集原基线重算门槛并检查抖动,不用日志六例临时标记晋升。GPU1监督器575698与真实子进程已存活,无错误。两任务都不使用Test1024;双模型同步推进要求保持。 ++ ++每组2GiB预算重新实写fsync通过;HY8小时/256MiB原子checkpoint上限,Kimodo4小时/128MiB轨迹上限,原磁盘/inode/目录项/日志轮转保护保持。主模型与文本GPU分配保持各自0/1,不共享Kimodo文本工作进程。配置差异artifacts/recovery_v25_checks.json,源码SHA/启动recovery_v25_launch.json。原始结果、恢复与实际写入检查recovery_hourly_20261110T09_analysis.json。 ++ ++清理两旧训练重复索引及HY已分析失败125快照,共786.78MiB;确认旧训练/监督器退出、无打开句柄、缓存SHA与保留副本一致、中间快照无配置输入引用。保留两模型最终完整断点、Kimodo125候选与best、正式/必要历史对照及全部轨迹/配置/指标。清单recovery_cleanup_hourly_20261110T09.json。本地约1.536GiB、共享约33560空闲inode,本地/共享1MiB写入fsync通过,不把共享df当配额。下次同时检查HY恢复/训练推进、Kimodo扩大验证结果。无正式训练,小时心跳416e9f54继续,下次10 UTC;任一正式训练健康启动后立即取消规则保持。 ++ ++ ++### 第二十六轮同步检查与历史诊断(2026-11-10 10 UTC) ++ ++HY v25持续GPU0训练,开始检查161步,结束169/250步;完整2步恢复通过,记录loss/梯度及初始/125完整轨迹有限,无错误/资源停止。125幅度0.272、接段2.299、姿态0.450 rad、根位置0.469 m,仍不合格。继续本轮有界预算,不重启,不碰正在使用的预留或断点。 ++ ++Kimodo v25扩大验证正常完成,无错误,预留释放。三组同24案例/两种子完整轨迹均有限、参考与case_keys匹配。v24的125步候选幅度0.966、接段2.234、姿态0.412 rad、根位置0.585 m、jerk971.45;明显位移失败11/24,相比v19候选8/24、根位置0.528、jerk906.78退化。加强stage1前两帧速度权重没有带来扩大集改善,不采用、不机械延长。不能根据小验证通过正式开训。 ++ ++Kimodo v26固定较好的v19 step-000125权重,比较predicted_history与expert_past两臂,目录`/root/autodl-fs/pangxinlei/models/hfd-recovery-history-v26-20261110/kimodo`,入口scripts/recovery_history_v26/supervise_pilot.py。保持原24validation案例/两种子、12帧执行、20步CFG2、因果全文与其余推理设置。第一臂必须逐元素复现v20;第二臂每次用截至当前decision帧的真实过去60帧替换生成历史,索引绝不超过decision、不读取未来GT条件。目标是区分反馈误差累积与单次条件预测不足,非训练配方或部署结果;真实过去在部署不可获得,重置也会制造拼接突变,不能把该组较低误差/jerk或资格标记当恢复成功。零更新、无Test1024,保留完整两臂轨迹。 ++ ++静态验证过去索引上界、共享runtime CONFIG切换和原组复现要求,收据artifacts/recovery_v26_checks.json;启动及源码SHA recovery_v26_launch.json。监督器576671及真实子进程已核验存活,无错误,GPU1启动前空闲;HY继续GPU0。2GiB实写fsync预留、4小时墙钟、主/文本GPU绑定校验、日志轮转/目录项/磁盘inode与128MiB轨迹上限保留。当前尚无诊断结论。 ++ ++清理已完成Kimodo扩大验证重复索引328.12MiB,核验旧训练/监督器退出、无打开句柄及与新副本SHA一致;正式/必要候选、完整断点和全部证据保留。清单recovery_cleanup_hourly_20261110T10.json;三组配对轨迹与HY健康分析recovery_hourly_20261110T10_analysis.json。本地约1.536GiB、共享33503空闲inode,本地/共享1MiB写入fsync通过,共享df不是账号配额。下次同时分析HY最终结果和Kimodo历史对照,不搁置任一模型。无正式训练,心跳416e9f54继续,下次11 UTC;正式训练健康启动后立即取消规则不变。 ++ ++### 第二十七轮双模型历史诊断(2026-11-10 11 UTC) ++ ++HY v25完成250步,完整恢复检查通过,loss/梯度记录和0/125/250全部轨迹有限;最终幅度0.283、接段2.224、姿态0.457 rad、根位置0.461 m,仍不合格。去掉旋转及根速度辅助损失没有恢复动作幅度,不机械延长。Kimodo v26两臂完成;生成历史组逐元素复现v20。真实过去组姿态0.176 rad、根位置0.093 m,但接段6.186、jerk5296.08,且每段重置使行走进度看似通过,不能作为恢复成功。按当前段真实起点计算,真实历史组段内根位移RMSE仍0.139 m;生成历史组按自己当前起点计算为0.078 m,二者锚点定义不同,不能据此单独下因果结论。完整证据见HFD_bench/artifacts/recovery_hourly_20261110T11_analysis.json;其中mean_root_jerk_rms是坐标分量RMS,和既有向量模jerk相差sqrt(3),不替换旧指标。 ++ ++两个模型同步启动v27,根目录/root/autodl-fs/pangxinlei/models/hfd-recovery-history-v27-20261110。HY固定v25最终权重,原6validation×2种子对照生成历史与截至decision的真实过去;首臂应复现v25最终轨迹。Kimodo固定v19 step125,在原24validation×2种子中分别只用真实历史根平移、只用真实历史旋转替换,分解条件误差来源。混合历史可能存在平移/姿态不一致,因此仅作诊断,不能视为可部署配方或正式资格。索引不超过decision,未来占位和原采样设置保持;零更新、无Test1024。入口scripts/recovery_history_v27_{hymotion,kimodo}/supervise_pilot.py,源码/启动及健康收据recovery_v27_launch_*.json、recovery_v27_health.json。下次检查必须核对HY原臂复现,并结合Kimodo两组的段内误差/接缝而非全局重置进度选择训练改动。 ++ ++GPU0/1分别确认真实子进程存活并占用显存,当前为加载及诊断阶段,不宣称结果已改善。每组2GiB预算实写fsync预留通过,4小时监督器、磁盘/inode/目录项/日志保护保留。本地约1.536GiB,磁盘本地/共享1MiB实写fsync通过,共享空闲inode超过33000;共享df不作为账号配额。确认旧进程已退出、无打开句柄及新副本SHA相同后,清理HYv25和Kimv26重复启动索引共656.24MiB;所有正式权重、完整断点和轨迹保留,清单recovery_cleanup_hourly_20261110T11.json。 ++ ++尚无正式训练,心跳416e9f54继续;下一次12 UTC。任一模型正式训练通过验收且真实健康启动后立即取消心跳,不关闭训练监督器。两模型继续同步推进。 ++ ++### 第二十八轮双模型恢复试验(2026-11-10 12 UTC) ++ ++v27两模型正常完成,无错误/资源停止,预留已释放;两臂所有完整轨迹有限。HY生成历史组逐元素复现v25最终结果,幅度0.283、接段2.224、姿态0.457 rad、根位置0.461 m、jerk334.62,2个幅度坍缩案例;4个明显位移组合均未严格达到0.5(其中一个0.499574,不能四舍五入算通过)。真实过去组幅度0.644、姿态0.146、根位置0.034,但接段12.068、jerk1878.39;重置产生的行走改善不计资格。Kimodo只替换历史平移:幅度1.235、接段2.397、姿态0.394、根位置0.100、jerk5448.41;只替换历史旋转:幅度1.217、接段6.417、姿态0.183、根位置0.383、jerk1256.41,行走仍失败7/24。真实历史及混合条件均非部署方法,不能用其较低误差正式开训。完整逐案例/轨迹核验在artifacts/recovery_hourly_20261110T12_analysis.json,不是FID。 ++ ++两边同步启动v28各250步有界恢复试验,根目录/root/autodl-fs/pangxinlei/models/hfd-recovery-pilot-v28-20261110,入口scripts/recovery_training_v28/supervise_pilot.py。HY以v25配方为对照,仅将生成历史训练比例0.25提高到1,检验真实历史和自主历史的分布差异;Kimodo以v19为对照,仅将监督标签过渡6帧延长至12帧,检验更缓的姿态/高度纠正能否减小接段冲击。后者不是直接改变XY轨迹或推理后处理,也不能预设有效。均从原正式best初始化,保留0/125/250同6案例×2种子及原始验收门槛;未达幅度/行走条件的平滑不算改善,若小验证通过仍须扩大验证。Test1024未使用。 ++ ++配置差异记录recovery_v28_config_diff.json;12帧过渡测试验证第12帧起与对齐标签完全一致、历史不被改写、数值有限,推理条件/部署实现不改。HY与Kimodo关键运行源码分别匹配v25/v19;启动先做2步optimizer/scheduler断点回读与恢复,随后监督器继续至250步,当前恢复检查尚未完成,不宣称已通过。两个真实子进程已核验,GPU分别0/1,不存在正式训练。HY2GiB/Kimodo12GiB实写fsync预留均通过;8小时墙钟、日志轮转、目录项/容量/inode检查、原子断点失败保留旧文件机制沿用。启动/源码SHA和健康证据recovery_v28_launch_*.json、recovery_v28_health.json。 ++ ++本轮清理已结束v27两模型重复启动索引656.24MiB,删除前确认旧进程/监督器退出、无打开句柄并与新副本SHA一致;保留全部权重、断点和原始轨迹,清单recovery_cleanup_hourly_20261110T12.json。本地约1.536GiB,本地/共享1MiB实写fsync通过,共享inode超过33000,共享df不作账号配额。小时任务416e9f54保持,下一次13 UTC重点核对两边恢复收据、loss/梯度和初始/中途轨迹;任一正式训练健康启动立即取消,不终止监督器。 ++ ++### 双模型检查及Kimodo扩大验证(2026-11-10 13 UTC) ++ ++HY v28仍正常训练,检查时153步、结束记录158/250步;完整2步optimizer/scheduler断点回读与恢复通过,无错误或资源停止。125步幅度0.480(未达0.5)、接段1.695、姿态0.418 rad(仍超过原初始1.1倍)、根位置0.430 m、jerk351.95。2327进度0.965/1.173改善,961仍-0.007/0.001,尚不合格;继续当前有界训练,不重启、不动其预留。full_predicted_history_fraction表示当前窗口已有60帧自身历史的比例,训练初始段不足60帧,因此不应把该日志低于1误读为配置没有启用100%预测链。 ++ ++Kimodo v28正常完成250步,恢复收据通过,预留已释放。125步通过原6案例两种子小验证:幅度0.859、接段2.352、姿态0.425 rad、根位置0.344 m、jerk652.17;2327进度0.770/0.947、961为1.388/0.929。250步退化:幅度0.734、根位置0.506 m、jerk277.79,但2327一个种子-0.006、961一个种子0.031,不能因jerk降低采用末步。保留125候选和完整最终断点,不续训退化配方。两模型已保存全部轨迹有限,case_keys/参考与对应配方对照一致,初始逐元素复现;训练loss/梯度记录有限,证据artifacts/recovery_hourly_20261110T13_analysis.json。这些诊断不是正式FID。 ++ ++立即启动Kimodo v29扩大验证:/root/autodl-fs/pangxinlei/models/hfd-recovery-broad-v29-20261110/candidate125,入口scripts/recovery_broad_v29/supervise_pilot.py,固定v28 step-000125.pt,沿用原24validation案例(各模式覆盖)、两种子、12帧执行与20步CFG2;零训练更新,无Test1024。必须与v18 expanded baseline、v20候选比较全部配对轨迹、行走/坍缩和jerk;日志中的原六案例门槛标记仍仅为临时诊断,不能据此正式开训。候选jerk仍高于原初始,扩大验证不能只看平均误差。GPU1启动前空闲,真实子进程580064和监督器580063已核验存活;HY继续GPU0。2GiB预算实写fsync预留通过,4小时墙钟/日志轮转/容量inode监督保留;源码SHA/启动/健康证据recovery_v29_launch_kimodo.json、recovery_v29_health.json。 ++ ++清理已结束Kimodo v28的重复启动索引328.12MiB,确认旧进程/监督器退出、无打开句柄、新副本SHA一致;所有权重和轨迹保留,HY运行资源未动。清单recovery_cleanup_hourly_20261110T13.json。本地约1.536GiB、共享inode超过37000,本地/共享实写fsync通过;共享df不作账号配额。尚无正式训练,小时任务416e9f54继续,下一次14 UTC同步分析HY完成结果与Kimodo扩大集;正式训练健康启动后立即取消规则保持。 ++ ++### 第三十轮双模型恢复试验(2026-11-10 14 UTC) ++ ++HY v28正常完成250步,2步恢复收据通过,全部训练loss/梯度及0/125/250轨迹有限,初始与原对照精确复现。最终幅度0.561、接段1.772、无低于0.1坍缩,姿态0.421 rad、根位置0.441 m、jerk359.06;相对原始姿态0.372仍超过10%恶化限,961两种子行走进度-0.057/-0.006,未达资格。100%生成历史使幅度改善,但未解决行走,不能将整体平均改善当正式开训依据。 ++ ++Kimodo v29扩大验证正常完成,无错误/资源停止、预留释放;原始扩大基线、v19step125和v28step125三组48条完整轨迹均有限、case_keys/GT完全匹配。v28候选幅度0.794、接段2.239、姿态0.408 rad、根位置0.568 m、jerk961.69,无坍缩、明显位移失败8/24;v19候选根位置0.528、jerk906.78、同样8/24失败,因此不采用延长标签过渡。完整配对指标及案例保存在artifacts/recovery_hourly_20261110T14_analysis.json。这些是原始SMPL-X诊断,不是正式FID。 ++ ++同步启动v30各250步,根目录/root/autodl-fs/pangxinlei/models/hfd-recovery-pilot-v30-20261110,入口scripts/recovery_training_v30/supervise_pilot.py。HY以v28为对照,仅将根速度辅助损失0改为0.1,保留全生成历史与旋转辅助0,检验在动作幅度恢复后明确根速度目标是否改善行走。Kimodo回到扩大集较好的v19配方(6帧标签过渡),仅统一把root/body学习率峰值和下限乘0.5,warmup/衰减时点及总250步保持;目的是检验后期退化与更新幅度的关系,不预设成功、不无限延长。两者仍从原正式best初始化,保持原始验收基线、0/125/250同案例两种子,Test1024未使用。 ++ ++配置差异和源码匹配检查recovery_v30_checks.json通过;没有新增监督表示/推理路径代码,恢复合同按独立源码/配置生成。HY2GiB/Kimodo12GiB预算实写fsync预留完成;GPU0/1启动前空闲,双模型真实子进程和监督器核验存活,目前执行两步恢复检查,尚未宣称该检查通过。启动SHA收据recovery_v30_launch_*.json,资源和PID见recovery_v30_health.json。保留8小时监督器、磁盘/inode/目录项检查、日志轮转和原子断点失败保留旧文件机制。 ++ ++清理已结束HYv28及Kimodo v29重复启动索引656.24MiB,确认旧进程/监督器退出、无打开句柄及新副本SHA一致,清单recovery_cleanup_hourly_20261110T14.json;正式权重、完整恢复断点、候选和全部轨迹保留。本地约1.535GiB,共享inode超过37000,本地/共享写入fsync通过;共享df不当账号配额。尚未正式开训,小时任务416e9f54继续,下次15 UTC检查两边恢复收据和训练/验证;任一正式训练健康启动立即取消规则保持。 ++ ++调度更新(2026-11-10):用户要求将检查频率改为每半小时。已通过paseo heartbeat update将现有任务416e9f54改为 `*/30 * * * *`(Etc/UTC),持久化记录核验通过;每小时整点及30分执行,下次2026-11-10 15:00 UTC。任务旧名称仍含“每小时”,实际频率以cron为准。继续同步推进两模型;任一正式训练健康启动立即取消的条件不变,不创建重复任务。 ++ ++### 半小时检查及Kimodo v31扩大验证(2026-11-10 15:00 UTC) ++ ++HY v30持续GPU0训练,检查157步,结束记录160/250;两步optimizer/scheduler断点回读及恢复通过,无错误/资源停止。125步幅度0.446、接段1.732、姿态0.427 rad、根位置0.414 m,仍不合格,继续当前有界预算。训练loss/梯度记录及已保存0/125完整轨迹有限,不重启不清理正在使用的资源。 ++ ++Kimodo v30完成250步,恢复检查通过,无错误/资源停止、预留释放。125步通过小验证:幅度0.730、接段2.124、姿态0.425 rad、根位置0.363 m、jerk628.92;2327进度0.751/0.761,961为1.129/0.507(后者接近下限,须关注扩大集稳定性)。250步幅度0.718、接段2.346、姿态0.414、根位置0.403、jerk384.19,但2327一个种子0.461、961一个种子0.365不合格。减半学习率没有消除后期退化,不续训失败末步。所有0/125/250完整轨迹有限、case_keys/参考与v19一致,初始精确复现;对照证据artifacts/recovery_hourly_20261110T1500_analysis.json,非正式FID。 ++ ++GPU1已启动v31:/root/autodl-fs/pangxinlei/models/hfd-recovery-broad-v31-20261110/candidate125,入口scripts/recovery_broad_v31/supervise_pilot.py,固定v30 step-000125.pt。同24validation案例、两个种子、原采样和12帧执行;零更新无Test1024。原6案例资格标记不能替代扩大集验收,须与v18原始扩大基线、v20及v29完整轨迹比较,检查行走失败、接段、幅度、jerk与姿态/位置误差。实际子进程582073及监督器582072已核验存活;2GiB实写fsync预留通过,4小时及容量/inode/日志保护沿用。启动和健康证据recovery_v31_launch_kimodo.json、recovery_v31_health.json。 ++ ++清理已结束Kimodo v30重复索引328.12MiB,确认旧进程/监督器退出、无打开句柄且新副本SHA一致;权重/完整断点/轨迹保留,HY预留和运行缓存未动。清单recovery_cleanup_hourly_20261110T1500.json。本地约1.535GiB,共享inode超过37000,本地/共享实写fsync检查通过;不将共享df作为账号配额。无正式训练,保持半小时cron */30,下一次15:30 UTC,优先分析任一已完成模型;正式训练健康启动后立即取消任务416e9f54的条件不变。 ++ ++### 半小时检查及v32采样覆盖对照(2026-11-10 15:30 UTC) ++ ++HY v30正常完成250步,两步恢复检查通过,无错误/资源停止;最终幅度0.512、接段1.707、姿态0.419 rad、根位置0.436 m、jerk302.88。961行走进度-0.048/-0.006,姿态仍超原初始10%限,未合格;增加根速度辅助降低jerk但没有恢复该行走案例。训练loss/梯度和0/125/250完整轨迹有限,参考和初始复现核验通过。 ++ ++Kimodo v31扩大验证正常完成,预留释放、无错误。v30半学习率候选幅度0.708、接段2.061、姿态0.407 rad、根位置0.575 m、jerk732.01,无坍缩但行走失败13/24;v19候选为8/24、根位置0.528、jerk906.78。降低学习率有平滑收益但伴随运动不足,不能正式晋级。与原扩大基线/v19全部48条轨迹有限且case_keys/GT匹配,原始及逐案例证据artifacts/recovery_hourly_20261110T1530_analysis.json;非FID。 ++ ++转查采样覆盖:代码确认预测历史链都从decision0开始,原链长50个12帧窗口最多覆盖开头20秒。v32两模型各只把链长50改25(10秒),在固定250更新内增加会话切换机会;HY对照v30,Kimodo回到扩大集较好的v19对照。它减少连续窗口相关性,但也缩短后半段覆盖,属于待证假设,不能预设修复;下一次需从训练locations检查独立会话数和实际时域覆盖并结合运动/行走/姿态。仍从原正式best初始化、全预测历史、初始单姿态启动,不引入GT历史替换。验证不缩短,保持原完整时长、原案例和两个种子;原门槛不变,无Test1024。 ++ ++新根目录/root/autodl-fs/pangxinlei/models/hfd-recovery-pilot-v32-20261110,入口scripts/recovery_training_v32/supervise_pilot.py。配置差异及生效字段检查recovery_v32_checks.json;HY的curriculum.chain_segments和pilot.chain_segments均25,Kimodo读取pilot.chain_segments。未改监督表示/推理代码。HY2GiB、Kimodo12GiB实写fsync预留通过,启动前GPU0/1空闲,真实进程/监督器已核验存活,当前执行两步恢复检查,尚不宣称完成。8小时监督/有限日志/目录项与磁盘inode检查/原子断点保护保留,启动及健康收据recovery_v32_launch_*.json、recovery_v32_health.json。 ++ ++清理已结束HYv30及Kimodo v31重复索引656.24MiB,旧进程/监督器退出、无打开句柄及新副本SHA一致后删除,清单recovery_cleanup_hourly_20261110T1530.json。所有权重、完整恢复断点及轨迹保留。本地约1.534GiB、共享inode超过36000,本地/共享实写fsync通过,共享df不作账号配额。尚无正式训练,半小时任务416e9f54继续,下一次16:00 UTC;任一正式训练健康启动立即取消规则不变。 ++ ++### v32运行中半小时检查(服务器时间2026-11-10 16:00 UTC) ++ ++本次核查以服务器date和运行收据为时间依据;对话环境日期标为11月11日,但实际服务器为11月10日16:00 UTC。最新仍是v32,没有发现正式训练启动。两模型真实进程和监督器存活,HY结束检查79/250步,Kimodo 202/250步;两步optimizer/scheduler恢复收据均通过,未出现错误或资源停止,训练loss/梯度记录和所有已保存完整轨迹有限,初始复现及参考配对通过。不重启、不抢占GPU。 ++ ++HY尚只有初始验证;Kimodo125幅度1.035、接段2.372、姿态0.455 rad、根位置0.454 m,未合格:2327种子12345进度0.430、961种子2026进度0.364低于0.5,其余两个1.013/0.956。等待预算内中途/最终轨迹,不仅看loss或幅度。链长改动实际增加了覆盖:相同79步HY训练独立会话8→16;相同200步Kimodo为18→33,所有比较更新的locations均有记录。最远起点588→288帧,验证了减少后段覆盖的代价;这些变化不等于质量改善,仍按原完整验证门槛判定。 ++ ++证据artifacts/recovery_hourly_20261110T1600_analysis.json。本地约1.534GiB、共享空闲inode36968,本地/共享1MiB实写fsync通过,当前每运行目录项低于500。两模型正在使用的预留、缓存、断点保持;本轮无新增已核验可删除项,清理检查写recovery_cleanup_hourly_20261110T1600.json(0字节删除)。不重复创建实验或扩大预算。文档和任务状态已更新,半小时调度下一次16:30 UTC,任一正式训练健康启动立即取消规则保持。 ++ ++### 半小时检查及Kimodo模块诊断(服务器2026-11-10 16:30 UTC) ++ ++HY v32继续GPU0,检查158步,结束记录177/250。恢复检查通过,无错误/资源停止,记录loss/梯度和已保存初始/125完整轨迹有限。125幅度0.419、接段1.962、姿态0.451 rad、根位置0.419 m,仍不合格,继续既定有界预算。 ++ ++Kimodo v32正常完成250步,恢复检查通过,预留释放,无错误。125和250均未合格;250幅度0.865、接段2.551、姿态0.426 rad、根位置0.477 m、jerk325.49;2327进度0.312/0.537、961为0.186/0.438,4组合3失败。增加会话覆盖未解决行走,不能因jerk降低晋级。完整0/125/250轨迹有限、参考匹配及初始复现通过,证据artifacts/recovery_hourly_20261110T1630_analysis.json。保持原正式基线与门槛,无Test1024,不是FID。 ++ ++下一步停止连续标量调参,启动冻结v33模块交换诊断:/root/autodl-fs/pangxinlei/models/hfd-recovery-modules-v33-20261110/kimodo,入口scripts/recovery_modules_v33/supervise_pilot.py。以v19 step125为候选,三组分别原候选、只换回原正式best根模块、只换回原正式best身体模块;原正式best不是官方原生权重。各组先严格重新加载完整候选,再覆盖一个模块,24validation案例两种子同协议,零训练更新、无未来GT条件。权重键/形状匹配通过,root/body各204键,检查recovery_v33_checks.json。第一组必须逐元素复现v20;模块交互可能导致分布不匹配,不能将混合权重改善直接当因果损坏证明或正式资格。结果用于定位下一训练改动。 ++ ++GPU1启动前空闲,新子进程583485和监督器583482存活,HY原进程持续。2GiB实写fsync预留、4小时墙钟、日志轮转及容量/inode检查保留;启动/源码SHA和健康收据recovery_v33_launch_kimodo.json、recovery_v33_health.json。清理已结束Kimodo v32重复索引328.12MiB,旧进程退出、无打开句柄、新副本SHA一致后删除,清单recovery_cleanup_hourly_20261110T1630.json;全部模型断点和轨迹保留,HY在用资源不动。本地约1.534GiB、共享inode36935,本地/共享实写fsync通过,目录项低于500。无正式训练,半小时任务继续,下次17:00 UTC;正式训练健康启动即取消条件不变。 ++ ++### 双模型运行中检查(服务器2026-11-10 17:00 UTC) ++ ++HY v32检查时249步,结束记录249/250;真实训练进程和监督器存活,尚无completed收据及最终验证结论,不将步数达到预算当成已完成。两步恢复检查通过,loss/梯度记录及已保存轨迹有限,初始/参考复现通过。保留当前预算,等待后台最终验证与保存,不重复启动。 ++ ++Kimodo v33仍运行,原候选组已完成:全部48条轨迹逐元素复现v20,幅度0.991、接段2.213、姿态0.416 rad、根位置0.528 m、jerk906.78,说明对照基线一致。模块交换组尚无完整结果,不预判根或身体模块的作用。GPU1真实主进程583485及文本进程运行,GPU0为HY;无错误/监督器停止。证据artifacts/recovery_hourly_20261110T1700_analysis.json。诊断不当正式FID或正式训练资格。 ++ ++本地约1.534GiB、共享inode36931,本地/共享1MiB实写fsync通过,两个运行目录项均低于500。当前预留/缓存/断点正在使用,本轮未发现新增已核验可清理项,0字节删除,清理检查recovery_cleanup_hourly_20261110T1700.json。两份文档与任务状态更新,未正式开训,半小时任务保持,下一次17:30 UTC;任一正式训练健康启动立即取消任务416e9f54。 ++ ++### 双模型采样诊断v34(服务器2026-11-10 17:30 UTC) ++ ++HY v32完成250步,无错误/资源停止,恢复检查通过;全部训练记录及初始/中途/最终轨迹有限。最终幅度0.476、接段2.025、姿态0.454 rad、根位置0.443 m、jerk318.87,2327进度1.015/1.698,961为-0.055/-0.005。缩短链后仍不合格,不延长该配方。Kimodo v33三臂完成,无错误,预留释放;48案例种子轨迹有限、参考配对一致,候选逐元素复现v20。候选行走失败8/24、根位置0.528、jerk906.78;原根模块替换失败10/24、根位置0.550、jerk864.14;原身体模块替换幅度0.453、失败17/24、根位置0.633、jerk590.09,接近原正式模型。说明恢复后的身体模块对幅度/行走改善很重要,简单回滚根或身体都不解决问题;模块交互影响仍在,不能断言单一模块损坏。证据artifacts/recovery_hourly_20261110T1730_analysis.json,非FID。 ++ ++同步启动v34固定权重采样步数诊断,目录/root/autodl-fs/pangxinlei/models/hfd-recovery-sampling-v34-20261110/{hymotion,kimodo},入口scripts/recovery_sampling_v34_{hymotion,kimodo}/supervise_pilot.py。HY固定较好幅度的v28最终权重,6validation×2种子比较10与50步;Kimodo固定v19 step125,24validation×2种子比较20与50步。先要求原采样组逐元素复现HYv28最终/Kimv20,再分析50步是否改善运动、行走和jerk及延迟代价。之前50步HY对照针对原正式权重,本轮针对恢复候选,非重复同一配方。未更改初始状态/执行窗口/因果文本/GT条件,不训练更新、不读取Test1024,验收基线不重置。更高步数即使改善也须记录计算成本,不能混同原等预算评测。 ++ ++两组各2GiB实写fsync预留完成,GPU0/1启动前空闲,真实子进程/监督器核验存活,启动/源码SHA及健康见recovery_v34_launch_*.json、recovery_v34_health.json。4小时墙钟、磁盘/inode/目录项限制、有限日志/缓存保留。清理已结束HYv32和Kimv33重复索引656.24MiB,删除前确认旧进程退出、无句柄占用、新副本SHA一致,清单recovery_cleanup_hourly_20261110T1730.json;所有权重、完整恢复断点及轨迹保留。本地约1.533GiB、共享inode超过36800,本地/共享实写fsync通过,共享df不是账号配额。无正式训练,半小时调度继续,下次18:00 UTC;任一正式训练健康启动立即取消任务416e9f54。 ++ ++### HY文本条件诊断及Kimodo运行检查(服务器2026-11-10 18:00 UTC) ++ ++HY v34两组正常完成,无错误/资源停止、预留释放。10步组逐元素复现v28最终;50步幅度0.887、接段1.671、姿态0.454 rad、根位置0.687 m、jerk368.75,较10步根位置0.441明显退化;2327进度2.055/3.613过冲,961为-0.283/-0.573反向,4组全不合格。采样更精细不等于恢复,不采用50步。两组完整轨迹有限、case_keys/GT相同。 ++ ++Kimodo v34仍在运行50步组;已完成20步组全部48条轨迹逐元素复现v20。尚无50步完整结果,不推断改善。真实进程584260及监督器584259存活,无错误/资源停止。完整核验和可写性证据artifacts/recovery_hourly_20261110T1800_analysis.json。 ++ ++HY继续推进冻结v35:/root/autodl-fs/pangxinlei/models/hfd-recovery-text-v35-20261110/hymotion,入口scripts/recovery_text_v35/supervise_pilot.py,固定v28最终权重与原10步采样,比较full_text和empty_text两臂、原6案例两种子。检查已恢复幅度候选的持续行走失败是否依赖文本条件;此前空文本实验使用别的权重,本轮不冒充相同对照。空文本经原编码器编码,不是零向量;只改变推理文本,轨迹/未来GT不输入条件。空文本偏离任务条件,仅用于机制诊断,不能作为正式评测或放宽资格。要求full_text精确复现v28最终。零权重更新,无Test1024。 ++ ++GPU0启动前空闲,HY真实子进程584643和监督器584642存活,GPU1继续Kimodo。2GiB实写fsync预留通过,4小时墙钟、日志轮转、容量/inode/目录项保护保留;启动与健康收据recovery_v35_launch_hymotion.json、recovery_v35_health.json。清理已结束HYv34重复启动索引328.12MiB,核验旧进程退出、无打开句柄、新副本SHA相同后删除;Kimodo在用预留不动,所有权重和轨迹保留,清单recovery_cleanup_hourly_20261110T1800.json。本地约1.533GiB、共享inode36828,本地/共享1MiB实写fsync通过。尚无正式训练,半小时任务继续,下次18:30 UTC,正式训练健康启动即取消规则保持。 ++ ++### HY v36执行窗口权重试验(服务器2026-11-10 18:30 UTC) ++ ++HY v35正常完成,预留释放,无错误。全文组完整轨迹逐元素复现v28最终;空文本组幅度0.544、接段1.874、姿态0.413 rad、根位置0.514 m、jerk339.82,2327进度0.082/0.011,961为-0.051/-0.010。去掉文本未恢复961且破坏2327行走,不能采用;仅为机制诊断。两臂轨迹有限、案例/参考相同,证据artifacts/recovery_hourly_20261110T1830_analysis.json。Kimodo v34仍正常运行50步组,原20步组精确复现已核验,目前没有50步完整结果,不重启或抢占GPU1。 ++ ++HY恢复训练v36,/root/autodl-fs/pangxinlei/models/hfd-recovery-pilot-v36-20261110/hymotion,入口scripts/recovery_training_v36/supervise_pilot.py。对照v28配方,仅把training.execution_weight从3改6,保持全生成历史、50窗口链、原正式best初始化及全部其他训练设置。代码核实该权重作用于未来60帧中实际执行的前12帧,并以加权总和归一化,因此并非简单放大总体loss。目的是检验近端执行目标的相对权重,不预设解决行走。原10步推理/正常因果全文/固定validation两种子/原验收基线不改,无Test1024,各250步,0/125/250比较。配置差异、入口源码一致核验recovery_v36_checks.json。 ++ ++GPU0启动前空闲,真实子进程585015及监督器585014存活,当前执行2步恢复检查,尚不宣称检查已完成;2GiB实写fsync预留通过,8小时墙钟、日志轮转/目录项/磁盘inode/原子断点保护保留。启动SHA和健康收据recovery_v36_launch_hymotion.json、recovery_v36_health.json。Kimodo主进程584260持续,监督器无错误或停止。 ++ ++清理已结束HYv35重复索引328.12MiB,确认旧进程退出、无句柄占用及新副本SHA相同;清单recovery_cleanup_hourly_20261110T1830.json,全部权重/断点/轨迹保留,Kimodo在用资源不动。本地约1.533GiB,共享inode36795,本地/共享实写fsync通过,目录项低于500。未正式开训,半小时任务保持,下次19:00 UTC;任一正式训练健康启动即取消规则不变。 ++ ++ ++### 半小时检查与Kimodo引导强度诊断(服务器2026-11-10 19:00 UTC) ++ ++HY v36仍在GPU0,真实进程585133、监督器585014,记录124/250步,正等待中途验证结果;控制台有数据读取进展,容量监督收据更新,无错误/资源停止。两步optimizer/scheduler恢复和读回通过,全部训练记录有限、已保存完整轨迹有限且初始/参考复现通过,不重复启动。当前尚不能判断本轮改进。 ++ ++Kimodo v34两组已完成且预留释放。全部48条/组完整轨迹有限、case_keys及参考一致,20步组精确复现v20。50步相比20步:幅度0.991→1.293、接段2.213→2.063、姿态0.416→0.447rad、根位置0.528→0.563m、根jerk906.78→1103.45,行走失败仍8/24;不能凭接段变小和幅度增加采用更高采样步数。证据artifacts/recovery_hourly_20261110T1900_analysis.json,非正式FID,无Test1024。 ++ ++Kimodo继续冻结v37引导强度单变量诊断:/root/autodl-fs/pangxinlei/models/hfd-recovery-guidance-v37-20261110/kimodo,入口scripts/recovery_guidance_v37/supervise_pilot.py。固定v19step125、20步采样、原24validation两种子、12帧执行及正常因果全文,只比较guidance2与1,排查引导放大是否贡献抖动及行走失败。guidance2必须精确复现v20;不预设低引导能恢复,不改变验收门槛。参数通过closed_loop的config.guidance实际传入采样器;零训练更新,不是正式训练。 ++ ++GPU1启动前空闲,真实子进程585526与监督器585525存活、初始化日志推进。2GiB实写fsync预留通过,4小时墙钟/容量inode/目录项/日志轮转保护保留,源码SHA与健康证据recovery_v37_launch_kimodo.json、recovery_v37_health.json。清理已结束Kimodo v34重复索引328.12MiB,旧进程退出、无打开句柄及新副本SHA一致后删除;所有权重和轨迹保留,HY在用资源未动,清单recovery_cleanup_hourly_20261110T1900.json。本地约1.532GiB,共享空闲inode36761,本地/共享1MiB实写fsync通过;共享df不作账号配额。半小时任务416e9f54继续,下次19:30 UTC;尚无正式训练,正式训练健康启动即取消规则保持。 ++ ++ ++### 双模型半小时运行检查(服务器2026-11-10 19:30 UTC) ++ ++HY v36实际推进到224/250,进程585133/监督器585014存活,两步恢复读回通过,全部loss/梯度记录及初始/125完整轨迹有限,参考/案例匹配、初始精确复现。125步幅度0.533、接段1.733、姿态0.420rad、根位置0.438m、jerk364.83;2327进度1.150/1.325,961为-0.030/-0.004。幅度达线但行走仍失败,姿态较原初始0.3723增约12.8%,超过10%门槛,尚不能正式开训。继续既定250步预算等待最终结果,不以平均数或loss作恢复结论。 ++ ++Kimodo v37运行中,真实进程585526/监督器585525存活,GPU1约17GiB。guidance2对照48条完整轨迹精确复现v20;guidance1组尚无完整结果,不预判改善或重复启动。HY继续GPU0。无错误/资源停止,证据artifacts/recovery_hourly_20261110T1930_analysis.json。当前两模型均为pilot或冻结诊断,没有正式训练。 ++ ++本地约1.532GiB,共享空闲inode36748,本地/共享1MiB实写fsync通过;共享df不是账号配额。两运行目录项低于500,现有预留与缓存正在使用,本轮无新增核验可删资源,0字节清理,记录recovery_cleanup_hourly_20261110T1930.json。全部断点和完整轨迹保留。半小时任务416e9f54保持,下次20:00 UTC;任一正式训练健康启动后立即取消规则不变。 ++ ++ ++### 双模型v38诊断(服务器2026-11-10 20:00 UTC) ++ ++HY v36完成250步,恢复读回通过,无错误,预留释放。全部训练记录及0/125/250完整轨迹有限,初始精确复现,参考/案例一致。最终幅度0.609、接段1.780、姿态0.41234rad(较原初始0.37227约+10.77%)、根位置0.452m、jerk355.50;2327进度1.489/1.855,961为-0.074/-0.071。仍不合格,不延长失败配方。Kimodo v37完成两臂,48条/组轨迹有限且参考一致,原引导2精确复现v20;引导1幅度0.980、接段2.258、姿态0.409、根位置0.614、jerk529.29(原906.78),行走失败8/24→16/24,不能以更平滑替代任务能力。证据artifacts/recovery_hourly_20261110T2000_analysis.json,无Test1024,非FID。 ++ ++两模型同步进入冻结v38,目录/root/autodl-fs/pangxinlei/models/hfd-recovery-guidance-v38-20261110/{hymotion,kimodo},入口scripts/recovery_guidance_v38_模型名/supervise_pilot.py。HY固定v36最终权重及10步,原6validation两种子比较guidance1与2;此前更强引导针对其他权重,本轮检查已恢复幅度候选。Kimodo固定v19step125、20步及24validation两种子,对照constant2与前10次去噪引导2/后10次引导1,检验强引导行走收益与抖动是否可按去噪阶段分开;不是根/身体模块分离。仅改变引导,无未来GT输入,原验收门槛不变,零训练更新。原臂必须精确复现HYv36最终/Kimv20。Kimodo本地采样副本撤销该切换后逐字节还原原实现,反向去噪迭代顺序验证通过,检查recovery_v38_checks.json;不修改上游或历史运行源码。 ++ ++两GPU启动前空闲,各2GiB实写fsync预留完成;HY真实进程586140/监督器586137、Kimodo586144/586142存活,初始化中,无错误。启动SHA及健康证据recovery_v38_launch_*.json、recovery_v38_health.json。4小时墙钟、容量/inode/目录项/日志保护保留。清理已结束HYv36和Kimv37重复启动索引656.24MiB,旧进程退出、无打开句柄及新副本SHA一致后删除;所有权重、恢复断点及轨迹保留,清单recovery_cleanup_hourly_20261110T2000.json。本地约1.532GiB,共享inode36697,本地/共享实写fsync通过,共享df不作账号配额。没有正式训练,半小时调度保持,下次20:30 UTC;任一正式训练健康启动即取消任务416e9f54。 ++ ++ ++### v38分析、故障修复及双模型v39(服务器2026-11-10 20:30 UTC) ++ ++HY v38完成、预留释放。两臂完整轨迹有限且参考/案例相同,引导1逐元素复现v36最终。引导2幅度0.645、接段1.893、姿态0.423rad、根位置0.729m、jerk560.34,2327进度2.581/1.443、961为0.079/0.011,3/4行走组合失败;原引导1根位置0.452、jerk355.50,仍不采用更强引导。证据artifacts/recovery_hourly_20261110T2030_analysis.json,非FID。 ++ ++Kimodo v38失败退出,原因是新增独立recovery_generate.py保留包内相对导入,make_window调用from .kimodo_causal_rep报ImportError。此前AST检查未覆盖实际导入路径,是本轮入口错误,未产生完整诊断结果,不归因于模型。监督器已释放预留,失败日志保留。v39独立复制源码只把两处包内相对导入改为hfd_bridge绝对导入;模块及依赖实际导入检查通过、AST确认无相对导入,recovery_v39_checks.json。保持v38的冻结权重、constant2/late1两臂和全部验证协议;原臂轨迹复现仍待完成,不把导入通过当效果验证。 ++ ++HY同步v39有界250步,从原正式best初始化,以v36全部配置为对照,仅根速度辅助权重0→0.1,保留execution_weight6、全生成历史50窗口链及10步推理。依据是v36改善幅度,而v30在execution_weight3下的小速度辅助曾减少jerk;本轮检验该作用能否在v36保留,不能预设改善行走,不机械延长或改门槛。固定原validation两种子,无Test1024。 ++ ++目录/root/autodl-fs/pangxinlei/models/hfd-recovery-v39-20261110/{hymotion,kimodo},入口scripts/recovery_v39_模型名/supervise_pilot.py。HY真实进程586697/监督器586694处于两步恢复检查初始化;Kimodo586701/586699处于诊断初始化,无新错误,未宣称恢复/诊断已通过。两GPU启动前空闲,各2GiB实写fsync预留通过;HY8小时/Kim4小时墙钟、日志/缓存/容量inode/断点保护保留,启动与健康见recovery_v39_launch_*.json、recovery_v39_health.json。 ++ ++清理已结束HYv38和已失败Kimv38重复索引656.24MiB,确认旧进程退出、无打开句柄、新副本SHA相同后删除;全部权重、恢复断点、轨迹及错误证据保留,清单recovery_cleanup_hourly_20261110T2030.json。本地约1.531GiB、共享inode36632,本地/共享实写fsync通过,共享df不作账号配额,目录项低于500。尚无正式训练,半小时任务继续,下次21:00 UTC;任一正式训练健康启动即取消规则保持。 ++ ++ ++### v39半小时运行检查(服务器2026-11-10 21:00 UTC) ++ ++HY v39推进108/250,真实训练进程586859/监督器586694存活;两步optimizer/scheduler恢复及断点读回通过,已记录loss/梯度全部有限,初始完整轨迹精确复现及参考配对通过。尚未到125步中途验证,不能判断新辅助项效果,不重复启动或扩大预算。 ++ ++Kimodo v39修复重试正常运行,进程586701/监督器586699存活,GPU1约17GiB。constant2对照完成,全部48条完整轨迹逐元素精确复现v20,有限性及案例/参考匹配通过,确认本次导入修复与采样副本没有改变对照输出。late1诊断仍运行,无完整结果,不预判改善;没有再次出现导入错误或资源停止。HY继续GPU0。证据artifacts/recovery_hourly_20261110T2100_analysis.json。 ++ ++本地约1.531GiB,共享空闲inode36605,本地/共享1MiB实写fsync通过,共享df不作账号配额。两个运行目录项低于500,预留/缓存/恢复断点正在使用,本轮没有新增已核验可清理项,0字节删除,清单recovery_cleanup_hourly_20261110T2100.json。尚无正式训练,保留半小时任务416e9f54,下一检查21:30 UTC;任一正式训练健康启动即取消规则不变。 ++ ++ ++### Kimodo v40与HY运行检查(服务器2026-11-10 21:30 UTC) ++ ++HY v39检查209步、结束更新210/250,真实进程586859及监督器586694存活。恢复检查通过、全部loss/梯度及初始/125完整轨迹有限,参考配对/初始复现通过。125幅度0.499、接段1.747、姿态0.423rad、根位置0.417m、jerk302.33;2327进度1.017/1.255,961为-0.047/-0.019,仍未达门槛,继续原250步预算。预留实写期间出现一次114.8秒训练步耗时,进程仍推进,无资源停止,不将此解释为模型性能变化。 ++ ++Kimodo v39两组完成,原臂精确复现v20、完整48条/组轨迹有限且参考一致。late1幅度0.978、接段2.351、姿态0.417、根位置0.542m、jerk892.22(对照906.78,仅降约1.6%),行走失败仍8/24,根位置误差增大,不采用该推理改动。证据artifacts/recovery_hourly_20261110T2130_analysis.json,无Test1024,非FID。 ++ ++Kimodo启动有界训练v40:/root/autodl-fs/pangxinlei/models/hfd-recovery-pilot-v40-20261110/kimodo,入口scripts/recovery_training_v40/supervise_pilot.py。从原正式best初始化,以v19为基线只把recovery.temporal_weights.decoded_root_acceleration从0.01提高到0.1,其余训练/验证/恢复配方保持,250步并保存0/125/250对照。该项跟踪解码后根位置二阶差分与参考加速度(包含fps及评分掩码),不是把所有加速度压零。此前引导改动无法兼顾行走与抖动,检验加强监督是否有帮助,不预设改善或放宽门槛。原入口/runtime源码一致、单变量配置核验recovery_v40_checks.json。 ++ ++GPU1启动前空闲,12GiB预算实写fsync预留通过(实际reserve约11.42GiB,非shared df配额推断);监督器587407与真实子进程存活,当前恢复检查初始化,尚未宣称通过。8小时墙钟、原子断点/日志/缓存/容量inode保护保留,启动SHA/健康recovery_v40_launch_kimodo.json、recovery_v40_health.json。HY GPU0继续运行。清理已结束Kimodo v39重复索引328.12MiB,确认旧进程退出、无打开句柄及新副本SHA相同;权重/恢复断点/完整轨迹保留,清单recovery_cleanup_hourly_20261110T2130.json。本地约1.531GiB,共享inode36585,本地/共享实写fsync通过,目录项低于500。未正式开训,半小时调度继续,下次22:00 UTC;正式训练健康启动即取消规则保持。 ++ ++ ++### HY扩展验证v41及Kimodo v40(服务器2026-11-10 22:00 UTC) ++ ++HY v39正常完成250步、恢复读回通过,预留释放;所有训练记录及0/125/250完整轨迹有限,原初始精确复现、参考/案例配对通过。最终幅度0.563、接段1.724、姿态0.418rad、根位置0.439m、jerk302.30;2327进度1.069/1.454,961为-0.069/-0.007,姿态及行走仍不达门槛。对比v36降低jerk但未恢复行走,不延长此配方。Kimodo v40检查196步、结束200/250,真实587571/监督器587407存活;完整3.394GB断点两步恢复通过、loss/梯度和初始/125轨迹有限,初始/参考匹配通过。125小验证幅度0.886、接段2.343、姿态0.442rad、根位置0.367m,现有小验证门槛通过;尚需最终结果及更大validation,不能据此正式开训。证据artifacts/recovery_hourly_20261110T2200_analysis.json,非FID。 ++ ++HY继续冻结扩展诊断v41:/root/autodl-fs/pangxinlei/models/hfd-recovery-broad-v41-20261110/hymotion,入口scripts/recovery_broad_v41_hymotion/supervise_pilot.py。同24validation案例、两种子比较原正式best、v36最终、v39最终三臂;6种交互模式各4例,选择不依赖模型结果并包含原6例,复用Kimodo扩展选样规则。先产生相同24例原正式基线,不沿用6例误差作比较门槛;逐臂使用已有asset/adapter键/shape校验完整重载权重,保持10步/引导1/12帧执行及因果文本,无训练更新/无Test1024。原6例子集需精确复现,各臂参考需一致。此轮为失败分布与覆盖诊断,不是已通过小验证后的正式晋级复核,不宣称候选合格。代码检查recovery_v41_checks.json。 ++ ++GPU0启动前空闲,监督器587842和子进程存活,仍在初始化;2GiB实写fsync预留通过,4小时墙钟、日志/缓存/容量inode保护保留,启动SHA和健康recovery_v41_launch_hymotion.json、recovery_v41_health.json。Kimodo GPU1继续训练,当前预留与恢复断点不动。清理已结束HYv39重复索引328.12MiB,旧进程退出、无打开句柄和新副本SHA一致后删除,所有权重/完整轨迹保留,清单recovery_cleanup_hourly_20261110T2200.json。本地约1.531GiB,共享inode36532,本地/共享实写fsync通过,目录项低于500,共享df不作账号配额。没有正式训练,半小时任务继续,下次22:30 UTC;任一正式训练健康启动即取消规则保持。 ++ ++ ++### Kimodo最终分析及扩大验证v42(服务器2026-11-10 22:30 UTC) ++ ++Kimodo v40正常完成250步,恢复读回通过、预留释放。全部训练loss/梯度以及0/125/250完整轨迹有限,初始精确复现、参考与案例匹配。125步幅度0.886、接段2.343、姿态0.442rad、根位置0.367m、jerk536.63;2327进度0.887/1.015、961为1.272/0.888,四组均达标。250步幅度0.665、接段2.671、姿态0.418、根位置0.493、jerk206.09,但2327进度0.004/0.788、961为0.045/0.162,三组行走失败且根位置超原基线10%。更低jerk伴随行走退化,不能晋级最终权重,不继续延长。证据artifacts/recovery_hourly_20261110T2230_analysis.json,非FID,无Test1024。 ++ ++保留125步候选,启动冻结扩大验证v42:/root/autodl-fs/pangxinlei/models/hfd-recovery-broad-v42-20261110/candidate125,入口scripts/recovery_broad_v42/supervise_pilot.py。同24validation×2种子、各模式4例,原20步guidance2/12帧执行不变。启动前baseline-validation.json替换为v18/baseline的48条原正式模型结果,避免旧6案例误差门槛与扩展样本错配;来源及复现操作写experiment.json和recovery_v42_checks.json,门槛未放宽。需核验原6子集精确复现v40step125、全量参考/案例一致和逐失败案例。小验证通过不能直接启动正式训练。 ++ ++GPU1启动前空闲,新监督器588139与子进程存活,初始化推进;2GiB实写fsync预留通过,4小时墙钟/日志/缓存/容量inode保护保留,recovery_v42_launch_kimodo.json、recovery_v42_health.json。HY v41进程587845/监督器587842存活,GPU0约18GiB,容量心跳更新,第一臂尚无完整输出,不声称已复现,不重启。原来较小HY验证耗时不能直接套到此次24例×三权重上。 ++ ++清理已完成Kimodo v40重复索引328.12MiB,确认旧进程退出、无打开句柄及新副本SHA相同后删除,所有权重/完整恢复断点/轨迹保留;recovery_cleanup_hourly_20261110T2230.json。HY在用资源不动。本地约1.531GiB,共享inode约36500,本地/共享实写fsync通过,共享df不作账号配额,目录项低于500。尚无正式训练,半小时任务保持,下次23:00 UTC;任一正式训练健康启动即取消规则保持。 ++ ++ ++### 扩展结果及Kimodo中点诊断v43(服务器2026-11-10 23:00 UTC) ++ ++HY v41原正式模型扩展基线已完成:幅度0.271、接段1.250、姿态0.373rad、根位置0.768m、jerk331.81,明显位移组合失败23/24。48条轨迹有限、原6例子集逐元素复现;恢复候选尚未完成,进程587845/监督器587842和容量心跳正常,不能只根据基线宣布改善。Kimodo v42完成,预留释放、无错误;完整48条轨迹有限、与扩展原基线参考/案例相同,原6例精确复现v40step125。扩展幅度0.741、接段2.356、姿态0.403rad、根位置0.558m、jerk725.25,行走失败9/24,未过原门槛。对比v19候选失败8/24、jerk906.78、姿态0.416,平滑/姿态有所改善但行走没有整体改善,不能正式开训。证据artifacts/recovery_hourly_20261110T2300_analysis.json,无Test1024,非FID。 ++ ++启动冻结v43诊断:/root/autodl-fs/pangxinlei/models/hfd-recovery-midpoint-v43-20261110/kimodo,入口scripts/recovery_midpoint_v43/supervise_pilot.py。两臂是原v40step125对照及v40/v19两个step125完整denoiser浮点张量等权中点,非浮点缓冲保持一致;两者同初始化/架构,仅加速度训练权重不同。固定0.5而非扫描参数,检验行走与平滑性折中是否存在,不假设权重平均必然改善。实际检查键/shape/dtype一致、中点数值有限、非浮点缓冲相同,recovery_v43_checks.json。使用相同24validation两种子/20步guidance2与原48例基线,原臂需精确复现v42,零训练更新,不导出新大权重;原门槛不变。 ++ ++GPU1启动前空闲,监督器588526和真实子进程存活、初始化中;2GiB实写fsync预留完成,4小时墙钟/日志/缓存/容量inode保护继续,启动及健康recovery_v43_launch_kimodo.json、recovery_v43_health.json。HY GPU0继续。清理已结束v42重复索引328.12MiB,确认旧进程退出、无打开句柄、新副本SHA一致,所有权重/恢复断点/轨迹保留;recovery_cleanup_hourly_20261110T2300.json。本地约1.530GiB,共享inode约36450,本地/共享实写fsync通过,共享df不是账号配额,目录项低于500。尚无正式训练,半小时任务保持,下次23:30 UTC;任一正式训练健康启动即取消规则不变。 ++ ++ ++### 双模型扩展诊断运行检查(服务器2026-11-10 23:30 UTC) ++ ++HY v41进程587845及监督器587842存活,累计约88分钟,GPU0约18GiB且有实际计算活动;原正式模型基线已核验,恢复候选仍未完成,没有新完整结果,不重复启动。Kimodo v43进程588528及监督器588526存活,GPU1约17GiB且计算活跃;v40对照臂48条完整轨迹逐元素精确复现v42,数值有限、原6例和全量参考匹配。中点臂仍运行,不预判权重平均收益。两组无错误或资源停止,证据artifacts/recovery_hourly_20261110T2330_analysis.json。均为冻结诊断,零训练更新,不是正式训练/FID。 ++ ++本地约1.530GiB、共享空闲inode36448,本地/共享1MiB实写fsync通过,共享df不作账号配额。目录项低于500,运行中预留与缓存保持,本轮无新增已核验可清理项,0字节删除,清单recovery_cleanup_hourly_20261110T2330.json。两份文档及任务状态已更新;半小时调度保持,下一次11月11日00:00 UTC,任一正式训练健康启动立即取消416e9f54规则不变。 ++ ++ ++### Kimodo根训练范围v44及HY扩展进展(服务器2026-11-11 00:00 UTC) ++ ++HY v41完成v36臂:幅度0.544、接段1.776、姿态0.402rad、根位置0.738m、jerk357.01,行走失败22/24(原基线23/24);幅度恢复不等于行走恢复,未过门槛。完整48轨迹有限、全量参考/案例匹配、原6例子集逐元素复现v36最终。v39仍运行,真实587845/监督器587842存活,无错误,不抢占GPU0。Kimodo v43两臂完成,预留释放、无错误,中点幅度0.798、接段2.325、姿态0.408、根位置0.564m、jerk732.78,行走失败10/24,较v40的9/24和v19的8/24更差,不采用。完整轨迹有限且参考匹配。证据artifacts/recovery_hourly_20261111T0000_analysis.json,非FID,无Test1024。 ++ ++Kimodo启动v44有界250步:/root/autodl-fs/pangxinlei/models/hfd-recovery-pilot-v44-20261111/kimodo,入口scripts/recovery_training_v44/supervise_pilot.py。以v19为对照、原正式best初始化,仅training.root_trainable由full改projections;根模块input_linear/output_linear训练,内部参数冻结,身体模块全量训练,全部学习率/损失(加速度仍0.01)/链/验证配置保持。基于身体适配必要且多次根全量训练后期行走退化,检验限制根更新自由度是否有帮助,不预设根更新就是原因。使用已有受支持模式,不修改上游。单变量配置及优化器排除冻结参数验证通过,recovery_v44_checks.json;真实模型训练计数/恢复检查待启动完成,不宣称已通过。 ++ ++GPU1启动前空闲,监督器589285和真实子进程存活,处于恢复检查初始化;12GiB预算实写fsync预留通过(实际reserve约11.42GiB),8小时墙钟、日志/缓存/目录项/容量inode/原子断点保护保留,recovery_v44_launch_kimodo.json、recovery_v44_health.json。HY继续扩展诊断。清理已结束v43重复索引328.12MiB,确认旧进程退出、无打开句柄和新副本SHA相同;所有正式权重/恢复断点/候选/轨迹保留,清单recovery_cleanup_hourly_20261111T0000.json。本地约1.530GiB,共享inode36432,本地/共享实写fsync通过,共享df不作账号配额,目录项低于500。无正式训练,半小时调度保持,下次00:30 UTC;正式训练健康启动立即取消规则不变。 ++ ++ ++### 双模型运行检查(服务器2026-11-11 00:30 UTC) ++ ++HY v41监督器587842、真实子进程587845存活,扩展v39臂继续;原基线及v36臂已完成,v36行走失败22/24,尚无新完整候选结果。Kimodo v44监督器589285、训练子进程589455存活,审计182/250步,loss/梯度及已保存完整轨迹有限,初始复现、参考匹配通过。两步恢复读回通过,optimizer/scheduler均为2,完整断点2268515405字节。实际根可训练565253、身体141529356参数,符合根projections模式。125步小验证幅度1.144、接段2.255、姿态0.462rad、根位置0.317m,既定小验证门槛通过;jerk773.16仍需关注。最终及扩展验证未完成,不据此宣布正式训练资格。证据artifacts/recovery_hourly_20261111T0030_analysis.json;非FID,无Test1024。 ++ ++两组无错误收据、容量心跳继续,HY约2小时28分钟,仍在4小时诊断上限内;Kimodo受8小时训练上限保护。本地约1.530GiB、共享空闲inode36397,本地/共享1MiB实写fsync通过,共享df不作账号配额。目录项均低于500;保留运行中预留、缓存、完整断点及轨迹,无新增可安全清理项,删除0字节,清单recovery_cleanup_hourly_20261111T0030.json。未重复启动。尚无正式训练,半小时调度继续,下一检查点01:00 UTC,检查HY扩展结果及Kimodo最终验证;任一正式训练健康启动即取消416e9f54规则保持。 ++ ++ ++### 完成结果及双模型v45(服务器2026-11-11 01:00 UTC) ++ ++HY v41三臂完成、无错误且预留已释放,所有完整轨迹有限、48例参考匹配,原6例子集精确复现各自候选。v39扩展幅度0.495、接段1.718、姿态0.400rad、根位置0.739m、jerk303.50,行走失败22/24;对比v36幅度0.544、jerk357.01、同样22/24失败,平滑收益没有解决广泛行走不足,未过原门槛。Kimodo v44正常完成250步、恢复读回通过、预留释放;0/125/250完整轨迹和训练数值有限,初始及参考匹配。最终幅度1.043、接段2.447、姿态0.439rad、根位置0.454m、jerk394.23,但961两种子进度0.352/0.298,不合格。125步通过小验证(幅度1.144、根位置0.317m、jerk773.16),保留该候选进行扩展复核,不延长退化终点。证据artifacts/recovery_hourly_20261111T0100_analysis.json,非FID,无Test1024。 ++ ++已启动共享根/root/autodl-fs/pangxinlei/models/hfd-recovery-v45-20261111:HY入口scripts/recovery_v45_hymotion/supervise_pilot.py,从原正式best重置进行250步有界实验,相对v36仅training.text_dropout由0.1改0;其余训练参数、执行权重6、因果预测历史及原门槛不变。依据扩展行走普遍不足、此前空文本消融损害行走,检验训练时完整保留文本条件的假设,尚不能归因或宣布改善。Kimodo入口scripts/recovery_v45_kimodo/supervise_pilot.py,冻结v44step125,固定24validation×2种子,沿用原正式48例基线,验证原6例子集精确复现及扩大覆盖后的行走/抖动表现。配置单变量、48例基线和语法检查通过,recovery_v45_checks.json。 ++ ++启动前GPU0/1均空闲;HY监督器590055/子进程590059、Kimodo监督器590058/子进程590060存活,无错误,处于初始化(HY恢复检查尚未完成,不宣称已通过)。各2GiB预算实写fsync预留通过,HY8小时/Kimodo4小时上限及原容量/inode/日志/原子断点保护保持,健康及源码SHA见recovery_v45_health.json和launch收据。清理已结束HYv41、Kimodov44重复启动索引共656.24MiB,确认进程退出、无打开句柄、新副本SHA相同;全部权重/恢复断点/诊断轨迹保留,清单recovery_cleanup_hourly_20261111T0100.json。本地约1.529GiB,共享空闲inode36348,实写fsync检查通过,目录项低于500,共享df不作账号配额。尚无正式训练,保留半小时调度,下一检查点01:30 UTC;任一正式训练健康启动立即取消416e9f54条件不变。 ++ ++ ++### HY v45继续、Kimodo位移监督v46(服务器2026-11-11 01:30 UTC) ++ ++HY v45真实训练590274/监督器590055存活,检查55步、轨迹审计时57/250,loss/梯度有限;原初始轨迹逐元素复现v36初始,2步完整恢复读回通过(optimizer/scheduler为2,136884553字节断点)。尚未到125验证,保留正在运行配方,不重复启动。Kimodo v45扩展验证完成,无错误、预留释放;48条完整轨迹有限、参考/case_keys与原扩展基线一致,原6例精确复现v44step125。幅度0.938、接段2.249、姿态0.426rad、根位置0.561m、jerk900.12,行走失败9/24,较v19的8/24及根位置0.528m未改善。限制根训练范围未解决泛化不足,不晋级正式训练。证据artifacts/recovery_hourly_20261111T0130_analysis.json,非FID,无Test1024。 ++ ++Kimodo实际启动v46:/root/autodl-fs/pangxinlei/models/hfd-recovery-pilot-v46-20261111/kimodo,入口scripts/recovery_training_v46/supervise_pilot.py。以v19原正式best初始化配方为对照,仅recovery.temporal_weights.stage1_root_displacement由1增至2,其余训练/损失/学习率/根全量训练/因果输入/推理不变。检验更强已有位移目标能否减少扩展行走不足,不预设有效,仍需防止抖动与误差退化。既有损失实现不变,只使用训练目标和有效帧掩码;单变量与原基线保持检查recovery_v46_checks.json。250步有界、固定0/125/250两种子验证、原门槛不变,不从退化断点延长。 ++ ++GPU1空闲后启动监督器590548和真实子进程,健康收据recovery_v46_health.json、源码SHA见launch收据;处于恢复检查初始化,不宣称新轮恢复或效果通过。12GiB预算实际写入fsync预留通过,8小时墙钟、容量/inode/目录项/日志/原子断点保护保持;HY GPU0继续。清理已结束Kimodo v45重复索引328.12MiB,确认旧进程退出、无打开句柄和新副本SHA一致;所有模型权重、完整断点及轨迹保留,清单recovery_cleanup_hourly_20261111T0130.json。本地约1.529GiB、共享空闲inode超过36000,本地/共享实写fsync通过,目录项低于500,共享df不作为账号配额。无正式训练,半小时调度继续,下次02:00 UTC;正式训练健康启动即取消416e9f54规则不变。 ++ ++ ++### 双模型中期健康检查(服务器2026-11-11 02:00 UTC) ++ ++HY v45真实进程590274/监督器590055存活,检查131步、审计135/250;Kimodo v46真实进程590704/监督器590548存活,检查196步、审计200/250。双方训练loss/梯度有限,初始/125完整轨迹有限、参考及案例匹配,原初始逐元素复现;两步恢复读回通过,optimizer/scheduler均为2,HY完整断点136884553字节、Kimodo3394249205字节。无错误或watchdog停止收据,GPU0/1分别约20.1/20.3GiB,监督器继续保护。 ++ ++HY125幅度0.644、接段1.729、姿态0.409rad、根位置0.504m(比原初始约恶化12.3%)、jerk383.25;2327两种子进度1.644/2.514,961为-0.267/-0.129,3/4行走组合失败,尚未合格。Kimodo125幅度1.194、接段2.224、姿态0.452rad、根位置0.424m、jerk766.11;2327进度0.750/0.876,961为1.434/0.524,现有小验证门槛通过,最终及扩展验证待完成,不能据此正式开训。证据artifacts/recovery_hourly_20261111T0200_analysis.json,非FID,无Test1024。保持原250步有界预算,不重复启动或修改运行源码。 ++ ++本地约1.529GiB、共享空闲inode36269,本地/共享1MiB实写fsync通过,共享df不作账号配额;两运行目录项低于500,预留及旧断点保护正常。没有新核验可删项,删除0字节,保留活动预留、缓存、权重及轨迹,清单recovery_cleanup_hourly_20261111T0200.json。无正式训练,半小时调度保持,下次02:30 UTC检查最终结果;正式训练健康启动即取消416e9f54规则不变。 ++ ++ ++### Kimodo v46完成及扩展v47(服务器2026-11-11 02:30 UTC) ++ ++HY v45进程590274/监督器590055存活,最新日志249/250,尚无完成收据或最终验证结果;训练数值及已保存初始/125完整轨迹有限,原初始复现和参考匹配通过,恢复检查通过。保留运行中预留及源码,未重复启动。Kimodo v46正常完成250步,预留释放、无错误;初始/125/250完整轨迹有限、参考/案例匹配、原初始精确复现,两步完整恢复通过。最终幅度1.219、接段2.329、姿态0.451rad、根位置0.437m、jerk478.29;2327进度0.864/0.585,961进度0.242/0.237,最终不合格。125小验证通过,但这不等于扩大覆盖合格。证据artifacts/recovery_hourly_20261111T0230_analysis.json,非FID,无Test1024。 ++ ++已在空闲GPU1启动冻结扩展v47:/root/autodl-fs/pangxinlei/models/hfd-recovery-broad-v47-20261111/candidate125,入口scripts/recovery_broad_v47/supervise_pilot.py。使用v46的step125候选,同24validation案例×两种子及原正式48例基线,20步/guidance2,零训练更新,保持原门槛。需核对完整参考及原6例精确复现后比较v19等扩展对照,不延长失败终点。2GiB预算实写fsync预留通过;监督器591195和真实子进程存活,初始化中,无错误,4小时墙钟/容量inode/日志缓存/原子保存保护保持,recovery_v47_checks.json、recovery_v47_health.json和launch收据。 ++ ++清理已结束Kimodo v46重复启动索引328.12MiB,旧进程退出、无打开句柄、新副本SHA相同;保留所有模型权重、完整恢复断点及轨迹,清单recovery_cleanup_hourly_20261111T0230.json。本地约1.529GiB、共享空闲inode超过36000,本地/共享实写fsync通过,目录项低于500,共享df不作账号配额。仍无正式训练,半小时调度保持,下次03:00 UTC检查HY最终结果和Kimodo扩展结果;正式训练健康启动即取消416e9f54规则不变。 ++ ++ ++### 完成分析及历史通道诊断v48(服务器2026-11-11 03:00 UTC) ++ ++HY v45正常完成250步,恢复检查通过、预留释放;训练及0/125/250完整轨迹有限、原初始精确复现和参考配对通过。最终幅度0.700、接段1.758、姿态0.406rad、根位置0.522m(较原初始恶化约16.2%)、jerk366.42;2327进度1.740/2.607,961为-0.326/-0.119,3/4行走组合失败,取消文本dropout未解决行为问题,不延长。Kimodo v47扩展完成、预留释放,无错误;48完整轨迹有限、原扩展参考匹配、原6例精确复现v46step125。幅度0.971、接段2.228、姿态0.421rad、根位置0.581m、jerk961.83,无低于0.1坍缩,但行走失败10/24,比v19的8/24、根位置0.528m及jerk906.78更差,不采用。证据artifacts/recovery_hourly_20261111T0300_analysis.json,非FID,无Test1024。 ++ ++两边实际启动冻结诊断v48,共享根/root/autodl-fs/pangxinlei/models/hfd-recovery-history-v48-20261111/{hymotion,kimodo},入口scripts/recovery_history_v48_{hymotion,kimodo}/supervise_pilot.py。HY固定v36最终候选、原6validation×两种子;Kimodo固定v19step125候选、扩展24validation×两种子。各含自主预测历史对照、只替换过去根平移、只替换过去旋转三臂。根平移是原始轨迹0:3通道,其余旋转3:;历史索引仅0..decision,未来占位仍由末历史帧构造。只作反馈误差定位,真实过去臂不可部署,位置重置造成的进度收益不算自主行走,所有oracle臂明确禁止用于正式晋级;自主对照需精确复现既有结果后才解释差异。不更新训练权重、不导出新大模型,保留原门槛。检查recovery_v48_checks.json,源码SHA见launch收据。 ++ ++GPU0/1启动前空闲;HY监督器591617/子进程591619、Kimodo监督器591621/子进程591622存活,初始化中,无错误,健康recovery_v48_health.json。各2GiB预算实写fsync预留通过,4小时墙钟/容量inode/目录项/日志缓存/原子保存保护保持。清理已结束HYv45、Kimodov47重复索引共656.24MiB,确认旧进程退出、无打开句柄、新副本SHA相同;全部模型权重、完整断点和轨迹保留,清单recovery_cleanup_hourly_20261111T0300.json。本地约1.528GiB、共享inode36185,本地/共享实写fsync通过,目录项低于500,共享df不作账号配额。尚无正式训练,半小时调度保持,下次03:30 UTC检查诊断对照复现与进度;正式训练健康启动即取消416e9f54条件不变。 ++ ++ ++### HY历史诊断细分v49、Kimodo v48继续(服务器2026-11-11 03:30 UTC) ++ ++HY v48三臂完成,无错误且预留释放,完整轨迹有限、同参考匹配,自主预测历史逐元素精确复现v36最终。只替换过去平移后根误差0.055m、行走进度接近1,但jerk2521.70,属于位置重置收益,不是自主行走改善;身体幅度0.607基本不变。只替换过去全部旋转后幅度0.832、姿态误差0.183rad、根位置0.279m,961两种子进度0.507/0.717;接段比10.516、jerk476.27,不可部署。表明旋转历史可能参与行走失败,但尚未区分根朝向和身体姿态,不能把oracle收益作为资格。Kimodo v48监督器591621/子进程591622存活,GPU1约17GiB、计算活动正常,尚在后续诊断;自主对照48条完整轨迹有限并逐元素精确复现v20,参考匹配,原有8/24行走失败不变。证据artifacts/recovery_hourly_20261111T0330_analysis.json,非FID,无Test1024。 ++ ++HY启动冻结v49:/root/autodl-fs/pangxinlei/models/hfd-recovery-history-v49-20261111/hymotion,入口scripts/recovery_history_v49_hymotion/supervise_pilot.py。相同v36最终候选、6validation×两种子,三臂为自主对照、仅替换过去根朝向3:6、仅替换过去身体关节旋转6:69;已核对原始3:69为22个轴角关节,保持根平移预测不重置,历史索引仅0..decision,不读未来GT。零训练更新,oracle臂禁止正式晋级;先要求自主对照精确复现,再据通道差异设计恢复实验。检查recovery_v49_checks.json。 ++ ++GPU0启动前空闲,监督器591942及真实子进程存活,无错误、初始化中;2GiB预算实写fsync预留通过,4小时墙钟/容量inode/日志缓存/原子保存保护保持,健康recovery_v49_health.json和启动SHA收据。Kimodo原运行继续、不重复启动。清理已结束HYv48重复索引328.12MiB,进程退出、无打开句柄、新副本SHA相同;所有权重/断点/轨迹保留,清单recovery_cleanup_hourly_20261111T0330.json。本地约1.528GiB、共享inode超过36000,实写fsync通过、目录项低于500,共享df不作账号配额。尚无正式训练,半小时调度继续,下次04:00 UTC;正式训练健康启动即取消416e9f54条件保持。 ++ ++ ++### 历史诊断完成及双模型混合历史v50(服务器2026-11-11 04:00 UTC) ++ ++HY v49和Kimodo v48均正常完成、预留释放;三臂完整轨迹有限、参考配对通过,自主对照逐元素精确复现HYv36最终及Kimodov20扩展。HY仅替换过去根朝向时961进度0.209/0.146仍不足,2327一seed超2;仅替换身体旋转时961仍反向-0.256/-0.108、接段10.516。与此前全部旋转替换相比,单通道修正不足,不能简单归因为某一独立通道。Kimodo替换平移使进度接近1但jerk5448.41,属于位置重置;替换旋转后姿态0.183rad、根位置0.383m,但接段6.417、jerk1256.41、行走失败7/24(自主对照8/24),仍不可部署。oracle臂不作为正式晋级依据。证据artifacts/recovery_hourly_20261111T0400_analysis.json,非FID,无Test1024。 ++ ++已启动双模型v50:/root/autodl-fs/pangxinlei/models/hfd-recovery-pilot-v50-20261111/{hymotion,kimodo},入口scripts/recovery_training_v50_{hymotion,kimodo}/supervise_pilot.py。分别以HYv36、Kimodov19为对照,从原正式best重置,仅pilot.predicted_probability从1降至0.75,使训练含25%真实历史窗口;其余训练参数、损失、学习率、链长、文本和0/125/250固定两种子验证均不变。诊断提示历史与姿态误差耦合,但并未证明混合历史训练有效,此轮检验是否有助于改善反馈误差。推理仍完全自主、无真实过去替换,无未来GT,无Test1024,原门槛保持;不延长失败断点。配置检查recovery_v50_checks.json。 ++ ++启动前GPU0/1空闲;两监督器及真实子进程存活、恢复检查初始化中,无错误,不宣称本轮恢复或效果已通过。HY2GiB、Kimodo12GiB预算实写fsync预留完成,8小时墙钟及容量/inode/日志缓存/原子断点保护保持,进程及源码SHA见recovery_v50_health.json和launch收据。清理已结束HYv49、Kimodov48重复索引共656.24MiB,旧进程退出、无打开句柄、新副本SHA一致后删除;正式权重、候选、完整断点及轨迹全部保留,清单recovery_cleanup_hourly_20261111T0400.json。本地约1.528GiB、共享inode超过36000,实写fsync检查通过、目录项低于500,共享df不作账号配额。无正式训练,半小时调度保持,下次04:30 UTC检查恢复和训练进度;正式训练健康启动即取消416e9f54条件不变。 ++ ++ ++### 双模型v50运行检查(服务器2026-11-11 04:30 UTC) ++ ++HY真实训练592427/监督器592244存活,审计98/250,尚未完成125验证;Kimodo真实训练592540/监督器592293存活,检查200步、审计205/250。双方两步完整断点恢复读回通过,optimizer/scheduler均为2,HY断点136888521字节、Kimodo3394253045字节;训练loss/梯度和已保存完整轨迹有限,初始逐元素复现原对照、参考/案例匹配。Kimodo日志已包含预测与真实历史链,混合训练路径生效;推理保持完全自主。无错误或watchdog停止,GPU0/1分别约20.1/20.3GiB,监督器正常。 ++ ++Kimodo125幅度0.764、接段2.262、姿态0.413rad、根位置0.520m(比原初始恶化约17.8%)、jerk590.96;2327两种子进度0.907/0.720,961为0.806/0.143,一组明显位移失败,未过门槛,不能把jerk下降视为已恢复质量。等待原250步有界预算的最终验证,不重复启动、不改动运行源码。证据artifacts/recovery_hourly_20261111T0430_analysis.json,非FID,无Test1024。 ++ ++本地约1.527GiB、共享空闲inode36041,本地/共享1MiB实写fsync通过,目录项低于500,共享df不作账号配额。活动预留、完整断点、缓存及轨迹保留,本轮无新核验可清理项,删除0字节,清单recovery_cleanup_hourly_20261111T0430.json。尚无正式训练,半小时调度继续,下次05:00 UTC检查中期/最终结果;正式训练健康启动即取消416e9f54规则不变。 ++ ++ ++### Kimodo v50完成与细分历史诊断v51(服务器2026-11-11 05:00 UTC) ++ ++HY v50进程592427/监督器592244存活,最新日志249/250、无错误、最终验证待返回。125幅度0.447低于0.5,姿态误差0.448rad较原初始恶化约20.3%,961进度-0.045/-0.001,尚不合格,不重复启动或覆盖运行源码。Kimodo v50正常完成250步、恢复通过、预留释放;双方训练及已保存初始/中期/最终完整轨迹有限,原初始精确复现、参考/案例匹配。Kimodo最终幅度0.658、接段2.562、姿态0.427rad、根位置0.409m、jerk372.26;2327进度0.373/1.001,961为0.771/0.221,2/4行走组合失败,混合历史未过门槛,不能以低jerk当恢复成功。证据artifacts/recovery_hourly_20261111T0500_analysis.json,非FID,无Test1024。 ++ ++Kimodo已启动冻结v51:/root/autodl-fs/pangxinlei/models/hfd-recovery-history-v51-20261111/kimodo,入口scripts/recovery_history_v51_kimodo/supervise_pilot.py。沿用v19step125、相同24validation×两种子及原正式48例基线,将v48全部旋转历史诊断细分为自主对照、仅替换过去根朝向3:6、仅替换过去身体旋转6:69;根平移保持预测,历史索引0..decision,未来占位不取GT。用于区分反馈误差来源,零训练更新,oracle臂不可部署且禁止正式晋级,自主对照必须精确复现。检查recovery_v51_checks.json。 ++ ++GPU1空闲后启动监督器593012及真实子进程,存活、初始化中、无错误;2GiB预算实写fsync预留通过,4小时墙钟/容量inode/日志缓存/原子保存保护保持,健康recovery_v51_health.json及launch源码SHA。HY GPU0继续。清理已结束Kimodo v50重复索引328.12MiB,旧进程退出、无打开句柄、新副本SHA一致后删除,所有模型权重、完整断点及轨迹保留,清单recovery_cleanup_hourly_20261111T0500.json。本地约1.527GiB、共享inode超过36000,本地/共享实写fsync通过、目录项低于500,共享df不作账号配额。尚无正式训练,半小时调度继续,下次05:30 UTC;正式训练健康启动即取消416e9f54规则不变。 ++ ++ ++### HY v50完成及标签过渡v52(服务器2026-11-11 05:30 UTC) ++ ++HY v50正常完成250步、无错误、预留释放、恢复通过;完整初始/125/250轨迹有限、参考配对、原初始精确复现。最终幅度0.557、接段1.887、姿态0.429rad(原初始+15.2%)、根位置0.450m、jerk358.13;2327进度1.316/1.838,961为-0.062/-0.006,姿态及行走失败,不延长混合历史配方。Kimodo v51真实593013/监督器593012存活,GPU1约17GiB并有计算活动,无错误;自主对照48条完整轨迹精确复现v20,参考一致且有限,后续诊断尚未完成,不预判oracle收益。证据artifacts/recovery_hourly_20261111T0530_analysis.json,非FID,无Test1024。 ++ ++HY启动v52有界250步:/root/autodl-fs/pangxinlei/models/hfd-recovery-pilot-v52-20261111/hymotion,入口scripts/recovery_training_v52/supervise_pilot.py。回到v36原正式best初始化对照,唯一语义变量标签SO3/高度过渡6帧缩至3帧,pilot和curriculum字段同步;其他训练参数、全预测历史、文本、损失、执行12帧及原验证保持。依据旋转历史误差耦合及混合历史无效,检验更快纠正监督姿态是否有效;存在接段恶化风险,原门槛不变,不预设改善。已有过渡实现不改,测试第三个未来帧起精确对齐目标、XY保持、有效掩码保持、输入不修改、真实历史恒等及数值有限通过,recovery_v52_checks.json。只修改训练标签,不增加未来GT输入,不改变自主推理,无Test1024。 ++ ++GPU0启动前空闲,监督器593503及真实子进程存活、恢复检查初始化中、无错误;2GiB预算实写fsync预留通过,8小时墙钟及容量/inode/日志缓存/原子断点保护保持,健康recovery_v52_health.json及launch源码SHA。不宣称新轮恢复通过。Kimodo原诊断继续。清理已结束HYv50重复索引328.12MiB,旧进程退出、无打开句柄和新副本SHA一致后删除,权重/断点/轨迹保留,清单recovery_cleanup_hourly_20261111T0530.json。本地约1.527GiB、共享inode超过35000,实写fsync通过、目录项低于500,共享df不作账号配额。无正式训练,半小时调度继续,下次06:00 UTC;正式训练健康启动立即取消416e9f54规则不变。 ++ ++ ++### HY v52继续及Kimodo标签过渡v53(服务器2026-11-11 06:00 UTC) ++ ++HY v52真实训练593626/监督器593503存活,124/250步、中期验证待返回;两步完整恢复读回通过,optimizer/scheduler均为2,断点136884553字节。训练数值及已存完整轨迹有限,原初始精确复现、参考配对,无错误。不重复启动或覆盖活动源码。Kimodo v51完成三臂、预留释放;48例完整轨迹有限、参考匹配、自主对照精确复现v20。自主行走失败8/24;仅替换过去根朝向失败10/24,幅度1.115、接段2.343、姿态0.387rad、根位置0.617m、jerk1000.12;仅替换身体旋转失败11/24,幅度1.251、接段6.582、姿态0.205rad、根位置0.549m、jerk1635.70。单通道oracle修正没有恢复自主行走,oracle不可晋级。证据artifacts/recovery_hourly_20261111T0600_analysis.json,非FID,无Test1024。 ++ ++Kimodo已启动v53:/root/autodl-fs/pangxinlei/models/hfd-recovery-pilot-v53-20261111/kimodo,入口scripts/recovery_training_v53/supervise_pilot.py。从原正式best初始化,以v19为对照,仅pilot.blend_frames从6改3,检验更快的标签姿态/高度纠正;此前12帧过渡和混合历史均未解决行走失败,因此不重复旧配方。250步有界、0/125/250固定两种子验证;训练、损失、学习率、推理、原门槛均不变。只改变标签不增加未来GT输入,存在接段恶化风险,不预设改善。准备时已移除误带的HY专属curriculum字段,未启动任何错误配置训练。三帧目标数值测试及单变量合同检查通过,recovery_v53_checks.json。 ++ ++GPU1空闲后监督器启动,真实子进程初始化中;12GiB预算实写fsync预留通过,8小时墙钟、磁盘/inode、日志缓存及原子断点保护保持。新轮恢复结果尚待返回,不宣称通过。健康及源码SHA见recovery_v53_health.json和launch收据。HY GPU0继续运行。仅清理已结束Kimodo v51重复索引328.12MiB,确认旧进程退出、无打开句柄、新副本SHA一致,权重/断点/完整轨迹保留,清单recovery_cleanup_hourly_20261111T0600.json。本地约1.526GiB,共享inode超过35000,实写fsync通过,目录项均低于500;共享df不作账号配额。无正式训练,半小时调度继续,下次06:30 UTC;任一正式训练健康启动即取消416e9f54的规则保持。 ++ ++ ++### 双模型标签过渡实验运行检查(服务器2026-11-11 06:30 UTC) ++ ++HY v52真实训练593626/监督器593503存活,审计190/250;Kimodo v53真实训练594424/监督器594263存活,125/250且中期验证尚待返回。双方两步完整断点恢复读回通过,optimizer/scheduler均为2,HY断点136884553字节、Kimodo3394249141字节。训练日志数值与已保存完整轨迹有限、参考和案例匹配、原初始精确复现,运行源码SHA与启动收据一致,无错误/监督器停止。GPU0/1占用约20.1/20.3GiB。HY日志full_predicted_history_fraction统计历史60帧是否全由模型生成,链启动阶段可低于1,不代表配置概率变为0.75;pilot.predicted_probability仍为1。 ++ ++HY125幅度0.538、接段2.124、姿态0.419675rad(原初始+12.7%)、根位置0.440253m、jerk368.13;2327两种子进度1.159/1.348,961为-0.031/0.001,姿态与行走不合格。Kimodo尚无中期结果,不能以初始结果判断本轮改善。双方继续已授权250步有界预算,不重复启动、不延长失败配方、不覆盖运行源码;下一次分析各自完成结果。证据artifacts/recovery_hourly_20261111T0630_analysis.json,非FID,无Test1024。 ++ ++本地约1.526GiB、共享空闲inode35910,本地及共享1MiB实写fsync检查通过,目录项均低于500,共享df不作账号配额。活动预留、缓存、完整断点及轨迹保留,无新增核验可删除项,本轮清理0字节,清单recovery_cleanup_hourly_20261111T0630.json。尚无正式训练,半小时调度保持,下次07:00 UTC;正式训练健康启动立即取消416e9f54的条件不变。 ++ ++ ++### 标签过渡实验完成及双模型v54(服务器2026-11-11 07:00 UTC) ++ ++HYv52与Kimodov53正常完成250步、旧进程退出、预留释放;恢复检查均通过,初始/125/250完整轨迹有限、参考配对、原初始精确复现。HY最终幅度0.609、接段2.452、姿态0.413428rad(原初始+11.1%)、根位置0.460978m、jerk363.53;2327进度1.552/1.887,961为-0.076/-0.070,仍不合格;较v36接段1.780恶化而姿态/行走未改善。Kimodo125幅度1.151、接段2.387、姿态0.44648rad、根位置0.43974m、jerk816.99,961一seed进度0.471低于0.5,不能放宽门槛;最终幅度1.210、接段2.652、姿态0.44417rad、根位置0.39969m、jerk514.15,2327进度0.814/0.455,961为0.794/0.195,两组失败。证据recovery_hourly_20261111T0700_analysis.json包括每例段内突变、根速度/jerk、坍缩及完整轨迹核对,非FID,无Test1024。 ++ ++启动双模型v54有界250步,根目录/root/autodl-fs/pangxinlei/models/hfd-recovery-pilot-v54-20261111,入口scripts/recovery_training_v54_{hymotion,kimodo}/supervise_pilot.py。HY回到v36原正式best初始化,仅旋转速度辅助权重0改0.05,原暖启动200步保持,检验姿态动态监督;不同于早期带根速度辅助/不同历史配方,不宣称该损失首次尝试。Kimodo回到v19原正式best,仅身体学习率曲线减半(1e-6/2e-7变5e-7/1e-7),根学习率2e-7/5e-8保持,检验身体更新过快与中后期退化的关系;区别于v30同时减半双模块。两者均非旧断点续训。其余配置、标签6帧、因果输入、推理、固定两种子验证及原验收保持;假设尚待验证。逐字段单变量比较及HY辅助损失有限/非零梯度检查通过,recovery_v54_checks.json。 ++ ++HY2GiB/Kimodo12GiB预算实写fsync预留通过后,在空闲GPU0/1分别启动监督器与真实子进程;新轮恢复检查初始化中,不宣称已通过。8小时墙钟、容量/inode、日志缓存、原子断点保护保持,健康和源码SHA见recovery_v54_health.json及launch收据。清理已结束HYv52/Kimodov53重复索引共656.24MiB,确认旧进程退出、无打开句柄、新副本SHA一致,权重/完整恢复断点/轨迹全部保留,清单recovery_cleanup_hourly_20261111T0700.json。本地约1.526GiB、共享inode超过35000,实写fsync检查通过,目录项低于500,共享df不作为账号配额。无正式训练,半小时调度继续,下次07:30 UTC;任一正式训练健康启动立即取消416e9f54规则不变。 ++ ++ ++### 双模型v54运行及Kimodo中期候选(服务器2026-11-11 07:30 UTC) ++ ++HY真实训练595667/监督器595433存活,审计90/250;Kimodo真实训练595769/监督器595479存活,检查184步、审计188/250。双方两步完整恢复读回通过,optimizer/scheduler均为2,HY断点136884553字节、Kimodo3394249205字节;训练数值有限,已保存完整轨迹有限、参考/案例匹配、原初始精确复现。源码SHA与启动收据一致,无错误/监督停止,GPU0/1占用约20.1/20.3GiB,日志持续更新。HY尚无125中期结果,不据初始结果推断本轮效果。 ++ ++Kimodo125在当前6案例×两种子上通过原门槛:幅度0.791598、接段2.058991、姿态0.423648rad、根位置0.314195m、jerk639.63;2327进度0.696/0.787,961为1.148/0.685,明显位移四组均在0.5–2。jerk仍高于原初始357.30,不能把小样本门槛通过等同于动作质量完全恢复。保留125候选,先完成当前250步有界预算;GPU1释放后优先对125进行相同24案例×两种子的扩展验证,与原正式48例基线及v19候选比较。未完成扩展验证前不启动正式训练,不重复占用GPU,不延长预算。证据recovery_hourly_20261111T0730_analysis.json包含完整轨迹核对及每例幅度、段内突变、速度、jerk、进度;非FID,无Test1024。 ++ ++本地约1.526GiB、共享空闲inode35821,本地及共享1MiB实写fsync检查通过、目录项低于500,共享df不作账号配额。活动预留、缓存、完整断点及候选保留,无新增核验可删除项,本轮清理0字节,清单recovery_cleanup_hourly_20261111T0730.json。尚无正式训练,半小时调度保持,下次08:00 UTC;正式训练健康启动立即取消416e9f54条件不变。 ++ ++ ++### 用户结束定时检查(2026-11-11T07:52:28.163857+00:00) ++ ++按用户明确指令,已执行 paseo heartbeat delete 416e9f54,CLI返回deleted,并核实调度记录不存在。任务状态已更新为cancelled_by_user,不再自动检查,不自动重建。此次操作仅取消心跳调度,不终止已授权训练,也不关闭监督器的容量/inode、断点保护及资源清理。 ++ ++ ++## 2026-11-11 三模型训练交付 ++ ++原训练工作区、代码差异、正式 latest/best、恢复候选、数据索引和特征缓存、环境锁、历史日志及学习曲线已上传并核验到以下私有仓库: ++ ++- [genmo](https://huggingface.co/starstone3/HFD-GENMO-SMPL-Training),commit `979e5130d6d2289582e8171830700d49dd0cdaf2`,清单 94 个文件。 ++- [hymotion](https://huggingface.co/starstone3/HFD-HY-Motion-SMPL-Training),commit `f858d43583e5897f850394fef886e7860f3a84f9`,清单 97 个文件。 ++- [kimodo](https://huggingface.co/starstone3/HFD-Kimodo-SMPLX-Training),commit `c9a71c47e7e16e9700779e986a607da6d920669e`,清单 107 个文件。 ++ ++三个模型均已实际短程续训并保存完整更新断点:HY 14000→14002(loss/LR/样本与原记录一致);Kimodo 98000→98002(loss/LR一致);GENMO 98000→98002(LR一致,loss略有差异,不能称为逐步精确一致)。GENMO 保存时触发容量受控停止,Kimodo 验收监督器异常和 HY 路径身份修复均已保留证据。原始数据和部分外部基础/文本/许可资产以固定来源与 SHA 清单交付,未声称已做全新环境重建。三个正式最新断点均到100K原预算,后续新阶段调度迁移尚未验证。 ++ ++Kimodo 原 best 仅权重;另交付与 best 参数逐位相同的同一步完整 last 对应文件。每仓库 README/acceptance-report/MANIFEST 记录具体命令、缺口和验证范围。心跳416e9f54已删除,不重建;既有训练未被终止。 ++ ++本地交付证据:`/root/autodl-fs/pangxinlei/handoffs/three-human-models-20261111/FINAL_DELIVERY.json`。核验时间:2026-11-11T08:35:41.057469+00:00。 ++ ++HY pilot54 已补交 step200 的最新完整断点与明确截止时间的日志快照;活动训练后续变化不包含在快照中。最终 MANIFEST 已从固定远端提交下载回读核验。最终核验时间:2026-11-11T08:39:04.787213+00:00。 ++ +diff --git a/SESSION_FORMAT.md b/SESSION_FORMAT.md +deleted file mode 100644 +index 03f5053..0000000 +--- a/SESSION_FORMAT.md ++++ /dev/null +@@ -1,82 +0,0 @@ +-# HFD session 格式与当前输入边界 +- +-更新:2026-10-21。正式默认为英文前景、`aligned_reference_context`、 +-`clock_policy: microphone`、`text_conditioning: none`:离线原文词对齐、最近30秒 +-已释放文本,不运行ASR或翻译。五模型的具体输入/输出、帧率与实现状态见 +-[当前协议](docs/HFD_HUMAN_BASELINE_INPUT_PROTOCOL.md),不再沿用旧中文翻译/ASR默认说明。 +- +-portable不携带原始tar、模型权重、缓存、视频或训练产物。HFD原始数据在 +-`/root/autodl-fs/pangxinlei/hfd_data`,按索引读取tar,保留分片,不全量解压。 +-原生训练session按配方动态组装并缓存在进程内;不是必须提前导出所有NPZ。 +-第三方读取已经组装的生产session。人体源动作也需要按计划组装,不能直接当完整session。 +- +-## 导出格式 +- +-`prepare_sessions.py`调用原生FullDuplexAudioCommandDataset;导出用于固定评测和短测, +-不改变任务、背景、监听/保持/过渡或事件时间。命令示例: +- +-```bash +-python prepare_sessions.py --config configs/sessions.yaml +-python prepare_sessions.py --config configs/sessions.yaml --set split=val +-python validate_sessions.py /data/hfd-sessions/train +-``` +- +-每个G1 session为NPZ及同名JSON: +- +-| 字段 | 内容 | +-| --- | --- | +-| `motion_q29_xyz_quat` | `[T,36]`:IsaacLab顺序29关节+根xyz+wxyz四元数;50Hz | +-| `audio` | 16kHz单声道混合PCM,保留原前景语音与背景 | +-| `control_labels` | 0=listen、1=motion、2=idle;监督/评测元数据,不作为输入 | +-| `turn_index_at_decision` | turn元数据,仅审计/展示,不作为输入 | +-| `valid_start_count` | 原生有效决策起点数;各模型再检查自己的窗口覆盖 | +-| `.json` | 源动作/音频、实际事件位置、关节/四元数约定、split、哈希 | +- +-`full_session`包含listen、motion、idle、过渡及事件前后的合法训练窗口。 +-`motion_only`仅保留作历史消融。源片段在session内部按配方组合,但训练窗口不跨session。 +-候选session数不是独立原始动作数,也不是已经缓存到磁盘的NPZ数量。 +- +-## 当前文本sidecar +- +-`data.asr_dir`是历史字段名;在当前协议下实际读取 +-`.reference.json`,不是`.asr.json`。动态训练则从同一离线对齐库和 +-当次session音频组件偏移直接生成相同记录,不需要在线ASR。 +- +-格式 `hfd-aligned-reference-v1` 包含NPZ/元数据哈希、源TTS片段、估计词末时间、 +-可用时间、首次释放时间、30秒上下文记录及异常fallback。已释放的完整句子可以输入; +-尚未释放的句尾、最终修正指令或整条GT全文不能提前提供。 +- +-动作时刻t对应麦克风t+0.2秒。词末加0.1秒guard,向上取整到麦克风 +-0.6/1.0/1.4…秒释放,模型只选择 `r_k <= t+0.2` 的记录。 +-0.6秒音频窗每0.4秒更新,是重叠窗口,不是每步新增0.6秒。 +-文本按释放点最近30秒滚动,过期可以为空;空文本不无限沿用旧命令。 +-无效词跨度在整个对应音频组件结束后释放。对齐误差和guard未完成逐词人工认证。 +- +-这是正确前景说话人的oracle转写实验:背景PCM保留,背景文字不输入。 +-旧 `causal_asr_context`、`stable_latest_sentence`、Whisper与中文→英文翻译仅用于 +-显式历史复现,不与当前结果混用。现有原始双语数据和历史产物不改写。 +- +-## 动作与坐标 +- +-原始NPZ不是MuJoCo/native关节顺序,不能直接切片当qpos。 +-`hfd_bridge/joints.py`按名称转换:TextOp原生23关节/57D,OMG原生29关节/125D。 +-当前TextOp未建模的腕关节固定在bootstrap姿态;指标只对其原生23关节计分。 +-根四元数在原始HFD文件为wxyz,进入各框架时显式重排。 +- +-GenMo读取配对人体trans、root_orient、pose_body、betas和sample_times, +-训练与推理不使用G1关节作为网络输入或目标。当前人体组装仍使用原生G1配方/ +-配对映射追踪源与时序;不能说整个数据构造程序与G1文件完全无关。 +-人体输入与retarget契约见[人体协议](docs/HFD_HUMAN_BASELINE_INPUT_PROTOCOL.md) +-及[评测转换说明](docs/HUMAN_RETARGET_EVALUATION.md)。 +- +-## 训练与连续评测 +- +-各模型使用自己的原生表示、归一化、损失和框架。OMG与GenMo训练提供GT历史, +-TextOp沿用原生分阶段历史课程。连续评测只初始化一次GT状态,之后反馈预测; +-真值历史诊断必须明确标注。GT事件仅用于选择/标注片段,不作为模型条件。 +- +-评测保留原混合音轨,将共同0.2秒microphone前置帧和bootstrap从动作评分中排除, +-不做DTW或事后根对齐。OMG/TextOp原生G1输出不做人体retarget;人体模型对完整轨迹 +-做一次retarget,并同时报告人体本身误差及人体GT重定向对照。 +-当前独立test未建立;val用于训练监测,不能称完全未见测试集。视频为运动学回放, +-不声称物理跟踪、端到端实时语音系统或语义任务成功率。 +diff --git a/UPSTREAM_LOCK.json b/UPSTREAM_LOCK.json +index 6c9873d..f30a85f 100644 +--- a/UPSTREAM_LOCK.json ++++ b/UPSTREAM_LOCK.json +@@ -2,11 +2,17 @@ + "external_repositories": { + "genmo": { + "url": "https://github.com/NVlabs/GENMO.git", +- "commit": "16bebf402d8893184249ee206d957b8248cd8310" ++ "commit": "16bebf402d8893184249ee206d957b8248cd8310", ++ "local_packaging_patch": "patches/genmo-numpy1264.patch", ++ "replay_numpy": "1.26.4" + }, + "kimodo": { + "url": "https://github.com/nv-tlabs/kimodo.git", + "commit": "e1c219c99b4b94ba48be1fb6a1cbc64ecee4c954" ++ }, ++ "hymotion": { ++ "url": "https://github.com/Tencent-Hunyuan/HY-Motion-1.0.git", ++ "commit": "4e426f5a1021cbcf7f375458c37b840ee7225229" + } + }, + "genmo_helper_assets": { +@@ -118,5 +124,16 @@ + "revision": "93ba4a7a9c66b3e1b7041e5f4ea038ef2418b948", + "metadata_manifest": "TRAINING_ASSETS.json", + "note": "Includes paired SMPL-X and all prior shards; do not bootstrap using the older pre-SMPL-X revision." ++ }, ++ "hfd_final_release": { ++ "data_repo": "Tinghong-Ye22/data_vqa_all_1025", ++ "data_revision": "bdca2fe7581c263e9ceb3a96204b314e199faddf", ++ "pipeline_repo": "Tinghong-Ye22/data_vqa_pipeline_1025", ++ "pipeline_revision": "9a2407adeed36a2645110390f1521304e36a2215", ++ "profile": "isolated_benchmark_1024_20261026", ++ "replay_arithmetic": "numpy2_float32_scalars", ++ "vendored_files": { ++ "upstream/hfd/src/psi/data/seed/third_party_dataset.py": "212fd08f50a6211caf3442d19772c8437b4de36ca0fb6de7f6865bfd621e84da" ++ } + } + } +diff --git a/configs/genmo.yaml b/configs/genmo.yaml +index 68fd000..a442cf6 100644 +--- a/configs/genmo.yaml ++++ b/configs/genmo.yaml +@@ -2,13 +2,15 @@ + model: genmo + genmo_root: /root/autodl-tmp/hfd-human-models/genmo + python: /root/autodl-tmp/envs/genmo/bin/python +-gpu: 1 +-output: /root/autodl-tmp/runs/hfd-genmo-english-v2 +-weights_output: /root/autodl-fs/pangxinlei/models/genmo/hfd-english-v2 ++gpu: 0 ++output: /root/autodl-tmp/runs/hfd-genmo-english-v2-split20261026-replay1264 ++weights_output: /root/autodl-fs/pangxinlei/models/genmo/hfd-english-v2-split20261026-replay1264 + init_checkpoint: /root/autodl-fs/pangxinlei/models/genmo/gem_smpl.ckpt + text_encoder: /root/autodl-fs/pangxinlei/models/genmo/t5-3b-local + body_models: /root/autodl-fs/pangxinlei/models/body_models + data: ++ replay_audit_status: /root/autodl-tmp/cache/hfd/genmo-replay1264/audit.json ++ require_replay_risk_gate: true + config: /root/autodl-tmp/hfd-baseline-portable/configs/sessions.yaml + reference_cache: /root/autodl-tmp/cache/hfd/reference-alignments.sqlite + language: en +@@ -43,6 +45,7 @@ strategy: + max_frames: 900 + seed: 12345 + training: ++ early_checkpoint_steps: [128, 256, 512, 1024] + # Total budget; review validation and closed-loop rollouts at 20K/40K/60K. + # Cosine T_max follows max_steps; resume must retain this original budget. + max_steps: 100000 +diff --git a/configs/genmo_legacy.yaml b/configs/genmo_legacy.yaml +index 57a7ea1..fa7f7aa 100644 +--- a/configs/genmo_legacy.yaml ++++ b/configs/genmo_legacy.yaml +@@ -3,8 +3,8 @@ model: genmo + genmo_root: /root/autodl-tmp/hfd-human-models/genmo + python: /root/autodl-tmp/envs/genmo/bin/python + gpu: 1 +-output: /root/autodl-tmp/runs/hfd-genmo-english-train +-weights_output: /root/autodl-fs/pangxinlei/models/genmo/hfd-english-train ++output: /root/autodl-tmp/runs/hfd-genmo-english-legacy-split20261026 ++weights_output: /root/autodl-fs/pangxinlei/models/genmo/hfd-english-legacy-split20261026 + init_checkpoint: /root/autodl-fs/pangxinlei/models/genmo/gem_smpl.ckpt + text_encoder: /root/autodl-fs/pangxinlei/models/genmo/t5-3b-local + body_models: /root/autodl-fs/pangxinlei/models/body_models +diff --git a/configs/kimodo.yaml b/configs/kimodo.yaml +index 5c27a61..4537176 100644 +--- a/configs/kimodo.yaml ++++ b/configs/kimodo.yaml +@@ -1,11 +1,11 @@ + # Full native English paired SMPL-X; text-only causal 30s context. + model: kimodo +-gpu: 0 ++gpu: 1 + model_dir: /root/autodl-fs/pangxinlei/models/kimodo/Kimodo-SMPLX-RP-v1 + session_config: /root/autodl-tmp/hfd-baseline-portable/configs/sessions.yaml + reference_cache: /root/autodl-tmp/cache/hfd/reference-alignments.sqlite +-output: /root/autodl-tmp/runs/hfd-kimodo-english-train +-weights_output: /root/autodl-tmp/models/kimodo/hfd-english-train ++output: /root/autodl-tmp/runs/hfd-kimodo-english-split20261026 ++weights_output: /root/autodl-fs/pangxinlei/models/kimodo/hfd-english-split20261026 + representation: + causal_root: true + future_frames: 60 +diff --git a/configs/omg.yaml b/configs/omg.yaml +index 9d973ce..c94cf3e 100644 +--- a/configs/omg.yaml ++++ b/configs/omg.yaml +@@ -1,4 +1,4 @@ +-# English foreground only; original indices/seed preserved. No ASR or translation. ++# English foreground only; isolated final split and recorded per-session seeds. + # Full-data English subset; long-run step budget still requires calibration. + schema_version: 1 + mode: train +@@ -7,9 +7,9 @@ framework: omg_lightning + paths: + data_disk: ${oc.env:HFD_DATA_DISK,/root/autodl-tmp} + upstream_root: upstream/omg +- session_root: ${paths.data_disk}/datasets/hfd-sessions ++ session_root: ${paths.data_disk}/datasets/hfd-sessions-split20261026 + cache_root: ${paths.data_disk}/cache +- output_dir: ${paths.data_disk}/runs/hfd-omg-english-train ++ output_dir: ${paths.data_disk}/runs/hfd-omg-english-split20261026 + runtime: + python: ${paths.data_disk}/envs/omg/bin/python + num_gpus: 1 +@@ -21,6 +21,8 @@ runtime: + distributed_timeout_seconds: 1800 + environment: + HFD_LOCAL_INDEX_CACHE: '1' ++ HFD_NATIVE_STARTUP_CACHE: '1' ++ HFD_STARTUP_INDEX_CACHE_DIR: ${paths.cache_root}/hfd/startup-indexes + data: + train_sessions: + - ${paths.session_root}/train/*.npz +@@ -28,7 +30,7 @@ data: + - ${paths.session_root}/val/*.npz + window_policy: full_session + max_windows: null +- cache_sessions: 512 ++ cache_sessions: 2048 + asr_dir: ${paths.session_root}/reference + asr_latency_seconds: 0.0 + audio_policy: hfd_past_2s +@@ -41,7 +43,7 @@ data: + prefetch_batches: 1 + native: + config: configs/sessions.yaml +- memory_mib: 2048 ++ memory_mib: 8192 + reference_cache: ${paths.cache_root}/hfd/reference-alignments.sqlite + reference_guard_seconds: 0.1 + refresh_batches: 256 +@@ -56,7 +58,7 @@ training: + precision: bf16-mixed + accumulate_grad_batches: 1 + gradient_clip_val: 1.0 +- save_every: 2000 ++ save_every: 256 + eval_every: 2000 + experiment: 50m + scheduler: +diff --git a/configs/omg_variants.yaml b/configs/omg_variants.yaml +index 87e9636..9d98296 100644 +--- a/configs/omg_variants.yaml ++++ b/configs/omg_variants.yaml +@@ -16,11 +16,11 @@ + experiment: 300m + checkpoint: checkpoints/paper/300m/sstep=055000.ckpt + sha256: df4ce9aeced84fd96a39583c7f6e2c5773b76f942cb43012c6c51dd9eb806806 +- batch_size: 32 +- accumulate_grad_batches: 16 ++ batch_size: 64 ++ accumulate_grad_batches: 8 + 500m: + experiment: 500m + checkpoint: checkpoints/paper/500m/sstep=050000.ckpt + sha256: f13bc58f07f104fcff67e35e79650def0fec5a6eb62a193fa1ae9f979fb911cf +- batch_size: 16 +- accumulate_grad_batches: 32 ++ batch_size: 32 ++ accumulate_grad_batches: 16 +diff --git a/configs/sessions.yaml b/configs/sessions.yaml +index c73045e..fa5652a 100644 +--- a/configs/sessions.yaml ++++ b/configs/sessions.yaml +@@ -2,7 +2,7 @@ + code_root: upstream/hfd + raw_root: /root/autodl-fs/pangxinlei/hfd_data + cache_root: ${oc.env:HFD_DATA_DISK,/root/autodl-tmp}/cache/hfd +-output_dir: ${oc.env:HFD_DATA_DISK,/root/autodl-tmp}/datasets/hfd-sessions/${split} ++output_dir: ${oc.env:HFD_DATA_DISK,/root/autodl-tmp}/datasets/hfd-sessions-split20261026/${split} + split: train + limit_sources: 0 + count: null +@@ -14,3 +14,11 @@ candidate_pool: false + recipe_overrides: + command_interruption_session_variants: 2 + include_third_party: true ++ ++# Final outer split selects sessions; upstream split is used only for replay. ++final_release: ++ profile: isolated_benchmark_1024_20261026 ++ revision: bdca2fe7581c263e9ceb3a96204b314e199faddf ++ directory: ${oc.env:HFD_DATA_DISK,/root/autodl-tmp}/cache/hfd/releases/bdca2fe7581c263e9ceb3a96204b314e199faddf/final_session_delivery ++ selection_index: ${oc.env:HFD_DATA_DISK,/root/autodl-tmp}/cache/hfd/releases/bdca2fe7581c263e9ceb3a96204b314e199faddf/motion-selection.sqlite ++ replay_support_root: ${oc.env:HFD_DATA_DISK,/root/autodl-tmp}/cache/hfd/releases/bdca2fe7581c263e9ceb3a96204b314e199faddf/runtime +diff --git a/configs/textop.yaml b/configs/textop.yaml +index 2235dab..49be7df 100644 +--- a/configs/textop.yaml ++++ b/configs/textop.yaml +@@ -1,4 +1,4 @@ +-# English foreground only; original indices/seed preserved. No ASR or translation. ++# English foreground only; isolated final split and recorded per-session seeds. + # Full-data English subset; long-run step budget still requires calibration. + schema_version: 1 + mode: train +@@ -7,9 +7,9 @@ framework: textop_native + paths: + data_disk: ${oc.env:HFD_DATA_DISK,/root/autodl-tmp} + upstream_root: upstream/textop +- session_root: ${paths.data_disk}/datasets/hfd-sessions ++ session_root: ${paths.data_disk}/datasets/hfd-sessions-split20261026 + cache_root: ${paths.data_disk}/cache +- output_dir: ${paths.data_disk}/runs/hfd-textop-english-train ++ output_dir: ${paths.data_disk}/runs/hfd-textop-english-split20261026 + runtime: + python: ${paths.data_disk}/envs/textop/bin/python + num_gpus: 1 +@@ -19,6 +19,8 @@ runtime: + offline: true + environment: + HFD_LOCAL_INDEX_CACHE: '1' ++ HFD_NATIVE_STARTUP_CACHE: '1' ++ HFD_STARTUP_INDEX_CACHE_DIR: ${paths.cache_root}/hfd/startup-indexes + data: + train_sessions: + - ${paths.session_root}/train/*.npz +@@ -26,7 +28,7 @@ data: + - ${paths.session_root}/val/*.npz + window_policy: full_session + max_windows: null +- cache_sessions: 512 ++ cache_sessions: 2048 + asr_dir: ${paths.session_root}/reference + asr_latency_seconds: 0.0 + audio_policy: released_asr +@@ -36,7 +38,7 @@ data: + backend: native_ondemand + native: + config: configs/sessions.yaml +- memory_mib: 2048 ++ memory_mib: 8192 + reference_cache: ${paths.cache_root}/hfd/reference-alignments.sqlite + reference_guard_seconds: 0.1 + refresh_batches: 64 +@@ -51,7 +53,7 @@ training: + precision: 32-true + accumulate_grad_batches: 1 + gradient_clip_val: 1.0 +- save_every: 2000 ++ save_every: 256 + eval_every: 2000 + num_primitive: 4 + stages: +diff --git a/docs/ASR_CONTEXT_OVERFIT.md b/docs/ASR_CONTEXT_OVERFIT.md +deleted file mode 100644 +index ce25532..0000000 +--- a/docs/ASR_CONTEXT_OVERFIT.md ++++ /dev/null +@@ -1,73 +0,0 @@ +-# OMG / TextOp 完整历史 ASR 过拟合测试 +- +-2026-10-18。结果目录:`/root/autodl-tmp/runs/archive/legacy-20261020/hfd-asr-context-overfit`。 +- +-本实验复用 `hfd-final/videos/omg_full_session_playable.mp4` 对应的原生双任务 session、混合音频和因果 Whisper sidecar。从相同官方初始权重重新训练,完整监督监听、执行、收尾动作;不把 GT 指令、任务阶段或控制标签作为输入。原来的 `hfd-final` 文件没有被覆盖。 +- +-## 输入变化 +- +-新增 `data.caption_policy: causal_asr_context`:每个释放点使用本次 ASR 对最近最多 30 秒已释放音频的**完整识别结果**;非空立即更新,空结果沿用上一份非空文本。不拼接历次重识别结果,不等两次一致,不用未来识别结果回填过去。这里保留的是 ASR 当前假设,最后一句可能尚未识别完整。 +- +-旧 `stable_latest_sentence` 只保留最近一句,并要求连续两次识别一致。本实验同时改变了文本覆盖范围和确认等待规则,因此不能把结果变化单独归因于历史句子的数量。本轮最初通过独立实验配置选择新策略;随后用户确认将其设为五个 baseline 的默认文本策略,见 [共同文本约定](BASELINE_TEXT_CONDITION.md)。旧 run 配置保留用于历史复现。 +- +-两个模型使用同一字符串,各自经过原生编码器。TextOp 保留 CLIP 的 77-token 限制和已有的超长文本最近后缀策略;OMG 保留 T5 的 50-token 限制。**本 session 的全部 37 个释放点,两个模型都没有截断。** 第二条指令出现后,输入可同时包含 `Taking an item from someone. Picking an item up from the floor.`,没有使用 metadata 修正 ASR。 +- +-`input_audit.json` 保存逐释放点的完整文本、旧稳定句、实际编码文本/令牌、截断标志;还核对 session 哈希、每个 ASR 音频切片哈希、释放边界与前缀不变性。ASR 仍是预计算的因果 sidecar,不计入实际识别计算延迟。 +- +-## 保持的训练与推理设置 +- +-| 项目 | OMG | TextOp | +-| --- | --- | --- | +-| 训练初始化 | 原官方 50M checkpoint | 原官方预训练 checkpoint | +-| 训练 seed | 42 | 42 | +-| 更新次数 | 1600 | 3200,阶段 `[800,1600,800]` | +-| batch / learning rate | 8 / 6e-5 | 4 / 1e-4 | +-| 额外条件 | 过去 2 秒已释放音频,10 帧动作历史 | 2 帧动作历史;不输入 PCM | +-| 动作预测 | 60 帧,30Hz | 8 帧,50Hz;训练链 4 个 primitive | +-| 推理 | DDIM 10 步、joint_one CFG、原生 async、24 帧余量、测得规划延迟 | CFG 5,每轮执行 8 帧 | +- +-保持原生网络、损失、优化器框架和动作表示。训练数据窗口与旧 run 相同:OMG 366 个窗口、TextOp 88 条连续链。只有开场使用 GT 初始化,随后完全反馈已执行的预测历史。TextOp 旧 run 保存的 `audio_policy: offline_target_aligned` 是未使用的历史字段,本轮按当前校验器写为 `released_asr`;两次都只使用因果 ASR。完整配置差异见 `experiment.json`。 +- +-OMG 的异步交接使用本次测量的规划耗时,旧新耗时可能不同,故误差差异包含实际调度影响,不是固定交接时刻的纯文本消融。主种子平均规划耗时旧版约 0.157 秒、新版约 0.175 秒。 +- +-## 结果 +- +-主视频沿用 seed=12345,没有挑选最优种子。关节 RMSE 排除开场初始化;OMG 统计 29 关节、TextOp 统计原生 23 关节,跨模型绝对值不能直接排名。 +- +-| 模型 | 旧稳定句关节 RMSE | 完整历史 ASR 关节 RMSE | 旧 / 新根坐标 RMSE | +-| --- | ---: | ---: | ---: | +-| OMG | 0.06191 rad | 0.07455 rad | 0.00844 / 0.01052 m | +-| TextOp | 0.29922 rad | 0.17019 rad | 0.04596 / 0.01603 m | +- +-| 模型 / 文本策略 | seed 12345 | seed 12346 | seed 12347 | 三种子均值 | +-| --- | ---: | ---: | ---: | ---: | +-| OMG 旧稳定句 | 0.06191 | 0.15385 | 0.05854 | 0.09143 | +-| OMG 完整历史 | 0.07455 | 0.13126 | 0.08651 | 0.09744 | +-| TextOp 旧稳定句 | 0.29922 | 0.36330 | 0.28990 | 0.31747 | +-| TextOp 完整历史 | 0.17019 | 0.30671 | 0.32361 | 0.26684 | +- +-TextOp 主种子的深弯腰峰值从旧版约 11.66 秒提前至 9.74 秒,接近 GT 的约 9.58 秒;但第一段动作仍有时序偏差,三个种子中仍有一个变差。其固定 GT 历史短窗口 RMSE 为 0.00976 rad,不能代替闭环结果。**完整历史文本对本条 session 的 TextOp 有帮助迹象,尚未实现稳定过拟合;OMG 没有整体改善。** +- +-三种子仅改变推理采样,每种文本策略只有一次训练、同一条训练 session;这不是独立训练重复或泛化评测。也不能据此认定模型已理解“旧指令被新指令替代”的对话语义。 +- +-## 产物与核验 +- +-- `videos/omg_full_session_playable.mp4`、`videos/textop_full_session_playable.mp4`:主种子视频,预测与原始 G1 GT 对照,保留混合音轨;均为运动学回放,没有物理 tracker。 +-- `comparison.png`、`video_contact_sheet.jpg`:旧新轨迹及视频帧对照。 +-- `summary.json`:全部采样种子、根误差、关节误差、跳变和规划耗时。 +-- `input_audit.json`、`experiment.json`、`verification.json`、`code_sha256.json`:输入、配置、检查与代码记录。 +-- 两个 `pipeline_check.json` 均通过有限 loss/梯度、参数更新、预期优化步数、原生 checkpoint 严格重载检查。28 项相关单元测试和两个数据合同检查通过。两个主视频完整解码及音频检查通过。 +-- OMG 冗余 checkpoint 先逐 tensor 比对最终权重与 global_step,确认一致后删除,保留最终 `last.ckpt` 与全部日志;记录在 `omg/checkpoint_cleanup.json`。 +- +-## 复现 +- +-以下训练输出目录应尚不存在;重跑时先覆盖 `paths.output_dir` 为新目录。 +- +-```bash +-/root/autodl-tmp/envs/omg/bin/python train.py --config configs/history/omg_asr_context_overfit.yaml +-/root/autodl-tmp/envs/textop/bin/python train.py --config configs/history/textop_asr_context_overfit.yaml +-/root/autodl-tmp/envs/omg/bin/python infer_overfit.py --run /root/autodl-tmp/runs/archive/legacy-20261020/hfd-asr-context-overfit/omg --continuous-only --skip-window-eval --task-seconds 0 14.8 +-/root/autodl-tmp/envs/textop/bin/python infer_overfit.py --run /root/autodl-tmp/runs/archive/legacy-20261020/hfd-asr-context-overfit/textop --continuous-only --task-seconds 0 14.8 +-/root/autodl-tmp/envs/omg/bin/python render_overfit.py --inference /root/autodl-tmp/runs/archive/legacy-20261020/hfd-asr-context-overfit/omg/inference --assets-root /root/autodl-tmp/models/OMG/OMG/assets --gpu 1 +-/root/autodl-tmp/envs/omg/bin/python render_overfit.py --inference /root/autodl-tmp/runs/archive/legacy-20261020/hfd-asr-context-overfit/textop/inference --assets-root /root/autodl-tmp/models/OMG/OMG/assets --gpu 1 +-``` +diff --git a/docs/BASELINE_TEXT_CONDITION.md b/docs/BASELINE_TEXT_CONDITION.md +deleted file mode 100644 +index 8d819f7..0000000 +--- a/docs/BASELINE_TEXT_CONDITION.md ++++ /dev/null +@@ -1,63 +0,0 @@ +-# 五个 baseline 的共同文本条件 +- +-2026-10-21 更新:按用户决定,当前训练仅选英文前景 session,使用 `aligned_reference_context`、`clock_policy: microphone`、`text_conditioning: none`。停止对中文样本做训练和在线翻译。原始中文数据及历史报告保留,中文文本缓存已清理。完整输入/输出与实现状态见[当前五模型协议](HFD_HUMAN_BASELINE_INPUT_PROTOCOL.md)。 +- +-模型接收截至当前释放点、最近 30 秒的英文前景文字。文字来自原始 TTS 文本的离线词级对齐,不再经过 ASR。词结束后加实验性的 0.1 秒 guard,再向上取整到 `0.6+0.4k` 秒释放;该 guard 不是已证明的对齐误差上界。无效词跨度整段结束后释放,禁止提前展示完整指令。背景 PCM 原样保留,背景文字不作为输入。这是 oracle 前景转写实验。 +- +-原生动作时刻 `t` 对应麦克风 `t+0.2s`;只允许使用已经到达的文字和音频。文字不跨 session 累积,不输入控制标签。编码器 token 预算仍有效:OMG T5 50 token、TextOp CLIP 77 token(75正文+两个边界)、GenMo T5-3B 50 token,超长时保留最近后缀。前端 30 秒不等于全部文字都能进入编码器。文本过期后允许为空,不沿用旧ASR“空结果保留最后非空”的行为。 +- +-| Baseline | 文本与额外输入 | 当前接入范围 | +-| --- | --- | --- | +-| OMG | 原生 T5-Base;过去2秒已释放混合音频+10帧动作历史@30Hz | 动态英文训练、连续验证与视频 | +-| TextOp | 原生 CLIP ViT-B/32;2帧57D动作历史@50Hz,不输入音频 | 动态英文300K训练、连续验证与视频 | +-| GenMo / GEM-SMPL | 原生 T5-3B;60帧/2秒人体历史@30Hz,当前音频关闭 | 动态原生英文人体训练、验证、完整resume、retarget视频;100K配置 | +-| HY-Motion | 原生 Qwen3 与 CLIP-L | 共同契约;完整 HFD 训练入口尚未实现 | +-| Kimodo | 冻结 LLM2Vec / Llama-3-8B,最多400正文token,mean pooling为1×4096 | 共同契约;固定英文配对子集训练/闭环/GMR入口已接通,见[Kimodo训练](KIMODO_HFD_TRAINING.md) | +- +-OMG/TextOp 的正式默认配置与 reference 配置均设置 `data.native.language: en`。原生索引、session ID 和 seed 不重排,仅将中文候选的采样权重置零;第三方已组装 session 同样按前景语言过滤。剩余 train **168,583**、val **5,109** 个候选 session。共享声明见 `configs/english_text_protocol.yaml`,详细时间戳与动态组装见 [动态 reference 训练说明](REFERENCE_DYNAMIC_TRAINING.md)。 +- +-GenMo只使用配对人体的原生英文子集:train **155,117**、val **4,719**,排除尚无人体真值的第三方生产session。全量session没有提前落盘;离线缓存的是源语音词对齐,原生session在训练时动态组装并做有上限的内存缓存。三模型的数据范围与音频可见性不同,不能仅按step或误差绝对值直接比较架构优劣。 +- +-GenMo 新缓存示例: +- +-```bash +-python cache_genmo_text.py --asr SESSION.reference.json --output NEW_TEXT_CACHE.pt --caption-policy aligned_reference_context --text-conditioning none +-``` +- +-旧 ASR 对照配置归档在 `configs/history/omg_bilingual_asr.yaml` 和 `configs/history/textop_bilingual_asr.yaml`。历史策略 `causal_asr_context` 使用混合音频的因果 ASR 及冻结 OPUS 中文→英文转换,曾有动作语义误译。旧配置、权重和日志不改写,但重跑中文历史实验需重建已删除的中文缓存。历史依据见 [ASR 过拟合报告](ASR_CONTEXT_OVERFIT.md) 和 [语义时钟修复](SEMANTICS_CLOCK_FIX.md)。 +- +-实际误差、吞吐和清理记录见 [纯英文短测报告](ENGLISH_ONLY_PROBE.md)。 +- +-## 历史双语盘点(清理前) +- +-### 本地中英文文本配对盘点(2026-10-19) +- +-原始 `audio/seed_tts/metadata.jsonl` 共 142,220 行,包含 5,220 个原始音频目录及 5,220 对不同的中英文完整指令;按字符串去重后为中文 5,060 条、英文 5,220 条。中英字段缺失均为 0。同一指令会对应多份动作和多个声音版本,142,220 行不能当作独立句子数。 +- +-加入停止、附和、犹豫、修正及拆分的指令片段后,已对齐的前景资产包含 9,313 种不同中文字符串和 9,140 种不同英文字符串,中文音频 61,860 个、英文音频 61,860 个。这不是动态 session 组合文本的数量,也不包含对背景说话者做完整中英配对的承诺。停止短语有 12 种中文及 12 种英文,附和短语各 15 种。 +- +-完整动作指令都有英文对应字段,但不能据此保证全部等义或一一对应: +- +-- 141 个不同中文字符串对应多个英文字符串,有些仅为措辞变体,也有“单手/双手”“轻/重”等实质区别。例如“左手持小发光物向左转身”同时对应英文的 one hand 和 two hands 描述。该计数不是已确认错误条数。 +-- 已知 `retro`(复古)被译为“后退”的问题涉及原始 594 行、30 个音频目录。原生规则会排除对应的 120 个中文声线版本;实际选定第三方 recipe 中也未发现这些错误中文资产。完整资产对齐库仍保留这些原始资源,不能将其存在误认为都进入训练。 +-- 中英指令独立拆分后,prefix/suffix 不一定语义对齐。例如“检查全身”对应 checking whole body,但中文 prefix 是“检查”,英文 prefix 是 checking whole;不能直接逐片替换。 +- +-现有完整配对英文可以作为复用与质量审查的起点,但在中文只说出一部分时不能提前释放完整英文。前面“转写正确”指文字与合成语音所用原文一致,不代表已完成全部中英等义人工校验。数据报告:`/root/autodl-tmp/runs/archive/legacy-20261020/reference-full-preparation-20261019/bilingual-inventory-audit.json`。 +- +-## 官方语言支持核查(2026-10-19) +- +-核查对象为 OMG(Omni-Modal Motion Generation)、TextOp、NVlabs GENMO/GEM、HY-Motion-1.0 和 Kimodo。区分动作模型的语言验证、文本编码器的语言说明、以及前端翻译/改写能力。“未找到中文声明”不等于证明中文绝对不可用,也不等于已证明只支持英文。 +- +-| 模型 | 官方可核查的说明 | 对中文输入的结论 | +-| --- | --- | --- | +-| OMG | [论文附录 B.3](https://arxiv.org/html/2606.10340v1) 明确要求 action 和 style 标注写成英文;C.3 使用冻结 T5-Base、最多 50 token。 | 查阅的论文和官方 README 未列中文支持或中文动作评测。英文标注规范与本地中文 token 坍缩相符。 | +-| TextOp | [论文 III-B/III-C](https://arxiv.org/html/2602.07439v1) 使用 BABEL 动作文本标注和 CLIP;[CLIP 官方模型卡](https://github.com/openai/CLIP/blob/main/model-card.md#out-of-scope-use-cases) 说明未专门训练或评估英文以外语言,限定英文用途。 | TextOp 未给完整语言清单,未找到中文动作评测或中文支持承诺。不能把中文可分词视为已支持中文语义。 | +-| GenMo / GEM-SMPL | [论文附录](https://arxiv.org/html/2505.01425v1) 使用 T5 文本编码;[官方配置](https://github.com/NVlabs/GENMO/blob/main/configs/exp/gem_smpl.yaml) 指定 `t5-3b`。 | 未找到动作模型语言清单、中文支持声明或中文动作评测。英文示例和 T5 选型是使用证据,不是所有其他语言被排除的证明。 | +-| HY-Motion-1.0 | [官方模型卡](https://huggingface.co/tencent/HY-Motion-1.0) 标注 English/Chinese,但其 Prompting Guide 要求使用英文,其他语言由 Text2MotionPrompter 改写。[论文第 3.3 节](https://arxiv.org/html/2512.23464v1) 的改写模块训练输入包含中英混合请求。 | 官方支持链路包含中文用户输入,但通过前置改写生成英文动作描述。未找到绕过改写后,核心动作模型直接读中文的明确质量保证或专门评测。 | +-| Kimodo | [技术报告第 4.2 节](https://research.nvidia.com/labs/sil/projects/kimodo/assets/kimodo_tech_report.pdf) 使用 LLM2Vec;[官方编码器配置](https://github.com/nv-tlabs/kimodo/blob/main/kimodo/scripts/run_text_encoder_server.py) 指向 Llama-3-8B-Instruct 的 LLM2Vec 适配器,其[模型卡](https://huggingface.co/McGill-NLP/LLM2Vec-Meta-Llama-3-8B-Instruct-mntp-supervised) 标注 English。 | Kimodo 技术报告和 SMPLX-RP-v1 模型卡未提供动作提示词语言清单或中文动作评测。Llama 能处理某些中文,不构成动作模型中文能力的验证。 | +- +-HY-Motion 的官方实现直接要求改写输出为英文,见 [prompt engineering 的系统提示](https://github.com/Tencent-Hunyuan/HY-Motion-1.0/blob/master/hymotion/prompt_engineering/model_constants.py)。因此模型卡的中英文标签与“动作生成推荐英文”应一起阅读,不能仅依据标签宣称核心网络原生中英双语,也不能把整套系统说成完全不能接收中文。 +- +-另需澄清:[T5-Base](https://huggingface.co/google-t5/t5-base) 和 [T5-3B](https://huggingface.co/google-t5/t5-3b) 的官方 NLP 语言列表为英语、法语、罗马尼亚语、德语,没有中文。这也不表示采用 T5 的动作模型已验证了法语、德语或罗马尼亚语动作生成;编码器能力与下游训练覆盖应分别报告。 +- +-对当前计划的含义:使用共享英文文本前端符合已查到的主要官方使用路径,尤其是 HY-Motion 已明确建议其他语言先改写。共享前端仍须验证翻译质量;直接中文训练或更换多语编码器应单列实验,不能以“官方已支持中文”跳过检查。本轮没有修改训练或推理配置。 +diff --git a/docs/ENGLISH_ONLY_PROBE.md b/docs/ENGLISH_ONLY_PROBE.md +deleted file mode 100644 +index 62771a8..0000000 +--- a/docs/ENGLISH_ONLY_PROBE.md ++++ /dev/null +@@ -1,90 +0,0 @@ +-# 纯英文训练切换与短测(2026-10-19) +- +-后续性能更新:已启用本地只读索引镜像并通过原生训练对照;旧正式训练已停止并清理,详见 [最新性能报告](LOCAL_INDEX_TRAINING.md)。下文耗时属于优化前记录。 +- +-清理说明:本文记录的一次性实验脚本已归档,复现前请按[脚本恢复说明](WORKSPACE_CLEANUP.md)取回;历史指标与日志保留。 +- +-已将当前 OMG/TextOp 正式默认、reference、smoke 和完整 session 配置切换为英文前景、离线词级对齐、最近 30 秒上下文、不翻译。GenMo 单 session 配置和新建文本缓存默认策略同步;HY-Motion/Kimodo 记录共同约定,尚无完整 HFD 训练适配。本轮实际训练及误差、吞吐验证仅覆盖 OMG/TextOp。没有启动 100K/300K 正式长训练。 +- +-英文路线可完成训练和严格重载,但本轮没有证据证明它一致改善动作质量。OMG 相对旧双语短测的关节误差下降 4.4%,TextOp 上升 1.8%;OMG 仍未优于官方初始模型。数据准备仍是主要效率瓶颈。 +- +-## 数据与清理 +- +-原生 train 155,117、val 4,719;选定第三方 train 13,466、val 390,合计 **168,583 / 5,109**。原始全量索引与 session ID、seed 不重排,中文候选权重置零;六种交互类型均保留。native train 覆盖 4,245 种英文指令和 99,099 个动作文件 ID,这两个计数不包含第三方的额外去重统计。 +- +-删除中文资产对齐条目 61,860 个、中文孤立对齐记录 27 个、非英文全文 recipe 13,695 个、翻译缓存 10,692 条、含中文文本的旧 ASR 缓存 109,323 条,以及 63 份中文 reference/ASR sidecar,约回收 0.8 GB。英文资产内容摘要删除前后相同,当前对齐库只含 61,860 个英文资产,完整性检查通过。旧英文 ASR 条目 100,509 个作为历史缓存保留,当前训练不使用。 +- +-原始 tar、原始双语 metadata、日志、历史报告、模型及英文缓存保留。recipe 的 ASCII 语言标签保留以过滤中文 ID,中文全文 recipe 已删除。背景 PCM 不变,过滤的是前景目标语言。清理记录:[cache-cleanup.json](/root/autodl-tmp/runs/archive/legacy-20261020/english-only-20261019/cache-cleanup.json)。 +- +-文字是原始 TTS 文本的 oracle 前景转写,按实际动态 session 拼接位置平移词时间戳;不需要在线 ASR。释放时钟为 `0.6+0.4k` 秒、词结束后加实验性 0.1 秒 guard,保留最近 30 秒。英文中 6,608/61,860(10.68%)个资产词边界无效,整段结束后释放;该 guard 未证明能覆盖所有对齐误差。OMG 输入过去 2 秒音频、已释放文本及动作历史;TextOp 输入已释放文本及动作历史。两者均未加载 ASR 或翻译模型。 +- +-## 误差对照 +- +-两次新训练均从官方权重初始化。OMG batch64、512 步、lr6e-5、bf16;TextOp batch64、768 步、lr1e-4、FP32,三阶段各 256 步。对照旧双语试验采用相同训练步数、batch、学习率和 seed42;采样范围按实验目的改变为英文。不是在旧双语 checkpoint 上续训。 +- +-在同一组三个英文 val session(普通指令、打断、急停)上,对官方初始、旧双语、纯英文三个 checkpoint 做连续闭环回放。各 checkpoint 使用完全相同的原始音频、oracle 文本、释放时钟、一次 GT bootstrap、推理参数和 seed12345。OMG 固定规划延迟 4 帧、原生异步规划;TextOp 原生 CFG5、预测历史。计分排除 bootstrap;OMG 计 29 关节,TextOp 按其原生覆盖关节计分,不能横向比较两模型绝对值。 +- +-以下为三个 session 各自 RMSE 的等权平均,单位 rad: +- +-| 模型 | 官方初始 | 旧双语 | 纯英文 | 纯英文相对双语 | +-| --- | ---: | ---: | ---: | ---: | +-| OMG | 0.4043 | 0.4257 | 0.4070 | -4.4% | +-| TEXTOP | 0.4848 | 0.4117 | 0.4193 | +1.8% | +- +-根节点 xyz RMSE 的相同等权平均,单位 m: +- +-| 模型 | 官方初始 | 旧双语 | 纯英文 | +-| --- | ---: | ---: | ---: | +-| OMG | 0.3002 | 1.5859 | 0.7177 | +-| TEXTOP | 0.7033 | 0.5317 | 0.6196 | +- +-逐 session 的关节 RMSE: +- +-| 模型 | 交互 | 官方初始 | 旧双语 | 纯英文 | +-| --- | --- | ---: | ---: | ---: | +-| OMG | turn-taking | 0.3616 | 0.3794 | 0.3654 | +-| OMG | interruption | 0.3404 | 0.3319 | 0.3097 | +-| OMG | emergency-stop | 0.5109 | 0.5657 | 0.5459 | +-| TEXTOP | turn-taking | 0.3750 | 0.3913 | 0.3873 | +-| TEXTOP | interruption | 0.3582 | 0.3401 | 0.3461 | +-| TEXTOP | emergency-stop | 0.7211 | 0.5037 | 0.5244 | +- +-OMG 新版比旧双语更接近 GT,但官方初始关节误差仍略低,根节点误差也更低。TextOp 新版比双语略差,虽然优于官方初始。这只是三个样本、单一随机种子的短测,没有统计显著性保证;单一 GT 的运动 RMSE 也不是指令执行成功率,不能据此宣称全双工能力改善。 +- +-OMG 英文固定验证损失为 0.05414 → 0.05118 → 0.04272 → 0.04288(128/256/384/512 步)。TextOp 为 0.13086、0.13235、0.13669、0.14283、0.13936、0.12966(每 128 步);阶段历史分布在 256/512 步改变,不能把整段曲线当作同一目标的收敛曲线。旧双语 native validation 含中文,新版仅英文,二者验证损失不能直接比较优劣。 +- +-## 训练效率 +- +-OMG 固定 GPU1,TextOp 固定 GPU0。另从官方初始权重做 batch512 测速:OMG64步、TextOp96步,分别与 batch64 对照处理相同的 32,768 / 49,152 个训练实例;优化器更新次数不同,batch512 权重不用于上面的质量结论。TextOp 的“实例”按每次更新的 primitive 样本计,OMG 为动作窗,不能拿两模型的样本/秒直接比架构效率。 +- +-“池内”是排除超过 2 秒的冷池/验证间隔后的连续步间隔中位数,包含热数据准备和训练。“训练阶段”包含组池、验证和周期保存;“总墙钟”还包括索引/模型初始化和结束检查。batch512 复用了前一阶段的 CPU 原始索引,因此它的总墙钟不能与含冷索引构建的 batch64 直接当作正式加速比。 +- +-| 模型 | Batch | 步数 | 池内秒/步 | 池内实例/秒 | 训练阶段秒 | 训练阶段实例/秒 | 总墙钟秒 | +-| --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: | +-| OMG | 64 | 512 | 0.118 | 543 | 547.4 | 59.9 | 729.1 | +-| OMG | 512 | 64 | 0.535 | 957 | 385.6 | 85.0 | 397.9 | +-| TEXTOP | 64 | 768 | 0.083 | 775 | 300.9 | 163.3 | 478.8 | +-| TEXTOP | 512 | 96 | 0.149 | 3430 | 399.6 | 123.0 | 410.3 | +- +-OMG batch512 的首池准备 341.6 秒,TextOp 为 374.2 秒。训练阶段分别约 88.6% / 93.6% 时间用于该首池准备。batch64 OMG 8 池共 431.5 秒,TextOp 3 池共 169.4 秒。实际观察到共享盘 tar 读取等待,GPU 在组池时基本空闲。两秒采样、排除各卡训练后的推理后,包含组池/验证的平均 GPU 利用率仅约 GPU1 3.1%、GPU0 3.9%;采样从索引初始化后开始,不能当成全程精确 GPU 工时。 +- +-batch512 峰值 PyTorch 分配显存 OMG **16.21 GiB**、TextOp **2.62 GiB**,均可在当前卡运行。TextOp 的 batch512 短测没有覆盖完整 256 更新的数据池刷新周期,只用一个池跑了96步;batch64 跑了三个64-session池,故新鲜 session 数不同。大 batch 的池内吞吐高,但首池成本和数据复用程度必须一起看。 +- +-OMG 两阶段启用逐步训练健康检查;TextOp matched 为匹配旧试验关闭额外健康 hook,batch512 开启,因此不能把全部吞吐差异归因于 batch。历史双语 OMG 还多准备了一个未使用的尾池,当前已修复;其 1063.8 秒历史 fit 时间与本轮 547.4 秒的差异包含此独立修复及缓存状态差异,不是单独“去中文”的严格加速比。 +- +-当前瓶颈仍在数据准备。后续效率工作应优先测后台组池、原始资产/动作特征缓存及共享盘读取,再决定正式步数和刷新频率;仅增大 batch 或把旧 100K/300K 步原样启动,不能保证合理训练时长或动作质量。 +- +-## 验证与产物 +- +-60 项相关单元测试、两套正式英文配置 dry-run 通过。OMG matched/batch512 分别记录到 512/64 次真实优化器更新,有限损失/梯度、参数变化和严格重载通过;TextOp batch512 记录96次真实更新并通过同样检查。TextOp matched 记录768步有限损失/梯度范数,108个 denoiser 状态张量变化且有限,实际回放完成严格加载;该阶段没有单独 hook 计数优化器。没有训练结束后的多余冷池。 +- +-- [完整指标与逐样本误差](/root/autodl-tmp/runs/archive/legacy-20261020/english-only-20261019/summary.json) +-- [英文验证曲线](/root/autodl-tmp/runs/archive/legacy-20261020/english-only-20261019/english-validation.png) +-- [GPU 利用率](/root/autodl-tmp/runs/archive/legacy-20261020/english-only-20261019/gpu-utilization.png) +-- [最终数据和前端检查](/root/autodl-tmp/runs/archive/legacy-20261020/english-only-20261019/final-audit.json) +-- [OMG 急停视频](/root/autodl-tmp/runs/archive/legacy-20261020/english-only-20261019/evaluation/omg/emergency-stop/english/inference/complete_task.mp4) +-- [TextOp 急停视频](/root/autodl-tmp/runs/archive/legacy-20261020/english-only-20261019/evaluation/textop/emergency-stop/english/inference/complete_task.mp4) +- +-视频保留原始混合音频,显示预测与 GT、麦克风时钟和实际释放的文本;两份均已完整解码检查。OMG/TextOp 直接输出原生 G1,无需 SMPL-X retarget。GenMo 等人体模型仍须遵循现有 SMPL-X→G1 retarget 与视频评测链路,本轮没有宣称已完成它们的新训练。 +- +-运行配置、日志及 checkpoint 位于 `/root/autodl-tmp/runs/archive/legacy-20261020/english-only-20261019/`。脚本为 `tools/run_english_training_probe.py`、`tools/compare_english_checkpoints.py`、`tools/summarize_english_probe.py`。已更新正式默认配置,但正式长训练未启动;100K/300K 仍属于待校准预算。 +diff --git a/docs/FORMAL_TRAINING_READINESS.md b/docs/FORMAL_TRAINING_READINESS.md +deleted file mode 100644 +index 426a887..0000000 +--- a/docs/FORMAL_TRAINING_READINESS.md ++++ /dev/null +@@ -1,117 +0,0 @@ +-# 正式训练准备检查 +- +-后续性能更新:已启用本地只读索引镜像并通过原生训练对照;旧正式训练已停止并清理,详见 [最新性能报告](LOCAL_INDEX_TRAINING.md)。下文耗时属于优化前记录。 +- +-2026-10-19 英文路线更新:正式默认与 reference 配置均仅采样英文前景 session,离线对齐、无 ASR/翻译;train 168,583、val 5,109。中文缓存已清理。当前有步数上限的双卡误差与吞吐短测见 [纯英文短测报告](ENGLISH_ONLY_PROBE.md),100K/300K 正式训练未启动,旧预算仍需校准。 +- +-2026-10-19 较早更新:离线前景转写对齐和动态 session 读取已接入独立 reference 配置,详见 [离线对齐与动态训练](REFERENCE_DYNAMIC_TRAINING.md)。当前是一卡 OMG、一卡 TextOp;下文的双卡 OMG、TextOp batch64 和 ASR 预算属于历史记录。中文翻译质量、TextOp 连续回放和新版训练预算仍需验证,不能直接沿用旧文中的启动结论。 +- +-2026-10-17:OMG 已完成一轮 100k 更新,训练循环 108.55 小时。实际结果、数据范围、性能瓶颈和新配置见 [正式训练复查](OMG_TRAINING_REVIEW.md)。当前增加 `data.batched_motion=true`、`training.fixed_validation=true` 和 `training.save_best=true`;早停默认关闭,可用 `training.early_stopping.enabled=true` 启用。新增最佳 checkpoint 约多占 1 GiB。下面 10 月 10 日的短测和资源规划为历史准备记录,不应当作此次完整训练的实测速度。 +- +-同日后续:OMG 默认 `data.prefetch_batches=1`。双卡 64 步同步/异步分段对照,排除边界两步后,池内平均每步由 1.07 秒降到 0.62 秒。该数值不包含下一池刷新和 ASR,也不是整轮加速承诺。主机内存预算需另留每卡一个 CPU batch;`num_workers=0` 保持不变。 +- +-2026-10-10 更新:419 个原始 tar 已全部下载且 SHA256 校验通过;正式配置改为 OMG 双卡、TextOP 单卡,每次只运行一个模型。全量原生 session 配方已通过索引构造,原有两模型 batch=2 的短测已通过。OMG 每卡 batch 512 的显存探针已通过;低吞吐正式运行已按用户要求终止并删除,当前仅进行有明确步数上限的性能短测。 +- +-## 已确认的配方修正 +- +-用户确认将 `configs/sessions.yaml` 中的 `command_interruption_session_variants` 从 1 调到 2。原值在当前数据版本只能产生 36,836 个打断候选,低于原生配比要求的 39,743;修改后有 73,672 个候选,原生生成、筛选和最长 session 剔除流程均通过。 +- +-不复制 session,不跳过最终配比,不改变原生任务类型目标。增加 variant 可能改变后续随机组装结果,应把这个 override 同步记录到论文 HFD 和 baseline 数据配方中。原始 recipe 源码保持不变,覆盖参数与源码哈希保存在检查报告中。 +- +-| 类型 | train | val | +-| --- | ---: | ---: | +-| turn-taking | 158,812 | 4,539 | +-| interruption | 39,703 | 1,134 | +-| emergency-stop | 26,468 | 756 | +-| backchannel | 39,703 | 1,134 | +-| hesitation | 34,071 | 1,016 | +-| pause | 23,486 | 693 | +-| 合计 | 322,243 | 9,272 | +- +-这些是原生筛选及每类最长 0.1% 剔除后的计数。interrupt/control/backchannel 的目标按基础任务 session 数定义;hesitation/pause 按窗口数加入,因此不能直接拿最终六类 session 百分比验证全部配比。train/val 沿用原生动作来源分组划分。原生长度估计对应约 3,304.3/94.6 小时,实际导出长度会随 idle、打断时刻等而变化。 +- +-## 全量数据准备的实际限制 +- +-- 检查开始时,两张 RTX 5090 32 GB 空闲;TextOP、OMG、CLIP、T5、Whisper 权重均已在本机。 +-- 原始分片共 419 个,完整 tar 约 49.8 GiB,另有索引和 metadata。2026-10-09 已有 419 个完整分片,大小和 SHA256 均通过清单校验,无需继续下载。原始目录保持在共享盘。后续使用 `tools/download_hfd_tars.py` 校验 SHA256、跳过完整文件并续传 `.partial`;命令及空间要求见 [迁移说明](MIGRATION.md)。实际 hesitation 样本仍读取 v1 分片中的资源,不能仅凭配方名为 v2 就忽略这些文件。 +-- 按原生估计长度,在每个交互类型中选一个中位附近 session,六个均完成导出、`validate_sessions.py` 检查和完整因果 ASR;这些仅用于容量与吞吐测量,没有进入训练。 +-- 六个样本的压缩 NPZ 约为每音频秒 37,765 字节。乘以原生总估计时长,train+val 约 **430 GiB**,还不含 JSON、ASR 和 checkpoint。此为旧完整导出方案的六样本推算,当前按需方案无需分配这部分空间;收尾时数据盘剩余约 44 GiB。 +-- 逐窗口 ASR 在五个热身后样本上约 0.048–0.091 秒/窗口。六个样本共 483 个因果窗口,逐条总耗时 57.9 秒,batch=32 总耗时 7.45 秒(约 7.8 倍);两个测量的热身条件并不完全一致,不能作为严格性能基准。16/483 个窗口的文本有差异,PCM 和截止时刻一致。保持混合音频,不用 GT 替代 ASR。 +-- 旧 NPZ 加载器只有 4-session 缓存且 OMG 保留所有周期 checkpoint。当前正式配置改用按需读取、OMG 512-session 特征缓存、每个 split 2 GiB 数组预算和最近 3 个周期 checkpoint,旧方案只保留用于已有小样本复现。 +- +-已选择按需组装方案:`configs/textop.yaml` 和 `configs/omg.yaml` 使用 `data.backend: native_ondemand`。每个 batch 槽位按合法动作窗口的长度加权抽 session,再从 session 内均匀取窗口;OMG 当前每卡 batch 为 512,正式值每 256 个 batch 刷新一次 session 池。相邻 batch 有相关性,但保持合法窗口的边际采样概率;每个 batch 仍会在已保留 session 内重新抽动作点。较长刷新间隔用于摊薄原生组装和 ASR 成本,若需更强 session 随机性可改回 64。OMG 缓存覆盖一个完整槽位池;批量音频特征计算保留逐窗口 35 维特征定义,独立对照容差为 rtol=atol=1e-6(存在浮点运算顺序差异)。原生配比定义的是候选 session 集合,不保证每个 batch 都满足该比例。增强使用固定 seed=42,缓存失效后重组同一 session 不改变其内容,也不每个 epoch 重新随机增强。ASR 结果以压缩 SQLite 缓存,Whisper 使用 batch=32 的因果窗口;批处理参数和实现哈希记录在 run manifest,实际识别结果位于缓存。它与逐条识别不是字节级等价,做论文复现时应一并保存匹配的缓存。 +- +-OMG 和 TextOP 的按需 smoke 都已通过(各 4 次更新、有限梯度/参数变化、验证读取、严格 checkpoint 加载)。六个样本的动作/PCM 与旧导出逐元素一致。少数窗口文字因 batch/半精度数值产生差异,这是已记录的预处理口径变化,不是未来音频或 GT 输入。 +- +-按需训练不生成数百 GiB 的派生 NPZ;当前数据盘只存 raw-range cache、ASR SQLite、日志和模型输出。建议先补齐原始 tar,再长时间训练;当前代码支持缺失文件的 range 回退,但受网络可用性与重试影响,不能按本地数据速度估算。`runtime.offline: true` 控制模型权重加载,不会关闭这个 raw tar 网络回退。 +- +-## 正式训练命令和资源预算 +- +-2026-10-09 配置改为**每次只训练一个模型**:OMG 使用 GPU 0、1 的 DDP;TextOP 沿用原生单卡训练器。OMG 的按需加载器在每个 rank 使用独立窗口采样随机流、对应 GPU 的 Whisper,以及分别命名的采样记录;session 构造种子保持一致。两个 rank 从完整原生集合独立有放回抽样,允许偶然抽到同一个 session,不是两卡复制同一个 batch。每个 rank 保留完整索引,模型梯度由 DDP 同步。 +- +-先在仓库目录检查配置(不启动训练): +- +-```bash +-cd /root/autodl-tmp/hfd-baseline-portable +-export HFD_DATA_DISK=/root/autodl-tmp +-/root/autodl-tmp/envs/omg/bin/python train.py --config configs/omg.yaml --dry-run +-/root/autodl-tmp/envs/textop/bin/python train.py --config configs/textop.yaml --dry-run +-``` +- +-在 screen 中运行 OMG: +- +-```bash +-screen -S hfd-omg +-cd /root/autodl-tmp/hfd-baseline-portable +-/root/autodl-tmp/envs/omg/bin/python -u train.py --config configs/omg.yaml +-``` +- +-OMG 结束后,另开一个 screen 运行 TextOP: +- +-```bash +-screen -S hfd-textop +-cd /root/autodl-tmp/hfd-baseline-portable +-/root/autodl-tmp/envs/textop/bin/python -u train.py --config configs/textop.yaml +-``` +- +-已经在 screen 内时只执行 `cd` 和 Python 命令,不嵌套创建 screen。`Ctrl+A` 然后 `D` 脱离,`screen -r hfd-omg` 或 `screen -r hfd-textop` 恢复。不要同时执行这两个训练命令。终端直接显示数据准备和训练进度,无需 `nohup`、`tee` 或输出重定向;run 中仍保留配置、checkpoint 和小型结构化指标,供复查实验。 +- +-| 项目 | TextOP | OMG | +-| --- | --- | --- | +-| GPU | 1 张 RTX 5090 32 GB,GPU 0 | 2 张 RTX 5090 32 GB,GPU 0、1 | +-| batch | 64 | 每卡 512,全局 1024 | +-| 精度 | FP32 | bf16 mixed | +-| 更新步数 | 300,000;三阶段各 100,000 步 | 100,000 | +-| 学习率 | 1e-4 | 6e-5 | +-| 数据读取 | 按需原生组装,完整 train/val | 同左,DDP 独立采样 | +-| ASR | Whisper tiny,batch 32,共用训练卡 | 每个 rank 在对应训练卡使用 Whisper tiny,batch 32 | +-| CPU | 4 线程,num_workers=0 | 每 rank 4 线程,num_workers=0;音频特征按 batch 向量化 | +-| 输出目录 | `/root/autodl-tmp/runs/hfd-textop-train` | `/root/autodl-tmp/runs/archive/legacy-20261020/hfd-omg-train` | +-| 保存策略 | 每 20,000 步,保留最新 3 份 | 每 2,000 步,保留最新 3 份 | +- +-输出目录必须尚不存在;启动新实验可加 `--set paths.output_dir=/root/autodl-tmp/runs/新的名称`。OMG 入口自动调用 torchrun,无需手工加启动器。默认沿用本机可用的 Gloo,DDP 超时 1,800 秒,给按需组装和 ASR 留余量。OMG 的 `training.batch_size` 是每卡 batch;当前每卡 512、总 batch 1024。真实 forward/backward 探针峰值约为 21.9 GiB/卡(另有 Whisper 常驻);没有把 640 作为正式值,因为 640 已达到约 27.3 GiB,给优化器和 DDP 临时张量的余量较小。改回 OMG 单卡时加 `--set runtime.num_gpus=1 'runtime.gpu_ids=[0]' training.batch_size=512`;保持学习率 6e-5 以便与 baseline 口径一致。 +- +-两模型都对完整 session 的动作窗口进行监督,保留 listen/idle 对应关节动作,不输入控制标签。OMG 使用过去 2 秒音频+因果 ASR;TextOP 使用因果 ASR。此次双卡改动不改变上述输入或数据配方。 +- +-训练是按优化更新步数停止的随机窗口采样,并不是逐个遍历全部 session 的 epoch:TextOP 每次取一个 batch 后依次对 4 个 primitive 做 4 次更新,共取约 75,000 组 batch;OMG 共取 100,000 组双卡 batch。OMG 槽位每 256 个 batch 刷新一次,TextOP 每 64 个 batch 刷新;全量集合可被抽样,但有限步数不保证每个 session 都被访问。论文应记录这一训练预算。 +- +-资源规划:OMG 两进程建议预留 8 CPU 核、32 GiB 主机 RAM;TextOP 建议至少 4 核、16 GiB RAM。OMG 每个 rank 的 train/val 各有 2 GiB 数组 LRU,另有 512-session 特征缓存和完整索引;2 GiB 不是整个进程内存上限。仅数据回放进程的 RSS 约 3.9 GiB;这不包含训练模型、Whisper、验证集与优化器,不能把它当作完整训练的 RAM 占用。ASR SQLite 共用 2 GiB 有效记录预算;原始 tar 齐全时无需网络 range 回退。无需生成约 430 GiB 的全量派生 NPZ。 +- +-数据盘在已有权重和环境之外建议预留 15–20 GiB。TextOP 输出约 1.2 GiB;OMG 最新 3 个周期 checkpoint 加最终 last.ckpt 约 4 GiB,多卡不会各保存一套 checkpoint。其余用于缓存、SQLite 页/WAL 和保存瞬间空间。当前约 44 GiB 空闲足够这份规划。原始 tar 约 49.8 GiB+索引保存在指定共享盘,已下载完成;共享盘 df 的整体空闲仍不能视作账号配额。 +- +-本轮已终止的 OMG batch=512 实跑暴露出旧配置的问题:64-session 缓存无法覆盖一个 512-session 槽位池,普通 batch 也会重复构造 session 及运动特征;每 64 个 batch 重新抽 session 时还需要处理 ASR,导致运行近 10.5 小时只完成约 500 步。配置已改为 512-session 缓存、每 256 个 batch 刷新一次。此前 session 回放基准(ASR 命中磁盘缓存)显示,缓存 512、数组预算 2 GiB 时首次准备约 50 秒;另一轮文件读取较冷的回放约 188 秒。后者包含不同缓存状态,不能把首次准备时间差全归因于代码优化。热缓存下 batch 准备约 1.7 秒;同样缓存容量、旧逐窗口音频实现的整个 batch 准备约 7.5 秒,改为 batch 向量化后降至约 1.7 秒。双卡真实 train mode 短测中,模型计算约 0.43 秒/步,缓存命中时完整 optimizer step 约 2.0 秒,首次装载一个 512-session 池约 72 秒;这些测试不包含完整 refresh 周期。按 100,000 步估算,纯缓存命中下约 56 小时,实际时间还取决于每 256 步的刷新、ASR 命中率和共享盘读取速度。正式重启前仍应做 3–5 个 optimizer step 的端到端短测,再估计总耗时。 +- +-OMG 每 2,000 步每 rank 抽取 16 个验证 batch(双卡共 32 个),TextOP 每 2,000 步运行原生 eval_steps=12(3 组四 primitive batch)。这是训练监测,不是全量独立测试集评测。统一入口目前仍未提供恢复优化器及步数的 resume;`weights.init_checkpoint` 仅表示初始化。断点恢复和按需 run 的论文评测入口是后续工程项,不能把重新加载初始化权重当作续训。 +- +-## 可复查的证据 +- +-小型 JSON 结果在 [formal_training_readiness.json](../reports/formal_training_readiness.json)。本机测量样本和日志位于 `/root/autodl-tmp/cache/hfd/readiness/`,不进入 Git,也不混入 `runs/hfd-final`。 +- +-重建索引检查(不导出 session、不启动训练): +- +-```bash +-/root/autodl-tmp/envs/omg/bin/python prepare_sessions.py --config configs/sessions.yaml --audit-only --report /root/autodl-tmp/cache/hfd/train_audit.json +-/root/autodl-tmp/envs/omg/bin/python prepare_sessions.py --config configs/sessions.yaml --audit-only --report /root/autodl-tmp/cache/hfd/val_audit.json --set split=val +-``` +- +-迁移时替换 Python、数据盘和原始 tar 路径;详见 [迁移说明](MIGRATION.md)。`--audit-only` 通过只说明索引构造与配比可行,不保证全部 tar、ASR 或训练就绪。全量 tar 已校验、独立 train/val 索引已构造;本轮正式 batch 的短测结果见本文开头及报告。长时吞吐和论文独立测试集评测须另行衡量。 +diff --git a/docs/GENMO_FORMAL_TRAINING.md b/docs/GENMO_FORMAL_TRAINING.md +deleted file mode 100644 +index d2b273f..0000000 +--- a/docs/GENMO_FORMAL_TRAINING.md ++++ /dev/null +@@ -1,150 +0,0 @@ +-> 本文保留旧正式配方的记录。当前默认训练入口采用 [GenMo v2](GENMO_V2_TRAINING.md),包括执行段损失、预测历史恢复、CPU预取及闭环选权重;旧配置见 `configs/genmo_legacy.yaml`。 +- +-# GenMo 原生英文 HFD 训练 +- +-2026-10-24:本轮100K正式训练已完成,耗时约85.8小时。best为92K,已对官方初始化、92K和最终100K按既有六类验证规则完成闭环、整段GMR及有声视频评测。当前六段结果以92K略好,后退/移动任务仍明显不足;完整数值与视频位置见[正式评测报告](HUMAN_FORMAL_VALIDATION_20261024.md)。以下命令和训练约定保留;新正式权重未修改或删除。 +- +-正式入口为 `train_genmo_full.py --config configs/genmo.yaml`。旧 `train_genmo.py` +-和 `run_genmo_overfit.py` 仍是单session检查入口,不用于全数据训练。 +-沿用锁定版本的原生 GEM `training_step`、regression→diffusion 顺序、AdamW、151D表示、 +-冻结预训练统计量及原生3D损失;适配层保留此前验证的坐标/位移逆变换、历史锁定和未来评分mask。 +- +-## 数据范围与输入 +- +-只使用原生英文 HFD:155,117条train候选session、4,719条val候选session。 +-第三方生产包目前只有G1时间线,没有经过验证的配对SMPL-X session,明确排除。 +-因此不能把这次训练与包含第三方数据的OMG/TextOp结果直接作为同数据量架构对比。 +- +-每次按相同索引、原生seed组装音频和G1轨迹,同时观察同次执行的源选择、裁剪、 +-持姿、对齐、过渡和拼接,依据权威配对表构造人体轨迹。源SMPL-X SHA256逐条核验, +-所有原生记录及stationary源的配对ID在首次使用时检查;未支持的操作或缺失源直接失败。 +-全双工人体session不是从G1反解出来的,也不要求预先把整套session解压存盘。 +- +-每条已组装人体session的30Hz重采样只计算一次并受CPU缓存容量限制。决策点同时位于 +-50Hz和30Hz网格;历史切片的插值支撑均不晚于决策时刻,末端无效目标仍保持原50Hz终帧。 +-优化前后实际batch32的前4步loss逐值一致;未来动作扰动和末端mask回归检查通过。 +- +-输入为已知/已执行的2秒人体动作历史,以及当前已释放的最近30秒英文GT前景文本。 +-词时间取现有离线对齐缓存,guard=.1秒,0.6+.4k秒释放,motion到microphone偏移.2秒。 +-冻结T5-3B以BF16编码,原生50token预算超长时保留最近token;空文本特征为零。 +-编码器常驻当前GPU,CPU特征缓存有256MiB上限。无需ASR、翻译或重新跑强制对齐。 +-本版本保留已采用的文本+动作历史条件,不混入独立的past-audio实验。 +- +-输出为未来2秒人体动作,30Hz,每次执行12帧(.4秒)再规划。不输入GT阶段、 +-listen/motion/idle标签或未来音频。训练窗口可有初始重复姿态和末端无效目标, +-已知历史不计入未来几何损失;未知边界位移保留监督。 +- +-## 采样、验证、续训 +- +-在原生session的合法.4秒决策窗口上带放回抽样,使用保守上界拒绝采样维持窗口边际。 +-一个槽位绑定一条session,默认16批后刷新槽位,每批重新抽该session中的决策点。 +-训练按step预算运行,epoch不表示无重复遍历全部session。 +- +-每2K步进行固定验证:原生GEM双模式目标、固定窗口和噪声,无梯度;网络保持eval, +-只切换diffusion调度器的训练目标标志,防止原生图像验证入口不兼容或BN统计更新。 +-该loss不是连续推理或任务成功率,应同时使用下面的retarget视频评测。 +-默认每次16批×16窗口,验证每批刷新session,使用固定的256个窗口;这仍是验证子集, +-不是遍历4,719条候选session。训练池的16批复用策略不用于验证。 +-GEM的时间步采样使用NumPy,故验证同时保存、固定并恢复Python/NumPy/Torch随机数, +-不能仅固定Torch高斯噪声。同一真实checkpoint重复验证两次loss均为7.5329627991, +-仅用于验证复现性,不代表正式训练效果。 +- +-checkpoint保存模型、AdamW、按step的cosine调度器、RNG和已消费的采样池状态。 +-先写临时文件再原子替换,保留最近两份和最佳验证权重。完整checkpoint单份约5.1GiB, +-默认放共享盘 `models/genmo/hfd-english-train`,日志放本地 `runs/hfd-genmo-english-train`。 +-常态约需3份权重空间,保存时还需要临时文件空间;共享盘写入错误会中止并保留原checkpoint。 +- +-续训核对输入配置、模型适配实现、原生配方、离线文本缓存、配对表和body model身份。 +-只允许在原总步数预算内续训;max_steps不是“再训练多少步”,不能直接修改总预算或学习率后冒充恢复。 +-可更换GPU,允许把同一checkpoint分支到新的空输出目录。 +- +-```bash +-cd /root/autodl-tmp/hfd-baseline-portable +-/root/autodl-tmp/envs/genmo/bin/python -u train_genmo_full.py --config configs/genmo.yaml +- +-# 中断后恢复完整状态 +-/root/autodl-tmp/envs/genmo/bin/python -u train_genmo_full.py --config configs/genmo.yaml --resume auto +-``` +- +-当前总预算100K步,20K/40K/60K为阶段检查点,不表示已验证必须训练到100K才收敛。 +-学习率2e-5起步,按全部100K步做cosine衰减,在100K降到1e-6;`T_max`直接读取 +-`training.max_steps`,20K不再是学习率衰减终点。BF16、GPU1、batch32,合计320万次窗口采样。 +-采样仍有放回、槽位每16批刷新,不保证每个session都出现,不把100K解释为固定epoch数。 +-每2K步固定验证和保存,保留最近两份及最佳验证checkpoint;阶段检查应及时运行连续评测, +-20K/40K/60K权重不会永久保留。评测需单独使用空闲GPU或在训练暂停后进行,避免抢占显存。 +-验证或rollout停止改善时可提前停止,并从最佳checkpoint评测;当前不自动按val/loss早停, +-因为它不是语义成功率。100K是总上限,不是达到20K后再额外训练100K。 +-若中断,用同一100K配置恢复;旧20K协议checkpoint不能通过修改YAML直接兼容续训。 +-本次预算调整时正式输出目录尚未创建、权重目录为空,无旧正式训练需要迁移。 +- +-实测含T5-3B的GPU峰值约23.7GiB。热批次数据准备约4.9秒降到0.05–0.07秒, +-计算约0.4秒;换session池仍有组装开销,不能把热批次加速当作整轮训练的加速倍数。 +-不会自动开始长训练。 +- +-## 连续推理、retarget与视频 +- +-```bash +-/root/autodl-tmp/envs/genmo/bin/python -u evaluate_genmo.py \ +- --run /root/autodl-tmp/runs/hfd-genmo-english-train \ +- --output /root/autodl-tmp/runs/genmo-english-validation \ +- --gpu 1 +-``` +- +-默认读取最佳完整checkpoint,六类原生英文val各选首条6–30秒样本,可用 `--checkpoint PATH` +-和 `--modes turn-taking interruption ...` 显式指定。样本选择在推理前,使用与训练相同的 +-配对provider与文本对齐;只初始化一次GT历史,随后反馈预测,不按窗口重置GT。 +-推理中的指令显示时间读取实际session元数据;GT文本出处传递到retarget与视频,避免误标ASR。 +- +-每条完整预测和完整人体GT各retarget一次,保留三种比较:预测→G1/原G1、人体GT→G1/原G1、 +-预测→G1/人体GT→G1。retarget误差是独立诊断,不能从模型误差直接相减。 +-视频保留原混合音轨与共同microphone前缀。默认生成三条视频,属于运动学回放,不是物理跟踪。 +-原生task源按train/val隔离;val用于训练监测,不是独立test,预训练数据重合未审计。 +- +-## 2026-10-20 训练链路复核 +- +-可复用检查入口: +- +-```bash +-/root/autodl-tmp/envs/genmo/bin/python -u tools/audit_genmo_training.py \ +- --config configs/genmo.yaml --gpu 1 \ +- --output /root/autodl-tmp/setup/genmo-training-audit-20261020/audit.json +-``` +- +-此次检查对六类交互各取首条合格train session,共核对483个决策点,其中465个非空文本。 +-逐点重建文本并核对词末时间、guard和离散释放时间,没有违反标注释放规则; +-相对强制对齐词末的披露延迟为0.110875–0.48秒,一处不合格对齐采用整段结束后释放。 +-强制对齐是事后估计,0.1秒guard不是人工验证的误差上界;这仍是oracle文本实验。 +-六条样本最长29个T5 token,未发生截断,不代表全量语料截断率为零。 +- +-实测改变未来GT位置和姿态,regression预测逐值不变;清空文本后未来预测发生变化 +-(标准化151D特征的平均绝对变化0.05681275;不是语义成功率)。 +-regression和diffusion分别反向传播,文本投影、动作输入投影、主干和相关输出头均有 +-有限非零梯度;固定历史的原始姿态输出梯度为零,未知边界位移保留梯度。 +-执行一次用于检查的diffusion优化器更新后,文本/动作输入投影权重及bias均变化; +-T5参数保持冻结且无梯度。检查不保存模型权重,不代表收敛或闭环质量。 +-配套44项回归检查通过;报告位于上述output路径,日志在同级目录的 +-`genmo-training-audit-20261020.log`。 +- +-此次修复:位移目标i表示i→i+1,原生编码会复制窗口末段位移。当前屏蔽不存在下一有效帧 +-的位移和静止标签,包括窗口末帧及session尾部填充边界,保留有效姿态和历史→未来边界 +-位移监督。该修改改变训练目标mask和实现身份,旧实验loss不可直接逐值对比,旧完整状态 +-不能伪装为新协议resume。此次修改前未启动GenMo正式训练。 +- +-评测新增retarget之前的人体根朝向、根速度误差,与已有身体关节旋转及根位置误差一起报告; +-`evaluate_human.py`的`human_metrics`及`evaluate_genmo.py`摘要都会保留这些结果。 +-它们不依赖原G1轨迹;retarget后三种比较仍用于单独诊断。误差均不做根对齐或DTW, +-排除GT bootstrap。速度RMSE按三个坐标分量求均方根,单位m/s。 +- +-本设计适合“已释放文本+动作历史”的条件生成对照,存在以下能力/衡量限制: +- +-- 文本释放前没有语音起点信息,不能要求模型准确判断倾听开始;语调、韵律和背景声也不可见。 +-- 最近30秒是文本候选范围,实际最多50个T5 token;没有显式词时间或当前说话状态输入。 +-- 训练历史为GT,闭环历史为预测;未来2秒内的新指令尚不可见,目标天然存在不确定性。 +-- 输出预测2秒,每0.4秒重新规划执行;原生位置损失使用GT朝向积分,实际推理使用预测朝向, +- 因而必须同时看根朝向和连续rollout,不能只依据val/loss判定部署误差。 +-- 原生loss沿用全窗口均值乘mask,尾部有效目标少时loss会偏小;固定验证的窗口/噪声可比, +- 该loss不能直接当成“每个有效未来帧”的误差或跨模型通用分数。 +-- 现有指标衡量运动学轨迹一致性,尚不衡量独立定义的指令成功率、停止延迟或物理跟踪稳定性。 +- +-当前结论是结构与梯度检查通过,可开展该文本条件baseline训练;输入信息不足造成的反应差异 +-不能简单归为训练未收敛。按既定设置保留文本条件,音频条件仍作为另行定义的实验。 +diff --git a/docs/GENMO_HFD_OVERFIT.md b/docs/GENMO_HFD_OVERFIT.md +deleted file mode 100644 +index 805e6b3..0000000 +--- a/docs/GENMO_HFD_OVERFIT.md ++++ /dev/null +@@ -1,105 +0,0 @@ +-# GenMo / GEM-SMPL 的 HFD 数据适配与过拟合实验 +- +-清理说明:本文记录的一次性实验脚本已归档,复现前请按[脚本恢复说明](WORKSPACE_CLEANUP.md)取回;历史指标与日志保留。 +- +-新增原生过去音频分支的匹配实验见 [GENMO_PAST_AUDIO_TEST.md](GENMO_PAST_AUDIO_TEST.md)。音频输入已接通,当前小样本表现未优于 ASR-only,未替换本文工作版本。 +- +-本轮使用官方 `NVlabs/GENMO` commit `16bebf402d8893184249ee206d957b8248cd8310`、`nvidia/GEM-X` 的 **`gem_smpl.ckpt`**。模型源码没有修改;适配代码位于 portable 仓库。官方训练入口是 `scripts/train.py exp=gem_smpl`,本地桥接入口 `train_genmo.py` 直接继承官方 `GEM.training_step`、`train_step`、AdamW 和 regression→diffusion 执行顺序,使用原生 Lightning Trainer、噪声过程、151D 编解码与 3D 损失。 +- +-## 当前验证范围 +- +-只验证一条 train split 的双指令 interruption session:接别人递来的物品,随后切换到捡地上物品。输入、音频与此前 OMG/TextOp 的 `train-002203-seed42` 相同。human session 为 740 帧 @50Hz;连续导出覆盖 444 帧 @30Hz,视频 14.8 秒。首帧为已知初始化,其余帧闭环生成。此实验不是全数据训练、未见样本评估或物理跟踪。 +- +-原 human 包仍是源动作。`prepare_genmo_session.py` 使用 `hfd_bridge/human_sessions.py` 观察同一次原生 HFD recipe 执行,记录 source ID、精确采样时间、随机裁剪、listen 持姿帧、过渡 envelope、截断/恢复和 append 计划。人体端从 SQLite/tar 索引读取 SMPL-X,不批量解压;根据权威配对表和 SHA256 对应源动作。平移按秒插值,旋转按 SO(3)/SLERP 处理,使用人体自身的根高度和 pelvis offset 重新对齐。idle 采用局部旋转残差,listen 采用所选人体持姿与连续过渡,整个 session 保持同一个 16D shape。 +- +-原生 G1 复现与旧结果的 PCM、控制标签和决策索引逐值一致。受环境浮点实现影响,G1 数组最大差异为 `2.3841858e-7`;输出保留旧 G1 数组作为参考。该误差已记录,不宣称 G1 数组逐字节一致。50Hz/源时间的末端舍入最多允许一个 50Hz 帧的边界保持,实际保持时间写入计划,超过即报错。其他任务的几何/事件组合尚未逐一验收;观察器遇到无法追踪的姿态会拒绝导出。 +- +-## 输入与因果边界 +- +-- 原生 GEM-SMPL 151D:21×6D body rotations、10D shape、camera/gravity 两组根旋转、3D 根局部逐帧位移。沿用 checkpoint 的统计量,不另拟合 HFD 归一化。 +-- 使用相同 Whisper 因果 sidecar 的 `stable_latest_sentence`,经原生 T5-3B 编码,最多 50 tokens。冻结编码器,以 BF16 计算并缓存 FP32 特征;空文本缓存为零。缓存核对原 session 哈希。 +-- ASR 释放仍为 `0.6+0.4*k` 秒,动作决策为 `0.4*k` 秒;保留既有 HFD 三时钟约定。每次固定预测未来 60 帧,执行 12 帧,再重规划。 +-- 对照输入历史为 12 帧(0.4s)或 60 帧(2s)。历史包含当前状态。闭环仅使用实际已生成的历史,不在每次规划读取真值;首帧不足的历史重复该已知姿态。 +-- 视觉、2D、相机、music、speech 分支全部标缺失。没有未来音频、GT 指令、listen/motion/idle 标签、未来 Stop、任务进度、GT turn ID 或 GT 总时长作为条件。 +-- 固定虚拟坐标映射为 AMASS Z-up→GEM Y-up,camera 和 gravity 使用同一固定坐标;平移转换包含 shape 对应的 SMPL-X pelvis offset,导出时返回 Z-up。 +-- 原生根“速度”使用前向帧差。历史单独重采样/编码,最后一个已知姿态的下一步位移保持未知;其 mask=0,作为未来边界监督。不能使用第一帧未来 GT 计算该历史特征。 +- +-沿用原生 observed-motion overwrite,不添加新的 mask embedding 或网络层。已知前缀在网络输入及输出均固定;原生根位移 loss 会积分整段窗口,因此仅锁输入、却让任意历史输出补偿未来位移是不一致的。输出固定通过无参数 forward hook 实现,并覆盖训练和 DDIM 采样。未知边界位移保留梯度。 +- +-## 训练损失与执行 +- +-保留原生 simple MSE、root-relative 3D joints/437 vertices、world translation、static-contact BCE;缺少真实图像时关闭相机平移和 2D/投影损失。simple loss 的 mask 排除复制的已知特征;几何/contact loss 只对未来有效帧评分。注意原生 masked mean 仍除以固定窗口尺寸,因此**不同历史长度的原始 loss 数值不宜直接比较**。比较应使用同样长度的闭环输出指标。 +- +-full model 共约 518M 参数,原始权重 709 个 state tensors 全部严格匹配。motion 模型全部微调,T5 冻结;batch=2,BF16,AdamW,gradient clip=0.5。本轮先做 3 步训练烟雾检查,再探索 300 步及低学习率继续训练。排查完 prefix 输出一致性后,两种历史长度都从同一官方权重、相同 seed=42、lr=2e-5 开始,分别训练 600 步。`locked-h12-fit600` / `locked-h60-fit600` 是用于比较的修正版本;此前 `pilot`、`fit300`、`fit1300`、`history60-fit600` 是排查记录。 +- +-## 位移表示往返检查 +- +-固定版本的 `get_local_transl_vel` 文档写 R^T,但实际代码为 R×world_delta;`rollout_local_transl_vel` 也乘 R,两者不互逆。真值完整 session 的原生往返误差约 0.01168m,正确逆变换 R^T 的误差约 6.4e-10m。桥接保留原生编码/预训练特征定义,修正解码与 world-translation loss 的积分方向;不修改上游 checkout。随机旋转的往返测试防止仅在近似单位旋转上“偶然正确”。 +- +-`final-h12` / `final-h60` 是输出前缀锁定后的匹配 600 步对照,保留了原生积分方向,用于记录历史长度实验。发现积分问题后,以 h60 的 600 步权重为初值,lr=5e-6 再训练 300 步(AdamW 状态重新初始化),保存为 `corrected-h60-fit900`;最终复现入口默认采用正确逆变换。`initial-corrected-h60` 是相同解码规则下的官方权重基线。修正这项确定性错误不等于消除了自由闭环的速度偏置/误差累积。 +- +-## 重要发现:短历史存在观测歧义 +- +-`input-ambiguity-h12.json` 找到两对完全相同的文本和历史姿态: +- +-| 动作时刻 | 当前文本 | 两份未来 body rotation 标签的 RMSE 差异 | +-| --- | --- | --- | +-| 2.8s / 3.2s | Taking an item from someone | 0.230 rad | +-| 7.6s / 8.0s | Picking an item up from the floor | 0.264 rad | +- +-同样的输入不能唯一确定两份不同的未来真值,不能把无法精确过拟合都归为优化问题。2s 历史在这条样例中消除了上述精确重复,但这不证明所有任务已经可观测,也不保证模型能记住更长的语义状态。若仍有事件响应问题,后续应单列原生 past-audio 或因果状态记忆实验,不能注入 GT 阶段计时器。 +- +-## Retarget 与视频 +- +-`infer_genmo.py` 导出标准 AMASS SMPL-X。GEM 不预测手指,预测端使用零残差手姿态;known shape 使用 session bootstrap 的 16D betas,GEM 原生特征取前 10 维。human GT 保留实际手残差。评测使用 `flat_hand_mean=False`、`use_pca=False`、`ext=npz`,与 human 包约定一致。 +- +-`evaluate_human.py` 将**完整连续轨迹**各 retarget 一次到 G1,每条轨迹使用独立、跨帧连续的 GMR IK 状态。GMR commit `bb1bbe40774794fceb2a7c579a3464a28e68c844`,solver=daqp,`offset_to_ground=False`。仅稀疏检出 GMR package 与 Unitree G1 资产,约 79MB;AutoDL 安装依赖失败后通过清空子进程代理的阿里云 HTTPS PyPI 国内镜像直连完成依赖安装。 +- +-同时保留三组比较和视频:prediction→G1 vs 原 G1、human GT→G1 vs 原 G1、prediction→G1 vs human GT→G1。人体真值转换与原 G1 的关节 RMSE 约 0.202 rad,是几何转换/表示的诊断,不能从模型误差中直接相减。视频沿用 `render_overfit.py` / OMG MuJoCo renderer,含原始混音声音,展示真实输入的因果 ASR 和初始化/预测历史标记。 +- +-## 文件和复现 +- +-实验根目录:`/root/autodl-tmp/runs/archive/legacy-20261020/genmo-overfit`。模型训练权重放 `/root/autodl-fs/pangxinlei/models/genmo/hfd-overfit`,数据盘只保留软链接。源数据仍保留 tar;原生 GenMo 和 GMR checkout 没有修改。 +- +-配置入口为 `configs/genmo_overfit.yaml`,一键运行: +- +-```bash +-cd /root/autodl-tmp/hfd-baseline-portable +-/root/autodl-tmp/envs/genmo/bin/python run_genmo_overfit.py \ +- --config configs/genmo_overfit.yaml \ +- --output /root/autodl-tmp/runs/genmo-new-overfit \ +- --prepared-session /root/autodl-tmp/runs/archive/legacy-20261020/genmo-overfit/session \ +- --text-cache /root/autodl-tmp/runs/archive/legacy-20261020/genmo-overfit/text.pt +-``` +- +-不提供后两个参数时,会重新准备同一条 paired session 和文本缓存。输出目录必须是新的。此 runner 调用的只是本轮已逐步验证的准备、训练、推理、retarget、渲染入口;组合 runner 自身不额外启动一轮重复训练。 +- +-契约检查:`tests/test_genmo_contract.py`、`tests/test_human_eval.py`。真实输入/梯度审计为 `tools/audit_genmo_inputs.py`:扰动全部未来 GT 后,observed motion、mask、f_cond/uncond/empty 和 T5 条件逐值不变;原生 regression/diffusion 两模式实际反向,532 个 denoiser 梯度 tensors 有限且非零。另检查 SO(3) 跨 ±π 插值、坐标/pelvis roundtrip、前缀损失掩码与未知边界位移梯度。 +- +-## 本轮结果(2026-10-18) +- +-正式比较使用同样的 2s 历史、固定坐标和修正后的位移逆变换,单次已知初始化后闭环生成,不使用真值重置: +- +-| 设置 | 对原始 G1 的 29 关节 RMSE (rad) | 根位置每坐标 RMSE (m) | +-| --- | ---: | ---: | +-| 官方权重 | 0.4786 | 0.3804 | +-| 微调 600+300 步,自身历史 | 0.1998 | 0.1285 | +-| 同一权重,真值历史诊断 | 0.1857 | 0.0125 | +-| human GT 经同一 retarget | 0.2016 | 0.0093 | +- +-最终模型与 **同样 retarget 的 human GT** 比较为关节 0.1384 rad、根位置 0.1236m;这更直接衡量相同转换后的生成误差。人体原始空间的闭环根位置每坐标 RMSE 为 0.1426m,真值历史诊断为 0.0070m。曲线显示根位置存在随时间积累的偏差,不能将本轮称作完整成功过拟合。动作阶段也仍有偏差。 +- +-前缀输出锁定、原生积分方向下的匹配 600 步对照:0.4s 历史的 G1 关节 RMSE=0.3420、2s 历史=0.1975 rad;两者根位置误差均约 0.34–0.35m。这组仅用于记录输入长度实验,与修正后 900 步结果分开。 +- +-结论:**数据适配和完整推理评测链路通过;当前小样本闭环过拟合未完全通过。** 后续建议从 2s 历史开始,分别验证闭环预测历史训练、原生 contact 后处理或显式因果状态记忆;每一项单独标注适配和对照,不加入 GT 任务阶段或手工真值位置校正。增加其他任务前,先冻结并验收对应人体/G1 session。 +- +-结果入口: +- +-- `/root/autodl-tmp/runs/archive/legacy-20261020/genmo-overfit/README.md`、`summary.json`。 +-- 主视频:`final/evaluation/rollout.mp4`;相同 retarget 比较:`final/evaluation/retargeted_human_comparison.mp4`;转换诊断:`final/evaluation/retarget_reference.mp4`。 +-- 最终三条视频均为 444 帧、30fps、14.8 秒,原始混音声音已解码核验;`videos.json` 保存详细检查。人工查看了 listen、第一动作、打断后的 listen、第二动作的代表帧。 +-- `trajectory_diagnostics.png` 展示关节误差和根漂移;视频是参考姿态播放,不代表物理控制成功。 +-- 最终权重:`/root/autodl-fs/pangxinlei/models/genmo/hfd-overfit/corrected-h60-fit900/fitted.ckpt`。 +-- 14 个相关契约测试通过;完整官方权重严格加载、真实 regression/diffusion 反向、未来 GT 扰动后的采样输出逐值不变、真实 GMR 和音视频集成已验证。GPU 训练峰值约 9.17GB。ASR/T5 为预计算缓存,采样耗时不是端到端在线延迟。 +- +-复现最终继续训练阶段:在 portable 目录执行 `GENMO_HISTORY_FRAMES=60 /root/autodl-tmp/envs/genmo/bin/python train_genmo.py --session /root/autodl-tmp/runs/archive/legacy-20261020/genmo-overfit/session --text-cache /root/autodl-tmp/runs/archive/legacy-20261020/genmo-overfit/text.pt --checkpoint /root/autodl-tmp/runs/archive/legacy-20261020/genmo-overfit/locked-h60-fit600/fitted.ckpt --output /root/autodl-tmp/runs/genmo-reproduce900 --steps 300 --batch-size 2 --lr 5e-6`。当前 `run_genmo_overfit.py` 默认是从官方权重直接运行修正后的 600 步实验配方,与本轮分阶段 900 步训练的优化器历史不同,应在报告中区分。 +diff --git a/docs/GENMO_LISTENING_DIAGNOSIS.md b/docs/GENMO_LISTENING_DIAGNOSIS.md +deleted file mode 100644 +index a7ab5de..0000000 +--- a/docs/GENMO_LISTENING_DIAGNOSIS.md ++++ /dev/null +@@ -1,25 +0,0 @@ +-2026-10-24:针对GenMo多个episode长期保持倾听姿态,使用当前92K权重做离线对照。正式权重未修改,训练损失、采样配方和推理规则均未改动。 +- +-已经确认:失败发生在人体生成阶段,早于retarget。按词释放的英文文本确实进入模型,模型也能产生变化的未来特征;当前证据不支持“没有读到文本”或“整个网络已经输出常数”。主要暴露的问题是:GT历史训练下能延续运动,反馈自身预测时却容易停留在倾听状态,动作启动不足。 +- +-| 后退样本val-backchannel-000057 | 人体姿态RMSE(rad) | 根位置RMSE(m) | 整段根路径(m) | +-|---|---:|---:|---:| +-| 正常预测历史闭环 | 0.536 | 1.252 | 0.136 | +-| 每0.4秒使用真实历史,仅作诊断 | 0.155 | 0.030 | 4.702 | +-| 人体GT | — | — | 3.652 | +- +-GT历史对照会不断重置条件及空间位置,因此位置误差天然受保护,其速度误差反而较大;它不能作为正式评测或有效修复。该对照说明运动历史对结果影响很大,并不证明模型能独立完成后退。 +- +-同一8秒决策点、相同英文caption和噪声,单窗口把真实运动历史改为固定倾听历史:CFG1时未来两秒的根路径由0.266m降至0.011m;CFG2.5时由0.206m降至0.093m。改变历史会明显影响运动启动。直接清空文本也会改变预测:t=999时未来151D特征平均绝对差约0.00963,真实文本特征范数约17.05;此时未来特征的帧间RMS约0.0271。因此不能把倾听姿态卡住等同于编码器完全失效。 +- +-曾怀疑原生CFG的无文本分支移除历史条件特征。实际检查当前`in_attr`不含该独立历史特征,历史主要由`motion_mask_3d/observed_motion_3d`约束;强制让两分支共用`f_cond`后,单窗口和完整闭环输出均保持相同。该候选原因已经排除,没有据此修改生产推理。 +- +-训练数据抽查来自真实`train_sampling`记录的最早、约三分之一、三分之二和末尾四个池,共128个窗口。按根速度>0.1m/s或平均身体角速度>0.3rad/s、活动帧占比至少25%分类:80个窗口的两秒未来有活动,63个窗口的实际12帧执行段有活动,49个窗口的历史较安静,其中13个执行段转入活动,122个窗口有非空文字。此抽查不是全语料统计,但排除了“训练全部读成静止样本”的解释。起动转换仅占该抽查约10.2%,不能据此推断全库比例。 +- +-当前训练完整地监督未来60帧,闭环只执行最前12帧;全部历史来自GT,没有预测历史混入。模型一旦在起动处继续预测倾听,下一次重规划又看到自己的倾听历史;caption不变时容易形成持续保持姿态的闭环。近真实未来加噪的去噪损失下降,不能保证从安静历史自主启动动作。这是目前最直接的失效机制,具体损失权重、历史课程、数据分布的独立贡献尚未完成受控消融。 +- +-后续训练策略应优先验证:提高已执行12帧及起动窗口的监督占比,加入真实解码运动的变化监督,并用train-only的预测历史训练恢复能力。输入仍保持原协议,不能用GT阶段、未来文本或评测时真实历史替代这些训练工作。本轮没有直接启用这些未验收配方,也没有重新训练100K。 +- +-本地诊断代码在被忽略的`test/diagnostics/`,数值与中间轨迹在本机cache;GitHub仅保留本文结论和复现当前正式训练所需的代码。当前训练源快照见`TRAINING_SNAPSHOT.json`及对应Git历史;本次迁移改动仅处理路径、资源定位和训练依赖拆分,没有改变训练算法。 +- +-后续已实现的训练适配改动见 [GenMo v2](GENMO_V2_TRAINING.md)。本文诊断数值仍对应旧正式权重,不能当作新策略全量训练后的效果。 +diff --git a/docs/GENMO_PAST_AUDIO_TEST.md b/docs/GENMO_PAST_AUDIO_TEST.md +deleted file mode 100644 +index 4d9c162..0000000 +--- a/docs/GENMO_PAST_AUDIO_TEST.md ++++ /dev/null +@@ -1,93 +0,0 @@ +-# GEM-ASR-past-audio:原生音频分支测试 +- +-2026-10-18。在已有 GenMo/HFD bridge 中启用官方 speech encoder,使用同一条冻结 interruption session、同一因果 ASR、同一人体真值和 GMR 链路。不是新增语音语义模型;ASR 文本仍经 T5-3B,音频由 checkpoint 自带的卷积前端编码。 +- +-## 输入构造 +- +-第 k 次决策的动作时刻是 `t_k=0.4*k`,按既有 HFD 协议,PCM 释放点是 `r_k=0.6+0.4*k`。取**释放点之前**的 `[r_k-2,r_k)` 两秒混合 PCM;不能把动作轴、PCM 轴和墙钟直接视为同一个零点。这里保留 OMG/TextOp/HFD 的释放规则。 +- +-1. 从原 16kHz session 混音按采样点截取,右端严格为 `9600+6400*k`。启动时不足两秒的左侧补静音,只有真实覆盖的历史位置 availability=1;首个决策实际覆盖 18 个音频 bin。 +-2. **先截断,再重采样和卷积。** SciPy polyphase FIR 按 9/8 重采样到 18kHz,保持波形幅度。未来样本没有进入 FIR 或卷积。 +-3. 原生六个 BasicBlock 将 `[B,1,36000]` 编码为 `[B,128,60]`;stride=600 samples,恰为 30Hz。卷积感受野为 18,635 samples(约 1.035s)。所有卷积支持都限制在本次已释放窗口及边界补零内;不是在完整 session 上先提特征再切片。 +-4. **编码后**补 60 帧未来特征,再用原生 `audio_embedder`、`cond_exists_embedder` 和 CFG 逻辑融合。120 帧条件中只有历史音频位置可用,未来 audio availability 恒为 0。未来的 f_cond 数值可以有“缺失条件”的学习偏置,但与波形内容逐值无关。 +-5. 特征按时间顺序放入 60 个动作历史槽位。卷积的中心时间为 `r_k-2+j/30`,最后一个中心比释放点早 1/30 秒;保留原生卷积的 phase,没有为对齐而读取释放点之后的样本。原生训练 BatchNorm 会在已截断的 batch/time 内统计;推理使用 eval/running statistics。 +- +-其它输入保持:60 帧已执行动作历史+T5 当前稳定 ASR;预测 60 帧,执行 12 帧,30Hz。训练使用真值历史,闭环使用模型自身历史;仅首帧初始化已知。不输入未来音频、GT 指令、状态/turn 标签或任务进度。music、视觉、2D 和相机条件仍缺失。 +- +-## 实现边界 +- +-- 新增 `hfd_bridge/genmo_audio.py` 负责释放窗口、重采样和卷积后 padding。 +-- `GENMO_PAST_AUDIO=1` 且 `GENMO_HISTORY_FRAMES=60` 启用音频;默认关闭,因此已有 ASR-only 入口保留。 +-- 在原生 `GEM.create_condition_mask` 外接入已截断的波形;encoder 的无参数 forward hook 仅在卷积之后补齐 future features。没有添加新的学习层、mask embedding 或改变 709 个 state tensor 的结构。上游 checkout 保持 clean。 +-- 保留原生音频条件 dropout=0.1、BatchNorm、卷积和音频投影的训练方式。模型主体及音频前端一起微调,T5 保持冻结。 +-- 之前修正的 prefix 输出锁定、未知边界位移和实际位移编码逆变换在两组实验中一致。 +- +-## 对照设计 +- +-两组均从 `/root/autodl-tmp/runs/archive/legacy-20261020/genmo-overfit/corrected-h60-fit900/fitted.ckpt` 初始化,再各训练 600 个 optimizer steps,batch=2、lr=5e-6、seed=42。motion 模型累计微调 1500 步;音频前端在本轮训练 600 步。优化器均重新初始化。两组的窗口、文本、监督、损失权重、DDIM 10 步、CFG 和采样种子相同;音频前端 dropout 会消耗额外随机数,因此不宣称两组训练噪声逐位一致。 +- +-| 训练目录 | 输入 | +-| --- | --- | +-| `asr-control-fit600` | ASR+2s 动作历史 | +-| `past-audio-fit600` | 同上+2s 已释放混合 PCM | +- +-另对音频模型使用 `infer_genmo.py --drop-audio` 做消融:清零波形并把 audio availability 全设 0,ASR、动作历史和权重不变。视频仍播放原 session 的真实声音,标题明确显示 `AUDIO OFF`;播放声音不等于模型收到了音频。 +- +-## 验证 +- +-- 完整模型严格加载仍为 709 个 state tensors。 +-- 扰动释放点之后的原始 PCM,重采样输入逐值相同;同时扰动未来 GT,原生条件及 DDIM 输出逐值相同。 +-- 过去 PCM 改变实际条件和未来采样结果,未来 audio 条件槽与波形内容无关;CFG unconditioned branch 对波形不变。 +-- 原生 regression/diffusion 两模式实际反向,音频 encoder 与 projection 的梯度有限且非零。 +-- 训练完成后的 checkpoint 检查:ASR-only 的音频 encoder/projection state 与初值完全相同;音频组有实际更新,BN batch counter 增加 1200,符合每步 regression/diffusion 各一次前端调用。 +-- 四个音频契约测试覆盖未来 PCM 隔离、启动补零与可用帧、16→18kHz 的频率保持、拒绝未覆盖的释放点。原有六个 GenMo 契约测试在音频开/关两种进程配置下通过。 +- +-测试根目录:`/root/autodl-tmp/runs/archive/legacy-20261020/genmo-audio-overfit`,真实前端审计为 `audio-audit.json`,权重更新审计为 `trained-audio-state-audit.json`。 +- +-## 配置与运行 +- +-配置:`configs/genmo_audio_overfit.yaml`;组合 runner 为 `run_genmo_overfit.py`,设置 `past_audio: true` 会在训练和推理子进程中显式启用对应环境变量。要做匹配的 ASR-only 对照,复制配置只把 `past_audio` 改为 false,保持初值、步数和学习率一致。 +- +-本轮直接调用的训练方式(两组用不同的新输出目录): +- +-```bash +-GENMO_HISTORY_FRAMES=60 GENMO_PAST_AUDIO=1 \ +-/root/autodl-tmp/envs/genmo/bin/python train_genmo.py \ +- --session /root/autodl-tmp/runs/archive/legacy-20261020/genmo-overfit/session \ +- --text-cache /root/autodl-tmp/runs/archive/legacy-20261020/genmo-overfit/text.pt \ +- --checkpoint /root/autodl-tmp/runs/archive/legacy-20261020/genmo-overfit/corrected-h60-fit900/fitted.ckpt \ +- --output /root/autodl-tmp/runs/genmo-new-audio-test \ +- --steps 600 --batch-size 2 --lr 5e-6 +-``` +- +-权重默认写 fs 的 `models/genmo/hfd-overfit//fitted.ckpt`,数据盘只保存软链接。推理时使用同样的两个环境变量。评测和视频继续调用 `evaluate_human.py`、`render_overfit.py`,对完整连续轨迹 retarget;不在每个规划窗口重置 IK。ASR、T5 为预计算缓存,报告的采样耗时不是端到端在线耗时。 +- +-## 实测结果 +- +-以原始 G1 为参考,全部排除首帧初始化;根位置为每坐标 RMSE,未做根对齐或 DTW: +- +-| 设置 | 采样 seed | 29 关节 RMSE (rad) | 根位置 RMSE (m) | +-| --- | ---: | ---: | ---: | +-| 匹配 ASR-only 对照 | 42 | 0.2061 | 0.0366 | +-| 原生过去音频,eval/running BN | 42 | 0.2893 | 0.1030 | +-| 匹配 ASR-only 对照 | 43 | 0.1992 | 0.0261 | +-| 原生过去音频,eval/running BN | 43 | 0.2773 | 0.1096 | +-| 音频模型,关闭 speech 条件 | 42 | 0.1992 | 0.1076 | +-| 音频模型,当前窗口 BN 诊断 | 42 | 0.2380 | 0.1085 | +- +-与同样 retarget 的 human GT 比较,seed 42 的 ASR-only 为关节 0.1395 rad、根位置 0.0350m,音频版为 0.2480 rad、0.0989m,方向一致。因此,退化不能仅归因于原始 G1 与 human GT 的转换差异。`summary.json` 另存 listen/motion/idle 三类 GT 阶段的评分;这些标签仅用于评估,未输入模型。 +- +-诊断中的 `--audio-normalization window` 只改变推理时音频 BatchNorm:其它模块保持 eval、Dropout 关闭,BN 使用**当前已截断的两秒窗口**统计,关闭 running-stat 更新,不使用未来 PCM 或整段 session 的统计。这是单独的归一化适配诊断,不是默认原生推理。姿态有所改善,但根漂移没有解决,尚不足以将归一化差异认定为唯一原因。未来 PCM 扰动的逐值不变性检查针对原生 eval 推理;训练保留原生跨 batch/time 的 BN 统计语义。 +- +-**结论:当前版本的因果性、维度、权重加载、梯度和完整评测链路通过,但小样本性能没有优于 ASR-only,音频版尚未成功过拟合。** 两个采样种子都显示相同方向;它们不是两次独立训练,更不是未见数据泛化验证。暂保留 ASR-only 为工作版本,音频模型作为实验结果保存。 +- +-可能的后续检查包括:固定音频前端/BN 的训练对照、原生逐帧音频预训练与仅历史条件之间的分布差异、闭环预测历史训练。当前结果不证明原生音频能力无效,也不能证明上述某一项已经解释了退化。 +- +-## 产物 +- +-- 汇总与曲线:`/root/autodl-tmp/runs/archive/legacy-20261020/genmo-audio-overfit/README.md`、`summary.json`、`comparison.png`。 +-- 音频版视频:`past-audio/evaluation/rollout.mp4`;ASR-only 对照:`asr-only/evaluation/rollout.mp4`。 +-- 两组均另有 `retarget_reference.mp4`(human GT→G1 vs 原 G1)和 `retargeted_human_comparison.mp4`(相同转换后的预测比较)。保留原始混音声音;视频为参考姿态播放,不是物理控制跟踪。 +-- 消融和归一化诊断保存在 `audio-disabled`、`audio-window-bn`,明确标注模型收到的输入;seed 43 复核保存推理数组与完整 GMR 数值,不重复渲染视频。 +-- 训练权重:fs 的 `models/genmo/hfd-overfit/asr-control-fit600/fitted.ckpt`、`past-audio-fit600/fitted.ckpt`。数据盘训练目录内为软链接。 +-- 未安装新依赖、下载新模型或复制主体资产;沿用已准备的 GenMo、T5、SMPL-X 和 GMR。 +diff --git a/docs/GENMO_V2_TRAINING.md b/docs/GENMO_V2_TRAINING.md +deleted file mode 100644 +index 0cc76f7..0000000 +--- a/docs/GENMO_V2_TRAINING.md ++++ /dev/null +@@ -1,52 +0,0 @@ +-# GenMo v2:执行段监督、预测历史恢复与 CPU 预取 +- +-本版本针对长期保持倾听姿态的失效机制改造训练适配。模型骨干、英文前景 GT 词时间释放、最近 30 秒上下文、2 秒动作历史、2 秒预测及每次执行 0.4 秒的输入输出协议不变。没有新增音频、未来文本或 GT 阶段条件。正式旧权重保留;短程链路检查不能替代新策略的全量质量评估。 +- +-## 训练目标 +- +-保留原生 regression + diffusion 两个训练分支及姿态几何损失;全局根平移监督改为按预测朝向解码的实际轨迹,避免用GT朝向掩盖推理时的方向误差。151D 特征误差中,实际执行的前 12 帧及历史末帧到第一未来帧的位移权重设为 3。逐样本、逐特征归一化权重,保持有效监督总权重与原配方一致,避免仅仅放大总 loss。无效尾帧及未知末帧位移仍不参与监督。 +- +-另加前 12 帧的解码根速度与身体关节旋转增量 SmoothL1 监督,默认权重分别为 1 和 0.05。根位移按原生编码器的实际逆变换 `Rᵀ @ local_delta` 解码,旋转来自预测本身;不以未来 GT 朝向代替预测朝向。旋转增量在 SO(3) 上计算,避免轴角跨分支的假速度。速度项不是鼓励一直运动,正常倾听/停止的 GT 速度仍是低速或零。 +- +-## 预测历史恢复 +- +-每个训练 batch 最多选 25% 的可用窗口,保持原始窗口采样分布。前 1,000 步不混入预测历史,接下来 4,000 步逐步增加,5,000 步起达到该上限。源窗口不足 0.4 秒历史时不参与恢复。 +- +-对时刻 `t` 的恢复样本,在 `t−0.4s` 决策点用当时已释放的文字及过去动作生成一次原生 10 步 DDIM 预测,只取实际执行的 12 帧,替换 `t` 输入历史的末 12 帧。生成所需的未知未来位置填成末个已知姿态,不读 GT 未来;预测过程关闭梯度及 dropout/BN 更新。未来GT关节姿态、根旋转、位移增量及时间不变;将未来根轨迹整体平移到实际历史的末位置,保持相同的局部运动。GEM的151D输入含根速度而不含绝对根位置,因此不能要求它纠正输入中不可观测的全局位置偏移。目标编码及速度/旋转增量的边界以这份一致的历史/未来为起点,避免监督矛盾。目标编码、解码和新增时序损失使用FP32几何计算。恢复样本仍需从偏离的自身预测回到正确轨迹。 +- +-这是一段预测历史的恢复课程,不等同于对无限长闭环的保证;必要时可另做多段恢复消融,不能把评测时使用 GT 历史当作修复。 +- +-## 验证与权重选择 +- +-继续记录固定噪声 `val/loss`。默认另选两个不同 interaction mode 的固定英文验证 session,各最多 900 帧。只用一个初始真实姿态启动,之后全部反馈自身预测,文字沿原时间轴释放。训练/验证 RNG 在检查后恢复。 +- +-默认 `best` 按 `val/closed_loop_score` 选择,定义为两个 session 的 `body geodesic RMSE(rad) + root coordinate RMSE(m)` 的平均值。另保存根速度误差与预测路径长度。这是训练选权重的代理指标,不是全测试集分数,也不替代人体到 G1 retarget 后的正式评测。`last` 仍按训练步保存完整恢复状态。 +- +-## 吞吐改动 +- +-- 单独的 CPU producer 负责 session 组装和窗口构造,最多预取 16 个 batch(一个样本池周期,深度上限 32);GPU 主进程负责冻结 T5 与训练。进程间只传配置、CPU batch 和采样状态,不共享 SQLite 连接或 CUDA 对象。 +-- checkpoint 记录已消费 batch 的池/RNG 状态,忽略预取但尚未消费的数据。producer 及主进程训练/验证 DataLoader 均使用独立随机生成器,不挤占模型的扩散随机流。 +-- GenMo 的 SO(3) 插值/混合批量化;人体几何 observer 只挂接实际需要的函数返回事件,避免全进程 `sys.setprofile` 的逐函数开销。原 HFD 文件及其校验哈希不修改。此优化需通过与旧 observer 的成对 session 对照,不能只看速度。 +-- 训练 session 内存预算为 4 GiB,验证预算为 8 GiB,文本特征 LRU 为 512 MiB。另有进程及索引开销;内存较少的机器可降低这些预算,采样规则不变。 +-- 启用内容校验的本地原生 session 索引缓存,缓存的是源组合索引,不是 GT session 全量导出或模型输入。原始 tar 继续留在共享盘,离线词时间缓存、索引缓存、代码和日志位于工作盘。 +-- 保持 `refresh_batches=16`,没有通过延长同一批 session 的重复使用来虚增吞吐。GT 几何在两个训练分支间复用一次计算。NumPy/OpenBLAS默认限制为1线程,避免小矩阵使用64线程的调度开销;正式配置关闭逐 batch 进度条,保留结构化指标输出。 +- +-耗时记录新增 `cpu_batch_seconds`、`text_encoding_seconds` 和样本池刷新标记。`data_wait_seconds` 是训练批次间的主进程等待,不能直接等同于磁盘 IO;验证耗时不再计入该字段。评估速度时应跨越多个样本池刷新,并单独报告初始化。 +- +-本机稳定阶段实测约快 1.4–3.1 倍,较保守复测为 1.44 倍;冷启动和共享存储波动仍明显。完整测量口径、正确性与小规模质量限制见 [GENMO_V2_VALIDATION.md](GENMO_V2_VALIDATION.md)。 +- +-## 命令与旧实验复现 +- +-现有服务器从仓库根目录正常前台运行: +- +-```bash +-/root/autodl-tmp/envs/genmo/bin/python -u train_genmo_full.py --config configs/genmo.yaml +-``` +- +-新输出为 `runs/hfd-genmo-english-v2`,权重为共享盘 `models/genmo/hfd-english-v2`;不覆盖旧的 `hfd-english-train`。跨机器先按 [PORTABLE_TRAINING.md](PORTABLE_TRAINING.md) 生成本机配置。 +- +-新 run 的完整断点使用同配置加 `--resume auto`。如要从旧正式权重初始化新策略,使用 `--init-checkpoint /path/to/old.ckpt`;这是权重 warm start,优化器、调度与步数重新开始,不能伪装成旧 run 的续训。默认仍从官方 `gem_smpl.ckpt` 开始。 +- +-`configs/genmo_legacy.yaml` 保留旧配方;配置生成器 `--genmo-recipe legacy` 可以复现该配方。若要审计旧训练的原实现,Git `0bcc49e` 保留了与旧 GenMo/Kimodo 正式训练源码记录匹配的快照,`fa5c1a5` 是本次策略修改前的便携版。恢复旧完整断点应使用对应原配置、代码和资产。跨版本/路径改动不承诺逐位相同。 +- +-测试、基准和过拟合脚本均在忽略的 `test/`,不随训练代码发布;实验日志、权重和视频不进入 Git。 +diff --git a/docs/GENMO_V2_VALIDATION.md b/docs/GENMO_V2_VALIDATION.md +deleted file mode 100644 +index ab834f4..0000000 +--- a/docs/GENMO_V2_VALIDATION.md ++++ /dev/null +@@ -1,45 +0,0 @@ +-# GenMo v2 本机验证记录(2026-10-24) +- +-本页记录训练实现和短程实测,不能视为新策略已经完成全量训练或测试集验收。正式配置与设计见 [GENMO_V2_TRAINING.md](GENMO_V2_TRAINING.md),跨机器准备见 [PORTABLE_TRAINING.md](PORTABLE_TRAINING.md)。诊断脚本、原始日志、测试权重和视频均不发布到 Git。 +- +-## 原耗时来自哪里 +- +-旧 100K run 耗时约 85.8 小时,记录的平均计算约 0.367 秒/步,平均批次等待约 2.718 秒/步;旧等待字段还混入验证间隔,不能直接当作 IO 时间。一次冷启动的 8-session CPU profile 中,实际读取约 625.6 MB 的累计耗时为 1.027 秒,而旋转插值累计约 6.644 秒。session 组装、旋转处理及主进程串行等待也是瓶颈,单纯搬动 tar 不能解释全部耗时。 +- +-本次保持原始 tar 在共享盘、缓存和派生索引在本地盘;采用向量化几何、定点 observer hook、本地索引缓存、单 CPU producer 和最多 16-batch 预取。保持每 16 batch 刷新样本池,未通过延长同一批 session 的重复使用换取吞吐。 +- +-## 吞吐实测 +- +-同机 RTX 5090 32 GB,batch 32,各执行 64 次真实优化更新,跨越 4 个样本池。基线与优化版的 session ID、决策帧和采样顺序一致;新策略开启恢复训练,GPU 计算负担比旧策略更高。吞吐检查关闭验证,不含正式闭环验证开销;总耗时栏是训练指标计时,另有进程初始化和退出开销。 +- +-| 测量 | 前 16 步(秒) | 64 步(秒) | 第 17–64 步平均(秒/步) | 同区间计算 / 等待(秒/步) | +-|---|---:|---:|---:|---:| +-| 旧策略、串行数据 | 51.61 | 284.00 | 4.84 | 0.35 / 4.49 | +-| 新策略、预取 16,较热缓存的一次 | 91.60 | 167.05 | 1.57 | 0.68 / 0.89 | +-| 修正监督后的新策略复测 | 111.50 | 272.33 | 3.35 | 0.68 / 2.67 | +- +-稳定阶段观察到约 **1.4–3.1 倍**吞吐提升;更保守的复测是约 **1.44 倍**。包含前期等待的 64 步复测只缩短约 4%,不能承诺全量训练固定快 3 倍。共享存储/主机负载与缓存状态使样本池生成耗时明显波动,预取仍未完全遮蔽 CPU 等待。新策略两次峰值显存约 20.7 GB(十进制),基线约 25.4 GB;这不是最大 batch 搜索结果。根轨迹监督最后增加了平移重锚,最终源码另通过参数更新与恢复检查,未再重复整套吞吐基准。 +- +-## 正确性检查 +- +-- 14 个成对 session 覆盖 6 种交互模式:音频、G1、文字、时间及参考元数据完全相同;人体数组最大绝对差约 `7.45e-9`。原始 HFD 文件及哈希不变,旧 profiler 状态保持。 +-- 7 项单元检查通过,覆盖 SO(3) 插值、执行段权重总量/有效 mask/梯度、姿态键、显式与隐式返回 hook、恢复样本条件、配置迁移与资源链接。 +-- 原生编码器检查通过:扰动历史后,保持未来局部运动并平移重锚的完美预测,其新增速度、根轨迹损失接近零;错误预测有有限且非零的梯度。未来 GT 关节姿态、根旋转和位移增量保持,源 GT batch 未改写。 +-- 训练恢复以预取队列中**已消费**的样本状态为准。最终源码连续 4 步与 2 步后中断恢复至 4 步的对照通过:709 个模型张量、Adam、调度器、采样状态及 Python/NumPy/Torch/CUDA RNG 均逐位一致。对照开启预取、预测历史恢复、固定噪声验证和闭环验证;这是短程恢复检查,不是长期质量保证。 +-- 外部资产清单中 141 个文件(约 53.6 GB)的 SHA256 全部核验通过。OMG 50M/100M/300M/500M 均严格加载并以配置中的微批次执行一次丢弃的真实参数更新;这不代表各尺寸已完成 HFD 正式训练。 +-- 从 Git 暂存内容导出的干净副本检查训练入口和模块依赖,不依赖忽略的 `test/` 或本机旧适配仓库。OMG/TextOp dry-run、GenMo 配置及新增模块导入检查通过;Kimodo 原生离线检查通过,真实参数更新成功,文本敏感性为 0.0632,根/身体梯度均有限且非零。 +- +-## 小规模闭环质量复核 +- +-使用旧 GenMo 正式 best 初始化,在两个**训练集** session 的全部合法窗口上训练 600 步,batch 8;恢复课程压缩为 warmup 40、ramp 160。此实验只检查新框架能否学习与闭环是否改善,不是默认 100K 配方的验证集比较,也不是与旧策略同预算的完整消融。 +- +-评估各自前 450 帧(15 秒):一个 GT 初始姿态,之后仅反馈预测历史,执行段为 12 帧;文字按原时间释放。无根对齐或时间扭曲。误差从第二帧开始统计,路径长度包含整个轨迹。 +- +-| 训练 session | 根坐标 RMSE,前→后(m) | 身体旋转 RMSE,前→后(rad) | 根路径,预测前→后 / GT(m) | +-|---|---:|---:|---:| +-| 53390 | 2.794 → 0.922 | 0.462 → 0.317 | 2.973 → 4.328 / 6.075 | +-| 28035 | 0.125 → 0.142 | 0.630 → 0.619 | 0.549 → 0.468 / 1.588 | +- +-首 20 步平均训练 loss 从 2.623 降至末 20 步的 1.499,身体、边界及未来速度输出均有非零梯度。一段恢复明显,另一段仍动作不足;此前其他短程窗口实验也有改善和退化并存的情况,不能声称倾听停滞已经彻底解决。 +- +-两段预测和人体 GT 分别对完整连续序列 retarget 一次到 G1,保存原 G1、人体 GT retarget 与预测的三种对照视频。视频本机目录为 `runs/genmo-v2-debug-20261024/videos/{53390,28035}/`;它们是训练集诊断,不用于报告测试集成绩。新正式 run 应在 20K/40K/60K 查看自主闭环与 retarget 视频后判断效果;默认 best 已改按固定自主闭环分数选择,旧正式权重保留。 +diff --git a/docs/HFD_HUMAN_BASELINE_INPUT_PROTOCOL.md b/docs/HFD_HUMAN_BASELINE_INPUT_PROTOCOL.md +deleted file mode 100644 +index 6579d20..0000000 +--- a/docs/HFD_HUMAN_BASELINE_INPUT_PROTOCOL.md ++++ /dev/null +@@ -1,149 +0,0 @@ +-# HFD 五个 baseline:当前输入、输出与训练约定 +- +-更新:2026-10-21(Asia/Taipei)。本文记录当前本地实现与约定;源码/运行配置优先。 +-原2026-10-17调研、上游commit、论文链接和待实现方案已完整保存在 +-[历史调研](history/HFD_HUMAN_BASELINE_INPUT_PROTOCOL_RESEARCH_20261017.md)。 +-历史ASR/中文翻译、12帧GenMo历史等配方不作为当前默认。 +- +-## 共同文本协议 +- +-五个baseline统一约定如下;OMG、TextOp、GenMo已实现,Kimodo已增加全量原生英文动态人体训练与闭环评测入口,HY-Motion仍待适配。Kimodo的小规模验收不代表全库正式训练已经完成。 +- +-```yaml +-foreground_language: en +-caption_policy: aligned_reference_context +-clock_policy: microphone +-text_conditioning: none +-context_seconds: 30.0 +-alignment_guard_seconds: 0.1 +-``` +- +-输入是已经释放的英文前景GT文本。原始TTS原文与对应源音频由 +-Qwen3-ForcedAligner-0.6B离线估计词时间,再按动态session内的真实音频拼接偏移平移。 +-原生指令、停止、附和、犹豫、修正片段均按各自的实际播放位置处理,不提前给出最终完整指令。 +-不运行ASR、不翻译、不用动作标签选择“当前任务”,也不做摘要或prompt rewrite。 +-背景混合PCM原样保留,但背景文字不输入。这是oracle foreground transcript实验, +-不代表已经解决部署时的ASR或目标说话人识别。 +- +-当前词级对齐库含61,860条英文前景资产;这是音频资产数,不是不同句子数或session数。 +-对齐跨度无效、越界、非单调或文本覆盖不完整时,在对应完整音频片段结束后释放整段文字。 +-0.1秒guard是实验缓冲,不是经过人工验证的对齐误差上界;事后强制对齐也不等同在线识别。 +-实现见 `hfd_bridge/reference_text.py`、`reference_runtime.py`,共享声明见 +-`configs/english_text_protocol.yaml`。 +- +-## 时钟与文本释放 +- +-动作源为50Hz,混合音频16kHz。动作时刻t对应麦克风时刻 `t+0.2s`。 +-文本释放点采用麦克风时间 `r_k=0.6+0.4k`,k从0开始。 +-词末估计为e时,先等到 `e+0.1s`,再向上取整到下一个r_k: +- +-```text +-r(word) = 0.6 + 0.4 * max(0, ceil((e + 0.1 - 0.6) / 0.4)) +-当前可见文本 = 最近已到达释放点r_k上的文本记录,要求r_k <= t + 0.2 +-``` +- +-因此麦克风0.6/1.0/1.4秒对应动作0.4/0.8/1.2秒;动作起点没有预先获得0.6秒文本。 +-这是当前microphone协议;旧 `legacy_hfd` 按动作组提前取记录的行为仅供历史复现。 +-旧材料“同一动作组直接读取0.6秒结果”不能继续套入当前正式实验。 +- +-每条记录只保留该释放点前30秒内、已经完成并获准释放的词;不等待句子完整, +-不拼接历次窗口字符串,过期后可以为空,不沿用任意久远的最后一句。session切换时清空。 +-例如词末0.85秒,加guard到0.95秒,在麦克风1.0秒释放,动作0.8秒决策才能看见。 +-0.6秒音频窗与0.4秒步长是重叠窗口:相邻只新增0.4秒,不能说每步新增0.6秒。 +- +-模型不接收未来PCM、未释放文字、GT turn ID、listen/motion/idle标签、未来事件时刻、 +-真实剩余动作时长、未来根路径或未来姿态约束。标签仅用于监督、样本选择、评测与视频标注。 +-扩散训练的带噪未来目标是学习变量;不能将干净未来GT作为条件。 +- +-## 实际输入与输出 +- +-共同的是信息释放规则,不要求模型拥有相同张量或动作历史长度。 +- +-| 模型 | 当前输入 | 当前输出及执行 | 本地接入状态 | +-| --- | --- | --- | --- | +-| OMG | 已释放英文文本→冻结T5-Base,最多50 token、保留最近后缀;过去2秒已释放混合PCM;10帧动作历史@30Hz | 原生125D动作特征,解码G1全29关节及根状态;计划60帧/2秒;当前原生异步缓冲剩24帧触发规划 | 正式英文训练及验证评测已接通 | +-| TextOp | 已释放英文文本→冻结CLIP ViT-B/32,77 token含两个边界token;2帧57D原生动作历史特征@50Hz;不输入音频 | 128D primitive latent,经冻结VAE解码8帧57D/0.16秒动作;G1原生23关节及根状态,未建模腕关节固定 | 正式300K训练完成;连续验证和视频见TextOp评测入口 | +-| GenMo/GEM-SMPL | 已释放英文文本→冻结T5-3B,最多50 token、保留最近后缀;60帧/2秒人体历史@30Hz;固定已知体型;当前past_audio=false | 原生151D表示,解码身体21关节旋转、根朝向及根位移;未来60帧/2秒,每次执行12帧/0.4秒后重规划 | 100K已完成,best92K与最终100K的闭环/GMR/视频见[正式评测](HUMAN_FORMAL_VALIDATION_20261024.md) | +-| HY-Motion | 共同英文文本计划接原生Qwen3 token与CLIP-L全局特征;流式历史条件需显式扩展 | 原生201D人体运动表示;连续输出、历史帧数与mask尚未验收 | 已有模型环境/调研;本地完整HFD训练尚未接通,不声称已按此协议训练 | +-| Kimodo-SMPLX | 已释放英文文本→冻结LLM2Vec,最近400正文token,mean pooling为1×4096;60帧/2秒人体历史@30Hz;203维历史姿态约束,无音频;当前配置采用因果root分解 | root→body两阶段,SMPLX22的273D表示;默认未来60帧/2秒,执行12帧/0.4秒;中性手部,保留源体型;12帧预测诊断未优于默认 | 修复版100K已完成,best即100K;冻结根骨干仅训练根投影/身体,见[训练修复](KIMODO_TRAINING_REPAIR.md)及[新评测](HUMAN_FORMAL_VALIDATION_20261024.md);未宣称闭环质量已达标 | +- +-30秒是候选文本范围,不等于30秒所有文字都进入编码器。TextOp的75个正文token +-加边界共77;T5为原生50 token。当前各自编码和缓存embedding,不共享不同模型的向量。 +-Kimodo的400正文token进入原生512总预算及模板;独立文本worker验证双向注意力和两份适配器权重。HY-Motion的token预算和前缀长度仍待实际adapter固定。 +- +-Kimodo使用归一化x0主损失,历史70维未观察辅助量在每个DDIM步保持训练噪声分布,避免未监督历史输出继续回灌。2026-10-23修复版增加原始根历史重建及实际解码运动的连续性监督;根/身体分别优化和裁剪,冻结根注意力骨干,未增加未来观测或推理修正。早期2K+1K、在线历史课程属于历史诊断。全量配置从完整英文索引按窗口均匀边缘抽样,不采用诊断中的70%活动窗口筛选;文本worker与动作模型同卡、独立进程,已编码特征在本地SQLite复用。共同文本释放规则及其余四模型输入未更改。 +- +-TextOp每8帧重新查询当前可见文字,文本本身只在0.4秒释放点改变;不同模型的执行周期 +-不必等于文本释放周期。其原生历史特征含相邻帧差分,2帧特征需读到当前决策状态, +-不能简单理解成两个完全独立的旧姿态。TextOp原生57D包含根姿态/位移、23个关节、 +-关节差分和接触;沿用官方统计与VAE,不直接拿G1原始36D训练。 +- +-OMG的60×35声学位置来自已经截取的过去PCM,不是未来2秒音频;文本和音频释放时钟一致。 +-初始不足两秒的音频左补零。TextOp和GenMo主实验没有声学分支,因此在文字出现前 +-没有显式语音起点、韵律或背景说话信息,不能要求它们仅靠文字精确判断倾听开始。 +- +-## 人体数据与session构造 +- +-源人体资产包含 `trans[T,3]`、`root_orient[T,3]`、`pose_body[T,63]`、 +-`pose_hand[T,90]`、`betas[16]`、`sample_times`。它们是源动作,不是完整全双工人体session。 +-以 `metadata/seed_g1_smplx_pairs.parquet` 的motion_id连接源,按原生配方的真实选取、 +-裁剪、持姿、朝向/位置衔接、过渡和事件时间构造人体时间线。 +-当前实现仍借用原G1 session组合程序及配对表追踪计划;不是从G1反解人体。 +-G1关节不是GenMo或Kimodo网络输入或训练损失目标。 +- +-GenMo训练实际使用trans、root_orient、21个身体关节及前10维betas; +-16维体型还用于SMPL-X pelvis/坐标转换。手指与面部不作为生成目标,推理导出中手部为中性残差。 +-151D由21×6D身体旋转、10D体型、两组6D根朝向和3D根局部相邻帧位移构成。 +-位移编码没有乘FPS,不能直接称为m/s;坐标转换和推理积分必须互为逆变换。 +- +-原生源动作约50Hz对齐成完整session后,按时间与SO(3)插值重采样到30Hz。 +-输入历史不能通过插值使用决策之后的GT;GenMo历史单独编码,最后一个历史帧指向 +-第一未来帧的位移未知,保留为生成目标。已知历史锁定,无下一有效帧的位移/静止标签不计损失。 +- +-## 数据范围与缓存 +- +-| 模型范围 | 英文train候选session | 英文val候选session | +-| --- | ---: | ---: | +-| OMG/TextOp:原生+已选第三方 | 168,583 | 5,109 | +-| GenMo:仅原生配对人体 | 155,117 | 4,719 | +- +-第三方生产包有G1时间线,但尚无经验证的配对人体session;GenMo排除该部分。 +-因此不能直接把OMG/TextOp与GenMo的结果差异完全归因于模型结构。 +-原生索引/seed保持,英文过滤不重排session ID;原生任务源train/val隔离。 +-当前没有独立test集,代码中test别名val不能当作未触碰测试集;预训练/共享stationary源重叠未排除。 +- +-离线完成的是源英文语音词对齐及索引,不是提前导出全部人体session。 +-OMG/TextOp/GenMo正式训练均按需组装原生session,组装结果放有上限的进程内缓存,退出即失效; +-第三方则读取已组装生产session。GenMo额外缓存30Hz重采样,train/val各512MiB, +-冻结文本特征共享256MiB CPU LRU;这不是全量持久session缓存。 +-原始tar在共享盘保留,本地只读索引镜像减轻共享盘随机读取。 +-采样有放回、槽位复用,step/epoch不等于逐条无重复遍历整个候选集。 +- +-## 训练与评测 +- +-正式配置:`configs/omg.yaml`、`configs/textop.yaml`、`configs/genmo.yaml`。 +-OMG预算100K、TextOp三阶段合计300K、GenMo100K;不按步数直接比较训练数据量。 +-OMG/GenMo训练使用GT历史,TextOp沿用原生分阶段GT/rollout/static-history课程。 +-正式连续评测只做一次初始GT bootstrap,之后使用预测历史;重复喂GT的结果必须标为oracle诊断。 +-TextOp当前根位置与朝向锚点取已知初始决策状态,不能每个primitive重新用GT根状态对齐。 +- +-评测入口:`evaluate_omg.py`、`evaluate_textop.py`、`evaluate_genmo.py`。 +-OMG/TextOp直接输出G1,不需要人体retarget。 +-人体模型先比较人体关节旋转、根位置、根朝向和根速度,再对整条预测及整条人体GT各做一次 +-SMPL-X→G1 retarget,不能在每个规划窗口重置IK。三种比较分别是预测→G1/原G1、 +-人体GT→G1/原G1、预测→G1/人体GT→G1;重定向误差不能直接从模型误差中相减。 +-统一保存原混合音轨、microphone前置帧、文本释放审计和bootstrap排除范围。 +- +-验证loss用于监测或选权重,但不等于连续生成质量或语义成功率。特别是GenMo原生位移loss +-使用GT朝向积分,闭环使用预测朝向,必须同时看朝向和长时漂移。 +-当前视频是运动学回放,不代表物理控制跟踪、独立任务成功率或真实端到端停止延迟。 +-GenMo的双分支梯度/因果性与44项检查见[正式训练说明](GENMO_FORMAL_TRAINING.md)。 +- +-更多入口:[共同文本条件](BASELINE_TEXT_CONDITION.md)、[session格式](../SESSION_FORMAT.md)、 +-[动态reference组装](REFERENCE_DYNAMIC_TRAINING.md)、[GenMo正式训练](GENMO_FORMAL_TRAINING.md)、 +-[人体retarget](HUMAN_RETARGET_EVALUATION.md)。历史研究及社区训练代码核查仍见归档, +-不把当年的待实现建议、旧ASR输入或旧缓存描述当作当前已经验收的实现。 +- +-## GenMo v2训练适配补充(2026-10-24) +- +-输入/输出与本协议相同。训练额外使用一段模型预测历史作为因果恢复条件,GT未来仅作监督,前0.4秒增加损失权重;这不增加模型推理时可见的信息。默认最佳权重用自主闭环验证选择,执行细节见 [GENMO_V2_TRAINING.md](GENMO_V2_TRAINING.md)。旧配方保留在 `configs/genmo_legacy.yaml`。 +diff --git a/docs/HFD_HUMAN_BASELINE_SOURCES.json b/docs/HFD_HUMAN_BASELINE_SOURCES.json +index 8752b9f..5cf7f2f 100644 +--- a/docs/HFD_HUMAN_BASELINE_SOURCES.json ++++ b/docs/HFD_HUMAN_BASELINE_SOURCES.json +@@ -175,10 +175,6 @@ + } + ], + "local_evidence": [ +- { +- "path": "/root/autodl-tmp/hfd-baseline-portable/SESSION_FORMAT.md", +- "sha256": "b478da021db41576b023de4ee057350e93539b89a9427fa680f16004abfb226b" +- }, + { + "path": "/root/autodl-tmp/hfd-baseline-portable/README.md", + "sha256": "f2b9adf6cf12bcbe7ebd3af72e74355d7a1910f0238c2744571a03a6193c7f40" +diff --git a/docs/HUMAN_FORMAL_VALIDATION_20261024.md b/docs/HUMAN_FORMAL_VALIDATION_20261024.md +deleted file mode 100644 +index 531d697..0000000 +--- a/docs/HUMAN_FORMAL_VALIDATION_20261024.md ++++ /dev/null +@@ -1,38 +0,0 @@ +-2026-10-24:GenMo、修复版 Kimodo 均已完成 100K 步;本轮按既有六类 native English validation 规则执行闭环生成、整段 GMR 和有声视频。没有训练、修改或删除正式权重。 +- +-输入为按既定词时间释放的最近30秒英文GT文本与60帧已执行人体历史,无音频输入。只使用第一帧GT初始化,随后完全反馈预测;规划60帧、执行12帧、30Hz。Kimodo采用DDIM20/CFG2,GenMo采用DDIM10/CFG2.5,均seed12345。人体GT和完整预测分别连续retarget一次,无轨迹对齐、DTW或位置修正。 +- +-GenMo依原入口每类选首个预计6–30秒的样本;Kimodo依原入口选6–20秒。打断、急停的样本ID不同,其余四类相同。所有样本均在生成前确定。下表用于比较各模型微调前后,不能据此直接排名架构;也不是全HFD验证集、独立test或物理跟踪/任务成功率评测。 +- +-| 模型/权重 | 人体姿态RMSE(rad) | 人体根位置RMSE(m) | G1关节RMSE(rad) | G1根位置RMSE(m) | G1根朝向RMSE(rad) | +-|---|---:|---:|---:|---:|---:| +-| GenMo official | 0.651 | 0.995 | 0.688 | 0.864 | 2.342 | +-| GenMo best92k | 0.407 | 0.701 | 0.373 | 0.604 | 0.373 | +-| GenMo last100k | 0.427 | 0.758 | 0.376 | 0.655 | 0.390 | +-| Kimodo official | 0.627 | 0.502 | 0.464 | 0.433 | 0.575 | +-| Kimodo last100k | 0.422 | 0.306 | 0.373 | 0.264 | 0.487 | +- +-六段分别评分后等权平均,排除GT初始化帧。人体姿态是21个身体关节的SO(3)角误差RMSE;根位置为三坐标逐元素RMSE。G1指标参考同流程retarget的human GT,原始G1仅作另一个参考;retarget误差不从模型误差中相减。 +- +-GenMo:优先保留92K的best作为当前评测候选。固定去噪验证loss在92K最低(2.409),100K为2.464;本轮六段连续生成的人体姿态、根位置及根速度误差也均为92K略低。训练耗时约85.8小时。批量评测复用模型但恢复原始初始化后的采样RNG;官方和92K各抽一个固定样本与未修改的infer_genmo.py命令复核,人体位置/根旋转/身体旋转均逐位相同。 +- +-Kimodo:best与last均为100K,训练耗时约18.7小时,固定验证loss为1.179→0.0669。根骨干200个参数张量仍与官方初始化逐位一致;同一后退诊断窗口t=0的末层时序RMS为0.0360、输入梯度范数0.00586,旧失效100K分别约1.08e-7和4.76e-12。本轮未重现此前根网络恒定输出失效。 +- +-与上一轮失败的Kimodo 100K历史报告相比,G1根位置误差从0.544降至0.264m,人体窗口首帧根速度误差从约0.839降至0.109m/s;但G1根朝向误差从0.266升至0.487rad。修复包含可训练参数范围、学习率、损失和注意力后端的组合修改,不能把这些收益归因于某个单独超参数。 +- +-两者相对官方初始化均有平均误差收益,但动作质量仍有明显限制:后退、长距离移动与部分指令跟随不足。backchannel样本的人体GT净后退位移约3.50m,Kimodo仅沿GT后退方向移动约0.37m;GenMo该样本的整段根路径只有约0.14m,而GT约3.65m,接近保持位置。低平均误差不能当成动作任务已经完成。 +- +-保持初始姿态的对照也保存在JSON中。Kimodo新100K的人体姿态平均误差0.422略低于静止对照0.431,但G1关节0.373仍略高于静止对照0.366。GenMo92K的人体姿态0.407低于静止对照0.429。静止对照无法执行指令,只用于识别静态区间对平均分的影响。 +- +-本轮导出42段有声MP4:Kimodo 18段、GenMo 24段。均核对帧数、30Hz、音轨和完整音视频解码;抽帧检查六类预测/参考、文本与时钟显示。runs只保留视频,工作数据和日志在cache。 +- +-- `/root/autodl-tmp/runs/kimodo-formal-validation-20261024/last100k/`:六类预测→G1 / human-GT→G1及六段retarget参考;`official/`含六段官方初始化对照。 +-- `/root/autodl-tmp/runs/genmo-formal-validation-20261024/best92k/`、`last100k/`、`official/`:各六段预测→G1 / human-GT→G1;last100k另含六段GT retarget参考。 +-- `retargeted_human_comparison.mp4`左侧预测,右侧同流程retarget的human GT;`retarget_reference.mp4`比较human-GT→G1与原始G1。 +- +-完整报告与曲线: +- +-- [Kimodo数值、逐例结果、根骨干检查和视频校验](../reports/kimodo_formal_validation_20261024.json) +-- [GenMo数值、逐例结果、命令一致性和视频校验](../reports/genmo_formal_validation_20261024.json) +-- [Kimodo验证曲线与后退轨迹](../reports/kimodo_formal_validation_20261024.png) +-- [GenMo验证曲线与后退轨迹](../reports/genmo_formal_validation_20261024.png) +diff --git a/docs/HUMAN_RETARGET_EVALUATION.md b/docs/HUMAN_RETARGET_EVALUATION.md +deleted file mode 100644 +index 3176e0b..0000000 +--- a/docs/HUMAN_RETARGET_EVALUATION.md ++++ /dev/null +@@ -1,98 +0,0 @@ +-# 人体模型的 retarget 评测入口 +- +-`evaluate_human.py` 共用于 Kimodo、GenMo、HY-Motion。输入为模型导出的 **完整连续 rollout**,先转换为标准 SMPL-X / AMASS,再通过 GMR 转到 G1。该入口不加载生成模型、不下载权重,也不改变现有 TextOP / OMG 推理逻辑。 +- +-2026-10-21更新:GenMo已接通动态原生英文人体训练及`evaluate_genmo.py`连续推理、完整轨迹GMR与带音轨视频;Kimodo已增加固定原生英文子集训练与`evaluate_kimodo.py`完整轨迹导出、GMR和视频。两者当前均输入已释放英文GT文本+60帧人体历史,不输入音频。Kimodo范围见[训练说明](KIMODO_HFD_TRAINING.md),HY仍待接入。共同输入协议见[当前约定](HFD_HUMAN_BASELINE_INPUT_PROTOCOL.md)。GMR位于 `/root/autodl-tmp/hfd-human-models/gmr-runtime`,主体资产在fs的`models/body_models`。 +- +-## 同时保存三组结果 +- +-| report.json 指标 | 左侧 | 右侧 | 用途 | +-| --- | --- | --- | --- | +-| `prediction_vs_original_g1` | 预测 SMPL-X→G1 | 原始 G1 真值 | 与机器人 baseline 相同目标上的轨迹误差 | +-| `human_gt_vs_original_g1` | human 真值→G1 | 原始 G1 真值 | 检查体型、session 几何渲染及 retarget 自身的偏差 | +-| `prediction_vs_retargeted_human_gt` | 预测 SMPL-X→G1 | human 真值→G1 | 相同转换后的预测误差 | +- +-第二项是转换链路的诊断,不能从第一项中直接相减,也不是数学意义的误差下界。所有指标包含 29 关节 RMSE(rad)、根位置每坐标 RMSE(m)、根朝向 SO(3) 测地角 RMSE(rad)。不做 DTW、逐帧根对齐、截短或补齐。`bootstrap_frames` 在三组评分中一致排除。teacher-forced 必须标为 oracle,不能作为闭环结果。 +- +-此外,`report.json.human_metrics`在retarget前直接比较人体身体旋转、根位置、根朝向及根速度,排除同一bootstrap。根速度按实际FPS计算,单位m/s;人体关节角采用SO(3)测地角。优先用人体自身指标判断生成误差,G1比较辅助诊断转换与机器人覆盖差异。 +- +-一个连续序列仅创建一次 GMR retargeter,内部状态跨帧保留;预测和 human GT 分别新建,不能共享 IK 状态,也不能每个 0.4s 规划块重置。保持旧 benchmark 的 `offset_to_ground=False` 和身高估计 `1.66 + 0.1 * beta[0]`;体型使用 16 个 betas、`use_pca=False`、`flat_hand_mean=False`。不根据未来 GT 调整预测的朝向或地面。该身高估计是否适用于当前体型应在真实集成测试中核对。 +- +-## 输入契约 +- +-路径可相对 manifest 所在目录。manifest 的三份动作文件必须携带同一个已冻结的 `session_plan_id`。**这个 ID 只用于发现误配,脚本不能证明 human/G1 已按同一个 plan 正确渲染**;生成端必须先完成该保证。不要仅凭相同随机种子或相同动作名配对。 +- +-预测和 human 真值两个 NPZ 都必须包含: +- +-| 字段 | 要求 | +-| --- | --- | +-| `trans` / `root_orient` / `pose_body` | `[T,3]` / `[T,3]` / `[T,63]`,轴角,有限值 | +-| `betas` | 恒定 `[10]` 或 `[16]`,10 维补零至 16;不从 GT 偷换预测体型 | +-| `pose_hand` | 可选 `[T,90]` 左右手轴角;相对 SMPL-X 手均值的残差;缺省零表示模型默认手姿态 | +-| `gender` | 可选 `neutral` / `male` / `female`,缺省 neutral | +-| `mocap_frame_rate` | 正标量;预测和 human GT 必须相同 | +-| `coordinate_system` | 字符串 `amass_z_up`;`trans` 是模型平移,不是 pelvis 世界坐标 | +-| `sample_times` | 可选;若存在必须等于 `arange(T)/fps`,时间不均匀时先在导出端显式重采样 | +-| `session_plan_id` | 冻结 plan 的标识,推荐使用内容哈希 | +-| `start_seconds` | 该 rollout 第 0 帧在 session 中的时间,与 manifest 一致 | +- +-Kimodo 的 Y-up joint positions / rotations、HY 的 201D 特征、GenMo 的 151D 特征均不能直接送入此入口,须通过各模型原生解码与坐标逆变换导出 AMASS。`trans` 不做隐式 pelvis 偏移修正。非零 `pose_jaw` / `pose_eye` 暂不支持。预测与 human GT 的体型/性别是否相同会记录在 provenance,差异需要在实验说明中解释。 +- +-原始 G1 session NPZ 必须包含: +- +-- `layout="xyz_wxyz_native29"` + `qpos_36[T,36]`,或 `layout="hfd_isaac29_xyz_wxyz"` + `motion[T,36]`。后者走现有按关节名排列的转换,不能直接切片当 MuJoCo qpos。 +-- `fps`、`session_plan_id`、`start_seconds=0`。 +-- `audio`(一维PCM)、`audio_sample_rate=16000`、`audio_start_seconds`。当前`microphone`协议下原PCM起点填0,动作时刻t显示在麦克风t+0.2秒;从动作0开始的回放给两侧共同增加0.2秒前置姿态。不能把当前PCM起点机械填成-0.2或再次偏移。历史协议需按原manifest显式复现。 +- +-G1 真值按真实时间线重采样到 human rollout 帧率:位置/关节线性插值,根四元数 SLERP。覆盖不足立即报错;音频也必须覆盖完整的 `T/fps` 播放区间。当前入口面向均匀 rollout,不接受拼接成连续序列的 disconnected windows。 +- +-manifest 示例(示意路径,不是现有数据): +- +-```json +-{ +- "schema_version": 1, +- "model": "genmo", +- "session_plan_id": "sha256-of-frozen-plan", +- "prediction": "prediction_amass.npz", +- "human_reference": "human_gt_amass.npz", +- "g1_reference": "g1_session.npz", +- "start_seconds": 0.0, +- "bootstrap_frames": 1, +- "history": "predicted", +- "clock_policy": "microphone", +- "caption_policy": "aligned_reference_context", +- "text_conditioning": "none", +- "text_source": "oracle_foreground_transcript", +- "inputs": {"text": "released English reference text <=30s", "history": "60-frame history; initial pose held at startup, then predictions"}, +- "audio_policy": "Aligned reference text <=30s; no raw audio conditioning", +- "use_asr_text": true, +- "command_seconds": [0.4, 5.2] +-} +-``` +- +-`command_seconds` 仅供叠字展示,不用于生成或 retarget。输入策略字段是生成端声明,评测脚本不能据此证明模型没有使用未来信息。 +- +-## 执行与渲染 +- +-在具有 NumPy / SciPy 的模型环境中,先验证导出数据: +- +-```bash +-python evaluate_human.py --manifest /path/to/manifest.json \ +- --output-dir /root/autodl-tmp/runs/kimodo-eval \ +- --gmr-root /path/to/GMR --body-models /path/to/body_models --preflight +-``` +- +-preflight 检查文件、布局、时间覆盖和 plan ID,只报告 GMR/资产目录是否存在;不会导入 GMR 或验证资产许可证、内容、依赖是否完整,也不写评测目录。具备真实依赖后去掉 `--preflight` 执行转换。输出目录已存在时拒绝覆盖。 +- +-输出 `report.json`、三组配对 NPZ,以及汇总 `retargeted.npz`(预测、human GT、原始 G1 GT 和绝对 timestamps),包含输入 SHA256、GMR commit/dirty 状态、路径、转换配置和体型一致性。真实资产存在时记录主体 NPZ 的 SHA256,并记录相关包版本;GMR 工作树仍需保持固定。 +- +-现有渲染环境中执行: +- +-```bash +-python render_overfit.py --inference /root/autodl-tmp/runs/kimodo-eval \ +- --assets-root /path/to/OMG/assets --gpu 0 +-``` +- +-三组视频均使用原 session 音频;标题区分预测、human GT retarget 和原始 G1 真值。视频为参考姿态播放,不是物理跟踪结果。human GT 转换对照不显示模型 ASR 输入或预测历史提示。 +- +-开发验证:`python -m unittest tests.test_human_eval tests.test_joint_mapping -v`。合成测试覆盖三组指标、bootstrap 排除、序列级后端调用、拒绝误配/不均匀时间/覆盖不足、四元数双覆盖和 SLERP,以及 HFD→G1 关节名映射。 +diff --git a/docs/KIMODO_CLOSED_LOOP_STABILITY.md b/docs/KIMODO_CLOSED_LOOP_STABILITY.md +deleted file mode 100644 +index ca51599..0000000 +--- a/docs/KIMODO_CLOSED_LOOP_STABILITY.md ++++ /dev/null +@@ -1,144 +0,0 @@ +-# Kimodo 训练策略与闭环稳定性实验 +- +-本轮从官方权重重新构建实验。此前及本轮的测试权重均已清理,复现需重训;报告、代码和视频保留,正式GenMo训练不受影响。 +- +-## 当前保留的诊断路线与最终状态 +- +-**训练与推理的关键一致性问题已修正,但六种模式混合闭环过拟合、少样本验证均未通过;不启动全量正式Kimodo训练。** +- +-目前较好的诊断路线是:官方初始化→归一化x0目标训练2,000步→只加载模型权重、重置Adam,再用GT历史训练1,000步。输入协议不变,不增加GT阶段、语音结束状态或动作时钟。两个公开配置为`configs/kimodo_overfit.yaml`与`configs/kimodo_refine_probe.yaml`,准备/复用输入与续训命令见[训练说明](KIMODO_HFD_TRAINING.md)。首阶段配置与实测`normalized-x0`除输出路径外完全相同。 +- +-| 修正采样后的实验 | 累计步数 | 训练身体RMSE(rad) | 训练根RMSE(m) | 训练根速度RMSE(m/s) | +-| --- | ---: | ---: | ---: | ---: | +-| 物理Huber,同一模型旧历史反馈规则 | 2,000 | 0.3703 | 0.2580 | 0.4307 | +-| 物理Huber,同一模型修正历史反馈规则 | 2,000 | 0.3747 | 0.2137 | 0.3638 | +-| 归一化x0,50%高噪声/70%运动窗口 | 2,000 | 0.3822 | 0.2055 | 0.4102 | +-| 上述模型再用GT历史训练1,000步 | 3,000 | **0.3105** | **0.0915** | **0.3183** | +-| 同一初始化,再用在线历史训练1,000步 | 3,000 | 0.4355 | 0.2826 | 0.5754 | +-| 归一化x0,均匀扩散步/均匀窗口 | 2,000 | 0.3898 | 0.5479 | 0.3337 | +- +-第一、二行已核对checkpoint SHA相同,唯一改动是推理历史辅助量规则:根位置/根速度误差分别降低约17%/16%,身体误差没有改善。第四、五行已核对初始化权重相同、各追加1,000步;在线课程更差,**不采用为默认**。归一化x0在2,000步时并未全面优于物理Huber,不能把第4行的改善全部归因于损失更换,训练预算和优化器重启也改变了。 +- +-GT历史3K模型的六条结果如下,均为完整预测历史rollout: +- +-| 模式 | 身体RMSE(rad) | 根RMSE(m) | 预测/GT身体角速度比 | +-| --- | ---: | ---: | ---: | +-| turn-taking | 0.4387 | 0.0398 | 1.747 | +-| interruption | 0.3192 | 0.0686 | 1.165 | +-| emergency-stop | 0.0564 | 0.0396 | 1.149 | +-| backchannel | 0.3220 | 0.0236 | 1.107 | +-| hesitation | 0.3703 | 0.1470 | 1.103 | +-| pause | 0.3564 | 0.2307 | 1.148 | +- +-逐条精确值以[机器报告](../reports/kimodo_closed_loop_stability_20261021.json)为准。行走/停止已能较准确复现,其他模式虽运动幅度不再明显不足,动作相位和多余运动仍存在;幅度正确不等于动作正确。六条验证session宏平均为身体0.5220rad、根0.8850m、根速度0.3673m/s,身体只有1/6好于冻结初始姿态,泛化没有通过。当前训练入口仍是固定小规模子集,不是全库动态训练器。 +- +-优先查看本轮较好模型的三组预测↔人体GT视频(双方各完整retarget到G1,均含音轨): +- +-- [行走/停止](/root/autodl-tmp/runs/kimodo-closed-loop-20261021/score-control/train-emergency-stop-000036/retargeted_human_comparison.mp4) +-- [动作切换,仍有相位偏差](/root/autodl-tmp/runs/kimodo-closed-loop-20261021/score-control/train-turn-taking-000013/retargeted_human_comparison.mp4) +-- [犹豫,仍未准确复现](/root/autodl-tmp/runs/kimodo-closed-loop-20261021/score-control/train-hesitation-000087/retargeted_human_comparison.mp4) +- +-## 修正历史辅助量的推理反馈 +- +-训练时,历史未观察的70维速度/接触辅助量始终是`epsilon * sigma_t`;它们不参与历史损失。旧推理却把模型对这些未监督历史字段的预测继续经DDIM反馈,低噪声阶段也保留其信号,与训练不一致。 +- +-当前`inference.prefix_aux_policy: training_noise`在每次去噪调用前恢复相应扩散步的历史纯噪声边缘分布,然后重新锁定已知203维姿态。噪声来自本轮初始epsilon,没有额外GT或随机数;未来采样变量不被这一步改写,也不对输出做平滑。训练与20步稀疏DDIM的sigma映射有独立一致性检查,旧行为仅通过`legacy`供对照复现。在线历史刷新也使用同一规则。 +- +-新增`normalized_x0`主损失直接在原生归一化特征上做Gaussian x0最小二乘;root/body两阶段分别平均,已知/填充帧和未定义末端速度排除。它是训练目标对照,不宣称复现未公开的官方训练配方。`physical_huber`保留供历史结果对照。 +- +-当前代码41项测试通过;最新完整模型续训重放含修正后的采样与在线缓存,在确定性模式下逐位一致。该完整重放使用physical_huber分支;normalized_x0共用持久化路径,并完成了实际全模型训练与梯度检查。本轮共16条训练路径(含共享初始化和后续阶段),不是16次独立随机实验。以下为历史实验明细;旧表均使用修正前的`legacy`历史辅助量反馈,重新运行需显式指定该规则。 +- +-## 视频审计修正 +- +-快速probe最初没有导出每轮`executed_frames`,导致显示层误报“文本未释放”。模型实际使用的caption、运动轨迹和误差均未受影响。现已补齐字段,核对执行帧总数等于预测总长减去bootstrap,并重渲染全部72个视频。现有retarget结果未重新计算;报告记录了仅显示元数据修正及前后manifest哈希。 +- +-## 改动与对照原则 +- +-输入协议不变:按时释放的最近30秒英文GT文本、60帧已知人体历史、预测60帧、执行12帧、无音频。推理仅用一帧GT初始化,后续全部接自己的输出,不加入GT阶段、未来指令、动作时钟或目标姿态,也不做平滑后处理。 +- +-`hfd_bridge/kimodo_closed_loop.py` 实现两项可独立开关的训练策略: +- +-1. `closed_loop.tail_weight`:实际执行的前12个未来帧权重为1,余下48帧为较低的正权重。无效填充帧依然排除,分母使用有效权重之和。这个实验检验远期监督是否稀释当前决策,不代表已经证明未来文字是退化的唯一原因。 +-2. `closed_loop.online_history`:按 `refresh_every` 用当前模型重新闭环生成训练session,混入原GT历史窗口。`max_probability` 在 `ramp_steps` 内逐渐增加。刷新使用与推理相同的DDIM步数、CFG、文字释放时钟和单帧初始化;只刷新训练集,不对验证集训练。预测历史只作为条件,GT未来保留原世界坐标,不再平移根位置或混合前6帧目标姿态。 +- +-历史刷新不反传穿过整个采样轨迹;随后训练步对网络两阶段正常反传。独立的历史选择随机数不会改变原窗口抽样顺序,刷新会恢复训练的全局随机数和train/eval模式。完整checkpoint包含历史缓存、缓存对应的policy step和历史选择随机数;恢复训练不重新生成另一份历史替代原缓存。 +- +-原 `prepare_kimodo_predicted_history.py` 为历史实验保留,不作为本轮在线策略入口。 +- +-可选 `reset_schedule` 只控制训练历史课程:例如 `[2,5,0]` 在前两阶段每2轮、5轮从GT过去重新启动短rollout,最后阶段从一个GT初始姿态生成完整session。每个窗口保存模型当时实际看到的60帧历史,不能从包含后续GT重启的拼接轨迹倒推历史。正式闭环评测入口没有这个开关,始终不重启、不补GT历史。`curriculum_stage_steps` 决定阶段长度,实际历史在下一次刷新时切换。 +- +-## 固定实验 +- +-数据保持六种模式各一条训练session与验证session,分别205与215个窗口。所有数据来自原生英文配对人体session。81种合法释放文本使用同一份冻结LLM2Vec缓存。 +- +-- 先按原方案训练2,000步,作为公共初始化。 +-- 三组分别从同一初始化再训练1,000步:原方案对照;`tail_weight=0.1`;`tail_weight=0.1`加在线历史(最大概率0.5,500步渐增,每250步刷新)。 +-- 三组保持batch8、LR1e-5→1e-6、随机种子、窗口抽样顺序、原七项损失、文本dropout0.1、DDIM20、CFG2一致。本轮不同时叠加高噪声抽样、活跃采样或额外速度损失。 +-- GT历史去噪loss只用于诊断,不能跨不同权重目标直接比较。主要比较六条训练session的完整闭环误差、静止基线、角速度幅度、根速度误差和重规划加速度误差;验证集另列,不作为训练过拟合的替代。 +- +-前三组之后增加一个同预算的课程恢复候选:短历史→长历史→完整历史,50%样本t=999,并在root输出进入body阶段前锁定已知root历史。它是组合配方,不把其结果归因于单一因素;仍不增加活跃抽样或额外速度损失。 +- +-快速多session评测入口为 `tools/probe_kimodo_closed_loop.py`,导出的manifest兼容 `evaluate_human.py` 和 `render_overfit.py`。`--teacher-forced`只能生成明确标记的诊断结果,不计入闭环成绩。每次评测必须使用新目录,避免误用其他权重的旧输出。 +- +-工作数据、临时日志与权重分别在 `cache/kimodo-stability-20261021`、`setup/kimodo-stability-*` 和 `/dev/shm/kimodo-stability-*`,不重新往 `runs` 堆积中间数据。`runs`只保存最终对照视频。 +- +-## 结果 +- +-**结论:混合闭环过拟合仍未通过,不推荐启动正式Kimodo训练。** 已落地表示一致性、导数标签与续训状态修正;在线历史、执行帧加权、动态损失和短预测时长没有形成可靠的混合训练配方。 +- +-下表为原始六条训练session完整预测历史rollout的等权宏平均,全部仅第一帧GT初始化、固定seed12345、DDIM20、CFG2,无后处理。累计步数包含公共初始化,表中不同累计预算不能解释为单因素同预算对照。 +- +-| 配方 | 累计更新 | 身体RMSE(rad) | 根RMSE(m) | 根速度RMSE(m/s) | +-| --- | ---: | ---: | ---: | ---: | +-| 原方案公共初始化 | 2000 | 0.4785 | 0.5938 | 0.2945 | +-| 原方案继续训练 | 3000 | 0.4186 | 0.5964 | 0.2944 | +-| 执行帧加权 | 3000 | 0.4353 | 0.5961 | 0.2946 | +-| 加权+整段在线历史 | 3000 | 0.3540 | 0.5789 | 1.2121 | +-| 高噪声+历史课程 | 3000 | 0.3673 | 0.2337 | 0.4767 | +-| 强动态监督/GT历史 | 5000 | 0.3583 | 0.2913 | 1.1430 | +-| 强动态监督/课程历史 | 5000 | 0.3470 | 0.3921 | 0.4033 | +-| 因果编码+强动态监督/GT历史 | 5000 | 0.4352 | 0.4061 | 0.9534 | +-| 因果编码+较轻动态监督/课程历史 | 6000 | 0.3231 | 0.4638 | 1.0686 | +-| 同预算追加/12帧预测 | 7000 | 0.3854 | 0.4401 | 1.3143 | +-| 同预算追加/60帧预测 | 7000 | 0.3387 | 0.3743 | 1.0379 | +- +-较轻课程组身体误差比3K原方案低约23%,但根速度误差从0.2944升至1.0686m/s,不能以身体误差下降宣称修复。其行走/停止样本根路径21.83m,GT5.29m,根速度误差4.7618m/s;backchannel运动速度仍只有GT约15%。强动态监督的在线版本相对稳定一些,但hesitation/pause仍偏静止、行走仍有额外速度。因果root编码解决了实际表示不一致,却没有独立带来混合动作质量提升。 +- +-12帧与60帧预测从同一个较轻课程组最终权重开始,均追加1,000步GT历史训练,LR1e-5→1e-6、batch8、运动抽样70%,仅预测长度不同。12帧组的三项宏平均均未优于60帧组,故默认保留60帧。该对照不是把不同预算或不同初始化混为预测时长的效果。 +- +-较轻课程组的六条验证session宏平均为身体0.4430rad、根0.4699m、根速度0.1859m/s,仅作小规模诊断,不是全库泛化或独立测试成绩。CFG1训练宏平均为0.3747rad/0.4654m/1.2224m/s,也没有改善本轮问题;验证结果与逐条指标见机器报告。 +- +-此处为旧采样规则下的历史对照;当前诊断配置和修正后的采样结果见文首。在线历史、执行帧加权及12帧预测未采纳为默认;其他四个baseline配置未改动。 +- +-[完整配置、逐条指标与检查记录](../reports/kimodo_closed_loop_stability_20261021.json)。旧采样规则阶段共11条训练路径(包含公共初始化),有共享初始化关系,不是11次独立随机实验。 +- +-视频均含原始音轨,分别先对完整预测和人体GT做一次GMR,再比较G1轨迹;不是逐窗口重置retarget。示例: +- +-- [原方案:行走/停止](/root/autodl-tmp/runs/kimodo-closed-loop-20261021/control/train-emergency-stop-000036/retargeted_human_comparison.mp4) +-- [较轻课程:行走/停止,仍有抖动](/root/autodl-tmp/runs/kimodo-closed-loop-20261021/causal-online-lite/train-emergency-stop-000036/retargeted_human_comparison.mp4) +-- [较轻课程:犹豫样本](/root/autodl-tmp/runs/kimodo-closed-loop-20261021/causal-online-lite/train-hesitation-000087/retargeted_human_comparison.mp4) +-- [12帧预测诊断:犹豫样本](/root/autodl-tmp/runs/kimodo-closed-loop-20261021/horizon12/train-hesitation-000087/retargeted_human_comparison.mp4) +- +-保留范围:报告和代码在仓库,工作缓存位于cache;runs中本轮只新增MP4。所有本轮过拟合权重已清理,正式OMG/TextOp权重及GenMo活跃输出保留。旧中间初始化权重不再可直接复用,复现需按报告中各阶段重建。 +- +-## 表示与速度监督的额外修正 +- +-复查原205个训练窗口,独立编码的历史末帧与整段监督中的同一帧,其虚拟平滑root平均相差0.0213m、最大0.1492m。这是特征分解差异,不是人体GT发生了15cm跳变;但把两份表示在边界拼接后,会影响root→body阶段计算的局部root速度。原先只检查“未来不能改变observed”,没有检查“known部分与target使用同一分解”。 +- +-新增 `representation.causal_root: true`:在原273D表示中使用单向、半径0.06m受限的根轨迹滤波(alpha0.2),并同步重算hips相对位移。FK、全局旋转、人体root物理位置及预训练统计保持一致。此滤波只用于特征分解,不对预测轨迹做事后平滑。前203维已知姿态在prefix与整段编码中一致,改变未来不改变输入,解码后保持原人体运动;已对全部420个train/val窗口检查。 +- +-`hfd_bridge/kimodo_causal_rep.py` 是可选编码实现;准备数据、在线历史、单session推理、多sessionprobe均读取相同配置。启用或切换编码后必须重新准备windows,不能直接把旧编码缓存标成新编码。严格训练合同会拒绝代码或数据已经变化的旧resume。 +- +-对于预测历史,位置目标仍为原绝对GT未来;另外的 `temporal_target` 只提供真实GT速度/加速度标签。构造这些标签时使用最后3个GT过去姿态,但它们不进入observed/mask或网络条件。这样不会把“从错误预测位置瞬间跳回GT”的位移当成期望速度。原GT历史样本的temporal target就是原target。新增边界root速度及实际解码root加速度监督,未配置的权重为0。 +- +-两组动态监督对照均从课程组第1,000步初始化,再训练2,000步、LR3e-5→3e-6:GT历史组与在线课程组使用相同强度的五项导数损失(1、0.5、0.1、2、0.05)。因果编码另做相同预算的GT历史对照。最后的较轻组合改为3,000步,权重0.2、0.2、0.05、0.2、0.005,70%抽样为运动窗口,75%历史混入上限,750步刷新,短2轮→短5轮→完整session;不降低远期监督权重。该组合改变了预算和多项训练策略,不能将其改善唯一归因于因果编码。 +- +-## 续训与确定性 +- +-包含因果编码、预测历史缓存和分开的导数标签的完整282,790,161参数模型,已在确定性CUDA/math attention模式下验证:第3步checkpoint两次恢复到第6步,模型、Adam、scheduler、全局随机状态、窗口采样、历史选择随机数、缓存及刷新元数据逐项完全相同。默认CUDA attention的重放出现过微小数值差异,但缓存和随机状态一致,不能把默认运行称为逐位可复现。确定性检查使用`CUBLAS_WORKSPACE_CONFIG=:4096:8`、`torch.use_deterministic_algorithms(True)`并关闭flash/memory-efficient SDPA;正常实验保留原性能配置。 +- +-## 输入时机的限制 +- +-训练中发现一个具体例子:`train-hesitation-000087` 的首段“performing ride the pony”在mic2.2s释放,最后一个词在mic5.4s释放;GT持续根运动起点约motion6.67s(按>0.15m/s持续0.2s检测)。一些候选在motion2.5s就开始运动。释放时钟没有错位,但文本模型看到可执行的前半句时,并没有显式的讲话/结束状态。这个例子提示输入信息可能限制时机复现,没有证明所有提前/静止问题均由它引起。 +- +-本轮没有加入GT阶段、语音结束标签、动作时钟或额外音频条件,也没有改动其他四个baseline的输入协议。后续若测试语音状态/文字稳定时长,需明确区分可实时计算的因果信号与GT阶段标签,并单独报告协议变化。 +diff --git a/docs/KIMODO_FORMAL_VALIDATION.md b/docs/KIMODO_FORMAL_VALIDATION.md +deleted file mode 100644 +index 0cf05ca..0000000 +--- a/docs/KIMODO_FORMAL_VALIDATION.md ++++ /dev/null +@@ -1,72 +0,0 @@ +-2026-10-24:修复后新100K的正式评测已完成,见[最新报告](HUMAN_FORMAL_VALIDATION_20261024.md)。本文记录旧失败训练的历史评测,不能把下文100K结果当成新权重的结果。 +- +-2026-10-23 更新:用户随后要求修复训练并删除旧失败目录;原run及其全部权重已删除,本文数值、分析缓存和视频保留。训练配置及修改见[训练修复](KIMODO_TRAINING_REPAIR.md)。当前同名run/权重目录已被新训练重新创建。 +- +-2026-10-23:`322190.textop` screen 实际运行的是 Kimodo,已正常完成 100,000 步,耗时约 19.6 小时。 +-本轮对六类未参与训练的 validation session 做了完整闭环生成、整段 GMR retarget 和有声视频评测。 +-100K 权重已有明确的人体姿态与朝向收益,但根位置、窗口衔接和部分交互仍较弱,暂不能认定全双工指标已经稳定可用。 +- +-同日进一步检查确认 100K 权重的根网络已退化为近似常数;因此撤回单凭平均姿态误差推荐 100K 整模的建议。诊断结论见[根网络退化检查](KIMODO_ROOT_COLLAPSE_DIAGNOSIS.md)。 +-`best.pt` 实际为 2K 权重:它的去噪验证 loss 最低;初轮六段连续生成的平均姿态误差逊于 100K,但根网络仍有动态响应。 +-初轮评测没有修改训练配置或重新训练;失败权重已在后续修复阶段按用户要求删除。 +- +-评测约定: +- +-- 每类取 native validation index 中第一个符合条件的英文、具有人体配对、预计 6–20 秒的 session。样本在生成前选定,没有根据效果挑选。 +-- 六类:turn-taking、interruption、emergency-stop、backchannel、hesitation、pause。源动作共 88.4 秒;排除每段第一帧后评分 2,646 帧。 +-- 输入为按既定时间因果释放的英文 GT 词、最近 30 秒上下文和已执行的人体历史。无音频输入、无未来文本或阶段标签。 +-- 仅以第一帧 GT 姿态初始化,之后完全使用预测历史。历史 60 帧,规划 60 帧,每次执行 12 帧,30 Hz。 +-- DDIM 20 步、CFG 2.0、seed 12345。官方、2K、100K 使用相同的适配、文本、随机种子和闭环流程。 +-- 每段完整预测及 human GT 各连续 retarget 一次到 G1;不在窗口之间重置 IK,不作位置对齐、DTW 或轨迹修正。 +-- 下表是六段分别计算后等权平均,单位为 rad 或 m。位置误差为三个坐标的逐元素 RMSE。 +- +-| 权重/对照 | 人体关节旋转 RMSE (rad) | G1 关节 RMSE (rad) | G1 根位置 RMSE (m) | G1 根朝向 RMSE (rad) | +-|---|---:|---:|---:|---:| +-| 官方初始化,采用相同 HFD 适配 | 0.628 | 0.464 | 0.433 | 0.600 | +-| best.pt,2K | 0.643 | 0.467 | 0.463 | 0.825 | +-| last.pt,100K | **0.454** | **0.390** | 0.544 | **0.266** | +-| 保持每段初始 GT 姿态 | 0.431 | 0.366 | 0.365 | — | +- +-G1 主指标的参考是同流程 retarget 的 human GT。100K 相比官方初始化,人体姿态误差降低 27.7%,G1 关节误差降低 15.9%,根朝向误差降低 55.7%,但根位置误差增加 25.7%。 +-人体原生空间的位置 RMSE 同样从 0.502 m 升至 0.629 m,所以位置问题并非只由 retarget 造成。 +-GT 活动帧的人体旋转误差从 0.633 降至 0.478 rad;保持初始姿态的相应误差约 0.447 rad。 +-动作幅度和视频均表明 100K 并非完全静止,但在本轮静止/倾听时间较多的样本上,平均误差仍未超过保持初始姿态的简单对照。 +-该对照不能完成动作任务;也不能把低姿态误差直接当成语义任务成功。 +- +-| 验证样本 | 100K 人体旋转 RMSE (rad) | 100K G1 关节 RMSE (rad) | 100K G1 根位置 RMSE (m) | +-|---|---:|---:|---:| +-| turn-taking:向后跳 | 0.321 | 0.247 | 0.385 | +-| interruption:摸头后切换姿态 | 0.533 | 0.478 | 0.226 | +-| emergency-stop:向左行走并收到停止文本 | 0.350 | 0.286 | 0.899 | +-| backchannel:拄拐后退并收到附和文本 | 0.541 | 0.469 | 1.311 | +-| hesitation:拿物体指令中插入犹豫 | 0.572 | 0.529 | 0.242 | +-| pause:拿物体指令中暂停 | 0.406 | 0.332 | 0.201 | +- +-主要闭环问题集中在窗口接缝。将每个 12 帧执行块的第一帧标为重规划边界,100K 的边界根速度误差为 0.839 m/s,块内部为 0.118 m/s。 +-各段边界帧占比平均 8.35%,却贡献平均 81.5% 的根速度平方误差。 +-官方初始化对应为边界 0.724、内部 0.169 m/s;100K 改善了块内运动,但边界仍有明显跳变。 +-曲线还可见部分样本约 2 秒一次的位置脉冲。这些现象在人体输出中已存在;目前证据只支持“生成窗口衔接未解决”,尚不能断定是某一行代码或单一训练因素。 +- +-训练去噪验证 loss:初始 0.371、2K 最低 0.198、100K 为 0.340。 +-这种 GT 历史下的短窗口去噪指标与执行预测历史的整段质量不一致,不能据此简单判断 2K 最好或 100K 只是过拟合。 +-后续应将固定闭环样本的姿态、位置和边界速度一起用于 checkpoint 选择,并优先排查历史/未来交界处的根表示与解码,再考虑连续性监督。 +-当前没有证据支持直接追加长训练即可解决这些问题。 +- +-视频保存在 `/root/autodl-tmp/runs/kimodo-formal-validation-20261023/`: +- +-- `last100k/`:六段预测→G1 与 human-GT→G1 对比,以及六段 human-GT→G1 与原始 G1 对比,共 12 段。 +-- `best2k/`、`official/`:各六段预测→G1 与 human-GT→G1 对比。 +-- `retargeted_human_comparison.mp4` 左侧为预测,右侧为 human GT 经相同 retarget 得到的参考。 +-- `retarget_reference.mp4` 用于查看 retarget 本身的差异。human-GT→G1 对原始 G1 的平均关节误差 0.226 rad、根位置误差 0.045 m;这不是可从模型误差中减去的常数。 +-- 共 24 个有声 MP4,帧数、帧率、音轨和完整音视频解码均通过检查。runs 中仅导出视频,工作数据在 cache 中。 +- +-这是一轮六段、单随机种子的离线轨迹检查,未运行物理跟踪、任务成功率或完整 HFD 验证集评测。 +-可以确认身体姿态训练有收益;进一步诊断已确认 100K 根阶段退化,需修复后再选择正式比较权重,当前不足以保证正式全双工指标“不难看”。 +- +-完整数据与曲线: +- +-- [数值和样本溯源](../reports/kimodo_formal_validation_20261023.json) +-- [训练验证曲线](../reports/kimodo_formal_validation_curve_20261023.png) +-- [六段闭环运动速度曲线](../reports/kimodo_formal_validation_motion_20261023.png) +-- [24 段视频的校验结果](../reports/kimodo_formal_validation_videos_20261023.json) +-- 本轮输入、原生预测、G1 retarget、完整文本释放记录在 `/root/autodl-tmp/cache/kimodo-formal-validation-20261023/`。 +diff --git a/docs/KIMODO_HFD_OVERFIT_RESULTS.md b/docs/KIMODO_HFD_OVERFIT_RESULTS.md +deleted file mode 100644 +index f9830e2..0000000 +--- a/docs/KIMODO_HFD_OVERFIT_RESULTS.md ++++ /dev/null +@@ -1,66 +0,0 @@ +-# Kimodo 英文人体过拟合验收(2026-10-21) +- +-复核结论:**本轮仅通过训练链路检查,闭环过拟合未通过。** 用户指出视频动作偏静止后,运动幅度审计确认三条训练样本的关节运动速度仅约GT的三分之一。训练急停样本保持初始姿态的关节/根误差为0.3607rad/2.0503m,微调闭环为0.3624rad/2.0289m,基本未超过静止基线。下文误差表作为原始记录保留,不应据其宏平均下降宣称动作复现成功。详见评测目录的 `motion_amplitude_audit.json` 和该样本 `motion_audit.json`。 +- +-训练框架、输入和复现命令见 [Kimodo训练说明](KIMODO_HFD_TRAINING.md)。运行目录为 `/root/autodl-tmp/runs/kimodo-english-overfit-20261021`;视频与逐样本指标入口为 [/root/autodl-tmp/runs/kimodo-english-evaluation-20261021/README.md](/root/autodl-tmp/runs/kimodo-english-evaluation-20261021/README.md)。 +- +-## 已完成的检查 +- +-- train/val 各六个原生英文配对人体 session,分别为 84.4/88.4 秒、205/215 个窗口,覆盖六种交互模式。选择规则在训练前固定;不是全库正式训练。 +-- 27 项单元/数据契约测试通过。实际 session 的坐标转换回环误差:平移不超过 6e-8m、旋转不超过 2.8e-6rad;扰动未来不改变输入历史或 heading。 +-- 文本基础模型无缺失权重;MNTP/supervised 两份适配器各 448 个张量与文件逐项一致。冻结编码得到 81 个文本特征,包含空文本的零向量;双向注意力探针通过。动作环境 `pip check` 通过,旧版文本依赖单独隔离。 +-- 完整 282,790,161 参数模型的 root/body 梯度分别非零且有限,一次 Adam 更新实际改变参数;移除文本会改变预测。 +-- 从第 4 步独立恢复两次至第 6 步,全部模型参数、采样器和两步 loss 完全一致;Adam step、scheduler step 均正确恢复。 +- +-## 1,000 步训练 +- +-使用 batch8、AdamW 1e-5→1e-6、bf16,7 项结构损失和 10% 文本 dropout。固定噪声评估每次抽取 64 个窗口,训练/验证独立抽样,不能理解为遍历全数据集。 +- +-| 指标 | 初始 | 1,000步 | +-| --- | ---: | ---: | +-| 固定训练窗口去噪损失 | 1.61369 | 0.24366 | +-| 固定验证窗口去噪损失 | 1.28607 | 0.60110 | +- +-最佳固定验证损失为第 800 步的 0.55017。主闭环对照明确使用最后第 1,000 步,避免把窗口损失最优当成闭环最优。每步中位耗时 0.0896s,峰值显存 5.33GiB;第 6→1,000 步续训进程共约 107 秒,含验证和保存。此吞吐只代表小型预制窗口与缓存文本,不代表全量动态数据加载和 8B 文本编码成本。 +- +-## 闭环对照 +- +-训练和验证各选 turn-taking、interruption、emergency-stop 三条 session;官方初始化/微调模型均使用相同 seed、DDIM20、CFG2、单帧 GT 初始化和后续预测历史。下表为各三条的逐 clip 宏平均,无根对齐或 DTW。 +- +-| 范围 | 关节旋转RMSE,初始→微调(rad) | 根位置RMSE,初始→微调(m) | +-| --- | ---: | ---: | +-| 训练三条 | 0.5711 → 0.3492 | 1.5246 → 0.7362 | +-| 验证三条 | 0.6580 → 0.4471 | 0.9332 → 0.4737 | +- +-验证宏平均关节误差下降约32%,根位置误差下降约49%。这是小规模探索性验证,三条样本不足以证明泛化性能。 +- +-| 样本 | 关节初始→微调(rad) | 根位置初始→微调(m) | +-| --- | ---: | ---: | +-| train turn-taking 13 | 0.5262 → 0.3260 | 0.8745 → 0.0323 | +-| train interruption 73 | 0.6752 → 0.3591 | 1.0668 → 0.1476 | +-| train emergency-stop 36 | 0.5118 → 0.3624 | 2.6323 → 2.0289 | +-| val turn-taking 3 | 0.5458 → 0.3664 | 1.0103 → 0.3108 | +-| val interruption 75 | 0.7042 → 0.4871 | 1.0318 → 0.0803 | +-| val emergency-stop 200 | 0.7240 → 0.4877 | 0.7574 → 1.0299 | +- +-每条完整预测与人体 GT 分别做 GMR,再与原始 G1 作三组比较;主对照共 36 条带音轨视频。人体指标先于 retarget 计算,GT→G1 的差异单独保留,不能把它当作可直接相减的误差下界。视频为姿态回放,不是物理跟踪或任务成功率评测。 +- +-加上两组历史/CFG诊断共42条视频,全部逐帧解码通过,共16,812帧,音轨也全部可解码。校验清单在评测目录 `video_checks.json`。这次20步DDIM的单轮规划平均约0.46–0.65秒,尚未满足每0.4秒执行窗口的严格实时要求;当前验收针对离线推理。 +- +-## 急停样本与历史诊断 +- +-训练急停样本的 GT 先行走约 5 米再停下,微调闭环预测大多停留在起点附近。它的问题主要是没有启动该段行走,不能简单解释成“听到停止后仍一直向前”。该样本原始动作描述为 `stop fast walk forward`,随后释放 `Pause for a moment.`;没有为改善结果改写文本。 +- +-| train emergency-stop 36诊断 | 关节RMSE(rad) | 根位置RMSE(m) | +-| --- | ---: | ---: | +-| 预测历史,CFG2,主结果 | 0.3624 | 2.0289 | +-| GT历史,CFG2,仅诊断 | 0.0896 | 0.0910 | +-| 预测历史,CFG1,仅诊断 | 0.4779 | 2.0421 | +- +-这说明模型在给定正确历史时能较好地拟合该轨迹,但闭环可能停留在静止状态;降低 CFG 未解决问题。GT历史结果不能替代主闭环指标,也不能独立证明唯一原因。验证急停样本的根误差还从 0.757m 升到 1.030m,因此不能宣称已经学会急停。 +- +-结论是微调框架与梯度路径可用,小规模训练已改善多数样本;当前仍有闭环启动、状态切换和根运动不足的问题。下一阶段应扩大训练覆盖,增加连续多轮训练/预测历史扰动,并单列启动与停止指标,再决定正式训练预算;只延长这六条 session 的训练不能保证解决上述问题。纯文本输入在词释放前也没有语音起点信息。 +- +-## 留存 +- +-运行目录保留配置、数据/文本审计、源码与环境快照、loss 曲线和完整 resume 检查。当前 `best.pt` / `last.pt` 在 `/dev/shm/kimodo-english-overfit-20261021`,属于临时测试权重,重启不保留;两份独立恢复测试的重复 checkpoint 已清理,日志仍保留。GenMo 的 GPU1 正式训练未停止。 +diff --git a/docs/KIMODO_HFD_TRAINING.md b/docs/KIMODO_HFD_TRAINING.md +deleted file mode 100644 +index 4810ba7..0000000 +--- a/docs/KIMODO_HFD_TRAINING.md ++++ /dev/null +@@ -1,112 +0,0 @@ +-# Kimodo HFD 全量微调与小规模验收 +- +-本地入口保留 Kimodo-SMPLX-RP-v1 的 root→body 两阶段网络与官方初始化,增加监督训练、完整断点恢复、连续推理、人体指标、整段 GMR 和视频评测。上游代码固定在 `e1c219c99b4b94ba48be1fb6a1cbc64ecee4c954`,不修改上游实现。 +- +-2026-10-24:修复版已从官方初始化完成新一轮100K训练;当前best与last均为100K,新正式权重保留。根骨干200个参数张量仍与官方初始化逐位一致,梯度和时序响应正常,未重现旧恒定输出失效。六段验证的G1根位置RMSE为0.264m,但后退与部分指令跟随仍不足。完整数值和有声视频见[本轮正式评测](HUMAN_FORMAL_VALIDATION_20261024.md)。 +- +-训练方式见[训练修复与验收](KIMODO_TRAINING_REPAIR.md):冻结根 Transformer 骨干,仅微调根输入/输出投影,身体保持训练;增加实际解码运动连续性监督,分阶段学习率及梯度裁剪。此前失败正式训练的run和权重已经删除,当前同名目录属于新训练。以下早期过拟合结果保留为历史对照。 +- +-历史混合闭环过拟合未通过。较好的早期诊断路线为归一化x0目标、2,000步初训加1,000步GT历史微调:行走样本达到约0.056rad/0.040m,但其他模式相位与泛化仍不足。历史实验见[静止退化复查](KIMODO_MOTION_RECOVERY.md),完整对照见[闭环稳定性报告](KIMODO_CLOSED_LOOP_STABILITY.md)。`kimodo_overfit.yaml`和`kimodo_refine_probe.yaml`复现这条诊断路线;`kimodo.yaml`现为全量正式训练配置。未默认启用在线历史混入。旧缓存需按当前编码重新准备。 +- +-## 输入、目标与因果边界 +- +-- 原生英文 human 配对样本,按实际 session 组合操作拼接人体动作;G1 只用于配对和评测参考,不进入 Kimodo 输入或损失。小规模验收固定 train/val 各六个 session;正式入口覆盖完整英文原生索引,按需组装配对人体 session。 +-- 文本沿用共同协议:对齐后的前景英文 GT 单词,词结束加 0.1s 后等待麦克风 `0.6+0.4k` 释放,保留最近 30s。动作时刻 `t` 对应麦克风 `t+0.2s`。无音频、翻译、未来文本、交互类型标签输入。 +-- 冻结 LLM2Vec/Llama-3-8B-Instruct,MNTP 适配器合并,再加载 supervised 适配器;逐项核对权重。原生 mean pooling 得到 `1×4096`,网络内部的 50 个文本 token 由其自身投影构造。文档预算 400 token,超长保留最近文本;模板总预算 512,编码内部 batch=1。 +-- 30Hz,历史60帧、默认预测60帧、执行12帧。12帧预测仅用于诊断,未采纳为默认。历史mask锁定203维姿态,70维速度/接触辅助量不作为已知约束。当前配置使用因果root特征分解,使历史在输入和监督中一致,不改变物理人体动作;首个已知姿态确定heading与平移坐标系。 +-- 输出原生 273D 特征,由原生 inverse 解码到 22 个 SMPL-X 关节和根轨迹,再逆变换到 AMASS;手部中性。保留 session 体型,显式处理其 pelvis 偏移。完整输出与人体 GT 分别 GMR 到 G1。 +- +-当前`training.objective: normalized_x0`在归一化273D特征上做x0均方误差,root与body分别取均值后相加;仅未来有效帧计分,末端无下一帧时速度/接触不计分。保留原生两阶段之间的detach,两阶段各有直接梯度。诊断配置使用50% t=999、其余均匀t,以及70%运动窗口抽样。原七项物理量SmoothL1配方通过`physical_huber`保留作对照,YAML的`training.weights`只在该模式生效;不宣称复现官方未公开训练配方。 +- +-推理从一个已知初始姿态启动,后续历史来自已执行预测;CFG两支保留相同历史。历史70维未观察辅助量在每个DDIM步恢复为训练时的纯噪声边缘分布,避免把未监督的历史预测继续反馈;`legacy`仅用于旧规则对照。默认DDIM20、CFG2,无轨迹后处理。`--teacher-forced`仅诊断使用。 +- +-## 环境 +- +-动作使用 `/root/autodl-tmp/envs/kimodo/bin/python`,保留上游 Transformers 5.1.0。原生 LLM2Vec 的双向 Llama 代码需要独立旧版依赖,在 `/root/autodl-tmp/envs/kimodo-text-compat` 保存 `requirements/kimodo-text.txt` 所列包;仅文本进程将其加入 Python 路径。可用 `KIMODO_TEXT_COMPAT` 指向其他安装目录。 +- +-重新安装文本 overlay 时,应先用用户要求的 `/root/.local/bin/turbo` 调用 pip,并用 `--no-deps --target <目录> -r requirements/kimodo-text.txt`,复用动作环境已有的 Torch 等依赖。不要直接将该 requirements 安装进动作环境。文本 worker 自检未来 token 能影响前面 hidden state,防止误用单向注意力;小规模离线缓存完成后动作训练不加载 8B 编码器。全量训练采用常驻独立文本进程,与动作训练共用所选 GPU;缓存命中时不重复编码。 +- +-## 全量正式训练(2026-10-23 修复版) +- +-```bash +-cd /root/autodl-tmp/hfd-baseline-portable +-/root/autodl-tmp/envs/kimodo/bin/python -u train_kimodo.py --config configs/kimodo.yaml +-``` +- +-中断恢复使用相同命令追加 `--resume`,读取完整 `last.pt`,恢复模型、Adam、余弦调度器、全局 RNG、抽样 RNG 和每个 batch 槽位的 session;100K 是总步数。配置/实现/数据身份变化会拒绝恢复。 +- +-- GPU 0;batch 8;100K 步;BF16。根仅训练输入/输出投影,AdamW 2e-6→2e-7,在20K衰减到底;身体AdamW 1e-5→1e-6,在100K衰减到底;500步warmup,两阶段分别裁剪到1。因果root、normalized x0、10%强制t=999(其余均匀t)、文本dropout 0.1;执行12帧权重1,剩余规划帧0.25。增加原始根历史重建及实际解码速度/加速度监督;根注意力采用数学后端,运行配置随checkpoint保存并恢复。全量所有合法窗口均匀边缘抽样,不启用在线历史课程。具体权重以`configs/kimodo.yaml`为准。 +-- train 原生英文 155,117 个 session;val 独立原生英文 4,719 个 session;没有6–20秒筛选、没有每模式数量上限、无第三方G1动作、无中文。初始化统计会打印真实覆盖数。 +-- 不预拼155K个session。每个 batch 槽位独立按保守长度上界拒绝抽样,再均匀选0.4s决策点;训练槽位复用16个batch,验证复用4个batch。所有session有被抽到的机会,预算内不保证每个session都被访问;100K不是一个epoch。 +-- paired SMPL-X 动态拼装与 GenMo 相同来源和时间轴;本地512MiB窗口LRU、每split512MiB原始sessionLRU;冻结LLM2Vec独立进程按需编码,每次内部batch1,SQLite缓存本地 `/root/autodl-tmp/cache/kimodo-text-full`,特征payload上限4GiB,清缓存不改变抽样顺序。编码器不反传;主模型正常反传。 +-- 文本仍是按单词释放的最近30s英文GT,60帧历史,60帧目标,12帧执行,无音频或未来语义输入。正式训练依然用GT历史;闭环推理用预测历史。小样本泛化差属于质量风险,继续通过全量独立验证和闭环视频观察。 +-- 首次启动要构建train/val原生索引、加载8B文本模型并编码首次抽样文本,启动阶段可能数分钟。`FULL_DATA`、`DATA`是正常准备进度。每10步打印`TRAIN`;`seconds`含数据等待和更新,`data_seconds`单独显示组装/文本等待。 +-- 每2K步执行固定种子16×8个验证窗口,并记录同种子训练诊断;这些窗口从完整val索引抽取,绝非全val遍历。噪声损失用于选`best.pt`,不能替代闭环动作评测。 +-- 日志/配置在 `/root/autodl-tmp/runs/hfd-kimodo-english-train`;权重在 `/root/autodl-tmp/models/kimodo/hfd-english-train`,保留完整last和最近2个完整快照,以及仅模型best。本地唯一临时文件完成flush/fsync后原子替换;写入失败最多重试3次,保留上一断点。快照用硬链接避免重复写入。 +- +-训练结束后,使用正式run选择六模式验证session,执行预测历史→整段GMR→带音频视频: +- +-```bash +-/root/autodl-tmp/envs/kimodo/bin/python -u evaluate_kimodo.py \ +- --run /root/autodl-tmp/runs/hfd-kimodo-english-train \ +- --output /root/autodl-tmp/runs/kimodo-full-validation \ +- --splits val --variants trained --gpu 0 +-``` +- +-全量评测按每模式第一个满足6–20s的原生英文session作固定可视化子集,动作G1仅评测导出,训练不加载。推理可额外使用`--checkpoint`选择其他完整快照。评测GPU需要容纳动作模型和独立文本模型;勿与该GPU上的训练并发。批量全val指标可后续扩大选择,当前命令只展示六例。 +- +-启动验收:44项单元测试通过;实际全量入口完成2步、保存完整断点后恢复到3步,核对下一批位置、session槽位、抽样RNG、Adam和调度器。运行统计身份误变问题已修复。临时测试权重已删除;这不是100K质量验收。报告见 [`kimodo_full_training_ready_20261021.json`](../reports/kimodo_full_training_ready_20261021.json)。 +- +-## 小规模诊断命令 +- +-```bash +-cd /root/autodl-tmp/hfd-baseline-portable +-PY=/root/autodl-tmp/envs/kimodo/bin/python +-$PY prepare_kimodo.py --config configs/kimodo_overfit.yaml +-$PY cache_kimodo_text.py --run /root/autodl-tmp/cache/kimodo-overfit --gpu 0 +-$PY train_kimodo.py --config configs/kimodo_overfit.yaml --offline-check +-$PY train_kimodo.py --config configs/kimodo_overfit.yaml +-# 中断后恢复,max_steps 是总预算: +-# $PY train_kimodo.py --config configs/kimodo_overfit.yaml --resume auto +-``` +- +-复现第二阶段时,复用首阶段的冻结输入,再从其权重warm-start(首阶段权重需暂时保留): +- +-```bash +-$PY - <<'PY' +-from pathlib import Path +-from omegaconf import OmegaConf +-base = OmegaConf.load('configs/kimodo_overfit.yaml') +-refine = OmegaConf.load('configs/kimodo_refine_probe.yaml') +-destination = Path(refine.output) +-destination.mkdir(parents=True, exist_ok=True) +-for name in ['data.json', 'text_features.pt', 'text_features.json', 'texts.json']: +- target = destination / name +- if not target.exists(): +- target.symlink_to(Path(base.output) / name) +-OmegaConf.save(refine, destination / 'resolved_config.yaml') +-PY +-$PY train_kimodo.py --config configs/kimodo_refine_probe.yaml +-``` +- +-完整checkpoint包含网络、AdamW、cosine scheduler、全局step、Torch/CUDA/Python/NumPy随机状态和窗口采样器;启用在线历史时还包含历史缓存、对应policy step与混入选择随机数。恢复核对数据、缓存文本、代码、官方权重/统计及配置。`--stop-after 4`可用于检查。`best.pt`按固定验证去噪损失选,`last.pt`保存最后状态;闭环质量另行评测。逐位复现需显式启用确定性CUDA配置,见最新报告。 +- +-```bash +-$PY evaluate_kimodo.py \ +- --run /root/autodl-tmp/cache/kimodo-overfit-refine \ +- --output /root/autodl-tmp/cache/kimodo-evaluation \ +- --checkpoint /dev/shm/hfd-kimodo-overfit-refine/last.pt --gpu 0 \ +- --modes turn-taking interruption emergency-stop backchannel hesitation pause +-``` +- +-默认分别报告训练/验证的 turn-taking、interruption、emergency-stop,每个对照官方初始化与微调模型。先计算人体根/旋转误差,再按每个完整轨迹做 GMR,保存 predicted human↔retargeted GT、predicted human↔原始 G1、retargeted GT↔原始 G1 三组比较。视频含原始音轨和文本释放记录。两种模型共用 seed、采样步数、CFG 与历史策略;验证集用于诊断,不作为独立测试集。 +- +-首阶段YAML为2,000步、batch8、AdamW 3e-5→3e-6;`kimodo_refine_probe.yaml`再从首阶段模型权重初始化,重置Adam,以1e-5→1e-6训练1,000步。第二阶段不是`--resume`,而是显式warm-start。工作数据和日志在cache,临时权重在/dev/shm,runs只存视频。正式训练仍需另行实现更大session范围并验证预算与泛化。 +- +-## 验收产物 +- +-cache中的工作目录记录数据、文本审计、完整模型梯度检查与训练指标;runs只保存筛选后的MP4。首轮失败诊断见[过拟合验收](KIMODO_HFD_OVERFIT_RESULTS.md),最新对照、视频与保留范围见[闭环稳定性](KIMODO_CLOSED_LOOP_STABILITY.md)。 +- +-2026-10-22 保存故障修复:旧训练在64K保存时无法在共享盘创建临时文件,62K完整断点及best已归档到 `/root/autodl-fs/pangxinlei/models/kimodo/archive/hfd-english-train-failed-20261022`。默认正式配置改用本地盘保存新权重,旧失败run的日志等已清理。标准启动命令从头训练;旧归档含匹配原断点契约的训练脚本副本,新代码不会自动绕过旧断点的实现校验。见[故障报告](../reports/kimodo_checkpoint_failure_20261022.json)。 +- +-2026-10-23 旧一轮正式训练完成 100K 步。六类 held-out session 的历史数值与 24 段有声视频见[正式训练验证](KIMODO_FORMAL_VALIDATION.md)。进一步确认 100K 根网络退化为近似常数;替换早期根权重也尚未恢复正确后退。随后加入的退化检测和自动停止已按用户要求全部撤回。本轮改为修复训练策略,旧失败run及其权重已全部删除,历史视频和分析数据保留;具体原因见[训练失效机制](KIMODO_ROOT_TRAINING_CAUSE_ANALYSIS.md)。 +diff --git a/docs/KIMODO_MOTION_RECOVERY.md b/docs/KIMODO_MOTION_RECOVERY.md +deleted file mode 100644 +index 6cc4edb..0000000 +--- a/docs/KIMODO_MOTION_RECOVERY.md ++++ /dev/null +@@ -1,78 +0,0 @@ +-# Kimodo 静止退化复查与恢复实验 +- +-本文保留历史恢复实验。后续在线历史课程、因果root表示修正和同预算预测时长对照见[最新闭环报告](KIMODO_CLOSED_LOOP_STABILITY.md);混合闭环验收仍未通过。 +- +-用户观看视频后指出预测几乎不动。复查确认:[原1,000步实验](KIMODO_HFD_OVERFIT_RESULTS.md)的闭环过拟合**没有通过**。宏平均误差下降不足以证明动作复现;先前报告对此解释不充分。 +- +-## 直接证据 +- +-三条原训练评测的预测关节运动速度仅为GT的32%–36%。`train-emergency-stop-000036` 的根路径为0.83m,GT为5.29m;一直保持初始姿态的关节/根误差为0.3607rad/2.0503m,模型为0.3624rad/2.0289m,几乎没有超过静止基线。 +- +-新增 `audit_kimodo_motion.py --inference <目录>` 同时报告原人体误差、静止初始姿态基线、根路径、关节角速度、GT活跃区间误差和持续运动启动时间。运动幅度不能单独证明正确性:抖动和错误动作也可能增加路径/角速度。启动检测仅为诊断,使用根速度>0.15m/s持续0.2s,不是通用任务成功判据。 +- +-## 训练改动与对照 +- +-输入协议保持不变:已释放英文文本、60帧历史、60帧未来、每轮执行12帧;无音频、GT阶段、未来姿态或剩余时长条件。文本缓存和预训练动作参数均沿用原实验。 +- +-`hfd_bridge/kimodo_recovery.py` 增加可选训练组合: +- +-- 在root阶段输出传给body阶段之前锁定已知历史root,保持两阶段条件一致。旧实现仅在两阶段完成后锁定最终输出。单独给旧权重加此约束,根误差仅2.0289→2.0265m,**它本身没有解决静止问题**。 +-- 原速度损失针对独立的velocity特征块,不能保证实际解码姿态在动。增加由解码root、FK关节和旋转增量计算的时序监督,权重为1、2、0.5;不是只拟合辅助速度输出。 +-- 每批50%样本使用t=999,增加从近纯噪声依靠文字/历史生成的训练;其余扩散步仍均匀采样。 +-- 训练按GT运动标签将70%抽样分给明显运动窗口,保留30%等待/停止窗口。六条原始session有83个活跃窗口、122个其余窗口。标签只用于训练抽样,不输入模型。 +- +-## 单样本、同预算对照 +- +-均从原六条训练的1,000步模型初始化,只训练同一条急停/行走session,新增1,000步、batch8、LR3e-5、相同cosine预算、text dropout=0、同seed。原配方对照保留七项损失与均匀窗口/扩散步;恢复配方同时启用上述改动。单条实验是定位问题,不是泛化评测。 +- +-| 配方(预测历史、CFG2) | 关节RMSE(rad) | 根RMSE(m) | +-| --- | ---: | ---: | +-| 原六条1,000步模型 | 0.3624 | 2.0289 | +-| 原配方单条新增1,000步 | 0.3164 | 0.5658 | +-| 恢复组合单条新增1,000步 | 0.0427 | 0.1332 | +- +-恢复组合开始真正行走,持续运动起点约3.17s,GT约3.33s;关节速度是GT的1.23倍。不过根路径7.31m仍高于GT5.29m,有抖动/多余位移。CFG1诊断为0.0437rad/0.0759m,根路径7.08m;不能只看较低的位置误差忽略多余路径。原配方更充分训练也改善根误差,且恢复实验同时改变多项因素,不能将改善完全归因于某一项损失。 +- +-[单条真实闭环视频](/root/autodl-tmp/runs/kimodo-motion-recovery-20261021/probe1000/evaluation/retargeted_human_comparison.mp4),仅第一帧GT初始化;[动作/位移曲线](/root/autodl-tmp/runs/kimodo-motion-recovery-20261021/single_motion_comparison.png)。 +- +-## 六条混合实验 +- +-`configs/kimodo_motion_recovery_six.yaml` 从原1,000步模型继续微调3,000步,恢复10%文本dropout,默认CFG2。六种训练模式均用模型自己历史做完整rollout。 +- +-| 训练模式 | 关节RMSE(rad) | 预测/GT关节角速度比 | +-| --- | ---: | ---: | +-| turn-taking | 0.3158 | 0.360 | +-| interruption | 0.3138 | 1.175 | +-| emergency-stop | 0.0912 | 1.354 | +-| backchannel | 0.3228 | 1.110 | +-| hesitation | 0.5201 | 0.269 | +-| pause | 0.4859 | 0.568 | +- +-因此单条恢复**没有自动推广到六条混合过拟合**。行走改善,部分舞蹈仍运动不足;幅度接近GT的样本也可能动作/相位错误。hesitation给GT历史时关节误差0.0630rad,而闭环为0.5201rad,说明历史状态分布差异仍然很大。GT历史只用于诊断,不计入验收结果。 +- +-## 预测历史训练 +- +-`prepare_kimodo_predicted_history.py` 仅对六条训练session增加一份由已完成模型闭环生成的历史视图,验证集保持原样。每个训练窗口的60帧输入来自缓存的真实预测历史;未来监督来自GT。专家相对位移从当前预测根位置出发,避免要求瞬移回已错过的GT位置;前6帧身体旋转用0.2s过渡接向专家姿态。此过渡属于训练目标构造,不是推理后处理。 +- +-原GT历史视图与预测历史视图各205个窗口,训练均使用最近已释放文字。数据记录输入预测的哈希与plan ID,并拒绝GT历史rollout;`augmentation`视图不当作额外独立评测session。新增 `configs/kimodo_predicted_history.yaml`,从六条3,000步模型初始化,再训练2,000步。不能把六条的两种历史视图称为12条独立训练session。 +- +-所有阶段均为固定小规模数据诊断,不代表全库训练已经准备就绪。后续应按闭环动作幅度、与静止基线的比较、动作一致性与启动/停止共同判断,不能以去噪loss或平均姿态误差单独宣布通过。 +- +-## 预测历史实验结果与结论 +- +-完成新增2,000步后,仍使用相同seed、DDIM20、CFG2、一个GT初始姿态和后续完整预测历史,评估原始六条训练session(不把增强视图当成新样本)。 +- +-| 训练模式 | 关节RMSE(rad) | 预测/GT关节角速度比 | +-| --- | ---: | ---: | +-| turn-taking | 0.4061 | 1.458 | +-| interruption | 0.3091 | 1.217 | +-| emergency-stop | 0.1933 | 1.910 | +-| backchannel | 0.4659 | 1.615 | +-| hesitation | 0.3127 | 1.197 | +-| pause | 0.3637 | 1.403 | +- +-角速度比的精确数值以各session的 `motion_audit.json` 和汇总 `motion_recovery_comparison.json` 为准。相比上阶段,hesitation/pause的静止退化缓解,但有些模式更不准确。急停根位置误差0.7804m、速度RMSE3.9659m/s,明显抖动;turn-taking虽然根位置误差只有0.0595m,关节误差仍0.4061rad。这再次说明移动更多或根误差更小,都不能代替动作复现。 +- +-**当前六条混合闭环过拟合仍未通过。** 不将这两个混合恢复版本替换为正式配方,不推荐直接扩大全量训练。单条真实动作已经可以学到,但混合数据的闭环启动、状态切换、动作相位和轨迹稳定性尚不可靠。实验支持历史分布差异的重要性,没有证明它是唯一原因;缓存一轮预测历史也没有充分解决问题。现阶段继续盲目增加正式训练步数,无法保证排除静止或错误动作。 +- +-本次30项回归测试通过;预测历史缓存的因果边界、输入预测哈希和已知物理历史一致性检查通过。链路正确性与动作验收分开报告。所有新实验保留数据盘日志和视频,测试权重仍暂存在`/dev/shm`;原GenMo GPU1训练未停止。 +diff --git a/docs/KIMODO_ROOT_COLLAPSE_DIAGNOSIS.md b/docs/KIMODO_ROOT_COLLAPSE_DIAGNOSIS.md +deleted file mode 100644 +index 6fa6855..0000000 +--- a/docs/KIMODO_ROOT_COLLAPSE_DIAGNOSIS.md ++++ /dev/null +@@ -1,55 +0,0 @@ +-2026-10-23:后退失败的直接原因已定位到 **100K 权重中的根轨迹网络退化为近似常数输出**。 +-窗口坐标转换与已知历史保留检查通过。单纯换回早期根网络仍未恢复正确后退,因此还需修复根网络训练稳定性与闭环动作续接,不能将本轮替换对照视为可用修复。 +- +-证据来自已有 validation 样本 `val-backchannel-000057`,未用该样本训练: +- +-- 输入正确:3.0 秒已释放完整 `walking backwards with crutches`,7.8 秒才追加 `Wow.`。去掉 `Wow.` 后仍反向漂移。 +-- GT 的人体坐标编码→解码→导出最大平移误差约 0.0000083 m;采样末尾的已知历史解码误差为 0。 +-- 在决策帧 0、132、240、396,使用预测历史或 GT 历史,100K 根网络都输出几乎相同的未来位置。 +-- 在几乎无噪声的输入(t=0,提供加噪后的 GT 未来)上也如此。t=526/999、FP32/BF16、train/eval 均复现,排除了仅由半精度或评估模式引起的异常。 +-- 官方初始化和 2K 网络在相同输入下输出随时间变化的根轨迹。 +- +-下面对八个 GT 存在明显位移的窗口进行额外检查,混合 t=0/526/999。运动 RMS 定义为有效未来帧相对其时间均值的三维位置偏差 RMS,使用根阶段的 smooth-root 通道,不让身体阶段的髋部残差掩盖根阶段退化。 +- +-| 网络 | 预测根轨迹运动 RMS (m) | GT RMS (m) | 近似常数窗口 | +-|---|---:|---:|---:| +-| 官方初始化 | 0.1594 | 0.1958 | 0 / 8 | +-| 2K | 0.2381 | 0.1958 | 0 / 8 | +-| 100K | **0.000057** | 0.1958 | **8 / 8** | +- +-现有窗口以已知历史第一帧确定原点/朝向,编码解码的逆变换本身成立。 +-但退化网络不断在该原点附近输出几乎相同的位置,而不是延续历史末尾位置。 +-当历史以 60 帧长度滚动时,原点跟随过去预测变化,从而出现约两秒一次的位置脉冲。 +-这是“近似常数预测 + 滚动窗口原点”共同产生的现象,不能简单认定为坐标轴翻转或平移加错一次。 +- +-本轮进行了仅替换根网络、保留 100K 身体网络的闭环消融,均保持原始因果文本、随机种子、窗口和采样配置,不添加轨迹修正或 GT 历史: +- +-| 组合 | 沿 GT 后退方向的净位移 (m) | 人体位置 RMSE (m) | 人体姿态 RMSE (rad) | +-|---|---:|---:|---:| +-| GT | +3.498 | — | — | +-| 100K 整模 | -0.780 | 1.513 | 0.541 | +-| 2K 根 + 100K 身体 | -0.075 | 1.300 | 0.570 | +-| 官方根 + 100K 身体 | +0.183 | 1.132 | 0.706 | +- +-替换缓解了漂移,但没有完成约 3.5 m 的后退;身体姿态误差还增大。 +-这表明常数根网络确实是问题之一,现有早期根权重也不足以解决完整 HFD 动作续接。 +-两个替换对照均已整段 retarget 并保存有声视频: +- +-- `/root/autodl-tmp/runs/kimodo-root-diagnosis-20261023/root2k-body100k/backchannel.mp4` +-- `/root/autodl-tmp/runs/kimodo-root-diagnosis-20261023/root-official-body100k/backchannel.mp4` +- +-训练日志提供了时间线索:1–6K 的平均 normalized-root loss 约 0.042,6.5–7K 升至约 0.149,7–8K 约 0.220,而身体 loss 仍在缓慢下降。 +-没有保留这个区间的逐步权重,无法确定精确退化时刻,也未分别消融学习率、高噪声比例或样本分布来确定触发因素。 +-可以确定训练后的根阶段失效;不能把“原因就是学习率过大”等猜测当作已验证结论。 +- +-用户随后明确要求分析原因,不加入退化检测。上一轮加入的检测模块、测试、配置和自动停止已全部撤回,训练脚本恢复为本次 100K 正式训练时的原版本。原因分析见[根网络训练失效机制](KIMODO_ROOT_TRAINING_CAUSE_ANALYSIS.md)。 +- +-正式模型权重与已有视频均保留。本轮没有追加训练,没有生成临时训练 checkpoint。 +-旧 100K 训练脚本快照保存在 `/root/autodl-tmp/models/kimodo/hfd-english-train/source-100k-20261023/train_kimodo.py`。 +-训练脚本已恢复到旧正式训练契约的相同内容;诊断推理仍可正常加载全部旧权重。 +- +-下一轮训练应从健康根权重恢复,分别控制根阶段与身体阶段的更新,加入实际解码位移/接缝监督,并用固定闭环验证选择权重。 +-100K 身体阶段的姿态收益可作为初始化候选;当前不能把 100K 整模或上述混合模型直接作为已经修复的正式 baseline。 +- +-[完整检查数据](../reports/kimodo_root_collapse_20261023.json)。 +-逐窗口特征、模式/精度对照与诊断推理保存在 `/root/autodl-tmp/cache/kimodo-formal-validation-20261023/root-probe/`。 +diff --git a/docs/KIMODO_ROOT_TRAINING_CAUSE_ANALYSIS.md b/docs/KIMODO_ROOT_TRAINING_CAUSE_ANALYSIS.md +deleted file mode 100644 +index aa08523..0000000 +--- a/docs/KIMODO_ROOT_TRAINING_CAUSE_ANALYSIS.md ++++ /dev/null +@@ -1,59 +0,0 @@ +-2026-10-23:本轮分析的核心结论是,微调后的根网络丢失了时序和条件信息,随后落入前层梯度接近零的常数回归状态。上一轮所说的“根网络退化”只是输出表现;这里补充网络内部如何失效、为何训练无法自行恢复,以及哪些训练设置最值得怀疑。 +- +-本文记录旧100K训练的原因分析。分析阶段按用户要求撤回全部退化检测及自动停止,恢复了旧训练代码和配置。用户随后授权修复训练框架并删除失败目录;当前代码和配置已按[训练修复](KIMODO_TRAINING_REPAIR.md)更新,失败run及其全部权重已删除,既有数值、分析缓存和评测视频保留。以下机制测量均来自删除前保存的诊断结果。 +- +-已确认的内部机制: +- +-1. 根 Transformer 的不同时间帧逐层变得相同。在同一 GT 历史、同一接近无噪声未来输入下,2K 网络最后一层的帧间特征 RMS 为 0.0489,100K 只有约 1.08e-7。100K 第 1 层仍有约 0.0601 的差异,到第 8 层约 0.000854,第 13 层约 3.56e-7。信息在 Transformer 内部被抹平,早于坐标解码。 +-2. 深层注意力同时趋于均匀。该输入共 172 个 token;均匀注意力的熵为 ln(172)=5.14749,100K 最后一层实测约 5.14749,2K 约 1.0776。不同帧的同质化与近似全局平均相互强化,最后输出不再区分时间或条件。 +-3. 前层几乎无法再获得有效梯度。以同一根 MSE 对照,2K 对输入的梯度范数为 0.0170,100K 为 4.76e-12,下降约 35.8 亿倍。100K 输入投影权重梯度约 9.50e-11,但输出层权重梯度仍约 8.14。因此“根分支总梯度非零”不代表整个根网络仍能学习;末端有更新,前层已经很难恢复时序信息。 +-4. 在这种状态下,当前 MSE 的最佳行为是调整末端,使固定输出接近训练目标的平均值。实际抽样验证了这一点,见下表。这是“进入近似常数状态后为什么会维持这个平均答案”的解释;不代表 MSE 的全局最优只能是常数。 +- +-网络内部的这些测量使用 FP32、普通数学注意力实现,避免把加速后端的额外差异混入梯度结论。之前的 train/eval、FP32/BF16 对照也都确认 100K 输出近似常数,因此最终失效本身并非单一评估模式的现象。 +- +-对原正式训练采样器以 seed 42 抽取前 1,024 个窗口,覆盖 64 个 session;维持 batch8、16 批刷新一次的原约定: +- +-| 项目 | 结果 | +-|---|---:| +-| 根轨迹运动 RMS <1 cm 的窗口 | 49.3% | +-| 根轨迹运动 RMS ≥3 cm 的窗口 | 39.3% | +-| 尚无已释放文本的窗口 | 6.7% | +-| 仅输出样本目标均值的最优常数根 MSE | 0.22739 | +-| 100K 模型固定输出在同一批目标上的根 MSE | 0.22990 | +- +-100K 的固定答案只比“完全忽略输入、始终回答平均值”的最优常数解差约 1.1%。这与均值回归状态相符。 +-同时约 39% 窗口有明显位移,2K 网络也能利用输入得到更低误差,不能把问题简单归结为“数据全是静止”或“输入没有动作信息”。这只是 64 个 session 的抽样,不是完整数据集比例普查。 +- +-当前训练实现让这种状态难以被其他目标纠正: +- +-- 根阶段和身体阶段是两个 Transformer。原生 `TwostageDenoiser.forward` 在训练时以 `torch.no_grad()` 将根预测转成身体条件,并显式 `detach()`。所以身体阶段即使发现行走不对,其损失也不能传回根网络。 +-- 正式目标是 `normalized_x0`:根阶段五个特征的 MSE,加身体阶段特征的 MSE。YAML 中 `weights.root_position`、`joint_velocity` 等物理损失权重属于另一个 objective,本次没有启用。 +-- 身体输出中的 `velocities` 是独立辅助通道。导出人体运动时实际位置由 smooth-root 和髋部位置残差解码,未通过积分这些速度通道生成。辅助速度通道学对,不等于实际导出的根位移正确。 +-- 本次未配置 `recovery.temporal_weights`,因此没有针对实际解码轨迹速度、接缝速度或加速度的额外监督。 +-- 已知历史在输出端会被覆盖回观测值,且不计入未来损失。因而“历史输出保留准确”并不能证明 Transformer 内部仍在利用历史:根网络内部可以已经忘记历史,输出检查仍通过。 +- +-这些结构事实解释了为什么身体 loss 可以继续下降、整体梯度也非零,而根网络的坏状态得不到修正。 +-它们并不单独证明某一条代码必然使模型退化;健康 2K 权重已经说明同一结构可以暂时正常学习。 +- +-最值得怀疑的触发设置,是较强的根网络全参数更新与高噪声目标组合: +- +-- 根和身体共享 AdamW、3e-5 初始学习率和按 100K 步拉长的 cosine 调度,没有根阶段独立的更新预算。 +-- 2K 时学习率仍为 2.997e-5,6.5K 为 2.972e-5,8K 为 2.958e-5,几乎仍处于初始强度。 +-- 日志中 1–6K 的平均根 loss 约 0.042;6.5–7K 约 0.149,7–8K 约 0.220。身体 loss 同期仍缓慢下降。这是优化恶化的信号,不是仅在未见过的验证集上变差。 +-- `high_noise_fraction=0.5` 将约 50.05% 样本设为最大噪声 t=999;均匀采样时这一比例只有 0.1%。这大幅增加了无法依赖未来 noisy motion、必须依赖历史和文本的样本权重。 +-- 最大噪声时历史和文本仍然存在,因此高噪声本身不要求模型输出常数。但一旦条件表征开始丢失,这部分目标对模型来说就变成高不确定性 MSE 回归,平均答案成为容易维持的方向。 +- +-最符合现有证据的解释是:在这套长程微调设置下,根网络的时序/条件表示被后续更新破坏,深层特征和注意力进入近似均匀状态;前层梯度随之消失;MSE 将尚能更新的输出端推向平均答案,而隔离的身体损失无法救回根网络。 +-随后滚动窗口以历史开头为原点,重复生成同一个非零局部位置偏差,把根网络的均值偏差转成约两秒一次的世界坐标漂移。 +- +-严格的因果边界:目前没有保存 6.5–8K 附近的连续权重,也没有在相同全量数据顺序和随机种子下,分别改变根学习率、最大噪声比例与实际轨迹监督重训。因此不能宣称“唯一原因就是学习率”或“唯一原因就是高噪声”。已实测的是内部信息同质化、梯度消失、均值回归状态及其不能被身体梯度纠正的机制。 +- +-补充数值排查:本环境 PyTorch 2.7.1+cu128、RTX 5090。FP32 的 efficient attention 在特定高噪声输入上出现明显异常的大梯度;普通数学后端没有这一现象。 +-但正式训练使用 BF16,profile 确认使用 efficient attention。对六个既有训练 session 重建的 32 批、每批八个窗口,按原噪声/dropout 策略对照 BF16 efficient 与 math,没有复现那种量级的梯度爆炸:全参数梯度范数比最大约 1.061。 +-这不足以覆盖整次训练或排除稀有数值问题,但目前不能把 FP32 的异常直接当作本次 BF16 训练退化的已证实触发原因。 +- +-参数替换对照中,仅恢复输入投影、文本投影或输出层都不能恢复时序变化。 +-仅将注意力参数恢复为 2K,而保留其他 100K 参数时,输入梯度从 4.75e-12 恢复到约 0.00407,根轨迹运动 RMS 从约 6e-7 m 提高到 0.00626 m。这是一个直接的内部干预,支持注意力参数更新与信息/梯度消失存在联系;但该组合仍不能生成完整正确轨迹,不能据此宣称已经修复。 +-这支持失效发生在 Transformer 的内部条件处理,而不是某一个坐标输出 bias 写错。本轮所有替换仅用于只读诊断,没有生成新的训练权重。 +- +-[完整原因分析数据](../reports/kimodo_root_training_cause_20261023.json)。原始逐层、梯度、参数替换、数据采样及后端记录均在 `/root/autodl-tmp/cache/kimodo-formal-validation-20261023/root-probe/`。 +diff --git a/docs/KIMODO_TRAINING_REPAIR.md b/docs/KIMODO_TRAINING_REPAIR.md +deleted file mode 100644 +index 013ccff..0000000 +--- a/docs/KIMODO_TRAINING_REPAIR.md ++++ /dev/null +@@ -1,50 +0,0 @@ +-2026-10-23:修复 Kimodo 的训练方式,针对根 Transformer 在此前长程微调中失去时序差异的问题。输入仍为按时间释放的英文 GT 文本、最近 30 秒上下文和已执行人体历史;没有新增未来文本、阶段标签、GT 推理历史或轨迹修正。 +- +-这次修改训练本身,未加入退化检测、阈值触发、自动停止或自动替换权重。 +- +-- 根阶段保留官方预训练的注意力、FFN、归一化、文本和时间嵌入,只训练输入/输出投影,共 565,253 个参数;身体阶段训练 141,529,356 个参数。冻结的根骨干仍参与前向和对输入投影的反向传播。 +-- 根和身体分组优化、分别裁剪梯度。根学习率 2e-6→2e-7,在 20K 步衰减到底;身体 1e-5→1e-6,在 100K 步衰减;均有 500 步 warmup。末端根学习率在 20K 后保持下限。 +-- 最大噪声强制采样比例从 0.5 降至 0.1,其余 timestep 仍均匀采样。 +-- 保留有效未来帧的 normalized-x0 监督,前 12 帧执行段权重为 1,剩余规划段为 0.25。 +-- 加入原始根网络对已知历史的重建损失,读取硬覆盖历史之前的网络输出。目标只来自已知过去,不使用未来作为观测。 +-- 对实际解码的根速度、关节速度、旋转变化、首帧接缝速度和根加速度监督;原来的独立速度辅助通道不能替代这些目标。 +-- 根注意力使用数学后端,身体保留原加速实现。运行配置保存进 checkpoint,推理加载时自动恢复。 +-- 完整 checkpoint 保留模型、分组 AdamW、调度器、随机状态及 native session 槽位,继续支持 `--resume`。 +- +-正式配置仍为全量 native 英文人体配对数据、batch8、100K 步、GPU0。缓存与官方预训练资产复用;新训练不继承失败的 100K 微调权重。 +- +-```bash +-cd /root/autodl-tmp/hfd-baseline-portable +-/root/autodl-tmp/envs/kimodo/bin/python -u train_kimodo.py --config configs/kimodo.yaml +-``` +- +-中断后在相同命令末尾追加 `--resume`。进程可在 screen 中前台运行。 +- +-用户要求删除的旧目录已经完整移除,包括其中的 best、last、98K/100K 快照、日志和脚本副本: +- +-- `/root/autodl-tmp/runs/hfd-kimodo-english-train` +-- `/root/autodl-tmp/models/kimodo/hfd-english-train` +- +-这两个路径后来由修复后的正式训练重新创建,当前包含新一轮100K产物。删除旧失败训练时,官方模型、原始数据、共用文本缓存、GenMo活跃训练及既有诊断视频均保留。 +- +-验收报告:[kimodo_training_repair_20261023.json](../reports/kimodo_training_repair_20261023.json)。 +- +-- 30 项单元测试通过,覆盖真实两阶段小网络的冻结/反传、原始历史监督、分阶段优化与恢复、因果表示、解码运动损失和 session 槽位恢复。全尺寸官方网络离线检查通过,根/身体均有有效梯度、参数实际更新、文本条件有效。 +-- 全量入口实际读取 155,117 个 train、4,719 个 val session,完成2步保存,再恢复到4步;AdamW和调度器均到第4步,8个native session槽位随断点恢复。短测试尚未触发2K周期验证,`best_validation`在正式JSON报告中记为null,表示尚未选best。 +-- 六段 train session 小样本训练到8K,1K时执行过完整恢复;六段val不参与更新。固定GT历史验证loss由0.577降至1K的0.222,8K为0.152。所有8,000步记录的loss有限。验证loss不是闭环动作成功率。 +-- 8K根骨干的200个参数张量与官方初始化逐位一致;四个根投影参数及204个身体参数张量均已更新。同一后退窗口、t=0时,根末层时序RMS为0.0364,输入梯度范数为0.00944;旧失败100K分别约1.08e-7、4.76e-12。本次检查未重现旧恒定输出失效,不能外推保证100K质量。 +-- 六类val和两段train完成单帧GT初始化后的全程预测历史生成,整段预测和人体GT分别GMR到G1;导出8段有声视频。帧数、30Hz、音轨及完整音视频解码全部通过,也抽帧查看了预测/参考和文本显示。 +- +-同六段val、同seed/DDIM20/CFG2的等权平均如下;1K只做人体空间诊断,未单独retarget: +- +-| 模型 | 人体姿态 RMSE(rad) | 人体根位置 RMSE(m) | 人体根速度 RMSE(m/s) | G1关节 RMSE(rad) | G1根位置 RMSE(m) | +-|---|---:|---:|---:|---:|---:| +-| 官方初始化+当前适配 | 0.627 | 0.502 | 0.269 | 0.464 | 0.433 | +-| 新训练1K | 0.411 | 0.476 | 0.150 | — | — | +-| 新训练8K | 0.418 | 0.602 | 0.328 | 0.367 | 0.519 | +- +-动作质量仍未通过完整闭环验收:8K姿态误差优于官方初始化,但根位置/速度相较1K恶化,行走与停止样本仍有轨迹漂移;此前后退任务也不能宣称已解决。该实验验证训练链路及旧根骨干失效是否重现,不把增加小样本训练步数视为质量保证。全量训练命令可执行,最终效果仍需独立闭环评测;这不是完整HFD benchmark。 +- +-视频目录:`/root/autodl-tmp/runs/kimodo-training-repair-20261023/`,每例只有`retargeted_human_comparison.mp4`,左侧预测、右侧同流程retarget的human GT。工作数据、曲线和校验日志位于`/root/autodl-tmp/cache/kimodo-training-repair-20261023/`。本次临时checkpoint验收后全部删除,未启动新的100K正式训练。 +- +-2026-10-24后续:用户已启动并完成修复版100K训练,新正式权重保留。根骨干及其时序/梯度复查通过;六类闭环、整段retarget及18段有声视频见[新100K评测](HUMAN_FORMAL_VALIDATION_20261024.md)。前面的8K数值属于修复时的小样本验收。 +diff --git a/docs/LOCAL_INDEX_TRAINING.md b/docs/LOCAL_INDEX_TRAINING.md +deleted file mode 100644 +index ba1d40e..0000000 +--- a/docs/LOCAL_INDEX_TRAINING.md ++++ /dev/null +@@ -1,71 +0,0 @@ +-# 本地索引镜像与训练性能(2026-10-19) +- +-优化已完成,两个旧训练进程已按用户授权停止,原输出目录已删除;screen 保留。旧日志和配置保存在本报告证据目录,新正式训练尚未自动启动。当前正常训练命令见文末。 +- +-## 原因与改动 +- +-64 次实际合法窗口采样的 CPU 剖析耗时 50.64 秒,其中 tar 索引查询累计 33.52 秒,SQLite execute/fetch 占 33.80 秒,实际 tar 字节读取仅约 2.88 秒。主要瓶颈是对共享盘 `index.sqlite` 的大量小查询,而不是显存不足或大文件吞吐。 +- +-新增 `hfd_bridge/index_cache.py`:把约 **433.59 MiB** 的原始只读索引镜像到数据盘 `cache/hfd/index-mirrors/<身份>/index.sqlite`。按源路径、大小、修改时间和 SQLite 头部身份区分版本,使用进程文件锁和原子替换;复制时检查源未变化,并核对源复制流和本地文件的 SHA256。明确指定 `HFD_INDEX_PATH` 时仍尊重该覆盖路径。 +- +-共享 tar 读取器和第三方 session provider 都使用同一个索引入口。原始 tar 保留共享盘;这是元数据索引的本地副本,不是把全量 session 或原始数据复制到内存/本地盘。它可以供后续使用相同读取层的 baseline 复用。 +- +-正式 OMG/TextOp 及 reference YAML 开启 `runtime.environment.HFD_LOCAL_INDEX_CACHE: '1'`;设为 `'0'` 可做旧路径对照。正式 batch512、刷新频率、语言/文本协议、学习率和100K/300K步计划保持原配置。当前优化没有更改原生组装算法、训练循环或数据采样。 +- +-## 对照证据 +- +-CPU ABBA 对照(远端、本地、本地、远端)同64个样本依次为 **48.81、3.95、3.61、10.46 秒**。两次远端时长差异说明文件/页缓存有明显影响;即便采用已经变热的最后一轮远端,本地仍更快。四轮采样ID、动作、音频、控制标签及 reference segment 内容哈希一致。 +- +-随后使用两个真实原生训练循环,batch128,每种路径经历3次数据池准备。OMG48步,TextOp96步;同模型两种路径使用相同初始化、seed、更新预算、采样ID与窗口帧。以下为训练段耗时,包含组池、更新和有界验证,排除索引/模型初始化: +- +-| 模型 | 远端索引秒 | 本地索引秒 | 加速比 | +-| --- | ---: | ---: | ---: | +-| OMG | 251.6 | 33.9 | 7.42× | +-| TEXTOP | 219.9 | 35.6 | 6.18× | +- +-OMG两条loss曲线逐项相同,TextOp最大绝对差为2.98e-8。两种路径的采样ID与窗口帧完全一致。第二种路径复用了原始CPU索引,故不能把包含冷启动的总墙钟差异全算成优化收益。此轮与旧正式训练并行运行,存在共享资源干扰;这些加速比只描述该短测,不能直接当成100K/300K完整训练加速承诺。 +- +-## 正式 batch512 复核 +- +-确认上述加速和135项回归通过后,先停止并清理旧正式训练,再独立运行两张卡各512次更新、每模型两个完整数据池刷新周期。采用当前正式batch512与刷新频率,短学习率/阶段预算和有界验证用于测速,权重不作为质量结论。 +- +-| 模型 | 训练段秒 | 两次组池秒 | 池内秒/步中位数 | 训练段实例/秒 | +-| --- | ---: | --- | ---: | ---: | +-| OMG | 283.8 | 38.0, 53.5 | 0.271 | 924 | +-| TEXTOP | 183.9 | 41.4, 40.3 | 0.143 | 1426 | +- +-训练段包含组池与最终验证/周期保存,初始化墙钟另记在JSON中。TextOp的实例按每次更新的primitive样本计,OMG为动作窗;两模型绝对实例/秒不能横向解释为架构优劣。池内中位数也不包含完整的冷组池时间,长程耗时应使用整体训练段与后续正式验证成本估算。 +- +-两模型各记录512次更新、有限训练loss,保存的Adam状态计数均为512,且恰好准备2个训练池,无多余尾池。临时测速checkpoint已清理,仅保留配置、采样、指标和日志。性能变好不代表动作质量变好;英文前景oracle文本及其已有局限保持原协议,仍需正式验证曲线与连续回放判断质量。 +- +-## 停止、清理与重启 +- +-仅停止已核验命令行和工作目录的旧OMG PID322001、TextOp PID322362。OMG收到SIGTERM后未及时退出,再使用SIGKILL;TextOp由SIGTERM结束。以下两个目录已删除: +- +-- `/root/autodl-tmp/runs/hfd-omg-english-train` +-- `/root/autodl-tmp/runs/hfd-textop-english-train` +- +-小体积日志、配置、数据manifest和采样记录保存在 [old-training](/root/autodl-tmp/runs/archive/legacy-20261020/data-pipeline-optimization-20261019/old-training),checkpoint未保留;原始数据、官方权重、其他历史run和screen未清理。测试用索引副本已去重,保留一份当前有效的本地镜像。 +- +-在两个screen中分别从头启动,不加 `--resume`: +- +-```bash +-cd /root/autodl-tmp/hfd-baseline-portable +-/root/autodl-tmp/envs/omg/bin/python -u train.py --config configs/omg.yaml +-``` +- +-```bash +-cd /root/autodl-tmp/hfd-baseline-portable +-/root/autodl-tmp/envs/textop/bin/python -u train.py --config configs/textop.yaml +-``` +- +-## 验证文件 +- +-- [加速与输入/loss对照](/root/autodl-tmp/runs/archive/legacy-20261020/data-pipeline-optimization-20261019/speed-gate.json) +-- [CPU ABBA与输入哈希](/root/autodl-tmp/runs/archive/legacy-20261020/data-pipeline-optimization-20261019/index-abba.json) +-- [CPU剖析](/root/autodl-tmp/runs/archive/legacy-20261020/data-pipeline-optimization-20261019/provider-profile.txt) +-- [OMG batch512](/root/autodl-tmp/runs/archive/legacy-20261020/data-pipeline-optimization-20261019/omg-full512.json) +-- [TextOp batch512](/root/autodl-tmp/runs/archive/legacy-20261020/data-pipeline-optimization-20261019/textop-full512.json) +-- [旧训练停止与目录删除记录](/root/autodl-tmp/runs/archive/legacy-20261020/data-pipeline-optimization-20261019/old-training-cleanup.json) +- +-分环境135项回归测试与两套正式配置dry-run通过。性能探针保存在仓库外的本报告证据目录,没有重新堆入日常 `tools/`。 +diff --git a/docs/MIGRATION.md b/docs/MIGRATION.md +deleted file mode 100644 +index d2d72fe..0000000 +--- a/docs/MIGRATION.md ++++ /dev/null +@@ -1,5 +0,0 @@ +-# 迁移入口 +- +-当前完整操作见[训练复现指南](PORTABLE_TRAINING.md):四模型环境、外部资源、原始tar、英文词对齐、路径生成和正式训练命令。 +- +-本版本发布训练代码、配置、资源版本元数据和文档。本机测试/评测在被忽略的`test/`。历史ASR、双语和过拟合配置不作为当前默认。旧迁移说明的本地副本在`test/docs/`,未随Git发布。 +diff --git a/docs/OMG_ENGLISH_VALIDATION.md b/docs/OMG_ENGLISH_VALIDATION.md +deleted file mode 100644 +index 24fe908..0000000 +--- a/docs/OMG_ENGLISH_VALIDATION.md ++++ /dev/null +@@ -1,56 +0,0 @@ +-# OMG 英文对齐文本训练后的评测 +- +-`evaluate_omg.py` 从正式训练的 `resolved_config.yaml`、`data_manifest.json` 和 +-`val_native_manifest.json` 恢复验证协议,调用相同的原生及第三方数据 provider。 +-它只支持当前英文 `aligned_reference_context` / `microphone` OMG 配置。 +- +-默认选六类原生场景各一条6–30秒英文 session,另选三条第三方英文验证 session。 +-原生按索引取首个符合条件的样本,第三方在合格索引中等距取样;选择发生在生成之前。 +-这不是按类别比例采样的全量验证估计。`test` 是 `val` 的别名,验证数据用于过训练监测。 +- +-```bash +-cd /root/autodl-tmp/hfd-baseline-portable +-/root/autodl-tmp/envs/omg/bin/python -u evaluate_omg.py \ +- --trained-run /root/autodl-tmp/runs/hfd-omg-english-train \ +- --output /root/autodl-tmp/runs/omg-english-validation-20261020 \ +- --gpu 1 +-``` +- +-默认对比官方初始化与 `best/best.ckpt`。可用 `--checkpoint PATH` 显式选择新权重, +-`--per-mode N` / `--third-party N` 控制样本数,`--stage prepare|run|summarize` +-分阶段运行。相同参数重跑会复用完整结果;不完整推理目录须检查后另选输出目录或移走, +-不会静默覆盖。不会启动训练、识别、翻译、强制对齐或下载。 +- +-## 与训练一致的输入 +- +-- 相同英文过滤、配方、原生种子、验证有序索引及第三方 selection。 +-- 原始混合 PCM;过去2秒音频。词级离线对齐结果来自训练用缓存。 +-- 截至当前释放点的最近30秒 GT 前景文本,0.1秒 guard,0.6+0.4k 秒更新。 +-- 原生 T5-Base 保留最近50 token;实际 token ID、截断方向、PCM截止和文字释放时间记录在规划日志。 +-- 动作时间到麦克风时间偏移0.2秒,音视频对预测和GT共同前置0.2秒显示。 +- +-评测冻结时逐值校验原生导出和 compact provider 的音频、动作、标签及有效起点数, +-逐个50Hz状态帧比较导出 sidecar 和训练 provider 的文字条件。第三方直接导出训练 provider 数组。 +-标签、完整原文和未来动作不传入规划器;标签仅用于计分分层,事件元数据只用于视频标注。 +- +-## 连续推理与指标 +- +-两份权重用相同噪声seed12345、DDIM10、joint_one、剩余24帧触发的原生异步缓冲。 +-公平对照固定规划延迟为4个30Hz帧(约133ms),同时单独记录实际规划耗时。 +-这不是包含前端和部署调度的实测端到端延迟。仅开头13帧动作作GT初始化,之后反馈已执行预测, +-不按窗口重置GT。初始化及共同显示前缀不计分。 +- +-统计29关节RMSE、根xyz分量RMSE、根朝向误差、FK链接位置误差,以及listen/motion/idle分层误差。 +-标签索引先减去视频共同前缀,再转换到50Hz原始动作时钟,避免使用未来标签。 +-没有DTW或事后根位置对齐;这些误差衡量参考跟随,不等于语义成功率。 +- +-OMG直接输出G1的29关节qpos,不需要人体retarget。渲染为运动学预测/GT对照,带原始混合音轨, +-不是物理tracker。所有视频完整解码,音轨核验;两份权重的GT和音频必须相同。 +- +-输出包含 `README.md` 视频索引、`metrics.csv`、`summary.json`、`verification.json`、 +-权重SHA256、冻结样本与对齐sidecar、逐次推理/渲染日志。 +-官方预训练重合、第三方与原生动作源重合未审计;小样本单seed不能替代完整评测。 +- +-视频封装保留完整H.264视频流,不再用 `-shortest`:实测相同长度PCM与含B帧的视频流封装时, +-该参数可能裁掉末尾视频包(508帧样本曾变为505帧)。新增实际编码/全帧解码回归测试; +-渲染中间视频只在最终MP4帧数、帧率和音轨通过检查后删除。 +diff --git a/docs/OMG_TRAINING_REVIEW.md b/docs/OMG_TRAINING_REVIEW.md +deleted file mode 100644 +index 9002c7c..0000000 +--- a/docs/OMG_TRAINING_REVIEW.md ++++ /dev/null +@@ -1,105 +0,0 @@ +-# OMG 正式训练复查与吞吐优化(2026-10-17) +- +-清理说明:本文记录的一次性实验脚本已归档,复现前请按[脚本恢复说明](WORKSPACE_CLEANUP.md)取回;历史指标与日志保留。 +- +-本轮输出:`/root/autodl-tmp/runs/archive/legacy-20261020/hfd-omg-train-20261010-v2`。已正常完成 **100,000 次优化器更新**,不是被 OOM 或异常中断。screen 的 `Trainer.fit stopped: max_steps=100000 reached` 和 checkpoint 的 `global_step=100000` 一致。训练循环耗时 108 小时 33 分 7 秒,整个命令约 108 小时 45 分,约 4.53 天。完成时间为 2026-10-15 01:30(Asia/Taipei)。 +- +-## 本轮结果 +- +-| 指标 | 结果 | +-| --- | --- | +-| 模型 | OMG 原生 50m 配方;denoiser 参数约 54.14M,另有冻结的 T5 等模块 | +-| 训练 | RTX 5090 × 2;每卡 batch 512,总 batch 1024;bf16 mixed | +-| 训练 loss,前 1,000 步平均 | 0.04943 | +-| 训练 loss,最后 10,000 步平均 | 0.02315 | +-| 首次验证,2,000 步 | 0.0365847 | +-| 最低日志验证 loss,90,000 步 | 0.0220401 | +-| 最后验证,100,000 步 | 0.0234458 | +-| 最终权重检查 | 可加载,394 个 state tensors 均无 NaN/Inf,含 optimizer/scheduler 状态 | +- +-![训练与验证曲线](../reports/omg_training_20261017.png) +- +-训练 loss 是 rank 0 的原生扩散损失;验证 loss 由两卡同步归约。它们不是厘米级动作误差,也不是任务成功率。`val/motion_loss=0` 来自原生 `diffusion_only` 配方,不代表动作没有监督;动作特征是扩散损失的预测目标。这里只确认训练收敛趋势和权重健康,尚未完成论文独立测试集或连续 rollout 评测。 +- +-最终文件是 `last.ckpt`。保留的周期权重是 `checkpoints/step-00096000.ckpt`、`step-00098000.ckpt` 和 `step-00100000.ckpt`;`checkpoints/last.ckpt` 是周期保存的链接。原策略按步数保留最近三份,**90,000 步的最低 loss 权重已被删除,不能恢复为 best**。现存三份记录的验证 loss 分别是 0.0222620、0.0223724、0.0234458;要据此选模型,应在同一固定验证样本上重新比较。 +- +-旧验证并不固定:每次验证每卡 16 个 batch,池每累计 256 个验证 batch 换一次。因此 2k、34k、66k、98k 步分别开始使用不同的验证 session 池;池内窗口和扩散噪声也不断变化。最后两个 loss 不能直接与 96k 比较后认定过拟合。 +- +-本轮 manifest 记录 train=322,243、val=9,272 个原生候选 session,`command_interruption_session_variants=2`。两卡记录合计访问 207,173 个不同训练 session,约占候选的 64.3%;有放回随机窗口训练不等于完整遍历一个 epoch。**本轮不包含后来加入的 third-party 包**:当前 `configs/sessions.yaml` 已有 `include_third_party: true`,不能用今天的配置文件反向描述旧实验。复查性能时使用关闭 third-party 的临时配方,重放旧 run 的 session 索引;当前正式数据配置未被改回旧值。 +- +-## 是否使用 OMG 原生脚本 +- +-入口是 portable 的 `train.py::omg`,没有直接调用上游 `omg.cli.generation.train.main`。入口组合上游 `exp=50m`,加载官方模型定义及初始化权重,再构造 `pytorch_lightning.Trainer`,训练、验证、优化器和 scheduler 仍调用上游 `MotionGenerator` 的实现。替换的是 HFD 数据适配器和训练管理 callback。准确表述是“使用 OMG 原生模型及训练步骤,通过 HFD 适配入口训练”。 +- +-本轮保持原生 100k 更新、2k warmup、6e-5 学习率及 cosine schedule。数据输入是过去两秒已释放音频 + 因果 ASR 当前指令,完整 session 动作监督;未改变动作映射、音频截止时间或模型目标。 +- +-## 慢在哪里 +- +-1. 每卡 batch=512 的 session 池每 256 个 batch 刷新。本轮 391 次刷新,取每次两个 rank 中较慢的准备时间累加约 **59.22 小时**,约占训练循环 54.6%。这包含共享 tar 读取、原生 session 组装、ASR、运动预处理和首个 batch,不能全算成磁盘 I/O。rank 0 的刷新中位数 493.6 秒。 +-2. 四次验证池刷新另需约 1.10 小时。完整训练平均每步 3.91 秒,旧日志没有每步数据/计算细分计时。 +-3. ASR SQLite 有 209,704 条记录、约 1.18 GiB 压缩有效内容。旧实现每插入一条都 `SUM(length(data))` 扫全表;复查时同一只读扫描缓存热后仍约 0.20–0.29 秒。这是可去除的元数据开销,不是模型计算。 +-4. 旧普通 batch 逐个调用 512 次 motion codec,许多小 CPU tensor 运算串行执行。音频特征此前已做批处理,动作特征仍有批处理空间。 +-5. tar `Member.readinto` 对小读请求仍取一个 256 KiB 块,且本地 tar 路径没有块复用。`BufferedReader`/NPZ 的小读和回读可反复读取同一块。 +- +-`num_workers=63/127` 是 Lightning 通用提示。当前数据对象持有 SQLite、状态采样器以及同卡 Whisper,且明确禁止 DataLoader worker;仅修改 `num_workers` 会被配置校验拒绝。若做并行预取,需要把 CPU 读盘/组装与 GPU ASR 分开,并管理 rank/worker 的采样状态,不能直接照提示填值。 +- +-## 已落地的改动 +- +-- 动作预处理:用同一个 OMG 原生 codec 一次计算整批 10 帧历史与 60 帧目标。`data.batched_motion=true`,可设为 false 回退逐样本实现。测试比较全部 batch 字段,容差 `rtol=atol=1e-6`,保持动作定义与时间对齐。 +-- ASR SQLite:增加由数据库 trigger 原子维护的容量计数,首次迁移扫描一次,此后不再每次插入扫全表。兼容已有 ASR 内容与 key,两卡连接共同维护;访问时间最多每分钟写一次,减少热缓存写锁。原有 2 GiB 有效内容上限仍生效;该上限不等于 SQLite 文件/WAL 的物理上限。 +-- tar I/O:每个打开的成员保留最多四个 256 KiB 块,减少小读放大;关闭时释放。原始 tar 留在共享盘,不解压,不复制完整数据集。 +-- 训练管理:固定验证窗口与扩散随机数,验证结束恢复训练 RNG;单独保存 `best/best.ckpt`,不受最近三份周期权重清理影响;记录每步等待/计算时间及最终训练摘要。时间字段是主机墙钟,`data_wait_seconds` 可能包含周期 checkpoint 等 callback 开销,不等于纯磁盘 I/O。 +-- 可选早停:原生 Lightning `EarlyStopping`,监测 `val/loss`,从 30k 步开始计数,`patience=10`,`min_delta=0.0002`。每 2k 步验证,因此对应连续 20k 步没有足够改进。patience 计算的是验证次数,见 [Lightning 官方说明](https://lightning.ai/docs/pytorch/latest/optional-api/api/lightning.pytorch.callbacks.EarlyStopping)。 +- +-默认 `training.fixed_validation=true`、`training.save_best=true`,但 `training.early_stopping.enabled=false`,保留原生固定 100k 预算。要选择早停训练,在原训练命令上加 `--set training.early_stopping.enabled=true paths.output_dir=/root/autodl-tmp/runs/新的目录`。这会成为不同的训练预算,论文中须记录。 +- +-固定验证复用同一组每卡 16 个 batch(双卡 16,384 次窗口抽样,窗口允许重复),仍只是监测子集。不能按旧随机验证日志选择最有利的早停阈值:事后机械套 patience=10 会在 88k 停,随后 90k 还有改善,说明不能保证早停明显省时且完全不损失质量。优化每步吞吐比激进缩短训练更有依据。 +- +-## 验证和性能测量 +- +-单元测试覆盖原生动作 codec 对照、固定验证样本、验证 RNG 恢复、真实 Lightning 早停/最佳权重保存、旧 SQLite 迁移与双连接容量计数、tar 小读/seek/EOF 等。性能测量仅重放旧训练池、运行限定步数,不保存更新后的训练权重,也不改写已完成 run。 +- +-本机测量记录集中在 `/root/autodl-tmp/cache/hfd/review-20261017/`,不提交原始数据或大文件。47 项相关测试与 12 个 subtests 通过,配置 dry-run 和 `git diff --check` 通过。 +- +-同一旧 run 的 512-session 池,4 CPU 线程,比较结果: +- +-| 测量口径 | 优化前 | 优化后 | +-| --- | ---: | ---: | +-| session 已缓存、仍有新音频窗口的 batch 数据准备 | 1.68 秒 | 0.79–0.80 秒 | +-| 同一进程、同一批窗口且音频已缓存,逐样本/批量动作 codec 对照 | 0.91–0.92 秒 | 0.048–0.078 秒 | +-| 209,704 条 ASR 记录的容量查询 | 约 0.16 秒/次全表扫描 | 约 2–44 微秒/次读计数 | +- +-真实 batch 对照比较了所有 tensor、caption 和索引,`rtol=atol=1e-6` 通过。热 session、新音频窗口的数据准备约快 **2.1 倍**;音频全命中的对照展示了消除逐样本动作计算后的收益,不能把该倍数当成整轮训练加速。 +- +-两次首次装入同一池分别耗时约 417 和 209 秒,数组读取/原生组装分别占约 407 和 200 秒。它们共享文件系统缓存且执行有重叠,不是受控冷盘对照,不能归因于代码修改。新 tar 块缓存有读取一致性与读放大量测试,但尚无隔离的整轮冷读测速。 +- +-双卡端到端短测已完成:加载本轮最终 checkpoint,在真实 HFD 数据上每卡 batch=512,各完成 6 次前向、反向与优化器更新;所有记录的 loss 有限,denoiser 确认为训练模式。第 2–6 步完整迭代为 **1.25–1.32 秒/步,均值约 1.28 秒**,模型计算约 0.42–0.50 秒/步。每卡 PyTorch 峰值 allocated 显存约 16.42 GiB;本次 ASR 已命中缓存,没有同时运行缺失文本的 Whisper 识别,不能据此提高正式 batch。 +- +-首次准备在 rank 0/1 分别是 68.9/371.2 秒;rank 0 第一轮又等待约 304 秒,而后两卡进入上述稳定吞吐。这个对照直接展示了共享盘读取/原生组装的不均衡及 DDP 等待。短测没有跨过 256-step 池刷新边界,不能用 1.28 秒直接乘 100k 预测完整训练时间。检查已结束,两张 GPU 已释放;没有启动新的正式训练,也没有覆盖本轮 checkpoint。 +- +-下一阶段若刷新仍占主导,优先将因果 ASR 从主训练循环拆成可复用的准备阶段,再考虑有界异步预取。单纯把 `refresh_batches` 从 256 加到 1024 虽然摊薄刷新,却减少相同步数下接触的新 session、增加相邻 batch 相关性,本次没有采用。当前数据盘不足以额外复制约 49.8 GiB 的全部原始 tar,因此也没有把“全部搬到本地”当作立即可行的方案。 +- +-## 异步预取与并行 I/O(2026-10-17 后续) +- +-增加 `data.prefetch_batches`(0 或 1)。开启后,OMG 在 GPU 训练当前 batch 时,用一个后台线程准备下一个已缓存的 CPU batch,包括动作编码和音频特征。最多额外保留一个 CPU batch;validation 仍在主线程执行。 +- +-后台线程不构造新 session、不访问 SQLite、不运行 Whisper,也不执行 GPU 工作。`WindowPool` 保留当前槽位的合法时刻范围,避免热 batch 为查询长度重新组装数组;缓存未就绪或到达刷新边界时,自动回到主线程。固定 seed 下,测试跨多个刷新边界逐一比较动作、音频、caption、采样时刻,并验证训练 RNG 不变。后台异常传回主线程,关闭 iterator 会结束任务。实现使用标准库 [ThreadPoolExecutor](https://docs.python.org/3.10/library/concurrent.futures.html)。 +- +-另外增加本地 tar 成员的有界并行预读:`runtime.environment.HFD_IO_READAHEAD_BLOCKS=4`,`runtime.environment.HFD_IO_WORKERS=4`。线程仅执行本地文件读取,不共享 SQLite 连接,不进行网络下载。大型/连续读取才触发预读,seek 会取消旧预读,每个成员的未来块数有界;原有四块缓存仍保留。它作用于原生 `ArchivePath/Member` 读取路径,third-party 已组装 WAV/NPZ 的直接大块读取路径没有替换。 +- +-8 个实际音频成员的同步→异步→异步→同步对照,全部 SHA256 相同,时间为 0.704、0.094、0.071、0.052 秒。第一次同步读取受冷缓存影响;最后一次同步比异步更快,因此不能把第一次与第二次相除宣称并行 I/O 提升。目前并行 tar 预读**默认关闭**,保留为不同文件系统上的可测选项。 +- +-CPU batch 异步预取则使用双卡 64 步的同步→异步→异步→同步分段对照;每段排除开头两步,以减少流水线切换影响。该测试仍重放旧训练池,不包含 third-party,也不执行新的正式训练。它测的是池内准备与 GPU 计算的重叠,不包含下一池的后台组装/ASR,不能视为解决全部刷新等待。 +- +-最终采用 **OMG 一个热 CPU batch 预取,tar 并行预读默认关闭**。两卡各 64 步均完成,所有 loss 有限,denoiser 确认为 train mode。每段排除两步后,按每一步较慢的 rank 汇总: +- +-| 模式 | 平均每步 | 分段结果(rank 0) | +-| --- | ---: | --- | +-| 同步 | 1.072 秒 | 第一段 1.244 秒;末段 0.897 秒 | +-| CPU batch 异步预取 | 0.617 秒 | 中间两段 0.656 / 0.572 秒 | +- +-该对照约提高池内吞吐 1.74 倍,单步时间减少约 42%。音频缓存会逐步变热,ABBA 仅减轻该影响,不能彻底消除;异步中段也比更晚、更热的同步末段快。CPU/GPU 并发时模型阶段本身略慢,但整体迭代更快。每卡 PyTorch allocated 峰值约 16.42 GiB。首次装载两卡分别约 227/223 秒,此部分未被热 batch 预取隐藏。 +- +-当前默认 `configs/omg.yaml` 中 `data.prefetch_batches: 1`,同步回退为 0;并行 I/O 的 `HFD_IO_READAHEAD_BLOCKS` 默认仍为 0。52 项相关测试及 12 个 subtests 通过,另有真实双卡 64 步检查。这里只运行了有限步数的性能探针,不保存更新权重;GPU 已释放,旧正式训练权重未改写。原始测量在 `ddp-async-abba/{rank0,rank1,summary}.json` 与 `tar-async-abba.json`(位于本机 review 缓存目录)。 +- +-复查命令可使用 `tools/benchmark_omg_training.py --steps 64 --compare-prefetch`(双卡需 torchrun,配合实际配置、`--replay-dir` 和 `--report-dir`);单独 tar 对照使用 `tools/benchmark_omg_io.py --members 8 --report /数据盘/新的测量.json`。这些命令是测试工具,不是正式训练或完整训练时长评估。 +diff --git a/docs/OMG_VALIDATION_DIAGNOSIS.md b/docs/OMG_VALIDATION_DIAGNOSIS.md +deleted file mode 100644 +index 9898751..0000000 +--- a/docs/OMG_VALIDATION_DIAGNOSIS.md ++++ /dev/null +@@ -1,77 +0,0 @@ +-# OMG 验证视频:提前动作与听完不执行的诊断 +- +-清理说明:本文记录的一次性实验脚本已归档,复现前请按[脚本恢复说明](WORKSPACE_CLEANUP.md)取回;历史指标与日志保留。 +- +-本次仅检查冻结的 100k checkpoint,没有训练、改权重、替换模型输入或覆盖原视频。诊断产物位于 `/root/autodl-tmp/runs/archive/legacy-20261020/omg-100k-validation/diagnostics/`。 +- +-## 1. 已确认的中文文本语义丢失 +- +-在实际使用的 `/root/autodl-tmp/models/OMG/t5-base-local` tokenizer 上,以下实际规划文本的 token IDs 完全相同: +- +-| 实际文本 | token IDs | token | +-| --- | --- | --- | +-| 慢跑向前 | `[3, 2, 1]` | `▁, , ` | +-| 小禮母 | `[3, 2, 1]` | 同上 | +-| 慢跑向前跳越中身手架 | `[3, 2, 1]` | 同上 | +-| 慢跑向前跳越中身手构想后跳 | `[3, 2, 1]` | 同上 | +- +-原生 `FrozenT5TextEncoder` 只根据这些 token IDs 和 mask 计算冻结 T5 特征,因此这些输入的文本条件相同,无法区分任务。打断样本第二、第三条中文指令到来后,ASR 字符串虽然变化,模型文本分支却收不到对应的语义变化。35D 声学分支包含频带能量等声学特征,不能假定它会自动补回丢失的指令语义。 +- +-这不只是 Whisper 的中文识别不准;即使正确识别“慢跑向前”,当前文本入口仍会丢失含义。之前只验证长度和截断,没有验证未知词覆盖,是适配验收的遗漏。保留完整历史 ASR 也不能修复该 tokenizer 问题;继续增加同配方训练步数不会恢复已经被映射成同一个 token 的指令信息。 +- +-`clock-and-token-audit.json` 保存六个视频每次规划实际文本及 token。用户指出的 `interruption-stable_latest_sentence.mp4` 对应 `val-000064-seed42`: +- +-- 视频任务语音起点为 1.6、9.2、14.8 秒。 +-- 进入模型的后续文本含“慢跑向前跳越中身手架”等 ASR 错误,且中文段均发生上述 token 合并。 +-- 预测根位置相对开场的最大水平位移仅 0.0242 m,GT 为 4.0543 m;模型没有跟随参考里的明显移动。 +- +-这些证据确认输入适配有问题,不能直接把该样本的失败归因为训练步数不足。 +- +-## 2. 提前动作包含生成模型自身的提前起步 +- +-六条视频的首次明显姿态变化几乎都在 1.6–1.7 秒,尽管语音起点不同。这里的“明显变化”定义为相对开场姿态、29 关节 RMS 偏移首次超过 0.05 rad,不是一个训练出的倾听分类器。 +- +-尤其是英文紧急停止样本 `val-000111-seed42`: +- +-- 视频指令语音到 2.4 秒才开始;GT 姿态超过阈值在约 3.47 秒。 +-- 预测在约 1.63 秒已经超过阈值,呈视频中的倾听/准备样式。 +-- 产生这些姿态的首轮规划只读到 PCM 1.0 秒,实际任务语音到 PCM 2.6 秒才开始;首轮 ASR 为背景上的幻觉 `you`。 +- +-对同一 checkpoint、同一开场历史、同一 seed,重新生成首个 60 帧计划,实际条件输出与保存视频的已执行部分匹配。分别去文本、去音频、两者都去掉,仍会产生相近幅度的动作: +- +-| 首轮条件 | 最大关节姿态偏移 RMS | +-| --- | ---: | +-| 实际音频+文本 | 0.5723 rad | +-| 去文本 | 0.5752 rad | +-| 去音频 | 0.5722 rad | +-| 两者都去掉 | 0.5756 rad | +- +-这是首轮计划诊断,不是完整 session 消融,也不表示屏蔽条件时原生扩散模型本应静止。它表明这些动作无需当前语音指令触发即可生成。当前模型只预测姿态,没有显式“检测到指令后才能进入倾听/执行”的门控;HFD 所需的等待和触发行为尚未被可靠学到。不能将其解释为模型提前听懂了后面的指令。 +- +-数值见 `silent-first-plan-audit.json`,轨迹见 `silent-first-plans.npz` 和 `early-motion.png`。 +- +-## 3. 现有动作轴与音频释放轴还会影响视频观感 +- +-现有评测沿用训练的 HFD 时钟,动作组 `k` 使用截止 `0.6+0.4*k` 秒的 PCM;视频按动作时间显示,播放的音轨却取原 PCM 的 `t+0.2` 秒位置。这不是统一麦克风墙钟的真实在线系统。 +- +-按实际计划交接记录换算后,打断视频有新计划开始执行时、输入音频内容比视频当时播放位置领先约 0.167 秒的区间;停止视频最大约 0.233 秒。它会影响“提前/延迟”的直观判断,当前视频不应直接用于真实响应延迟评估。 +- +-这个小幅度时钟领先不能解释停止样本中约 0.77 秒的提前起步:首轮规划根本未包含任务语音,且无条件首轮仍产生动作。两件事应分开处理。修正评测墙钟后需要重新核对动作标签和训练时的释放映射,不能只把音轨随意平移来美化视频。 +- +-## 4. 仍有生成/闭环能力问题,但尚未完全归因 +- +-英文跳跃样本文本能正常编码;预测根高度范围仅约 0.0633 m,GT 为约 0.3907 m。因此中文入口并非全部问题。现有模型、训练窗口和闭环规划组合仍有动作幅度与阶段推进不足。 +- +-异步规划执行旧缓冲、替换未执行的计划后半段,模型是否反复把动作推迟到未来,以及真实历史与预测历史的差异,尚需固定条件、固定交接时刻和历史模式对照才能分解。不能仅凭当前视频断言优化器错误、模型架构无能力或某个单一训练参数是根因。 +- +-后续优先级:先修复中文语义入口并审计五个模型的实际 tokenizer 覆盖;再统一评测墙钟;最后区分模型等待/触发学习和闭环规划问题,再决定重训配方。完整历史 ASR 仍是用户确认的共同策略,但不能把“完整文本存在”当作“模型获得了文本语义”。 +- +-复现主诊断: +- +-```bash +-# diagnostics 目录必须尚不存在;保留原诊断时改用新的评测根目录副本。 +-/root/autodl-tmp/envs/omg/bin/python tools/diagnose_omg_validation.py --root /root/autodl-tmp/runs/archive/legacy-20261020/omg-100k-validation +-``` +- +- +-后续输入与时钟修复已实现,旧诊断结果保持不变,见 [修复记录](SEMANTICS_CLOCK_FIX.md)。 +diff --git a/docs/OMG_VALIDATION_ROLLOUT.md b/docs/OMG_VALIDATION_ROLLOUT.md +deleted file mode 100644 +index a7cff48..0000000 +--- a/docs/OMG_VALIDATION_ROLLOUT.md ++++ /dev/null +@@ -1,67 +0,0 @@ +-# 已完成训练的 OMG:六类验证 session 连续推理 +- +-清理说明:本文记录的一次性实验脚本已归档,复现前请按[脚本恢复说明](WORKSPACE_CLEANUP.md)取回;历史指标与日志保留。 +- +-结果目录:`/root/autodl-tmp/runs/archive/legacy-20261020/omg-100k-validation/`。 +- +-后续诊断确认了 **T5 中文输入映射为相同 `` token、无任务语音时提前起步,以及视频/释放时钟差异**。详见 [提前动作与不执行诊断](OMG_VALIDATION_DIAGNOSIS.md)。本轮指标包含这些输入适配问题,不能单独用于判断动作模型能力。 +- +-## 数据和权重 +- +-使用 `/root/autodl-tmp/runs/archive/legacy-20261020/hfd-omg-train-20261010-v2/last.ckpt`,确认 `global_step=100000`。没有重新训练或改写权重,评测目录仅用软链接引用原 checkpoint。`evaluation.json` 保存 SHA256;每次推理使用原生模型的严格权重加载。 +- +-仓库有 HFD 原生 train/val 构造入口,数据来自本机 `/root/autodl-fs/pangxinlei/hfd_data` 的 tar 分片。旧训练对应 322243 条 train、9272 条 val session。任务动作源按哈希隔离;验证集曾用于训练监测。**源码中的 `test` 是 val 的别名,没有第三份独立、未使用过的 test 划分。** 公共 idle/listen 素材可能共用;官方 OMG 预训练数据与这些动作的重合没有审计,不能声称模型整个训练过程从未见过它们。 +- +-本次重建的 val 索引指纹 `dc6cb1533a00d2fd983d220628a999cad60023339fbf5abba106cd704d2421dd` 与训练保存的 manifest 一致,任务源哈希划分验证通过。每类选择第一条估计时长 6–30 秒的 session,按类别交替优先选择英文和中文;选择完成后才运行模型,没有根据输出挑样本。具体索引和动作身份见 `selection.json`。 +- +-| 场景 | session index | 语言 | 长度 | +-| --- | ---: | --- | ---: | +-| 轮流指令 | 3 | 英文 | 10.8 s | +-| 任务打断 | 64 | 中文 | 21.6 s | +-| 紧急停止 | 111 | 英文 | 16.8 s | +-| 附和 | 60 | 中文 | 15.6 s | +-| 犹豫 | 28 | 英文 | 16.8 s | +-| 停顿 | 41 | 中文 | 27.2 s | +- +-## 推理设置 +- +-旧权重采用 `stable_latest_sentence` 训练,因此主评测保持原策略;另以同一 checkpoint 测试当前默认 `causal_asr_context`。后者属于输入分布变化对照,不代表已用完整历史重训。 +- +-两组共用相同的因果 ASR sidecar、音频、GT 和 seed=12345。ASR 使用本地 Whisper-tiny、auto 语言、最多 30 秒历史、batch 32;每条输入切片的哈希与释放边界均核验。OMG 输入过去两秒已释放音频及文本,保持原生 10 帧动作历史、60 帧未来、30Hz、DDIM 10 步、joint_one CFG、剩余 24 帧触发的 async 缓冲。异步规划耗时实测,旧新策略的耗时和交接时刻可能不同;ASR 识别计算耗时不计入。 +- +-开场 13 帧 GT bootstrap 不计分,之后始终反馈已执行预测,没有逐窗口 GT 重置。listen/motion/idle 标签仅用于评测分层,不进入推理。输出直接是 G1 qpos,无需人体 retarget;视频仅为运动学回放,没有物理 tracker。 +- +-## 结果 +- +-关节 RMSE 为原生 29 关节,根坐标 RMSE 为 xyz 分量 RMSE;它不是逐帧根位置欧氏距离的均值。所有误差按共同时间轴直接计算,没有 DTW 或事后根位置对齐。 +- +-| 场景 | 原稳定句关节 RMSE | 完整历史关节 RMSE | 原稳定句根坐标 RMSE | +-| --- | ---: | ---: | ---: | +-| 轮流指令 | 0.2501 rad | 0.2264 rad | 0.0915 m | +-| 任务打断 | 0.3253 rad | 0.3244 rad | 1.6924 m | +-| 紧急停止 | 0.4356 rad | 0.4614 rad | 0.6104 m | +-| 附和 | 0.2860 rad | 0.2625 rad | 0.0133 m | +-| 犹豫 | 0.4297 rad | 0.4278 rad | 0.0437 m | +-| 停顿 | 0.2919 rad | 0.2695 rad | 0.5386 m | +-| 六条等权均值 | **0.3364 rad** | **0.3287 rad** | **0.4983 m** | +- +-按计分帧合并的关节 RMSE 为 0.3436 / 0.3393 rad;两种汇总的含义不同,完整记录在 `summary.json`。另保存根朝向误差、FK 链接位置误差、listen/motion/idle 分层误差及实际规划文本非空比例。 +- +-这些验证动作的参考跟随偏差明显高于之前单条 session 的过拟合结果。完整历史 ASR 的平均改善很小,紧急停止样本的关节误差反而上升。部分中文转写有明显识别偏差,例如打断样本的末次文本为“曼跑向前跳躍中身手架我想后跳”;保留实际转写,不用 GT 修正。这是 ASR+OMG 链路的结果,不能把所有偏差归因于动作模型。 +- +-这是六条验证样本、一个采样 seed 的初步检查,不是全量测试或任务成功率评测。不同动作也可能满足同一句话,因此 GT RMSE 不能单独判断语义是否正确。未提供停止响应延迟、任务成功率或物理稳定性结论。 +- +-## 查看和核验 +- +-打开结果目录的 `README.md` 可逐类访问两种策略的 12 条视频,均为左侧预测、右侧 GT,带原始混合音轨;`comparison.png`、`metrics.csv` 和 `summary.json` 分别为图表、表格和完整数值。`video_contact_sheet.jpg` 是中段视频帧概览。部分中文视频字幕受现有字体支持限制显示为问号,实际 Unicode 输入保存在各条 `report.json` 和 ASR sidecar 中。 +- +-12 条视频共 6528 帧完整解码,音轨检查通过;两种策略的 GT 数组逐值一致,ASR PCM 切片哈希通过。13 项相关缓冲、关节映射、文本因果性测试通过,记录在 `verification.json`。此次只读取本地 tar,没有下载数据。 +- +-## 复现 +- +-```bash +-# output 必须为新目录;从原训练记录恢复验证配方 +-/root/autodl-tmp/envs/omg/bin/python tools/prepare_omg_validation.py --trained-run /root/autodl-tmp/runs/archive/legacy-20261020/hfd-omg-train-20261010-v2 --output /root/autodl-tmp/runs/archive/legacy-20261020/omg-100k-validation +-/root/autodl-tmp/envs/omg/bin/python tools/run_omg_validation.py --root /root/autodl-tmp/runs/archive/legacy-20261020/omg-100k-validation +-/root/autodl-tmp/envs/omg/bin/python tools/summarize_omg_validation.py --root /root/autodl-tmp/runs/archive/legacy-20261020/omg-100k-validation +-``` +diff --git a/docs/PORTABLE_TRAINING.md b/docs/PORTABLE_TRAINING.md +deleted file mode 100644 +index c73bd69..0000000 +--- a/docs/PORTABLE_TRAINING.md ++++ /dev/null +@@ -1,167 +0,0 @@ +-# 在另一台机器复现训练 +- +-当前四个正式入口及配方见README。HFD/OMG/TextOp原生源码在`upstream/`;GenMo和Kimodo使用`UPSTREAM_LOCK.json`指定的外部源码。模型、原始数据和离线词时间缓存单独准备。本版本不发布测试/评测脚本、正式权重或日志。 +- +-以下以`/data/hfd_work`为工作盘、`/data/hfd_assets`为外部资产盘。已验证环境是Linux、Python3.10、PyTorch2.7.1+cu128、RTX5090 32GB。GenMo训练峰值约25GB;Kimodo还需要同卡常驻的8B冻结文本worker。已有资源可由路径参数复用,不必搬迁。 +- +-## 代码及环境 +- +-AutoDL下载/安装命令先加`/root/.local/bin/turbo`,不要全局source代理。下面下载命令展示该前缀;其他机器使用自身已授权的网络方式。HF失败可使用已配置的`hfmirror`,不能自动转用个人代理。 +- +-```bash +-/root/.local/bin/turbo git clone https://github.com/starstone3/Dual-hfd-portable.git /workspace/hfd-baseline-portable +-cd /workspace/hfd-baseline-portable +-mkdir -p /data/hfd_work/envs /data/hfd_work/hfd-human-models +-``` +- +-创建独立Python3.10环境。完整冻结依赖在`requirements/`;GenMo、Kimodo和forced-align的Torch版本另列于`runtime-versions.json`: +- +-```bash +-python3.10 -m venv /data/hfd_work/envs/genmo +-/root/.local/bin/turbo /data/hfd_work/envs/genmo/bin/python -m pip install \ +- --index-url https://download.pytorch.org/whl/cu128 torch==2.7.1 torchvision==0.22.1 +-/root/.local/bin/turbo /data/hfd_work/envs/genmo/bin/python -m pip install -r requirements/genmo.txt +- +-python3.10 -m venv /data/hfd_work/envs/kimodo +-/root/.local/bin/turbo /data/hfd_work/envs/kimodo/bin/python -m pip install \ +- --index-url https://download.pytorch.org/whl/cu128 torch==2.7.1 torchvision==0.22.1 +-/root/.local/bin/turbo /data/hfd_work/envs/kimodo/bin/python -m pip install -r requirements/kimodo.txt +-/root/.local/bin/turbo /data/hfd_work/envs/kimodo/bin/python -m pip install --no-deps \ +- --target /data/hfd_work/envs/kimodo-text-compat -r requirements/kimodo-text.txt +- +-python3.10 -m venv /data/hfd_work/envs/omg +-/root/.local/bin/turbo /data/hfd_work/envs/omg/bin/python -m pip install \ +- --extra-index-url https://download.pytorch.org/whl/cu128 -r requirements/omg.txt +-/root/.local/bin/turbo /data/hfd_work/envs/omg/bin/python -m pip install --no-deps -e upstream/omg +- +-python3.10 -m venv /data/hfd_work/envs/textop +-/root/.local/bin/turbo /data/hfd_work/envs/textop/bin/python -m pip install \ +- --extra-index-url https://download.pytorch.org/whl/cu128 -r requirements/textop.txt +-/root/.local/bin/turbo /data/hfd_work/envs/textop/bin/python -m pip install --no-deps \ +- -e upstream/textop/deps/CLIP -e upstream/textop/deps/isaac_utils -e upstream/textop/TextOpRobotMDAR +-``` +- +-固定外部代码revision,不换成最新分支: +- +-```bash +-/root/.local/bin/turbo git clone https://github.com/NVlabs/GENMO.git /data/hfd_work/hfd-human-models/genmo +-git -C /data/hfd_work/hfd-human-models/genmo checkout 16bebf402d8893184249ee206d957b8248cd8310 +-/root/.local/bin/turbo git clone https://github.com/nv-tlabs/kimodo.git /data/hfd_work/hfd-human-models/kimodo +-git -C /data/hfd_work/hfd-human-models/kimodo checkout e1c219c99b4b94ba48be1fb6a1cbc64ecee4c954 +-/root/.local/bin/turbo /data/hfd_work/envs/genmo/bin/python -m pip install --no-deps -e /data/hfd_work/hfd-human-models/genmo +-env SKIP_MOTION_CORRECTION_IN_SETUP=1 /root/.local/bin/turbo /data/hfd_work/envs/kimodo/bin/python -m pip install \ +- --no-deps --no-build-isolation -e /data/hfd_work/hfd-human-models/kimodo +-``` +- +-## 外部资源 +- +-完整身份清单在`TRAINING_ASSETS.json`。HFD索引与人体分片固定为`93ba4a7a9c66b3e1b7041e5f4ea038ef2418b948`,不能再用不含SMPL-X的旧版本。先准备已授权的Llama base及SMPL-X主体;其余缺失资产可按清单下载,不覆盖已有文件: +- +-```bash +-/root/.local/bin/turbo /data/hfd_work/envs/omg/bin/python tools/training_assets.py \ +- --models-root /data/hfd_assets/models --raw-root /data/hfd_assets/hfd_data \ +- --models omg textop genmo kimodo --omg-sizes 50m 100m 300m 500m --download +-/root/.local/bin/turbo /data/hfd_work/envs/omg/bin/python tools/download_hfd_tars.py \ +- --root /data/hfd_assets/hfd_data +-``` +- +-已有资产去掉`--download`进行SHA256核验;`--quick`只检查大小,`--plan`只列出所需文件及大小。HF源可使用授权的`hfmirror`重试;CLIP的Azure下载不能借此自动转用个人代理。只训练某一模型可缩小`--models`/`--omg-sizes`。不要在缺少原始tar时直接启动训练,训练缺分片会报错,不再隐式联网补齐。若复用既有完整资产,直接核验即可,不必重新下载。 +- +-准备下面的相对布局;权重和数据不在Git中: +- +-```text +-/data/hfd_assets/ +- hfd_data/ +- index.sqlite, MANIFEST.json, PATH_MAP.json +- metadata/seed_g1_smplx_pairs.parquet +- 原始tar分片(保留,不批量解压) +- models/ +- body_models/smplx/SMPLX_NEUTRAL.npz +- genmo/gem_smpl.ckpt +- genmo/t5-3b-local/ # 完整tokenizer/config/权重 +- kimodo/Kimodo-SMPLX-RP-v1/ # model.safetensors、config.yaml及完整stats/ +- kimodo/Meta-Llama-3-8B-Instruct/ # 完整base及分片 +- kimodo/LLM2Vec-Meta-Llama-3-8B-Instruct-mntp/ +- kimodo/LLM2Vec-Meta-Llama-3-8B-Instruct-mntp-supervised/ +- OMG/official/checkpoints/paper/50m/sstep=090000.ckpt +- OMG/t5-base-local/ +- TextOp-Data/TextOpRobotMDAR/logs/pretrained/checkpoint/{ckpt_200000.pth,vae.pth} +- TextOp-Data/TextOpRobotMDAR/dataset/PRIVATE-DATA/meanstd.pkl +- TextOp-Data/TextOpRobotMDAR/description/robots/g1/ +- TextOp-Data/clip/ViT-B-32.pt +- asr/Qwen3-ForcedAligner-0.6B/ # 重建对齐时需要 +-``` +- +-GenMo权重来自`nvidia/GEM-X` revision `5b9a69d5ed4713ac4ba2a63b335e66db87279a3e`的`gem_smpl.ckpt`。Kimodo权重来自`nvidia/Kimodo-SMPLX-RP-v1` revision `1419ba56b734c48bbafb41fefa84088ca94583b5`;两个LLM2Vec适配器revision和SHA256在`UPSTREAM_LOCK.json`。Llama本机来源为ModelScope的`LLM-Research/Meta-Llama-3-8B-Instruct`,也可按已有授权获取对应HF base,不能用量化版或不同base替代。SMPL-X主体需自行从已授权来源获取;本机neutral-2020文件SHA256为`bdf06146e27d92022fe5dadad3b9203373f6879eca8e4d8235359ee3ec6a5a74`。 +- +-GenMo需要四个额外身体回归/顶点表,URL、大小、revision和hash已固定: +- +-```bash +-/root/.local/bin/turbo /data/hfd_work/envs/genmo/bin/python tools/fetch_genmo_helpers.py \ +- --genmo-root /data/hfd_work/hfd-human-models/genmo +-/data/hfd_work/envs/genmo/bin/python tools/prepare_human_runtime.py \ +- --models-root /data/hfd_assets/models --genmo-root /data/hfd_work/hfd-human-models/genmo +-``` +- +-Kimodo安装时跳过setup中的MotionCorrection自动构建,避免其另行联网拉取构建依赖;随后由准备器使用已安装的Eigen和pybind11构建,并关闭CMake联网。准备器也建立本地base/adapter链接,不复制权重。构建需要C++17编译器和CMake>=3.15: +- +-```bash +-/data/hfd_work/envs/kimodo/bin/python tools/prepare_human_runtime.py \ +- --models-root /data/hfd_assets/models --kimodo-root /data/hfd_work/hfd-human-models/kimodo \ +- --build-correction --build-dir /data/hfd_work/cache/kimodo-motion-correction-build +-``` +- +-## 目标机器配置 +- +-```bash +-/data/hfd_work/envs/omg/bin/python tools/configure_training.py \ +- --work-root /data/hfd_work --assets-root /data/hfd_assets \ +- --raw-root /data/hfd_assets/hfd_data \ +- --external-code-root /data/hfd_work/hfd-human-models \ +- --env-root /data/hfd_work/envs --gpu 0 +-source /data/hfd_work/training-configs/training.env +-``` +- +-生成器按所选配方保留batch、预算、损失、调度、随机种子及数据规则,改成本机路径和GPU。默认四模型均使用GPU0,分别运行。已有OMG/TextOp资源可用`--robot-models-root /existing/models`复用;`--output`可改配置目录。它还生成`omg-{50m,100m,300m,500m}.yaml`和`genmo-{v2,legacy}.yaml`;`--omg-size`和`--genmo-recipe`选择默认项。默认OMG仍为50M,GenMo改为v2;更大OMG的微批次/累积与各原生损失见[复现矩阵](TRAINING_REPRODUCTION_MATRIX.md)。 +- +-## 英文词时间缓存 +- +-优先把现有`cache/hfd/reference-alignments.sqlite`迁移到新工作盘。它是源语音词对齐及第三方recipe索引,session仍动态组装。复制应在写入进程结束后保留完整SQLite状态。 +- +-没有缓存时创建独立Python3.10环境`forced-align`,按`runtime-versions.json`安装Torch,再安装`requirements/forced-align.txt`。准备`Qwen/Qwen3-ForcedAligner-0.6B` revision `c7cbfc2048c462b0d63a45797104fc9db3ad62b7`后运行: +- +-```bash +-/data/hfd_work/envs/forced-align/bin/python tools/align_reference_corpus.py \ +- --root /data/hfd_work/cache/reference-preparation \ +- --cache /data/hfd_work/cache/hfd/reference-alignments.sqlite \ +- --model /data/hfd_assets/models/asr/Qwen3-ForcedAligner-0.6B \ +- --language en --gpus 0 --batch-size 16 +-/data/hfd_work/envs/forced-align/bin/python tools/index_reference_recipes.py \ +- --cache /data/hfd_work/cache/hfd/reference-alignments.sqlite --language en \ +- --report /data/hfd_work/cache/reference-recipes.json +-``` +- +-双GPU可用`--gpus 0 1`。完成后应有61,860条英文源资产、缺失0;OMG/TextOp另用选定第三方英文recipe,GenMo/Kimodo仅用具有人体配对的原生英文session。重新对齐可能有数值差异;恢复原完整断点时应保留原缓存及代码/配置身份。原始tar、索引、PATH_MAP和配对metadata需完整;`tools/download_hfd_tars.py`可按已有manifest补齐分片。 +- +-## 预检、训练、恢复 +- +-从仓库根目录执行: +- +-```bash +-/data/hfd_work/envs/omg/bin/python train.py --config /data/hfd_work/training-configs/omg.yaml --dry-run +-/data/hfd_work/envs/textop/bin/python train.py --config /data/hfd_work/training-configs/textop.yaml --dry-run +-/data/hfd_work/envs/genmo/bin/python train_genmo_full.py --config /data/hfd_work/training-configs/genmo.yaml --check-only +-``` +- +-上述预检不更新参数。Kimodo的`--offline-check`会执行一次丢弃的梯度更新,加载全量索引及冻结文本worker。正式命令分别为: +- +-```bash +-/data/hfd_work/envs/omg/bin/python -u train.py --config /data/hfd_work/training-configs/omg.yaml +-/data/hfd_work/envs/textop/bin/python -u train.py --config /data/hfd_work/training-configs/textop.yaml +-/data/hfd_work/envs/genmo/bin/python -u train_genmo_full.py --config /data/hfd_work/training-configs/genmo.yaml +-/data/hfd_work/envs/kimodo/bin/python -u train_kimodo.py --config /data/hfd_work/training-configs/kimodo.yaml +-``` +- +-可在screen中前台执行。恢复时加`--resume auto`,不改变总预算;版本、路径或数据身份变化会拒绝伪装成同一次训练。`TRAINING_SNAPSHOT.json`与Git历史保留迁移前实现。跨机器从头复现使用生成配置;逐位复现还取决于缓存、完整随机状态、CUDA确定性及硬件,不能只凭seed保证。 +- +-GenMo旧权重的倾听姿态问题见[诊断说明](GENMO_LISTENING_DIAGNOSIS.md)。当前默认采用[GenMo v2训练策略](GENMO_V2_TRAINING.md),旧配方和原代码快照仍保留,未覆盖旧正式权重。新策略完整恢复使用`--resume auto`;旧权重迁入新策略使用`--init-checkpoint`,这是新run的权重初始化。 +diff --git a/docs/REFERENCE_DYNAMIC_TRAINING.md b/docs/REFERENCE_DYNAMIC_TRAINING.md +deleted file mode 100644 +index 99edc89..0000000 +--- a/docs/REFERENCE_DYNAMIC_TRAINING.md ++++ /dev/null +@@ -1,84 +0,0 @@ +-# 离线对齐与动态 session 训练 +- +-清理说明:本文记录的一次性实验脚本已归档,复现前请按[脚本恢复说明](WORKSPACE_CLEANUP.md)取回;历史指标与日志保留。 +- +-2026-10-19。当前实现保留原生 HFD 动态组装流程:缓存原始语音资产的词时间戳,训练时按实际组装位置平移时间戳。无需把全部 session 导出为 NPZ,也无需在线运行 ASR。 +- +-## 已落实的数据链路 +- +-1. `tools/align_reference_corpus.py` 从本地 tar 索引枚举原始任务、停止、附和、犹豫与修正语音,按唯一资产对齐。最初完成 123,720 个双语资产;按英文训练决策,当前库仅保留 61,860 个英文资产,英文缺失 0。没有批量解压或复制原始音频。 +-2. `/root/autodl-tmp/cache/hfd/reference-alignments.sqlite` 保存资产路径、原始音频 SHA256、16 kHz 长度、原文与词时间戳。训练时只读查表;不是把全部 session 常驻内存。进程内仍有受容量限制的 session/特征缓存。 +-3. 原生组装器在实际音频拼接时记录源资产和起止采样点,包括犹豫、停顿、修正、停止及附和。该元数据不改变动作、PCM、控制监督或随机组装流程。 +-4. 第三方已组装 session 按本地 `session_recipes.jsonl` 的前景目标语音位置映射,检查原始音频哈希与长度。背景音频保留,背景说话者的文本不作为条件。索引含 27,584 个 recipe、82,590 个目标片段;实际 train/val 仍以原选择清单的 26,737/783 个 session 为准。 +-5. 词结束后加实验性的 0.1 秒 guard,再向上取整到 `0.6, 1.0, 1.4, ...` 秒释放。模型只看到当前时刻已释放、过去 30 秒内的文字。原生动作时钟与麦克风时钟的 0.2 秒偏移按 `clock_policy: microphone` 统一处理。 +- +-24 个双语冻结样例的动态 PCM、动作与控制监督逐元素相同;动态生成的完整文本 records 也与先前导出的 `.reference.json` 完全相同。该检查覆盖 train/val、中英及六种交互类型,属于代表性回归检查,不表示已逐一实例化所有原生 session。 +- +-这属于 **oracle 前景转写条件**,不能作为真实 ASR 识别准确率实验。强制对齐也是时间估计,0.1 秒不是已证明的误差上界。历史双语库 8,342/123,720(6.74%)个资产存在无效跨度;当前保留的英文资产中为 6,608/61,860(10.68%)。这些资产保守地整段结束后释放文字;不猜测缺失词边界,不丢弃文本。 +- +-## 模型实际输入 +- +-| 模型 | 条件 | 监督与输出 | +-| --- | --- | --- | +-| OMG | 当前可用的过去 2 秒音频、已释放的最近 30 秒文字、原生动作历史 | 原生 G1 动作表示与原生训练损失 | +-| TextOp | 已释放的最近 30 秒文字、原生动作历史 | 原生 G1 动作表示、四 primitive 的原生三阶段 DAR 训练 | +- +-不输入控制标签、未来音频或未来原文。当前只采样英文前景 session,`text_conditioning: none`,不加载 ASR 或翻译模型。背景音频仍原样保留。中文翻译实验属于历史对照。 +- +-“保留 30 秒文字”是前端上下文约定,不代表编码器具有无限 token 容量。TextOp 的 CLIP 保留预算内的最近文本后缀;OMG 原生 T5 为 50 token,新 reference 配置通过适配工厂设置 `truncation_side=left`,超长时保留最近文字。架构、长度、参数与严格 checkpoint 加载不变;旧 run 不追溯修改。 +- +-## 配置和复查 +- +-新配置为 `configs/omg_reference_train.yaml`、`configs/textop_reference_train.yaml`,使用 `native_ondemand` 和 `aligned_reference_context`。`omg.yaml`、`textop.yaml` 也已同步为英文 reference 默认;旧 ASR 对照移至 `configs/history/`。OMG 使用物理 GPU 1,TextOp 使用物理 GPU 0。 +- +-```bash +-/root/autodl-tmp/envs/omg/bin/python train.py --config configs/omg_reference_train.yaml --dry-run +-/root/autodl-tmp/envs/textop/bin/python train.py --config configs/textop_reference_train.yaml --dry-run +-``` +- +-`--dry-run` 检查完整对齐索引、第三方 recipe 表和权重路径;完整读取与训练仍需短测。构建索引约需数分钟 CPU,首次组池和共享盘读取不能算作纯 GPU 更新吞吐。当前英文路线不做翻译。 +- +-本轮证据目录:`/root/autodl-tmp/runs/archive/legacy-20261020/reference-full-preparation-20261019/`。其中 `coverage.json`、`alignment-quality.json`、`dynamic-verification.json`、`provider-verification.json`、`translation-*.json` 分别记录覆盖、回退、动态文本、实际 provider 和翻译诊断。配置中的 100K/300K 步仍是旧预算;batch 改变后不能直接当作已验证的推荐预算。 +- +-TextOp 旧单 session 过拟合的 teacher-forced 短窗误差下降,但连续闭环误差变差。当前同 seed 引导系数 1 的补充检查仍是初始 0.3004 → 训练后 0.3121 rad;单纯把系数从 5 改成 1 没有解决。该证据不支持宣称全双工或长程动作能力已经验证。 +- +-全量正式训练尚未启动。当前英文短测报告将保存在 [纯英文短测报告](ENGLISH_ONLY_PROBE.md);历史双语结果不能直接与英文验证损失比较,必须使用同一英文回放集合。 +- +-## 历史双语短测结果 +- +-两模型都从官方权重初始化,batch=64,使用完整原生加第三方集合的随机窗口采样;不是逐一训练完所有 session。 +- +-| 检查 | OMG | TextOp | +-| --- | --- | --- | +-| 更新预算 | 512 步 | 768 步,三阶段各 256 步 | +-| 实际训练 session | 512 个槽位、510 个不同 ID | 192 个槽位、192 个不同 ID;每组四 primitive | +-| 原生验证损失 | 0.07212 → 0.04290,固定窗口和噪声 | 0.14213 → 0.14530,原生随机验证 | +-| 权重检查 | 512 次优化更新、梯度和损失有限、参数变化、严格重载通过 | 记录的梯度范数和损失有限,108 个状态张量变化,严格重载通过;未另挂 hook 独立计数 optimizer.step | +- +-OMG 池内单步中位耗时 0.115 秒,而包含冷池、验证和保存的训练阶段共约 1,064 秒;未包含此前索引与模型启动,不能用池内速度直接外推。OMG 每个 64-session 池准备中位约 103 秒;TextOp 三个池约 105/97/125 秒。当前耗时主要仍在共享盘读取、session 组装、首次翻译及特征准备,尚未证明正式 batch512 的端到端效率。 +- +-OMG 最后一轮因框架 lookahead 多准备了一个未训练的池,约 94 秒。这项开销已通过训练/验证 batch 上限修正;真实 CPU Lightning/DataLoader 的四步边界检查确认只请求四个源 batch,未触发下一冷池。57 项相关测试通过。本轮真实模型曲线和耗时来自修复前,未人为扣除这段耗时,也未重跑整个短测。 +- +-孤立批量翻译探针:64 条真实中文释放文本,batch16 与逐条结果 64/64 完全相同,纯模型翻译约 0.49 秒对 2.41 秒。这是后续优化依据,尚未改训练默认翻译策略,也没有解决译错动作的问题。 +- +-证据:运行目录下 `summary.json`、`pilot-curves.png`、`gpu-utilization.png`、`textop-checkpoint-verification.json`、`lightning-batch-bound-verification.json`;原生 run 的配置、采样记录、日志、验证曲线及 checkpoint 均保留。短测固定验证损失下降不等于连续闭环动作质量通过;旧 100K/300K 预算仍待校准。 +- +-## 正确中文直接输入的检查 +- +-2026-10-19 补充:文本转写正确与编码器支持中文是两项独立条件。原生冻结编码器实测如下: +- +-- OMG 的 T5-Base 将“向前走”“向后走”“向左转”“向右转”“停止动作”“继续动作”等 16 条中文对照指令全部编码为 `[3, 2, 1]`,其中 2 为 unknown token。相同 token 与 mask 的文本条件相同,后续仅训练动作网络无法恢复丢失的指令区别;音频及动作历史仍可能驱动输出。直接解冻这个编码器也不能区分已经被 tokenizer 合并的输入。 +-- TextOp 的 CLIP 字节 BPE 能保留不同中文字符,不能因此断言直接中文训练不可能。但当前冻结 CLIP 在这组手写 16 类中英文本匹配上仅 1/16 top1 正确。这是编码空间的小型诊断,不是中文动作准确率。 +- +-用上述短测 checkpoint,在同样两条中文验证 session(打断、急停)上保持种子、PCM、初始动作历史及文本释放时间相同,各回放中文原文、OPUS 英文和空文本,共 12 次。两条 session 的等权平均关节 RMSE: +- +-| 模型 | 直接中文 | 翻译英文 | 空文本 | +-| --- | ---: | ---: | ---: | +-| OMG | 0.4406 | 0.5121 | 0.4069 | +-| TextOp | 0.4355 | 0.3804 | 0.4307 | +- +-单位 rad,单种子 12345。OMG 的中文误差确实低于译文,但空文本更低;结合 token 坍缩,不能把这一结果解释成读懂中文。TextOp 译文在两个样例中均比中文误差低,但闭环关节误差仍不等于执行指令语义的准确率。 +- +-这只是已按英文条件训练的 checkpoint 的推理消融,尚未做直接中文与翻译英文的等预算训练比较。OMG 如需原生读懂中文,必须调整 tokenizer/encoder;TextOp 可单独做直接中文训练实验。更换中文/多语编码器还需要对齐原动作条件空间,并应标为独立模型变体。正式配置本轮未改变。 +- +-复现工具 `tools/probe_direct_chinese.py`;证据目录 `/root/autodl-tmp/runs/archive/legacy-20261020/direct-chinese-probe-20261019/`,其中 `encoders.json`、`omg-rollout.json`、`textop-rollout.json` 和 `summary.json` 保留完整结果。 +diff --git a/docs/ROLLOUT_DEBUG.md b/docs/ROLLOUT_DEBUG.md +deleted file mode 100644 +index fc94a97..0000000 +--- a/docs/ROLLOUT_DEBUG.md ++++ /dev/null +@@ -1,82 +0,0 @@ +-# 视频问题调试:输入、监督范围与执行节奏 +- +-清理说明:本文记录的一次性实验脚本已归档,复现前请按[脚本恢复说明](WORKSPACE_CLEANUP.md)取回;历史指标与日志保留。 +- +-本轮使用同一个原生双任务 session:接取别人递来的物品尚未完成时,输入捡起地上物品的新指令。数据未手工拼接,混合录音、背景音、ASR 与关节映射未改变。所有输出位于 `runs/hfd-final/`;主视频目录只展示两条当前结果,所有对照集中在 `debug/`。 +- +-## 结论与已采用修改 +- +-1. **监听姿态缺失首先是监督范围问题。** 之前把“不输入 listen 等控制标签”扩大成了“不训练这些帧的关节动作”。两个模型没有控制分类头,但都能拟合这些关节姿态。经用户确认改成 `full_session`:完整动作进入原生损失,控制标签仍不作为模型输入。原来 TextOP 仅 35 条四段连续训练链,OMG 仅 84 个两秒窗口;补回后分别为 88 条与 366 个窗口。监听、起步和收尾不再被过滤。 +-2. **OMG 重复动作与缺少收尾监督有关;卡顿还与异步计划交接有关。** 原先深弯腰在约 2.07、7.93、10.30、12.77 秒出现四次,当前主结果只有 9.57 秒一次,与真值相符。异步策略计算时执行旧计划,完成后跳过新计划前缀;两个计划在交接时的姿态不保证一致。最大跳变原来正好落在交接帧,并非播放器掉帧。该策略参照原生 `upstream/omg/src/omg/cli/pipeline/modes.py` 的 bridge/skip 流程,直接参考姿态回放没有物理跟踪器去缓和突变。 +-3. **TextOP 连续生成仍不合格,不能靠拉长文本窗口修好。** 它执行全部原生 8 帧(50 Hz 下 0.16 秒),且 2 帧动作历史是预测反馈,没有每个窗口重新初始化。固定真值历史时误差很小,但用自己的历史后逐渐停滞或延迟。完整监督后能出现监听姿态,却长时间停留在那里;捡取弯腰峰值约 11.66 秒,真值为 9.58 秒。延长训练反馈链也没有可靠解决。 +- +-当前 OMG 采用完整动作监督、过去两秒音频+持续因果 ASR、原生 DDIM 10 步、`joint_one` CFG、60 帧计划、剩余 24 帧时异步预取,使用实际测得规划耗时。修改的是原生采样参数,没有换网络、损失或添加手工平滑。TextOP 保留完整监督和原生 0.16 秒执行节奏,当前结果明确标记尚未收敛。 +- +-## 同 checkpoint 的对照 +- +-所有主对照固定 seed=12345。RMSE 为整段连续关节误差,排除开场真值初始化。TextOP 统计其 23 个关节,OMG 统计 29 个,不能把二者绝对值当作直接模型排名。 +- +-| 旧 motion-only checkpoint 的推理变化 | 连续 RMSE (rad) | 说明 | +-| --- | ---: | --- | +-| OMG 原推理 | 0.3294 | 50 步、测量延迟、异步 | +-| OMG 假设零延迟 | 0.3510 | 仅诊断;最大跳变从 0.4288 降到 0.1854 | +-| OMG 每 0.4 秒截断重规划 | 0.3708 | 不模拟计算耗时,仍差 | +-| OMG 执行完整 2 秒计划 | 0.3226 | 不模拟计算耗时,末尾仍重复 | +-| OMG 音频改为 0.6 秒 | 0.3204 | 零延迟,与训练 2 秒输入不匹配,仅诊断 | +-| OMG 音频改为完整前缀 | 0.3389 | 零延迟,前缀压成 60 个声学位置;仍无改善 | +-| OMG 去掉 ASR | 0.3867 | 零延迟,保持音频;文本条件有作用 | +-| OMG 去掉音频 | 0.3615 | 零延迟,保持 ASR;不能替代联合输入 | +-| TextOP 原推理 | 0.2434 | 原生 8 帧输出 | +-| TextOP 文本每 1.92 秒刷新 | 0.2441 | 仍执行原生 8 帧,改善几乎为零 | +- +-音频窗口和条件屏蔽是同 checkpoint 的敏感性检查;由于改变了训练时的输入分布,不能据此决定新训练的最佳窗口,更不能将诊断零延迟作为实时结果。 +- +-## 从官方权重重训,仅改为完整监督 +- +-TextOP 3200 次更新,阶段 `[800,1600,800]`,每次原生训练链 4 段;OMG 1600 次更新。其余参数与旧对应 run 保持一致。两者都通过有限 loss/梯度、参数更新、原生 checkpoint 严格加载检查。 +- +-| 完整监督后的设置 | 连续 RMSE (rad) | 最大逐帧跳变 RMS (rad) | +-| --- | ---: | ---: | +-| OMG 50 步+实际延迟 | 0.1865 | 0.3166 | +-| OMG 50 步+假设零延迟 | 0.0886 | 0.1242 | +-| OMG 10 步+实际延迟(当前) | 0.0619 | 0.1160 | +-| OMG 20 步+实际延迟 | 0.0622 | 0.1431 | +-| OMG 50 步+固定 5 帧延迟 | 0.0795 | 0.0976 | +-| OMG 10 步+固定 15 帧延迟 | 0.0753 | 0.2395 | +-| OMG 每次完整执行 2 秒,不模拟耗时 | 0.2293 | 0.0839 | +-| OMG 每 0.4 秒截断,不模拟耗时 | 0.2995 | 0.2400 | +-| TextOP 原生连续反馈(当前) | 0.2992 | 0.0197 | +-| TextOP 每块提供真值历史(oracle) | 0.0125 | 不作为连续性能 | +-| TextOP guidance=1 | 0.3247 | — | +-| TextOP 原生默认站姿初始化 | 0.4385 | — | +-| TextOP 去掉 ASR | 0.3260 | — | +-| TextOP 连续训练链从 4 段加到 16 段 | 0.2735 | 0.3357 | +- +-固定延迟交叉对照表明,采样轨迹与交接延迟都会影响结果,不能把全部误差改善只归因于速度。10 步采样平均规划约 0.157 秒,20 步约 0.230 秒,先前 50 步约 0.5 秒。没有手工插值、滤波或利用真值重置来美化当前视频。 +- +-OMG 10 步在 seed=12345/12346/12347 下 RMSE 分别是 **0.0619 / 0.1538 / 0.0585 rad**。主视频沿用原先 seed=12345,没有替换成最优种子。仍有随机性和残余接缝;真值最大逐帧跳变是 0.0355 rad,当前 OMG 的 0.1160 仍大于真值。 +- +-## 为什么 TextOP 的窗口很准,完整视频却不对 +- +-同一 ASR、同一个种子下,使用任务中的真值历史,8 帧误差约 0.0037–0.0054 rad;使用开场 HFD 站姿,预测的短期动作幅度约 0.003 rad。该旧 checkpoint 从原生默认站姿可以产生更大动作,但完整回放并未因此变好,说明不是简单换初始化就能解决。 +- +-完整监督的 TextOP 在每块真值历史诊断中误差 0.0125 rad,预测历史闭环却为 0.2992 rad,支持“训练短片段可拟合,闭环历史分布与任务阶段跟踪不稳定”的判断。缺少 KV cache 不是这里已经证实的唯一原因。TextOP 只有 ASR 条件:同一稳定句子可跨监听、执行、完成多个阶段;ASR 本身不显式提供动作是否已完成。保留完整标签用于动作监督不会自动增加阶段记忆。 +- +-16 段连续训练(2.56 秒)比 4 段(0.64 秒)误差略低,却引入大跳变,因此未采用。当前 TextOP 仍需更系统的原生训练与闭环验证;此次不将扩充模型状态或增添控制头偷偷并入 baseline。 +- +-## 复现与查看 +- +-当前配置:`configs/omg_full_session.yaml`、`configs/textop_full_session.yaml`。重训时改 `paths.output_dir` 为新目录。正式模板也已改为 `full_session`。当前 OMG 配置保存 `inference.omg_sampling_steps: 10`;TextOP smoke 保留此前诊断的 `static_prob=0.05`,正式模板仍是上游设置,应在论文里独立记录这一超参数。 +- +-```bash +-# 在对应模型环境运行;输出目录需尚不存在 +-python infer_overfit.py --run /root/autodl-tmp/runs/archive/legacy-20261020/hfd-final/omg --output-dir /data/runs/omg-replay --continuous-only --skip-window-eval --task-seconds 0 14.8 +-python infer_overfit.py --run /root/autodl-tmp/runs/archive/legacy-20261020/hfd-final/textop --output-dir /data/runs/textop-replay --continuous-only --task-seconds 0 14.8 +-# 仅作反事实诊断:--omg-latency-frames 0、--textop-history-mode teacher_forced +-# 其他对照:--omg-sampling-steps 50、--omg-execute-frames 60 --omg-replan-mode receding +-python tools/summarize_rollout_debug.py --root /root/autodl-tmp/runs/archive/legacy-20261020/hfd-final +-``` +- +-主视频:`videos/omg_full_session_playable.mp4`、`videos/textop_full_session_playable.mp4`。曲线:`debug/trajectory_comparison.png`。全部数值:`debug/summary.json`;旧结果在 `debug/previous_motion_only/`。这些均为有原始音频的关节参考姿态回放,没有物理跟踪器。 +- +-本轮沿用 HFD 名义动作时钟和音频释放映射,ASR 为预计算的因果 sidecar,未加入实际 ASR 计算延迟。OMG 仅计入本次声学特征、文本编码和模型规划耗时,不等同于完整麦克风到机器人端到端实时系统。所有测试来自同一个训练 session,不能作为独立测试集或论文泛化结论。 +diff --git a/docs/SEMANTICS_CLOCK_FIX.md b/docs/SEMANTICS_CLOCK_FIX.md +deleted file mode 100644 +index 82611b7..0000000 +--- a/docs/SEMANTICS_CLOCK_FIX.md ++++ /dev/null +@@ -1,59 +0,0 @@ +-# 中文语义与麦克风时钟修复 +- +-清理说明:本文记录的一次性实验脚本已归档,复现前请按[脚本恢复说明](WORKSPACE_CLEANUP.md)取回;历史指标与日志保留。 +- +-2026-10-18。修复已接入新 OMG/TextOp 训练配置、共享按需/third-party 数据入口、GenMo 文本缓存与窗口、连续推理和视频评测。未重训模型;历史 run、权重、ASR 和旧视频保持不变。HY-Motion/Kimodo 继续以同一输入契约为后续适配目标,其 HFD 训练器尚未实现。 +- +-## 语义条件 +- +-保留 Whisper-tiny `task=transcribe` 的原始 ASR,先按 `causal_asr_context` 获取已释放的完整假设,再通过冻结的 [Helsinki-NLP/opus-mt-zh-en](https://huggingface.co/Helsinki-NLP/opus-mt-zh-en) 转成英文。纯英文和空文本原样通过,仍由模型各自的原生文本编码器编码。没有使用 GT 指令或动作标签修正文本。 +- +-`慢跑向前` → `Jogging forward.`;`向后跳` → `Jump back.`。原先这些中文在 OMG T5-Base 中容易合并为 `[3,2,1]`,其中 `2` 是 UNK。六条指令的回归测试验证了 T5 token 可区分、无 UNK,保留左右手、否定和次序。实际三个样本共 42 次规划,翻译后文本均无 UNK,最长 19 token,未触发原生 50 token 截断。 +- +-实现:`hfd_bridge/semantics.py`。模型 eval、无梯度、beam=4、无采样,逐个 ASR 假设独立转换;长输入按 128 源 token 分块,输出遇到长度限制显式失败,不静默丢弃历史。初始化保护训练/采样 RNG。SQLite 缓存按模型身份、解码参数、规范化器版本和源字符串区分;缓存有容量限制,不跨释放点积累翻译记忆。 +- +-模型唯一存放在 `/root/autodl-fs/pangxinlei/models/translation/opus-mt-zh-en`,权重 312,087,009 字节。HF metadata 经 AutoDL 获取,权重在 AutoDL 超时、HF 镜像限流后从 ModelScope 国内直连下载(`trust_env=False`),未回退个人代理;本地 `download_identity.json` 记录版本和 SHA256。OMG/TextOp 依赖锁新增所需 tokenizer/规范化依赖;GenMo 环境已安装 `sacremoses==0.2.0`,其余依赖沿用现有环境。可用 `HFD_TRANSLATOR_ROOT` 指定相同身份的本地模型。 +- +-试验过 Whisper 直接 `task=translate`,但本例翻译误差明显,已放弃。输出目录下 `asr-en/` 是被拒绝的候选,三个修复视频均使用原始 `omg-100k-validation/asr` 的 transcribe sidecar,再做上述独立文本翻译。 +- +-## 信息可用时刻与视频 +- +-HFD 原始动作时刻 `t_motion` 对应麦克风时刻 `t_mic=t_motion+0.2s`。原始音频窗口 `[0.4k,0.4k+0.6]` 在 `r_k=0.6+0.4k` 到达。新 `microphone` 策略只读取满足 `r_k<=t_mic` 的最近完整窗口和 ASR。首个窗口还未到达时输入为空文本及静音。音频先裁剪再做归一化、重采样和卷积特征,未来 PCM 不参与这些运算。 +- +-例如原生动作 0.4s 请求对应麦克风 0.6s,只能用到 PCM 0.6s;历史 `legacy_hfd` 会选择到 1.0s。共享实现为 `hfd_bridge/clocks.py`,30Hz/50Hz 数据与推理共用;原始 PCM 和 sidecar 不需要重做。 +- +-完整视频从麦克风 0s 播放,预测与 GT 同时前置 0.2s 首帧(OMG/GenMo 6 帧、TextOp 10 帧)。这是显示时间变换,原生参考轨迹和计分范围保持不变;不是单独平移预测去匹配 GT。GenMo 仍只对完整预测和完整人体 GT 各 retarget 一次,原生 `retargeted.npz` 保持原始时间,显示前缀只写入视频比较文件。 +- +-新配置显式记录: +- +-```yaml +-caption_policy: causal_asr_context +-clock_policy: microphone +-text_conditioning: asr_to_english +-``` +- +-OMG/TextOp 位于 `data` 下,GenMo 位于顶层。无新字段的旧 run 保留 `legacy_hfd`、不翻译;新 GenMo 缓存记录翻译策略并包含启动所需空文本 embedding,不允许将旧缓存误当成新输入。 +- +-**延迟口径:** ASR/翻译为逐释放点因果离线缓存,计算耗时没有加入回放时间;OMG 规划仍使用实测生成耗时。信息因果性已验证,但不代表测得了包含前端的端到端实时延迟。 +- +-## 验证与同权重回放 +- +-输出:`/root/autodl-tmp/runs/archive/legacy-20261020/omg-semantics-clock-fix`。使用原始 100k OMG 权重、seed12345、DDIM10、native async、过去 2s PCM;对照旧 `causal_asr_context` 视频,同时修复语义与时钟,不能用这组三样本区分两项修复各自的效果。 +- +-| 样本 | 旧关节 RMSE(rad) | 新关节 RMSE(rad) | 旧根位置 RMSE(m) | 新根位置 RMSE(m) | +-| --- | ---: | ---: | ---: | ---: | +-| interruption | 0.3244 | 0.3539 | 1.6918 | 1.5484 | +-| backchannel | 0.2625 | 0.2776 | 0.0100 | 0.0112 | +-| emergency-stop | 0.4614 | 0.3864 | 0.4912 | 0.1547 | +- +-检查结果: +- +-- 70 项共享回归测试通过,另有 11 项 GenMo 几何/音频测试通过。 +-- OMG/TextOp 的真实数据契约检查通过:未来动作、未来 ASR/标签、未来 PCM 不影响当前条件,标量/批量音频特征一致。 +-- 三个视频共 1,638 帧完整解码且有音轨;NPZ 音频严格等于原始 PCM 从 0 开始的相应片段。 +-- 新视频去掉共同显示前缀后,GT 与旧评测逐元素完全相同;三个样本均链接同一原始 checkpoint。 +-- 42 次规划的 PCM 截止和 ASR 释放时刻均不晚于请求的麦克风时刻;保存每次原始 ASR、英文条件及实际 T5 token。 +- +-可复查 `verification.json`、`tests.log`、`genmo-tests.log`、`omg-contract.log`、`textop-contract.log`。逐次验证脚本为 `tools/verify_semantics_clock.py`。 +- +-**当前限制:** 输入已能区分中文,但 Whisper 错字、静音幻觉和翻译误差仍在;完整 ASR 也不保证模型理解纠错。旧权重曾在中文 UNK/旧时钟下训练,此次没有重训,动作仍有明显偏差,指标并非全面改善。不能据此宣称提前倾听、不执行指令等模型行为已经修好。 +diff --git a/docs/TEXTOP_ENGLISH_VALIDATION.md b/docs/TEXTOP_ENGLISH_VALIDATION.md +deleted file mode 100644 +index ecfd00c..0000000 +--- a/docs/TEXTOP_ENGLISH_VALIDATION.md ++++ /dev/null +@@ -1,83 +0,0 @@ +-# TextOp 300K 英文验证评测(2026-10-21) +- +-训练目录:`/root/autodl-tmp/runs/hfd-textop-english-train`,使用完整训练后的 +-`ckpt_300000.pth`,与官方初始化checkpoint做同条件对照。没有继续训练或改动已有权重。 +- +-结果根目录:[textop-english-validation-20261021](/root/autodl-tmp/runs/textop-english-validation-20261021/README.md)。 +-包含18条有原混合音轨的视频、原始qpos、文本/token审计、指标CSV、汇总JSON、比较图及验证记录。 +- +-## 样本与协议 +- +-- 六类原生英文验证各一条:普通指令、打断、急停、附和、犹豫、暂停;另取三条第三方验证。 +-- 使用训练时相同的val provider、配方、seed42、原生与第三方选择摘要和离线文本库; +- 导出动作、音频及逐帧文字与动态provider逐值相同,原生任务source hash划分核对通过。 +-- 样本在生成前选定:各类首条估计6–30秒的合格英文session,第三方按合格ID等距选取。 +- 不是按结果挑选。当前没有独立test,val曾用于训练监测。 +-- 两组生成seed均12345;相同一次初始GT历史/根锚点,之后只反馈预测。 +- 原生2帧57D历史、8帧primitive、50Hz、CFG5,使用训练配置中的5步原生扩散。 +- 覆盖所选session内的完整primitive,末尾不足8帧不额外补齐或重置GT。 +-- 文本为`aligned_reference_context`:已释放最近30秒英文GT前景文本,CLIP77最近后缀; +- microphone时钟、0.1秒guard,不运行ASR/翻译,不输入音频或控制标签。 +-- TextOp直接生成G1,不做人体系retarget;未建模腕关节固定,主关节指标只算原生23个关节。 +-- 视频左预测右GT,保留0.2秒共同microphone前置帧;评分排除前置及GT bootstrap, +- 不做DTW或事后根对齐。同步运动学回放,不把实测规划耗时模拟为控制器延迟。 +- +-## 结果 +- +-以下按session等权平均,关节单位rad,根xyz分量RMSE单位m: +- +-| 范围 | 初始关节RMSE | 300K关节RMSE | 初始根RMSE | 300K根RMSE | +-| --- | ---: | ---: | ---: | ---: | +-| 全部9条 | 0.4380 | 0.4152 | 1.2243 | 1.4877 | +-| 原生6条 | 0.4481 | 0.4569 | 0.7812 | 0.9121 | +-| 第三方3条 | 0.4178 | 0.3318 | 2.1105 | 2.6390 | +- +-整体关节RMSE降低约5.2%,根RMSE增加约21.5%;原生六类未整体改善。 +-附和、暂停样本改善,急停和打断仍有明显问题,不能以局部或平均关节误差下降宣称任务能力提升。 +- +-| 原生场景 | 初始关节RMSE | 300K关节RMSE | 初始根RMSE | 300K根RMSE | +-| --- | ---: | ---: | ---: | ---: | +-| 普通指令 | 0.4346 | 0.4541 | 0.2770 | 0.1805 | +-| 打断 | 0.4084 | 0.3946 | 1.3258 | 2.6228 | +-| 急停 | 0.5908 | 0.6995 | 1.4652 | 2.1198 | +-| 附和 | 0.3017 | 0.2429 | 1.5385 | 0.5089 | +-| 犹豫 | 0.4440 | 0.5042 | 0.0399 | 0.0328 | +-| 暂停 | 0.5091 | 0.4462 | 0.0410 | 0.0079 | +- +-优先查看: +- +-- [急停300K](/root/autodl-tmp/runs/textop-english-validation-20261021/videos/emergency-stop-trained.mp4) +-- [打断300K](/root/autodl-tmp/runs/textop-english-validation-20261021/videos/interruption-trained.mp4) +-- [附和300K](/root/autodl-tmp/runs/textop-english-validation-20261021/videos/backchannel-trained.mp4) +-- [全部对照与指标](/root/autodl-tmp/runs/textop-english-validation-20261021/README.md) +- +-急停样本`val-000111-seed42`中,`Stop.`在麦克风10.6秒已经释放且实际送入CLIP; +-12–16.88秒区间,预测平均根速度约1.008m/s、GT约0.014m/s,模型仍持续运动。 +-这是单条样例的停止后运动诊断,不是经过统一定义的停止延迟指标;可以排除该例是因 +-停止文字未释放、被token截断或ASR漏识别造成。不能仅凭此确定训练目标、历史依赖或文本理解 +-中哪一项是主要原因。详见结果目录的`emergency-stop-diagnostic.json`。 +- +-训练日志最终`eval_total=0.10740`;最后100K阶段最低约0.10626(250K)。 +-训练验证样本/阶段历史分布与连续闭环不同,不能把该loss解释为上述视频误差或用全程曲线 +-单独证明过拟合。完整抽取的150个验证点见结果目录`training-curve.json`。 +- +-## 复现与验证 +- +-```bash +-cd /root/autodl-tmp/hfd-baseline-portable +-/root/autodl-tmp/envs/textop/bin/python -u evaluate_textop.py \ +- --trained-run /root/autodl-tmp/runs/hfd-textop-english-train \ +- --output /root/autodl-tmp/runs/textop-english-validation-NEW \ +- --gpu 0 +-``` +- +-入口支持`--checkpoint`、`--stage prepare/run/summarize`。原生模型、冻结VAE、CLIP、统计量 +-均复用本地资产;不重新下载,不把验证样本写入训练数据。 +-推理入口已支持显式TextOp checkpoint,并可跳过无关训练窗口评测;每个primitive记录原文、 +-实际编码后缀、CLIP token ID、条件帧、释放时刻和规划耗时。 +-视频长文本按列宽换行,过长仅显示最近150字符;完整模型输入仍保留在report,不改变条件。 +- +-18条视频共16,504帧,完整解码、原音频一致、两权重GT一致、计划时钟/文本/token一致、 +-CLIP77与释放因果边界检查通过。配套25项数据/时钟/文本回归检查通过。 +-这些验证说明评测链路一致,不证明语义成功或统计显著改善。 +-共享stationary池、第三方与原生的源重叠及上游预训练重叠未排除,不能称完全未见动作泛化。 +diff --git a/docs/TRAINING_REPRODUCTION_MATRIX.md b/docs/TRAINING_REPRODUCTION_MATRIX.md +deleted file mode 100644 +index e420551..0000000 +--- a/docs/TRAINING_REPRODUCTION_MATRIX.md ++++ /dev/null +@@ -1,44 +0,0 @@ +-# 训练复现覆盖范围 +- +-当前发布四种 baseline 的训练入口,HY-Motion 不在已完成训练框架范围内。源码、正式配置、依赖版本、外部源码 revision 和必要资产身份均在仓库;数据与模型文件按单独授权准备,不上传权重/日志/测试脚本。 +- +-## 现有配方和 OMG 多个初始化权重 +- +-| 配方 | 入口/生成配置 | 训练预算 | 微批次 × 累积 | 说明 | +-|---|---|---:|---:|---| +-| OMG 50M paper | `train.py` / `omg-50m.yaml` | 100K optimizer steps | 512 × 1 | 已有 HFD 正式实验使用此初始化 | +-| OMG 100M paper | `train.py` / `omg-100m.yaml` | 100K | 64 × 8 | 独立输出目录;有效 batch 512 | +-| OMG 300M paper | `train.py` / `omg-300m.yaml` | 100K | 32 × 16 | 独立输出目录;有效 batch 512 | +-| OMG 500M paper | `train.py` / `omg-500m.yaml` | 100K | 16 × 32 | 独立输出目录;有效 batch 512 | +-| TextOp | `train.py` / `textop.yaml` | 300K,三阶段各 100K | 512 × 1 | DAR/VAE/CLIP/57D 统计及机器人资源配套 | +-| GenMo legacy | `train_genmo_full.py` / `genmo-legacy.yaml` | 100K | 32 × 1 | 保留原正式配方及旧代码快照 | +-| GenMo v2 | `train_genmo_full.py` / `genmo-v2.yaml` | 100K | 32 × 1 | 新恢复课程、执行段损失、闭环选权重 | +-| Kimodo | `train_kimodo.py` / `kimodo.yaml` | 100K | 8 × 1 | 当前因果动作表示、根锚点约束、预测历史恢复及冻结文本 worker | +- +-其他 OMG 尺寸是可复现的初始化与微调配置,不代表已经各自完成 HFD 全量训练。各尺寸的实际文件名和 SHA256 见 `configs/omg_variants.yaml`,下载来源固定为 `THU-MARS/OMG` revision `06a88419da9f9125d0b5480cc10966741a189358`。100M/500M 需显式打开与权重相符的音频/人体分支结构;人体条件的可用性mask仍为false,HFD适配不输入未来人体GT。500M 仅移除已知的 52 个废弃 `visual_film` 参数,其他缺失/多余参数仍拒绝加载。保留各原生 experiment 的损失配置,100M/500M 的 full loss 与 50M/300M 的 diffusion-only loss 不能混称完全相同的训练目标。 +- +-上述 OMG 微批次按单卡 32GB 保守设置,均保持有效 batch 512;实际峰值还受文本长度与数据影响,不能把单次检查当作最大 batch 调优。50M/100M/300M/500M 均已用对应微批次完成严格权重匹配和一次丢弃的参数更新。历史 HFD run 中的 `best`、98K、100K 等是同一次训练的不同 checkpoint,均由原生验证与保留策略产生;它们不是新的模型尺寸,也不在 Git 中。 +- +-## 生成与运行 +- +-按便携指南准备资产后: +- +-```bash +-/data/hfd_work/envs/omg/bin/python tools/configure_training.py \ +- --work-root /data/hfd_work --assets-root /data/hfd_assets \ +- --raw-root /data/hfd_assets/hfd_data --omg-size 300m --genmo-recipe v2 --gpu 0 +-source /data/hfd_work/training-configs/training.env +-/data/hfd_work/envs/omg/bin/python -u train.py --config /data/hfd_work/training-configs/omg-300m.yaml +-``` +- +-生成器始终输出四种 OMG 配置及两种 GenMo 配方;`omg.yaml` 和 `genmo.yaml` 保存所选默认项。所有配置默认单卡执行,不能把多条命令同时塞到同一块卡。 +- +-## 外部资产、数据与环境 +- +-`TRAINING_ASSETS.json` 记录必要模型、统计、机器人资源、SMPL-X 主体、原始索引和配对表的相对路径、大小、SHA256 以及已确认的下载来源。`tools/training_assets.py` 默认只验证;显式 `--download` 才会下载缺失文件,不覆盖已有文件。Llama 需按记录哈希获取同一份已授权 base,SMPL-X 主体仍需自行获取。派生 runtime adapter 路径与本机原生扩展由准备器重建。 +- +-HFD 原始数据版本必须使用 `93ba4a7a9c66b3e1b7041e5f4ea038ef2418b948`;旧 `7745110...` 缺少新增人体元数据/分片。先下载同一版本的索引、PATH_MAP、manifest 和人体配对表,再按 manifest 补齐 tar,保留 tar 不解压。训练不自动联网补缺失文件。 +- +-源文本时间戳使用 `reference-alignments.sqlite`。迁移已完成的数据库并校验身份可保持既有释放时间;也可按固定 ForcedAligner、依赖和离线对齐命令重建。重建可能有数值差异,不应声称与旧缓存逐位相同。 +- +-原始数据、外部模型和离线缓存齐备后,训练不再依赖当前服务器的 `setup/`、其他适配仓库、测试目录或运行日志。四模型的完整断点均包含优化器/调度/采样或课程所需状态。严格恢复仍要求相同代码、配方和资产;Git 历史保留原训练快照。完整操作见 [PORTABLE_TRAINING.md](PORTABLE_TRAINING.md),GenMo v2 见 [GENMO_V2_TRAINING.md](GENMO_V2_TRAINING.md)。 +diff --git a/docs/TRAINING_RESUME.md b/docs/TRAINING_RESUME.md +deleted file mode 100644 +index ba9c43b..0000000 +--- a/docs/TRAINING_RESUME.md ++++ /dev/null +@@ -1,43 +0,0 @@ +-# OMG / TextOp 训练续训 +- +-2026-10-19:两个原生训练入口已接入完整 checkpoint 恢复,适用于当前一卡 OMG、一卡 TextOp。正式配置 `training.resume: null` 表示新训练;已有输出目录在没有显式 resume 时仍会拒绝覆盖。 +- +-从仓库根目录运行,自动选择原输出目录的 checkpoint: +- +-```bash +-/root/autodl-tmp/envs/omg/bin/python train.py --config configs/omg.yaml --resume auto +-/root/autodl-tmp/envs/textop/bin/python train.py --config configs/textop.yaml --resume auto +-``` +- +-`--resume` 不带参数等价于 `--resume auto`。找不到文件会报错,不会退回从头训练。OMG 在根目录 `last.ckpt`、`checkpoints/last.ckpt` 和周期 checkpoint 中按保存的步数选择最新文件,避免旧的根目录 last 覆盖较新的周期保存点;TextOp 同样按保存步数选择。忽略原子保存中的 `.partial` 文件。 +- +-显式指定恢复点: +- +-```bash +-/root/autodl-tmp/envs/omg/bin/python train.py --config configs/omg.yaml --resume /path/to/step-00002000.ckpt +-/root/autodl-tmp/envs/textop/bin/python train.py --config configs/textop.yaml --resume /path/to/ckpt_2000.pth +-``` +- +-也可在 YAML 设置 `training.resume: auto` 或 checkpoint 绝对路径。显式命令行参数优先。可以追加 `--dry-run` 先检查 checkpoint、配置与输出目录。 +- +-## 恢复内容与边界 +- +-- OMG 由 Lightning 恢复模型、优化器、scheduler、训练步数、precision 和原生 callback 状态。 +-- TextOp 恢复 denoiser、AdamW、全局步数及启用时的 EMA;学习率退火和三阶段位置由原生逻辑按已恢复步数继续计算。 +-- 新 checkpoint 额外保存 Python、NumPy、Torch/CUDA 随机状态、动态池的 session 索引、批次位置和采样 RNG。OMG 保存实际已消费 batch 的池快照,不使用可能已被预取推进的现场状态。 +-- 不保存大型音频/动作内存缓存。恢复时重建当前池所需内容,首次读取可能仍慢。 +-- TextOp 必须在完整 primitive 链边界恢复;配置检查要求保存间隔能被 `num_primitive` 整除。 +- +-`max_steps` 是原训练计划的总目标,不是恢复后再增加的步数。续训会检查模型、数据语言、文本时钟、batch、学习率、总步数、scheduler/阶段等配置一致;允许更换单张 GPU、调整保存/验证频率,或指定新的空输出目录。当前没有接入多 GPU 的完整 HFD 采样恢复。 +- +-官方预训练权重使用 `weights.init_checkpoint` 初始化,不能当作此次 HFD 续训点。历史 HFD checkpoint 如缺少新增随机/数据池状态,仍可恢复其完整优化器和步数,但会明确提示采样不能精确接续。新 checkpoint 的训练恢复也不承诺 GPU 逐位一致;原生验证的执行与随机数消耗会影响后续数值轨迹。 +- +-只能恢复到最近成功保存的位置,断电前尚未保存的更新不会恢复。两模型正式配置现在均每 **2000 步**保存,保留最近 **3** 份;OMG 另存最佳验证 checkpoint。100K/300K 总预算未改动,正式长训练未启动。 +- +-## 日志与检查 +- +-恢复后标量日志追加写入;原始 `resolved_config.yaml` 保留。每次恢复在 `resume_history//` 保存本次配置、checkpoint 来源、代码摘要、数据清单及恢复前的原生配置/训练摘要。显式回退到较早 checkpoint 时,标量日志可能出现重复步数,可用恢复记录区分。 +- +-已通过两个真实 GPU 原生训练入口的 **4 步停止 → 恢复到 8 步**检查:恢复后各完成4次更新,Adam状态计数从4到8,OMG scheduler接续,TextOp跨阶段接续,有限损失/梯度与严格重载通过。该 smoke 使用已导出的英文 session;动态池与预取的一致性另用 CPU Lightning 回归覆盖,不代表正式全数据训练已完成。 +- +-分环境共 **132 项回归测试**通过,包括预取启用下的消费位置、优化器与 scheduler 连续性、TextOp 原生 manager 跨阶段、RNG 和采样池恢复。证据见 [验证记录](/root/autodl-tmp/runs/archive/legacy-20261020/resume-validation-20261019/verification.json)。本次验证生成的临时 checkpoint 已清理,配置、日志和结果保留。 +diff --git a/docs/WORKSPACE_CLEANUP.md b/docs/WORKSPACE_CLEANUP.md +deleted file mode 100644 +index 1551b42..0000000 +--- a/docs/WORKSPACE_CLEANUP.md ++++ /dev/null +@@ -1,78 +0,0 @@ +-# 工作区清理(2026-10-19) +- +-检查了 `git diff --stat`、逐文件 diff、`git status --short` 的未跟踪文件,以及运行时代码/测试的引用。仅用 `git diff` 会漏掉大量未跟踪脚本。新增的 bridge、动态数据读取、对齐、GenMo 和 retarget 模块仍是实际功能,不能按“未跟踪”直接删除。 +- +-已将 **43 个一次性实验脚本**移出工作区;包括旧 ASR/翻译探针、不同阶段的预算和吞吐 benchmark、中英 pilot、已完成过拟合与验证的专用汇总脚本。旧的 `omg_asr_context_overfit.yaml`、`textop_asr_context_overfit.yaml`、`omg_budget60k.yaml` 三份配置移入 `configs/history/`,引用已更新。未重置或提交用户已有改动。 +- +-以下保留在日常工作区: +- +-- 正式训练入口、当前配置与 resume 实现。 +-- 动态 session/缓存适配、文本时钟、音频和动作特征、映射等运行时模块。 +-- OMG/TextOp/GenMo 推理、retarget、视频评测入口及必要回归测试。 +-- 原始数据、模型、已有实验权重、日志和报告。 +-- 七个数据准备/审查工具:`align_reference_corpus.py`、`audit_reference_text.py`、`download_hfd_tars.py`、`index_reference_recipes.py`、`prepare_reference_sessions.py`、`verify_dynamic_reference.py`、`verify_reference_release.py`。 +- +-## 归档与恢复 +- +-脚本移出前已打包,并逐文件验证归档内容 SHA256,清单保存原路径、大小和哈希。归档在服务器工作区外,若迁移后还需复现实验,应同时复制该归档。 +- +-- [归档包](/root/autodl-tmp/hfd-code-archive/20261019-resume-cleanup/historical-experiments.tar.gz) +-- [逐文件清单](/root/autodl-tmp/hfd-code-archive/20261019-resume-cleanup/cleanup.json) +-- [清理前 Git 状态](/root/autodl-tmp/hfd-code-archive/20261019-resume-cleanup/git-status-before.txt) +-- [清理前已跟踪文件 diff](/root/autodl-tmp/hfd-code-archive/20261019-resume-cleanup/git-diff-before.patch) +- +-按需恢复某个脚本,例如: +- +-```bash +-tar -xzf /root/autodl-tmp/hfd-code-archive/20261019-resume-cleanup/historical-experiments.tar.gz -C /root/autodl-tmp/hfd-baseline-portable tools/run_english_training_probe.py +-``` +- +-历史报告中的脚本路径属于当时的实验记录,已添加归档提示;当前报告、指标和视频链接保留。归档 SHA256:`1fc40fe878610f4c37610d77a56aee9a5db70bfb32ed9bf7c30d25ae71fba06c`。 +- +-## 移出工作区的脚本 +- +-| 原路径 | 字节 | +-| --- | ---: | +-| `tools/analyze_omg_budget.py` | 5110 | +-| `tools/audit_asr_context.py` | 3829 | +-| `tools/audit_genmo_audio.py` | 4258 | +-| `tools/audit_genmo_inputs.py` | 3353 | +-| `tools/audit_reference_translation.py` | 3706 | +-| `tools/benchmark_omg_data.py` | 6149 | +-| `tools/benchmark_omg_io.py` | 2526 | +-| `tools/benchmark_omg_resident.py` | 4317 | +-| `tools/benchmark_omg_training.py` | 8946 | +-| `tools/benchmark_textop_batch.py` | 7413 | +-| `tools/benchmark_translation_batch.py` | 2800 | +-| `tools/check_genmo_ambiguity.py` | 1781 | +-| `tools/compare_asr_models.py` | 7332 | +-| `tools/compare_english_checkpoints.py` | 6201 | +-| `tools/diagnose_omg_validation.py` | 6358 | +-| `tools/eval_bilingual_pilot.py` | 3519 | +-| `tools/eval_reference_overfit.py` | 3580 | +-| `tools/initial_bilingual_validation.py` | 2183 | +-| `tools/prepare_bilingual_pilot.py` | 3325 | +-| `tools/prepare_omg_validation.py` | 4446 | +-| `tools/probe_asr_latency.py` | 2183 | +-| `tools/probe_bilingual_condition.py` | 4904 | +-| `tools/probe_direct_chinese.py` | 7190 | +-| `tools/probe_reference_alignment.py` | 15248 | +-| `tools/prune_chinese_caches.py` | 6225 | +-| `tools/render_bilingual_pilot.py` | 2202 | +-| `tools/run_bilingual_pilot.py` | 4289 | +-| `tools/run_english_training_probe.py` | 1537 | +-| `tools/run_omg_validation.py` | 5170 | +-| `tools/summarize_asr_context.py` | 5454 | +-| `tools/summarize_asr_throughput.py` | 6969 | +-| `tools/summarize_bilingual_pilot.py` | 8655 | +-| `tools/summarize_english_probe.py` | 7373 | +-| `tools/summarize_genmo_audio.py` | 5254 | +-| `tools/summarize_genmo_overfit.py` | 4932 | +-| `tools/summarize_omg_validation.py` | 9952 | +-| `tools/summarize_reference_overfit.py` | 9336 | +-| `tools/summarize_reference_preparation.py` | 6832 | +-| `tools/summarize_rollout_debug.py` | 4354 | +-| `tools/summarize_textop_batch.py` | 5570 | +-| `tools/verify_bilingual_pilot.py` | 3282 | +-| `tools/verify_native_batch_limit.py` | 2224 | +-| `tools/verify_semantics_clock.py` | 4106 | +diff --git a/docs/history/BASELINE_METHOD_BEFORE_SEMANTIC_FIX.md b/docs/history/BASELINE_METHOD_BEFORE_SEMANTIC_FIX.md +deleted file mode 100644 +index 0ea8331..0000000 +--- a/docs/history/BASELINE_METHOD_BEFORE_SEMANTIC_FIX.md ++++ /dev/null +@@ -1,95 +0,0 @@ +-# 最终采用的 baseline 输入与推理方式 +- +-本轮完成了用户提出的累计音频、原生推理节奏对照。两模型仍使用原生网络、损失、训练循环和官方初始化权重;没有新增 KV 模块、控制分类头、任务进度真值或动作平滑。测试样本为原生双任务 interruption session `train-002203-seed42`:接取物品未完成时输入捡起地上物品的新指令。原始声音、背景音和动作时间线保留。 +- +-## 最终选择 +- +-| 模型 | 输入 | 动作生成与衔接 | +-| --- | --- | --- | +-| OMG | 保存原始录音;模型每次编码截至 HFD 当前释放时刻的最近两秒 PCM。开头左侧补零,保留原生 60 个声学条件位置 | 每次预测原生 60 帧 / 两秒;模拟原生异步参考缓冲,剩余 24 个 30 Hz 参考帧(0.8 s)时请求新计划。计算期间继续执行旧参考,新计划跳过已耗时对应帧后接入 | +-| TextOP | 由已释放音频生成的因果 ASR;当前实验保持最新稳定 ASR 句子,文字内容连续两次一致才更新,标点变化不算新命令;继续使用原生 CLIP token 上限 | 原生 8 帧 / 0.16 s primitive,约 6.25 Hz;持续反馈预测历史,换文本时不重置姿态;原生 guidance=5 与三阶段训练 | +- +-音频到达频率、语义条件变化、动作生成频率是三个独立事件。音频仍按 HFD 0.6 秒窗口 / 0.4 秒释放间隔提供可用边界;模型可以重新编码更长的已听历史,但不能读取边界之后的录音。OMG 不强制每 0.4 秒抛弃未来计划,也不等整个任务做完才处理新指令。 +- +-OMG 缓冲参数对应原生 tracker 50 Hz 下剩余 40 帧的 0.8 秒时间,当前在 30 Hz 的参考轨迹上实现为 24 帧。两秒计划减去 0.8 秒缓冲,对应大约 1.2 秒的规划请求间隔;新计划真正接入还受计算时间影响。 +- +-## 为什么没有采用“全历史直接输入” +- +-累计录音已经保留。另一个 OMG 模型从官方初始化,用 `hfd_full_prefix` 在同一 session 上训练 1600 步:每个动作点读 `[0,当前释放截止)` 的全部 PCM,再把原生声学特征线性插值到固定 60 个条件位置。训练和推理策略一致,未来录音不会影响当前特征。 +- +-这个适配没有带来更好的结果,且随着历史增长降低时间分辨率、增加重新编码成本。OMG 原生音频编码为 35D 声学特征,不是语言语义编码器。此结果只能否定本次“全历史压缩到固定位置”的适配效果,不能概括为完整音频历史没有价值。 +- +-最终采用两秒历史,是为了保留原生特征的时间排列和一致的训练/推理接口。超过两秒的指令仍可能被截断,这是明确的 baseline 上下文限制。改成可变长语义编码器将涉及网络改造,本轮没有这样做。 +- +-TextOP 输入的是文字,不是原始 PCM。此 14.8 秒样本短于 ASR 的 30 秒历史上限,ASR 本来就在每个截止时刻重新识别从开始到当时的全部录音。真正改变的是供给 CLIP 的指令组织和稳定保持方式。`stable_latest_sentence` 是词法预处理,不使用 GT turn、事件标签或真值任务文本;它也不是可靠的语音任务边界检测器。多句复合指令、复杂犹豫或 ASR 幻觉可能使末句不代表完整任务;可用 `rolling_transcript` 做保留完整转写的对照,正式训练需在既定验证集核对这一限制。 +- +-## 结果 +- +-所有误差来自该训练 session 的连续生成,开头真值初始化帧被排除;不代表独立测试集或完整全双工评测。TextOP 指标为 23 个建模关节,OMG 为 29 个,不能直接横向排名。 +- +-| 实验 | 连续关节 RMSE(rad) | root RMSE(m) | +-| --- | --- | --- | +-| OMG 两秒历史,0.4 s 重规划(前轮) | 0.25026 | 0.04641 | +-| OMG 两秒历史,同步两秒块执行(前轮) | 0.18776 | 0.03093 | +-| OMG 两秒历史,模拟原生异步缓冲,本轮最终测时 | 0.09436 | 0.01764 | +-| OMG 全历史压缩,模拟原生异步缓冲 | 0.22073 | 0.03660 | +-| TextOP 滚动转写、原生周期、guidance=5(前轮) | 0.40908 | 0.03739 | +-| TextOP 稳定最新句、原生周期、guidance=5 | 0.30987 | 0.02933 | +- +-OMG 两秒历史模型已在前轮完成同预算 1600 步训练,本轮复用其唯一最终权重测试缓冲;累计历史模型和新 TextOP 均重新从官方初始化训练 1600 步。没有为不同执行周期挑选不同检查点。 +- +-TextOP 真实历史窗口 RMSE 为 0.00700 rad,但连续生成仍未充分拟合。相同静止历史下切换两条指令,8 帧输出差异约 0.00066 rad;相同文本下换成真实动作历史,输出差异约 0.22 / 0.47 rad。这说明此次小样本拟合仍主要依赖姿态续接,不能把稳定文本和原生周期描述为已经解决任务启动问题。 +- +-OMG 固定时延敏感性对照:异步缓冲用零计算时延时为 0.10596 rad,固定 0.2 秒为 0.08031 rad,最终实测时延为 0.09436 rad。人工固定时延仅用于诊断,最终没有选择最有利的 0.2 秒去改善指标。 +- +-## 推理耗时与时钟边界 +- +-本机 RTX 5090、FP32 原生 PyTorch,模型加载/预热后: +- +-| 输入 / 模型 | 每次规划平均耗时 | P95 | +-| --- | --- | --- | +-| OMG 两秒历史 | 0.479 s | 0.489 s | +-| OMG 累计历史 | 0.523 s | 0.555 s | +-| TextOP primitive | 0.066 s | 0.112 s | +- +-OMG 计时包含本轮声学特征、历史 FK/编码和原生生成,并在 CPU 取回预测后结束计时。TextOP 包含当前 CLIP 条件读取/编码及原生生成,不包含之前生成 ASR sidecar 的 Whisper 计算。渲染与真实机器人控制器不计入。 +- +-在当前未优化后端,OMG 本身已超过 0.4 秒预算,不能宣称 0.4 秒内完成完整推理和执行。异步缓冲允许旧参考在规划期间继续输出,最终实测无缓冲耗尽。模型输出回到缓冲时,按实测时间向上取整到 30 Hz 帧数,跳过新计划中已过去的帧;旧参考提供这段 bridge,而不是把未执行的新预测当作实际历史。 +- +-这是离线源时间参考轨迹上的执行模拟,使用真实测得的规划耗时;视频仍是原始音频内容对齐的姿态回放。它不是实时麦克风、ASR、机器人 tracker 的完整墙钟部署,不从视频直接推导论文中的硬件反应时延。HFD 音频释放截止、源动作时间、真实计算耗时分别记录在报告中。 +- +-## 配置、复现与迁移 +- +-采用的短训练配置:`configs/history/omg_adopted.yaml`、`configs/history/textop_adopted.yaml`。正式训练模板 `configs/omg.yaml`、`configs/textop.yaml` 也记录相同输入和推理策略,但独立数据划分、正式训练和论文评测尚未完成。按根目录 README 安装两个环境并准备外部权重、ASR、机器人 assets;更新 YAML 中路径、GPU 和 batch size 后训练到新的 output_dir。 +- +-```bash +-python train.py --config configs/history/omg_adopted.yaml +-python train.py --config configs/history/textop_adopted.yaml +-python infer_overfit.py --run NEW_OMG_RUN --continuous-only --task-seconds 0 14.8 +-python infer_overfit.py --run NEW_TEXTOP_RUN --continuous-only --task-seconds 0 14.8 +-``` +- +-当前已验证权重与输出: +- +-- OMG 权重:`/root/autodl-tmp/runs/hfd-omg-context-debug/omg-past2-train/last.ckpt`;推理输出 `hfd-native-context-final/omg-past2-adopted`。 +-- TextOP 权重:`/root/autodl-tmp/runs/hfd-native-context-final/textop-stable-command/ckpt_1600.pth`;推理输出该 run 的 `inference`。 +- +-复现最终 OMG 用 `--omg-replan-mode native_async --output-dir NEW_OUTPUT`,因为该保留权重的旧 resolved_config 仍记录前轮重规划设置。新 adopted 配置把这些推理设置放在 `inference` 中,无需 CLI 覆盖。 +- +-重要配置项:`data.audio_policy`、`data.caption_policy`、`inference.omg_replan_mode`、`inference.omg_buffer_remaining_frames`、`inference.textop_guidance_scale`。`--omg-latency-frames` 是诊断覆盖,正式采用实测时间,不在配置中固定有利的延迟。 +- +-全历史对照用 `configs/history/omg_native_context.yaml`,其中 `hfd_full_prefix` 不作为最终默认。缓存仅用于训练时重复窗口的确定性声学特征,单 source 最多 128 组;推理不复用未接收的未来条件。缓存与直接提取逐项一致。 +- +-## 原生依据与核验 +- +-- [TextOP 原生部署论文](https://arxiv.org/html/2602.07439v1#S3.SS5):生成器 6.25 Hz,tracker 50 Hz。 +-- [OMG 原生执行文档](https://github.com/Tsinghua-MARS-Lab/OMG/blob/main/docs/generation.md):同步与异步参考缓冲。 +-- 本地固定版 `upstream/omg/src/omg/cli/pipeline/modes.py`:旧参考 bridge、新参考跳过 bridge 时段;原生 tracker 参数剩余 40 帧 / 50 Hz。 +- +-新两个训练的优化器更新、参数变化、有限 loss/梯度和严格检查点重载均通过。数据合同检查和 37 项回归测试通过,覆盖累计音频未来不可见、缓存一致性、ASR 稳定策略的因果性、缓冲 bridge/跳过位置与 underrun 检测。全历史模型的三个等价重复检查点已清理,唯一最终权重与对照记录保留。 +- +-## 当前语义保持方案(2026-10-07) +- +-OMG 的正式 smoke 配置使用 `motion_only` 监督,并把原生 ASR 作为因果文本条件接入。每个推理时刻仍使用最近 2 秒 PCM;ASR 的最后一个已稳定句子在后续背景音阶段持续保留,因此任务描述结束后不会因为音频窗口滑走而丢失目标。ASR 文本只使用当时已经释放的转写,不能看到未来文本或真值动作。 +- +-OMG 代码保留两种 guidance:`joint_one` 是当前可复现实验默认值(单次采样,适合异步缓冲);`native` 启用原生文本/音频分支引导。后者在本机一次规划约需 1.37 秒,在 0.8 秒异步余量下会缓冲欠载,已记录为诊断结果,不能宣称实时可用。TextOP 的 3200-step 起步诊断连续误差为 0.2434 rad,窗口误差为 0.0094 rad;这说明训练链路有效,但整段闭环仍未达到理想过拟合。 +diff --git a/docs/history/DUPLEX_TASK.md b/docs/history/DUPLEX_TASK.md +deleted file mode 100644 +index 8ba0556..0000000 +--- a/docs/history/DUPLEX_TASK.md ++++ /dev/null +@@ -1,79 +0,0 @@ +-> 历史实验记录;当前入口见 [README](../../README.md)。旧运行目录已移动到 `/root/autodl-tmp/hfd-runs-archive-20261007/`。 +- +-# 两个正常任务的完整全双工 session +- +-本例选用原生 HFD `train-002203-seed42`:任务一是接取别人递来的物品(`neutral_item_take_R_001__A545`),任务二是弯腰捡起地上的物品(`neutral_item_pick_up_floor_R_002__A544`)。英文女声 `female_02_en`;保留原始录音与空调背景音。两项任务都包含实质动作,原有 listen/idle 过渡仍保留在时间线上,标签没有作为模型的额外控制输入。 +- +-这是原生 interruption session,第二条指令到达时第一项动作尚未完成。第一项动作会被中断,随后完成第二项动作;不是声称两个动作都完整执行结束。两任务之间的姿态过渡由原生 HFD 构造器处理,没有人为把两个独立视频拼在一起,也没有对预测做平滑。 +- +-## 时间线 +- +-运动时间线长 14.8 秒,50 Hz 共 740 帧。 +- +-| 时间 | 原生内容 | +-| --- | --- | +-| 0–1.2 s | 开场 idle;第一条录音实际在运动时间 0.4–2.16 s 可听见 | +-| 1.2–3.2 s | 第一条指令的 listen/姿态过渡 | +-| 3.2–6.0 s | 第一项动作;第二条录音在运动时间 5.2 s 到达,动作仍在执行 | +-| 6.0–8.0 s | 第二条指令的 listen/姿态过渡 | +-| 8.0–12.0 s | 第二项实质动作:捡起地上的物品 | +-| 12.0–14.8 s | 收尾 idle | +- +-`turn_speech_start_frames=[20,260]`。音频数组相对运动有原生 0.2 秒前置偏移:未裁剪音频中的实际发声时间为 0.6–2.36 s、5.4–7.16 s。导出的视频按此偏移对齐,所以视频中是 0.4–2.16 s、5.2–6.96 s。原生中断记录显示第一项动作剩余 86 帧,即 1.72 秒;不是由视频外观猜测。 +- +-## 训练与视频口径 +- +-两个模型从官方初始化权重开始,用各自原生训练框架在此 session 上训练 1600 步。训练目标窗口覆盖该 session 的时间范围 `[0,730)`,只跳过尾部 0.2 秒 idle;`max_windows: null`,不局限于此前八个稀疏窗口。TextOP batch size 4、原生四 primitive 链;OMG batch size 8、原生 60 帧未来动作窗口。OMG 使用完整目标区间音频,属于之前确认的离线口径;TextOP 使用原始录音的因果 ASR,不使用真值指令文本。少样本选择跳过正式训练的大规模交互配比,不改变原生构造算法。 +- +-视频覆盖整个 session,从开场连续推理到收尾。只在开头用一次真值历史初始化,随后使用预测历史,没有在第二条指令到达时重置姿态。TextOP 以原生 8 帧 primitive 为单位,可能略少于最后一个不完整 primitive;OMG 保留整段 30 Hz 时间线。前缀真值初始化单独标注,并从连续推理误差中排除。左右分别为预测和真值,视频只是关节姿态播放,不包含物理控制和实际物体交互。 +- +-视频统一为 H.264 + MP3 的 MP4,使用 `_playable.mp4` 新文件名。主入口为 `/root/autodl-tmp/runs/hfd-overfit-videos-jointfixed/README.md`。结果指标和音频验证写入 `duplex_task_summary.json`;窗口指标使用真值历史,连续整段指标使用预测历史,二者不能混为模型完全学会全双工任务。 +- +-## 迁移后复现 +- +-先按根目录 README 准备两个环境、官方权重、外部机器人 assets 与 HFD 原始 tar 分片路径,再更新 YAML 中绝对路径。依次执行(工作目录是代码文件夹): +- +-```bash +-/root/autodl-tmp/envs/omg/bin/python prepare_duplex_task.py --output /root/autodl-tmp/runs/hfd-duplex-task-overfit/sessions +-/root/autodl-tmp/envs/omg/bin/python transcribe_sessions.py /root/autodl-tmp/runs/hfd-duplex-task-overfit/sessions/train-002203-seed42.npz --device cuda:0 --language auto --output-dir /root/autodl-tmp/runs/hfd-duplex-task-overfit/asr +-/root/autodl-tmp/envs/textop/bin/python train.py --config configs/history/textop_duplex_task.yaml +-/root/autodl-tmp/envs/omg/bin/python train.py --config configs/history/omg_duplex_task.yaml +-/root/autodl-tmp/envs/textop/bin/python infer_overfit.py --run /root/autodl-tmp/runs/hfd-duplex-task-overfit/textop --continuous-only --task-seconds 0 14.8 --textop-guidance-scale 1 +-/root/autodl-tmp/envs/omg/bin/python infer_overfit.py --run /root/autodl-tmp/runs/hfd-duplex-task-overfit/omg --continuous-only --task-seconds 0 14.8 +-``` +- +-渲染每个 `RUN/inference`: +- +-```bash +-/root/autodl-tmp/envs/omg/bin/python render_overfit.py --inference RUN/inference --assets-root /root/autodl-tmp/datasets/eval_textop_omg/OMG_code_20260927/OMG/assets --gpu 0 +-``` +- +-选择脚本默认构建 512 条原始 source 的候选池,筛选指定两动作的原生 interruption session,保存 `selection.json` 和原生 metadata。输出目录必须不存在,防止混用旧样本和检查点。 +- +-## 清理检查点 +- +-`cleanup_checkpoints.py` 只清理 OMG 同一训练步、每一个 state_dict 张量完全相同的重复保存。正式初始化权重、唯一训练结果、TextOP 所需 VAE 都保留。先检查,然后执行: +- +-```bash +-/root/autodl-tmp/envs/omg/bin/python cleanup_checkpoints.py RUN +-/root/autodl-tmp/envs/omg/bin/python cleanup_checkpoints.py RUN --apply +-``` +- +-每个 run 保留 `last.ckpt`,删除等价 `pipeline.ckpt` 和 `checkpoints/*.ckpt`,记录 `checkpoint_cleanup.json` 并更新训练检查报告。不同参数或不同 global_step 会拒绝删除。可复现代码目录不包含这些大文件。 +- +-## 本次检查结果 +- +-两模型的 1600 次优化更新、参数变化、有限 loss/梯度和严格检查点重新加载均通过。23 项回归测试通过。 +- +-| 模型 | 训练窗口初始化关节 RMSE | 训练窗口训练后关节 RMSE | 整段连续关节 RMSE | 整段 root RMSE | +-| --- | --- | --- | --- | --- | +-| TextOP | 0.03169 rad | 0.00516 rad | 0.32428 rad | 0.02954 m | +-| OMG | 0.30846 rad | 0.10780 rad | 0.10088 rad | 0.01108 m | +- +-TextOP 的短窗口明显拟合,但整段预测仍未跟随两项动作。OMG 在整段中能看到接取和弯腰动作,但仍有姿态误差。这符合训练链路测试的用途,不作为两个 baseline 充分拟合或正式评测的结论。TextOP 的指标只计其建模的 23 关节,OMG 计 29 关节,不能直接据此排名。视频初始化分别占 2 帧 / 13 帧,已从整段指标中排除。 +- +-视频全部帧和音轨均完整解码通过,配音在编码前与原生 session 对应音频切片逐样本完全相等。两条视频音轨均非静音。TextOP 为 738 帧、14.76 s;OMG 为 444 帧、14.8 s。 +- +-本次共移除三个 OMG run 中的 9 份等价重复检查点,释放 9,535,716,429 bytes(约 8.88 GiB)。每个 run 的唯一最终 `last.ckpt` 都保留。清理记录汇总位于 `/root/autodl-tmp/runs/diagnostics/checkpoint_cleanup_summary.json`。 +- +-本页旧 TextOP 数值和视频使用 guidance=1。当前默认推理已补齐原生 guidance=5;复现旧结果须显式指定 `--textop-guidance-scale 1`。当前窗口输入与新检查见 [HFD_WINDOW_PROTOCOL.md](HFD_WINDOW_PROTOCOL.md)。 +diff --git a/docs/history/HFD_HUMAN_BASELINE_INPUT_PROTOCOL_RESEARCH_20261017.md b/docs/history/HFD_HUMAN_BASELINE_INPUT_PROTOCOL_RESEARCH_20261017.md +deleted file mode 100644 +index 4c560a7..0000000 +--- a/docs/history/HFD_HUMAN_BASELINE_INPUT_PROTOCOL_RESEARCH_20261017.md ++++ /dev/null +@@ -1,318 +0,0 @@ +-> 历史调研快照,归档于2026-10-21,包含2026-10-17起调研及后续历史补充。当前约定见[最新输入协议](../HFD_HUMAN_BASELINE_INPUT_PROTOCOL.md),以下旧默认与待实现状态不用于当前训练。 +- +-# HFD 人体 baseline:模型结构、因果输入与训练约定 +- +-当前输入修复:新训练配置使用 `clock_policy: microphone` 和 `text_conditioning: asr_to_english`,保留 `causal_asr_context`。按 `r_k <= t_motion+0.2s` 选择已释放 PCM/ASR,中文 ASR 经冻结翻译后交给原生文本编码器;视频完整保留麦克风开头。无新字段的历史 run 仍保留旧行为。详见 [中文语义与时钟修复](../SEMANTICS_CLOCK_FIX.md)。 +- +-调研日期:2026-10-17(Asia/Taipei)。状态:**代码与论文调研完成;以下新模型配方是待实现、待验证的建议,不是已经跑通的训练链路。** +- +-2026-10-18 实现进展:GenMo 的单条成对 session、原生训练桥接、连续推理、GMR 与视频链路已实现;短/长历史实验及限制见 [GENMO_HFD_OVERFIT.md](../GENMO_HFD_OVERFIT.md)。下文保留调研时的原始建议,不代表全部任务已验收。 +- +-2026-10-18 用户确认:五个 baseline 默认文本策略统一为 **`causal_asr_context`**;OMG/TextOp 已完成同 session 过拟合对照。此决定覆盖下文保留的历史建议。当前共同规则见 [五模型文本条件约定](../BASELINE_TEXT_CONDITION.md)。旧训练权重和结果仍按其原始 manifest 标识。 +- +-## 1. 推荐结论 +- +-先冻结同一份全双工 session 计划,生成经过验收的人体动作时间轴;再做各模型的表示转换。**统一信息可用时间,不强求模型接收相同形状的张量。** 0.6 秒是 HFD 每次新增的音频窗口,不是所有 baseline 的语义记忆长度,也不是动作预测长度。 +- +-第一版建议保留各模型的文本编码器,以与现有 TextOP/OMG 一致的因果 ASR 提供语义;用当前已知动作提供连续性。原生支持掩码条件的模型沿用该接口;没有该接口的模型必须明确标注为适配版本。不要把三者改造成同一个 Qwen 音频前端模型后仍称作原生 baseline。 +- +-| 模型 | 核实到的原生结构与条件 | 建议的 HFD 输入 | 适配性质 | +-| --- | --- | --- | --- | +-| 当前 HFD | Qwen 流式语音上下文;动作端有状态/历史和 RTC | 每次 0.6s PCM、0.4s 更新,session 内积累语义历史 | 参照系统;不等于独立短窗分类 | +-| 当前 TextOP | 文本条件的动作 primitive 生成 | 最近 30s 完整因果 ASR+原生 2 帧动作历史 | 本地已实现;50Hz,8 帧未来 | +-| 当前 OMG | 文本/声学条件的动作生成 | 同一因果 ASR+过去 2s 已释放 PCM+原生 10 帧历史 | 本地已实现;30Hz,60 帧未来 | +-| GenMo / GEM-SMPL | Transformer 扩散;T5 文本、逐帧音频/音乐/视觉条件;有 3D motion overwrite 接口 | 主实验:ASR 文本+已知动作前缀,视觉/音乐/语音分支标为缺失;另做 ASR+过去音频实验 | 有原生训练入口,但前缀训练分布需适配验证 | +-| HY-Motion 1.0 | 双流/单流 MMDiT,flow matching;Qwen3 token 特征+CLIP-L 全局特征 | ASR 文本经两个原生编码器;流式版另加显式动作前缀条件 | 没有公开的原生历史动作入口;流式版是方法扩展 | +-| Kimodo | root→body 两阶段 Transformer 扩散;LLM2Vec 文本+observed motion+mask | ASR 文本+历史姿态约束及 mask;未来无真值约束 | 原生约束接口可用;连续前缀微调和训练循环需实现 | +- +-建议先实现 **Kimodo-SMPLX 的输入转换与短样本闭环**,其次 GEM,最后 HY-Motion 的流式扩展。Kimodo 的 SMPLX 路线转换最直接,但官方提示该版本有 retargeting artifacts;这里只把它作为接入起点,不预先认定质量最好。若后续采用 SOMA checkpoint,需要单独验证人体到 SOMA 的转换,不能把当前 SMPL-X 数组换名当 SOMA。 +- +-## 2. 本次依据与版本 +- +-用户引用的两个线程已从本机保存记录核对: +- +-- `codex://threads/01a13a26-2519-7861-a188-c7adaa9667c1`:人体源动作交接与下载。 +-- `codex://threads/01a101b8-f059-72d1-90b4-9749bf524531`:OMG/TextOP 训练、third-party、I/O。 +- +-本地现行约定以 [SESSION_FORMAT.md](../../SESSION_FORMAT.md)、[README.md](../../README.md)、[SMPLX_HANDOFF.md](/root/autodl-fs/pangxinlei/hfd_data/SMPLX_HANDOFF.md) 为准,不从历史实验文档恢复旧策略。 +- +-官方源码通过 AutoDL 下载到 `/root/autodl-tmp/downloads/hfd-human-research-20261017/`,跳过 LFS 内容和子模块;未安装环境、下载模型权重或启动训练。 +- +-| 上游 | 本次固定 commit | 论文/官方说明 | +-| --- | --- | --- | +-| [NVlabs/GENMO](https://github.com/NVlabs/GENMO) | `16bebf402d8893184249ee206d957b8248cd8310` | [GENMO,ICCV 2025](https://arxiv.org/abs/2505.01425);仓库已改称 GEM | +-| [Tencent-Hunyuan/HY-Motion-1.0](https://github.com/Tencent-Hunyuan/HY-Motion-1.0) | `4e426f5a1021cbcf7f375458c37b840ee7225229` | [HY-Motion 1.0](https://arxiv.org/abs/2512.23464) | +-| [nv-tlabs/kimodo](https://github.com/nv-tlabs/kimodo) | `e1c219c99b4b94ba48be1fb6a1cbc64ecee4c954` | [Kimodo](https://arxiv.org/abs/2603.15546)、[官方文档](https://research.nvidia.com/labs/sil/projects/kimodo/docs/) | +- +-逐文件 SHA256 和源码定位另见 [调研来源清单](../HFD_HUMAN_BASELINE_SOURCES.json)。本次是静态核对,不声称已验证 checkpoint 兼容性、显存或推理速度。 +- +-## 3. 先构造正确的人体 session +- +-### 3.1 当前 human 包的边界 +- +-交接包提供 142,220 条成对 SMPL-X 源动作,而非已经验证的完整全双工人体训练集。本地看到 128 个 `seed_smplx/*.tar`,本次未重新做全量分片哈希审计。 +- +-按动作身份连接:本机实际读取的 `metadata/seed_g1_smplx_pairs.parquet` 使用 `motion_id` 列,对应 G1 metadata 的 `filename`;人体路径直接取表中的 `smplx_path`。保留原 split、八音色、背景、Stop、backchannel、hesitation、pause 等音频原料。禁止按行号、G1 帧下标或文件名字串猜对应关系。 +- +-人体源含 `trans[T,3]`、`root_orient[T,3]`、`pose_body[T,63]`、手/脸、`betas[16]` 和 `sample_times`。`sample_times` 是时间依据;源帧率并非严格 30Hz。`trans` 是 AMASS 模型平移,不能直接解释成世界 pelvis 位置。 +- +-### 3.2 两种身体共用计划,各自合成几何 +- +-应拆成 `SessionPlan → render_g1(plan) / render_human(plan)`。计划至少固定:split、motion/source ID、每段源时间映射、1–4 轮顺序、音色、背景偏移与增益、语音事件、动作开始/截断/保持/恢复、idle/listen 素材、过渡长度,以及所有随机选择。仅同 seed 或同 source index 不足以保证两版成对。 +- +-人体 renderer 采用共同 50Hz session 时间轴;根平移按秒插值,旋转用 SO(3)/SLERP。根据当前人体末位置和朝向拼接下一段,使用一致体型;不能复制机器人根高度、关节限制或保持姿态数组。idle/listen 和 Stop 过渡也必须重新构造人体几何。 +- +-验收后再把**已合成的整条人体 session**转到模型的 30Hz 时间网格。派生输入的最后一个历史采样点不得通过插值使用决策边界后的 GT 帧:在已知前缀上单独插值/边界保持;未来标签可按完整 GT 重采样。记录两者的取样规则。 +- +-### 3.3 third-party 不能自动宣称已有人体版 +- +-当前 portable 已接入 `hfd_third_party_multi_v5`,但其读取器加载的是生产包的 `[T,36] states`,即 G1 动作。新 SMPL-X 源动作不自动补出这些生产 session 的人体几何。必须恢复其完整源片段/时间映射,或从冻结计划同时渲染两版;不能从 G1 36D 直接构造 SMPL-X 真值。 +- +-若暂时无法恢复,应在所有参加该轮对比的模型上使用共同的、已验证的人体 session 子集,注明不含该部分 third-party。不要用新模型的较小数据集与既有 OMG 训练结果直接归因模型优劣。旁人说话的 PCM 应保留;数据层不能用旁人 GT 标签替模型过滤输入。 +- +-## 4. 因果边界与共同数据契约 +- +-### 4.1 三个时钟分开保存 +- +-当前 portable 对第 `k` 个语义组规定: +- +-```text +-新增音频窗口:W_k = audio[6400*k : 6400*k + 9600] # 16kHz +-音频释放时刻:r_k = 0.6 + 0.4*k # 音频轴,秒 +-原生动作起点:i = 20*k + j,j=0..19 # 50Hz 动作索引 +-该组内条件:仅使用截至 r_k 释放的音频/ASR +-``` +- +-这来自 `hfd_bridge/sessions.py::SessionWindows`。`i/50`、`r_k`、实际墙钟不是可随意替换的同一时间。接入人体系时必须保留这张映射;不要直接以 `audio[:i/50]` 替代现行约定,也不要因为 `r_k` 数值大于动作名义时间就擅自移位动作监督。原始语音与动作内容的约 0.2 秒偏移,也不能当作这里的释放延迟使用。 +- +-如果另建“麦克风物理时间零点”的真实在线评测,需明确初始预缓冲和动作轴到墙钟的映射,所有模型共用;它与复现现行 portable 是不同的时间协议。 +- +-### 4.2 共同可用信息 +- +-在一次决策处,允许输入:已释放的混合 PCM、由这些 PCM 得到的 ASR、已知当前姿态及过去已执行姿态、固定的身体/坐标定义,以及事先指定的计算窗口长度。 +- +-不允许输入:未来音频、整句 GT 指令、最终纠正后的文本、GT turn ID、listen/motion/idle 标签、未来 Stop 时刻、未来动作长度/进度、未来轨迹/落脚点、从未来 GT 动作算出的历史特征。GT 标签可用于监督、分层抽样和评估,不可变成推理时不可得的条件。 +- +-噪声化的未来动作是扩散/flow matching 的训练变量;这属于目标学习过程。问题在于把干净未来特征送入条件,或在评估时提供它们。未来生成 token 相互注意本身不违反在线输入因果性,因此不应把所有模型强制改成逐帧 causal attention。 +- +-### 4.3 建议缓存字段 +- +-```text +-session_id, split, plan_hash, recipe_hash, pcm_hash +-decision_index_50hz, semantic_group, audio_release_sample +-motion_sample_times, known_motion_end, native_fps +-asr_cache_hash, asr_policy, asr_text, asr_audio_end_sample +-body_model_id, shape_id, coordinate_transform, representation_id +-history_features, history_feature_mask, valid_history_mask +-future_target_features, valid_future_mask +-``` +- +-训练专用的目标和控制标签与输入字段隔开;ASR 缓存绑定 PCM 和识别配置,不绑定某种模型的姿态。G1/human 共用 PCM 后可以复用 ASR;人体特征缓存必须独立。 +- +-## 5. 文本条件、语义记忆与全双工限制 +- +-### 5.1 五个 baseline 默认采用完整历史 ASR +- +-复用现有 Whisper-tiny、`language=auto`,每个语义组仅识别截至 `r_k` 的最多 30 秒历史混合音频。原始 sidecar 保留中英文和背景,不翻译、不从 metadata 修正;当前模型条件另经冻结中文→英文转换。离线预计算必须逐释放点截断后识别;先识别完整 session 再按词时间戳切分会利用未来声学上下文。 +- +-用户已确认五个 baseline 统一采用 `causal_asr_context`:每次使用截至当前释放点、最近最多 30 秒音频的完整识别结果;非空立即更新,空结果沿用上一份非空文本,session 结束清空。不要把历次重识别结果不断拼接,也不要用未来识别结果回填过去。最后一句允许是不完整的当前假设,不再等待连续两次一致。 +- +-五个模型共用同一字符串,再各自通过原生 tokenizer/encoder:OMG 的 T5、TextOp 的 CLIP、GenMo 的 T5-3B、HY-Motion 的 Qwen3/CLIP、Kimodo 的 LLM2Vec。embedding 不共享,长度预算按原生配置,缓存并报告截断;不得静默换成摘要、GT 指令或“当前任务”。共同实现为 `hfd_bridge.textop_condition.caption_contexts`。 +- +-### 5.2 对照结果与历史兼容 +- +-OMG/TextOp 的相同 session 对照见 [完整历史 ASR 过拟合测试](../ASR_CONTEXT_OVERFIT.md)。TextOp 有改善但仍不稳定,OMG 没有整体改善。两种策略同时改变了历史覆盖范围和两次确认延迟,不能只归因于上下文长度;保留多个句子也不保证理解否定、指令替换或第三方话语。 +- +-`stable_latest_sentence` 和 `rolling_transcript` 只供显式历史复现/消融使用,不能混称新默认。已保存的旧权重、结果和缓存不改写。GenMo 新缓存记录策略;当前 runner 复用缓存时必须匹配请求策略,不匹配则要求重新生成。HY-Motion 和 Kimodo 的 HFD 训练适配仍待实现,但文本入口必须遵循此默认约定。 +- +-不建议第一版新增 GT 驱动的“当前任务选择器”,也不把所有 PCM 统一变成 Qwen hidden 后接新头。需要意图/说话人解析时,另立适配系统实验,单独报告其训练数据、错误和延迟。 +- +-## 6. GenMo / GEM:原生结构决定如何接入 +- +-2026-10-18:ASR+过去 2s 原生 PCM 分支已实现并完成小样本匹配实验,当前未优于 ASR-only;实际因果边界、维度、BN 诊断和结果见 [GENMO_PAST_AUDIO_TEST.md](../GENMO_PAST_AUDIO_TEST.md)。 +- +-### 6.1 已核实的能力 +- +-本次针对官方 **GEM-SMPL**,不是视频生成模型 Mochi。原生 `gem_smpl` 配方有 regression/diffusion 双模式,16 层、1024 隐藏宽度的 RoPE Transformer,文本通过 T5-3B 的 1024D token 特征做 cross attention。检查点和实际配置优先于论文中的概念表示。 +- +-公开编码器的动作输入/预测核心是 151D:身体 21×6D 旋转、10D betas、camera-view 根旋转 6D、gravity-view 根旋转 6D、根局部位移 3D。末三维虽叫 velocity,代码计算的是相邻帧位移,**没有乘 FPS**;不能直接套米/秒速度。见 [原生 encoder](https://github.com/NVlabs/GENMO/blob/16bebf402d8893184249ee206d957b8248cd8310/gem/network/endecoder.py#L122)、[位移计算](https://github.com/NVlabs/GENMO/blob/16bebf402d8893184249ee206d957b8248cd8310/gem/utils/motion_utils.py#L25)。 +- +-PCM 分支是多层卷积,输出逐帧 128D 特征,再投影进动作条件;BEAT2 loader 把音频重采样至 **18kHz**、动作 30Hz。它是逐帧对齐语音条件,不能把 HFD 16kHz 波形原样塞入并假定时长不变,也不能把 music 的 35D 特征当语音语义。见 [音频编码](https://github.com/NVlabs/GENMO/blob/16bebf402d8893184249ee206d957b8248cd8310/gem/gem.py#L175)、[BEAT2 loader](https://github.com/NVlabs/GENMO/blob/16bebf402d8893184249ee206d957b8248cd8310/gem/datasets/beat2/beat2.py)。 +- +-denoiser 有 `observed_motion_3d`、`motion_mask_3d`,可把已知动作写入 noisy motion。但这不等于已具备 HFD 的 KV/RTC。默认 `dual_mode` 的 `in_attr` 没有启用 observed-motion embedder;代码有相关分支引用,也不能仅凭该引用就认定实际 checkpoint 含对应参数。见 [overwrite 接口](https://github.com/NVlabs/GENMO/blob/16bebf402d8893184249ee206d957b8248cd8310/gem/network/gem_denoiser.py#L214)、[默认条件配置](https://github.com/NVlabs/GENMO/blob/16bebf402d8893184249ee206d957b8248cd8310/configs/pipeline/dual_mode.yaml)。 +- +-### 6.2 推荐训练构造 +- +-主实验命名 `GEM-ASR-prefix`。30Hz 下建议起始设定为 12 帧已知历史+60 帧未来;12 帧包含当前状态,所有 future 条件只使用当前决策的文本。未使用的视觉、2D、camera-motion、music、speech 条件必须通过原生 availability mask 标缺失,不从 GT 人体投影出“视觉输入”。保留固定虚拟相机/重力坐标以完成表示转换,不根据未来 GT 轨迹设置相机。 +- +-历史经原生 encoder 转成 151D,送入现成 overwrite 接口;未知未来 observed 值清零、mask=0。首版不为缺失的 observed-motion embedder 虚构权重,也不自动扩展 `in_attr`。先严格加载原始 checkpoint,再检查前缀分支和梯度;如必须增加 mask embedding,明确命名为额外结构扩展。 +- +-保留原生训练入口和双模式执行关系。其 `forward_train(mode='diffusion')` 要求 `regression_outputs`,不能直接删除 regression 调用而声称仍用原训练流程。HFD 有完整 3D 监督,无需伪造 2D-only 样本或启用无来源的相机重投影监督。采用原生噪声过程与 3D loss;将已知前缀/补齐帧从未来预测指标中排除,并审计每个损失的有效 mask,避免已知帧复制降低表观 loss。见 [训练路径](https://github.com/NVlabs/GENMO/blob/16bebf402d8893184249ee206d957b8248cd8310/gem/network/gem_diffusion.py#L73)。 +- +-### 6.3 语音增强实验 +- +-命名 `GEM-ASR-past-audio`,与主实验分开。只截过去 2 秒已释放 PCM,再重采样至 18kHz;编码前就截断,防止卷积跨边界。将过去音频的逐帧特征放在窗口的历史位置,未知未来位置设 availability=0,不能把未来动作对应的真实音频给模型。 +- +-如果仅保留 0.4 秒动作历史而希望使用整段 2 秒音频,现有逐帧对齐接口需要额外设计。第一版增强实验建议使用 **60 帧过去上下文+60 帧未来**,已知动作范围同步扩为两秒;另做相同两秒动作历史但关闭音频的对照。否则不能把收益单独归因音频。音频有效长度、卷积输出帧数、填充边界必须用真实 forward 验证,不能假设输出恰好逐点对齐。 +- +-## 7. HY-Motion:文本条件可以复用,流式历史需要扩展 +- +-### 7.1 原生能力与公开程度 +- +-原生动作表示是 SMPL-H 身体 22 关节的 **201D**:根平移 3、根旋转 6、21 个局部旋转 126、22 个局部位置 66。它不等于 HumanML3D 263D。网络结合双流/单流 Transformer,语义条件为 Qwen3-8B token 特征和 CLIP-L 全局向量,flow matching 预测速度。论文中的局部 attention 是双向局部性,不是在线语义记忆。见 [论文模型章节](https://arxiv.org/html/2512.23464v1#S3)、[文本编码器](https://github.com/Tencent-Hunyuan/HY-Motion-1.0/blob/4e426f5a1021cbcf7f375458c37b840ee7225229/hymotion/network/text_encoders/text_encoder.py)。 +- +-核查的 `generate()` 接口接收文本、seed、duration;`hidden_state_dict` 是预编码文本特征,并非历史动作状态。`x_mask_temporal` 表示有效序列位置,也不是 known-history mask。源码有 `mask_mode='causal'` 分支同样不等于已训练的流式预测器。公开仓库以推理为主,本次未发现可直接套用的完整官方训练循环。见 [生成接口](https://github.com/Tencent-Hunyuan/HY-Motion-1.0/blob/4e426f5a1021cbcf7f375458c37b840ee7225229/hymotion/pipeline/motion_diffusion.py#L477)、[forward](https://github.com/Tencent-Hunyuan/HY-Motion-1.0/blob/4e426f5a1021cbcf7f375458c37b840ee7225229/hymotion/network/hymotion_mmdit.py#L409)。 +- +-### 7.2 两种结果必须分开命名 +- +-`HY-native-window`:每次使用当前 ASR 文本、固定两秒计算窗口做文本到动作。保持原生网络,用于判断文本/动作表示是否接通;它没有历史姿态条件,不能冒充连续全双工 baseline,也不能靠 GT 对齐每个预测片段掩盖跳变。 +- +-`HY-ASR-prefix`:推荐的流式适配。采用 12 帧历史+60 帧未来、30Hz;增加显式 known-mask 和历史条件通路,例如零初始化的 mask 投影,保留原有 201D 输入投影权重。未来真值只用于 flow target。新增参数及加载差异必须写入 manifest;明确这是我们实现的扩展。 +- +-一种可实现的训练约定是:已知前缀始终为干净历史;未知部分按 `x_s=(1-s)ε+s*x_GT` 插值;网络得到 known mask 和当前 ASR,预测未知部分的 `x_GT-ε`,只在未知有效帧上计算 flow loss。采样每个 ODE 子步锁定历史,未知位置从噪声积分,结束后只执行未来段。训练/采样必须采用同一种前缀参数化。它改变了原始无前缀训练分布,需通过微调和闭环验证,不能只改推理器硬覆盖就宣称完成适配。 +- +-禁用 duration estimation 和 prompt rewrite,计算时域预先固定;不能用 GT 总动作长度或下一事件时间决定输出长度。保留双编码器,不把 Qwen 单独替换成 CLIP。可先冻结文本编码器、只训练 motion 模型及新增投影;显存/LoRA 策略在真实加载后决定。 +- +-原生解码默认包含旋转/平移平滑和按整段最低顶点落地。流式版若对整个“历史+预测”执行这些操作,会修改历史和世界根高度。需锁定已执行前缀、固定 session 地面/世界原点;后处理只作用于未执行计划,另外报告开启/关闭结果。利用自己预测的未来做后处理可以是因果规划,但不能回写已执行帧或利用 GT 未来。 +- +-## 8. Kimodo:优先复用原生约束机制 +- +-### 8.1 结构与版本选择 +- +-Kimodo 先预测 root,再把 root 转为局部表示供 body denoiser 使用。`motion_mask_mode='concat'` 时,已知值覆盖 noisy motion,特征 mask 与动作拼接进入网络;这是真正的约束接口。见 [两阶段 denoiser](https://github.com/nv-tlabs/kimodo/blob/e1c219c99b4b94ba48be1fb6a1cbc64ecee4c954/kimodo/model/twostage_denoiser.py#L73)。 +- +-文本默认是 LLM2Vec/Llama-3-8B 系列特征,不是 CLIP/T5;本次看到的配置可加载本地编码器或编码服务。训练实验应固定权重版本并本地缓存已释放文本的 embedding,服务不可静默换模型。 +- +-官方分别提供 SOMA、G1、SMPLX checkpoint。SOMA 内部 30 关节、公共输出可为 77 关节;SMPLX 为身体 22 关节。本批数据先对接固定的 `nvidia/Kimodo-SMPLX-RP-v1`,不使用会随版本变化的默认短名。官方说明 SMPLX 版本可能有较明显重定向伪影。见 [骨架说明](https://research.nvidia.com/labs/sil/projects/kimodo/docs/key_concepts/skeleton.html)。 +- +-源码的 AMASS→Kimodo 转换包含 pelvis offset、Y-up/Z-up 与额外朝向旋转;优先使用 `amass_arrays_to_kimodo_motion()` 并做往返检验,不能直接照搬 GEM 的坐标矩阵。源 16D betas 与 Kimodo 固定骨架体型也需核对;必要的形体标准化必须训练/验证一致,不能丢掉 betas 后声称几何完全相同。见 [官方转换](https://github.com/nv-tlabs/kimodo/blob/e1c219c99b4b94ba48be1fb6a1cbc64ecee4c954/kimodo/exports/smplx.py#L74)。 +- +-### 8.2 推荐输入与训练 +- +-命名 `Kimodo-SMPLX-ASR-prefix`。起始采用 12 帧历史+60 帧未来、30Hz;历史来自当前可知的姿态。优先按原生 full-body/末端约束格式构造历史,未来所有约束 mask=0;用 `first_heading_angle` 和可知根位置承接世界状态。不能输入未来全身关键帧、真实根路径、未来接触标签,即使原生 API 支持这些条件。 +- +-原生 multi-prompt 代码已经用上一段生成结果构造 transition constraints,说明无需另起一套条件网络;但原 UI/批处理可以看到全部 prompt,在线 wrapper 必须在每次释放后单步调用,不一次传入整条未来 prompt 时间线。原实现还会重新生成拼接重叠区,在线版不能修改已经执行的部分。见 [transition 构造](https://github.com/nv-tlabs/kimodo/blob/e1c219c99b4b94ba48be1fb6a1cbc64ecee4c954/kimodo/model/kimodo_model.py#L140)。 +- +-沿用 checkpoint 的扩散 schedule、两阶段网络和运动表示,新增 HFD 数据训练循环。论文目标是 clean-motion 分量的加权 smooth-L1 与 FK 一致性,不能用简单的统一 MSE 替换后称原生损失复现。连续历史前缀也不是原稀疏关键帧分布,需在微调中覆盖。历史姿态约束下的未约束辅助量仍可能需要损失;具体 mask 按语义/维度确定,不笼统把前缀全部当成未知目标。见 [论文训练目标](https://arxiv.org/html/2603.15546v1#S4.SS3)。 +- +-公开代码提供 denoiser、噪声过程与采样,不等于已提供完整训练器;优化器、EMA、loss 权重和约束采样需按论文/配置实现并标明复现范围。本轮不承诺两卡 batch 或训练时长。 +- +-### 8.3 平滑根与速度的未来泄漏 +- +-Kimodo 原生表示包含平滑根、heading、局部位置、全局旋转、速度和接触。SMPLX22 的该表示按源码组成是 `3+2+22*3+22*6+22*3+4=273D`,实际仍须以 checkpoint 的 `motion_rep_dim` 断言,不能套用 SOMA 的维数。见 [运动表示](https://github.com/nv-tlabs/kimodo/blob/e1c219c99b4b94ba48be1fb6a1cbc64ecee4c954/kimodo/motion_rep/reps/kimodo_motionrep.py)。 +- +-`get_smooth_root_pos()` 是整段轨迹平滑;速度是前向差分。**禁止对“GT 历史+GT 未来”整体编码后,把前 12 帧当输入。** 原生平滑甚至会影响历史段中更早的位置,单独屏蔽最后一帧速度不足以消除泄漏。见 [平滑根](https://github.com/nv-tlabs/kimodo/blob/e1c219c99b4b94ba48be1fb6a1cbc64ecee4c954/kimodo/motion_rep/smooth_root.py#L202)、[速度计算](https://github.com/nv-tlabs/kimodo/blob/e1c219c99b4b94ba48be1fb6a1cbc64ecee4c954/kimodo/motion_rep/feature_utils.py#L36)。 +- +-建议从已知前缀独立构造 full-body/end-effector constraints,而不是锁定整段 GT 编码的所有 273D 特征。若使用完整历史特征,只能在过去前缀上编码,边界速度用最近已知位移延拓;推理可以保留前次已生成计划的特征,但必须记录来源为模型预测。未来目标可以离线整体编码;输入中的平滑根与目标平滑根可能不同,需通过可观测姿态约束而非强求两者逐值相等来处理。 +- +-## 9. 新模型共同的窗口与 rollout 建议 +- +-以下是起始实验参数,不是三篇论文的原始训练配置: +- +-| 项目 | 第一版建议 | +-| --- | --- | +-| 新人体模型动作 FPS | 30Hz,严格按时间重采样 | +-| 动作历史 | 0.4s / 12 帧,包含当前状态;不足时重复首姿态并记录有效范围 | +-| 未来计划 | 2s / 60 帧;不依赖 GT 剩余动作长度 | +-| 名义条件更新/执行 | 0.4s / 12 帧后可重规划;每次仅用当时已释放条件 | +-| 语义条件 | 五者共享 `causal_asr_context`;旧稳定句仅显式历史对照 | +-| 训练覆盖 | `full_session`,包含 listen、idle、保持、过渡及事件前后 | +-| 外部控制 | 第一版不添加 GT stop gate、事件识别头或手工姿态重置 | +- +-预测两秒不表示必须播放完两秒。新语义组到来后可重规划,避免 Stop 要等完整计划结束。理想零计算延迟回放和实测延迟回放分别报告;如果规划/ASR 耗时超过 0.4 秒,不能声称达到该响应频率。 +- +-训练窗口在 session 内按合法起点取样,不能跨 session 连接;未来段即使覆盖下一条指令,也只提供当前可用文本。未来事件尚不可知时,监督天然存在多解,应保留并报告这种不确定性,不以未来 ASR 消除它。可增加只监督近期执行段的消融,但它改变训练损失,不能混为原生方法。 +- +-第一阶段使用 GT 过去姿态做 teacher-forcing,未来按模型原目标训练;正式评估必须用自身实际执行结果作为下一步历史,只允许 session 初始姿态来自 GT,不能每组重新喂 GT。后续可加入 rollout/noisy-history 微调,单独标注训练配方,不把它当作已经实现的能力。 +- +-统计归一化先沿用官方 checkpoint 的对应统计以保留初始化语义;人体分布若明显失配,再设计重估和权重/输入变换的成套方案。统计只能来自 train,不能把 SMPL-X 原始轴角、G1 36D、TextOP/OMG 的统计互相套用。 +- +-## 10. 实现前后的验收与比较口径 +- +-1. **人体 session**:逐任务覆盖 1–4 轮;与 G1 版比较计划、PCM 哈希、事件边界、split。人体单独检查旋转合法、无 NaN、拼接连续、固定体型与带声音视频。当前尚未完成。 +-2. **表示往返**:已知姿态→原生特征→解码,检查左右手、根朝向、位移单位、pelvis offset、帧率与固定形体;预先记录转换误差。所有模型按共同身体 22 关节评估,手指不纳入主体指标。 +-3. **因果反事实**:固定当前 PCM/历史,仅替换释放点后的音频,当前输入必须不变;仅替换当前姿态后的 GT 动作,历史特征/文本/约束必须不变;修改未来事件元数据也不得改变当前输入。 +-4. **原生模型接通**:严格加载官方 checkpoint,核对特征维数与实际前向;检查 loss、有效 mask、有限梯度、参数更新与 checkpoint 恢复。HY 新增参数单独列出;Kimodo 新训练器说明 loss 复现范围。 +-5. **短样本闭环**:包含普通指令、Stop、backchannel、hesitation、pause、third-party 的已验证人体样例;比较 teacher-forcing 与自主 rollout,检查“听完才动”、停止延迟、被打断后保持和下一指令切换。 +-6. **正式比较**:固定 session manifest、共同数据范围、ASR 版本、初始状态和随机种子;报告源动作/事件覆盖、训练更新数及有效目标帧,不强求不同架构拥有相同 batch/epoch。固定验证集,测试集不参与调参。 +-7. **结果**:同时报告动作质量、事件响应、语义错误、边界跳变、长时漂移、ASR 稳定等待、模型规划延迟和执行缓冲。文本输入不能完全覆盖声学/说话人信息,ASR 系统与原始音频 HFD 属于不同输入管线,表格应标明。 +- +-上游预训练数据与 HFD 源动作也要核查重叠。Kimodo 当前 README 明确说明部分旧 SOMA checkpoint 存在 BONES-SEED benchmark overlap,v1.2 修正;这不自动证明 SMPLX-v1 对 HFD 测试完全独立。无可核实的预训练划分时,写明“下游 session/source 隔离,预训练重叠未排除”,不能宣称完全未见动作泛化。 +- +-本说明不改变现行 OMG/TextOP 配置。下一步需要实现的是 session 计划冻结与 human renderer,再是三个 adapter;不是直接把新源动作交给既有训练命令。 +- +-## 11. 2026-10-18:官方与社区训练脚本综合核查 +- +-本节回答“是否有可用训练脚本”,不代表已经接通 HFD 训练。核查了官方当前源码树、官方训练代码 issue、定向搜索得到的社区仓库及实际训练代码;没有穷举所有 fork。固定版本、文件哈希与小规模 probe 的记录见 `reports/human_training_scripts_20261018.json`。选取的上游文本源码快照保存在 `/root/autodl-fs/pangxinlei/downloads/hfd-training-script-research-20261018/`,未替换三个原有模型仓库。 +- +-### 11.1 结论与选择 +- +-| 模型/候选 | 实际公开内容 | 能否用于当前 HFD 逐帧监督微调 | 验证程度 | +-| --- | --- | --- | --- | +-| GenMo/GEM 官方 | Lightning/Hydra 训练器、数据模块、双模式 diffusion/regression、损失、优化器与 checkpoint | **优先复用官方框架**,补 HFD 数据/条件 bridge 和专用配置 | 源码核对;现有环境及小型原生网络反向已通过;未跑真实 HFD 官方训练步 | +-| HY-Motion 官方 | 201D 模型、CLIP/Qwen 文本编码、采样、评测 | **没有官方训练入口**,监督 flow-matching 循环仍需补 | 官方源码与作者 issue 回复核对 | +-| HY-Motion:MotionRFT | 对 Full/Lite 的 LoRA、可微采样、奖励模型与微调循环 | **可参考训练组件,但不是逐帧监督训练的直接替代品** | 静态核对训练/梯度代码;未运行其完整奖励训练 | +-| Kimodo:Calix-L 后训练项目 | 特征数据集、文本缓存、训练步、优化器、梯度累积、checkpoint、七项结构化损失 | **适合作为轻量监督训练起点**;默认 SOMA 配置和离线约束采样必须替换 | 七项损失及其原样 train_step 已对 tiny 官方 SMPLX22 两阶段网络做反向 probe | +-| Kimodo:windydauy 的 scq 分支 | 完整 G1 微调目录、DDP/AMP、EMA、AdamAtan2、七项损失;另有蒸馏工作流 | **可复用训练设施**,G1 数据适配与损失须重审;不直接采用改变层数的蒸馏版本 | 损失模块在当前 SMPLX22 表示上反向通过;完整训练循环仅静态核对 | +- +-“可复用”指可以基于现有代码补齐当前任务,**不是修改数据路径即可开训**。共同缺口仍是第 4 节的冻结 SessionPlan 与 human renderer,以及因果 ASR/姿态历史、闭环推理和第 10 节验收。 +- +-### 11.2 GenMo/GEM:官方训练框架最完整 +- +-来源:[NVlabs/GENMO](https://github.com/NVlabs/GENMO),本地与远端核对版本均为 `16bebf402d8893184249ee206d957b8248cd8310`。这里是 NVIDIA 人体动作项目,不能误用同名 Genmo/Mochi 视频生成训练器,也不把 GEM-X 的另一套网络混入当前 baseline。 +- +-主要入口: +- +-- `scripts/train.py`:实例化模型和 datamodule,调用 Lightning `Trainer.fit`,支持断点恢复、日志和验证。 +-- `gem/gem.py`:`training_step`、`configure_optimizers`,原生 regression/diffusion 的执行次序。 +-- `gem/pipeline/gem_pipeline.py`、`gem/network/gem_diffusion.py`:扩散训练及几何损失。 +-- `configs/exp/gem_smpl.yaml`:包含文本/音频生成的完整模型配方;`gem_smpl_regression` 是回归配方,不能拿它代替文生动作训练。 +- +-官方命令 `python scripts/train.py exp=gem_smpl` 是框架入口示例;原配方会读取多套原始训练/评测集,并非 HFD 命令。默认 exp 引用缺失的 mixed,需显式指定 exp。模型构造还依赖主体模型、归一化与辅助资产,CLI --help 成功不证明这些运行资产已完整准备。 +- +-接 HFD 时应保留官方模型、噪声过程和合适的 3D 损失,接入 HFD datamodule 与原生条件 availability/overwrite。不要给模型伪造未来视觉观测;世界/重力/虚拟相机变换、body shape、语音特征与文本缓存仍须按第 6 节核对。diffusion 路径依赖 regression 输出,首轮不直接裁掉 regression 调用。输入历史的速度边界和未来 mask 需通过反事实检查。 +- +-### 11.3 HY-Motion:已有社区训练,但主要是奖励微调 +- +-官方 [issue #2](https://github.com/Tencent-Hunyuan/HY-Motion-1.0/issues/2) 中,协作者于 2025-12-31 明确回复没有开源该项目训练代码的计划。当前官方版本 `4e426f5a1021cbcf7f375458c37b840ee7225229` 的源码树也未找到训练入口。README 的预训练→高质量微调→RL 是方法说明,不能当成已发布训练器。网上部分教程给出 `train.py`/`configs/finetune.yaml` 却没有对应仓库文件,不作为可执行证据。 +- +-找到的主要可用社区实现是 [Xiaofeng-Tan/MotionRFT](https://github.com/Xiaofeng-Tan/MotionRFT),固定版本 `84543efec6bf8c0d819bfa75665f5c820c35daae`: +- +-- `RFT_HY/motionrft_hy.py` 包含 `LoRALayer`、`apply_lora_to_model`、`DifferentiableFlowMatchingSamplerV2`、训练器和 checkpoint;支持 Full/Lite 配置。 +-- 它先无梯度采样到随机时间,再通过一个可微末步产生动作,使用 `loss = -combined_reward` 更新 LoRA;不是标准配对动作的 flow-matching 速度场监督。 +-- 模型采样仍使用原生 **201D**。奖励/匹配模型使用的 **135D** 是取根平移与旋转部分的另一条数据通路,不可把 Mean_135/Std_135 或 HumanML263 统计替代 HY 的 201D 模型统计。 +-- 依赖 HumanML3D 风格数据、SPM/Sentence-T5 与可选 MotionReward checkpoint。代码即使选择 MotionReward 也会构造 SPM;不能假定仅准备 HY 权重即可启动。改为 HFD ASR/全 session 还需修改数据、目标和时间条件。 +- +-建议把它作为 LoRA 注入、可微网络调用和保存恢复的参考;第一版 HFD 过拟合仍采用原生 201D 与配对 GT 的监督 flow-matching 目标,不额外引入奖励模型作为第一阶段前提。若采用 MotionRFT 的奖励目标,应独立标为后训练实验,并评估事件响应是否被全片段奖励掩盖。原生 HY 接口没有经过训练的姿态历史约束,MotionRFT 也没有直接补齐我们的因果前缀接口。 +- +-另有 [FlowHMR](https://github.com/flowhmr/flowhmr/blob/main/docs/training.md) 复用 HY MMDiT backbone,提供 `train_v2m.py` 的监督训练与 GRPO;但它是视频到动作任务,数据要求相机、bbox 和 SAM 3D Body 特征,不能将其训练器/权重原样称为 HY 文本动作微调。另查了 5 个 HY mirror/streaming 候选,未发现额外可直接用的监督训练器;具体结果见快照 `hy-other-candidates.json`。 +- +-### 11.4 Kimodo:社区监督训练确实可以复用,但配方有差别 +- +-官方当前版本 `e1c219c99b4b94ba48be1fb6a1cbc64ecee4c954` 仍主要提供推理、约束与评测。[训练代码请求 #25](https://github.com/nv-tlabs/kimodo/issues/25) 的社区回复只描述尝试微调,没有给出官方训练发行版。 +- +-**候选 A:Calix-L/Kimodo-SOMA-RP-final-showcase**,固定版本 `2ece5366f8cee2a54999763f372c19bf3e20f5e2`。 +- +-- [train/train.py](https://github.com/Calix-L/Kimodo-SOMA-RP-final-showcase/blob/main/train/train.py) 包含完整的加噪→原生 denoiser→loss→backward→optimizer→保存循环,不只是展示网页。 +-- `train/kimodo_seed_repro/{data,kimodo_io,train_utils}.py` 提供特征读取、缓存文本、加载与恢复、AdamW/AdamAtan2、masked MSE/结构化 Smooth-L1;还有 `scripts/train_with_val.py`。 +-- 2026-10-18 本地实测:使用当前 Kimodo **SMPLX22 / 273D 官方统计**、tiny 随机两阶段 denoiser、synthetic cached text,原样调用其 `train_step`,打开关键帧条件与七项结构化损失;root 和 body 两分支均得到有限、非零梯度。另单独验证其 loss 反向。 +-- 该 probe 只验证接口和基本梯度流,不加载 full-size 预训练权重,不使用真实 HFD,不验证收敛、混合精度、checkpoint 恢复或官方论文复现。 +- +-必须修改/复核的点:现有 YAML 默认 SOMA、150 帧离线裁剪;关键帧可从完整 GT 有效区间随机抽取,并包含整体编码的 smooth root,不能直接用于 HFD 因果历史。`first_heading_angle` 固定零需与规范化策略一致。缓存文本路径下,`text_dropout` 仅修改文本字符串,已有缓存向量不会因此置空,使用缓存 CFG 时必须单独修正。基本 train.py 未见 EMA 更新;若声称沿用论文配方需要补齐。FK loss 当前比较 `FK(pred)` 与 `FK(target)`,没有直接使用目标的 position feature;应与论文的 FK→目标位置定义核对,不能只因七个权重相同就称一致。 +- +-**候选 B:[windydauy/kimodo_my,scq 分支](https://github.com/windydauy/kimodo_my/tree/scq/kimodo/training)**,固定版本 `59e12423f4a313aba8339f065bba7b326e7e7db2`。 +- +-其 `kimodo/training/train.py`、`dataset.py`、`loss.py`、`optimizers.py`、`ema.py` 已形成训练工作流,包含 DDP/AMP、断点、CFG dropout、EMA、AdamAtan2 和关键帧课程;比 Calix 的最小循环具备更多训练设施。七项 loss 在当前 SMPLX22 表示上的 CPU 合成反向 probe 通过,但完整循环还未执行。 +- +-它默认是 G1 CSV/机器人配置,不能直接读取 SMPL-X;七项 loss 实现使用 **普通 L1**,与当前约定的 Smooth-L1 不完全一致。默认 loss 权重与 YAML 指定权重也不同,运行必须显式锁定。其 `distillation` 目录另有 16→8 层、100→20 步等学生训练,改变了架构/目标;基础 human baseline 不直接采用。对应官方 [PR #44](https://github.com/nv-tlabs/kimodo/pull/44) 已关闭且未合入,不能标注官方训练代码。 +- +-其他容易误判的线索:`LavaTap/qwen-kimodo-train-fork` 的 `train_text_projection.py` 训练的是 Qwen→LLM2Vec 投影,不是动作模型。Blender/ComfyUI 接入器主要是推理或 retarget。近期 [Controllable-Motion-Exaggeration](https://github.com/AHHHZ975/Controllable-Motion-Exaggeration) 论文有训练适配,但当前仓库仅 README、标记 coming soon,暂没有可拿来运行的训练器。 +- +-### 11.5 对当前任务的执行建议 +- +-1. **GenMo**:保留官方 Lightning/Hydra,写 HFD bridge 与单独训练配方;不另造通用训练框架。 +-2. **Kimodo**:按用户既定的优先顺序,先用 Calix 的已验证训练步作为参考,沿用官方两阶段网络/扩散/统计,吸收另一项目的 EMA/DDP 设施;先统一损失、缓存 CFG、初始朝向与仅过去约束。所有复用的第三方代码保留来源与许可,不覆盖官方原文件。 +-3. **HY-Motion**:复用官方网络和编码器,参考 MotionRFT 的可训练调用/LoRA,补最小监督 flow-matching 循环;不要把奖励微调当过拟合测试,不能声称用了不存在的官方训练框架。 +-4. 三者接同一个冻结的 HFD session plan,分别编码为原生表示;训练和评测独立保留 teacher-forced / predicted-history。评测统一走已有 SMPL-X→G1 链路和 human GT retarget 对照。 +- +-本轮只完成调研、源码快照与上述合成 probe,没有启动正式训练或修改三者的实际训练配方。 +diff --git a/docs/history/HFD_WINDOW_PROTOCOL.md b/docs/history/HFD_WINDOW_PROTOCOL.md +deleted file mode 100644 +index 61916a7..0000000 +--- a/docs/history/HFD_WINDOW_PROTOCOL.md ++++ /dev/null +@@ -1,68 +0,0 @@ +-> 历史实验记录;当前入口见 [README](../../README.md)。旧运行目录已移动到 `/root/autodl-tmp/hfd-runs-archive-20261007/`。 +- +-# HFD 当前音频窗口口径与 TextOP 连续推理排查 +- +-旧 `omg_duplex_session_playable.mp4` 使用此前确认的 `offline_target_aligned`:动作目标未来 2 秒与这 2 秒对应音频一起输入模型。它可用于离线训练链路检查,不能据此宣称实时全双工能力;连续关节 RMSE 为 0.10088 rad,并非完美拟合。旧结果保留为离线对照。 +- +-## 新 OMG 输入时序 +- +-`data.audio_policy: hfd_current_window`。HFD 原始动作时钟为 50 Hz,每 20 个动作点使用一个 0.6 秒音频窗口;窗口每 0.4 秒更新: +- +-- 原始动作帧 `u` 的语义组 `k = u // 20`。 +-- 当前 PCM 范围 `[0.4*k, 0.4*k+0.6)` 秒,16 kHz 时为 `[6400*k, 6400*k+9600)`。 +-- 音频可用截止为 `0.6+0.4*k`。这沿用 HFD 释放时钟,与运动/音频内容偏移 0.2 秒不同,也不表示零延迟在线系统。 +-- OMG 保留原生 30 Hz 动作格式;状态帧 `t` 对应组 `k=floor(t/12)`。训练在每个可用 30 Hz 动作点取样,同组复用同一窗口。 +- +-OMG 当前音频接口需要每个预测位置一个 35D 声学特征,不接收原始 PCM,也没有 Qwen 的语言语义 KV 缓存。转换器只对当前 9600 个样本调用原生 `current35` 提取器,再将得到的特征序列线性插值到原生 60 个条件位置。**这 60 个位置都编码同一个当前窗口,不能解释为读取了未来 2 秒音频。** 包括峰值归一化在内的全部特征运算仅访问该窗口。没有额外的语义缓存前向,原生声学 embedding 和扩散采样仍然执行。 +- +-训练仍用原生 Lightning、125D 动作 codec、60 帧未来目标和 10 帧状态历史。未来动作是真值监督,未来音频不是输入。修改当前可用窗口以外的任何 PCM 不会改变输入;测试与数据合同检查覆盖此约束。 +- +-## 连续推理 +- +-只在开头使用一次真值历史。每轮用当前 0.6 秒音频窗口预测原生 60 帧(2 秒),执行前 12 帧(0.4 秒),再用已执行的预测姿态更新历史,读取新窗口并重新规划。未执行的预测丢弃,不能拿未执行的未来预测初始化下一轮。 +- +-不能把一整段 session 的各个未来音频窗口一次性传给 `generate(num_frames=整段)`:这会让当前规划提前接触之后才可用的条件。新推理代码逐轮调用原生生成器,并在报告中保存每轮 PCM 起止位置、状态帧和执行帧数。 +- +-这是沿用 HFD 音频释放规则的 OMG 适配,保留 OMG 原生预测长度;不是宣称 OMG 拥有 HFD 的语义记忆。当前窗口可能只包含长指令的一部分,缺少 KV 历史确实可能降低表现。音频在正式训练中也保留中文和背景音,不根据真值指令/任务边界提取音频。 +- +-## TextOP 为什么几乎不动 +- +-旧小样本配置 `stages: [1600]` 只进入原生 stage 0。虽然 `use_rollout: true`,`DARManager.should_rollout()` 在 stage 0 始终返回 False,所以训练全程使用每个 primitive 的真实历史。整段视频则只在开头给一次真实历史,随后反馈模型预测,训练和这种验证存在历史分布差异。 +- +-旧 TextOP 训练窗口关节 RMSE 为 0.00516 rad,整段连续 RMSE 为 0.32428 rad。两者不矛盾:从真实动作姿态继续生成 8 帧比从静止姿态独立启动并持续任务容易。文本在 listen/执行/收尾阶段也可能保持相同,而 TextOP 没有额外的真值控制标签或任务进度输入,容易停留在接近静止的历史状态。这个现象不能仅凭一条 loss 断定是映射问题;映射沿用已独立验证的关节名称转换。 +- +-本次保留原生训练循环与损失,用 `stages: [400,800,400]` 进入原生反馈预测历史的训练阶段。原生正式训练模板本来就有三阶段;此次是修正小样本实验只运行首阶段的局限。新配置为 `configs/history/textop_duplex_causal.yaml`,新 OMG 配置为 `configs/history/omg_duplex_causal.yaml`。两者均从官方初始化重训,没有复用旧离线 OMG 权重。 +- +-## 复现 +- +-已有原生 `train-002203-seed42` 与完整因果 ASR 后,先更新配置中的环境、session、权重和输出路径,再执行: +- +-```bash +-/root/autodl-tmp/envs/textop/bin/python train.py --config configs/history/textop_duplex_causal.yaml +-/root/autodl-tmp/envs/omg/bin/python train.py --config configs/history/omg_duplex_causal.yaml +-/root/autodl-tmp/envs/textop/bin/python infer_overfit.py --run /root/autodl-tmp/runs/hfd-duplex-causal/textop --continuous-only --task-seconds 0 14.8 +-/root/autodl-tmp/envs/omg/bin/python infer_overfit.py --run /root/autodl-tmp/runs/hfd-duplex-causal/omg --continuous-only --task-seconds 0 14.8 +-``` +- +-用原有 `render_overfit.py` 渲染两目录的 inference,默认 H.264/MP3 MP4。本页对应的旧 OMG 实验使用 `hfd_current_window`;后续正式模板已统一为 `hfd_past_2s`,详见 [OMG_CONTEXT_DEBUG.md](OMG_CONTEXT_DEBUG.md);旧离线 smoke/duplex 配置保留用于复现旧结果。 +- +-## 推理设置核验与本次结果 +- +-进一步对照原生 TextOP 可视化发现,旧导出使用普通条件采样(相当于 CFG guidance=1),原生 `robotmdar/config/vis_dar.yaml` 默认 guidance=5。当前 `infer_overfit.py` 已使用原生 `ClassifierFreeWrapper` 并默认 guidance=5;可用配置 `inference.textop_guidance_scale` 或 CLI `--textop-guidance-scale` 指定。旧视频/诊断的 guidance=1 设置明确保留为对照,不与新设置的误差直接混用。 +- +-同一静止历史、同一噪声种子,替换实际已释放的两条 ASR 条件:guidance=1 的输出差异约 0.00014 rad;guidance=5 约 0.00069 rad。guidance=5 下,保持文本不变而换成真实动作历史时,输出差异约 0.22 / 0.47 rad,真实历史窗口的误差约 0.00575 / 0.01026 rad。说明此次小样本拟合主要依赖动作历史续接,文本对从静止姿态启动的作用仍很弱。 +- +-三阶段加原生 guidance=5 后,TextOP 整段仍未拟合,不能将补齐这些设置描述为已解决几乎不动。旧单阶段配置是实验局限,guidance=1 是与原生默认可视化不同的设置;二者都不是静止现象的充分解释。诊断只针对本次两个任务和小样本预算,不断言整个 TextOP 架构没有能力。 +- +-OMG 当前窗口训练后窗口 RMSE 为 0.12468 rad,整段预测为 0.30954 rad;旧离线整段为 0.10088 rad。新旧同时改变音频范围、窗口取样密度与重新规划方式,这一对比不能单独量化未来音频的因果贡献。新口径不会向当前预测提供之后才释放的 PCM,而缺少语义缓存仍可能限制长指令处理。 +- +-两个 1600 步原生训练链路、严格权重重载与两个数据合同检查通过,27 项回归测试通过。新视频均保留原始音轨,音频和全部视频帧完整解码通过。OMG 推理每轮的窗口边界和执行历史记录在 `hfd_window_summary.json`。新 OMG run 的三份等价重复检查点已经清理,释放约 2.96 GiB,最终 `last.ckpt` 保留。 +- +-新视频入口: +- +-- `/root/autodl-tmp/runs/hfd-overfit-videos-jointfixed/textop_duplex_rollout_playable.mp4`:三阶段、原生 guidance=5、已释放 ASR。 +-- `/root/autodl-tmp/runs/hfd-overfit-videos-jointfixed/omg_duplex_hfd_window_playable.mp4`:当前 0.6 秒音频、0.4 秒重新规划。 +- +-诊断可复现:`python diagnose_textop.py --run RUN --frames 242 490 --guidance-scale 5`。这只比较同种子下原生 8 帧生成结果,不使用真值任务文本;真值姿态仅用于明确标注的历史对照。 +- +-最终原生 guidance=5 的 TextOP 训练窗口 RMSE 为 0.00687 rad,整段为 0.40908 rad(root 0.03739 m)。这再次表明短窗口拟合不能证明整段任务能力;新 OMG 整段为 0.30954 rad(root 0.06111 m)。两模型的关节维度不同,仍不能据此排名。 +diff --git a/docs/history/NORMAL_TASK.md b/docs/history/NORMAL_TASK.md +deleted file mode 100644 +index 90a97d6..0000000 +--- a/docs/history/NORMAL_TASK.md ++++ /dev/null +@@ -1,73 +0,0 @@ +-> 历史实验记录;当前入口见 [README](../../README.md)。旧运行目录已移动到 `/root/autodl-tmp/hfd-runs-archive-20261007/`。 +- +-# 正常单任务的小样本检查 +- +-推荐观看当前视频目录中的 `textop_complete_task_playable.mp4` 和 `omg_complete_task_playable.mp4`。这是连续播放的完整单任务,不是 `*_fitted_windows_playable.mp4` 的窗口拼接。 +- +-## 样本与任务 +- +-原始样本为 `sessions/train-000094-seed42.npz`,由 HFD 原生 session 构造器生成。元数据确认 `interaction_mode=turn-taking`、`turn_count=1`,没有打断、停止或后续任务。动作名是 `grab_walk_ff_180_002__A552_M`,即向前走并作抓取动作。英文女声为 `female_02_en`,ASR 识别为 “Character walks forward while grabbing something.”。没有额外场景物体或抓取物理模拟,播放的是机器人参考动作。 +- +-完整 session 长 17.6 秒: +- +-| 动作时间 | 内容 | +-| --- | --- | +-| 0.00–1.20 s | 初始待机;指令开始 | +-| 1.20–4.40 s | 原生听指令阶段与过渡 | +-| 4.40–16.00 s | 连续向前走并抓取,共 11.6 秒 | +-| 16.00–17.60 s | 动作结束后的待机 | +- +-原始麦克风语音约在 0.60–3.32 秒;与动作时钟相差 0.2 秒,按适配器既有时钟关系播放。数据中的听指令姿态也保留,但两个模型不接收 `listen/motion/idle` 控制标签。初始真实历史用于模型启动并在画面中标注;之后自由反馈预测。初始化画面不计入连续推理误差。 +- +-## 训练与导出 +- +-两模型从各自官方初始化权重重新训练 800 步,使用原生 DAR / Lightning 框架。配置选择整个 session 的 `[0,870)` 原始 50 Hz 帧,并保留这一连续区间内所有合法窗口,`max_windows=null`;不再均匀抽八个分散窗口。实际视频结尾根据模型帧网格取整到约 17.3–17.4 秒,只省略末尾少量待机。 +- +-TextOP 使用这个 session 新生成的逐释放窗口 ASR;OMG 使用完整离线目标段音频。音频、原始轨迹和关节映射保持不变,没有平滑、改速度、补帧或拼接其他任务来改善观感。 +- +-`textop-single/` 和 `omg-single/` 保存有效检查点、配置、指标与推理结果;`sessions/` 保存原始单轮任务及选择依据;`asr/` 保存新生成的完整转写;`summary.json` 保存本次训练和推理指标。视频和任务说明在 `/root/autodl-tmp/runs/hfd-overfit-videos-jointfixed`。 +- +-窗口拟合和自由连续推理分别报告。800 步只是训练链路的小样本检查,不保证完整任务的自由推理已拟合;模型偏离真值时不能当作视频拼接问题。右侧真值用来检查正常任务的实际动作连续性。 +- +-## 之前的视频为什么跳跃 +- +-旧 `textop_fitted_windows_playable.mp4` 拼接八段,每段仅 0.64 秒,总长 5.12 秒。八段中约 82% 的帧标为 motion、约 18% 为 idle,没有 listen 帧。但采样覆盖很偏:第一段是初始待机,其余几乎来自另一条长 session,没有覆盖原先向前走抓取任务的完整执行。训练窗口从不同时间点直接连接,必然出现时间与姿态跳变,不能当作完整任务视频。 +- +-新样本选择、训练和导出可以复现: +- +-```bash +-# 所有 output 必须指向未使用的新目录;选择仍使用原生构造,仅跳过小样本不适用的大规模最终配比。 +-python prepare_normal_task.py --output /data/normal-task/sessions --motion grab_walk_ff --limit-sources 128 +-# 转写选中的 NPZ,并按新路径修改 configs/*_normal_task.yaml。 +-python transcribe_sessions.py /data/normal-task/sessions/sample.npz --device cuda:0 --output-dir /data/normal-task/asr +-python train.py --config configs/history/textop_normal_task.yaml +-python train.py --config configs/history/omg_normal_task.yaml +-python infer_overfit.py --run /data/normal-task/textop --continuous-only --task-seconds 0 17.4 --textop-guidance-scale 1 +-python infer_overfit.py --run /data/normal-task/omg --continuous-only --task-seconds 0 17.4 +-python render_overfit.py --inference /data/normal-task/textop/inference --assets-root /data/OMG/assets --gpu 0 +-python render_overfit.py --inference /data/normal-task/omg/inference --assets-root /data/OMG/assets --gpu 1 +-``` +- +-`data.frame_range` 是原始 50 Hz 的 `[start,end)` 采样边界,仅允许 smoke 模式、单个 session,用于明确小样本检查范围;正式训练仍走完整数据配置。相机跟随使用机器人身体尺寸计算距离,避免走得远时机器人越画越小;这只影响视角,不改变原始世界坐标和误差。 +- +-本次 800 步结果(rad / m): +- +-| 模型 | 训练窗口关节 RMSE:初始 → 训练后 | 完整自由推理关节 RMSE | 完整自由推理根位置 RMSE | +-| --- | --- | --- | --- | +-| TEXTOP | 0.0448 → 0.0089 | 0.3164 | 2.5040 | +-| OMG | 0.2926 → 0.2252 | 0.2522 | 0.8357 | +- +-两模型训练检查均通过,但这些连续误差说明完整任务自由推理尚未充分拟合,不能作为已成功复现任务的结论。 +- +-## 当前视频格式与 VS Code 播放 +- +-当前交付统一为 MP4 容器、H.264 视频、48 kHz 双声道 MP3 音轨;文件名以 `_playable.mp4` 结尾。根据官方文档选择 H.264、MP4 与 MP3,避开不支持的 AAC。此前 VP8 WebM 在用户当前预览中仍无法打开,因此已清理。 +- +-官方支持清单:https://github.com/microsoft/vscode/issues/180780 +-官方 Webview 说明:https://code.visualstudio.com/api/extension-guides/webview#supported-media-formats +- +-所有视频检查过完整声画解码、非静音音轨、帧数和帧率;H.264 码流与原始视频的 SHA256 一致,转换只更换音频编码和封装,没有改动作。检查记录见视频目录的 `playable_mp4_check.json`。远程服务器上的解码检查不等于本地 VS Code 实测,首次打开请使用新文件名。 +- +-后续 `render_overfit.py` 与 `remux_video_audio.py` 默认导出 H.264/MP3 MP4,无需 WebM 或额外的 `--vscode-preview` 参数。 +- +-本页旧 TextOP 数值和视频使用 guidance=1。当前默认推理已补齐原生 guidance=5;复现旧结果须显式指定 `--textop-guidance-scale 1`。当前窗口输入与新检查见 [HFD_WINDOW_PROTOCOL.md](HFD_WINDOW_PROTOCOL.md)。 +diff --git a/docs/history/OMG_CONTEXT_DEBUG.md b/docs/history/OMG_CONTEXT_DEBUG.md +deleted file mode 100644 +index 92bcc43..0000000 +--- a/docs/history/OMG_CONTEXT_DEBUG.md ++++ /dev/null +@@ -1,79 +0,0 @@ +-> 历史实验记录;当前入口见 [README](../../README.md)。旧运行目录已移动到 `/root/autodl-tmp/hfd-runs-archive-20261007/`。 +- +-# OMG 两秒历史上下文与动作启动调试 +- +-样本仍是原生 `train-002203-seed42`:接取物品尚未完成时,输入捡起地上物品的新指令。完整原生音轨、中文/英文与背景音处理策略没有改变。此样本的两条录音在视频时间 0.4–2.16 s、5.2–6.96 s,视频到 14.8 s。最后一句话结束后已有约 7.8 s,所以不能直接归因于视频结束太早。 +- +-## 输入策略 +- +-新策略 `data.audio_policy: hfd_past_2s`:以 HFD 当前组的释放截止 `hi=9600+6400*k` 为准,读取 `[max(0,hi-32000),hi)` 原生 PCM。不足两秒时左侧补零。原生 OMG 仍保持 30 Hz、125D 动作、10 帧历史、60 帧预测与 `current35` 音频特征。对完整两秒 PCM 提取原生逐帧特征,取前 60 行,不做旧 0.6 秒到 60 位置的插值。 +- +-训练和推理都采用同一策略。没有额外 KV 预填充,每次重新编码已听过的最近两秒;未来音频不输入。两秒是有限历史上下文,不保证所有长指令都完整保留,也不是完整的 ASR/语言理解编码器。开头补零和窗口内峰值归一化在训练/推理中一致。 +- +-新增测试检查:截止时刻保持 HFD 时钟、左侧补零长度、原生特征顺序、修改未释放音频不影响当前特征,以及之前的录音确实能影响两秒条件。数据合同检查覆盖动作/音频因果边界。 +- +-## 对照设计 +- +-所有实验使用同一个 native session、相同初始随机种子与一次真实历史初始化。训练时动作点和未来目标保持一致;新模型从官方 50m 初始化重新训练 1600 步,没有从旧离线模型续训。 +- +-| 实验 | 训练音频 | 推理音频 | 每轮执行 | +-| --- | --- | --- | --- | +-| A | 当前 0.6 s | 当前 0.6 s | 0.4 s | +-| B | 当前 0.6 s | 最近两秒 | 0.4 s | +-| B2 | 当前 0.6 s | 最近两秒 | 2 s | +-| C | 最近两秒 | 最近两秒 | 0.4 s | +-| C2 | 最近两秒 | 最近两秒 | 2 s | +- +-B 就是用户提出的“训练 0.6 秒,推理两秒”,保留为诊断对照。B/C 的音频时间都不超出当前 HFD 释放边界。B2/C2 检查原生两秒预测能否在后半段启动;它们两秒才更新一次规划,不能与 0.4 秒重新规划混为同一个实时反应口径。 +- +-每轮报告记录两秒预测前 0.4 秒与之后 1.6 秒相对上一已执行姿态的关节变化。不能只看执行片段而忽略计划后半段:反复重规划可能把稍后的启动动作丢弃。执行后的状态才反馈到下一轮,没有使用未执行未来姿态。 +- +-## 配置与复现 +- +-新配置 `configs/history/omg_duplex_past2s.yaml` 包含原生框架、官方初始化权重、dataset、GPU、batch size、训练步数、音频策略与 `inference.omg_execute_frames: 12`。迁移时按根目录 README 准备环境和外部权重,修改 YAML 路径后运行: +- +-```bash +-python train.py --config configs/history/omg_duplex_past2s.yaml +-python infer_overfit.py --run RUN --continuous-only --task-seconds 0 14.8 +-python infer_overfit.py --run RUN --continuous-only --task-seconds 0 14.8 --omg-execute-frames 60 --output-dir RUN/inference-execute2 +-``` +- +-复现 B:将 RUN 指向已有当前 0.6 秒训练结果,再加 `--omg-audio-policy hfd_past_2s --output-dir NEW_OUTPUT`。推理报告分别记录训练和推理策略,防止把输入不匹配的结果误认为匹配训练。 +- +-`--gpu-id` 只改变推理用的物理 GPU,不改训练配置或权重。输出目录必须不存在。渲染仍使用原有 `render_overfit.py`,交付为 H.264/MP3 MP4,保留对应原生音轨。 +- +-## 本次结果 +- +-| 实验 | 整段关节 RMSE(rad) | root RMSE(m) | +-| --- | --- | --- | +-| A:0.6 s 训练 / 0.6 s 推理 / 执行 0.4 s | 0.30954 | 0.06111 | +-| B:0.6 s 训练 / 2 s 推理 / 执行 0.4 s | 0.30978 | 0.04566 | +-| B2:0.6 s 训练 / 2 s 推理 / 执行 2 s | 0.26777 | 0.04292 | +-| C:2 s 训练 / 2 s 推理 / 执行 0.4 s | 0.25026 | 0.04641 | +-| C2:2 s 训练 / 2 s 推理 / 执行 2 s | 0.18776 | 0.03093 | +- +-新两秒上下文模型的真实历史训练窗口 RMSE 从初始化 0.31303 rad 降到 0.09066 rad。两种执行长度使用相同最终权重,不是分别训练或按较好视频选择检查点。 +- +-只把旧 0.6 秒训练模型的推理输入延长到两秒,没有解决几乎不动。匹配训练与推理上下文后,第一项动作能启动;0.4 秒重规划下第二项仍较弱。以原生两秒块执行时,两段实质动作都可见,但捡取动作晚于真值,仍未充分拟合。 +- +-计划检查表明,某些静止历史下模型将明显动作安排在预测后半段。例如 B 在 source 2.4 s 的计划:前 0.4 s 相对上一姿态变化约 0.00518 rad,后 1.6 s 约 0.42096 rad。频繁重新规划会丢掉这部分计划,是此样本启动不足的一个因素。这是当前样本的诊断,不把两秒执行视为更好的实时策略;两秒执行降低指令更新频率,正式论文必须明确执行周期与反应延迟。 +- +-新视频: +- +-- `omg_duplex_past2_exec04_playable.mp4`:训练/推理统一两秒历史,0.4 秒重规划,14.8 s。 +-- `omg_duplex_past2_exec2_playable.mp4`:相同权重、相同两秒历史,原生两秒块执行,14.8 s。 +-- `omg_duplex_past2_exec2_tail4_playable.mp4`:相同两秒块推理追加四秒静音,18.8 s。 +- +-视频都在 `/root/autodl-tmp/runs/hfd-overfit-videos-jointfixed`。左侧预测、右侧原生真值,保留原始任务录音和背景音,H.264/MP3 MP4。第三条 14.8 s 之后右侧仅保持最终姿态,不存在真实真值,画面有明确标注。 +- +-延长尾段时,原始 14.8 s 的预测与原音轨逐项完全不变;新增部分全为诊断静音,并从误差计算中排除。因此延长版仍为 0.18776 rad,而不能借延长真值待机降低指标。新增四秒主要是收尾/待机,没有补出新的任务动作。原始模型的几乎不动不能归因于没有多给四秒等待时间。 +- +-本次 1600 步原生训练、严格最终权重重载、数据合同和 29 项回归检查通过;三条新视频完整声画解码通过。新 run 三份等价重复检查点已清理,释放约 2.96 GiB,唯一 `last.ckpt` 保留。完整数字、每轮窗口边界和执行片段记录见 `omg_context_debug_summary.json`。 +- +-延长尾段只用于诊断,必须跟在完整原生 session 后;复现命令: +- +-```bash +-python infer_overfit.py --run RUN --continuous-only --task-seconds 0 14.8 --omg-execute-frames 60 --tail-seconds 4 --output-dir RUN/inference-tail4 +-``` +- +-当前训练/推理两秒上下文模板仍默认执行 12 帧(0.4 s);两秒块对照用 CLI `--omg-execute-frames 60`。窗口长度、预测长度、执行长度是三个不同参数,不能把扩大上下文和延长等待混为同一件事。 +diff --git a/docs/history/README.md b/docs/history/README.md +deleted file mode 100644 +index c6d32cb..0000000 +--- a/docs/history/README.md ++++ /dev/null +@@ -1,9 +0,0 @@ +-> 最新调试改为 full_session 动作监督;OMG 10 步采样,TextOP 连续生成仍未解决。以根目录 README 和 docs/ROLLOUT_DEBUG.md 为准。 +- +-# 历史记录 +- +-本目录和 `configs/history/`、`reports/history/` 保留过去实验的文档、配置与检查记录。当前方案以根目录 README 为准。 +- +-这些文件没有删除,也没有移出 Git 仓库。引用的旧运行路径属于当时证据;实际产物现归档到 `/root/autodl-tmp/hfd-runs-archive-20261007/`。 +- +-旧运行产物后来经用户确认永久删除;历史配置和报告仍保留,但旧产物路径不再可用。当前样本与模型保存在 `runs/hfd-final/`。 +diff --git a/docs/history/VIDEO_SAMPLES.md b/docs/history/VIDEO_SAMPLES.md +deleted file mode 100644 +index 5a47fd5..0000000 +--- a/docs/history/VIDEO_SAMPLES.md ++++ /dev/null +@@ -1,92 +0,0 @@ +-> 历史实验记录;当前入口见 [README](../../README.md)。旧运行目录已移动到 `/root/autodl-tmp/hfd-runs-archive-20261007/`。 +- +-推荐查看新的完整单任务视频 `textop_complete_task_playable.mp4`、`omg_complete_task_playable.mp4`,任务与训练说明见 [NORMAL_TASK.md](NORMAL_TASK.md)。下文是此前八窗口及事件短片的记录。 +- +-# 修正关节映射后的推理视频与任务说明 +- +-七个 MP4 均已重新导出声音,使用 H.264 画面、48 kHz 双声道 MP3 音轨;逐个完整解码,检查非零音量、音频时长和视频帧数。音轨来自原始 session 的混合录音,保留语音、背景音和静音间隔。这批视频只对播放声音施加固定增益(不削波),没有改变模型输入、预测动作或训练结果。逐文件增益、音量和哈希见 `audio_verification.json`。 +- +-除 `gt_before_after_playable.mp4` 外,左侧是模型预测,右侧是真值。真值对照视频左侧是旧的错误关节解释,右侧是正确解释;两侧均为同一段真值,没有模型预测。 +- +-## 原生 session 中的任务 +- +-任务描述依据 session 的原始动作名与事件元数据,便于理解动作。下面引用的普通任务英文为已保存 ASR 的识别结果,可能有识别误差;`Stop now.` 是原始控制事件中记录的口令。GT 描述只用于本文解释,没有作为额外模型输入。 +- +-| session | 任务与交互 | 声音 | +-| --- | --- | --- | +-| `train-000002-seed42` | 第一轮从站立进入吸烟动作(`idle_to_smoke_idle_R_001__A552`);第一任务尚未结束时,第二轮要求向前走并抓取物体(`grab_walk_ff_180_001__A549`)。新任务打断旧任务。 | 英文男声;ASR 为 “from standing to standing and smoking.”,随后 “Character walks forward while grabbing something.”;保留空调背景音。 | +-| `train-000014-seed42` | 四轮分别为转身、站立喝水后扔瓶、再次转身、向前走并抓取物体。第二轮喝水/扔瓶任务未完成时插入 `Stop now.`,然后再进行后续任务。 | 英文女声;第二轮 ASR 为 “Drinks from bottle while standing, then throws bottle to the left.”;停止口令来自原始元数据;保留空调背景音。 | +- +-原生数据支持中文和背景音,本次选中的这两个 session 恰好是英文,未把中文样本转成英文。动作名和 ASR 的左右方向可能不一致,因此描述不据 ASR 推断精确转向;真值动作以原始轨迹为准。 +- +-## 连续事件视频 +- +-以下时间范围均为原始 session 的动作时间。原生适配器的麦克风时间比动作时间多 0.2 秒;音频片段按这一关系裁切,与事件时间对齐。 +- +-| 文件 | 动作时间范围 | 关键指令在视频中的时间 | 要观察的任务 | +-| --- | --- | --- | --- | +-| `textop_interruption_playable.mp4` | 7.72–13.64 s | 第二轮开始于约 1.08 s | 吸烟任务进行中输入向前走、抓取的新任务;听到的是第二轮指令,第一轮完整指令在更早处。 | +-| `omg_interruption_playable.mp4` | 7.63–13.63 s | 第二轮开始于约 1.17 s | 同上。 | +-| `textop_emergency-stop_playable.mp4` | 81.00–86.92 s | `Stop now.` 开始于约 1.00 s | 喝水/扔瓶任务中的急停;第二轮完整任务指令位于录音 73.8–77.0 s,不包含在这一短片中。 | +-| `omg_emergency-stop_playable.mp4` | 80.83–86.83 s | `Stop now.` 开始于约 1.17 s | 同上。 | +-| `gt_before_after_playable.mp4` | 7.00–13.00 s | 第二轮开始于约 1.80 s | 同一段吸烟被新任务打断的真值,仅用于对照错误/正确的关节映射。现在也包含对应原始音频。 | +- +-指令前后短暂只有背景音是 session 原本的内容。连续推理开头使用一次真实历史,之后反馈模型预测。不是所有事件窗口都在八窗口训练子集中,因此这些片段不是完整的全双工能力评测。 +- +-## 训练窗口拼接视频 +- +-`*_fitted_windows_playable.mp4` 拼接八个用于训练检查的窗口,并非完整连续任务。TextOP 每段 0.64 秒,总长 5.12 秒;OMG 每段 2 秒,总长 16 秒。窗口常落在动作执行或待机阶段,可能只含背景音,也可能截到半句话;不添加额外解说或补播片段外的指令。 +- +-### TEXTOP:`textop_fitted_windows_playable.mp4` +- +-| 样本 | 视频时间 | 原始 session | 原始动作时间 | +-| --- | --- | --- | --- | +-| 0 | 0.00–0.64 s | `train-000002-seed42.npz` | 0.04–0.68 s | +-| 1 | 0.64–1.28 s | `train-000014-seed42.npz` | 4.20–4.84 s | +-| 2 | 1.28–1.92 s | `train-000014-seed42.npz` | 33.32–33.96 s | +-| 3 | 1.92–2.56 s | `train-000014-seed42.npz` | 62.44–63.08 s | +-| 4 | 2.56–3.20 s | `train-000014-seed42.npz` | 91.72–92.36 s | +-| 5 | 3.20–3.84 s | `train-000014-seed42.npz` | 120.84–121.48 s | +-| 6 | 3.84–4.48 s | `train-000014-seed42.npz` | 149.96–150.60 s | +-| 7 | 4.48–5.12 s | `train-000014-seed42.npz` | 179.24–179.88 s | +- +-### OMG:`omg_fitted_windows_playable.mp4` +- +-| 样本 | 视频时间 | 原始 session | 原始动作时间 | +-| --- | --- | --- | --- | +-| 0 | 0.00–2.00 s | `train-000002-seed42.npz` | 0.43–2.43 s | +-| 1 | 2.00–4.00 s | `train-000014-seed42.npz` | 5.63–7.63 s | +-| 2 | 4.00–6.00 s | `train-000014-seed42.npz` | 34.43–36.43 s | +-| 3 | 6.00–8.00 s | `train-000014-seed42.npz` | 62.83–64.83 s | +-| 4 | 8.00–10.00 s | `train-000014-seed42.npz` | 91.63–93.63 s | +-| 5 | 10.00–12.00 s | `train-000014-seed42.npz` | 120.03–122.03 s | +-| 6 | 12.00–14.00 s | `train-000014-seed42.npz` | 148.83–150.83 s | +-| 7 | 14.00–16.00 s | `train-000014-seed42.npz` | 177.63–179.63 s | +- +-## 结果与重现 +- +-两个模型从官方初始化权重重新训练 800 步;修复前关节映射错误的检查点、视频和动作指标作废。TextOP 预测 23 个关节,其余腕关节固定在初始化姿态;OMG 预测全部 29 个关节。OMG 使用离线目标段音频,TextOP 使用已释放窗口的 ASR。视频是参考姿态播放,不是物理跟踪执行。 +- +-训练、误差与映射核验位于数据盘的视频目录 `/root/autodl-tmp/runs/hfd-overfit-videos-jointfixed`,见 `summary.json`、`textop_report.json`、`omg_report.json`、`mapping_audit.json`;新的声音核验见 `audio_verification.json`。 +- +-后续导出使用代码仓库的 `render_overfit.py`,其默认生成 48 kHz 双声道 MP3 并实测解码音量。真值对照使用 `check_joint_mapping.py`,现在也导出原始音频。已有画面可用 `remux_video_audio.py` 重建音轨,例如: +- +-```bash +-python remux_video_audio.py \ +- --video /data/runs/textop/inference/interruption_playable.mp4 \ +- --audio-npz /data/runs/textop/inference/interruption.npz \ +- --output /data/runs/textop/inference/interruption_with_audio.mp4 +-``` +- +-若从完整 session NPZ 裁切音频,需传 `--audio-start-seconds`(动作起点 + 0.2 秒)。默认播放增益为 1;`--gain` 仅影响导出的声音,不影响模型。脚本在增益会造成削波时拒绝导出。 +- +-## 当前视频格式与 VS Code 播放 +- +-当前交付统一为 MP4 容器、H.264 视频、48 kHz 双声道 MP3 音轨;文件名以 `_playable.mp4` 结尾。根据官方文档选择 H.264、MP4 与 MP3,避开不支持的 AAC。此前 VP8 WebM 在用户当前预览中仍无法打开,因此已清理。 +- +-官方支持清单:https://github.com/microsoft/vscode/issues/180780 +-官方 Webview 说明:https://code.visualstudio.com/api/extension-guides/webview#supported-media-formats +- +-所有视频检查过完整声画解码、非静音音轨、帧数和帧率;H.264 码流与原始视频的 SHA256 一致,转换只更换音频编码和封装,没有改动作。检查记录见视频目录的 `playable_mp4_check.json`。远程服务器上的解码检查不等于本地 VS Code 实测,首次打开请使用新文件名。 +- +-后续 `render_overfit.py` 与 `remux_video_audio.py` 默认导出 H.264/MP3 MP4,无需 WebM 或额外的 `--vscode-preview` 参数。 +diff --git a/hfd_bridge/genmo_prefetch.py b/hfd_bridge/genmo_prefetch.py +index e9445d1..b37742f 100644 +--- a/hfd_bridge/genmo_prefetch.py ++++ b/hfd_bridge/genmo_prefetch.py +@@ -15,6 +15,8 @@ class CPUBatches(IterableDataset): + + def __iter__(self): + os.environ['CUDA_VISIBLE_DEVICES']='' ++ from .genmo_runtime import require_replay_runtime ++ require_replay_runtime() + torch.set_num_threads(2) + from .genmo_data import PairedNativeProvider,GenmoBatches + provider=PairedNativeProvider(**self.options) +diff --git a/hfd_bridge/genmo_training.py b/hfd_bridge/genmo_training.py +index 3f094ed..eba28ac 100644 +--- a/hfd_bridge/genmo_training.py ++++ b/hfd_bridge/genmo_training.py +@@ -13,6 +13,19 @@ def evaluation_mode(model): + for module,state in states.items():module.training=state + + ++@contextmanager ++def sampling_autocast(device_type): ++ """Do not cache no-grad weight casts inside the outer training autocast. ++ ++ Reusing those casts in the subsequent loss forward silently removes the ++ denoiser's parameter gradients. Preserve the caller's AMP dtype and state. ++ """ ++ with torch.autocast(device_type=device_type, ++ enabled=torch.is_autocast_enabled(device_type), ++ dtype=torch.get_autocast_dtype(device_type),cache_enabled=False): ++ yield ++ ++ + def execution_mask(mask,history,execute,weight): + """Preserve each feature's total loss mass while emphasizing executed frames.""" + result=mask.float().clone() +@@ -61,7 +74,7 @@ def consistent_translation_loss(inputs,outputs,pipeline): + def generate_parameters(model,batch): + """Generate from past parameters only. Future placeholders contain no GT.""" + from .genmo import HISTORY,FUTURE +- with evaluation_mode(model): ++ with sampling_autocast(batch['smpl_params_w']['transl'].device.type),evaluation_mode(model): + model.prepare_batch(batch,'regression') + model.create_condition_mask(batch,{},'diffusion',train=False) + result=model.pipeline.denoiser3d.forward_test(batch) +diff --git a/hfd_bridge/human_sessions.py b/hfd_bridge/human_sessions.py +index 52fb0c1..89175d9 100644 +--- a/hfd_bridge/human_sessions.py ++++ b/hfd_bridge/human_sessions.py +@@ -65,11 +65,11 @@ class NativeHumanObserver: + def __init__(self, ds, module, root): + # Share the bridge's verified local metadata mirror. Raw SMPL-X assets + # remain indexed in their original shared-storage tar shards. +- from .storage import ArchivePath ++ from .storage import ArchivePath,store + class IndexedReader: + def read(self, path):return ArchivePath(path).read_bytes() + self.store = IndexedReader() +- pairs = pd.read_parquet(Path(root)/'metadata/seed_g1_smplx_pairs.parquet') ++ pairs = pd.read_parquet(store().metadata_path('metadata/seed_g1_smplx_pairs.parquet',root)) + self.pairs = {r.motion_id:r for r in pairs.itertuples()} + self.ds, self.module = ds, module + self.rows = {} +diff --git a/hfd_bridge/kimodo.py b/hfd_bridge/kimodo.py +index 67cb18d..e9742a6 100644 +--- a/hfd_bridge/kimodo.py ++++ b/hfd_bridge/kimodo.py +@@ -25,6 +25,8 @@ def setup(gpu=0): + HFD_BODY_MODELS_ROOT=os.environ.get('HFD_BODY_MODELS_ROOT','/root/autodl-fs/pangxinlei/models/body_models'), + TEXT_ENCODERS_DIR=os.environ.get('TEXT_ENCODERS_DIR',str(Path(os.environ.get('KIMODO_MODELS_ROOT','/root/autodl-fs/pangxinlei/models/kimodo'))/'runtime-text-encoders')), + TEXT_ENCODER_DEVICE='cuda:0',OMP_NUM_THREADS='4',MKL_NUM_THREADS='4') ++ os.environ.setdefault('HFD_NATIVE_STARTUP_CACHE','1') ++ os.environ.setdefault('HFD_STARTUP_INDEX_CACHE_DIR',str(Path(os.environ.get('HFD_DATA_DISK','/root/autodl-tmp'))/'cache/hfd/startup-indexes')) + torch.set_num_threads(4) + + +diff --git a/hfd_bridge/loader.py b/hfd_bridge/loader.py +index 31321d0..124b42d 100644 +--- a/hfd_bridge/loader.py ++++ b/hfd_bridge/loader.py +@@ -30,7 +30,7 @@ def _instrument_returns(text,filename): + tree=Returns().visit(ast.parse(text,filename)) + return compile(ast.fix_missing_locations(tree),filename,'exec') + +-def load_hfd(): ++def load_hfd(replay_numpy2=False): + root=Path(os.environ['HFD_CODE_ROOT']);sys.path.insert(0,str(root/'src')) + source=root/'src/psi/data/seed/full_duplex_audio_command_dataset.py' + text=source.read_text() +@@ -74,5 +74,9 @@ def load_hfd(): + module._hfd_return_hooks=hooked;module._hfd_current_observer=None + module._hfd_observer_return=_return_hook + exec(_instrument_returns(text,str(source)) if hooked else compile(text,str(source),'exec'),module.__dict__) ++ if replay_numpy2: ++ from .replay_numerics import quat_to_yaw_wxyz,wrap_angle ++ module.quat_to_yaw_wxyz=quat_to_yaw_wxyz ++ module.wrap_angle=wrap_angle + module.SOURCE_SHA256=hashlib.sha256(source.read_bytes()).hexdigest() + return module +diff --git a/hfd_bridge/omg_checks.py b/hfd_bridge/omg_checks.py +index 17e6574..7784fa9 100644 +--- a/hfd_bridge/omg_checks.py ++++ b/hfd_bridge/omg_checks.py +@@ -37,8 +37,9 @@ class WarmupEarlyStopping(EarlyStopping): + + + class OMGTrainingObserver(pl.Callback): +- def __init__(self,output,check_pipeline=False,audio_policy='hfd_past_2s',use_asr_text=False,input_text_source='asr'): ++ def __init__(self,output,check_pipeline=False,audio_policy='hfd_past_2s',use_asr_text=False,input_text_source='asr',checkpoint_root=None): + self.output=Path(output);self.check_pipeline=check_pipeline;self.optimizer_updates=0;self.validation_batches=0 ++ self.checkpoint_root=Path(checkpoint_root) if checkpoint_root else self.output + self.audio_policy=audio_policy + self.use_asr_text=use_asr_text + self.input_text_source=input_text_source +@@ -52,6 +53,7 @@ class OMGTrainingObserver(pl.Callback): + self.optimizer_updates+=1 + def on_train_start(self,trainer,module): + self.initial_step=trainer.global_step ++ self.logged_step=trainer.global_step;self.micro_losses=[];self.micro_compute=[];self.micro_wait=[] + def on_before_optimizer_step(self,trainer,module,optimizer): + if self.check_pipeline: + grads=[p.grad for p in module.parameters() if p.grad is not None] +@@ -60,12 +62,18 @@ class OMGTrainingObserver(pl.Callback): + def on_train_batch_end(self,trainer,module,outputs,batch,batch_idx): + loss=outputs['loss'] if isinstance(outputs,dict) else outputs + if not torch.isfinite(loss).all():raise RuntimeError('Nonfinite OMG training loss') +- if trainer.is_global_zero: ++ self.micro_losses.append(float(loss.detach())*trainer.accumulate_grad_batches) ++ self.micro_compute.append(time.monotonic()-self.batch_started) ++ if self.data_wait_seconds is not None:self.micro_wait.append(self.data_wait_seconds) ++ if trainer.is_global_zero and trainer.global_step>self.logged_step: + with (self.output/'metrics.jsonl').open('a') as f: +- f.write(json.dumps(dict(step=trainer.global_step,loss=float(loss.detach()), ++ f.write(json.dumps(dict(step=trainer.global_step,loss=sum(self.micro_losses)/len(self.micro_losses), ++ micro_batches=len(self.micro_losses),learning_rate=trainer.optimizers[0].param_groups[0]['lr'], + elapsed_seconds=time.monotonic()-self.started, +- data_wait_seconds=self.data_wait_seconds, +- compute_seconds=time.monotonic()-self.batch_started))+'\n') ++ data_wait_seconds=sum(self.micro_wait),compute_seconds=sum(self.micro_compute), ++ peak_allocated_mib=torch.cuda.max_memory_allocated(module.device)/2**20 if module.device.type=='cuda' else 0))+'\n') ++ if trainer.global_step>self.logged_step: ++ self.logged_step=trainer.global_step;self.micro_losses=[];self.micro_compute=[];self.micro_wait=[] + self.previous_batch_end=time.monotonic() + def on_train_batch_start(self,trainer,module,batch,batch_idx): + self.batch_started=time.monotonic() +@@ -91,7 +99,7 @@ class OMGTrainingObserver(pl.Callback): + elapsed_seconds=time.monotonic()-self.started,best_checkpoints=best,early_stopping=early),indent=2)+'\n') + if not self.check_pipeline:return + changed=any(not torch.equal(self.initial[n],p.detach().cpu()) for n,p in module.named_parameters() if n in self.initial) +- file=self.output/'pipeline.ckpt' ++ file=self.checkpoint_root/'pipeline.ckpt' + trainer.save_checkpoint(str(file)) + trainer.strategy.barrier() + ckpt=torch.load(file,map_location='cpu',weights_only=False) +diff --git a/hfd_bridge/ondemand.py b/hfd_bridge/ondemand.py +index 0fd7144..4d00440 100644 +--- a/hfd_bridge/ondemand.py ++++ b/hfd_bridge/ondemand.py +@@ -35,8 +35,15 @@ def native_dataset(config,split): + os.environ[env]=str(resolve(c[field])) + os.environ['HFD_STARTUP_INDEX_CACHE']=os.environ.get('HFD_NATIVE_STARTUP_CACHE','0') + os.environ['HFD_RANGE_CACHE_MAX_BYTES']=str(1024*2**20) +- ds=build_dataset(c.limit_sources,split,OmegaConf.to_container(c.recipe_overrides),candidate_pool=c.candidate_pool) ++ selection=None ++ if c.get('final_release'): ++ from .final_split import FinalSelection ++ selection=FinalSelection(OmegaConf.to_container(c.final_release,resolve=True),split) ++ selection.identity['third_party_replay_source_sha256']=sha256(resolve(c.code_root)/'src/psi/data/seed/third_party_dataset.py') ++ upstream_split=selection.upstream_split if selection is not None else split ++ ds=build_dataset(c.limit_sources,upstream_split,OmegaConf.to_container(c.recipe_overrides),candidate_pool=c.candidate_pool,replay_numpy2=selection is not None) + ds.bridge_compact_output=True ++ if selection is not None:ds.bridge_final_selection=selection + _datasets[key]=ds + return _datasets[key],c + +@@ -113,6 +120,24 @@ def legal_times(arrays,model,num_primitive=4): + return range(12,max(12,stop)) + + ++def final_window_times(session,prefix_frames,model,num_primitive): ++ """Use published clocks for rejection tests, without replaying a candidate. ++ ++ Native dense starts stop prefix_frames+1 before the final motion frame. ++ The pinned third-party replay keeps that same complete native timeline. ++ This supplies the exact legal grid before loading any motion/audio arrays. ++ """ ++ count=int(session['dense_steps']);groups=int(session['semantic_windows']) ++ geometry=dict(valid_start_count=count, ++ motion_q29_xyz_quat=range(count+int(prefix_frames)+1), ++ audio=range((groups-1)*6400+9600)) ++ return legal_times(geometry,model,num_primitive) ++ ++ ++def final_window_count(session,prefix_frames,model,num_primitive): ++ return len(final_window_times(session,prefix_frames,model,num_primitive)) ++ ++ + class NativeSessionProvider: + def __init__(self,config,split,model,seed=42,num_primitive=4,memory_mib=512, + asr_cache=None,asr_cache_mib=2048,asr_model=None,asr_device='cuda:0',asr_batch_size=32, +@@ -121,6 +146,8 @@ class NativeSessionProvider: + if language not in ('all','en','zh'):raise ValueError('Unknown foreground language filter') + self.language=language + self.ds,c=native_dataset(config,split) ++ self.split=split ++ self.final_selection=getattr(self.ds,'bridge_final_selection',None) + if c.candidate_pool or c.limit_sources or c.count is not None: + raise ValueError('On-demand formal sampling requires the full balanced native recipe') + self.model=model;self.seed=int(seed);self.num_primitive=num_primitive +@@ -152,6 +179,10 @@ class NativeSessionProvider: + upper=np.asarray([self.ds._session_max_stream_steps(s)*20+self.ds.prefix_frames+1 for s in self.ds.sessions],dtype=np.int64) + self.bounds=(upper+7)//8 if model=='textop' else (upper*3+4)//5 + self.eligible=np.array([language=='all' or s[0].get('_session_voice_name','').endswith('_'+language) for s in self.ds.sessions],dtype=bool) ++ if self.final_selection is not None: ++ allowed=self.final_selection.native ++ if allowed and max(allowed)>=len(self.ds.sessions):raise ValueError('Final native index exceeds the pinned source inventory') ++ self.eligible &= np.array([i in allowed for i in range(len(self.ds.sessions))],dtype=bool) + if not self.eligible.any():raise ValueError('No native sessions in requested foreground language') + self.bounds[~self.eligible]=0 + self.rank=int(os.environ.get('RANK','0'));self.world_size=int(os.environ.get('WORLD_SIZE','1')) +@@ -174,27 +205,39 @@ class NativeSessionProvider: + for session in self.ds.sessions: + digest.update(json.dumps([(int(r['row_id']),r.get('_session_interaction_mode'),r.get('_session_voice_name'),r.get('_session_voice_variant')) for r in session],sort_keys=True).encode()) + self.audit['ordered_session_selection_sha256']=digest.hexdigest() ++ if self.final_selection is not None: ++ self.audit['final_release']=self.final_selection.identity ++ self.audit['ordered_session_selection_sha256']=self.final_selection.ordered_identity(language) + def arrays(self,index): + if hasattr(self,'eligible') and not self.eligible[int(index)]:raise ValueError('Session excluded by foreground language filter') + cached=self.cache.get(index) + if cached is not None:return cached + raw=fixed_session(self.ds,int(index),self.seed);arrays=raw['arrays'] + raw['times']=legal_times(arrays,self.model,self.num_primitive) ++ if getattr(self,'final_selection',None) is not None and self.model in ('textop','omg'): ++ if raw['times']!=self.window_times(index):raise ValueError('Final native window grid changed') + if len(raw['times'])>self.bounds[index]:raise RuntimeError('Native conservative length bound violated') + raw['size']=sum(v.nbytes for v in arrays.values()) + self.cache.put(index,raw,raw['size']);return raw ++ def window_times(self,index): ++ if getattr(self,'final_selection',None) is not None and self.model in ('textop','omg'): ++ return final_window_times(self.final_selection.native_record(index)['session'], ++ self.ds.prefix_frames,self.model,self.num_primitive) ++ return self.arrays(index)['times'] ++ def window_count(self,index): ++ return len(self.window_times(index)) + def choose(self): + for _ in range(10000): + slot=int(self.rng.integers(int(self.cumulative[-1]))) + index=int(np.searchsorted(self.cumulative,slot,side='right')) +- raw=self.arrays(index);n=len(raw['times']) ++ n=self.window_count(index) + if self.rng.random()*self.bounds[index]self.bounds[index]:raise RuntimeError('Replayed third-party length bound violated') ++ raw_item=dict(arrays=arrays,times=times,size=sum(x.nbytes for x in arrays.values()),source_id='third-party::'+sid) ++ self.cache.put(cache_key,raw_item,raw_item['size']);return raw_item + import io, wave + with np.load(io.BytesIO(self._asset('timelines',sid)),allow_pickle=False) as z: + raw=np.asarray(z['states'],dtype=np.float32) +@@ -347,7 +418,7 @@ class ThirdPartySessionProvider: + for _ in range(10000): + slot=int(self.rng.integers(int(self.cumulative[-1]))) + index=int(np.searchsorted(self.cumulative,slot,side='right')) +- n=len(self.arrays(index)['times']) ++ n=self.window_count(index) + if self.rng.random()*self.bounds[index]self.capacity:self.files.popitem(last=False) ++ data=os.pread(handle.fileno(),count,start) ++ while len(data)1 and not ondemand, +- val_check_interval=c.training.eval_every,check_val_every_n_epoch=None) ++ val_check_interval=validation_interval(c.training),check_val_every_n_epoch=None) ++ from hydra.utils import instantiate ++ trainer_options['plugins']=[instantiate(cfg.checkpoint_io)] + if ondemand: +- trainer_options['limit_val_batches']=int(c.training.validation_batches) ++ trainer_options['limit_val_batches']=len(val_loader.dataset) + elif not val: + trainer_options['limit_val_batches']=0 + if os.environ.get('LOCAL_RANK','0')=='0': +@@ -316,7 +333,7 @@ def omg(c, train, val, output): + model.train() + if model.text_encoder is not None:model.text_encoder.encoder.eval() + trainer.fit(model,train_dataloaders=train_loader,val_dataloaders=val_loader,ckpt_path=c.training.get('resume')) +- trainer.save_checkpoint(str(output/'last.ckpt')) ++ trainer.save_checkpoint(str(weights_output/'last.ckpt')) + + + def native_options(c): +diff --git a/train_genmo_full.py b/train_genmo_full.py +index cd3a7cc..43bb8f7 100644 +--- a/train_genmo_full.py ++++ b/train_genmo_full.py +@@ -37,6 +37,20 @@ def main(): + for name in ['init_checkpoint','text_encoder','body_models']: + if not Path(c[name]).exists():raise FileNotFoundError(c[name]) + out=Path(c.output).resolve();weights=Path(c.weights_output).resolve() ++ from hfd_bridge.genmo_runtime import require_replay_runtime ++ runtime=require_replay_runtime() ++ from hfd_bridge.genmo_replay_audit import read_status ++ audit_path=c.data.get('replay_audit_status') ++ if audit_path: ++ audit_status=read_status(audit_path,runtime, ++ require_risk=bool(c.data.get('require_replay_risk_gate',True))) ++ from hfd_bridge.asr import sha256 as config_sha256 ++ if audit_status['identity']['config_sha256']!=config_sha256(Path(c.data.config)): ++ raise ValueError('Replay audit uses another session configuration') ++ for name,digest in audit_status['identity']['implementation'].items(): ++ if config_sha256(Path(__file__).parent/name)!=digest: ++ raise ValueError('Replay audit implementation changed: '+name) ++ if audit_status.get('failed'):raise RuntimeError('Replay audit failed: '+str(audit_status.get('failure'))) + if a.check_only: + print(OmegaConf.to_yaml(c));return + os.environ.update(CUDA_VISIBLE_DEVICES=str(c.gpu),GENMO_ROOT=str(c.genmo_root), +@@ -69,9 +83,11 @@ def main(): + protocol=OmegaConf.to_container(c,resolve=True) + for key in ['output','weights_output','gpu','python']:protocol.pop(key,None) + protocol['training'].pop('resume',None) ++ protocol['replay_runtime']=runtime + protocol['data_implementation']={name:sha256(Path(__file__).parent/name) for name in + ['train_genmo_full.py','hfd_bridge/genmo.py','hfd_bridge/genmo_data.py','hfd_bridge/genmo_validation.py', + 'hfd_bridge/genmo_training.py','hfd_bridge/genmo_geometry.py','hfd_bridge/genmo_prefetch.py', ++ 'hfd_bridge/genmo_runtime.py','hfd_bridge/genmo_replay_audit.py','hfd_bridge/genmo_safety.py', + 'hfd_bridge/ondemand.py','hfd_bridge/training.py','hfd_bridge/textop_condition.py', + 'hfd_bridge/human_sessions.py','hfd_bridge/reference_text.py','hfd_bridge/reference_runtime.py', + 'hfd_bridge/loader.py','hfd_bridge/sessions.py','hfd_bridge/storage.py','hfd_bridge/clocks.py']} +@@ -149,11 +165,17 @@ def main(): + finally:Path(temporary).unlink(missing_ok=True) + + class Audit(Callback): +- def __init__(self):self.start=time.monotonic();self.grad_checks=0;self.last_end=None ++ def __init__(self):self.start=time.monotonic();self.grad_checks=0;self.denoiser_grad_checks=0;self.last_end=None + def on_train_batch_start(self,trainer,module,batch,batch_idx): + self.started=time.monotonic();self.wait=None if self.last_end is None else self.started-self.last_end + self.data_profile=batch.pop('_hfd_data_profile',{}) + def on_before_optimizer_step(self,trainer,module,optimizer): ++ # Recovery sampling must never silence the trainable Transformer. ++ core=module.pipeline.denoiser3d.denoiser.blocks ++ missing=[name for name,param in core.named_parameters() if param.requires_grad and param.grad is None] ++ if missing: ++ raise FloatingPointError(f'GenMo Transformer gradients missing for {len(missing)} tensors: {missing[:3]}') ++ self.denoiser_grad_checks+=1 + if trainer.global_step>=4 and trainer.global_step%100 and self.grad_checks>=4:return + gradients=[v.grad for v in module.parameters() if v.requires_grad and v.grad is not None] + if not gradients or not all(torch.isfinite(g).all() for g in gradients):raise FloatingPointError('Missing/nonfinite gradients') +@@ -180,11 +202,13 @@ def main(): + def on_save_checkpoint(self,trainer,module,state):state['hfd_contract']=protocol + + audit=Audit() ++ from hfd_bridge.genmo_safety import GenmoReplaySafety ++ replay_safety=GenmoReplaySafety(weights,c.training.get('early_checkpoint_steps',[]),audit_path,runtime) + from hfd_bridge.genmo_validation import GenmoClosedLoopValidation + closed=c.get('strategy',{}).get('closed_loop_validation',{}) + monitor='val/closed_loop_score' if closed.get('enabled',False) else 'val/loss' + callbacks=[HFDResumeState(train,val),GenmoValidationNoise(int(c.training.seed)+1000003), +- GenmoClosedLoopValidation(val.provider,text,closed,out),audit, ++ GenmoClosedLoopValidation(val.provider,text,closed,out),audit,replay_safety, + ModelCheckpoint(dirpath=str(weights/'checkpoints'),every_n_train_steps=int(c.training.save_every), + save_top_k=int(c.training.keep_last_checkpoints),monitor='step',mode='max',save_last='link', + filename='step-{step:08d}',auto_insert_metric_name=False,enable_version_counter=False), +@@ -204,9 +228,12 @@ def main(): + DataLoader(val,batch_size=None,num_workers=0,generator=val_generator), + ckpt_path=str(checkpoint) if checkpoint else None) + summary=dict(global_step=trainer.global_step,elapsed_seconds=time.monotonic()-audit.start, +- checked_gradient_updates=audit.grad_checks,weights_output=str(weights),gpu_peak_bytes=torch.cuda.max_memory_allocated()) ++ checked_gradient_updates=audit.grad_checks,checked_denoiser_updates=audit.denoiser_grad_checks, ++ weights_output=str(weights),gpu_peak_bytes=torch.cuda.max_memory_allocated(), ++ replay_runtime=runtime,replay_audit_failure=replay_safety.failure) + (out/'training_summary.json').write_text(json.dumps(summary,indent=2)+'\n') + print('COMPLETE',json.dumps(summary),flush=True) ++ if replay_safety.failure:raise RuntimeError('Stopped at a complete update after replay audit failure') + + + if __name__=='__main__':main() +diff --git a/train_kimodo.py b/train_kimodo.py +index 75219f0..aba64ab 100644 +--- a/train_kimodo.py ++++ b/train_kimodo.py +@@ -51,6 +51,8 @@ def main(): + p.add_argument('--resume',nargs='?',const='auto');p.add_argument('--stop-after',type=int) + p.add_argument('--offline-check',action='store_true',help='Full pretrained gradient/text/causality checks only') + a=p.parse_args();c=OmegaConf.load(a.config);out=Path(c.output);weights=Path(c.weights_output) ++ # Model imports can initialise CUDA; select the physical GPU before them. ++ os.environ['CUDA_VISIBLE_DEVICES']=str(int(c.gpu)) + if c.get('kimodo_root'):os.environ['KIMODO_ROOT']=str(c.kimodo_root) + if c.get('body_models'):os.environ['HFD_BODY_MODELS_ROOT']=str(c.body_models) + if c.get('text_models_root'):os.environ['KIMODO_MODELS_ROOT']=str(c.text_models_root) diff --git a/handoff-v1/inventory/workspace-status.txt b/handoff-v1/inventory/workspace-status.txt new file mode 100644 index 0000000000000000000000000000000000000000..e02f3ef5e8126ca9153363f0d1393f8349abf734 --- /dev/null +++ b/handoff-v1/inventory/workspace-status.txt @@ -0,0 +1,109 @@ + D FINAL_BASELINE_METHOD.md + D JOINT_MAPPING.md + M MANIFEST.json + M README.md + D SESSION_FORMAT.md + M UPSTREAM_LOCK.json + M configs/genmo.yaml + M configs/genmo_legacy.yaml + M configs/kimodo.yaml + M configs/omg.yaml + M configs/omg_variants.yaml + M configs/sessions.yaml + M configs/textop.yaml + D docs/ASR_CONTEXT_OVERFIT.md + D docs/BASELINE_TEXT_CONDITION.md + D docs/ENGLISH_ONLY_PROBE.md + D docs/FORMAL_TRAINING_READINESS.md + D docs/GENMO_FORMAL_TRAINING.md + D docs/GENMO_HFD_OVERFIT.md + D docs/GENMO_LISTENING_DIAGNOSIS.md + D docs/GENMO_PAST_AUDIO_TEST.md + D docs/GENMO_V2_TRAINING.md + D docs/GENMO_V2_VALIDATION.md + D docs/HFD_HUMAN_BASELINE_INPUT_PROTOCOL.md + M docs/HFD_HUMAN_BASELINE_SOURCES.json + D docs/HUMAN_FORMAL_VALIDATION_20261024.md + D docs/HUMAN_RETARGET_EVALUATION.md + D docs/KIMODO_CLOSED_LOOP_STABILITY.md + D docs/KIMODO_FORMAL_VALIDATION.md + D docs/KIMODO_HFD_OVERFIT_RESULTS.md + D docs/KIMODO_HFD_TRAINING.md + D docs/KIMODO_MOTION_RECOVERY.md + D docs/KIMODO_ROOT_COLLAPSE_DIAGNOSIS.md + D docs/KIMODO_ROOT_TRAINING_CAUSE_ANALYSIS.md + D docs/KIMODO_TRAINING_REPAIR.md + D docs/LOCAL_INDEX_TRAINING.md + D docs/MIGRATION.md + D docs/OMG_ENGLISH_VALIDATION.md + D docs/OMG_TRAINING_REVIEW.md + D docs/OMG_VALIDATION_DIAGNOSIS.md + D docs/OMG_VALIDATION_ROLLOUT.md + D docs/PORTABLE_TRAINING.md + D docs/REFERENCE_DYNAMIC_TRAINING.md + D docs/ROLLOUT_DEBUG.md + D docs/SEMANTICS_CLOCK_FIX.md + D docs/TEXTOP_ENGLISH_VALIDATION.md + D docs/TRAINING_REPRODUCTION_MATRIX.md + D docs/TRAINING_RESUME.md + D docs/WORKSPACE_CLEANUP.md + D docs/history/BASELINE_METHOD_BEFORE_SEMANTIC_FIX.md + D docs/history/DUPLEX_TASK.md + D docs/history/HFD_HUMAN_BASELINE_INPUT_PROTOCOL_RESEARCH_20261017.md + D docs/history/HFD_WINDOW_PROTOCOL.md + D docs/history/NORMAL_TASK.md + D docs/history/OMG_CONTEXT_DEBUG.md + D docs/history/README.md + D docs/history/VIDEO_SAMPLES.md + M hfd_bridge/genmo_prefetch.py + M hfd_bridge/genmo_training.py + M hfd_bridge/human_sessions.py + M hfd_bridge/kimodo.py + M hfd_bridge/loader.py + M hfd_bridge/omg_checks.py + M hfd_bridge/ondemand.py + M hfd_bridge/ondemand_data.py + M hfd_bridge/resume.py + M hfd_bridge/sessions.py + M hfd_bridge/storage.py + M prepare_sessions.py + M requirements/genmo.txt + M requirements/runtime-versions.json + M tools/configure_training.py + M train.py + M train_genmo_full.py + M train_kimodo.py +?? configs/hymotion.yaml +?? configs/hymotion_c24.yaml +?? configs/omg_l.yaml +?? configs/omg_xl.yaml +?? docs/CURRENT_RUNS.md +?? docs/HYMOTION_ASSETS.json +?? hfd_bridge/final_split.py +?? hfd_bridge/final_third_party.py +?? hfd_bridge/genmo_replay_audit.py +?? hfd_bridge/genmo_runtime.py +?? hfd_bridge/genmo_safety.py +?? hfd_bridge/hot_assets.py +?? hfd_bridge/hymotion_autoregressive.py +?? hfd_bridge/hymotion_data.py +?? hfd_bridge/hymotion_evaluation.py +?? hfd_bridge/hymotion_metrics.py +?? hfd_bridge/hymotion_model.py +?? hfd_bridge/hymotion_rep.py +?? hfd_bridge/hymotion_rollout.py +?? hfd_bridge/hymotion_text.py +?? hfd_bridge/hymotion_training.py +?? hfd_bridge/hymotion_validation.py +?? hfd_bridge/omg_scaling.py +?? hfd_bridge/replay_numerics.py +?? hfd_bridge/textop_text_cache.py +?? patches/ +?? requirements/hymotion.txt +?? tools/audit_genmo_replay.py +?? tools/download_hymotion_assets.py +?? tools/index_hymotion_text_cache.py +?? tools/prepare_hot_assets.py +?? train_hymotion.py +?? train_hymotion_autoregressive.py +?? upstream/hfd/src/psi/data/seed/third_party_dataset.py diff --git a/handoff-v1/inventory/workspace-submodules.txt b/handoff-v1/inventory/workspace-submodules.txt new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/handoff-v1/original-run/checkpoint-cleanup-20261029.json b/handoff-v1/original-run/checkpoint-cleanup-20261029.json new file mode 100644 index 0000000000000000000000000000000000000000..d833ded8fd1e16a3c104771a8692dbb48e4e94d9 --- /dev/null +++ b/handoff-v1/original-run/checkpoint-cleanup-20261029.json @@ -0,0 +1,39 @@ +{ + "time": "2026-10-29T14:20:18.808937+08:00", + "removed": [ + { + "path": "/root/autodl-tmp/runs/hfd-omg-english-split20261026/checkpoints/step-00001024.ckpt", + "bytes": 1059547872, + "allocated_bytes": 1059549184 + }, + { + "path": "/root/autodl-tmp/runs/hfd-omg-english-split20261026/checkpoints/step-00099328.ckpt", + "bytes": 1059548064, + "allocated_bytes": 1059549184 + }, + { + "path": "/root/autodl-tmp/runs/hfd-omg-english-split20261026/checkpoints/step-00099584.ckpt", + "bytes": 1059548064, + "allocated_bytes": 1059549184 + }, + { + "path": "/root/autodl-tmp/runs/hfd-omg-english-split20261026/checkpoints/step-00099840.ckpt", + "bytes": 1059548064, + "allocated_bytes": 1059549184 + } + ], + "removed_symlink": { + "path": "/root/autodl-tmp/runs/hfd-omg-english-split20261026/checkpoints/last-v1.ckpt", + "target": "step-00099840.ckpt" + }, + "retained": [ + "/root/autodl-tmp/runs/hfd-omg-english-split20261026/best/best.ckpt", + "/root/autodl-tmp/runs/hfd-omg-english-split20261026/last.ckpt", + "/root/autodl-tmp/runs/hfd-omg-english-split20261026/checkpoints/last.ckpt" + ], + "removed_file_bytes": 4238192064, + "removed_allocated_bytes": 4238196736, + "local_free_before_bytes": 9968586752, + "local_free_after_bytes": 14206783488, + "observed_free_increase_bytes": 4238196736 +} diff --git a/handoff-v1/original-run/code_sha256.json b/handoff-v1/original-run/code_sha256.json new file mode 100644 index 0000000000000000000000000000000000000000..802bccff197fda4cacc13bfbc399a323eac9292d --- /dev/null +++ b/handoff-v1/original-run/code_sha256.json @@ -0,0 +1,142 @@ +{ + "hfd_bridge/storage.py": "8b553d204c19ddad7cec223c96f42146865266bc578e7207ad80eb9c7c8f396d", + "hfd_bridge/__init__.py": "885528eb35015189a0bcd2dc62b60c61878efcf959caef68ff69e93501cafd5c", + "hfd_bridge/loader.py": "13772137dc5cbb036bd5b5e4ee6ffe8cb45ba8d033ad580887b4dd63b253ac9b", + "hfd_bridge/sessions.py": "c271aecf69e52acaee178b1381e465f22af82732f606d4e0bd5f891e2bb86076", + "hfd_bridge/asr.py": "4baaf4faecfdce2054844a41326760220af301eb11d6fad892bed8c7380dceb0", + "hfd_bridge/training.py": "b513a632bcf010c83c23a35bcfd09e189149ff8c0d7d667d8966ec09e6e55970", + "hfd_bridge/native.py": "72ca9f675c2dfc7a2e304191da7b4d4016cb6bc226ccc22f9e7d0aeb2889c9f4", + "hfd_bridge/paths.py": "2aaee82ecad571dcc1186c275e270af6c854c670db4ae354ee984d23a50b8bf3", + "hfd_bridge/configured_data.py": "f916e0f9d4cdd2a4120fe94a9f1dfe59bbd24606d71ce1941023f04f97f181cd", + "hfd_bridge/textop_checks.py": "00662346781a638298cfa344b9ce8a58b6e8bba11261cd5783fdca956f4df7c8", + "hfd_bridge/omg_checks.py": "1c80361412892538400cc225ab28aa03d5357ccab7a53e3a6d46deb5da9d32e8", + "hfd_bridge/textop_condition.py": "4fca0f4a02df3b843a908b28b64766877585aa3300977359d21fdb3adb5a3c28", + "hfd_bridge/joints.py": "7827c82ca3463a026db6007c41a66528e1294485b4acb081f27f09449da2d154", + "hfd_bridge/audio_condition.py": "cab6795bdddc36aa63d4f54c3c8e8aaee7958c86882abe35ae76f67a79b4b945", + "hfd_bridge/batch_asr.py": "9d4f0ba1d382f45c7a0adf562bc008485fce7cd00e86e29607528676f503f8f1", + "hfd_bridge/ondemand.py": "ddf6c76a6b3d736de9dbf5e27c9d8d6e9384156cee68fc96bba0ce08237aa7c5", + "hfd_bridge/ondemand_data.py": "24c61e36623eaca0bdc2f7bff2c8ce307946c4a5afef18f0427b96c06c9ed57e", + "hfd_bridge/prefetch.py": "b227251cc6470f980231999bbecbeb1fc583c754eda56e70d4f3c84ed1845d64", + "hfd_bridge/human_sessions.py": "8b93bbe9450fef5eabd05c827b596a41dfa4748aeaaf93a63849877f5e8df38e", + "hfd_bridge/genmo.py": "3c898d2828267f38f8ba9f41aae62c4a52cbae87cf63dd2cef2b6634f7160060", + "hfd_bridge/genmo_audio.py": "1cdfa6ebc19a75af7540b2280f5dc9c3ecc584ed1c386c434de32e3777020c6f", + "hfd_bridge/clocks.py": "27054c0b2d3e7d4087eb8fc3c59c6bcf5a581ca2e3022341e306732dcbe92637", + "hfd_bridge/semantics.py": "42ed59d58609095b8053d0598a9c575d52af9b8508e699162743cc11599f03b4", + "hfd_bridge/reference_text.py": "50da93fed3be2bad05035bf978c9be41ba7b48aaf3ce7c43a879519e7386a121", + "hfd_bridge/reference_runtime.py": "2b7fe129152d0f7ea3a4a4e3f5fc9ae799e805138a0d76487e9f59065fe7d975", + "hfd_bridge/resume.py": "e8f9c06323415cbdf011272efa4504f12e0c44079d4648a92277f27a5e5d0cc3", + "hfd_bridge/resume_lightning.py": "69e4c6bf11a199f3a979227eb551870c44e371fcae0af924ec64d2d58d3ce7a6", + "hfd_bridge/index_cache.py": "d75cbcb0105448c61c841e525a1f17e63c55c83ad9201104e8118e22f7728b17", + "hfd_bridge/genmo_data.py": "4b102d1eecb0af9757a76d9600ede345c3f21f92166276a22cbf8e144fa9a38c", + "hfd_bridge/genmo_validation.py": "fa265cf9df51a4b36db977be945547dd78bdf105db9aed8ccba940b1a977f5e4", + "hfd_bridge/kimodo_data.py": "35b3b7d8fc8f36fe65eb311e6e65e667fe1fdbcdfc23e58f37fb75c1c9e9feae", + "hfd_bridge/kimodo.py": "196f8ff99caec99470b04adb5e24efac83b5f4239aa2fdc3083b4438fa4cf460", + "hfd_bridge/kimodo_recovery.py": "b9177ff13b55b753430dd14d46778b723d2a03ce8927313c246c5d7ce780809b", + "hfd_bridge/kimodo_closed_loop.py": "38d7e022b1450a715b35542cd23a70b1491a6dbd307dbb4be730b9f562100f91", + "hfd_bridge/kimodo_causal_rep.py": "c0c40003a3d36ec35e8490fa1165ac88723178b15a0e118d2d1d4c9f7ed7c537", + "hfd_bridge/kimodo_diffusion_training.py": "078fe8cf1ee39d910a6ca5ef94d67f20c4398cfe505895a089acb80c05336539", + "hfd_bridge/kimodo_text_worker.py": "a275c121a9364ae85b6233a67f3acb63ceddb2c9da490c2ea7d476f4cc710751", + "hfd_bridge/kimodo_full_data.py": "c62b4bb2673f1875792edb3403347f76bb52fa2360d92a9c1091eae10d43f905", + "hfd_bridge/kimodo_checkpoint.py": "b9cd24df225c91a6212c354983e3f51acb495373747f303c6096000af37f7d59", + "hfd_bridge/kimodo_finetuning.py": "a2c54f04af2b516c14d00025c8b695093ff09da6f04b0b73eb3269817bc0a544", + "hfd_bridge/reference_assets.py": "e977a1d804807629fe925d511acdd25b488a11acc3767c0234020af8f699a43c", + "hfd_bridge/omg_model_loader.py": "f972080dd6140e1574124be615c9351001b51a73c5774b63bd1c063671bd56aa", + "hfd_bridge/genmo_geometry.py": "4d3e912a858b8d3d4ddf73941c3120d5304ed70187c02fc969730816e155b2cb", + "hfd_bridge/genmo_training.py": "8276d4f266b44e2c112e14021eb29c442753f2de07abf29e9adb321226dc87d8", + "hfd_bridge/genmo_prefetch.py": "3375352f7668dfd5c76ff629853aa8e4d0bcd2e2294a6e96932280f2eafa68e4", + "hfd_bridge/hymotion_rep.py": "ee2f8282720d311b98d3ceb8cc72f63219622b5f61a138af6eb3221181a89f7a", + "hfd_bridge/hymotion_model.py": "db0968384b87929c6495b87647a37e92e13729a94a34f6a5347f4d23ed05be5c", + "hfd_bridge/hymotion_data.py": "06170526a18b60fa486f084864e2e65f2608274ed967ef367b29de269f8c9147", + "hfd_bridge/hymotion_text.py": "f0238ffc3656e2a90a26360c5e9bc8c6de3d2d6be0af6e87787c83237914e98b", + "hfd_bridge/hymotion_rollout.py": "a7fefe8d72f9e3614103632d68d1d63ce3ee072650739683131f5315d0a41181", + "hfd_bridge/hymotion_validation.py": "cf6fd4b14fdbafc65a8a65abaade620b43d5b1f95e7a7a6194a21247f2ad4127", + "hfd_bridge/hymotion_training.py": "67f40e79cc43950f6876f8c0c2c67fe1f606f1b67e692a7dd239457fa4b96a2c", + "hfd_bridge/hymotion_clock.py": "1eaf969e9198f6902759457488ff8c80fc20c973b72388443b3ef0af7c8282a2", + "hfd_bridge/hymotion_recovery.py": "e47b841f85dfc19b37a646a6cf6e865dd2642718073a3419a97195f16024c379", + "hfd_bridge/human_eval.py": "c943c332315f16f41cb53beee8303f2d94a4b522e678cfd346e8a6c9ec203b71", + "hfd_bridge/omg_rollout.py": "24ecdec4a2c698bd8cdc2b4ac6e1c0ad960c252b5a9f9a15d57a7716545fc4d7", + "hfd_bridge/final_split.py": "d0f0adef291d17d3b73477258b5de74bb8ed477ca1952c9b83ddb115efd34305", + "hfd_bridge/final_third_party.py": "3926a0f72d419bb5006d8a62136a8f5ffd5ddca439c39dd7768f7cf03422d716", + "hfd_bridge/replay_numerics.py": "827c70aa9b2308ac33b6fccee2c4dc88df58d6b89d902d8c215adfdf918ab231", + "hfd_bridge/textop_text_cache.py": "3d5c8a187e461144aa97cc539661e6e90b475800379c0c096b243bcd337a0e38", + "upstream/omg/download_data.py": "a247118ce874864affc5153a23d94788871d3b6d327bcc33ceeba45279d7732c", + "upstream/omg/download_models.py": "811132811b35321174adc69a754cf64f9437eeba5d19f6ad02625ef9d9f73b57", + "upstream/omg/a.py": "66e1c68c0a631d6fef0d7d13948299671baf72ec813fe025c1cac1e222fd4c63", + "upstream/omg/smoke.py": "8008c82e92250f148512e3127b939df3aded7025a89473738d024444d21edcfd", + "upstream/omg/validate_model.py": "36023f739cd8dfe24179f9247b5de04ca425d7a0d73c99b166bbeaea778a975f", + "upstream/omg/validate_setup.py": "4e086d4615c621386cb54885aad0ede341c24a71a1668c3e5386f368c02f3741", + "upstream/omg/paper_table.py": "97c16c1c044639362f46cac439159384a1d79dc80a90c839c5e2ce471d8060d2", + "upstream/omg/paper_main_table.py": "8f95f11b00882e4574df37e2f44d5843a15073004f0b9f78450895c65939196f", + "upstream/omg/scripts/prepare_omg_inputs.py": "99bbc364b5f380ae7c2cc0fa53657722f896d3613d8b98937881d06197beccb4", + "upstream/omg/scripts/extract_omg_paper_row.py": "f64cb31b5a19efba6f3e1b1771d3d5175f6610392a588d118014e98790b82503", + "upstream/omg/scripts/psi0_omg_generate.py": "38ccbd4c8f1d62b470193ae6e3598445f22408e2df6edc6b267ed46b4873f579", + "upstream/omg/scripts/verify_omg_artifact_match.py": "46815bd446b956583ae7feeb0ed03a58b6dde8cc78ecb4a1d4fd1a3a6b737ee7", + "upstream/omg/src/omg/__init__.py": "e06eff3bd6a146e0082518eb791164e26b80b548f9bdcd21c9f0aca1220bca34", + "upstream/omg/src/omg/callbacks/__init__.py": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", + "upstream/omg/src/omg/callbacks/divergence_guard.py": "c1ec2773dddda3bd800247bca0abd3eca46aeae3c78ebae8050b98739292b61a", + "upstream/omg/src/omg/callbacks/global_step_progress.py": "bd21180de0140a38fa2601f51b56c94aef346c757c140786e88b907720f3eb15", + "upstream/omg/src/omg/callbacks/grad_norm.py": "c41e7e0bac01c984c85c38baf6de4b9e10ded95875a79ddc1c9d85ded5b9b742", + "upstream/omg/src/omg/callbacks/speed_timer.py": "29ae0ec76a0545576bc31f240a4beab74f5c5e9ae50a0078c1ed46b8326fc360", + "upstream/omg/src/omg/callbacks/training_diagnostics.py": "6e348e5cca9e690379cdb7a3364678b666d7c53d0a7405d5269ca713574a6917", + "upstream/omg/src/omg/cli/__init__.py": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", + "upstream/omg/src/omg/cli/data/__init__.py": "052be341b9c1b05585d3a27073a741d768974cfdb2a91659a687ed1d3d6291f9", + "upstream/omg/src/omg/cli/data/materialize_episode_cache.py": "c8e61cbcf17f55e17b47ad2a3e5c2aa0919d9f5bcc2a93edbf5a519e79f9793a", + "upstream/omg/src/omg/cli/data/validate_episode_cache.py": "6840e28c774964583d88c6402c201ab36c0150ec67139403a66782364173c878", + "upstream/omg/src/omg/cli/generation/__init__.py": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", + "upstream/omg/src/omg/cli/generation/compute_stats.py": "849ed1ac16612e23e09a887e5a030c5e0c496610e9afe882ce98e99d570bd6ca", + "upstream/omg/src/omg/cli/generation/train.py": "a39cc0c6cfccd7b6660bbd9ab4f88f90bc634bb95ce30e22492c8a6cb74a7c8c", + "upstream/omg/src/omg/cli/generation/benchmark.py": "ff813f344953da4b02388837f5a510230f8d4430393a88f205261faaf28c5dee", + "upstream/omg/src/omg/cli/generation/physical_benchmark.py": "f48b2b9d2d312c44d0e33eebf0c7fd95d8f2ce65a871fe253ef8ff9ceee30dd0", + "upstream/omg/src/omg/cli/generation/generate.py": "66eda839732e76a3b0fc39a767e6a238aa93bbf13ef2f3ddb6e1f9021948bc4a", + "upstream/omg/src/omg/cli/generation/export_onnx.py": "9cacee46fea5e18f8a9f9ffd52b25110f89e80be20d21acce85fe066392c00c1", + "upstream/omg/src/omg/core/__init__.py": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", + "upstream/omg/src/omg/core/logging.py": "86c57564f7d8614d71d95e0ce377724642c8c3aa44501d9c317c33bf9ba7dce7", + "upstream/omg/src/omg/core/paths.py": "6bb254b7219e70b58fc8581479857bfb7ee77a9f8937faf5dcec784dd1173a4d", + "upstream/omg/src/omg/core/tensor.py": "998e6f24948983c15f3257f055a452e07b96cc0f13b357cb4881f4545bddb7d0", + "upstream/omg/src/omg/data/__init__.py": "4e8cd9d824b5b834a673011d1cd95cf0227a1deb60ee917ba85ec364d51db21f", + "upstream/omg/src/omg/data/datamodule.py": "94cbaf84b019d559dd7810494694c156803d318f23a5ec95549315c59ea795d6", + "upstream/omg/src/omg/data/episode_cache.py": "96639991cb75a3db602c608d929e4f742caef615166198e66731c406f068207e", + "upstream/omg/src/omg/data/episode_cache_inspect.py": "b7f43c2a5f7a1c63dcb9c243e79018ebff3ecce056fba22ac9b1948e754feb20", + "upstream/omg/src/omg/data/lerobot_dataset.py": "78172eb344a37cab8457133a957eb25691ad78054f71db0c39171f241be12831", + "upstream/omg/src/omg/data/windowing.py": "88710f0c24d6c696670b702b9ce77609c59acd2106389fa769a00d39a98f2afb", + "upstream/omg/src/omg/generation/__init__.py": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", + "upstream/omg/src/omg/generation/architecture.py": "f77548e52394c78b813ad36dcd5fb8c8981ba393e5c723126491df6c171aeabf", + "upstream/omg/src/omg/generation/conditions/__init__.py": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", + "upstream/omg/src/omg/generation/conditions/t5.py": "67c0f834106ff860e612a978bdba6ef503d2cf0dfb24e831af0bba723e95c890", + "upstream/omg/src/omg/generation/denoisers/__init__.py": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", + "upstream/omg/src/omg/generation/denoisers/interface.py": "c1f8b5f2a93e6cf0a4b5df730f9f744c95d9b58bc5245e3cb15176a61065e1dd", + "upstream/omg/src/omg/generation/denoisers/transformer.py": "ae968a16ba6193607721f5c0c1910122818797407df3ee0f0fafa3728c0e398a", + "upstream/omg/src/omg/generation/diffusion/__init__.py": "fab4af0ec5b7b38892ffb309d6073841de00f8f37e5b4c2da1657e1b30c92e1c", + "upstream/omg/src/omg/generation/diffusion/forcing.py": "935b2f38db1d1d3c496df0b03cadaeb032ef1539522169f2731b87a8724362b5", + "upstream/omg/src/omg/generation/diffusion/guided.py": "3fd42c7ae9d01497223b08421ae3becc0a7be81846c1cf08c8bb944d8949b870", + "upstream/omg/src/omg/generation/diffusion/objectives.py": "db05bcbc32268e5d21298e19130e9e8f4b23eff7d1b679886306cd7c94195675", + "upstream/omg/src/omg/generation/diffusion/schedules.py": "a06c3f3ab1d81e07f680d44f8fb9ef56e8bd2a6e4c5e4ac911c316217f6e92e0", + "upstream/omg/src/omg/generation/export/__init__.py": "0ee7ed60aa43848dad0a4ba0e5dd29d815aab6dc37f4ae1c46d8e56a2ebc1770", + "upstream/omg/src/omg/generation/export/onnx.py": "168ad808867669b69cd69a419d5b8d2c3fd124c69c20daeca53b6fc59a0a21dd", + "upstream/omg/src/omg/generation/losses/__init__.py": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", + "upstream/omg/src/omg/generation/losses/contact.py": "75927b00fc5584988f2ecedbcfbba9af3ba80206966b0a91642ec133db5e4267", + "upstream/omg/src/omg/generation/losses/fk.py": "75927b00fc5584988f2ecedbcfbba9af3ba80206966b0a91642ec133db5e4267", + "upstream/omg/src/omg/generation/losses/motion.py": "00ee200046658dafcae2b9ec439ed0c3203ab533f4dd9d4a8475541df74c2fa5", + "upstream/omg/src/omg/generation/losses/terrain.py": "75927b00fc5584988f2ecedbcfbba9af3ba80206966b0a91642ec133db5e4267", + "upstream/omg/src/omg/generation/metrics/__init__.py": "f686c86d57bab24aa1dd27cfae2bbab88757d7909e74cd7d4b9525a892ba025e", + "upstream/omg/src/omg/generation/metrics/physical.py": "7aa2d8fc7d29b9d3ec836c3d23d0b5a7fc64a1eea5359c7300f6e6e191904f29", + "upstream/omg/src/omg/generation/models/__init__.py": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", + "upstream/omg/src/omg/generation/models/motion_generator.py": "05947c8e73646f634e9bbedf0a60f2becaf1ded21b9d1093f1084eaaefef2a3b", + "upstream/omg/src/omg/motion/__init__.py": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", + "upstream/omg/src/omg/motion/feature_codec.py": "3d148816baf239dbc938210c959954c3c2fac4c6d9914ad48b851976ed7e2960", + "upstream/omg/src/omg/motion/representation.py": "457328078b2a2652e06c39e5fa06c8a0d8ba9223dcf24e508a33cdd1806f45a7", + "upstream/omg/src/omg/pipeline/__init__.py": "3806048000a9bdc918ace57c68aabcd3c56bb503c4998b103d1593553a23a895", + "upstream/omg/src/omg/pipeline/audio.py": "d88a0b51fd72d80fdf7da1344082e7799f481fb114a246ed4e66c0b9745f3d09", + "upstream/omg/src/omg/pipeline/condition_sequence.py": "e831393b0b843219520ee8a836b7cbc0b406295e60947cba1a443f1f78a81377", + "upstream/omg/src/omg/pipeline/human.py": "d5be4338bc69308450c18331da4cf9a93fc9e02bd5596dce92a3e58248ff14d0", + "upstream/omg/src/omg/pipeline/planner.py": "fa7aa07eebf9f56287301a6ab13fb7a42bf8b310e152b0916033019d7d55ce0c", + "upstream/omg/src/omg/robots/__init__.py": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", + "upstream/omg/src/omg/robots/g1/__init__.py": "dc46237fccb843612112274586e9e00dee6a0e472f4ce6b93da267b64f8f4c19", + "upstream/omg/src/omg/robots/g1/constants.py": "1f2d0fd495de8765349ad24bdb8e8b327ffdf80c16b13c665fabfe8e9b1a39ea", + "upstream/omg/src/omg/robots/g1/kinematics.py": "c8adbae0d96ef85ba948e27f1fb16bed35d8529c08a6d71d1943d31644585db4", + "upstream/omg/src/omg/runtime/__init__.py": "9728869f63d93ada09a0f77fc10e98f2add27f4ab84e999d228becca102315fe", + "upstream/omg/src/omg/runtime/checkpoint_io.py": "1b9ce178d3e5c1bba39c77d3098ea7e64788d8c81c48acdebaec09cc1be88e5e", + "upstream/omg/src/omg/runtime/onnx_providers.py": "5ed87e442ebf9c56b4065fc011c89d7745ac5bf163158fbf0732db160d0e5673", + "upstream/omg/src/omg/utils/rotation_conversions.py": "627ead38f2e15b0001eea501f6be09e7b6c309c21f1938ef419f65f8e1eb7ea5", + "train.py": "0724b3db6d374259db9f6378e9b663ef8438e5b969093f3b4d1e46ee56456ddc" +} diff --git a/handoff-v1/original-run/data_manifest.json b/handoff-v1/original-run/data_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..780d5d70a75ec94cc59ed73927e0867d7b0538d3 --- /dev/null +++ b/handoff-v1/original-run/data_manifest.json @@ -0,0 +1,31 @@ +{ + "text_frontend": { + "policy": "none" + }, + "clock_policy": "microphone", + "backend": "native_ondemand", + "train": [], + "val": [], + "text_source": "oracle_foreground_transcript", + "caption_policy": "aligned_reference_context", + "reference_sidecars": [], + "native_options": { + "config": "/root/autodl-tmp/hfd-baseline-portable/configs/sessions.yaml", + "memory_mib": 2048, + "reference_cache": "/root/autodl-tmp/cache/hfd/reference-alignments.sqlite", + "reference_guard_seconds": 0.1, + "refresh_batches": 256, + "language": "en", + "seed": 42, + "asr_device": "cuda:0", + "caption_policy": "aligned_reference_context", + "asr_latency": 0.0, + "clock_policy": "microphone", + "text_conditioning": "none", + "include_third_party": true + }, + "split_manifests": [ + "train_native_manifest.json", + "val_native_manifest.json" + ] +} diff --git a/handoff-v1/original-run/native_config.yaml b/handoff-v1/original-run/native_config.yaml new file mode 100644 index 0000000000000000000000000000000000000000..8cd5e71590cc52ac479ea43510274d781c72a3e9 --- /dev/null +++ b/handoff-v1/original-run/native_config.yaml @@ -0,0 +1,220 @@ +data: + _target_: omg.data.datamodule.GenerationDataModule + limit_each_trainset: null + train_subset_ratio: null + rotation_representation: ${representation.rotation_representation} + dataset_opts: + train: + omg_lerobot_train: + _target_: omg.data.lerobot_dataset.LeRobotG1MotionDataset + dataset_root: ${paths.data_root} + repo_id: THU-MARS/OMG-Data + revision: 6e0dfbc1c5298bff14d4e2b1459ad678af0a38e7 + manifest_sha256: be443885018180dda0f88ed874efc15cb3776a91148148baf49001d56a5ec855 + sequence_duration: 2.0 + num_prev_states: ${representation.num_prev_states} + canonical_frame_idx: ${representation.canonical_frame_idx} + fps: 30.0 + use_text: true + use_audio: false + audio_dim: 35 + use_human_motion: false + human_motion_dim: 66 + split: train + val: + omg_lerobot_val: + _target_: omg.data.lerobot_dataset.LeRobotG1MotionDataset + dataset_root: ${paths.data_root} + repo_id: THU-MARS/OMG-Data + revision: 6e0dfbc1c5298bff14d4e2b1459ad678af0a38e7 + manifest_sha256: be443885018180dda0f88ed874efc15cb3776a91148148baf49001d56a5ec855 + sequence_duration: 2.0 + num_prev_states: ${representation.num_prev_states} + canonical_frame_idx: ${representation.canonical_frame_idx} + fps: 30.0 + use_text: true + use_audio: false + audio_dim: 35 + use_human_motion: false + human_motion_dim: 66 + split: val + test: + omg_lerobot_test: + _target_: omg.data.lerobot_dataset.LeRobotG1MotionDataset + dataset_root: ${paths.data_root} + repo_id: THU-MARS/OMG-Data + revision: 6e0dfbc1c5298bff14d4e2b1459ad678af0a38e7 + manifest_sha256: be443885018180dda0f88ed874efc15cb3776a91148148baf49001d56a5ec855 + sequence_duration: 2.0 + num_prev_states: ${representation.num_prev_states} + canonical_frame_idx: ${representation.canonical_frame_idx} + fps: 30.0 + use_text: true + use_audio: false + audio_dim: 35 + use_human_motion: false + human_motion_dim: 66 + split: test + loader_opts: + train: + batch_size: 192 + num_workers: 0 + prefetch_factor: null + val: + batch_size: 16 + num_workers: 2 + prefetch_factor: 1 + test: + batch_size: 16 + num_workers: 2 + prefetch_factor: 1 +representation: + _target_: omg.motion.representation.G1MotionRepresentation + stats_path: assets/stats/g1_125d_stats.json + kinematics_path: assets/robots/g1/g1_kinematics.json + num_prev_states: 10 + canonical_frame_idx: 9 + feat_dim: 125 + sequence_length: 60 + rotation_representation: rot6d + rot6d_gradient_mode: canonical +denoiser: + _target_: omg.generation.denoisers.transformer.MotionTransformerDenoiser + input_dim: 125 + hidden_dim: 512 + num_layers: 10 + num_heads: 8 + mlp_ratio: 4.0 + text_dim: 768 + dropout: 0.1 + rope_base: 10000.0 + audio_local_attn_window: 8 + human_motion_local_attn_window: 2 + use_human_motion_local_attn: false + self_attention_qk_norm: true + cross_attention_qk_norm: true + frame_cond_injection: per_layer_film +diffusion: + _target_: omg.generation.diffusion.guided.GuidedDiffusion + timesteps: 1000 + objective: pred_x0 + beta_schedule: cosine + train_timestep_respacing: '' + test_timestep_respacing: '50' + sampler: ddim + ddim_eta: 0.0 + cfg_scale: 2.5 +loss: + _target_: omg.generation.losses.motion.MotionLoss + simple_root_pos: 0.0 + simple_root_rot: 0.0 + simple_joint_dof: 0.0 + simple_body_pos: 0.0 + vel_root_pos: 0.0 + vel_root_rot: 0.0 + vel_joint_dof: 0.0 + body_pos_consistency: 0.0 + fk_body_pos: 0.0 + fk_body_rot: 0.0 + body_vel: 0.0 + terrain_penetration: 0.0 + contact_height: 0.0 + contact_velocity: 0.0 + seam_root_pos: 0.0 + seam_root_rot: 0.0 + seam_joint_dof: 0.0 + seam_body_pos: 0.0 +model: + _target_: omg.generation.models.motion_generator.MotionGenerator + _recursive_: false + representation: ${representation} + denoiser: ${denoiser} + diffusion: ${diffusion} + loss: ${loss} + text_encoder: + _target_: hfd_bridge.native.recent_t5_encoder + model_name: ${paths.models_root}/t5-base-local + max_length: 50 + output_dim: 768 + diffusion_target: future + condition_dim: ${denoiser.hidden_dim} + text_mask_prob: 0.3 + history_mask_prob: 0.1 + use_audio: true + audio_dim: 35 + audio_mask_prob: 0.1 + use_human_motion: true + human_motion_dim: 66 + human_motion_mask_prob: 0.5 + frame_cond_injection: per_layer_film + log_loss_term_grad_norms: false + loss_term_grad_norm_every_n_steps: 100 + loss_term_grad_norm_norm_type: 2.0 + optimizer: + lr: 6.0e-05 + weight_decay: 0.01 + scheduler: + type: linear_warmup_cosine + warmup_steps: 2000 + warmup_start_factor: 0.0001 + t_max: 100000 + eta_min: 1.0e-06 +trainer: + accelerator: auto + devices: 1 + max_steps: 100000 + val_check_interval: 2000 + log_every_n_steps: 10 + precision: bf16-mixed + gradient_clip_val: 1.0 + enable_checkpointing: true + num_sanity_val_steps: 0 + check_val_every_n_epoch: null + use_distributed_sampler: false + accumulate_grad_batches: 1 + limit_val_batches: 200 +logger: + enabled: false +callbacks: + checkpoint: + _target_: pytorch_lightning.callbacks.ModelCheckpoint + dirpath: ${output_dir}/checkpoints + filename: s{step:06d} + every_n_train_steps: 2000 + save_top_k: -1 + save_last: true + global_step_progress: + _target_: omg.callbacks.global_step_progress.GlobalStepProgressBar + lr_monitor: + _target_: pytorch_lightning.callbacks.LearningRateMonitor + requires_logger: true + logging_interval: step + grad_norm_monitor: null + training_diagnostics: null + divergence_guard: + _target_: omg.callbacks.divergence_guard.DivergenceGuard + dirpath: ${output_dir}/diagnostics + warmup_steps: 2000 + max_loss: 10.0 + loss_monitor: loss + loss_patience: 4 + max_grad_norm: 1000.0 + capture_loss_term_grad_norms: false + norm_type: 2.0 + speed_timer: + _target_: omg.callbacks.speed_timer.SpeedTimer +paths: + repo_root: . + output_root: outputs + data_root: ${oc.env:OMG_DATA_ROOT,${paths.repo_root}/data/OMG-Data} + materialized_root: ${oc.env:OMG_MATERIALIZED_ROOT,${paths.data_root}/materialized} + models_root: ${oc.env:OMG_MODELS_ROOT,${paths.repo_root}/models} +seed: 1234 +task: fit +ckpt_path: null +init_weights_only_ckpt: null +init_weights_strict: true +exp_name: g1_transformer_guided_50m_mixed_modal_all_original_rot6d_diffusion_only_4gpu_materialized +output_dir: /root/autodl-tmp/runs/hfd-omg-english-split20261026 +checkpoint_io: + _target_: omg.runtime.checkpoint_io.DirectTorchCheckpointIO diff --git a/handoff-v1/original-run/resolved_config.yaml b/handoff-v1/original-run/resolved_config.yaml new file mode 100644 index 0000000000000000000000000000000000000000..9754d88fd8dd370d7ded744767e9da11f3aea39d --- /dev/null +++ b/handoff-v1/original-run/resolved_config.yaml @@ -0,0 +1,86 @@ +schema_version: 1 +mode: train +model: omg +framework: omg_lightning +paths: + data_disk: /root/autodl-tmp + upstream_root: upstream/omg + session_root: /root/autodl-tmp/datasets/hfd-sessions-split20261026 + cache_root: /root/autodl-tmp/cache + output_dir: /root/autodl-tmp/runs/hfd-omg-english-split20261026 +runtime: + python: /root/autodl-tmp/envs/omg/bin/python + num_gpus: 1 + gpu_ids: + - 1 + cpu_threads: 4 + offline: true + distributed_backend: gloo + distributed_timeout_seconds: 1800 + environment: + HFD_LOCAL_INDEX_CACHE: '1' +data: + train_sessions: + - /root/autodl-tmp/datasets/hfd-sessions-split20261026/train/*.npz + val_sessions: + - /root/autodl-tmp/datasets/hfd-sessions-split20261026/val/*.npz + window_policy: full_session + max_windows: null + cache_sessions: 512 + asr_dir: /root/autodl-tmp/datasets/hfd-sessions-split20261026/reference + asr_latency_seconds: 0.0 + audio_policy: hfd_past_2s + use_asr_text: true + caption_policy: aligned_reference_context + clock_policy: microphone + text_conditioning: none + backend: native_ondemand + batched_motion: true + prefetch_batches: 1 + native: + config: configs/sessions.yaml + memory_mib: 2048 + reference_cache: /root/autodl-tmp/cache/hfd/reference-alignments.sqlite + reference_guard_seconds: 0.1 + refresh_batches: 256 + language: en +training: + resume: null + seed: 42 + batch_size: 512 + max_steps: 100000 + learning_rate: 6.0e-05 + num_workers: 0 + precision: bf16-mixed + accumulate_grad_batches: 1 + gradient_clip_val: 1.0 + save_every: 2000 + eval_every: 2000 + experiment: 50m + scheduler: + type: linear_warmup_cosine + warmup_steps: 2000 + warmup_start_factor: 0.0001 + t_max: 100000 + eta_min: 1.0e-06 + keep_last_checkpoints: 3 + validation_batches: 16 + fixed_validation: true + save_best: true + early_stopping: + enabled: false + start_step: 30000 + patience: 10 + min_delta: 0.0002 +inference: + omg_guidance: joint_one + omg_replan_mode: native_async + omg_buffer_remaining_frames: 24 + omg_sampling_steps: 10 +native_overrides: + model: + text_encoder: + _target_: hfd_bridge.native.recent_t5_encoder +weights: + models_root: /root/autodl-tmp/models/OMG + init_checkpoint: /root/autodl-tmp/models/OMG/official/checkpoints/paper/50m/sstep=090000.ckpt diff --git a/handoff-v1/original-run/train_native_manifest.json b/handoff-v1/original-run/train_native_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..472c5fc3c4c202f5e0ea9e85b75d10ca695516b1 --- /dev/null +++ b/handoff-v1/original-run/train_native_manifest.json @@ -0,0 +1,206 @@ +{ + "split": "train", + "sessions": 126950, + "model": "omg", + "caption_policy": "aligned_reference_context", + "clock_policy": "microphone", + "text_conditioning": "none", + "recipe_overrides": { + "command_interruption_session_variants": 2 + }, + "native_source_sha256": "6e30d7f9e125e5b0cf618a3a1966b50b59c5b4fdc62116501ef80ab141a2665a", + "seed": 42, + "asr": { + "format": "hfd-native-aligned-reference-v1", + "inventory": { + "assets": 61860, + "by_language": { + "en": 61860, + "zh": 0 + }, + "inventory_sha256": "dac2d9befa68a693d8b6595574a786f53f37647cb30389413a1217809b31cea9", + "raw_root": "/root/autodl-fs/pangxinlei/hfd_data", + "complete": true, + "model": { + "repo": "Qwen/Qwen3-ForcedAligner-0.6B", + "revision": "c7cbfc2048c462b0d63a45797104fc9db3ad62b7", + "route": "AutoDL" + }, + "missing": 0, + "parent_inventory_sha256": "bb586e33fa2db3d0b49dbbf28419282cb411645560cadb65a9ca744df11a499b", + "foreground_language": "en" + }, + "guard_seconds": 0.1, + "context_seconds": 30.0, + "text_source": "oracle_foreground_transcript", + "implementation": "2b7fe129152d0f7ea3a4a4e3f5fc9ae799e805138a0d76487e9f59065fe7d975" + }, + "rank": 0, + "world_size": 1, + "sampling_seed": 9000043, + "asr_device": "cuda:0", + "sampling": "uniform legal action-window marginal over native and third-party sessions via conservative-bound rejection", + "recipe_sha256": "ed7da387ed766c1e54dc2a54fcc68e6df813085eee50a0d39f481f8e3334bc54", + "native_config": { + "online_audio": true, + "continuous_audio_root": "/cpfs01/thscc/sharestorage/AIGC-06-01/Hithink_VideoGen_Group/user_workspace/yetinghong/videogen/hfd_audio_command_continuous_office_v1", + "metadata_jsonl": "/cpfs01/thscc/sharestorage/AIGC-06-01/Hithink_VideoGen_Group/user_workspace/buwendong/user/code/qwen3-tts/metadata.jsonl", + "motion_metadata_csv": "/cpfs01/thscc/sharestorage/AIGC-06-01/Hithink_VideoGen_Group/user_workspace/yetinghong/code/Psi0/dataset/seed_all/metadata/seed_metadata_v004.csv", + "motion_root": "/cpfs01/thscc/sharestorage/AIGC-06-01/Hithink_VideoGen_Group/user_workspace/yetinghong/code/Psi0/dataset/seed_all", + "background_assignment_path": "/cpfs01/thscc/sharestorage/AIGC-06-01/Hithink_VideoGen_Group/user_workspace/yetinghong/videogen/hfd_audio_command_continuous_office_v1/manifests/source_background_assignments.jsonl", + "background_source_name": "hfd_continuous_office_background", + "background_task_name": "audio_command_text", + "action_exec_horizon": 20, + "motion_fps": 50, + "motion_metadata_fps": 120, + "audio_sampling_rate": 16000, + "audio_window_seconds": 0.6, + "audio_sample_step_frames": 20, + "action_sample_step_frames": 1, + "motion_cache_size": 256, + "fixed_voice_per_record": true, + "max_voice_windows": 30, + "max_voice_median_ratio": 3.0, + "limit_sources": 0, + "max_motion_steps": 0, + "initial_idle_max_steps": 7, + "stationary_blend_frames": 20, + "stationary_heading_alignment": true, + "safe_stationary_anchors": true, + "stationary_idle_residual_retargeting": true, + "stationary_listen_hold": true, + "dynamic_task_entry_blend": true, + "multi_turn_sessions": true, + "causal_listen_timing": true, + "interruption_sessions": true, + "independent_task_pools": true, + "session_min_turns": 1, + "session_max_turns": 4, + "session_voice_variants": 2, + "turn_taking_session_variants": 4, + "command_interruption_session_variants": 2, + "emergency_stop_session_variants": 2, + "safe_hold_session_variants": 0, + "backchannel_session_variants": 2, + "hesitation_session_variants": 3, + "pause_session_variants": 2, + "interruption_task_session_ratio": 0.15, + "control_task_session_ratio": 0.1, + "backchannel_task_session_ratio": 0.15, + "hesitation_task_window_ratio": 0.12, + "pause_task_window_ratio": 0.08, + "control_followup_ratio": 0.5, + "control_post_idle_min_steps": 3, + "control_post_idle_max_steps": 20, + "control_post_idle_sampling": "short_heavy_long_tail", + "control_min_start_seconds": 1.2, + "control_max_start_seconds": 4.0, + "control_min_remaining_seconds": 2.0, + "control_activity_aware_timing": true, + "filter_known_tts_translation_mismatches": true, + "control_command_manifest": "/cpfs01/thscc/sharestorage/AIGC-06-01/Hithink_VideoGen_Group/user_workspace/yetinghong/videogen/hfd_control_commands_v1/manifest.jsonl", + "backchannel_command_manifest": "/cpfs01/thscc/sharestorage/AIGC-06-01/Hithink_VideoGen_Group/user_workspace/yetinghong/videogen/hfd_backchannel_commands_v2_short/manifest.jsonl", + "hesitation_pause_manifest": "/cpfs01/thscc/sharestorage/AIGC-06-01/Hithink_VideoGen_Group/user_workspace/yetinghong/videogen/hfd_hesitation_pause_v2/manifest.jsonl", + "new_task_max_rows_per_instruction_train": 32, + "new_task_max_rows_per_instruction_val": 8, + "hesitation_filled_ratio": 0.6, + "self_repair_implicit_ratio": 0.35, + "backchannel_min_start_seconds": 1.2, + "backchannel_max_start_seconds": 6.0, + "backchannel_min_remaining_seconds": 1.2, + "safe_hold_stability_window_frames": 20, + "interruption_ratio": 0.5, + "interruption_min_start_seconds": 2.0, + "interruption_max_start_seconds": 6.0, + "interruption_min_remaining_seconds": 1.5, + "turn_transition_frames": 50, + "max_speech_onset_steps": 3, + "listen_reaction_steps": 1, + "max_stream_steps": 1919, + "long_session_drop_ratio": 0.001, + "canonical_motion_manifest": "", + "hidden_root": "/__hfd__/unused", + "split": "train" + }, + "text_source": "oracle_foreground_transcript", + "asr_enabled": false, + "foreground_language": "en", + "indexed_sessions": 348980, + "eligible_modes": { + "pause": 7708, + "hesitation": 11658, + "turn-taking": 60206, + "backchannel": 14290, + "emergency-stop": 9290, + "interruption": 14138 + }, + "eligible_unique_instructions": 4241, + "eligible_unique_motion_ids": 91830, + "native_indices_preserved": true, + "ordered_session_selection_sha256": "60a9b4e673b5628e309aed968b78c0b2f296578a611b45d2ee04c0ed9e49e67e", + "final_release": { + "profile": "isolated_benchmark_1024_20261026", + "revision": "bdca2fe7581c263e9ceb3a96204b314e199faddf", + "split": "train", + "upstream_split": "train", + "selection_sha256": "4480d6df798000127ac56b265cccca5afca292b1e4968c491adc9b61ead2b3e9", + "manifest_index_sha256": "79980aadae886bbe03058210083ff3b540d8ee743c193f1c80d3dcf1f300e294", + "implementation": { + "final_split.py": "d0f0adef291d17d3b73477258b5de74bb8ed477ca1952c9b83ddb115efd34305", + "final_third_party.py": "3926a0f72d419bb5006d8a62136a8f5ffd5ddca439c39dd7768f7cf03422d716", + "replay_numerics.py": "827c70aa9b2308ac33b6fccee2c4dc88df58d6b89d902d8c215adfdf918ab231" + }, + "vqa_inputs": false, + "seed_policy": "recorded_per_session", + "replay_arithmetic": "numpy2_float32_scalars", + "counts": { + "test": 1024, + "train": 258661, + "val": 512 + }, + "third_party_replay_source_sha256": "212fd08f50a6211caf3442d19772c8437b4de36ca0fb6de7f6865bfd621e84da" + }, + "mixed_native_sessions": 117290, + "mixed_third_party_sessions": 9660, + "third_party": { + "package": "hfd_third_party_multi_v5", + "split": "train", + "sessions": 9660, + "caption_policy": "aligned_reference_context", + "selection_path": "audio/hfd_third_party_multi_v5/training_scale_all_commands/selection.json", + "selection_sha256": "71ed6921a3e223dcb9db584b2dd55206039ca8144b95a4b759c07ae686fbfe19", + "hours": 291.87272222222464, + "motion_fps": 50, + "audio_sr": 16000, + "motion_order": "q29,xyz3,quat_wxyz4", + "sampling": "uniform legal action windows across both packages", + "foreground_language": "en", + "indexed_sessions": 26737, + "upstream_split": "train", + "final_release": { + "profile": "isolated_benchmark_1024_20261026", + "revision": "bdca2fe7581c263e9ceb3a96204b314e199faddf", + "split": "train", + "upstream_split": "train", + "selection_sha256": "4480d6df798000127ac56b265cccca5afca292b1e4968c491adc9b61ead2b3e9", + "manifest_index_sha256": "79980aadae886bbe03058210083ff3b540d8ee743c193f1c80d3dcf1f300e294", + "implementation": { + "final_split.py": "d0f0adef291d17d3b73477258b5de74bb8ed477ca1952c9b83ddb115efd34305", + "final_third_party.py": "3926a0f72d419bb5006d8a62136a8f5ffd5ddca439c39dd7768f7cf03422d716", + "replay_numerics.py": "827c70aa9b2308ac33b6fccee2c4dc88df58d6b89d902d8c215adfdf918ab231" + }, + "vqa_inputs": false, + "seed_policy": "recorded_per_session", + "replay_arithmetic": "numpy2_float32_scalars", + "counts": { + "test": 1024, + "train": 258661, + "val": 512 + }, + "third_party_replay_source_sha256": "212fd08f50a6211caf3442d19772c8437b4de36ca0fb6de7f6865bfd621e84da" + }, + "pose_source": "current pinned recipe replay; stored poses are not targets" + }, + "refresh_batches": 256, + "batch_size": 512 +} diff --git a/handoff-v1/original-run/training_summary.json b/handoff-v1/original-run/training_summary.json new file mode 100644 index 0000000000000000000000000000000000000000..8c28cbc4173778a0a7c5e16ff8433c7eee0e5680 --- /dev/null +++ b/handoff-v1/original-run/training_summary.json @@ -0,0 +1,11 @@ +{ + "global_step": 100000, + "elapsed_seconds": 53949.8710324727, + "best_checkpoints": [ + { + "path": "/root/autodl-tmp/runs/hfd-omg-english-split20261026/best/best.ckpt", + "score": 0.021027423441410065 + } + ], + "early_stopping": [] +} diff --git a/handoff-v1/original-run/val_native_manifest.json b/handoff-v1/original-run/val_native_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..be2bb06993c47a2483417e6b92bba615c8fac667 --- /dev/null +++ b/handoff-v1/original-run/val_native_manifest.json @@ -0,0 +1,206 @@ +{ + "split": "val", + "sessions": 302, + "model": "omg", + "caption_policy": "aligned_reference_context", + "clock_policy": "microphone", + "text_conditioning": "none", + "recipe_overrides": { + "command_interruption_session_variants": 2 + }, + "native_source_sha256": "6e30d7f9e125e5b0cf618a3a1966b50b59c5b4fdc62116501ef80ab141a2665a", + "seed": 42, + "asr": { + "format": "hfd-native-aligned-reference-v1", + "inventory": { + "assets": 61860, + "by_language": { + "en": 61860, + "zh": 0 + }, + "inventory_sha256": "dac2d9befa68a693d8b6595574a786f53f37647cb30389413a1217809b31cea9", + "raw_root": "/root/autodl-fs/pangxinlei/hfd_data", + "complete": true, + "model": { + "repo": "Qwen/Qwen3-ForcedAligner-0.6B", + "revision": "c7cbfc2048c462b0d63a45797104fc9db3ad62b7", + "route": "AutoDL" + }, + "missing": 0, + "parent_inventory_sha256": "bb586e33fa2db3d0b49dbbf28419282cb411645560cadb65a9ca744df11a499b", + "foreground_language": "en" + }, + "guard_seconds": 0.1, + "context_seconds": 30.0, + "text_source": "oracle_foreground_transcript", + "implementation": "2b7fe129152d0f7ea3a4a4e3f5fc9ae799e805138a0d76487e9f59065fe7d975" + }, + "rank": 0, + "world_size": 1, + "sampling_seed": 9000049, + "asr_device": "cuda:0", + "sampling": "uniform legal action-window marginal over native and third-party sessions via conservative-bound rejection", + "recipe_sha256": "ed7da387ed766c1e54dc2a54fcc68e6df813085eee50a0d39f481f8e3334bc54", + "native_config": { + "online_audio": true, + "continuous_audio_root": "/cpfs01/thscc/sharestorage/AIGC-06-01/Hithink_VideoGen_Group/user_workspace/yetinghong/videogen/hfd_audio_command_continuous_office_v1", + "metadata_jsonl": "/cpfs01/thscc/sharestorage/AIGC-06-01/Hithink_VideoGen_Group/user_workspace/buwendong/user/code/qwen3-tts/metadata.jsonl", + "motion_metadata_csv": "/cpfs01/thscc/sharestorage/AIGC-06-01/Hithink_VideoGen_Group/user_workspace/yetinghong/code/Psi0/dataset/seed_all/metadata/seed_metadata_v004.csv", + "motion_root": "/cpfs01/thscc/sharestorage/AIGC-06-01/Hithink_VideoGen_Group/user_workspace/yetinghong/code/Psi0/dataset/seed_all", + "background_assignment_path": "/cpfs01/thscc/sharestorage/AIGC-06-01/Hithink_VideoGen_Group/user_workspace/yetinghong/videogen/hfd_audio_command_continuous_office_v1/manifests/source_background_assignments.jsonl", + "background_source_name": "hfd_continuous_office_background", + "background_task_name": "audio_command_text", + "action_exec_horizon": 20, + "motion_fps": 50, + "motion_metadata_fps": 120, + "audio_sampling_rate": 16000, + "audio_window_seconds": 0.6, + "audio_sample_step_frames": 20, + "action_sample_step_frames": 1, + "motion_cache_size": 256, + "fixed_voice_per_record": true, + "max_voice_windows": 30, + "max_voice_median_ratio": 3.0, + "limit_sources": 0, + "max_motion_steps": 0, + "initial_idle_max_steps": 7, + "stationary_blend_frames": 20, + "stationary_heading_alignment": true, + "safe_stationary_anchors": true, + "stationary_idle_residual_retargeting": true, + "stationary_listen_hold": true, + "dynamic_task_entry_blend": true, + "multi_turn_sessions": true, + "causal_listen_timing": true, + "interruption_sessions": true, + "independent_task_pools": true, + "session_min_turns": 1, + "session_max_turns": 4, + "session_voice_variants": 2, + "turn_taking_session_variants": 4, + "command_interruption_session_variants": 2, + "emergency_stop_session_variants": 2, + "safe_hold_session_variants": 0, + "backchannel_session_variants": 2, + "hesitation_session_variants": 3, + "pause_session_variants": 2, + "interruption_task_session_ratio": 0.15, + "control_task_session_ratio": 0.1, + "backchannel_task_session_ratio": 0.15, + "hesitation_task_window_ratio": 0.12, + "pause_task_window_ratio": 0.08, + "control_followup_ratio": 0.5, + "control_post_idle_min_steps": 3, + "control_post_idle_max_steps": 20, + "control_post_idle_sampling": "short_heavy_long_tail", + "control_min_start_seconds": 1.2, + "control_max_start_seconds": 4.0, + "control_min_remaining_seconds": 2.0, + "control_activity_aware_timing": true, + "filter_known_tts_translation_mismatches": true, + "control_command_manifest": "/cpfs01/thscc/sharestorage/AIGC-06-01/Hithink_VideoGen_Group/user_workspace/yetinghong/videogen/hfd_control_commands_v1/manifest.jsonl", + "backchannel_command_manifest": "/cpfs01/thscc/sharestorage/AIGC-06-01/Hithink_VideoGen_Group/user_workspace/yetinghong/videogen/hfd_backchannel_commands_v2_short/manifest.jsonl", + "hesitation_pause_manifest": "/cpfs01/thscc/sharestorage/AIGC-06-01/Hithink_VideoGen_Group/user_workspace/yetinghong/videogen/hfd_hesitation_pause_v2/manifest.jsonl", + "new_task_max_rows_per_instruction_train": 32, + "new_task_max_rows_per_instruction_val": 8, + "hesitation_filled_ratio": 0.6, + "self_repair_implicit_ratio": 0.35, + "backchannel_min_start_seconds": 1.2, + "backchannel_max_start_seconds": 6.0, + "backchannel_min_remaining_seconds": 1.2, + "safe_hold_stability_window_frames": 20, + "interruption_ratio": 0.5, + "interruption_min_start_seconds": 2.0, + "interruption_max_start_seconds": 6.0, + "interruption_min_remaining_seconds": 1.5, + "turn_transition_frames": 50, + "max_speech_onset_steps": 3, + "listen_reaction_steps": 1, + "max_stream_steps": 1919, + "long_session_drop_ratio": 0.001, + "canonical_motion_manifest": "", + "hidden_root": "/__hfd__/unused", + "split": "val" + }, + "text_source": "oracle_foreground_transcript", + "asr_enabled": false, + "foreground_language": "en", + "indexed_sessions": 10055, + "eligible_modes": { + "turn-taking": 127, + "backchannel": 32, + "hesitation": 68, + "pause": 42, + "emergency-stop": 17, + "interruption": 11 + }, + "eligible_unique_instructions": 350, + "eligible_unique_motion_ids": 401, + "native_indices_preserved": true, + "ordered_session_selection_sha256": "7731ed049f43edadde535a321dd70e80f248a1641ceaccefe428f25439babc06", + "final_release": { + "profile": "isolated_benchmark_1024_20261026", + "revision": "bdca2fe7581c263e9ceb3a96204b314e199faddf", + "split": "val", + "upstream_split": "val", + "selection_sha256": "4480d6df798000127ac56b265cccca5afca292b1e4968c491adc9b61ead2b3e9", + "manifest_index_sha256": "79980aadae886bbe03058210083ff3b540d8ee743c193f1c80d3dcf1f300e294", + "implementation": { + "final_split.py": "d0f0adef291d17d3b73477258b5de74bb8ed477ca1952c9b83ddb115efd34305", + "final_third_party.py": "3926a0f72d419bb5006d8a62136a8f5ffd5ddca439c39dd7768f7cf03422d716", + "replay_numerics.py": "827c70aa9b2308ac33b6fccee2c4dc88df58d6b89d902d8c215adfdf918ab231" + }, + "vqa_inputs": false, + "seed_policy": "recorded_per_session", + "replay_arithmetic": "numpy2_float32_scalars", + "counts": { + "test": 1024, + "train": 258661, + "val": 512 + }, + "third_party_replay_source_sha256": "212fd08f50a6211caf3442d19772c8437b4de36ca0fb6de7f6865bfd621e84da" + }, + "mixed_native_sessions": 297, + "mixed_third_party_sessions": 5, + "third_party": { + "package": "hfd_third_party_multi_v5", + "split": "val", + "sessions": 5, + "caption_policy": "aligned_reference_context", + "selection_path": "audio/hfd_third_party_multi_v5/training_scale_all_commands/validation.json", + "selection_sha256": "4706df72f4d4faea9f2839451d5896b5a94e13d7b4d91f8366067721c840d573", + "hours": 8.490611111111106, + "motion_fps": 50, + "audio_sr": 16000, + "motion_order": "q29,xyz3,quat_wxyz4", + "sampling": "uniform legal action windows across both packages", + "foreground_language": "en", + "indexed_sessions": 783, + "upstream_split": "validation", + "final_release": { + "profile": "isolated_benchmark_1024_20261026", + "revision": "bdca2fe7581c263e9ceb3a96204b314e199faddf", + "split": "val", + "upstream_split": "val", + "selection_sha256": "4480d6df798000127ac56b265cccca5afca292b1e4968c491adc9b61ead2b3e9", + "manifest_index_sha256": "79980aadae886bbe03058210083ff3b540d8ee743c193f1c80d3dcf1f300e294", + "implementation": { + "final_split.py": "d0f0adef291d17d3b73477258b5de74bb8ed477ca1952c9b83ddb115efd34305", + "final_third_party.py": "3926a0f72d419bb5006d8a62136a8f5ffd5ddca439c39dd7768f7cf03422d716", + "replay_numerics.py": "827c70aa9b2308ac33b6fccee2c4dc88df58d6b89d902d8c215adfdf918ab231" + }, + "vqa_inputs": false, + "seed_policy": "recorded_per_session", + "replay_arithmetic": "numpy2_float32_scalars", + "counts": { + "test": 1024, + "train": 258661, + "val": 512 + }, + "third_party_replay_source_sha256": "212fd08f50a6211caf3442d19772c8437b4de36ca0fb6de7f6865bfd621e84da" + }, + "pose_source": "current pinned recipe replay; stored poses are not targets" + }, + "refresh_batches": 256, + "batch_size": 512 +} diff --git a/handoff-v1/resources/EXTERNAL_ASSETS.json b/handoff-v1/resources/EXTERNAL_ASSETS.json new file mode 100644 index 0000000000000000000000000000000000000000..4c5e2d4d3cc488aceb4efdd00d4459abc37cd9fa --- /dev/null +++ b/handoff-v1/resources/EXTERNAL_ASSETS.json @@ -0,0 +1,198 @@ +{ + "version": 1, + "description": "External file identities only. Model groups use models-root; hfd-metadata uses raw-root. Llama and licensed body model require authorized manual provisioning.", + "files": [ + { + "group": "omg-100m", + "path": "OMG/official/checkpoints/paper/100m/sstep=100000.ckpt", + "bytes": 1933267854, + "sha256": "5e25dc7fd7251ec6df9dda8d3e008ff2540a5c8d1d691c576d4c64f9916eedaa", + "source": { + "url": "https://huggingface.co/THU-MARS/OMG/resolve/06a88419da9f9125d0b5480cc10966741a189358/checkpoints/paper/100m/sstep%3D100000.ckpt", + "repo": "THU-MARS/OMG", + "revision": "06a88419da9f9125d0b5480cc10966741a189358", + "repo_type": "model", + "filename": "checkpoints/paper/100m/sstep=100000.ckpt", + "local_prefix": "OMG/official" + } + }, + { + "group": "omg-300m", + "path": "OMG/official/checkpoints/paper/300m/sstep=055000.ckpt", + "bytes": 5065534010, + "sha256": "df4ce9aeced84fd96a39583c7f6e2c5773b76f942cb43012c6c51dd9eb806806", + "source": { + "url": "https://huggingface.co/THU-MARS/OMG/resolve/06a88419da9f9125d0b5480cc10966741a189358/checkpoints/paper/300m/sstep%3D055000.ckpt", + "repo": "THU-MARS/OMG", + "revision": "06a88419da9f9125d0b5480cc10966741a189358", + "repo_type": "model", + "filename": "checkpoints/paper/300m/sstep=055000.ckpt", + "local_prefix": "OMG/official" + } + }, + { + "group": "omg-500m", + "path": "OMG/official/checkpoints/paper/500m/sstep=050000.ckpt", + "bytes": 8332144642, + "sha256": "f13bc58f07f104fcff67e35e79650def0fec5a6eb62a193fa1ae9f979fb911cf", + "source": { + "url": "https://huggingface.co/THU-MARS/OMG/resolve/06a88419da9f9125d0b5480cc10966741a189358/checkpoints/paper/500m/sstep%3D050000.ckpt", + "repo": "THU-MARS/OMG", + "revision": "06a88419da9f9125d0b5480cc10966741a189358", + "repo_type": "model", + "filename": "checkpoints/paper/500m/sstep=050000.ckpt", + "local_prefix": "OMG/official" + } + }, + { + "group": "omg-50m", + "path": "OMG/official/checkpoints/paper/50m/sstep=090000.ckpt", + "bytes": 1106432846, + "sha256": "389fae1aa2ba31234e0ef621706362a19387ac485dce1f929408e06a3b012b2c", + "source": { + "url": "https://huggingface.co/THU-MARS/OMG/resolve/06a88419da9f9125d0b5480cc10966741a189358/checkpoints/paper/50m/sstep%3D090000.ckpt", + "repo": "THU-MARS/OMG", + "revision": "06a88419da9f9125d0b5480cc10966741a189358", + "repo_type": "model", + "filename": "checkpoints/paper/50m/sstep=090000.ckpt", + "local_prefix": "OMG/official" + } + }, + { + "group": "omg-shared", + "path": "OMG/t5-base-local/README.md", + "bytes": 8477, + "sha256": "c43c23f205b6839a6adb10f0757265f76c1c6f96b7793a4e53582d3a61cbfb23", + "source": { + "url": "https://huggingface.co/google-t5/t5-base/resolve/a9723ea7f1b39c1eae772870f3b547bf6ef7e6c1/README.md", + "repo": "google-t5/t5-base", + "revision": "a9723ea7f1b39c1eae772870f3b547bf6ef7e6c1", + "repo_type": "model", + "filename": "README.md", + "local_prefix": "OMG/t5-base-local" + } + }, + { + "group": "omg-shared", + "path": "OMG/t5-base-local/config.json", + "bytes": 1208, + "sha256": "46dd7cb62d29c81fb551e0ef1ea274c24a46ba441eeb948897706252933df033", + "source": { + "url": "https://huggingface.co/google-t5/t5-base/resolve/a9723ea7f1b39c1eae772870f3b547bf6ef7e6c1/config.json", + "repo": "google-t5/t5-base", + "revision": "a9723ea7f1b39c1eae772870f3b547bf6ef7e6c1", + "repo_type": "model", + "filename": "config.json", + "local_prefix": "OMG/t5-base-local" + } + }, + { + "group": "omg-shared", + "path": "OMG/t5-base-local/generation_config.json", + "bytes": 147, + "sha256": "f5a1c7e2be8092018d8835128987edf0111637dd98e90599cc80310fef75d95a", + "source": { + "url": "https://huggingface.co/google-t5/t5-base/resolve/a9723ea7f1b39c1eae772870f3b547bf6ef7e6c1/generation_config.json", + "repo": "google-t5/t5-base", + "revision": "a9723ea7f1b39c1eae772870f3b547bf6ef7e6c1", + "repo_type": "model", + "filename": "generation_config.json", + "local_prefix": "OMG/t5-base-local" + } + }, + { + "group": "omg-shared", + "path": "OMG/t5-base-local/model.safetensors", + "bytes": 891646390, + "sha256": "a90903540cc02cbeb7ff9f823f1a80eb778c7e22426a0e620b01c77a5ec8f5b4", + "source": { + "url": "https://huggingface.co/google-t5/t5-base/resolve/a9723ea7f1b39c1eae772870f3b547bf6ef7e6c1/model.safetensors", + "repo": "google-t5/t5-base", + "revision": "a9723ea7f1b39c1eae772870f3b547bf6ef7e6c1", + "repo_type": "model", + "filename": "model.safetensors", + "local_prefix": "OMG/t5-base-local" + } + }, + { + "group": "omg-shared", + "path": "OMG/t5-base-local/spiece.model", + "bytes": 791656, + "sha256": "d60acb128cf7b7f2536e8f38a5b18a05535c9e14c7a355904270e15b0945ea86", + "source": { + "url": "https://huggingface.co/google-t5/t5-base/resolve/a9723ea7f1b39c1eae772870f3b547bf6ef7e6c1/spiece.model", + "repo": "google-t5/t5-base", + "revision": "a9723ea7f1b39c1eae772870f3b547bf6ef7e6c1", + "repo_type": "model", + "filename": "spiece.model", + "local_prefix": "OMG/t5-base-local" + } + }, + { + "group": "omg-shared", + "path": "OMG/t5-base-local/tokenizer.json", + "bytes": 1389353, + "sha256": "d2acde0d8d71dd30a711834b07781b9c89feaac33fd332f60507699282740066", + "source": { + "url": "https://huggingface.co/google-t5/t5-base/resolve/a9723ea7f1b39c1eae772870f3b547bf6ef7e6c1/tokenizer.json", + "repo": "google-t5/t5-base", + "revision": "a9723ea7f1b39c1eae772870f3b547bf6ef7e6c1", + "repo_type": "model", + "filename": "tokenizer.json", + "local_prefix": "OMG/t5-base-local" + } + }, + { + "group": "hfd-metadata", + "path": "MANIFEST.json", + "bytes": 283191, + "sha256": "a7d91a52dabd142d73a8e93ed86403a5763385b3115ff2217777bf82d2bb492e", + "source": { + "repo": "Tinghong-Ye22/hfd_data", + "revision": "93ba4a7a9c66b3e1b7041e5f4ea038ef2418b948", + "repo_type": "dataset", + "filename": "MANIFEST.json", + "local_prefix": "" + } + }, + { + "group": "hfd-metadata", + "path": "PATH_MAP.json", + "bytes": 1151, + "sha256": "0e89066ad6758bfbd2811ba497e541aaf2d8f51aeeda3959379b765930f53098", + "source": { + "repo": "Tinghong-Ye22/hfd_data", + "revision": "93ba4a7a9c66b3e1b7041e5f4ea038ef2418b948", + "repo_type": "dataset", + "filename": "PATH_MAP.json", + "local_prefix": "" + } + }, + { + "group": "hfd-metadata", + "path": "index.sqlite", + "bytes": 454651904, + "sha256": "e41b7ba92d5ccd57092b7c3a190828cc0e8a545c79a8bce575633a68f2bb19bf", + "source": { + "repo": "Tinghong-Ye22/hfd_data", + "revision": "93ba4a7a9c66b3e1b7041e5f4ea038ef2418b948", + "repo_type": "dataset", + "filename": "index.sqlite", + "local_prefix": "" + } + }, + { + "group": "hfd-metadata", + "path": "metadata/seed_g1_smplx_pairs.parquet", + "bytes": 26580927, + "sha256": "63fcd6eb1deb1c270d93a7944f9f2fe95a4433c3bf557cb4ca8671bd18197b10", + "source": { + "repo": "Tinghong-Ye22/hfd_data", + "revision": "93ba4a7a9c66b3e1b7041e5f4ea038ef2418b948", + "repo_type": "dataset", + "filename": "metadata/seed_g1_smplx_pairs.parquet", + "local_prefix": "" + } + } + ] +} \ No newline at end of file diff --git a/handoff-v1/resources/RAW_MANIFEST.json b/handoff-v1/resources/RAW_MANIFEST.json new file mode 100644 index 0000000000000000000000000000000000000000..a7a0161ebde6dd3bddf834c6a4e9a11875c9de43 --- /dev/null +++ b/handoff-v1/resources/RAW_MANIFEST.json @@ -0,0 +1,6752 @@ +[ + { + "file": "shards/hfd_audio_command_continuous_office_v1/00000.tar", + "files": 309, + "samples": 309, + "raw_bytes": 40928796, + "archive_bytes": 41472000, + "sha256": "edc572e1618859c3bfc0222779798067814e48d656d7c0e8f41b2e32be6f159d", + "source_signature": "37e378cef35465bccc9534542eea9074d98ecc3272aaf8190d9a7308ee8d9275", + "index_sha256": "60cd75683c4099fb6b23592c06691d646caaec30704b86acb8a74f4a31419cf4" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00001.tar", + "files": 314, + "samples": 314, + "raw_bytes": 37722616, + "archive_bytes": 38277120, + "sha256": "48adba5d312b2294c84dff371f6035765843c80691eb7e3f37ae289e18d1b717", + "source_signature": "003393a0c54c26567ffcc9c757c8c8feee7e1fd933bf8ea81eab062fbffa425d", + "index_sha256": "c636d9716d7e894ad91082d833e13c9766f2c24d398281dea647b2c3ed3ff81e" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00002.tar", + "files": 348, + "samples": 348, + "raw_bytes": 44328912, + "archive_bytes": 44943360, + "sha256": "ffbb4f34de637beced4f9d77fb461397a92073137b1b8988def1eda2c335b6fc", + "source_signature": "e6738eb193b7ddbbecffd5358e37d200f72bea2b810d8e76a00ea67d6d1a2ad3", + "index_sha256": "ebdab87db67eba278c012bda28b15b254580c85402e0c0eb6fd18a98341e7822" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00003.tar", + "files": 329, + "samples": 329, + "raw_bytes": 42260876, + "archive_bytes": 42844160, + "sha256": "8505a4adbf2c097fe9623ee02208948f00fa70111a4b4054860c54cc4493b9d9", + "source_signature": "321bd34dc7f292f30a8737f7865f33115e08c52272a06a1359bcda12ebeb7550", + "index_sha256": "84274f85485636fb1ad90a661d7271d0edd73b18345684d9d1cf1f9dc8ef713d" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00004.tar", + "files": 340, + "samples": 340, + "raw_bytes": 41397360, + "archive_bytes": 41994240, + "sha256": "7cbc76258f5c97f4f9521675c702e9f62e6dda8a6d6269b14d1c44e014c8a120", + "source_signature": "8057a6810885a069a2580513991a2f1f63c6ce6969f292b5bcefb559a82c643b", + "index_sha256": "ec51e3265048b9feb01fecca85255688a8203cfa36f2f652baebadc68a94c374" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00005.tar", + "files": 321, + "samples": 321, + "raw_bytes": 40353324, + "archive_bytes": 40919040, + "sha256": "73db027a323025f5ab1fa692a2b49c8c4a56fbb04ed9767b66688095523bd9c4", + "source_signature": "991e587407a771370065a498539f6c5d8d42ddcc25f8416d670fbb27bb33c70b", + "index_sha256": "7c12c18d4fef7fdfa638c7394e5224bb23648c82577c6b721e8b443d91ca984a" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00006.tar", + "files": 333, + "samples": 333, + "raw_bytes": 41365052, + "archive_bytes": 41953280, + "sha256": "637182c81d39415b614beb5f51c652747f0ec27a5080f6d1421bd9c7a74aa3e5", + "source_signature": "a649b89a415b7394f78d3bbad6b2e233ea8d760ad8f7b8e9c78ff40e06794617", + "index_sha256": "f450e70bc01ad58dcc7db59ecddc2bdfcfbbb66a933c2fb32a6570a9b0201a22" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00007.tar", + "files": 311, + "samples": 311, + "raw_bytes": 39866484, + "archive_bytes": 40417280, + "sha256": "40e304f756efdbc92c6a7529ee1a0ee25946c882ddc0fb42c96815f9bff273cf", + "source_signature": "bdb33f57dacff55e946e701a7883f86af5f3ea371d37ddcfe3cceb015c15d245", + "index_sha256": "d986680a54c157c7fd24cce809d5eea28ceb1e291d601c94bade8d798a1140bc" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00008.tar", + "files": 326, + "samples": 326, + "raw_bytes": 42356744, + "archive_bytes": 42936320, + "sha256": "37ff9d9fc933151d2488fe89e2e128381deb70a18c1362aa1e6bcd635d841e48", + "source_signature": "f2fe5b70489a6bae0917cc7eb454310bdee8eb32f6f2914ce631cd2fd3b3ec04", + "index_sha256": "ce6bc39e062a5db3f51cceaad41f3af2b2695612c761e1f7f4fcb3b4967d4335" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00009.tar", + "files": 326, + "samples": 326, + "raw_bytes": 42625544, + "archive_bytes": 43202560, + "sha256": "0e331a96c0319595445874e497c0251e47f17a9e937d8ce19bde4e159d845199", + "source_signature": "1703aabc86a19664a3097940178174c48c37c7262085ba49a95bc24a83d0bdbc", + "index_sha256": "af99fcc0fe042093233eed1cb879fed2d3869593b67d817eadada198baea8646" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00010.tar", + "files": 307, + "samples": 307, + "raw_bytes": 38227908, + "archive_bytes": 38768640, + "sha256": "05340ebc28eecb9edbf72e811969f5e1ac5b371cae6846121b97d40c577b712c", + "source_signature": "cd9700aa0eac5ed4cfcc7b7a783d64b91bdc9a383b3414c4ac04ba469b357ee2", + "index_sha256": "1e1dca0cb8bed3a1346b1e9457c2bd4dd0475ca9335ff312e5bcd97f6bbb61af" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00011.tar", + "files": 332, + "samples": 332, + "raw_bytes": 41320208, + "archive_bytes": 41902080, + "sha256": "ee7eccf466d7359ad848fb6900c596793c69e5275b9282939ea03013fc228681", + "source_signature": "2b069dc7f33830920c7af9b7d38e356794854f613263d1578747df19e1e58833", + "index_sha256": "0e8361a79813964c5a2d7dc07bf10e1e7d6104c6a8d0e52f488c11cedb11426e" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00012.tar", + "files": 319, + "samples": 319, + "raw_bytes": 40135636, + "archive_bytes": 40704000, + "sha256": "a6c4f86dac1bc08371ea25b61fd0f2af3258156d180400c7bcdb5a19fc265ee0", + "source_signature": "e857ad635856e39e18a04ebab8d7af0eded10ea5d7879fffe113769dc9b81c85", + "index_sha256": "380bf7d6305c9cab3c2ced0f847ce38a7afd7eb85912fd64f3d69aa9af7ab5ad" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00013.tar", + "files": 337, + "samples": 337, + "raw_bytes": 41186028, + "archive_bytes": 41779200, + "sha256": "be5e8a931aacbd1a3ff45739484d1583f65e3debce50c7b4dcdbba27f9566b3e", + "source_signature": "e8655f4970b3083721bc332c738aa78a49653c902615090a33ce5cd4e66833d9", + "index_sha256": "817c50fa2bf4d5ffbb6834b8ad1713d31aebc23d6fef481a209f3cde4e4f6b4c" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00014.tar", + "files": 332, + "samples": 332, + "raw_bytes": 40629008, + "archive_bytes": 41216000, + "sha256": "e9b0c8fafb1a632d43efd8547f7e9007c06d2a259e14b0a310b795892c2918ee", + "source_signature": "9db70f03b5552cb390f6f616200246851d94a507e5aae3e149fe45137adb405e", + "index_sha256": "c484d8603e8be654286b199102809d3ec4f71aa3bf977f2252caa541fced82aa" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00015.tar", + "files": 310, + "samples": 310, + "raw_bytes": 43072200, + "archive_bytes": 43622400, + "sha256": "bc43e1804a64e13b30856d1ece422122176fae4070eb97de7f0d3be137d8dbd0", + "source_signature": "a4b63ebc3dea21497bfe0b31cfe39464378f14644488c09338f4517cda5fd7f1", + "index_sha256": "e6cafd8bb50e06de5b562e28628f5c80e47a40af868302d69586162d8f57fa27" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00016.tar", + "files": 303, + "samples": 303, + "raw_bytes": 37050132, + "archive_bytes": 37580800, + "sha256": "f8f11a7fb8b5ecac915fa7c56b3d1000ba85736dad30b64e3512d37d2cefa8e0", + "source_signature": "75bda19727255277596e1ded2de93bf715c7b52accc45a433895c66dadf6d239", + "index_sha256": "f8de1fb85883bf9eb174ff876ba930f3e475e5f7778b8920f7e8e6357f3cce73" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00017.tar", + "files": 331, + "samples": 331, + "raw_bytes": 39969764, + "archive_bytes": 40550400, + "sha256": "778473972e4c65dac29f6aa4ab4f3178acad333e56bccc31903bf898dfb0532c", + "source_signature": "f920e9df95e95de5586d1ee83e2e0568b01f3da465566ae0e1540bdf4112e288", + "index_sha256": "24d2bcfb034440c1a89f9cd8a24c183a1c7778d8e0f90b016b8c92ea292f7403" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00018.tar", + "files": 335, + "samples": 335, + "raw_bytes": 41923208, + "archive_bytes": 42516480, + "sha256": "08421ed611e54982a7019caa09dad9c3a0ee73f2bc7e93828f18027a56c91d96", + "source_signature": "ad3f2fa2005a408431c6081906443f6ecb6df8f419899a7d7e09e1bff5d2631c", + "index_sha256": "26ce2bb487bd686c36dbca8e9473c54dcc2993bb75b10b115cf4bf3c6c7092a5" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00019.tar", + "files": 332, + "samples": 332, + "raw_bytes": 42843408, + "archive_bytes": 43427840, + "sha256": "351c532d3fa29c7580e1389295a51009d67a7102efb089dc5127daae1f878609", + "source_signature": "075dc6c7823994c5aa5eee4ef7e1a60aefa64352d75af16fefef408f2d0e6622", + "index_sha256": "ee867c75a445b1f17affda2c3a70ca830205d3c9b1eb82b08ddd5e42855e8a18" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00020.tar", + "files": 307, + "samples": 307, + "raw_bytes": 39495108, + "archive_bytes": 40038400, + "sha256": "d91e183aaa8643ead7dbae2d8225f05579f2d5ab4386d315a25a5235b1eebe2d", + "source_signature": "0fc48141327a46e6da5fae5ea49812b53040c97836e5b93e97d76977e187b1be", + "index_sha256": "e556759e4b232c9eb9ded88518b8a1c2923b1ac69341e18d8286cfe3168a7499" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00021.tar", + "files": 324, + "samples": 324, + "raw_bytes": 41294256, + "archive_bytes": 41871360, + "sha256": "87ae2ed5a7df09e1439fe0c6016f440cf6388f19030af300f4b95e7521089153", + "source_signature": "c0f448ec2739b90883956014ef6de24744118d71f42c4dcf464edd6d547409a1", + "index_sha256": "56cc6af3a16555272c1eeb4f7967156dfc369995e88527c27478ca06d7a5ee37" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00022.tar", + "files": 283, + "samples": 283, + "raw_bytes": 34924452, + "archive_bytes": 35420160, + "sha256": "8813ecde7e8d1d2a76d47c94155b8aa99319e1aed985b14803ba4bcc58a05055", + "source_signature": "0f03c708aa1c785c9570dddefc377983a2ebcf7ff75a9ac3a744a16015e40c82", + "index_sha256": "593d93c5a365d7ee563dd73170edfb540c9b3f6996cb9972ce9bf11c0934f977" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00023.tar", + "files": 288, + "samples": 288, + "raw_bytes": 34982272, + "archive_bytes": 35491840, + "sha256": "2baf666f0d99ace2c70b993a3d86fa58a2e7167971060fd22450cf7601c8cfcd", + "source_signature": "492086838beeb0e23ea1a07697d2061b37fe1b0df03222f895c6f6805d53b6d8", + "index_sha256": "89d053064c566ab531e77999c948e7d1cdc8f68c28d10a61735fd1117c2c553c" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00024.tar", + "files": 362, + "samples": 362, + "raw_bytes": 45795845, + "archive_bytes": 46438400, + "sha256": "bf91d285b5d7caa4939fd7aad56758be2163c1501238f8dab56dad87e7bc9ad8", + "source_signature": "d6e3b2f53bdd50f16786fef9f0395197070d0c89871746ecc133112f5af80618", + "index_sha256": "bf31a06cee94bb75f09fc6c33098395669b5a87824c0a1abc45428e1f05d577d" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00025.tar", + "files": 287, + "samples": 287, + "raw_bytes": 37011028, + "archive_bytes": 37519360, + "sha256": "cee03cda337df0d03b086d27d890221ef54b99957269eb4d8d35c14218422dae", + "source_signature": "cf6449c99378851971200fefd1ea1d8f5c40d40aa7b8e9d747c9b6396de9e79a", + "index_sha256": "b0e573d486ffe30954dc3b6db1a4a273d73321fb9b4cf4e1aee770a8f5f89c25" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00026.tar", + "files": 313, + "samples": 313, + "raw_bytes": 41568972, + "archive_bytes": 42117120, + "sha256": "e914a78d22e8b9f1b87cbd0e8295f88fcab53c5d3727edbc874735ca7184056a", + "source_signature": "8dc8684fa8326224449a33d7e290aeb846821be278f6007ccd1424691e180636", + "index_sha256": "f7e52f62aac8d3a7b8ead61aaa5b998f74017e3c253c26d65de5f6d93d450a92" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00027.tar", + "files": 349, + "samples": 349, + "raw_bytes": 42556156, + "archive_bytes": 43171840, + "sha256": "6b9dbe8e9c05480799e806cffb727416d76767a6bc1674b5218a6231264e257b", + "source_signature": "6e2a5ad4d0925d917c5087264f1843a4e8bfdb9ff4ae2564c58a60f5d22f289b", + "index_sha256": "5ddd0e3811182f1ef27271d655e3c5ae1bc57cc9b1e76278865b594e34ee6386" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00028.tar", + "files": 334, + "samples": 334, + "raw_bytes": 39246696, + "archive_bytes": 39833600, + "sha256": "e6a113ba70b6d4eb0c756808db92cda84e22edebe2a2e901ba8e11e47b78ad10", + "source_signature": "2053a630683817baa3e69ba2f836c1e34ac4050f468e186c0699952f85a4829a", + "index_sha256": "2220073e187c23bd430951af51a24e3f96ad6ff46f199034b23762ce2699cf5d" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00029.tar", + "files": 331, + "samples": 331, + "raw_bytes": 39867364, + "archive_bytes": 40448000, + "sha256": "8426fcd42016303d4326b3baa4968318a311c09ba9779d528d97c68dc4969c70", + "source_signature": "e83302832e750a2f5ca414583ee54bd0cf99f5a19709a97e756c9d6a85b5ad96", + "index_sha256": "2a1457a886add1f6e967410591ed763e8c2fb64446a5d70b4bc1da8e928ea1bd" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00030.tar", + "files": 319, + "samples": 319, + "raw_bytes": 39726036, + "archive_bytes": 40294400, + "sha256": "22ec6bf5bd0ccccb93b11199d821601681a13473b246ca83d4e825abc5bdae6a", + "source_signature": "7fa1abd5c61ef3df804976228abd719db0d8d037fa8f4ba7b0bc3ff920b082d3", + "index_sha256": "ea864b3c9e9b7cf9669239b776aed86e4089dd4714c76353d6217390dbbbbf7b" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00031.tar", + "files": 343, + "samples": 343, + "raw_bytes": 42991092, + "archive_bytes": 43591680, + "sha256": "72bf55d2c6b2d983aee13b13db10ff85e277cb6c8e158ec2823351cf5e06ba8a", + "source_signature": "d0d5f67253af077e41f9b2cc3d8651f775dda5ea5ab670b7a927692e986d98b2", + "index_sha256": "af8b1adac06df238878d64789821afb02e39709ebc0fac9f8735b4e2e11a39e6" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00032.tar", + "files": 329, + "samples": 329, + "raw_bytes": 41160076, + "archive_bytes": 41738240, + "sha256": "4f283ae1a4d187da989098dd9913fcc8296d28fbabb37d4b1f702abc39b9b519", + "source_signature": "ab028b83ba618310ed16877a439747579b2703b5e67b3768fcc000864caa231b", + "index_sha256": "b709138a1559496905a14e1dfa258a4d20fe773cd5b0242f7a0d88f2c3a4ef56" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00033.tar", + "files": 333, + "samples": 333, + "raw_bytes": 43093052, + "archive_bytes": 43683840, + "sha256": "97e2c3783d67edf47ea1fc2a05416709727df703540322f0932dfead1841b692", + "source_signature": "9ebe84ff910cf0d851194848aca99fbe159c2b13dc3defe770ed0ea31b9335d6", + "index_sha256": "1008d3b65e13c4448861c50c8698eb208751386bd8e0f31a0e7c531b9b4b38ea" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00034.tar", + "files": 335, + "samples": 335, + "raw_bytes": 43413140, + "archive_bytes": 44001280, + "sha256": "945539f498ddb717abec6d29b8705024c86eff8e5dce0baeaa0011a5782bb2cd", + "source_signature": "2f54572da9e982077e14851018501796031bfdaeb75e20a3ab97408293e0fbf1", + "index_sha256": "a5a5fb382fede45e06f1da45eca468b5edf13eae9d7aa4c4dd8698b303898c5a" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00035.tar", + "files": 336, + "samples": 336, + "raw_bytes": 62958784, + "archive_bytes": 63549440, + "sha256": "4e4146a507f665b867f7ab0834e6057f54f34a9f29a794ee5dc99f6d1c2a11a7", + "source_signature": "dd85b92a161c7e499e36006a7affb4729e132ee1395bbde4dadeaa67db00ab9e", + "index_sha256": "a33763d421e61ba389c284aea6f3a01b50637e28e3b1ee457f275178bf421087" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00036.tar", + "files": 318, + "samples": 318, + "raw_bytes": 40129192, + "archive_bytes": 40693760, + "sha256": "45fdb33aeece62165daeb1a642ec232577f93bca909be39a3a73fcf9007e13ba", + "source_signature": "45e0b45a4c68dd50b3539b335840b3eea75b2f28e7af045bb984159e737e90ab", + "index_sha256": "83ad613f417a0ba4207ed327fac69ffee23a74e719cd9ad05601fb44b64a87df" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00037.tar", + "files": 302, + "samples": 302, + "raw_bytes": 38714650, + "archive_bytes": 39249920, + "sha256": "3cd66bd03656b3a12233e4cdc1c72e0012d5cdf6122cd441389d4b688b5669f1", + "source_signature": "2924a24ebf02166d6cb5146abe3f2383eab3f0c3478ee53880b36671a6a8a108", + "index_sha256": "c992edba846d88245d393fa1a1489cac43d04dae0a1a293571968ea31f67ec3b" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00038.tar", + "files": 346, + "samples": 346, + "raw_bytes": 43266424, + "archive_bytes": 43878400, + "sha256": "e3d82343e93766f15a91d814a29635af5fb043935b6551ff2bd7888e4efcec21", + "source_signature": "38d25188f0be0feff7f9601502ee06ae923c3ca8b601a15ecca5a966abfd6bc1", + "index_sha256": "37a09f61f600d6e4afd4c2a51b789b5b7561a8ee96758264bb9501aa650e89f2" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00039.tar", + "files": 321, + "samples": 321, + "raw_bytes": 39508524, + "archive_bytes": 40079360, + "sha256": "b8439316f65c8c9eeac2c006447887dc2f688fe6a9a81bfd40dcc7574302eb00", + "source_signature": "2f1c9fd02be888433b32b4fec2274725f45bbfa18fa92f86f7fab48395511c0c", + "index_sha256": "4e8b730f613fdabd2618deefa1689c2d505a7899cb010aa9949cf1e20e602094" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00040.tar", + "files": 321, + "samples": 321, + "raw_bytes": 39790124, + "archive_bytes": 40355840, + "sha256": "93715b6b0417beec03ab3f262ef21ab94fadd5e8bc20051cf9e91be2c4bda3d3", + "source_signature": "cd7207b710c0fb5848e15a9e3ab84f74bd68a567ed9d34bb20458ac73716d308", + "index_sha256": "930144eaaf03c02dbbee7149a4e918517018c1da2156a1c090048fedfe1caa54" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00041.tar", + "files": 320, + "samples": 320, + "raw_bytes": 40321280, + "archive_bytes": 40888320, + "sha256": "e2ee393fec34f04412b8f042a0672936759dc8e05d77f95c37ff028cbfce80f4", + "source_signature": "fdb954291d8d0f0eb254997143d066d2a78b71fdd8c3f282503bd3631f2080a9", + "index_sha256": "9b40acb73da9ddfb0abbe381949e4c144a897260083dd050adb94fa890eec059" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00042.tar", + "files": 330, + "samples": 330, + "raw_bytes": 41204920, + "archive_bytes": 41789440, + "sha256": "da97f8de1a9e03bacb919b1a8a0a6ef08a93276d4678d7b85f44356c6ee60855", + "source_signature": "2da9ec0824e9cf77c86a961798b1b35deb8f884f9aaa4f1c3eb65a465d3780b1", + "index_sha256": "589a50845b3ac6ced0b2b0952f0bc2f70c8a1f4c32e22ac25799414d39980515" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00043.tar", + "files": 349, + "samples": 349, + "raw_bytes": 43388156, + "archive_bytes": 44001280, + "sha256": "fc6239a33b3de72d7493735536b3b56e992a1a46ced8896ff69937bcb51111d3", + "source_signature": "deaeea2886b684ca114df0d2943e1906227ff673234f5149661cd6c96e37f0d5", + "index_sha256": "e8c2b46c6b45862754b85d557bb441cfe819ac3bff9597668104ede64f2e8d0d" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00044.tar", + "files": 360, + "samples": 360, + "raw_bytes": 46633440, + "archive_bytes": 47267840, + "sha256": "8b52ae33f153c1e151c89879a721d56ddff7f8cd3a68b7479dee2e3485e9b3c7", + "source_signature": "b5a23fa5d32126fda27cd0f3239d3a8fa0c727e1ea559d5b3eded74b9b1b75f5", + "index_sha256": "2f347e88b1037f3e52280a91b87398fe0357f7d7e148814858f80280449186d8" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00045.tar", + "files": 293, + "samples": 293, + "raw_bytes": 36512092, + "archive_bytes": 37027840, + "sha256": "0f8467a200b004acd9611000866e5405e1e35c0ab28a331e1f1556fd79f3c3cb", + "source_signature": "c697828fe3a333ae1e6929b0013d4921c753c08c91e377f881d9f994f7948a4e", + "index_sha256": "037999e1508e36d9580f8bcd1d82671e95d58a0bca0a4998886b8151920152b4" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00046.tar", + "files": 351, + "samples": 351, + "raw_bytes": 44117844, + "archive_bytes": 44738560, + "sha256": "12c02c76fe1f4f4a01c142adcd7527c099d92597af940162ce035abf1a1bba36", + "source_signature": "e1d768c10c527e9e07c3b11875b379cbeb5fba4394d87a65f1895eb735efa746", + "index_sha256": "4a1b54eb7bcb8987a6b15a7b83e4668ec1d446ea80bffb9ce59e3fbc7b72709a" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00047.tar", + "files": 337, + "samples": 337, + "raw_bytes": 41134828, + "archive_bytes": 41728000, + "sha256": "602e4c90756feb899a46ea47185b94eb4c6eb8ab71f0e1229683d335d1304532", + "source_signature": "37962b7c344e1816b70059a6098d477d24870f9e64bef0918bd562c7e3f28a34", + "index_sha256": "eee0cd2aa7dfa5dd5a1b72de87003d944c9ed4126b4499590c5b9659fe3aac27" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00048.tar", + "files": 323, + "samples": 323, + "raw_bytes": 40775812, + "archive_bytes": 41349120, + "sha256": "978ff0520d79d4a710286de4de5362fd48d6142c59f7d679459c4507f5e29fec", + "source_signature": "d1bac0b904b13065edc41f3e4eebf724b31b0e1866d258b6b35a84f3c0602640", + "index_sha256": "3d9e45e29077bb7dcfb2b65cc7c0ab99ea4e3655c39cedfdd1e546caf5a9a193" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00049.tar", + "files": 349, + "samples": 349, + "raw_bytes": 44322556, + "archive_bytes": 44943360, + "sha256": "45ce369858b176730b69e5a3764bcfc09cd155c37cf07721dbd3a5b0d3bb9939", + "source_signature": "8874148005e51d487cb6379087d0cf5be9458d52aabb2ca907e524ee27943d4c", + "index_sha256": "eada85f482d19e1d858189ed3ce286605b3a04a32c33113fc5bd45e0f62b3977" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00050.tar", + "files": 319, + "samples": 319, + "raw_bytes": 38259847, + "archive_bytes": 38819840, + "sha256": "2c7e6761a7a5986ecb3aa53d58ead04f7311e5c2dc15f72adaa67dbef13bb560", + "source_signature": "9bc681261ab40191e48ec449c45b7bf1c2fcbb43ce931f5a44fde675fc75ac90", + "index_sha256": "799020da0e618c1cbd6ab607d59c747aa6e72ba7f604c53205d6d90de8b29934" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00051.tar", + "files": 312, + "samples": 312, + "raw_bytes": 38196128, + "archive_bytes": 38748160, + "sha256": "5639f7fe8f0ac4f6b412ad180327f9e1aac58a5f560e6a72eec416edc0543c90", + "source_signature": "38be6691770fb13a005ce64f99123690d181fa0bcb2b0002bb12b626060c8fbf", + "index_sha256": "22bb929db1e3c05206603a2d7d5f1cc123c5f0d3dbfdff347eb5d7059acee069" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00052.tar", + "files": 313, + "samples": 313, + "raw_bytes": 40288972, + "archive_bytes": 40837120, + "sha256": "a6b80039645a4e5f973e3617e5997885a2463897a7529240e0a158891558d27b", + "source_signature": "e8209a75f3cfb0d252fbad9448fad41256f51fb767ae7e605ad4f794472572a3", + "index_sha256": "17759180daf51352102d1c7136946cb3547432fa5e0ad625acc91e955c248c73" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00053.tar", + "files": 319, + "samples": 319, + "raw_bytes": 38689236, + "archive_bytes": 39249920, + "sha256": "eb9646a3a6e8295b74a482c5937a07e2854ca740d9b4efd38d82b7cb3ba72ae0", + "source_signature": "d354720ffb26838e08936e2bdf59ceee5807430da92d64e25f3848863ef24c1d", + "index_sha256": "241e34a89e17d694a213fba0c7c5ac3ad17c495b696b92d27bbc63dd6411bb9f" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00054.tar", + "files": 311, + "samples": 311, + "raw_bytes": 58240884, + "archive_bytes": 58787840, + "sha256": "43cb1b9bad8d93c75f55ee64f497dd4524650b4e3f5f10a304b31fc55262da1c", + "source_signature": "b2083bcf85dfe2a6cf2e56262b82e4aca7c3f59a3dfe6f31449f71bea2ce5425", + "index_sha256": "4065fe4c34c8d1e605bdcb5b2fdc63098bf4b2cf1f67fd95acf690a9a2e10c33" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00055.tar", + "files": 319, + "samples": 319, + "raw_bytes": 40980436, + "archive_bytes": 41543680, + "sha256": "594293e2b454e020f2241ef35dc4d54083c30e1e68be7e36168e851bda4b7bb6", + "source_signature": "8245712da6bf06b0a4430bbac15f87cabf4ddd2dd9c1a3a4dea8f4e8b19a6279", + "index_sha256": "ea05c4000f42dd3bc6777571cdb259d2db6f91da6f21c37588e938cac8e8d66e" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00056.tar", + "files": 317, + "samples": 317, + "raw_bytes": 38996348, + "archive_bytes": 39557120, + "sha256": "c0184deee534d40d20c498e9a54740046ded3a977e0c888883b285de25c51231", + "source_signature": "6e0263e95e5a996529e4d3508f0b62d1c07c9de3a10efb0b5cebed4ca93cb43a", + "index_sha256": "8cbabe12f6abbb7f0ce182ff00081afa5cee4b340fe2a6011a485c28ee83eac9" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00057.tar", + "files": 310, + "samples": 310, + "raw_bytes": 40052040, + "archive_bytes": 40601600, + "sha256": "2bb72b2f5a79632475c7e91294d8192a0f101ff6c4a3f7813e64ebbf902076e4", + "source_signature": "9899d7585c26fbec963f6a06e0abef5071a50767fc1722345890945794608ba3", + "index_sha256": "cd1b03c2d9285344f9842a55bdc0f75cd6963a0c86412d394b8283eab2262708" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00058.tar", + "files": 309, + "samples": 309, + "raw_bytes": 69572646, + "archive_bytes": 70123520, + "sha256": "969c7520b93ad5e36cbc169fa0b0fa699245c62d09d7ee82705fff436bffdd8f", + "source_signature": "0d163a7ac9b25508d41bb413cce3a7822548b6b998ed6442f293e579fd9c08dc", + "index_sha256": "834dd735c6a30d214fd70a503585c2f9f4d8c993c223e650cb1fddf4130fab09" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00059.tar", + "files": 315, + "samples": 315, + "raw_bytes": 38881060, + "archive_bytes": 39434240, + "sha256": "bffb34ec454ca4dd677264b2da5d53976c0db61862b027ee2b7c1e66895c57df", + "source_signature": "a2bd2853034efa8f4e1fc14bf3ad68fd1c317c3d75ffff1bd6424de99710db62", + "index_sha256": "7c8afccad568ea986e2082cbe9006c907c50f29b05775c453e8374c17879d2c1" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00060.tar", + "files": 319, + "samples": 319, + "raw_bytes": 40353236, + "archive_bytes": 40919040, + "sha256": "3b192f005ff5cb00e235a4e0a7f2eaa8f1617ee259589516985932f327b698ef", + "source_signature": "9ec3cfc9e835a45f142a874362c4b8a875d1c3e23ca7883727082f575a47dee9", + "index_sha256": "3e7772718db855216ac541c8935673341143076bad31529d170db41f53511de4" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00061.tar", + "files": 329, + "samples": 329, + "raw_bytes": 39304076, + "archive_bytes": 39884800, + "sha256": "20f155fadb145c348311681cf82acdc57f478628986fc42f44295866a3b7eb3c", + "source_signature": "b9695883dfdd6cc8e939b60501da215bcd2c367ebf5876e2d84790914266e966", + "index_sha256": "da08ffafc484cd848ce9db4ced2ca8c5d7ae86eadc384a562391fa6e6e75fec3" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00062.tar", + "files": 301, + "samples": 301, + "raw_bytes": 36627644, + "archive_bytes": 37160960, + "sha256": "32f51e566fbbebc7a89ed10f37cc4162913362cc4ba1b8427c3aac99f837f8e8", + "source_signature": "9a8186cc5d0c284fcb9f2cfdc16a341b3333d6d2dd949ce32d9366bc6afa8e49", + "index_sha256": "56f23cb6266c570002bc05a4f274e1765c1f4dab0d90e68637f9474c2fd3ddca" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00063.tar", + "files": 315, + "samples": 315, + "raw_bytes": 39047460, + "archive_bytes": 39608320, + "sha256": "b6136c399c8b46a1e14f84341c3fa37c53747df052713139f4519ea3fe10a4fb", + "source_signature": "ca711aca8300665cf22ab1445351852b1895c4fc91ca52be548b4755d815e0cf", + "index_sha256": "f8d18d4a1df61626a7988ee528d78fb897163a53f22bece963e53e386f9ae99d" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00064.tar", + "files": 369, + "samples": 369, + "raw_bytes": 46947436, + "archive_bytes": 47595520, + "sha256": "71fcb02436ac976d21d07022132df2951852b4eb9a7fbb0f8b2263a991968dea", + "source_signature": "ba6e2813fe074e4606b2dd5ad07eb1fbb1b51e8606acb269e997bfd9e1ab0804", + "index_sha256": "af4ba15a5ec342f6ee7750546636fb8fb94358a70ee6a877ac24648cd1023816" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00065.tar", + "files": 316, + "samples": 316, + "raw_bytes": 38593104, + "archive_bytes": 39147520, + "sha256": "18f884d8b5a21d7b960650ae238900897beeafae85918431ea7dbdf8e07cb463", + "source_signature": "1c88be333f9f2e39241ee4f18d3a2e07f821810daaace357daad27b0d85ad4bf", + "index_sha256": "f1d050df3da97d45d7369f746959800844e62c09ed86c519609e2a2dc775af7b" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00066.tar", + "files": 352, + "samples": 352, + "raw_bytes": 45813888, + "archive_bytes": 46438400, + "sha256": "e01646ec62dfafe4d711a8e0c6b49814a984e05b31ca6853cff5db3e03f39144", + "source_signature": "ee65be9d60a140a9e03b5aa0fd7a14068d43ef0b04246eb8f3a4310d1715e691", + "index_sha256": "5cdd1bb5b7e1e5c9b2b2838e87f6c7a79909dd65b95ab07c9095f5a99436f6fc" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00067.tar", + "files": 345, + "samples": 345, + "raw_bytes": 48776780, + "archive_bytes": 49387520, + "sha256": "f16fac3abd419f287b53249993a0f0c1637f7df24e8bcf428cdd8dacd16309be", + "source_signature": "fd862857b1174bc3689d540a6cbaf926c4d893282428ebb81b79edc18152c80d", + "index_sha256": "18dda576936bb334245e24e97e5e5accb8306ca3abfd2d53f29d011791a3c3aa" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00068.tar", + "files": 335, + "samples": 335, + "raw_bytes": 41518740, + "archive_bytes": 42106880, + "sha256": "6beadcbf216402f3b14c3e2770f00b1a07a0923d1ee6fd0f853f83f494bbd335", + "source_signature": "6f1e9f90084c66a7a77e6b23c612c86317d1698307cc440f31ffb5b7b2cc82ab", + "index_sha256": "a2e8c0867ea302b697aae452cf8e85e5cc8a71b4d5d741e97d3b274d3461d33e" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00069.tar", + "files": 296, + "samples": 296, + "raw_bytes": 38029024, + "archive_bytes": 38553600, + "sha256": "abaace22b7c89b32060f802742fe03e895849204efbb1ffcdcaa55096d8d79de", + "source_signature": "eaa7e32e3468c6be7d74d017cb81aacaa6ef64e77ecea666891fcb89b20b4711", + "index_sha256": "2edf42619dad07bd4b2ce685a8235686df3985fc89526dd159563409b3ee5dcc" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00070.tar", + "files": 346, + "samples": 346, + "raw_bytes": 43381624, + "archive_bytes": 43991040, + "sha256": "223e4d79664172045ac043581c4d7c76b2c23250134749cbac63b13a07da36ed", + "source_signature": "2d0c35f3675a1d0a27d61c7433c83632e6ea3522acfd8a1a25235775f1941656", + "index_sha256": "dfce06baeb37ae9420cb1f3f8494a15b98af37effee00230ea1161feba6be15f" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00071.tar", + "files": 350, + "samples": 350, + "raw_bytes": 45301800, + "archive_bytes": 45916160, + "sha256": "01649b88a4883a71207105df06d4fd18acb5a7109a4085ceb643921fb42b151a", + "source_signature": "2b506f16372d2a49b15852790d5ad8c0a73358d9c4ddf40728761394c94a40f4", + "index_sha256": "6bca05fcf46653c550673e8d520ca5f907a78a077947d148464ffcc721f78dbf" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00072.tar", + "files": 342, + "samples": 342, + "raw_bytes": 42549448, + "archive_bytes": 43151360, + "sha256": "448343d454f4e05f61cbe1161623c7e1ad45c38c00361bab71f356b0d561cad5", + "source_signature": "7435b065e874938f8617f74ab2cc98d7a01c3790a601e892266a685867f56a92", + "index_sha256": "81d02d1cd9fda8f6ad43e9a0dfb96519d7d2faa699b3d1f867729d9e838145ff" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00073.tar", + "files": 304, + "samples": 304, + "raw_bytes": 38771776, + "archive_bytes": 39311360, + "sha256": "d098cc02497e16396f28283a7d3586a345be5175fa7ed644bc865cef662c2316", + "source_signature": "604a0c4c483e7e9448d6b7812aa243c1b7bd8bf932e0afa394a9cdea5a731737", + "index_sha256": "6ae6ddf36cff32c1d8b075eb4f56e9567fce8599399eacfaf410671b7c90fffb" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00074.tar", + "files": 325, + "samples": 325, + "raw_bytes": 75310585, + "archive_bytes": 75888640, + "sha256": "d071f984dc13e707aaf2684f0db7d6b19f8f9be39c831aa7fe7143fae446d18b", + "source_signature": "6948a02444accbe8665d6574d1fe81432d57a7b95753f7ad6b8e1a138130b3c7", + "index_sha256": "96f74d172bea75fa2bc8f06de38e68d0a06275543367e33f39e9db557295195f" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00075.tar", + "files": 335, + "samples": 335, + "raw_bytes": 40827540, + "archive_bytes": 41420800, + "sha256": "b4092c8b245b8000ddc54595626c97f4f1461bc3aaa4038cd15081363f62963e", + "source_signature": "856b22028d04c825da535e78ac54385b43bbee53cf605494bfeae566a6ed256d", + "index_sha256": "16f53f26a5343aa779f62d28f0e07e2a2dbd99b27529f606ebdddc43663f98ab" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00076.tar", + "files": 344, + "samples": 344, + "raw_bytes": 42434336, + "archive_bytes": 43038720, + "sha256": "a19e9986aa349d9afa65ef740e4cea5e8d8b7e5c4cc42e1bde3cdd434d2ea0d5", + "source_signature": "3a1d237f8c6c13e5334b8c4700f2d8e46f41f769a2192f32833452630d27ea7f", + "index_sha256": "216b6e4f47660a79238fe361dafc23b5899a5511edad8d93a96f4e2f235fd699" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00077.tar", + "files": 318, + "samples": 318, + "raw_bytes": 42138792, + "archive_bytes": 42700800, + "sha256": "9239b7655b44b17bf87cbb04aa455df6684765b82552ec80dba6bf0113c7d74d", + "source_signature": "dba543435f7ef3e461062a3e9e3557bfeb096b2c4cb9096e288db8c3b0a9f933", + "index_sha256": "40cbb5bc52421b328025bfffb1c4fe67f4bfa4acebd843fd5813d0429b546920" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00078.tar", + "files": 329, + "samples": 329, + "raw_bytes": 43729164, + "archive_bytes": 44308480, + "sha256": "f2d3f57be0d1c3ae2881496d54e38ab176fb4c57630a149a5a883f5d6b3c1898", + "source_signature": "26c524c6cd79fe5180d8cf01662172f7c84158b1de41d5e25e6469cb6ec7c07e", + "index_sha256": "4a46427d1f26a0a8820c7bea10c47b826d99032059a841274740c8d37b51668e" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00079.tar", + "files": 319, + "samples": 319, + "raw_bytes": 40481236, + "archive_bytes": 41041920, + "sha256": "82bac1648b4b439109310ae93392ffedf25b028be1f5d3970441692724493c77", + "source_signature": "6c16ed2eff766bf0037aaa323024c9f91095bc4969222cd95e1f6cebe5393a88", + "index_sha256": "e9d708ab295124b230c3b64bbea4cf8f9ba191633fe85e706de37f602af5e702" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00080.tar", + "files": 324, + "samples": 324, + "raw_bytes": 39566256, + "archive_bytes": 40140800, + "sha256": "d07687b31500b57ff57082ff4363c24b63f811a379da6fe894f4b3ccae52a458", + "source_signature": "90d2ce5c4456663dfe537b80d6c110b1de5b5626cd38714352c312a563a4ff5d", + "index_sha256": "fc754c34e2754cfae9b14f967565c172f7ece26a8d2ea0343f2ae594a7beba51" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00081.tar", + "files": 356, + "samples": 356, + "raw_bytes": 45506864, + "archive_bytes": 46131200, + "sha256": "fa201091f166f5ab939f633edc3e01858f99f07a4811a49525edcc02ea31eb44", + "source_signature": "890e6cad3312619234a1c053a80bf38ed53e739902aab72c7a2643d9d09467f7", + "index_sha256": "cce07749dd8a667571080bcc9f7da71b92765430611bc222dd1bcf0cc1985e24" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00082.tar", + "files": 322, + "samples": 322, + "raw_bytes": 40705368, + "archive_bytes": 41277440, + "sha256": "d29bd533254574772d4f7a7243f0bca5ea3e8b9046eb7dbad942ac5da22c90c6", + "source_signature": "3de1d64eb73b3aab956dc0b369465b84e10183de0bf000d6a60b49622e76f6d4", + "index_sha256": "435801d3db4c87d00058176080cad18ec4fb5549cb72770d5763d1333a3a4199" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00083.tar", + "files": 333, + "samples": 333, + "raw_bytes": 62273852, + "archive_bytes": 62863360, + "sha256": "923f03cba37138f88709d92ad6ed98d2d056dc1e566602b54a02106288d289eb", + "source_signature": "dd26a614366ef046ef75eab1125f9e2a1b8175649774910c0b499e04afb1b188", + "index_sha256": "f481ba2f4e349c5830e0ac6c0f5fce4e804ba4c687009c50ce920780f7be1d5d" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00084.tar", + "files": 323, + "samples": 323, + "raw_bytes": 40251012, + "archive_bytes": 40816640, + "sha256": "ed803810e3e7601bfb4fabe2e48ca43b75d0335d16eb6491aacf3e888f756c82", + "source_signature": "d3d5c607d1705ca4ca7695634a44264c7fe5d5af3184cb36fcd59826958d38ed", + "index_sha256": "d1026c62601cfa59dfd2eac7d47a909426414b594af423e2985506b0c382891d" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00085.tar", + "files": 317, + "samples": 317, + "raw_bytes": 39521148, + "archive_bytes": 40079360, + "sha256": "14b31ab861b555c49570309a242df0e18493ccebce09877105167420d69ef7f9", + "source_signature": "8d188fb2b8f8d1a95f57aa69e3fbdae79520549e83ffbf3f454739d61aa32e59", + "index_sha256": "a46bb408dc735dd6c264ea3889e905ebb2ed649bf7ed5661abe4a0af8b7934b5" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00086.tar", + "files": 365, + "samples": 365, + "raw_bytes": 46128060, + "archive_bytes": 46776320, + "sha256": "05231c8bcad79c362e9a6ae1ff0dc2fde89c7dc7a106880e97a8d9b1df6e2c64", + "source_signature": "e2a1146f50d1633417e82072d55c2707d3b821198f85d495027834043f4c5181", + "index_sha256": "bebf5cda75dd8cfa67bf25d733c128a57b424d60510d826478a8f29527f1c658" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00087.tar", + "files": 324, + "samples": 324, + "raw_bytes": 41895856, + "archive_bytes": 42465280, + "sha256": "d53a7f427f9c3dc2effe79789be216bbf6311258a9947bef2e341a75699adc3d", + "source_signature": "1173f0588de659918bbd1396ef81c3b5244c2f733de9bc3789ef3ccbf5143800", + "index_sha256": "cc13a097c51c360dd5f942622efc935c942a4333490487ebaa29395a8111eceb" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00088.tar", + "files": 335, + "samples": 335, + "raw_bytes": 42965140, + "archive_bytes": 43560960, + "sha256": "98a82dabf77bcebbc01fbc81894654058f67e680af05ed709bea3a0dfaf2839b", + "source_signature": "7359d73829a976d7e846779d512dc91d94076c8cc899084d8591423d06a54692", + "index_sha256": "c82c60c93488dbee5f2c921f28188c082c90bb74bce142c055bd8d45b2ecc12f" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00089.tar", + "files": 329, + "samples": 329, + "raw_bytes": 41646476, + "archive_bytes": 42229760, + "sha256": "4152bc085746365052fe54dc3346c3b9780320de6d42ee1028c404a81c7df14b", + "source_signature": "26e413dc3a3f7f57832524e45f60fce240f7bdea276f423cb5e1a2e33e6764d5", + "index_sha256": "cf8ab52c65d76e7d2b84e0cf81f51bfd5f5786fd47958a8daf9189ae64c74670" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00090.tar", + "files": 326, + "samples": 326, + "raw_bytes": 41921544, + "archive_bytes": 42496000, + "sha256": "23923f495694833ebc9bcec4805846f3295a2add1dcdfe3b38f7014891840c3b", + "source_signature": "0ec7cfaf92ca46c74f5809281a19e2003bb57d299a9c363089b0d4ae02bc7409", + "index_sha256": "9febf73e364d0eaf5631891a62c6d822bed83109570d73165ddcec47bb57754f" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00091.tar", + "files": 306, + "samples": 306, + "raw_bytes": 38119064, + "archive_bytes": 38656000, + "sha256": "6723d8f0dcde70cd25451836309d135edf9e481a8a8c827bfd8c66bf9f0d8d8f", + "source_signature": "8df0fcbdadc95ed31b350f8838f15cac541f4d9268a5b1e033c81147aedd4f53", + "index_sha256": "b95502eb31ac11f007258e3da6a5f5c68a13bb288dd056cf70b9feebd128f1a5" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00092.tar", + "files": 313, + "samples": 313, + "raw_bytes": 40570572, + "archive_bytes": 41123840, + "sha256": "49559f0c42628020a5f0d96ad9296321d740ed3c20c3ccec68b091956a5bbf41", + "source_signature": "96bec28855c47436b719fb29e9376bc9f684c3dbf8afd403cb23a06cc21a5826", + "index_sha256": "7277aee42b7f83c813257dad0338d7db7d0963a98ab088d918bc70d6a1255a54" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00093.tar", + "files": 312, + "samples": 312, + "raw_bytes": 40602528, + "archive_bytes": 41154560, + "sha256": "a48ad37706aeb963e9849535d2c86c46bb6dc20e5afb4571e0a755488da1e950", + "source_signature": "10be6063dbf3f57322525e49b5516f15cb8b39999cde8027e9305d51b4923017", + "index_sha256": "2b86ca18204ffc3307d694b73dcdcce649853a29521a2beec778fbd25fc0061f" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00094.tar", + "files": 335, + "samples": 335, + "raw_bytes": 41557140, + "archive_bytes": 42147840, + "sha256": "c9da41d6a27f32a8ea4a93e890cc7f0f175b89e8941e5f1fd1e7a2ff5a9d61df", + "source_signature": "23a0890819c98aa11727a6d06b341e9306d7f4d373d47d0f91d9c39c47ad014d", + "index_sha256": "986ff1860b460b5ee42ca056ea28fff89893081d74f6bc71a91c74759cfdc8ee" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00095.tar", + "files": 350, + "samples": 350, + "raw_bytes": 42481408, + "archive_bytes": 43100160, + "sha256": "067f2fd10ea207b0fcf6532518f9340142f0d2ffe6655adc48ea913273f58277", + "source_signature": "9141c3787d775242f576db9eabef7708f07114e9936e1a08cbbf84777265251c", + "index_sha256": "f8fa2fba311e592596288369dd9dc5452d2a3207120c1227189c8997f6f2187f" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00096.tar", + "files": 331, + "samples": 331, + "raw_bytes": 40712164, + "archive_bytes": 41297920, + "sha256": "89db4237b791793c89857aaf2d4f3383bf60a1f117d0e4f7923606c2c35adb41", + "source_signature": "5e140a355eacc0b99ad461e876dfe6407a928024b49ec52f761991c8c7352517", + "index_sha256": "584de8152092f99e00681be586e3b9edeeac1e901b4d056448a558ac0bffbc0b" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00097.tar", + "files": 298, + "samples": 298, + "raw_bytes": 37133112, + "archive_bytes": 37662720, + "sha256": "b2b31d169c73febb3bc57f4edcbaa1ae151ad3789fb0ba90f460102893e87ec8", + "source_signature": "6450930a29c082fa21bbae67283f01f738e0ee2216c5dba4620eb14f040937b3", + "index_sha256": "1163f34928126951c9b3051b0781fa65fd1d4f8e0fef579630d45a6eca7027cc" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00098.tar", + "files": 301, + "samples": 301, + "raw_bytes": 54502844, + "archive_bytes": 55040000, + "sha256": "3d2519e4566a715186efa5eca08556c3c46e592e8be5c64df8da44aca188261f", + "source_signature": "2194cb6b45cc9a2028dd5a0e7c4bc676fc9c18e5909386e2dd4036083fca7f9f", + "index_sha256": "9cfcc9ccd7186dedcdd9ab03b9b52763951e13edea9ce9a36173118c899af9cd" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00099.tar", + "files": 326, + "samples": 326, + "raw_bytes": 41460744, + "archive_bytes": 42035200, + "sha256": "ece552defec9edce614cc97b2da44edf9118a029e566c8c9af48fd796e68ba9a", + "source_signature": "a2ebf70272d8f0e2dc2731997e035d87e530ade1bcce96b540c6470f4b934876", + "index_sha256": "a4850f3a9ad6d9485812d0d8668c41926550883bf02743d6155ec73c3fa436a5" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00100.tar", + "files": 318, + "samples": 318, + "raw_bytes": 39117992, + "archive_bytes": 39680000, + "sha256": "43232dde4494ea3aa9bac35824c19a65a2784599292a601e0c4ad0cacfa89faf", + "source_signature": "214566bafc3302f58954e7a9ea1d6209af32957adbef1e5ff249a71d9b7409b1", + "index_sha256": "f56977684bf69a58cb0e8eb526fc68214208434c654bad1e797d0e6f316660dd" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00101.tar", + "files": 327, + "samples": 327, + "raw_bytes": 42222388, + "archive_bytes": 42803200, + "sha256": "1990375066bac9ff97a52db8eb52f11f83bf2e210482a2749e29af3232b5410b", + "source_signature": "09d081b95e232c1b6f62284644b56a628e422b12986555981e6f8572fc5dac5b", + "index_sha256": "0357fd7fbf62779883ad70b53c95aae3e9efe3768699c6959f58231cd3403d0d" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00102.tar", + "files": 316, + "samples": 316, + "raw_bytes": 40077904, + "archive_bytes": 40632320, + "sha256": "789ef419af768dafd3a278d2c22576144206da2d039151c41a9af2a30af379e7", + "source_signature": "fdb84967d203b084012f4532b1b88ab092cdfd64ef01738ad5a992e15acb8072", + "index_sha256": "64929865485a80234bd40229c1e894ff90e65efcbc870ce2f6cfa18ddf108a42" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00103.tar", + "files": 349, + "samples": 349, + "raw_bytes": 57376703, + "archive_bytes": 57989120, + "sha256": "eeb72aa4f1f304d45e4a5d12509ec92441ee86af984fe68d6397084921d88157", + "source_signature": "a346b472df75c70362a10a946e9a6b0ae37cdf44e5b75c3734ae57fa78228049", + "index_sha256": "568b996d6a2d100c935184b47fd59a13ab6e7a7f8ba358bcd0091cec27ab69af" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00104.tar", + "files": 332, + "samples": 332, + "raw_bytes": 39425808, + "archive_bytes": 40007680, + "sha256": "4e726608043f83c504f8fb4f26672d3db1f6c65bb1b02048ebf407b9ccf30de8", + "source_signature": "f6c5d792a20cca49d0e6c608a63c44cb81284a2434be7771ce7bac1d4117b675", + "index_sha256": "70cc9e198f5217ae723af2abc5c072419e292327fa58358075556c0f028bd5b7" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00105.tar", + "files": 315, + "samples": 315, + "raw_bytes": 40289060, + "archive_bytes": 40847360, + "sha256": "3556fe5e90aad3a199297564e49f252645c2141d60945e724a9bbea9a375007c", + "source_signature": "a8754d8b2b4d12861e37a3a36b4aa396468c5e1acf8e44dd845528f064fdbf4e", + "index_sha256": "4a8b1419981c67940c58067a6d0f28666c599ef3234ac6becfd0e3d3e2ca9bc5" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00106.tar", + "files": 324, + "samples": 324, + "raw_bytes": 40027056, + "archive_bytes": 40601600, + "sha256": "997a200db4a74be3c2074b938f30018499c26f7b6042b8a54557d933f4f78293", + "source_signature": "e12392934c99617bd4c6a15dbb638fb3d020386023ad19b30fc6d61e698543c2", + "index_sha256": "05730faef92e686b14baea6672417af10aaef326a10a460ead01507ced9901cc" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00107.tar", + "files": 340, + "samples": 340, + "raw_bytes": 42805360, + "archive_bytes": 43407360, + "sha256": "4eb564a14ff017642b519d6c28d939a8ee516a06b46d235a55142028844b5dc0", + "source_signature": "a9bffbdb7492c91f0bc0f36b1c8c18c50966701dd044f7ea7832557d493a9e2e", + "index_sha256": "b4516b13efa7a70e891458989a721681a5357ab5e7cdded8b1efb7bbfaf2fcef" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00108.tar", + "files": 291, + "samples": 291, + "raw_bytes": 36537604, + "archive_bytes": 37048320, + "sha256": "76f443b78b67068156e8d67f4910d787733e49b70ed2d1ac58f9193b3fb3b1e5", + "source_signature": "3d1f39398a245e68ac5c43f16af77b792b340924ba78f4bbd0d773c8159a4de1", + "index_sha256": "7e9ad3c89ed37bbc3ed6d1ac25ab72cd4f23a6d660f3ee1fde33c3e8047014dd" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00109.tar", + "files": 325, + "samples": 325, + "raw_bytes": 41723100, + "archive_bytes": 42301440, + "sha256": "ce06f2e54a0c327a265f17c40d54153e05782713abc839cecd3876b22d51e8da", + "source_signature": "b31da148a97603e0b6ef97d9a3c67f3dc17a8e9067e7ec504128ed4e5d750149", + "index_sha256": "60a3451ada4777c9bdcf233370eb786dd59bb5608042d75778d39ec06c595d8b" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00110.tar", + "files": 341, + "samples": 341, + "raw_bytes": 42184604, + "archive_bytes": 42782720, + "sha256": "0470ed9e94514ca7d8bdc28a65588fbb7ce5b0902d59f37ed1ece4cf332e483b", + "source_signature": "940fddff92964d26e16e915870d647203953f53c32d6e21da9449e02a376461e", + "index_sha256": "b36ff9aafdf8d94f0732f47c83fadec6611a7ba52b13eabc1a042fa1632fbfbc" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00111.tar", + "files": 343, + "samples": 343, + "raw_bytes": 43451892, + "archive_bytes": 44052480, + "sha256": "46f810f6d5ad672c2f29711be6e6676febd036f453b0084caf2ae1f110aeff07", + "source_signature": "15814bdb3e59243dea4967fd283095e2ae6b37e664a7a43e7b52af46addfea8c", + "index_sha256": "4719c40eaaeb68491970ce19eae46fe81deec51eb9878fe23fc7c9d65a04830f" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00112.tar", + "files": 347, + "samples": 347, + "raw_bytes": 42671268, + "archive_bytes": 43284480, + "sha256": "e0bcc570e0abc5af3bb1296565188f0f27fbaa703e9d32cb5e7118a5e861676b", + "source_signature": "1616755c6d13cdae4f0e4d01f0a36c099d3246de8f1b45d782abd680535f89d6", + "index_sha256": "cb8163f55062055a185d473e5e66835c514ab6962f0ced47347fe7450cdfae3d" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00113.tar", + "files": 327, + "samples": 327, + "raw_bytes": 39764788, + "archive_bytes": 40345600, + "sha256": "aeca80e0e283c7af5198c52ce88d2d7439a517f279f0f218c5a9eefc104da0f7", + "source_signature": "7ee820cbb076cd3b81274f212d9a9c7626a7d192272c5b0cba527b1c55e5caaf", + "index_sha256": "9edf9b95516aed5794bb9af402e9e31c4e063e4803ef160c47847f42d31a55e3" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00114.tar", + "files": 345, + "samples": 345, + "raw_bytes": 44258380, + "archive_bytes": 44871680, + "sha256": "1a2b8a51cf49a512c4355875add743c403b1944336e66a8a5db69b6421ce734a", + "source_signature": "8abba4c589f5604d47641e015664a5f73e59811f9532ceabb6e91e0b6313fea3", + "index_sha256": "deb37c55e88a786ed3475cd102d1199edd29c49855e28cc95f7d9d1932814ce9" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00115.tar", + "files": 320, + "samples": 320, + "raw_bytes": 39694080, + "archive_bytes": 40263680, + "sha256": "3f1bdd4d33c18269283cb5a9a9eb9f1a727abd41348b3d2120e92e15f54b567a", + "source_signature": "c71bcfeaf87d99a6002a250a470bc7d0b3f63a355a06f460e06f2a8e61ce20a7", + "index_sha256": "1776f9712c695bca698d32cdc963049bc5ddc25b1fd1b0aef90b5859d8047c5d" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00116.tar", + "files": 322, + "samples": 322, + "raw_bytes": 40206168, + "archive_bytes": 40775680, + "sha256": "229db215622b0b1880abdec66bdac1b8a996ed2f9b7889e5b93df34f74dc53e5", + "source_signature": "3fcd156e0a1698f8750485884c14e62eab67564cc8c3583363614a2816d4a9ee", + "index_sha256": "da46e178c1b12881e10590884333351197710d82bab442daa24b8b7e98ec98dd" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00117.tar", + "files": 322, + "samples": 322, + "raw_bytes": 41332568, + "archive_bytes": 41902080, + "sha256": "e86b54618e621b61bf8d32740fcdb81cba22e83bfd6a41f75fecda09021e54d4", + "source_signature": "92e5c372eae31a6afe842f644828a7fe3adab940ee31f588853b9f294239e2c5", + "index_sha256": "da075408a492530f0995fa9ad4002538dca95faf0730d5c4e439c31b078e97eb" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00118.tar", + "files": 325, + "samples": 325, + "raw_bytes": 42411228, + "archive_bytes": 42987520, + "sha256": "499d6c22010c4c9c01fba3fea7e2cfe96bd9cf7ea14ba4906700ea4cc4e754bf", + "source_signature": "f6b1185d6be4e786fef46a0d073c2e34b668acc8d617631803367e284a104394", + "index_sha256": "138045a2e83824fdfcb646a6761dcc54c6ffa87e943b51ca8bacb05c9bbb9c95" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00119.tar", + "files": 334, + "samples": 334, + "raw_bytes": 42549096, + "archive_bytes": 43141120, + "sha256": "5fffb4e6348fec257a194e5a5aadf46f83693bb808b92a41dfd472cad9b084e2", + "source_signature": "ae1f4a9f34253429bf9a854bc918d85e5c7562e2598f18489e9bd9eba18480c6", + "index_sha256": "10f1a5ae880c8cdb7d6c2cea6cfb61290ad5fe771fdc2362c69d0d66bb5f79a9" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00120.tar", + "files": 337, + "samples": 337, + "raw_bytes": 42747628, + "archive_bytes": 43345920, + "sha256": "9d7ff1263e7e2db30beb53e09f91fdaebf0e60f060e1a2e16c05a56126dbf94a", + "source_signature": "8f64dbfdf01bdfaf50a8e74b64f18c6b7695400d75565405d6724a21f3cb1cf8", + "index_sha256": "b2a8692fa485254328cfd6da56010ab55ad009206358ac0bf436989736c5a01b" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00121.tar", + "files": 332, + "samples": 332, + "raw_bytes": 40795408, + "archive_bytes": 41379840, + "sha256": "1e8e690796fac27f42d7912610fbc5a9e1f1e8c745766f15ea4d1f10fd0dffc4", + "source_signature": "c5bad4e660350df3fb96bac076153bebd68ddedbee283cdd60cca3752158d022", + "index_sha256": "0d6d6eba47f26d82d87b1c758b2af1d0ccc7a54d0a5c182b4d915c3d3244b318" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00122.tar", + "files": 325, + "samples": 325, + "raw_bytes": 39214300, + "archive_bytes": 39792640, + "sha256": "ecbbeeb16f1d1a6adcc46e142d93c1a5cb932924ca7b3a537643223d065a25b5", + "source_signature": "0b5c999d72a90da8d9fc1bac35db440a9b31d67a9de452176ecdf07dae2dbb3f", + "index_sha256": "d602155c4dc9e5f5657cf06bd3a6f7db5f466c6c68254099f29690e202d3d006" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00123.tar", + "files": 356, + "samples": 356, + "raw_bytes": 45603377, + "archive_bytes": 46233600, + "sha256": "d6c8cd91debbe72a09084e1486da3dc6c35c5064eeb20dffc1934beffba9734f", + "source_signature": "550a326ab5ff91318039e5a91264b517e9c349b5d2e7b6de11048cb5aafbd497", + "index_sha256": "5cf459e0e45df6e3b3326c617c14a08d781ec5d23045cc3a31f1005503039353" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00124.tar", + "files": 334, + "samples": 334, + "raw_bytes": 42459496, + "archive_bytes": 43048960, + "sha256": "592e96be49810c79254232a94cac250bd799da6950109a5db96125069d2e729c", + "source_signature": "3341222b57de4171262c96835ccf99be670662a298c899b053521d45bfa72d72", + "index_sha256": "5f560bdab944b48e9ef687817d32885d4870e7e9685d2299b0d68dc386b2a704" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00125.tar", + "files": 310, + "samples": 310, + "raw_bytes": 36890440, + "archive_bytes": 37437440, + "sha256": "87497e9131a30052fab22487c6c902a9b1f8c01c8047b276601696324b9386c8", + "source_signature": "a865c87bd9c8cfafa1df239553aae0411fa8d77813a98628ed3b2b881ec3f349", + "index_sha256": "ba4c40e8ac2db0912c6b2b903435d1dc140b9630477baee65ffa675925caff0e" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00126.tar", + "files": 299, + "samples": 299, + "raw_bytes": 37651556, + "archive_bytes": 38184960, + "sha256": "f882065bf4f6d11fc6c369b5d6b1b448fee89904eee177556ad6d843263cacbe", + "source_signature": "e71ab32d923b9481f1004be764035a1543da09ee3e6ff7cb1cdbb881ce9ada78", + "index_sha256": "7126f1c07672adb362cca491ebbab0972ba5eec93138a0eb73805c2f020af94f" + }, + { + "file": "shards/hfd_audio_command_continuous_office_v1/00127.tar", + "files": 310, + "samples": 310, + "raw_bytes": 60378120, + "archive_bytes": 60928000, + "sha256": "9871cab79cd56608dc454b80806f820444ba67655193fa0177dc248a4f489d2e", + "source_signature": "6c0cf355a8e1bffe3bedd3f94aa6fb087ca323e6ac7b8fb9e6aeafff282ff2b4", + "index_sha256": "ffca30e005faf32ef78b1276a12c6d3315fc1e08e80147cf74bb2fcf489f3d8b" + }, + { + "file": "shards/hfd_backchannel_commands_v2_short/00000.tar", + "files": 121, + "samples": 121, + "raw_bytes": 2325773, + "archive_bytes": 2560000, + "sha256": "e87a6cf29bd174f62239fba65ec71c06aaa28f145c519fc9b4dc55ff3219b886", + "source_signature": "b96d0b78657955a275b4810e8c5b97e57de5bea01cf8e8e1a59934c8070a0a90", + "index_sha256": "fab4e3803c6dd4e5839c34ea365db1064b996c79d7968e6a24f75b79f4731b0f" + }, + { + "file": "shards/hfd_control_commands_v1/00000.tar", + "files": 97, + "samples": 97, + "raw_bytes": 3481689, + "archive_bytes": 3686400, + "sha256": "931ed66438de3bb950af8b0b950e0515b9eeaaccd334d78088bfb1c51c09de32", + "source_signature": "457fd03c8a85176166d084b2dd7cb9a6cf1ba792f2384eedbd36654bb24c92c8", + "index_sha256": "0eae7dc7294fd0ca19d5733be8c476a377404d5655d7b8b2c37d64f947ba053e" + }, + { + "file": "shards/hfd_hesitation_pause_v1/00000.tar", + "files": 2553, + "samples": 2553, + "raw_bytes": 140619796, + "archive_bytes": 145408000, + "sha256": "011993688d55f5b0f0c54f8d402abd82957b382d94e94fd788737142df117d4b", + "source_signature": "004348f75df98ceed5a0c75bb7b136d4759e0ab09051658912204519008a56d7", + "index_sha256": "fb8654bcb31f5bd56631e2ccba078af0893fb9ee85e61418f0ce8915f64be2d9" + }, + { + "file": "shards/hfd_hesitation_pause_v1/00001.tar", + "files": 2522, + "samples": 2522, + "raw_bytes": 139527914, + "archive_bytes": 144261120, + "sha256": "77e21dd06fac3dab3fe80d4b36b8a648d36deb222dc5e33f86be55abdf2353de", + "source_signature": "49e699db83e5363dfd4b571d6f30e96bbe5f5cfd14882d2c852eff05aefa0815", + "index_sha256": "6520d0c7458d5d393bf38815edd4683c7b1b29b526cc39e29717754687b6bbd6" + }, + { + "file": "shards/hfd_hesitation_pause_v1/00002.tar", + "files": 2513, + "samples": 2513, + "raw_bytes": 137574268, + "archive_bytes": 142284800, + "sha256": "bc461a0b4a156301109af850d9fc90b802a8f2f4dd777d43f01127fb9d93846e", + "source_signature": "dac7cfd232e26f441a85f8c3466f89d27cdb82a0c1baf5c21da26dfd87a25e38", + "index_sha256": "027225d701fac9c7df73c0260810173d5b167ad3c96e44a5d3f226bcb685206f" + }, + { + "file": "shards/hfd_hesitation_pause_v1/00003.tar", + "files": 2541, + "samples": 2541, + "raw_bytes": 137777112, + "archive_bytes": 142551040, + "sha256": "22453fb9b1268170eef0f518d4c3ae9cd2ef441589493f3b4ec3fad05c9eb66b", + "source_signature": "3701076e0b72d008fecfa5669d250775d3554168a73af919bb621bebd552c3c6", + "index_sha256": "95b6b45dc60b20ed5e3293ae3895b1f61d41fa7e3f8ac64bbe9068e8fd623b35" + }, + { + "file": "shards/hfd_hesitation_pause_v1/00004.tar", + "files": 2593, + "samples": 2593, + "raw_bytes": 141045664, + "archive_bytes": 145909760, + "sha256": "6509f7416d93de7bceb2cbc89053da36342e8bef1d43e9a7eda3d4534866ad4a", + "source_signature": "43292ae174fb65e0b5dd21035e2531fc84c248c834d1565af45b5bfa752be18f", + "index_sha256": "abe34e7d07acbbd45f81b7598cf12509985c1137bf594745da9436591dc25bd0" + }, + { + "file": "shards/hfd_hesitation_pause_v1/00005.tar", + "files": 2612, + "samples": 2612, + "raw_bytes": 144318362, + "archive_bytes": 149227520, + "sha256": "ffc3758689d1e9ac16249910838f1dfe79b6f0d022e8b888130de62676c8a028", + "source_signature": "9fb1d660c2ba2b8f1a2c72f994a0e84ba487751ca28e2c4238b4bbccc323b81c", + "index_sha256": "d4b315c134c78961712163d4f1f09e0589fb3507112a5c8389faacd88bf49257" + }, + { + "file": "shards/hfd_hesitation_pause_v1/00006.tar", + "files": 2609, + "samples": 2609, + "raw_bytes": 210200579, + "archive_bytes": 215091200, + "sha256": "26fef958c65f1f44634a44a0f38b00e2505af44f1b6988f2206f42b4d6359464", + "source_signature": "af49f534f8fbe84fdadd3a96f2ebd6bdcc771f68e37de8dfbdd81b092672cb84", + "index_sha256": "36914d8d36d58b06ade2631cd2212295210db5c6c53672327a550f7fbfd888ae" + }, + { + "file": "shards/hfd_hesitation_pause_v1/00007.tar", + "files": 2552, + "samples": 2552, + "raw_bytes": 142577594, + "archive_bytes": 147374080, + "sha256": "1b5271e59827d30fc06e5628b0aad017bf6c400f0bbcff8459690d6044dc144e", + "source_signature": "f9de2fb5393d88ca45834cb83ed163e0773459b11c3c08550a14e7e60fc89e05", + "index_sha256": "29fae2daf286c09acf15a8c52a72bbd80014bc9ea646515b6c1ed209c87658d5" + }, + { + "file": "shards/hfd_hesitation_pause_v1/00008.tar", + "files": 2602, + "samples": 2602, + "raw_bytes": 143146098, + "archive_bytes": 148039680, + "sha256": "dd70aed4ccbd59699f99ba10dcaac7f63dc9ae4cb79323239f71f41879605b26", + "source_signature": "e4af217907f1acedc923ffafb3ba740cf8b5c89851758daab2e8a0c68c3a0bc7", + "index_sha256": "3e610df4e17d7c059406d9294b7d69213e9df58d69a7669b42cfcd74b7ffc7a2" + }, + { + "file": "shards/hfd_hesitation_pause_v1/00009.tar", + "files": 2578, + "samples": 2578, + "raw_bytes": 142338704, + "archive_bytes": 147189760, + "sha256": "3efab2954680ebbe3447efd7e71dcaa632790bb30eb32904778d53f227047912", + "source_signature": "46d2c1b5841ee87ebf57393345bece0ca6162a4b8013be8544b5074ad2d33567", + "index_sha256": "a8b49253b4db99ed484cb0eecdeb558b23cc27a4335880ce8b184f377e6476c4" + }, + { + "file": "shards/hfd_hesitation_pause_v1/00010.tar", + "files": 2560, + "samples": 2560, + "raw_bytes": 139684532, + "archive_bytes": 144496640, + "sha256": "e5d5a9a238b9bb6f0740524ec90522730e2fe87d047aa7858b21d0581b941f69", + "source_signature": "1ab3e67b676aa34d9b5d3cd4a960ba2cb83c8102513fa1838df307e690037995", + "index_sha256": "eea265508bcc6ba91aa44b6f89f1469adc4b00010c1dbcc058bfbcbb76a8ca62" + }, + { + "file": "shards/hfd_hesitation_pause_v1/00011.tar", + "files": 2482, + "samples": 2482, + "raw_bytes": 137955706, + "archive_bytes": 142622720, + "sha256": "19015558a10a31329dd01f7c82c8513cebcadc8870c49b3d66880fd98ef61210", + "source_signature": "273edf3166cadcc7fd3fdc99dd381f5729600689abc1bffc2cee80d180ed2e9e", + "index_sha256": "43619be4a7a56b4cb72a6d3fe4979b9da4d7afddf94a3f034003e6705deb14bd" + }, + { + "file": "shards/hfd_hesitation_pause_v1/00012.tar", + "files": 2556, + "samples": 2556, + "raw_bytes": 140481622, + "archive_bytes": 145285120, + "sha256": "7f769b427299d26c54bc04cb0caae40e7d1e671774ade2ac64d547c026a855f7", + "source_signature": "50b5e403bc64ef427baa9b77ab56b1daa9502b78cc483f87d78559b7ac0234b9", + "index_sha256": "8b4f045471ddc617d5ebac0debfe0b528a3c6800f2029dcbc4fc841a49a124ed" + }, + { + "file": "shards/hfd_hesitation_pause_v1/00013.tar", + "files": 2608, + "samples": 2608, + "raw_bytes": 141977706, + "archive_bytes": 146882560, + "sha256": "02f79dcc1d83a06712499c84628b688f54e10a269bce733550a1cf628766be36", + "source_signature": "d3b76cfdd739264804e813310d5d320d36a25888f4e336cf21ad1d4ee0caccf3", + "index_sha256": "562fc0f188324325f91d43e3ebed4f0dd81a1efb79f7877d7310426b3aad7ec6" + }, + { + "file": "shards/hfd_hesitation_pause_v1/00014.tar", + "files": 2587, + "samples": 2587, + "raw_bytes": 140110114, + "archive_bytes": 144967680, + "sha256": "86d3636ffac3b47cac59b07b3bc923beebaa25f17f3188db786bd39e62115d40", + "source_signature": "ff58c03506fa3dadee2ad1f3932372366832d537ffecaeccb9f79bcdf6b344dc", + "index_sha256": "01de77287551543fd3671842a6ba2580128932c2d16e6b9eabc4f3a4061516a2" + }, + { + "file": "shards/hfd_hesitation_pause_v1/00015.tar", + "files": 2592, + "samples": 2592, + "raw_bytes": 143239606, + "archive_bytes": 148101120, + "sha256": "4fd8bdfa4f853e78dd4c7b06d4269d1789e76c3af6f6a6ec5a67df76f8e23e5e", + "source_signature": "6c843365d6ea60758b0d41e042a5cd6107578cc73ae448085f43921754509aef", + "index_sha256": "126fb2c936520588413f491c74a6aae1639aec97526f99b5ca2fb235c50a5048" + }, + { + "file": "shards/hfd_hesitation_pause_v1/00016.tar", + "files": 2586, + "samples": 2586, + "raw_bytes": 142764434, + "archive_bytes": 147619840, + "sha256": "f2aa704fad58a877eb32cf309ff0cc0cb9fbc47c5a56a65c5b97a1522fe2d832", + "source_signature": "b478a9cc4add4bf0cf1fbeda648b4f86b3a3a42e34716b35ccabc6f027141498", + "index_sha256": "d015d60f490f5dce0e9f039bd119aa93215c7760c52ed2d569293b23ae45f3b9" + }, + { + "file": "shards/hfd_hesitation_pause_v1/00017.tar", + "files": 2570, + "samples": 2570, + "raw_bytes": 141883762, + "archive_bytes": 146708480, + "sha256": "18efd732851496c53f75f5d44a2da76a8e3d062f4abf214c836b244de7e7238d", + "source_signature": "712c1d0a0c9175fc38a82e6ee624f024f0ca92ff8fd0404497a250b65fe82d10", + "index_sha256": "8031046ff19212dafe4aa4e285a727b9798de3283d56c7fb4eab0a60ead3c93b" + }, + { + "file": "shards/hfd_hesitation_pause_v1/00018.tar", + "files": 2557, + "samples": 2557, + "raw_bytes": 140629906, + "archive_bytes": 145438720, + "sha256": "84a4be2df65fed861892fefe25c7c5ad441ff0f0d87d4ada379649742e393c21", + "source_signature": "75a73a72d9c4dadea978a311bcefae748d6daa210cffb06c55ba6f615ee1f899", + "index_sha256": "8151b724d890def4fa15a11479df720182e5d25e8fac86f355e630e3c67af407" + }, + { + "file": "shards/hfd_hesitation_pause_v1/00019.tar", + "files": 2495, + "samples": 2495, + "raw_bytes": 140763025, + "archive_bytes": 145448960, + "sha256": "89ef29d2f8569b6738e1eff723e36218c73112f1ac7c35992a164a8870f44a3c", + "source_signature": "9948b154777889142ea4b6aceb6926f02b0148ed0f9b1e983a387a87a0e48ee4", + "index_sha256": "19a14047a86ef577ab419702f0f02c0a415125d31857950be7d6922dba1c2986" + }, + { + "file": "shards/hfd_hesitation_pause_v1/00020.tar", + "files": 2560, + "samples": 2560, + "raw_bytes": 139402803, + "archive_bytes": 144199680, + "sha256": "3cafec0fa7c33d67440c4c44ec6743373953b3c65547a3eaf264412eb3e5f40a", + "source_signature": "b4ad6228cdc5cf9b592fe5a75ac7864c5c3e84d36b4c505cb5223e2d76f6d6fc", + "index_sha256": "4c333a40c6db196b82577bfcb8bbdd4755ee869d67fe2031aadf5ed97b545d9e" + }, + { + "file": "shards/hfd_hesitation_pause_v1/00021.tar", + "files": 2541, + "samples": 2541, + "raw_bytes": 137844434, + "archive_bytes": 142612480, + "sha256": "6a29792a46cedc7e63104d080e7621457dfbb6af022b1f7063d7556a5c2deec3", + "source_signature": "f4ee5e7840330243cd8162e1edd35091d7f5848927458d1218ad12ccc4282daa", + "index_sha256": "b64020a71cdb2c38fcba74ea074cccce56c31eb5c484e717016f551239349ab2" + }, + { + "file": "shards/hfd_hesitation_pause_v1/00022.tar", + "files": 2554, + "samples": 2554, + "raw_bytes": 141483026, + "archive_bytes": 146278400, + "sha256": "81f8deb18d47404583fe461bda3d5fdad6800dd8311b180ea091f45ba1d74687", + "source_signature": "da3eb78498751315685674ac80ab962119ac103a115f36a9733e99cd48964dea", + "index_sha256": "976f7dab85a82ca121d8f28ad8e2024aee4d3ffd1308f54f4bcab95940b80049" + }, + { + "file": "shards/hfd_hesitation_pause_v1/00023.tar", + "files": 2614, + "samples": 2614, + "raw_bytes": 144190472, + "archive_bytes": 149114880, + "sha256": "f667666f5b7e8abd71c6019d1f886e905296c6ac2711af8d821cbd469c546ece", + "source_signature": "fc894b28b3cbc048a4404ae82192cdaeb6963d6c97aa7c224d06f14b24947e4b", + "index_sha256": "4a9a9bb044a368d0dc9cd5d8a9e303d50e9ec37f85245f82a7e7265f2530e0c9" + }, + { + "file": "shards/hfd_hesitation_pause_v1/00024.tar", + "files": 2563, + "samples": 2563, + "raw_bytes": 140038364, + "archive_bytes": 144844800, + "sha256": "f6d04acb28c1d065d840ea66f1b60794adef947905efb5a53f31a9d1744c9afa", + "source_signature": "d95cdae397d820e4e4e4be4a6c189157e9f70d94b80f01b61cf1c8e495eb4544", + "index_sha256": "8c22b9f0863197d0e416f8a46e06a52640a3008e325b110563596ca62fbcc1f6" + }, + { + "file": "shards/hfd_hesitation_pause_v1/00025.tar", + "files": 2547, + "samples": 2547, + "raw_bytes": 140134371, + "archive_bytes": 144916480, + "sha256": "be51f5d3e190ee72d273feb21e516bf5471eeb5996dff4954796c561c2928d42", + "source_signature": "2fe3d1e0a76e32edb4ede46f965599667606fc6486c0735e62234f7b86a64c19", + "index_sha256": "3dd56f5a504520ef2dc9b4d3931219de7aa5c4a06994eab8d10f58c328fb69be" + }, + { + "file": "shards/hfd_hesitation_pause_v1/00026.tar", + "files": 2597, + "samples": 2597, + "raw_bytes": 140302684, + "archive_bytes": 145172480, + "sha256": "94a043687cde1ddf9885f064482c476cba515ca464c11d64d0dc34d660dd2149", + "source_signature": "c1d957517f146367fbb16bcbbc179a1c4bf521ef6f98b497a2cb7f402eaf3396", + "index_sha256": "362c11cc532258029c6dc7dfd93a46463f6b3bcc787b67b294d2ef0d397b7357" + }, + { + "file": "shards/hfd_hesitation_pause_v1/00027.tar", + "files": 2546, + "samples": 2546, + "raw_bytes": 139372264, + "archive_bytes": 144158720, + "sha256": "445cf676beb1e7260b737aa56b76ed894e636db7056aafaeaf96a197e03286c1", + "source_signature": "cdf64a0ea8965bdbe9ccac97a593edfd4c0d8c07982843871b3383d06adccc1f", + "index_sha256": "b527a6079cc5b6a9bc876ecb1c0ab18c9dd31698bca3f1fd0b4c6c4682ef01dd" + }, + { + "file": "shards/hfd_hesitation_pause_v1/00028.tar", + "files": 2546, + "samples": 2546, + "raw_bytes": 140693112, + "archive_bytes": 145479680, + "sha256": "2269c653264e9f7e03ed161f257e354b97bcb82e2c2610760caf6e8d0fee971f", + "source_signature": "65dde20e41f311750733fa3b83b86ca3186a81b42556df4869adbdbf0ed2ecb5", + "index_sha256": "070fb08dc4f899aafd0ecc5f2590d5d39a5b913a0cc0204490ed09dbd2ccccd8" + }, + { + "file": "shards/hfd_hesitation_pause_v1/00029.tar", + "files": 2448, + "samples": 2448, + "raw_bytes": 131726108, + "archive_bytes": 136335360, + "sha256": "631ab8736d60fbb3e69a7dc9f52dd1f870cf049938000bd8db521f61aa9a009e", + "source_signature": "264ac2fd5a744de07f8921bf20a5b4fbc4977120df438d3ac9c77dd773a8fc87", + "index_sha256": "bcab7b0adecdfb9553c59dc25901c6f13d8756b9b70b36ab415702dda63ad7a9" + }, + { + "file": "shards/hfd_hesitation_pause_v1/00030.tar", + "files": 2467, + "samples": 2467, + "raw_bytes": 136999536, + "archive_bytes": 141629440, + "sha256": "8a52a90ca65d319a09ddb7f37bc94b95dfc313a8072a9fbca283c09eee668650", + "source_signature": "875c5eaf376d63d250e5bfc716e254d7cae14d49dcbcac9aabb003ea69c43dce", + "index_sha256": "070b654eb8fdece72dba5764f55b5d603f4f86b1dd2af17754b465b2891b0190" + }, + { + "file": "shards/hfd_hesitation_pause_v1/00031.tar", + "files": 2504, + "samples": 2504, + "raw_bytes": 136644626, + "archive_bytes": 141352960, + "sha256": "8dc406b85228bc7d715674a804cb47f5a23f2794cb856a06e3914bd0288fa54a", + "source_signature": "90106d30822faec14c6c602c40794a32447313551b14f1ebdf62ade4e56a9780", + "index_sha256": "632606bf8a1725deb129a35245dd2a87741c85cd4f8295430f98aef8662f92af" + }, + { + "file": "shards/hfd_hesitation_pause_v2/00000.tar", + "files": 91, + "samples": 91, + "raw_bytes": 74569500, + "archive_bytes": 74752000, + "sha256": "61fb5be780e093a3909c11afa72cefcfb1e37cce7c14cee21c76e89a2c202bc5", + "source_signature": "d2c74640204e3082c9758da9f6ead0ca8ccb8b437fdebaecf034cc1f64fe3719", + "index_sha256": "4f86c8ce7e857bb4025e74862e5c756779efe3dcc0ae3bb615bf6a9668bf8b6b" + }, + { + "file": "shards/hfd_third_party_multi_v5/00000.tar", + "files": 415, + "samples": 415, + "raw_bytes": 293213068, + "archive_bytes": 293949440, + "sha256": "1c9ad74856af5e2cae5983021f8b3111740c8916bf54c0ee0e9cdb79782e1253", + "source_signature": "27abfeab713befc2395cf2202027160a5fae29b1952ef1c918a46c30c4fc9dbe", + "index_sha256": "572354a55a98bacca550f007a723ca18f0ea73e57754daf9a7f98257ac365c4d" + }, + { + "file": "shards/hfd_third_party_multi_v5/00001.tar", + "files": 437, + "samples": 437, + "raw_bytes": 313897991, + "archive_bytes": 314675200, + "sha256": "55f9bae2a707115bfc3350348598ed588361526f59e9ee80781d28ed06f34c16", + "source_signature": "22e9fd0a8928738f8184e3043232a84641d46cda1df55cdd05272f14d05550ca", + "index_sha256": "7c4bec6e73ebce53e8489bce4a2fd937e8449023b7185a76f8b163021129566c" + }, + { + "file": "shards/hfd_third_party_multi_v5/00002.tar", + "files": 419, + "samples": 419, + "raw_bytes": 306964214, + "archive_bytes": 307712000, + "sha256": "308b66356270aebb88db350f2e116f26692b13b20a3ee789e623da6fda0f29df", + "source_signature": "dd3c5acfe6a7a134e1be32e77d99d2b70e0ce99155b392f9273522becaa237d0", + "index_sha256": "a60473a2e91a368cb7d4e0193d6ca0895e1748bca886748ee46e0501580e1d4a" + }, + { + "file": "shards/hfd_third_party_multi_v5/00003.tar", + "files": 407, + "samples": 407, + "raw_bytes": 311447477, + "archive_bytes": 312176640, + "sha256": "a47267a9787e6a1ef5c05487ee7cedf5334d7fec2f5067f45d4df27bfd618483", + "source_signature": "406d03bf7a3804643642073cc5aa9146e587484fdd0cb9136d7cedf0a7abbe81", + "index_sha256": "942918e222f4921206bf329532ae36395473010303e74fee5e6341bcb00df01d" + }, + { + "file": "shards/hfd_third_party_multi_v5/00004.tar", + "files": 419, + "samples": 419, + "raw_bytes": 307917277, + "archive_bytes": 308664320, + "sha256": "1e81c204064fa2415cda1b0647dacd302ab510890580c9c28888e5291f05f7b5", + "source_signature": "2406410401ba11027b46e81ce5a7d4d3d52138d975825b8e2abdb473475d0ee4", + "index_sha256": "4c9b900ed296738ee26a11bca11fe1bb30b7b494d9f7366d5d8caab12d99a155" + }, + { + "file": "shards/hfd_third_party_multi_v5/00005.tar", + "files": 434, + "samples": 434, + "raw_bytes": 320342638, + "archive_bytes": 321116160, + "sha256": "96310618b1003049fc9287c53ced265ee94bf59538da0d3d0a0e546b4748dc8c", + "source_signature": "52ddcdd148c42e2ac34df678de96e7c78fe57389b3798315a9c9ae7ff35406f0", + "index_sha256": "208297a3351a275cbee7637332605352e28e56788a7338d740dbdb88a8d13c06" + }, + { + "file": "shards/hfd_third_party_multi_v5/00006.tar", + "files": 424, + "samples": 424, + "raw_bytes": 337292857, + "archive_bytes": 338053120, + "sha256": "1f18db221a4daabfbfc0c452181f87ee9baacc59b11a04a16617ff0fe31aac57", + "source_signature": "859f83c25d56c302079779d6064b17a11bc3fc854b5ff12cb7d8231dd1556998", + "index_sha256": "c8da4b61e4a5730a8b1d2f43334698aaa60101bc0d379028171103b6b46a74d4" + }, + { + "file": "shards/hfd_third_party_multi_v5/00007.tar", + "files": 455, + "samples": 455, + "raw_bytes": 343688438, + "archive_bytes": 344504320, + "sha256": "9aa75097318f7c978838541223cfd3c3ec28c8365f39ee7da03646636d9b748e", + "source_signature": "5586d27988bb95e857c0a246ac62a908bd757aa364c452e7150e394082fcd203", + "index_sha256": "94307cb5e287474466d8e3a7f946c4699ddb7c746c334a158e7d810a32bfce58" + }, + { + "file": "shards/hfd_third_party_multi_v5/00008.tar", + "files": 487, + "samples": 487, + "raw_bytes": 356590457, + "archive_bytes": 357457920, + "sha256": "4496597f9aecde5d258fe8f82400eccc58cf64c4e89aa79976a96c3a2455e799", + "source_signature": "d03a07061d99488535aa4b5459a345424585d8ac8ea2eff50bc1b3c8d0553e59", + "index_sha256": "8aa999cf49e36c5853c760235e9fa84e49bcdd5bb7463304998aaf94255445a1" + }, + { + "file": "shards/hfd_third_party_multi_v5/00009.tar", + "files": 478, + "samples": 478, + "raw_bytes": 370310085, + "archive_bytes": 371159040, + "sha256": "734e52471a065b147c774fa3765b5908ad368500ee85e7ed5e82d9d02ca027ee", + "source_signature": "6d02ddb20bcb7113f82483eee08fd06c6f60d6fd422291e9571c7b9699f35793", + "index_sha256": "b0bb17bc3e3d427e8f60e642055530ee67f811201dc81e100d044bda47cb4855" + }, + { + "file": "shards/hfd_third_party_multi_v5/00010.tar", + "files": 472, + "samples": 472, + "raw_bytes": 350895988, + "archive_bytes": 351733760, + "sha256": "ba9f9dbce8a2963e5241f15f9653ff2a5bab71af45f8d29efead466d02a41e17", + "source_signature": "bbcdac9dac755a1a20cf56cc2860dbf427bdc33ebe764ef0ae1f971f7c0fb287", + "index_sha256": "99f8f70db806fa636301145ca8ad15f8b080c3b16b0b3932a95c513602a5dac5" + }, + { + "file": "shards/hfd_third_party_multi_v5/00011.tar", + "files": 454, + "samples": 454, + "raw_bytes": 319355245, + "archive_bytes": 320174080, + "sha256": "1f08984f9c747716bd791ee66818bd26a1c0f5cdc21e8e2905acd2279e0a0f39", + "source_signature": "b1804060f346eb0bf69e397b00e88965509657cf5714206b1f76c08aa83dfdb8", + "index_sha256": "220946b598c8ea2729a0d1c52e5067f23178a998966c99dd98a3a95cd0effab5" + }, + { + "file": "shards/hfd_third_party_multi_v5/00012.tar", + "files": 430, + "samples": 430, + "raw_bytes": 353566076, + "archive_bytes": 354334720, + "sha256": "d001d40c0361e7bcb58793c769230706442a77c45a859853c8c94d2ae3568752", + "source_signature": "2b12a414fce164293d477e684f2e720bbeb1b355b4e784d6806f6f1bbb4a0881", + "index_sha256": "63a6fc34971724f91f8d57e8996c6cd5227b209062f7cae114ad54c365ec8c72" + }, + { + "file": "shards/hfd_third_party_multi_v5/00013.tar", + "files": 425, + "samples": 425, + "raw_bytes": 336321863, + "archive_bytes": 337080320, + "sha256": "36708e2e45744e223b269f5f29631f3d8bb38c91b241854de58185731c3e608c", + "source_signature": "9a645f47920fc30485d07669206e5ff00a6d0491eb82c117f6ce7edb91b6f8b6", + "index_sha256": "bf0c95acc94bb0ecd81f053f9c56005e749209b176eb0dd2385dd8876f20e5d7" + }, + { + "file": "shards/hfd_third_party_multi_v5/00014.tar", + "files": 448, + "samples": 448, + "raw_bytes": 333043515, + "archive_bytes": 333844480, + "sha256": "aee98fc516e10edc93590adcc265381c07f51277d31498c7a54fd62b1d62d994", + "source_signature": "a137357ab52953bade9fcbbd70ed840b18674c1ac4754bfebcd9a2bc78e4d97f", + "index_sha256": "9b21dbc039142b6bf4667b9a486279d8fee2a68c4616df2fb25b0995818cbbd1" + }, + { + "file": "shards/hfd_third_party_multi_v5/00015.tar", + "files": 405, + "samples": 405, + "raw_bytes": 317089233, + "archive_bytes": 317808640, + "sha256": "f5ac09d3a98e1ec3c86943f9774462ea07ed01d4a165787d32ecce7e96065a18", + "source_signature": "0928cecf85352eca85152b0da7a623c2e45a79804f01f4b74822d2040b4ec9c1", + "index_sha256": "0e8e8ee3538ac48eff4186a6cf2fb8cf50819be93fc23f18bfc147bb7f9b9fad" + }, + { + "file": "shards/hfd_third_party_multi_v5/00016.tar", + "files": 461, + "samples": 461, + "raw_bytes": 337526430, + "archive_bytes": 338350080, + "sha256": "53cb60eef6e7edf98f2729e8dd94f01889a5a812cef7c79e7bdfac8e4db4dc58", + "source_signature": "b7aed42030dab675b32ac2d27a62b7e8b890b2ccb6e71b285c0512e3bae40ff6", + "index_sha256": "5c02078300faf6acaadd2bf6917d6af05b39729a5edbef49218bf2306e6f0208" + }, + { + "file": "shards/hfd_third_party_multi_v5/00017.tar", + "files": 421, + "samples": 421, + "raw_bytes": 307294962, + "archive_bytes": 308049920, + "sha256": "5efaa10e12feb6e8268de9af55da43903efa3cd47f0ca717e6aa40db69235c75", + "source_signature": "f75ab4ae758b33a42844637222a8fc5fb84d58f4e60fbe8a9bcf3620c4d37c30", + "index_sha256": "4db9b2ac33a8bcea148075db88384204a5a73d312484305f9bd31aab750702f7" + }, + { + "file": "shards/hfd_third_party_multi_v5/00018.tar", + "files": 387, + "samples": 387, + "raw_bytes": 288955184, + "archive_bytes": 289638400, + "sha256": "72e3ad262204413634f2975d1019e7531ba4ff0f8893d5495011cf8b9a17606f", + "source_signature": "27b5948ca0eacff11b886a7240a663d2ef5de1f561a462e7ad6f619036719f2c", + "index_sha256": "c2d713e27b5763807e41f372b3563aeeb1590ed7facbc11faa3b3ac1f778ed0a" + }, + { + "file": "shards/hfd_third_party_multi_v5/00019.tar", + "files": 434, + "samples": 434, + "raw_bytes": 351158271, + "archive_bytes": 351928320, + "sha256": "75660261661d19ef82c7ae7f3719aa14939f440a8777818adc0bd38e856c85cb", + "source_signature": "b8cf6559bb45d0af6b3b3a630f38026153428d1a4112a92caad6563ecda1fc27", + "index_sha256": "047a62bc11ebf44998dc185177a4141e65eb45fdee45b92d4aa849252ca86045" + }, + { + "file": "shards/hfd_third_party_multi_v5/00020.tar", + "files": 432, + "samples": 432, + "raw_bytes": 340791483, + "archive_bytes": 341565440, + "sha256": "d2303a912d7e4f20eb7b372e1298e6222410934fdec4ba9908695887567dc856", + "source_signature": "777f193b43b90773ade44f13189bb9ad959ff85933d214134ffcb59fa982f63c", + "index_sha256": "6d3d402da6a8d79eacfec6adc507e3650025cf2b79ef5c95fd6213b10f627bf4" + }, + { + "file": "shards/hfd_third_party_multi_v5/00021.tar", + "files": 442, + "samples": 442, + "raw_bytes": 341692206, + "archive_bytes": 342476800, + "sha256": "122f8b711d6d18e89178b3a8688707749b4581b6b2dfc702f529bd9121555ec4", + "source_signature": "67ac576715dbb154d28e6665aaf0b5a24790b43ae1f27081438bd392f7677fbb", + "index_sha256": "f705695a213775f58b9ca0afb8453f1e181cfcbc3ed1365aad651327f83b100b" + }, + { + "file": "shards/hfd_third_party_multi_v5/00022.tar", + "files": 413, + "samples": 413, + "raw_bytes": 293418217, + "archive_bytes": 294154240, + "sha256": "5fdc4c60bf8c79b22d782d6ebc3e977b5ee40b268249f3dbd754641d342986b8", + "source_signature": "cd134345975c8bdd6ea4626c357a3b421a6ac607313f60c02e941ff35d035238", + "index_sha256": "fa2ddb591ddcde4fa9e5da6c26c8f81de2f30d80dcadcee9f676b917fc7ed344" + }, + { + "file": "shards/hfd_third_party_multi_v5/00023.tar", + "files": 440, + "samples": 440, + "raw_bytes": 337265996, + "archive_bytes": 338053120, + "sha256": "ccb2a0bd081c4e8309e8b4a470e586b1b37c2485d5ada709e75af81602f90285", + "source_signature": "b10831a361af60a630b19c0ced985f79ac5f2f7dc3ec9c726d84e263388206a9", + "index_sha256": "d22bebc07587df351fd891f2539fbb7b27fbb8f03b362c2339940db0a1ab9d4b" + }, + { + "file": "shards/hfd_third_party_multi_v5/00024.tar", + "files": 415, + "samples": 415, + "raw_bytes": 317020493, + "archive_bytes": 317757440, + "sha256": "60b0e1f3ad4a99b000b147d122e0ec677b1a3475cb9975cf61e455f7c9672694", + "source_signature": "f8ba5cccb342a53f1012a784366395a3e769a350a61b61494f9cc0bc729b3522", + "index_sha256": "580b50a09de270ec30197fa629a341cd8e9e87ce085006fcbeac19c7053037f4" + }, + { + "file": "shards/hfd_third_party_multi_v5/00025.tar", + "files": 440, + "samples": 440, + "raw_bytes": 338569158, + "archive_bytes": 339353600, + "sha256": "a2ba05e545d85b88bc5e2480c212f929c8440fea0834d7e1b5f381ab726ccc9b", + "source_signature": "7ca97238f721a9ff78a1d5977f64ed7f1da7a7e75a8cf11b30705d0f80f388c8", + "index_sha256": "d0e145b723ef07d0687ead6e25027951b66428cdcd221b544ad5ce0a98d1eeda" + }, + { + "file": "shards/hfd_third_party_multi_v5/00026.tar", + "files": 421, + "samples": 421, + "raw_bytes": 336513639, + "archive_bytes": 337264640, + "sha256": "9da5009d95e3d5c3077d3abaddda61218356b04f36c1f27b73709ffef4eea133", + "source_signature": "724242fee87a5cf28ea6831c9dec0f6376070baf5610f50008176e1f8a505344", + "index_sha256": "3a1bc1446244912e94def259a959f06a98dc688c406546bcf7a600e74f445597" + }, + { + "file": "shards/hfd_third_party_multi_v5/00027.tar", + "files": 446, + "samples": 446, + "raw_bytes": 334793805, + "archive_bytes": 335585280, + "sha256": "0dfc7b642f4a8831e9f1929b37cc2a204f8e1f8e536e8f3d3f0a94e4695bae84", + "source_signature": "cf8cde06fc8aa3c887cd7fa914108d4a201d04b3596ea79fe85999b0463d1501", + "index_sha256": "8510b775122e53435f5e78eed38a047cc01c138c168f20c0b8bc002793561358" + }, + { + "file": "shards/hfd_third_party_multi_v5/00028.tar", + "files": 435, + "samples": 435, + "raw_bytes": 323348291, + "archive_bytes": 324126720, + "sha256": "d3cbaadbec108a5957a74b6c1ad502ed6468a77ff2d72418282d6f057b49c193", + "source_signature": "63ef002d873bc188402c47bd2543e3f0be5ea8ae60f217cfbb02c3cbaa54e490", + "index_sha256": "7db0487ddfe962c22d850955d5f7c79b537f61cf5bb2c4e2650f31125522568c" + }, + { + "file": "shards/hfd_third_party_multi_v5/00029.tar", + "files": 426, + "samples": 426, + "raw_bytes": 318982123, + "archive_bytes": 319744000, + "sha256": "408b5420b884930ae570aa47e5df155de535f0c31c57f8f62f507455a83911e0", + "source_signature": "600aae7a032ce924341caaccf9c664e2f04405099bdce60c599e68912276dccd", + "index_sha256": "6103e20cf77f5e19c4a9c79f1cc8b0f6748d8e04f13fd26b6b408546d0186b9f" + }, + { + "file": "shards/hfd_third_party_multi_v5/00030.tar", + "files": 452, + "samples": 452, + "raw_bytes": 313982962, + "archive_bytes": 314787840, + "sha256": "530c047ffa98f2eb76b57fa66f03d2c669ee26c37cb1fbf0bc89a4ed5e243820", + "source_signature": "2faa66b81ad6c701ffb87619f75c7ca324a208dd39a14e1df327ef25fd1cfa0d", + "index_sha256": "2828e85969f200d8cf49d25b26f360092b1cbd634caf6b3dd08eafe783d5b7fe" + }, + { + "file": "shards/hfd_third_party_multi_v5/00031.tar", + "files": 402, + "samples": 402, + "raw_bytes": 307959600, + "archive_bytes": 308684800, + "sha256": "da57045f971935630a5f9d3f0daf808aa2763f2664cb7e1979ee40ba25d2c451", + "source_signature": "c0e3d3b0449273a4dd8d1d35ac2f0ecd9384299c2216e4f5b472cb9d945330df", + "index_sha256": "49d9476a8d82c87ddd309cd533aae1fffc71ff39566688e279d981dec817cec1" + }, + { + "file": "shards/hfd_third_party_multi_v5/00032.tar", + "files": 445, + "samples": 445, + "raw_bytes": 340420181, + "archive_bytes": 341217280, + "sha256": "1dd72089903e036e029c048b453d7124f1cd60b08ed6a86b3938b24685212a01", + "source_signature": "d1648a93d451dedc096c28d1f51c5020b942f9f0feb92c31d3b443775bf80174", + "index_sha256": "44b28174973a68b2b89899b603652b69a3cba57cb4b98fe712184eb6c02fadf1" + }, + { + "file": "shards/hfd_third_party_multi_v5/00033.tar", + "files": 446, + "samples": 446, + "raw_bytes": 326165571, + "archive_bytes": 326963200, + "sha256": "451d605759f71271e0225db4c70f8849e4eeeb21b175a9c922c0fb2825d797d2", + "source_signature": "b5b29d1969c04777ac634d69ea145fab8c9844a301769bc86d022d3df86ee8f0", + "index_sha256": "d1377b3e908b9c07c0ad20e2397eccecabd90b8b1f3dccb85155169d3a31e741" + }, + { + "file": "shards/hfd_third_party_multi_v5/00034.tar", + "files": 431, + "samples": 431, + "raw_bytes": 312970369, + "archive_bytes": 313743360, + "sha256": "ec3abadae348565c90eff18823fc404291625ef0d51a869ccd7d357c9e9b9a30", + "source_signature": "4983e925c91548def544622533cd2a74f409d06f3686ae3225456aeceb7019a6", + "index_sha256": "fbe98a38ba8edeee2912a37955dd95b8d8bda3d675d759bd63f4b8e9946a0a07" + }, + { + "file": "shards/hfd_third_party_multi_v5/00035.tar", + "files": 378, + "samples": 378, + "raw_bytes": 288046695, + "archive_bytes": 288727040, + "sha256": "870bc04d5bc0d04937e17e8aa19538edb8492d2ce8dbf05db5b0798294a2d46e", + "source_signature": "f3768aa468a6d59a73c292f7c740063c77d285effd95385c2af72935eb652866", + "index_sha256": "e06cd07b2130089ec4832aed8d5d8ef5c4ab8258a003c88fecbb1f95a5875999" + }, + { + "file": "shards/hfd_third_party_multi_v5/00036.tar", + "files": 465, + "samples": 465, + "raw_bytes": 353412532, + "archive_bytes": 354242560, + "sha256": "b1dfb1580954f1a698408d75ed9ce2fa40ffa80bb98bab94059d741b13cb9a44", + "source_signature": "8422c221528d31d64cb42f0fc0af12623259148564169c76e19ef0b73807f48d", + "index_sha256": "9083df6d8d3c60d03744653f67bb2e6f50bab45843b11e7e3e94870ea5652e80" + }, + { + "file": "shards/hfd_third_party_multi_v5/00037.tar", + "files": 417, + "samples": 417, + "raw_bytes": 303290003, + "archive_bytes": 304035840, + "sha256": "c9baa10f19d5aa40204dadb00f883bbb3996fca24caa8a1140ccadb18e1352fe", + "source_signature": "f319f0be921989b935a88fa6831e470d4a5fc37d0b84d170bd6e73a0188e14df", + "index_sha256": "8b8aa690373ed628288c531fdbf1c21c2cbd82143e62175a5a04c76f431273b7" + }, + { + "file": "shards/hfd_third_party_multi_v5/00038.tar", + "files": 449, + "samples": 449, + "raw_bytes": 350124814, + "archive_bytes": 350924800, + "sha256": "1ee34cb4229c289ff96ea8f8a4325ff51763d9778a4c6545ca2713a8c8ac7862", + "source_signature": "cc1e86ca4e0bdf0e96cda71ab8aa44ce85d89e39576c75b2ecfc9d59ed6f26c6", + "index_sha256": "d3c7ec2af9efd93054a5107e0f431d152399ec2be9c970c3e809c4f597e794f6" + }, + { + "file": "shards/hfd_third_party_multi_v5/00039.tar", + "files": 441, + "samples": 441, + "raw_bytes": 318512622, + "archive_bytes": 319293440, + "sha256": "de6d29e96367a594b43acaaa9a71463194ad7a7af26d1cc089a7daafccff39ca", + "source_signature": "412b1f2c8fce4d086d262184127572f23b125049c08a56feae910970ef70dc6b", + "index_sha256": "5ce824fa8b3772dead7623fac9987ebfb29db779ea5d7a35db1d77b3fb309139" + }, + { + "file": "shards/hfd_third_party_multi_v5/00040.tar", + "files": 454, + "samples": 454, + "raw_bytes": 303929367, + "archive_bytes": 304742400, + "sha256": "2b5b14fd83f8572de3589a19f8983753acc65afbc88b9954419df3867b8c03cc", + "source_signature": "e6ef6da0796695b039d78c5e3f98fd6f7bd2a8be44f1f696a2a8504e405d7f87", + "index_sha256": "e4530bd56a377584fabd506d91c724d3e3954aa00417540469a860813792b131" + }, + { + "file": "shards/hfd_third_party_multi_v5/00041.tar", + "files": 424, + "samples": 424, + "raw_bytes": 313375040, + "archive_bytes": 314132480, + "sha256": "783c3781403f651364e7f5e2de2624d06cd2d9bd15cb1406f45886a97963bda3", + "source_signature": "310022c658b8a53c209ce2f30abf68cd3534af81566ecde76c4737ad686641ad", + "index_sha256": "4faed4dac5930a78fadae22586f5f74ee6722f75cd458da62e4d90b45c6278b0" + }, + { + "file": "shards/hfd_third_party_multi_v5/00042.tar", + "files": 436, + "samples": 436, + "raw_bytes": 324141165, + "archive_bytes": 324915200, + "sha256": "fc887574c022d4f94780d5739b87a7322a464f6e1a2192cf05442d43bd399327", + "source_signature": "9685907836b1484bc2b6afb4f222c845ca8622ac0afe27da942fb39eb6fe4168", + "index_sha256": "5f99ea91c2f8c1a06b0d293f42f7ddfb0120b1c7735ea8a399eb2670a493aa83" + }, + { + "file": "shards/hfd_third_party_multi_v5/00043.tar", + "files": 442, + "samples": 442, + "raw_bytes": 344041521, + "archive_bytes": 344832000, + "sha256": "ea90500972f46d0962a71198ffc49b8926b73772bac4f6b13e850cc554a5c499", + "source_signature": "072763a748b3fdb3600e268232360c09943ee9356d99c5adeb496a47f2328a03", + "index_sha256": "cdd9b9576a712dd069db26d021fcfcc414a74863d30aca26c9370903339b8d21" + }, + { + "file": "shards/hfd_third_party_multi_v5/00044.tar", + "files": 378, + "samples": 378, + "raw_bytes": 299770781, + "archive_bytes": 300441600, + "sha256": "c8ec7991acac016961e2ae2520b7bbb5006b4ad4718589b86085eaa75dde754d", + "source_signature": "024e6c49bbaea5f43dd22c5ec65da4377bc5961bd6636aca66dbdb9525a6d67a", + "index_sha256": "779a854271a3c552f891d12356a3539cd6c721b62dd9c0df9944d50b52d1b1df" + }, + { + "file": "shards/hfd_third_party_multi_v5/00045.tar", + "files": 418, + "samples": 418, + "raw_bytes": 319753556, + "archive_bytes": 320501760, + "sha256": "d2f9de075f06d47128dc48d751a4d2584cc73f329ea9c628bd1151af9a222714", + "source_signature": "daca34ecfeeebfde38cf8393067b5242cba272f3eab380da5daddad7b6e854fa", + "index_sha256": "0be0a08090d6ab89e8ef005ce9516d29329ba2a4a08d8cb891b99304f9facb7a" + }, + { + "file": "shards/hfd_third_party_multi_v5/00046.tar", + "files": 431, + "samples": 431, + "raw_bytes": 382344839, + "archive_bytes": 383119360, + "sha256": "06237f7b6491e03b5d9fb5311487cfe8e9a63e35ae22fd47242d3ea2c4b75e28", + "source_signature": "14be16516c1935405f68a3c5e30e9ad874fde3cbe0b023cabda8624723402560", + "index_sha256": "9e9b434f2f783ba162079405b62fc6c31bc1fd8914faf7d184cd836d2371797e" + }, + { + "file": "shards/hfd_third_party_multi_v5/00047.tar", + "files": 415, + "samples": 415, + "raw_bytes": 332175334, + "archive_bytes": 332912640, + "sha256": "b8079593abe90a4c79c0c5a01699e07d35092d2f7916cc26fb258618b3632418", + "source_signature": "f743088b8f864dad3ef9b57599801fd2c4638eaa2a15436e9a55fb1b8637927a", + "index_sha256": "ce83f12f111b3ae8f4aa30dbd44829e37ce36dd3828fc54e467b507b61566ac4" + }, + { + "file": "shards/hfd_third_party_multi_v5/00048.tar", + "files": 423, + "samples": 423, + "raw_bytes": 310942609, + "archive_bytes": 311695360, + "sha256": "3b964d2ee4868d8cfa81ddd0f068abe64f0906da3d4708db24d629cfb88d2684", + "source_signature": "e8278b7b0bef22a9e6ec1744428b4b608807e359e7bf2e08c26d9c7dd7e94b3d", + "index_sha256": "43d2c4edc1ee985fda8c5679953357ec521a3201cb0d5345ef7cf1f7c7b508d7" + }, + { + "file": "shards/hfd_third_party_multi_v5/00049.tar", + "files": 436, + "samples": 436, + "raw_bytes": 342261512, + "archive_bytes": 343040000, + "sha256": "ee60f36ccf5b309d2b9c96d52179daead4412c9924b459bc9f7dababb50defc6", + "source_signature": "419a6caba601954f310d1596f69316b73b9e72854dd376b8ab0809fcdefcc4a9", + "index_sha256": "7cdb2316bd98dc441331342c8a845b1fffb23dde030123c79d5116ba36285741" + }, + { + "file": "shards/hfd_third_party_multi_v5/00050.tar", + "files": 434, + "samples": 434, + "raw_bytes": 312318151, + "archive_bytes": 313088000, + "sha256": "319c0fa4bb6be2130a063ba93fd7eef3017dc45470028b7ed76ff473511c52a3", + "source_signature": "b7380bf92f9bd103e281a6ea669369994e4840c79ada39bdb95a95d6bb1a1032", + "index_sha256": "2b8d698ad2a7366b641a62373ceb16ddba77deb53918b5469d96595bc8c72ad8" + }, + { + "file": "shards/hfd_third_party_multi_v5/00051.tar", + "files": 416, + "samples": 416, + "raw_bytes": 300857780, + "archive_bytes": 301598720, + "sha256": "d0ed32b0982ad69fbffcb853daf69091f4756c8de3dd61d093402a9a298075dc", + "source_signature": "3775c04c063ff148523e88af3cd5822d2327122152b7c6dc4e5b5f4ee5d5d577", + "index_sha256": "30ad35274dc07651a2506d4658c951c5f06f2fff2d17e0e8d8dc33ac8484d0b7" + }, + { + "file": "shards/hfd_third_party_multi_v5/00052.tar", + "files": 429, + "samples": 429, + "raw_bytes": 305755819, + "archive_bytes": 306524160, + "sha256": "f6c66439c86803c1566bf49945fa0dd2f584d494696f6bfba3eec37f186e844f", + "source_signature": "8f2ce3042a91c4b644a4cdd87150d18d27ac3aba49b7ddf2ef10afa229a0b0b6", + "index_sha256": "ac698aa84e2d57041ac98578c93a72607a467cb7bd5978f18f97ea9a8a46e1e4" + }, + { + "file": "shards/hfd_third_party_multi_v5/00053.tar", + "files": 418, + "samples": 418, + "raw_bytes": 316783373, + "archive_bytes": 317532160, + "sha256": "0cafd8695f4916638cafbd7484edeb90c7bb7660bf0c4694d3803df4643d4ef1", + "source_signature": "975d66a7d455dcf5b20cc62841c72b0496ddde5061c2241ea1e9e0914d702a6e", + "index_sha256": "ad6da3887e598ff60ed96f3c03128ccf213e5c32cd381917c94b49da59dab552" + }, + { + "file": "shards/hfd_third_party_multi_v5/00054.tar", + "files": 446, + "samples": 446, + "raw_bytes": 333415361, + "archive_bytes": 334213120, + "sha256": "53979336f63d5738efa0af5d1a944d6fbe57943fc23b329bbb834696d479c282", + "source_signature": "9e4afe57f699ea9e08df6c146bcec9522be2da0d2e4c80d2a00b542c1c68d84c", + "index_sha256": "3ccd41b4ab732384657758eab598a493b1e9c58792458af3ead1ef061b6c17a7" + }, + { + "file": "shards/hfd_third_party_multi_v5/00055.tar", + "files": 458, + "samples": 458, + "raw_bytes": 336433117, + "archive_bytes": 337254400, + "sha256": "1c60239d1744f4c93122a50a3cd39352384af7da4bf4b1fec666ac05b476eb2e", + "source_signature": "2bab67c516737ae2f15091f7127da7541ffa8db4bb193d4b86fb11f26987e744", + "index_sha256": "62c231a78f75eb6a6ad878535efe11e2cd2af249822139a62be49ca41a83a91a" + }, + { + "file": "shards/hfd_third_party_multi_v5/00056.tar", + "files": 420, + "samples": 420, + "raw_bytes": 330484838, + "archive_bytes": 331233280, + "sha256": "2826f6946920258f62dc915c47805496fcc2f30c81b2afa4a3dfacefc9e97fe0", + "source_signature": "329faee382316f80fde8971522a9ca72dbbb565492605fa1d84f8ea462f307a3", + "index_sha256": "88e47cdb5ec411a139393b59f2cc5e1ac99d7d9a651a9b2c0a707addcbf5a8d9" + }, + { + "file": "shards/hfd_third_party_multi_v5/00057.tar", + "files": 422, + "samples": 422, + "raw_bytes": 329458708, + "archive_bytes": 330209280, + "sha256": "bb0d119a6130e9b296881e74c6eb61edc451eb8b14ad0c0e77c1cb320c32cbba", + "source_signature": "52dc1af7f4272501b36c2824d2d192fc7fea9f31b2850b92fdb4078f60cffb87", + "index_sha256": "28f5c6fbf946aafe00f1b18d500d2f0eaffb256c68fa9c285395327d740f68a5" + }, + { + "file": "shards/hfd_third_party_multi_v5/00058.tar", + "files": 391, + "samples": 391, + "raw_bytes": 293788202, + "archive_bytes": 294481920, + "sha256": "7faf38b8b8a802cf807680d5a167038d5b208b4fd824e910cdb9e7f4d789e951", + "source_signature": "a9c463452aedf540de6268b970d443b7053316f71ec15a40bdceca37ac390f4e", + "index_sha256": "3e81a7dd2e9f9a1734e4d92a19b4c9e2c82d1fae344114e68de6445cac66181c" + }, + { + "file": "shards/hfd_third_party_multi_v5/00059.tar", + "files": 418, + "samples": 418, + "raw_bytes": 290024277, + "archive_bytes": 290775040, + "sha256": "0c9cf7b2d4e1ddd13cb8ffc027fd74e6307d2c2e88908be4087ff1284684769e", + "source_signature": "aa6c29a1ef4c943363e96985c61a8ad3e77e0fd3e75d57bcb7a02d23efd2b91d", + "index_sha256": "6cee283ea686abd1e6a9fc4cc143b0bbed59a71759928715a0097be99e3f38d4" + }, + { + "file": "shards/hfd_third_party_multi_v5/00060.tar", + "files": 423, + "samples": 423, + "raw_bytes": 361286874, + "archive_bytes": 362045440, + "sha256": "df506e0c05a43939acd6380b7d71e2b525fecdd6cf72718c13a1e8b97af52957", + "source_signature": "619aed0d025649cb85283ddde206836a40714760227c19c78dda29063cfe497a", + "index_sha256": "daf385975134ef5d3afe9ad0960db2b4c576d4cc2c8afbebce8b9f6d2a134966" + }, + { + "file": "shards/hfd_third_party_multi_v5/00061.tar", + "files": 479, + "samples": 479, + "raw_bytes": 377368554, + "archive_bytes": 378224640, + "sha256": "c910aceb128d873d4b9dee0e77a5e2916929ec79ab0619d1df005132ba2098a6", + "source_signature": "4f87221fd54742ba6d4d1d6f6dc65216637cd2ed632223f1d18f6d23e9443b97", + "index_sha256": "362c60061263b316cf676017bdd0355ede6a466cf266e09007b8717b03344bb4" + }, + { + "file": "shards/hfd_third_party_multi_v5/00062.tar", + "files": 399, + "samples": 399, + "raw_bytes": 294053285, + "archive_bytes": 294768640, + "sha256": "8af54d0b5d34ced1943381260adb34101b5e2b19e0da7a571cafbc427627ae86", + "source_signature": "60e78b318f4fd1953561757f41f94d2024a258554d1f35272018d8a0c2065a3b", + "index_sha256": "6bb5718b214566a45a7b9a003472fc16f6054e7d5654a50678c477908fc4ef9d" + }, + { + "file": "shards/hfd_third_party_multi_v5/00063.tar", + "files": 424, + "samples": 424, + "raw_bytes": 330825200, + "archive_bytes": 331581440, + "sha256": "6736c26a668ac8ee870aa7217d558bbdb34dcdcd1157ef21fb148aca5588054d", + "source_signature": "c21dbd478ef979d435bee178fb4f39804e26d1263454b1a648aa3fde6979dd97", + "index_sha256": "003a79b1cc9a895d105c8c2505b841e7d8ccb67c34530676f7ff01a5dae11cb7" + }, + { + "file": "shards/hfd_third_party_multi_v5/00064.tar", + "files": 414, + "samples": 414, + "raw_bytes": 319954339, + "archive_bytes": 320686080, + "sha256": "0718f9c61bcb321f5aaa43239102428d887b5ebc832c43f4cab0b89576b33dc3", + "source_signature": "ee2c9d54ccdaf9f4969dbfa295fe5d8d5d7be321c9d059592d2268d57b4d6e49", + "index_sha256": "c5df8a454d65fd1dbe62d803fe0e8e2bf250f72f4365d56841e03384233134c2" + }, + { + "file": "shards/hfd_third_party_multi_v5/00065.tar", + "files": 443, + "samples": 443, + "raw_bytes": 374302139, + "archive_bytes": 375091200, + "sha256": "6ee30fce9882e0adb037981dd92179d304c0946dcf841ee78aa160388bb8a716", + "source_signature": "e9e60e40f8e2ba4d7802a5a731c39afbdc8c312b558e39b0d65c71561d672848", + "index_sha256": "e649fd622d9fea6b9d625a4d678199b0dcb59cde4dc9da53ca6ec7f3facc86ce" + }, + { + "file": "shards/hfd_third_party_multi_v5/00066.tar", + "files": 431, + "samples": 431, + "raw_bytes": 319549428, + "archive_bytes": 320317440, + "sha256": "12bc796f01af4f920c22c8f6fff10e85d59b5e06b08c096abf24496eb40f0a14", + "source_signature": "062c057bc8fd14897b57827d76d434a3327b9069f2521579d6082f0c1d01760b", + "index_sha256": "0caed2b5d052772ac5317e06ab36616c2b98869ea265183b4e6f3e225700323e" + }, + { + "file": "shards/hfd_third_party_multi_v5/00067.tar", + "files": 436, + "samples": 436, + "raw_bytes": 337156352, + "archive_bytes": 337940480, + "sha256": "86f1ffb18c7e1ca2b057d8aa8877d559e9eac90c544bb5d96730d354ef7f3172", + "source_signature": "afce993cd6e765e16168f215bedccf7959cff8f69b5310dd686031018d84bb15", + "index_sha256": "dc823c665a2b4361b4fa06f41df0ce81e0dc5be6231f02e401d392dc172e158b" + }, + { + "file": "shards/hfd_third_party_multi_v5/00068.tar", + "files": 440, + "samples": 440, + "raw_bytes": 332283154, + "archive_bytes": 333066240, + "sha256": "4718392b80bc55fc6c2dd74ccab2ad271e4fe2b302fb2305ebcc8d2d692e6c3f", + "source_signature": "8129073ec1b3edb8a7a3aa9d5d1c80e0eec16c4c228c2aa064d9d4a84176752d", + "index_sha256": "f2017ad65f96892f883875dedefb1d0c522372259c3b56892db564151a98fbd8" + }, + { + "file": "shards/hfd_third_party_multi_v5/00069.tar", + "files": 413, + "samples": 413, + "raw_bytes": 339427514, + "archive_bytes": 340162560, + "sha256": "1c4948bb65c546ae42405dfb6199c2e1326e9b6c98e4af3e1a173b712eb27e1a", + "source_signature": "26e7e04b91533f291a8d290fe52678a1a48624ff496bd5249382166e924f0754", + "index_sha256": "4634a2564af9d103bd27eb481ecf8766ff00763c7068e8cda256cbbf68d47e76" + }, + { + "file": "shards/hfd_third_party_multi_v5/00070.tar", + "files": 420, + "samples": 420, + "raw_bytes": 305078796, + "archive_bytes": 305827840, + "sha256": "0bf61911897478814f197295b5cd6ccf43ce3676ee1b185a1d0bac87164ea871", + "source_signature": "612cdc22e2103ae1648a65e705be9a1d7fa1787073cec1da8178cf65b1aa74e8", + "index_sha256": "27ede076aa0a1be6bc021650be2d8da5fe0a9889f67e75a552c96c59131dc529" + }, + { + "file": "shards/hfd_third_party_multi_v5/00071.tar", + "files": 475, + "samples": 475, + "raw_bytes": 364179715, + "archive_bytes": 365025280, + "sha256": "20abc8ea01326cf3470d6a2a64bda1c7bf54b06094075488a98c588e541fe80d", + "source_signature": "3158f5833b53b542a2a91b890b1009b9bb7dd3e1e5e7f117a221a469fdcc1754", + "index_sha256": "8feb82c7ab4cbf83a1b4d8b8e261ef40b096ae6b360d67271a9522b94d55d044" + }, + { + "file": "shards/hfd_third_party_multi_v5/00072.tar", + "files": 422, + "samples": 422, + "raw_bytes": 318704713, + "archive_bytes": 319457280, + "sha256": "d64fb6fc7821b61a13de248ce42c1ca2de18bc29474e319c9fd90477cdeecd16", + "source_signature": "6fda69e020928d58da07b4aa6625564c1763825753911d89d7eca1063d329e7e", + "index_sha256": "65955099b70d5ccc7ec83082bb643e31aaef4ccf2b409250e073bc92932de83d" + }, + { + "file": "shards/hfd_third_party_multi_v5/00073.tar", + "files": 449, + "samples": 449, + "raw_bytes": 341889376, + "archive_bytes": 342691840, + "sha256": "d93bc4e91644241e2ab4e5aa14d0cb24a23e46c515a22ebd1a7a72afc05ef8dd", + "source_signature": "c970da67294f3ba594a328cba1214cc3f8b7797b0a9a0b53542a9b893d16290f", + "index_sha256": "bf991ab5b4205e7d9eff642b71a00e3688e30900d5bf82e3efe264815c3ee704" + }, + { + "file": "shards/hfd_third_party_multi_v5/00074.tar", + "files": 476, + "samples": 476, + "raw_bytes": 335926428, + "archive_bytes": 336773120, + "sha256": "1aa0611545419f89622248db0edab678c1d03212149d89519fc92d3d0546c80f", + "source_signature": "5c756fea4b649069e926983da11f618dd761097ce725982568c8044c74b3851d", + "index_sha256": "5bab6d54a3a429f763d59a7d2cc8f4e178bebc0aa67504e9e411eaf6a17b18fb" + }, + { + "file": "shards/hfd_third_party_multi_v5/00075.tar", + "files": 444, + "samples": 444, + "raw_bytes": 879627979, + "archive_bytes": 880414720, + "sha256": "4c50161a37d5bc8a7f0f25ac2bb68c19bc42329eeb0d46c663cb7cdfbe8b0eeb", + "source_signature": "b12e434d95f9aaf2ebb85fd624bd54acd27b2de25f1b1c7475226f1f6bf3600c", + "index_sha256": "df3e767fa8ddb517403a9d21c284f1ade0fcf37ec4e162224e4a33e9f969c239" + }, + { + "file": "shards/hfd_third_party_multi_v5/00076.tar", + "files": 434, + "samples": 434, + "raw_bytes": 660161453, + "archive_bytes": 660930560, + "sha256": "d2cfd39fbc700e851bc3fc394b2a018bc7c29f5bc0aa46b3c3e2a8332554d85b", + "source_signature": "100333be5bacd1fa88c364ded00aa2295b3b8e49aff38819e9f175d4bc32c0e7", + "index_sha256": "1ba82dd2bd83c56e1dbc51cd0b1da7bd7562718b3db8e536cf6eda53f9007e2f" + }, + { + "file": "shards/hfd_third_party_multi_v5/00077.tar", + "files": 422, + "samples": 422, + "raw_bytes": 337745905, + "archive_bytes": 338493440, + "sha256": "943aad8ccf298dc51076d26e8923e5fa0ef1bc305a0321765818e5e7c8a4e483", + "source_signature": "23213318d58c88e2ba1f90f111f396c3a80704c2eafb8b1c42e3c3573b47cba6", + "index_sha256": "83c6e8bb6d560dba5355025b383a4193fa733b1feff72f950a1b40cd70b93126" + }, + { + "file": "shards/hfd_third_party_multi_v5/00078.tar", + "files": 444, + "samples": 444, + "raw_bytes": 375356119, + "archive_bytes": 376145920, + "sha256": "73833e6967801ce6b67a8fabb378f38c88b78b81425086425f47c9f6de713746", + "source_signature": "fe55713b7cce8eb8188acd07c263eb6c414dd5bd85dec9644c928fb5ffcc98ee", + "index_sha256": "09b04bf5c3c9aab19ba82a6f277001b6d85dc00262f76529c6ed560cc980ddd7" + }, + { + "file": "shards/hfd_third_party_multi_v5/00079.tar", + "files": 424, + "samples": 424, + "raw_bytes": 318954799, + "archive_bytes": 319713280, + "sha256": "46a26496bbf1eb956fa24669f4a20becbea17213b10fe34a60a93769b0cdd83b", + "source_signature": "add4811db8fe00f894a95fd9c720dd71f37f0d1dc4b4d6dee4c16babcb3d9546", + "index_sha256": "b60e1f8fe5a417e9a63e8a6ce65b26ea010e00cb2193e021900aa00d94988902" + }, + { + "file": "shards/hfd_third_party_multi_v5/00080.tar", + "files": 409, + "samples": 409, + "raw_bytes": 304845248, + "archive_bytes": 305571840, + "sha256": "16caf6267f03c79dc0f1f1d4fa0c62efabe02e2e7656d52459ea33f8701872da", + "source_signature": "608a46cd859bc54995af38ca9d8272c2926b306a45b983a287d73f812386a9e3", + "index_sha256": "1cf63e74208d29d814abccdc5628020f27268bd39575bba990189a67451b6c23" + }, + { + "file": "shards/hfd_third_party_multi_v5/00081.tar", + "files": 427, + "samples": 427, + "raw_bytes": 316657238, + "archive_bytes": 317419520, + "sha256": "29b2af650a82e214da22a53a458f4876d7c3dc08051c8b0784e20b6bd4fb4be8", + "source_signature": "0fad8b8dcfe7971d3572bd6e233b19eb46a25e4eeee46ee90e77d56ee4eb6b2c", + "index_sha256": "9a1c3b08cba46d9063d1dcc2e4dacfc016362c2d7cdbb12e89e631855d223413" + }, + { + "file": "shards/hfd_third_party_multi_v5/00082.tar", + "files": 425, + "samples": 425, + "raw_bytes": 337977201, + "archive_bytes": 338739200, + "sha256": "07ee3554badbb246ef26d4eb6143e8b5900c1aac3106629a629402b054155249", + "source_signature": "e78a7545dfd69da6a5a5f1eed620d7f4153866c3a2d0091f42f082fcfe7ceea2", + "index_sha256": "8b8f5026c76cd7e80feb9a447145486324f4be7f5d750e73df8b6cab23c37652" + }, + { + "file": "shards/hfd_third_party_multi_v5/00083.tar", + "files": 394, + "samples": 394, + "raw_bytes": 282219732, + "archive_bytes": 282920960, + "sha256": "14dd731408b707bc481303754614584052daee43c2c5299cc287fd0c192aa240", + "source_signature": "b70959829b1cd020f21b5e53f1a9df33b741c72e0e05c57ba14b63908b8b571c", + "index_sha256": "c98ff659f833229a46234e90b5676538167fd79edd8d764c252f292d6f23c14b" + }, + { + "file": "shards/hfd_third_party_multi_v5/00084.tar", + "files": 438, + "samples": 438, + "raw_bytes": 321689523, + "archive_bytes": 322467840, + "sha256": "a8058395f87919342e195cb5eb3cdbba9eb76d9a2a0d75cd245b9cfaefde1bfa", + "source_signature": "526e43ccf3baaf3e31e9b3d4af294fb0ba417be470e3188c05fc43aa85fc5dad", + "index_sha256": "d7207cb8bb271bc0760ed7adf314eaac60c6d93f2af2b49c44e4f641bbc675e3" + }, + { + "file": "shards/hfd_third_party_multi_v5/00085.tar", + "files": 443, + "samples": 443, + "raw_bytes": 348257719, + "archive_bytes": 349050880, + "sha256": "bb83da45305d5f5e9a5609f14f4eaae54fc831387ded8ded934efceefc64b3a2", + "source_signature": "11520473ef3c0f34e490d6a7605995e6248813630788ee74793dc257fa43e4b6", + "index_sha256": "f10f2488890b605acf2058f583f797ab4a9256885d6a980f92e683fee6885a78" + }, + { + "file": "shards/hfd_third_party_multi_v5/00086.tar", + "files": 451, + "samples": 451, + "raw_bytes": 333939538, + "archive_bytes": 334735360, + "sha256": "6d254737bf80078b32feaa76804868603fe3f3bc20f2a298dd3fd92376137084", + "source_signature": "8ff793d4f5e410891e2a5388c97e02b09abad1a756affd9db4b0dcef70654e5b", + "index_sha256": "4ff28c88af928de7af6d2cc6797b591e6642aff780b90cf0c2b442e311ac697d" + }, + { + "file": "shards/hfd_third_party_multi_v5/00087.tar", + "files": 445, + "samples": 445, + "raw_bytes": 333019177, + "archive_bytes": 333813760, + "sha256": "b2f33b342fa65cf5b2eb53d71a46856255ab5c79959985a23ed73d60a1e89e06", + "source_signature": "56c8ed68d5543ac71d739e34e66dcf00d46479b98852581e786275152eef5fac", + "index_sha256": "f71c7c6786a9d2cb14155be019b9001832ba62a9ae4e83ab756316375feedb12" + }, + { + "file": "shards/hfd_third_party_multi_v5/00088.tar", + "files": 419, + "samples": 419, + "raw_bytes": 305450342, + "archive_bytes": 306196480, + "sha256": "324dcb11c27ba0db6a49629115ac2d6909055880e245fc48afb2affdf3d19d2a", + "source_signature": "d0b9fc3a156272421bb4b6593b991982edba451ccd0efe1d1ba65a54240ded7a", + "index_sha256": "81989700ef172e5686377b094ef7e0bf40a4682fe501095d0664eb898f3ab5df" + }, + { + "file": "shards/hfd_third_party_multi_v5/00089.tar", + "files": 435, + "samples": 435, + "raw_bytes": 326063705, + "archive_bytes": 326840320, + "sha256": "3e8807a87cff8036d6cbb20ab87951485641e3184e8d652b13391e14d0d577d5", + "source_signature": "c91a8a930f9807c00b22c1b6bcd88d1d737d7f22473805696df084ae47d7697d", + "index_sha256": "813c5a30d297b4004a96d120dbb4f38ddddbf714191c3efee7ec1dd3e5e86335" + }, + { + "file": "shards/hfd_third_party_multi_v5/00090.tar", + "files": 416, + "samples": 416, + "raw_bytes": 317770515, + "archive_bytes": 318504960, + "sha256": "5a435445d686904cd601fbff9c3115d34919757da07a7d7bd19e727d23cb4c42", + "source_signature": "ed6de1b3e1b4c1d3ed6032cff6bf557e078a20961e8321bb3d575efa23b4ebc2", + "index_sha256": "9ef4ddb71735586544698a2e2cdc95e2bcc711d2ca0fa9dd897567e748bd8ac7" + }, + { + "file": "shards/hfd_third_party_multi_v5/00091.tar", + "files": 407, + "samples": 407, + "raw_bytes": 295735208, + "archive_bytes": 296458240, + "sha256": "7e16396c8e81d4eb41fb0103e013cfc8ce1cf8980c5332615996e5538adee7e4", + "source_signature": "687581956a0bf6ccf25381e29a948099f905052ddf15c97d0ee67e57fc88f56d", + "index_sha256": "046a9b652de23fc8da2cd2506fe9da3dca835e30a64d2e48520db91f14a65c6a" + }, + { + "file": "shards/hfd_third_party_multi_v5/00092.tar", + "files": 427, + "samples": 427, + "raw_bytes": 329201109, + "archive_bytes": 329963520, + "sha256": "34367e7d484fc3aefe4215173c1101785e9ebb65da566c351c99d53d8f028871", + "source_signature": "f7d9bc61048d1c71aa83b04b1950bb6121f214b86f76ac95070f280fff888dfb", + "index_sha256": "9a9bee512cf1bc9e62431e3271547283155680984568a07f5b257afb24936b1c" + }, + { + "file": "shards/hfd_third_party_multi_v5/00093.tar", + "files": 421, + "samples": 421, + "raw_bytes": 322820043, + "archive_bytes": 323573760, + "sha256": "92c1c77f7f05e2e929089d635006feb045fe45c7edf2010c7c75e8177e8e0c7f", + "source_signature": "2563fd647dd7fb5d1118920315f3a8b9eb9bedff9621bd3c4fde2d0f8bee8ee5", + "index_sha256": "5456faf576e9c1720218da706ada2ef8379513906fb5b32a0018534073643ef7" + }, + { + "file": "shards/hfd_third_party_multi_v5/00094.tar", + "files": 426, + "samples": 426, + "raw_bytes": 294094725, + "archive_bytes": 294850560, + "sha256": "455336914fb1e3821534106f941f2546e13cf07f8454f812e1a3fed0ec1e4843", + "source_signature": "1197ae62bb81a193436362b614b791fd00011523ebb28107e9357e298f304dde", + "index_sha256": "740d1bb82c3562aeb83990c9bf7b02daabe6dbc7f17773cc2d49e8c413e8b80f" + }, + { + "file": "shards/hfd_third_party_multi_v5/00095.tar", + "files": 420, + "samples": 420, + "raw_bytes": 331501870, + "archive_bytes": 332247040, + "sha256": "172856a0749ad3f406812faa3878097632622c7869d5a14e76e33505104b27ec", + "source_signature": "f01820f0599b0c0da5d7c792111636fc5d7f0559f12a78f6c54c6efea0b2eaba", + "index_sha256": "d6f4c20e67f356981b910762c10e08fc1a0377366123b9e5cb047351621f3060" + }, + { + "file": "shards/hfd_third_party_multi_v5/00096.tar", + "files": 408, + "samples": 408, + "raw_bytes": 304151438, + "archive_bytes": 304885760, + "sha256": "8c7df12f3bc98e70660e95a689dc974114f27db74e8e61052dcac8bd8f68995a", + "source_signature": "38088b0343c3e78f1880b455a61d20276703591866526237d17fb5d3a11bd0d1", + "index_sha256": "ba16835b167b57f02ce9f902c17eadf66523b9b7beea6271fc9d20609c949593" + }, + { + "file": "shards/hfd_third_party_multi_v5/00097.tar", + "files": 450, + "samples": 450, + "raw_bytes": 351573686, + "archive_bytes": 352378880, + "sha256": "5689e37f949062daa7268258b07e796c45060c12dfa2dc05a392f228a66743c7", + "source_signature": "a72eafe5bbe84aedc20da196243330ff5807a9b2c7c1a3323a010a1531eaa171", + "index_sha256": "67abeb84ee3decc6c98a053e377216580b5d972055cfd94ee1cc68602e41df77" + }, + { + "file": "shards/hfd_third_party_multi_v5/00098.tar", + "files": 459, + "samples": 459, + "raw_bytes": 342589749, + "archive_bytes": 343408640, + "sha256": "f130d0cb5661873cc4734a8d6ac3ed18bd0d950e77e43544a0d72c341ef2a86e", + "source_signature": "84c2f0e94e9a3b00339410c11794c8b400b66f1092f13cbf53a18576d1926c7d", + "index_sha256": "6c444999fe80e5dbabb5775ef060bc9835696aa81a9a09f7e7512af37b11c37e" + }, + { + "file": "shards/hfd_third_party_multi_v5/00099.tar", + "files": 433, + "samples": 433, + "raw_bytes": 352968700, + "archive_bytes": 353740800, + "sha256": "fa0b85e837fd126d41337607570ecf164123f9871923f07650f4ae5bd22e1fdc", + "source_signature": "1e5584b930233367392f90342c691a498bd8ec14c573c8fe3a330e6eac6a3376", + "index_sha256": "9e293dddccc75aa0c298f1e759ad0843306b615252bcf6701d92da9217997826" + }, + { + "file": "shards/hfd_third_party_multi_v5/00100.tar", + "files": 376, + "samples": 376, + "raw_bytes": 276552601, + "archive_bytes": 277217280, + "sha256": "41084fab083711d71a75c0e55e18135efd9f3f1e6208954e0929223925c36ccf", + "source_signature": "c4653d61a7273c604698035c2147499fceba2ca0b847f199e3833db0ada8069e", + "index_sha256": "fcaa7cb1c6e690ceb4a0b3b02a0387e7054aa8ce0ebaf1f3fe2281791365656f" + }, + { + "file": "shards/hfd_third_party_multi_v5/00101.tar", + "files": 418, + "samples": 418, + "raw_bytes": 318683330, + "archive_bytes": 319426560, + "sha256": "2d7eb1e76cb19cc4785a7d6abbeaac401e0a9c79077e0c9d347beedfb5bc8b17", + "source_signature": "3fa91adcf5d9b64d61af8616c86ecdacd519e8c8e7d96af932137ad8ee27dfc8", + "index_sha256": "a45290c33dce8fb89ba8dea60c4381e9ff9b6ff6b701b7c6054ba921c8e3f3a5" + }, + { + "file": "shards/hfd_third_party_multi_v5/00102.tar", + "files": 436, + "samples": 436, + "raw_bytes": 322700304, + "archive_bytes": 323481600, + "sha256": "9009b46e01572019721fffc81f75f96e9acfd1e7dfd291483e2129500db2740e", + "source_signature": "6aa118081d116de88ad9c2c494f3e34bfbfe50bbe2cf01117011ee9807f40c4b", + "index_sha256": "1af6cdb2a5678a3658b1209b0006e4d0b71144ecf68c8cf065fc23609b978a38" + }, + { + "file": "shards/hfd_third_party_multi_v5/00103.tar", + "files": 419, + "samples": 419, + "raw_bytes": 320659383, + "archive_bytes": 321402880, + "sha256": "0e278403b429cde12926b37b2bc03b19e0bfa96574e2fcdacd95b754f744d936", + "source_signature": "ef34ebf67f3f57c41661ad107520ca91e8ff551c2fa5893883391d1692b2e57b", + "index_sha256": "0a8a7c51c34309b5c800fa022baa84a3aafe1ff5e48ff6db6147bc7ff5e60f06" + }, + { + "file": "shards/hfd_third_party_multi_v5/00104.tar", + "files": 412, + "samples": 412, + "raw_bytes": 320702509, + "archive_bytes": 321433600, + "sha256": "e4c44beae821cd55db0dc195bbfad4bbb64577000f90eddfa054bc5b97994414", + "source_signature": "b43fecbab0c209fac9950bac471c1160ab008b5d03f06a2b8bfb9ad1a5f22b25", + "index_sha256": "c6e0aea9a16fd19891410db1c6c166a9aa48f23f624a360735de19e6fe2bce70" + }, + { + "file": "shards/hfd_third_party_multi_v5/00105.tar", + "files": 433, + "samples": 433, + "raw_bytes": 349325338, + "archive_bytes": 350105600, + "sha256": "e43a251edd9fc1425189dfb4658f1b1e9f07e79b7a6259da3b84a21b1408ed07", + "source_signature": "917cb61ceea6e883579535b026d635dbc0a82debe16db1ccc29057865c423b25", + "index_sha256": "2f779bb92fc43b1505c23610a12e13c44ccc1df8835dab26b2906b2c93d3020e" + }, + { + "file": "shards/hfd_third_party_multi_v5/00106.tar", + "files": 447, + "samples": 447, + "raw_bytes": 351883120, + "archive_bytes": 352686080, + "sha256": "77fe843e912f1c1682f075a9feaaeb85aec3ee2b5bacabe337d9d92dafc31303", + "source_signature": "1769bb3a3cf7bea6c5b35d3fae8531b0580ebe07d4fef55f7dc80819bc90347b", + "index_sha256": "20451cb79c451e305132e876c57b9489529d137956717fb6288736d7765b409c" + }, + { + "file": "shards/hfd_third_party_multi_v5/00107.tar", + "files": 451, + "samples": 451, + "raw_bytes": 318887624, + "archive_bytes": 319692800, + "sha256": "432ab61c66614615c4545999f15f67b5e63f81f2fc885e24b923db4809efe176", + "source_signature": "92a716e3a4217afeaf131f8baf31d09f4fdc0216acb8013edefca41c66840c48", + "index_sha256": "a3a5ff4e197b8ec9f74ffe0ce3f029ca595a6653aa50f0158e74febb9a649d40" + }, + { + "file": "shards/hfd_third_party_multi_v5/00108.tar", + "files": 443, + "samples": 443, + "raw_bytes": 330947538, + "archive_bytes": 331735040, + "sha256": "478dedfb846ad42e5df1da3560538882882e2040c97824ad940ddf6d818476c3", + "source_signature": "30d9dc3db0f6abdb53603e46c0a1c18b5401c1a6f13a4cceaa4ed31fd6c26305", + "index_sha256": "953f8aaab285b0cf077697ef49986d7b71ec2d69abc772ea0afd980c6a7d98d0" + }, + { + "file": "shards/hfd_third_party_multi_v5/00109.tar", + "files": 464, + "samples": 464, + "raw_bytes": 338853120, + "archive_bytes": 339681280, + "sha256": "1bd91bc9e20e28284ab80688416e5f4810e2018e4b2a89bfe54af5a9d49ab9f1", + "source_signature": "67a2fb852d977b2745c22ea69d00fe81431272a49137c0b17960b109079c0226", + "index_sha256": "b7bea9b85729a24cd94138850fbbecceb57b071655735545496c9d071210359a" + }, + { + "file": "shards/hfd_third_party_multi_v5/00110.tar", + "files": 407, + "samples": 407, + "raw_bytes": 311962280, + "archive_bytes": 312688640, + "sha256": "00565ea5b764375bc987c764b0d685bbedadf88df84abab5148724dba52fc839", + "source_signature": "7baf2081a8056be19b2c2b3626a8ef5e59179083535f572e158e4928bac68da1", + "index_sha256": "b6de6edff082cf4de9da189f0e2fee48823df3be70e6a5cbca56f593acb0b735" + }, + { + "file": "shards/hfd_third_party_multi_v5/00111.tar", + "files": 413, + "samples": 413, + "raw_bytes": 336636286, + "archive_bytes": 337377280, + "sha256": "03870790c50d943cc37487ad5cfeca0bb37290b3d730ca19c65fe38239607276", + "source_signature": "e37f04ce559195dec71c93a76a4452f2c554ec72fe64b0abd45e1b1ca9290695", + "index_sha256": "d71bde98bfa031f87a4573709626103faaf82a12250ff23c6b83cbb5507c239b" + }, + { + "file": "shards/hfd_third_party_multi_v5/00112.tar", + "files": 447, + "samples": 447, + "raw_bytes": 355346689, + "archive_bytes": 356147200, + "sha256": "91e52fbb379e9e7e967b1f2b68f9b234a0d8abc93990e7523c4a2de37a7712e0", + "source_signature": "55d8c32c0b9ff90dbe6542fcb7188e7f6238afb726e17e07e7a8fad61f522985", + "index_sha256": "cdb6c30cc7fedcbd24d06fdf59093edb1f3a0338ee7038df1e135da151024712" + }, + { + "file": "shards/hfd_third_party_multi_v5/00113.tar", + "files": 413, + "samples": 413, + "raw_bytes": 325580403, + "archive_bytes": 326318080, + "sha256": "0a3b9de5e18ff311689c1c8589fe4cebb348548edd5149274f4c61466a9beada", + "source_signature": "1e74c8fa24ec96db2f916eb2a83b2dca78f50df71353b5a7ae1a2ea21d0e04db", + "index_sha256": "20e0750c73f22f38d0ba6d4bf7b69103c94f9e8a4eb22b4392e55248dd30c8fd" + }, + { + "file": "shards/hfd_third_party_multi_v5/00114.tar", + "files": 399, + "samples": 399, + "raw_bytes": 319905496, + "archive_bytes": 320614400, + "sha256": "d2753ebb97ecdb76c38b0aa6e740b987d14314e9abd46c21172e54595a113482", + "source_signature": "aae7ede39da7b5643d5e7bc5081803b9d965e2cbf76ed344679bb3a71df3ef41", + "index_sha256": "47d309e0cac98f1f7d6db8e5759142af26c85030eb54e7f4222eb19236b8df8f" + }, + { + "file": "shards/hfd_third_party_multi_v5/00115.tar", + "files": 412, + "samples": 412, + "raw_bytes": 295265431, + "archive_bytes": 295997440, + "sha256": "ffd07ee7cedd3c50f9a3603aa90f236db6e6c649d2f561ef2a5f532104171aa3", + "source_signature": "b395cbbd8091867730eb5abf3031d015b30ca2cd90c83a380ea21d97973abc44", + "index_sha256": "468a3befba53da4b88076544f5daf0210d55e55033f0b7cf6502f7f08e217b64" + }, + { + "file": "shards/hfd_third_party_multi_v5/00116.tar", + "files": 436, + "samples": 436, + "raw_bytes": 330154363, + "archive_bytes": 330936320, + "sha256": "d1d12233bdc20e3690664474456c2c041d3118b6a914f065f512b2cb8f7356da", + "source_signature": "239d9fb37b4607169475eb124d2685d11dc77ddcaddcaf3112d7503a71fa3778", + "index_sha256": "e94dd9e91ee955fe18760ddc861494babcedeeb70bd62fe4ffdb76fdb5d4f67b" + }, + { + "file": "shards/hfd_third_party_multi_v5/00117.tar", + "files": 436, + "samples": 436, + "raw_bytes": 310954281, + "archive_bytes": 311736320, + "sha256": "54660cad25086216a0169743b8d16f39d8b34ab0baed1ebd4b369cf8b5ac1f63", + "source_signature": "15315e62dcf4a5a4db175ac1e0a352a2784538eb21b4f6aa3018bbbda106440d", + "index_sha256": "f3678e18cf08be378011472f2d33540f1e5f88f9cbcb4bd1bdfeff14de8b1380" + }, + { + "file": "shards/hfd_third_party_multi_v5/00118.tar", + "files": 420, + "samples": 420, + "raw_bytes": 325491824, + "archive_bytes": 326236160, + "sha256": "b6d42e0d5cb852d4e47072df4ebee01437934c4a75d3f09868af5de7bb2b6cf0", + "source_signature": "68e79be24a49c3b8bd7c4115b27757407f63860cbef225077754b0923b64ac42", + "index_sha256": "cbc6f5ee586fb5516b91e9c59179a0767f2e347730153630ea713be598444796" + }, + { + "file": "shards/hfd_third_party_multi_v5/00119.tar", + "files": 429, + "samples": 429, + "raw_bytes": 322443057, + "archive_bytes": 323205120, + "sha256": "7217213264c4094adf23e18566abc44e3e98fc1844749877d3e72d11577d3293", + "source_signature": "d5ec3f45b8034660ab69ee6611640a302b728a452d2803ba0a57ee1bc89f9dbf", + "index_sha256": "697234705e57bf8fc6ce3ad0de952745f543f45fd057c622a701ab57b078bd58" + }, + { + "file": "shards/hfd_third_party_multi_v5/00120.tar", + "files": 448, + "samples": 448, + "raw_bytes": 350061846, + "archive_bytes": 350853120, + "sha256": "c0e4e0b4e809c2e8147aff4122eb5c281497f5da7af081696b8e3a884484c095", + "source_signature": "33a126e48a399d83382c6e582ea04116f8aaa18afc3ae9d291939d50e8df1457", + "index_sha256": "6e6a47b8404c08e27ce34901f7553097f5276b36a343487d36f2beb76981f33c" + }, + { + "file": "shards/hfd_third_party_multi_v5/00121.tar", + "files": 444, + "samples": 444, + "raw_bytes": 336770997, + "archive_bytes": 337561600, + "sha256": "770a17dda0904f80c0df41774b12a161d9c8ddcb2040518efb7564b182134e2d", + "source_signature": "0a0bbedc5fc815fa66fdd2ee39b1111ff31c85a24cb574b62122662c37a3ab1f", + "index_sha256": "bb6d5b99fa188707790bffadeab7b5c0ec5af2d9ea622c106b23c4145f37107e" + }, + { + "file": "shards/hfd_third_party_multi_v5/00122.tar", + "files": 394, + "samples": 394, + "raw_bytes": 321937989, + "archive_bytes": 322641920, + "sha256": "9fc9130e5696b47e228de04ee2b8d64a5bb459c0ca313bccb5cd209ee734535c", + "source_signature": "5038a7a53e5ec0b043320013f7206533384808f93dccc8d672f446ae66f09db6", + "index_sha256": "c1a632e064416ffbeffff8298374aadb3baadba9979b52936a0326144500dd77" + }, + { + "file": "shards/hfd_third_party_multi_v5/00123.tar", + "files": 426, + "samples": 426, + "raw_bytes": 310417780, + "archive_bytes": 311173120, + "sha256": "92a88fe95bf95b266d0864c87d1116cd8987b5d5d1ded312eaaa779632bfe747", + "source_signature": "8d53974612638e583f12ebef6460a396aa3fa54a6700a20f1000f9ebaf6970a9", + "index_sha256": "d6c7db084a57cd1e99e4c91b91c7ede517ed49537ab449fd2e890d03b3f5a0df" + }, + { + "file": "shards/hfd_third_party_multi_v5/00124.tar", + "files": 435, + "samples": 435, + "raw_bytes": 317858304, + "archive_bytes": 318638080, + "sha256": "e7483ec65364dffa4011e214c2a2c96b06e21dabb95c555c6a9b2b111cdbabc6", + "source_signature": "f2f3d92de4d569ad2520e63f912ee74b70175abf25f1d0c42910e28505d29295", + "index_sha256": "320c59c5a5e9188ecadbe20cef7b56c4ec3193c7c1ffd6888f52ab2c9dc344ab" + }, + { + "file": "shards/hfd_third_party_multi_v5/00125.tar", + "files": 427, + "samples": 427, + "raw_bytes": 328831942, + "archive_bytes": 329594880, + "sha256": "5eacf25f9fccaff30b85173cb48e424b8b125a35164adbaeab2450f83f5037d0", + "source_signature": "d187d841846caadebb3f6dcc3833b8ed8b163bd40534f8e2156c2b80906748da", + "index_sha256": "bbbf219100475fa278154ba115423d9ac092cae96cac14ce2100377e7676d64c" + }, + { + "file": "shards/hfd_third_party_multi_v5/00126.tar", + "files": 436, + "samples": 436, + "raw_bytes": 328998577, + "archive_bytes": 329779200, + "sha256": "8bdbd02eed859696fb3457744ed1a4753c3fb8630a5e4063576289c3c2cb9099", + "source_signature": "38b001069e62bb524671e81edfd315c60309a6f34805ccddf3bfe35af35048a2", + "index_sha256": "9f1536c26b60ebabf257f5276d245dcc0ef8ae0252a8f5b1ae265b651b9abd46" + }, + { + "file": "shards/hfd_third_party_multi_v5/00127.tar", + "files": 413, + "samples": 413, + "raw_bytes": 308043940, + "archive_bytes": 308776960, + "sha256": "00b5da5ef62f52c5f6238164c65f43b45c5bae057ed4fdc56eadac9b1f49278a", + "source_signature": "1ee465617a4f9bf485fea3887d8c8292495860fc64c667734ae286e4e368bb09", + "index_sha256": "097e33d93255d6092ce9ef161474708da9e2410968f887165546bba34404cc18" + }, + { + "file": "shards/seed_motion/00000.tar", + "files": 9408, + "samples": 1176, + "raw_bytes": 293670429, + "archive_bytes": 303155200, + "sha256": "21cf81d19a18a6c492b9c7ee5a5dd6614f7ff0256aae19ba76f9a9e667bf3f80", + "source_signature": "caad1263137291cb2085fea45ad4520656939ec39faf06b1da13cd9023181efb", + "index_sha256": "98a956b8a8aa9b144c1b92c1d7360a1d9317683caad1f510219e40ffe6a84ce9" + }, + { + "file": "shards/seed_motion/00001.tar", + "files": 9056, + "samples": 1132, + "raw_bytes": 272919346, + "archive_bytes": 282050560, + "sha256": "5338b56a86b2b3ee40900ce03d614c049158ad610f7de47432fd6188ce0ae86f", + "source_signature": "de7d98ab024b9f451e65b2474bfe2848e82dcb2cf80d76d69129b30e3c4f35f8", + "index_sha256": "f0cf15fea7b55f66b7a8bdc09dc47f042a0cbed07cafdb3fd3516380bc676225" + }, + { + "file": "shards/seed_motion/00002.tar", + "files": 9080, + "samples": 1135, + "raw_bytes": 280781763, + "archive_bytes": 289925120, + "sha256": "9f0e540a37ff741b10ea27ad70162454da2fc4cb65008ff2f485c6259e826ca1", + "source_signature": "2e4a5d2aebb9931521cd5880c63a682af052c629626146b764515c61664e50ed", + "index_sha256": "420eefaa022c56d379f70946d92349b065b111c7d99fc6d77f03f55677fa92ef" + }, + { + "file": "shards/seed_motion/00003.tar", + "files": 9176, + "samples": 1147, + "raw_bytes": 284362822, + "archive_bytes": 293601280, + "sha256": "3747593145025aae1eb3c96e0a122f24628d7ac790a0f642357def7d59cf3d2e", + "source_signature": "c972f1f2907e4b07fcd5f10957ca19eae76d338fc69815cf32e3af37178d2542", + "index_sha256": "dfd431dea980e82d2fee81c028e6e25291ad3d842995f8649a6480d999955131" + }, + { + "file": "shards/seed_motion/00004.tar", + "files": 9024, + "samples": 1128, + "raw_bytes": 272186462, + "archive_bytes": 281292800, + "sha256": "e71c6a11bd9255d1e11d865b30d7e97a84055baf786aa0f91f4c9e72ac1ce5f6", + "source_signature": "600824f9cd7274b17e73cedf53583e28d879c992d95af68842063e9fa8ac43db", + "index_sha256": "002d41adac971e6352be8a8e2395092da61d76447ba1f3590b2cbc065b167be6" + }, + { + "file": "shards/seed_motion/00005.tar", + "files": 8288, + "samples": 1036, + "raw_bytes": 254926820, + "archive_bytes": 263249920, + "sha256": "7d5e4314cf486e077b3e83a0dc52be13a9e9704efd30cbf62f98502adeeeb3e2", + "source_signature": "cc03c591bdb04056afc250e7e71a4fab5062a861d0d69a4e446da1d34ce5b355", + "index_sha256": "dcdb8225dc68e432dbc15cc9460b5eda62173a83669fd666223a07eb5de037f0" + }, + { + "file": "shards/seed_motion/00006.tar", + "files": 9112, + "samples": 1139, + "raw_bytes": 278789582, + "archive_bytes": 287989760, + "sha256": "09aa3d009dd1a9c07772499233e6275139de29d72ecaf3911069c8df28b651c2", + "source_signature": "8b3a4a457e2cb46eac119fb9d32808a9ee8fac985e3feed926dbaa4cf416ea48", + "index_sha256": "cb5e7ea1d75c52f863960e3d856ccf794f8de55a72f2a2c756c41e7704604f2a" + }, + { + "file": "shards/seed_motion/00007.tar", + "files": 8808, + "samples": 1101, + "raw_bytes": 278081477, + "archive_bytes": 286976000, + "sha256": "f7f397b4d88d1b584b0918d86e70534206ced6d46788d6dac914a1794de5f22e", + "source_signature": "0f1ed716c7e603dabcd44e8ed8d2c051b0e4a285b645c9aa888696ea15492f28", + "index_sha256": "f705eb77f0c23f01bdfd7a03535e2ba5d69226f919551ac969432ee4ea61a31c" + }, + { + "file": "shards/seed_motion/00008.tar", + "files": 9184, + "samples": 1148, + "raw_bytes": 280175554, + "archive_bytes": 289433600, + "sha256": "a6fbb12e25f74b4070ceb9f087025712c49f547893e0a2c3ccc40e840e626a82", + "source_signature": "1095b22eb645f868e9857518f78ecd9894510893fa452d4049409ebc3b1d679b", + "index_sha256": "eb35053d62ab2fc73a0f9bfc7050a52c314bf99ad649f03d22ffd0512f94a0f5" + }, + { + "file": "shards/seed_motion/00009.tar", + "files": 8712, + "samples": 1089, + "raw_bytes": 272673137, + "archive_bytes": 281425920, + "sha256": "ceec1b6144f47a47eca74377eac2f51db1bc33524c2ba7b26c4b0794fdf7fb1b", + "source_signature": "a4a20fcab263eb1070fe6935c9f1c749b51af8e55c871d731074b0ad064c9e40", + "index_sha256": "64d9cbd87c62e1c04ce80f8cc274033aeff1e0901bce0836fc6200b92a48f538" + }, + { + "file": "shards/seed_motion/00010.tar", + "files": 8928, + "samples": 1116, + "raw_bytes": 279643114, + "archive_bytes": 288655360, + "sha256": "2b179da6ee30f38018500ff77dbdedae956d68bb66d82104303d2880ee1a2083", + "source_signature": "00fb4c3cc44b8c764d1e8ee536b90654727cda1e3fe54244522dc59810f08280", + "index_sha256": "82b060ca7721a9f226f8f5934c9221d3194daa47bc072be6b925e3de3a6dcb5d" + }, + { + "file": "shards/seed_motion/00011.tar", + "files": 9048, + "samples": 1131, + "raw_bytes": 276528533, + "archive_bytes": 285655040, + "sha256": "d247c6349aaa75b24a14878506026422c8d65a30b886e88230033908d4c0d350", + "source_signature": "4183681565eaff55dd559a5a2c2b69bf85945c07f652115acedd532cbb905078", + "index_sha256": "a587b02432238dc12f80a852150007f2adf971f3a65bc68c674d3d7ab1b4e9ad" + }, + { + "file": "shards/seed_motion/00012.tar", + "files": 9048, + "samples": 1131, + "raw_bytes": 279208119, + "archive_bytes": 288348160, + "sha256": "3a92184192bf798e061d27d168120e09ce907c2f47dbb38a2d2a7b7c17b80916", + "source_signature": "d511223a9383c6fe0d36e17784fcaa160d6912c1dcb70622cd5c07d69296c92d", + "index_sha256": "f15668cb2c3c2e95ba66185769145e6c1d0e1ff6fc197f78c1d898db261fb7f8" + }, + { + "file": "shards/seed_motion/00013.tar", + "files": 8576, + "samples": 1072, + "raw_bytes": 256130871, + "archive_bytes": 264785920, + "sha256": "24277149065d929068e2a4e4036e23f1349a98c935e53e882d12b97c3f56a11a", + "source_signature": "54e8f636c9ad9e921dc17e3fcf6c6ceeadaa56df4de2c6ce17c4994b950570b2", + "index_sha256": "42fac84dced21470ccb5ac85b9ff1778338b3ce4df92dd486885d69443d80081" + }, + { + "file": "shards/seed_motion/00014.tar", + "files": 8840, + "samples": 1105, + "raw_bytes": 257082367, + "archive_bytes": 266004480, + "sha256": "3b14e0d4a3d5b36494a57d10a6163394b73e8e86801b44d2b028143754476475", + "source_signature": "4cf588217eff4c3430ba119a8a243dc37758409d7815b7a4c8a6e198ff85a77f", + "index_sha256": "30b8d130ca4842f6c042e2a6d793e3fdfe392af14bcc98b0585d4d9635b9170b" + }, + { + "file": "shards/seed_motion/00015.tar", + "files": 8736, + "samples": 1092, + "raw_bytes": 270033971, + "archive_bytes": 278824960, + "sha256": "306ca7f9d1e2d5648cbb38a7bbe1344298eaa4c68a8bc58e24c3965352e1c26c", + "source_signature": "c38ca89c2dcde692e8b63f1272c6723eacfc7dc6874fb9050ee6a0b798db0911", + "index_sha256": "be984ac426881ebdc4bee48e843d75e7c2703975722bf1a8d686d1e912561a88" + }, + { + "file": "shards/seed_motion/00016.tar", + "files": 8776, + "samples": 1097, + "raw_bytes": 270117413, + "archive_bytes": 278958080, + "sha256": "fab5f20b1d342303f50463df765574886c5e8f611abdadde1a5ccddbbbdeae3c", + "source_signature": "037832fc4c380a0e3af1aaa869ad857bf08175c9d18c8198aaa8122db2ade65f", + "index_sha256": "67b73a4456dc73ead6633d4b48d2c288b463aa726f07a46c8514bb939383759a" + }, + { + "file": "shards/seed_motion/00017.tar", + "files": 8712, + "samples": 1089, + "raw_bytes": 283960397, + "archive_bytes": 292720640, + "sha256": "f43cd15b0c9954a6e94185281c3746c9ef13b8d14ef7cfe7eda5120d7ba55778", + "source_signature": "db5b8636431663b6aa5e09be0d1074079b4913cc3b4072252119032c5dec0ac8", + "index_sha256": "89aa3e418cb44e2fa0ec2e268233920b836dd7aa369485d2a431d19398700229" + }, + { + "file": "shards/seed_motion/00018.tar", + "files": 8712, + "samples": 1089, + "raw_bytes": 273077879, + "archive_bytes": 281886720, + "sha256": "715abd22cedc53d023f2c04a740bf965fcd4bfb0439f7a0a1eac5c25e43f8582", + "source_signature": "edf53dc77f6db51dafac3aef2e532a10974e530376190210123d3f8beded445a", + "index_sha256": "0d2c84ed6ac037f955b5cdc2c03e679238867fc6aa21ef22863099bb649e103a" + }, + { + "file": "shards/seed_motion/00019.tar", + "files": 9088, + "samples": 1136, + "raw_bytes": 271770785, + "archive_bytes": 280944640, + "sha256": "2dd11c497b5459fd71cd56517f10ca4f1555867edf60ff40c816307777e7a755", + "source_signature": "521a53d191f88ab3969fa6145fe2bf508597258d47a4f6f9e274507749901645", + "index_sha256": "9c48c696b39d994c5af3007aff90d5685e53924801c260688469be0af16f6812" + }, + { + "file": "shards/seed_motion/00020.tar", + "files": 9232, + "samples": 1154, + "raw_bytes": 270083131, + "archive_bytes": 279367680, + "sha256": "a242daceb0f6d4a68cc1082a28f2a44c5e366464c5a0d366107e49ebf0f03189", + "source_signature": "3956a485428b3bcf8eaef6b4906cf0561c3e76f499fb498e91f9ae46a5800fe9", + "index_sha256": "66facc9256afbb013624232cceb932232d543bd894bd57d5c7a0697119e377fe" + }, + { + "file": "shards/seed_motion/00021.tar", + "files": 8640, + "samples": 1080, + "raw_bytes": 260520934, + "archive_bytes": 269240320, + "sha256": "063135bbe0033f926b8ced9b574ddd28525a0365097831c13613302bbf3fb815", + "source_signature": "8ef34bef104c684233410a19a72425199edec047f7d4de2e3e2e0be0cccacfb6", + "index_sha256": "aefa0f566746070ac8852c2e794fc45b14ec964e95942deea15668a47371fd36" + }, + { + "file": "shards/seed_motion/00022.tar", + "files": 9016, + "samples": 1127, + "raw_bytes": 270403651, + "archive_bytes": 279511040, + "sha256": "e84eefaa4b6d7cc8e9440c3dda402fdc0546c0c6ddd4dd0dbfb6cc2dccc1280a", + "source_signature": "ffa8c011d1e542772e3cf709f01475f9c0521e7e7bbfba34edcfd549230b7184", + "index_sha256": "5e1c50a8c590124f2944085effc5206258aaeb866430110524af0ac958465d14" + }, + { + "file": "shards/seed_motion/00023.tar", + "files": 9072, + "samples": 1134, + "raw_bytes": 274955071, + "archive_bytes": 284108800, + "sha256": "f33b60722621eaf17db88364c9b8e0da2400018dfcc2451cc6c1f3a6bdb94a36", + "source_signature": "c098155e745838c0b3806beb91e0923174208512e23b124eb4fc06a64028eaeb", + "index_sha256": "5075cdb061dca48d14b19fbd55addeb53aaea85bb6562973e9529be26ea2fb00" + }, + { + "file": "shards/seed_motion/00024.tar", + "files": 8721, + "samples": 1090, + "raw_bytes": 281710940, + "archive_bytes": 290508800, + "sha256": "99f54ede1f37785557c575fc3db068abd8ed1ef976af19147e8654680f313a85", + "source_signature": "308a3b9f6b687014773e0d4250f7e17cd0295a7cb4f6d6ab046da3dd97f54b5c", + "index_sha256": "ad05a8bc7a9645ed7843de696d087faa5c311e0f10aeab597ce16406570f8665" + }, + { + "file": "shards/seed_motion/00025.tar", + "files": 8976, + "samples": 1122, + "raw_bytes": 272812604, + "archive_bytes": 281856000, + "sha256": "5d10d202a0908ebc8f50b2e601a05fd8c7ae55e9ba508b7cf2ff23691f530cae", + "source_signature": "380757d3d60ca57e2153f99d64700707a585a0730999a006eb5a361b87e31440", + "index_sha256": "3ec3e7340664c64238b245649c545492966db37ff99a393cb1add0c7d141a3a1" + }, + { + "file": "shards/seed_motion/00026.tar", + "files": 8728, + "samples": 1091, + "raw_bytes": 279586226, + "archive_bytes": 288399360, + "sha256": "df9a36ff3abce1683c0d39a7a7ff2eaa0aed75a60692071f3faa3f7d4d68a653", + "source_signature": "876168d131d39b921045582db61b806e9650041403acb3126908ea883a904f5c", + "index_sha256": "495d49602a51f8e0b6215d7523218b814b51b524ba5b26355371860a31e7bd12" + }, + { + "file": "shards/seed_motion/00027.tar", + "files": 9496, + "samples": 1187, + "raw_bytes": 298667454, + "archive_bytes": 308244480, + "sha256": "c9d307b00141b84d26f81d56fa67ca621f5253a1309f3076b221dc7a95395d20", + "source_signature": "5d2f8605f1a0ce57173e36b7911c1e983dd23f37dd444ebbf97c88a1b81fd741", + "index_sha256": "a4abf9b96c1bc58d2e6b7d67f57720fb4ac315668392e354a23f64b9f13e1b6b" + }, + { + "file": "shards/seed_motion/00028.tar", + "files": 9234, + "samples": 1154, + "raw_bytes": 272183333, + "archive_bytes": 281456640, + "sha256": "88771cc284eca82e4af16d3513345f0fbe85c75bc858bd7dfb6c23d0b1bb16a5", + "source_signature": "b06c98d24195c9e675e4bb70673cf9e0eff9fee5af6499313afdfc1329026a1f", + "index_sha256": "1816800160178c0f3b92ca6fd4456fb0971ed04a106fd5b29ada4ddc4a5015bd" + }, + { + "file": "shards/seed_motion/00029.tar", + "files": 8592, + "samples": 1074, + "raw_bytes": 267316261, + "archive_bytes": 275998720, + "sha256": "06114bc60070ed7190cfa49738fc0c7d1e65e8143d01251cf105920d7d63bdae", + "source_signature": "f3385855fa7a198ef5cf562e472fed08dd798ba020427abb2f9edbc50674ce96", + "index_sha256": "702e5254414abb82ad5beb48f855df6c539db5dd08b6d0f0dc7f8fa14cd74272" + }, + { + "file": "shards/seed_motion/00030.tar", + "files": 8704, + "samples": 1088, + "raw_bytes": 260274058, + "archive_bytes": 269056000, + "sha256": "be64e5cafb3e0dab83ee66f4efaa0d835378650ae20c149d50cfb218321cd9f3", + "source_signature": "d42b1104993e2736e2e3f5c342a087f9dc236062bc5dd0664fa519c2c165e880", + "index_sha256": "f057dac40f7fe4d63ad8c1246a081076265ae841fd78ac0f1a70a4840ca0a625" + }, + { + "file": "shards/seed_motion/00031.tar", + "files": 9080, + "samples": 1135, + "raw_bytes": 281003811, + "archive_bytes": 290160640, + "sha256": "da6a6564e9bec49a371653a6b6954646caf29e9afb8f986a33976131c62a318a", + "source_signature": "bdd4206092d1da59cd6c311efa8ae6ae6ee27259429f998c4455abec82a87792", + "index_sha256": "039d05659a42a0b3e6877797beadd69534b01935078fe636b54ecf81b60ebf7f" + }, + { + "file": "shards/seed_motion/00032.tar", + "files": 8504, + "samples": 1063, + "raw_bytes": 287556838, + "archive_bytes": 296110080, + "sha256": "b1edae9a0e2a4c79c3aa5e6b8e69d704dd6a2e18f281676803c5395267c4fb5e", + "source_signature": "9b41a13ce1dd9abd8d5f51b6ae63c4ac8dac78cf22ac625cd5105cb6248fab80", + "index_sha256": "8cfd684f54375afd3e8a86d644117be3d3aab3f22000d9a9a3174c3ae905705c" + }, + { + "file": "shards/seed_motion/00033.tar", + "files": 8824, + "samples": 1103, + "raw_bytes": 271632346, + "archive_bytes": 280504320, + "sha256": "be88e19574c38cab69dc8e5bda988b83654bd4a0761f9bca4ee0950e9a97b527", + "source_signature": "0f84ab353daf1444d4250c0fae43b603f16605723dc816c2cbdd966e6acfae4e", + "index_sha256": "a1ff5aeb72f518b5f52ff43f8bdf3f7ef6a95ddebfc44044bfcd1ef3ea13bcd7" + }, + { + "file": "shards/seed_motion/00034.tar", + "files": 9384, + "samples": 1173, + "raw_bytes": 299641244, + "archive_bytes": 309084160, + "sha256": "3f9c354420959e58736ad1d7e720993936d2cff7ee3a52b9e4f868d73a094b2e", + "source_signature": "1a5e305acb2314826551a31651dd77e224d3e4210c27398d2fd01235cf2286a5", + "index_sha256": "47a6f0866248c4d47a979f35fefa1eb707349a7d50305d0243199c1eb919ab7a" + }, + { + "file": "shards/seed_motion/00035.tar", + "files": 8736, + "samples": 1092, + "raw_bytes": 272286347, + "archive_bytes": 281118720, + "sha256": "7ba66f29f78494a4b6b78525fbdfe6ed6fdea05cd8c77d97a2fd557209a6746f", + "source_signature": "1903d0cc7cb16d57bbc6b602a3ca833eb012ae930eb13bdc3a00d586701f55cd", + "index_sha256": "185cd7f4ddeb0356ce146cb1e053a8568f9908216cce5114d4ccc7cb441b8883" + }, + { + "file": "shards/seed_motion/00036.tar", + "files": 8880, + "samples": 1110, + "raw_bytes": 272445567, + "archive_bytes": 281374720, + "sha256": "145852f101081648c524136ac8a72d96996000ed3804b79d753e1bfd0745b6e5", + "source_signature": "b4876f864111d5c174a7a2ef66b2f686f6506198cf41bb80183092f9d2dcedc6", + "index_sha256": "dae7ce283f283ae16fc3fc79c54b6a68679836b7c058775248aba80a8c0eeed4" + }, + { + "file": "shards/seed_motion/00037.tar", + "files": 8816, + "samples": 1102, + "raw_bytes": 284203041, + "archive_bytes": 293099520, + "sha256": "68bb4966d8385a156f16d3c20ef92cd45e84cafd405d5f7b549716cb6ed495e4", + "source_signature": "e65faf4484259b7dcd8cba811f1bfd1755f6ad2e4abcd4140f44df20a30ec5bf", + "index_sha256": "f59a8d6273eeafee7d00084b95cbdd60ac966c8bd9ce91200fb0704545076965" + }, + { + "file": "shards/seed_motion/00038.tar", + "files": 8840, + "samples": 1105, + "raw_bytes": 270323513, + "archive_bytes": 279244800, + "sha256": "53f19ecddd03c8f5042c67f1ad0be76d5c5d7d7aa70b9dc489e698c9b2e27b44", + "source_signature": "f7e7698f8334569fad10d4b71fd576b371c88205adbb2306e9162912d997e236", + "index_sha256": "834700dd8ffa641fefe82ba370a118d52aa00b23168101175fbf724f24de05cc" + }, + { + "file": "shards/seed_motion/00039.tar", + "files": 9232, + "samples": 1154, + "raw_bytes": 289816950, + "archive_bytes": 299151360, + "sha256": "ab84d94fccd047b6354650b3d05e5eddc650219e5c1a8c9495d86a1f13aed381", + "source_signature": "ab4baabac56e5bacb63b5b4bdc4223826d548e43ff75d8cd7e91e8bbe92965b5", + "index_sha256": "e717bffc8ce0ecf122649f5e545ced03f460f31c832b0d1df65b71270156efbb" + }, + { + "file": "shards/seed_motion/00040.tar", + "files": 8912, + "samples": 1114, + "raw_bytes": 280768183, + "archive_bytes": 289730560, + "sha256": "ab8c0448548c96fada47ffe292b2be144d600657c1f3b8cef8b63e0d41322a3d", + "source_signature": "db52319b63c2894a4da2c67e86b05f7188c6047f3d017820458c8c25b7578a63", + "index_sha256": "ad36a71cc0289fc7e382617f601e45b2e96b990e76917c4f6251d25435381a27" + }, + { + "file": "shards/seed_motion/00041.tar", + "files": 8832, + "samples": 1104, + "raw_bytes": 268846373, + "archive_bytes": 277739520, + "sha256": "f66eb6e2ae4c4a6cc6216260770ad5cb0149e5cda302d0ab669daef63775b9cb", + "source_signature": "5daba0f4b0299d591d3054379e43589636dfb113414fb05ebbdab9527688791f", + "index_sha256": "72a94284d583f1054d1976a224d6aecd68559aa2f70e6d1d0362f8184fb2f366" + }, + { + "file": "shards/seed_motion/00042.tar", + "files": 8648, + "samples": 1081, + "raw_bytes": 278848031, + "archive_bytes": 287559680, + "sha256": "9cea05baffb1355ea46b4ac7d7266d1f93aede04a03e3404bd5e4bad79ba50da", + "source_signature": "213df505a62e3a7a3dfc99f341abfcc96da3a44eb94f4bdcf5c497e992108860", + "index_sha256": "e948cae72c37dca0bec74c6e155a594c6a1356be53be0f6a2ef87842754084f9" + }, + { + "file": "shards/seed_motion/00043.tar", + "files": 9104, + "samples": 1138, + "raw_bytes": 292182723, + "archive_bytes": 301363200, + "sha256": "1d6db1ef9b5cfcb33cc8d1be69a6f682a47fdf32eb30b016172efd807ed872a1", + "source_signature": "4842868deef2cd804210c5b5aea29f8edf706d0aa3c9d8ead4253e0300e2ed71", + "index_sha256": "87ed8b304fc805aba2860f2ceb86e082939751ffc41e5a267767f5f3946bbe7e" + }, + { + "file": "shards/seed_motion/00044.tar", + "files": 8888, + "samples": 1111, + "raw_bytes": 274927709, + "archive_bytes": 283883520, + "sha256": "5a97edae9a88532f813d4fc2e8773dc617f8d975a13d621f68c2e6521c35cda8", + "source_signature": "5f65a661f06c77f6b61b579ed4af44c67e673d8c6c29f792fcb276efb8deb36a", + "index_sha256": "0309437376cd5551003aa87ba265c1ad7ab6040f5d18409a149bd03d89a4cee9" + }, + { + "file": "shards/seed_motion/00045.tar", + "files": 8544, + "samples": 1068, + "raw_bytes": 257268507, + "archive_bytes": 265871360, + "sha256": "a7cd4e8ff1f888bd04d328df206328214bc58281704d55ac0f8f532cf7fca51f", + "source_signature": "8170f92602520cec254268f52d824e0973744e3064d955f97d52b640d5d86143", + "index_sha256": "b7679902954b6a9db175072392a02cb6ca8bd246243afaa335f37dcfe5172ba2" + }, + { + "file": "shards/seed_motion/00046.tar", + "files": 8984, + "samples": 1123, + "raw_bytes": 269977328, + "archive_bytes": 279040000, + "sha256": "fd878c0fd0ccb0ebc48f35852d427e9fe1774260f8b185b8086793b5ddc9f7c0", + "source_signature": "9e1b39a778b67d90b6f062fc4135d297ce575971af87e0fe3c32b1037ddba840", + "index_sha256": "65265f2abd1c9d90b76e6b412848da74325de7b3451cc4741571bea2fe639a28" + }, + { + "file": "shards/seed_motion/00047.tar", + "files": 9432, + "samples": 1179, + "raw_bytes": 280545720, + "archive_bytes": 290037760, + "sha256": "3380b93ad872e9b00f159051ab722af7bc5d976cd054ec6e3966fc04427495ed", + "source_signature": "790561b691cb5a86ae4a496c8d543b7b224433f78fd58c9fc45ffb960e68f896", + "index_sha256": "8fdbde5047c376b7bb824d4936736b30ed29f5af94a4b81bee2e1a25e5433875" + }, + { + "file": "shards/seed_motion/00048.tar", + "files": 8872, + "samples": 1109, + "raw_bytes": 275976805, + "archive_bytes": 284938240, + "sha256": "58741f5c277ebd6d79e8f33e66f8afb8a4347e917963448208340ee607306f8d", + "source_signature": "c6efc0a4b779205546f075bf8ae6afcd88082cb2512c37b6ac951f25631503b3", + "index_sha256": "81661e3c19d80655bf9ae88d54346e1eb2f2b5c42b4bff883c5a60c869d2af84" + }, + { + "file": "shards/seed_motion/00049.tar", + "files": 8632, + "samples": 1079, + "raw_bytes": 257205013, + "archive_bytes": 265912320, + "sha256": "81881903048e4591e5aa3f5faf9c45a56fa43d5e4f4f494077e22ed4f1016ded", + "source_signature": "98345fd381cb5a0bb0d38f38dd1100072b421085a710d847950b6c1f621adf3a", + "index_sha256": "bb6c32636efccbaca25165c5d0f0797e65678dccd30237cdb2dea18c240ad025" + }, + { + "file": "shards/seed_motion/00050.tar", + "files": 8936, + "samples": 1117, + "raw_bytes": 265757877, + "archive_bytes": 274780160, + "sha256": "482a2441cca75764a7aa3b3d4f6053dfc94efe4a0ca36c67df461aa19a7ce320", + "source_signature": "be78b54e0f3bce67eb25d896de20b35d1a9168e6d20ff01ee26b5f55c08d78ff", + "index_sha256": "21f71e3c373f7e58c5951f557a001edfa15d2cb57234d19702c633b0567e267c" + }, + { + "file": "shards/seed_motion/00051.tar", + "files": 8312, + "samples": 1039, + "raw_bytes": 261000647, + "archive_bytes": 269373440, + "sha256": "f4977cb655e30b396d0c8516958db7a23c31ea9da48e24e7e3c74ea097583397", + "source_signature": "e78235e9f43aabb4a7b5c92e709701a2ce1d335c48cce1fff8d3b5046915cd56", + "index_sha256": "32e2e38ae621740af7ffb4c5c081441234b208d67eee23c1e16c46f9d61fd483" + }, + { + "file": "shards/seed_motion/00052.tar", + "files": 8704, + "samples": 1088, + "raw_bytes": 269865740, + "archive_bytes": 278661120, + "sha256": "aa61e749a048f4a2f83daf577aa8e5097b1a7300bdc2a83aacba871c48b93968", + "source_signature": "341bcddd73c96a6b25e8077933b1b811a9f020aecd87069a3ffe1468e9793deb", + "index_sha256": "b74468ae04c45336591b86bcb6ebc865cdcd42db111a329ebaad48f0f6255131" + }, + { + "file": "shards/seed_motion/00053.tar", + "files": 9112, + "samples": 1139, + "raw_bytes": 283251460, + "archive_bytes": 292423680, + "sha256": "da7f67fa49b28faf9efaf0882765cbe094cc08347e8f9c3bea20ae16824574dd", + "source_signature": "086fda06896d8973a9ce261c1c1a59d1f0220b0d47a0239f7a556d3324c3f2db", + "index_sha256": "05250b75836aa97d9d7c89bf31d26693ac99229c2d19f1c94aac05d8f136040c" + }, + { + "file": "shards/seed_motion/00054.tar", + "files": 8992, + "samples": 1124, + "raw_bytes": 276779448, + "archive_bytes": 285870080, + "sha256": "61c082536825c0ce7f44a4ae1670d9be233794cca9e8aa43a60c3f371a6aa22c", + "source_signature": "9bbe67492bbde42f6171715e19f181ce8e9125991742c0bd519f1fe93d9cd952", + "index_sha256": "eda3ffcbf17afe5689fdef745903acf78011c63c065d913a5ad541295403fd50" + }, + { + "file": "shards/seed_motion/00055.tar", + "files": 9257, + "samples": 1157, + "raw_bytes": 290429832, + "archive_bytes": 299755520, + "sha256": "d92577fcc1803f49f7d8fcd6a83d7a0b517659b2ab50cb80bd6e32f403148579", + "source_signature": "c27e4e2143953cfc9612edeff68e31c02d1de4d79e2709e451200ecfedec0d94", + "index_sha256": "5bed63aac259aa265b685c2f430de2515f3662922752f8f1726a4d07d4c06382" + }, + { + "file": "shards/seed_motion/00056.tar", + "files": 9192, + "samples": 1149, + "raw_bytes": 286286729, + "archive_bytes": 295546880, + "sha256": "97c5efa2e49927cede60611c78a9fdeb9c57dab769c1417f9786d4123d2d2e02", + "source_signature": "55ca4a6dc4b0ef20f2636566cafe6dcc22b4aa173d7ba6dfee53ee421ceae730", + "index_sha256": "f21b65384c48198822f382d319e7bb0a72253cfe2a3bb7fd697c7a25c249d136" + }, + { + "file": "shards/seed_motion/00057.tar", + "files": 8736, + "samples": 1092, + "raw_bytes": 275824004, + "archive_bytes": 284610560, + "sha256": "f5167ae8b4c79aa65c2317212ead0932fbc1d085a54b8e48e2a1fe0b1907db35", + "source_signature": "3cbbf97ff9dd4dada1be0789537e30445e5c0c5cb90ad02851f7a291ed62b3aa", + "index_sha256": "b5dba453ddd634ef7f91847346b5afddd3d1224799bc2a3750866d285328cba7" + }, + { + "file": "shards/seed_motion/00058.tar", + "files": 9288, + "samples": 1161, + "raw_bytes": 289312217, + "archive_bytes": 298711040, + "sha256": "b839affc1abfb61815417badae9f801f7bbb69678f0ebbe769ab8c75f88ae91a", + "source_signature": "18bfb5d1b369ac551ee2fdad2a1611c17ddf9b44d8c376cc7af5c470b238b88b", + "index_sha256": "7550feb283ed375521a0512045d14aef9041b9b5d1db08cb0a8a304385c3f028" + }, + { + "file": "shards/seed_motion/00059.tar", + "files": 8984, + "samples": 1123, + "raw_bytes": 270077871, + "archive_bytes": 279152640, + "sha256": "7387a08d744e0d8be6cb6e1d01786e5c8b5a3f79e282c37765655b857de1e33c", + "source_signature": "f54bea0e05a500bf8b1845a2f6878c1ed86afbf9ea9185fd1d99ea92957e09b0", + "index_sha256": "420a08b993a4e39d5ebcdf91dcd824b90cf9fbbdfec72db6c862b4e5b0e0f7b5" + }, + { + "file": "shards/seed_motion/00060.tar", + "files": 9040, + "samples": 1130, + "raw_bytes": 266267537, + "archive_bytes": 275394560, + "sha256": "31307130775d7adbe3bcec12260b59f69bf6a43bc3fe8c8c671a3f95833747d2", + "source_signature": "5664a08943e3e0ca72e560b3e55dfd2e06f12db95ee6bb3fda58c9d12c7e9e85", + "index_sha256": "03fbbb4865256c74133534d013e6b0c39899a7c4b946637f09b0941398d5d615" + }, + { + "file": "shards/seed_motion/00061.tar", + "files": 8984, + "samples": 1123, + "raw_bytes": 288202398, + "archive_bytes": 297246720, + "sha256": "4540c3b9e4dc7bb59cdab4e3ddf5ade0bca5b889b6d590cb1fb63d0d06e37a78", + "source_signature": "30a8daaecbeba6a656de97167714c34b274906b6e43745962ce02b6e3e462fb1", + "index_sha256": "3b33334cc02b7407c739fd7f3d0a569238783d7acc978b4fc2bd661ea480defb" + }, + { + "file": "shards/seed_motion/00062.tar", + "files": 8592, + "samples": 1074, + "raw_bytes": 280941250, + "archive_bytes": 289597440, + "sha256": "df296fbff59b26ebe47ffa2b61a62f9ba90465adee1ad8988c9f99d7b9db6792", + "source_signature": "8cb528c1f1619efa21817cd1429a780250551b9e06233fb6c3c339322bb979c4", + "index_sha256": "40def65ed482fb59c0b7491fdc53ac1826b13d82352d53995bdc034d24049851" + }, + { + "file": "shards/seed_motion/00063.tar", + "files": 9120, + "samples": 1140, + "raw_bytes": 276335532, + "archive_bytes": 285521920, + "sha256": "0f1ae643872ecdccf412c490974868d07ad9d95b02874694a4686b87f7b902ed", + "source_signature": "3421895133a1b1267a75864bb3c7eda3f3642cf4338eec52be9caaae9f6d22c4", + "index_sha256": "aa41275abdd3663a4090f3aab647ea2161294e928994cc98c034042247ec20f7" + }, + { + "file": "shards/seed_motion/00064.tar", + "files": 8864, + "samples": 1108, + "raw_bytes": 269953879, + "archive_bytes": 278896640, + "sha256": "8e41f4719ee115be17eeb7b02554014bfc084012a5d24fa75d563d2a065b4598", + "source_signature": "d7ec893c4501febda9e775345ac3d1461dadb7fa5771d6a9c3d7a6f53f5188ca", + "index_sha256": "fb0a13112ef5686500d1ac7606eb59034c3bd083812616543cb60f84d6523b01" + }, + { + "file": "shards/seed_motion/00065.tar", + "files": 9544, + "samples": 1193, + "raw_bytes": 305444080, + "archive_bytes": 315054080, + "sha256": "c08987a76ca1e6b3de56465fcfa73d7a8cd20070dc46cb56b27aad51df7c1d2d", + "source_signature": "c6ebb02273927f61dd4fd7de67d562651dd0bdae9201eda3b6488cd5531a0575", + "index_sha256": "a81651c068a32d8e70a2976a787499f40ff644e66a0d3ef51798e54a0329c8d8" + }, + { + "file": "shards/seed_motion/00066.tar", + "files": 9104, + "samples": 1138, + "raw_bytes": 277558409, + "archive_bytes": 286760960, + "sha256": "d530920fafa72e8a607252efa844ae1881f1fe41be6a1d8014476edafb01f68d", + "source_signature": "5487a95c29978aa4a8e989f5734e9aea329fc656c1f7ef14cc35f4bccb831792", + "index_sha256": "1469ad7005385cf4319969c7e5a1217af1ef00ecb47ecb2bf1c050c3c76cc605" + }, + { + "file": "shards/seed_motion/00067.tar", + "files": 9728, + "samples": 1216, + "raw_bytes": 292719602, + "archive_bytes": 302520320, + "sha256": "bd0a4f6b8a6c7b83cebd9c6283478e7e80b210ea1ec5ad1a2fcd74ca051727f8", + "source_signature": "3d9e0b9a9974ed6aaf3d3bba96a1af707df3baf88a4775f5204d3c696e3a1fae", + "index_sha256": "c683ae171db52c24228cc591ca87bd16948943dbf2d21bfa2f368273a4f718b2" + }, + { + "file": "shards/seed_motion/00068.tar", + "files": 8544, + "samples": 1068, + "raw_bytes": 256933587, + "archive_bytes": 265533440, + "sha256": "503dd2afa88eedaf9d18022b4bc01aa42cfd0a46e2df87034cd4c74fa3f7fa1b", + "source_signature": "93f4608fe343cb30fc6035dda0fb06f82f2a2385cf6b1953d6f3fde1d97af145", + "index_sha256": "47d09566114c6a36573a63d3f2d4c44e5844a5f073aaf08525318d5d2a80325b" + }, + { + "file": "shards/seed_motion/00069.tar", + "files": 8920, + "samples": 1115, + "raw_bytes": 283812048, + "archive_bytes": 292792320, + "sha256": "890980c5d504999f2c8dc49c199d211b693385f89c48cbbc74b455285e2b63f6", + "source_signature": "8379d3059c87f287f9854b561b9426eb02b3e76a6e8660c0e651b58b33f6ae8a", + "index_sha256": "f70fc99d5887473840191726f2b8886f03fb0eff6c95537fcb8dd1dc3abfd9c9" + }, + { + "file": "shards/seed_motion/00070.tar", + "files": 8496, + "samples": 1062, + "raw_bytes": 259094436, + "archive_bytes": 267673600, + "sha256": "1c4d78b7e3a82fbe6b76d64653a7b3adb2f51ec9f9b3766bbfc9598ff256a7c6", + "source_signature": "14305fc10606fbe7399cd4eaf7b519ccb53b430ff96b6190a48a855044c4f375", + "index_sha256": "eb68086cd34f4c80cb9586fe34543ccf4bea5fa38a56141db8bb68100ba36a33" + }, + { + "file": "shards/seed_motion/00071.tar", + "files": 8944, + "samples": 1118, + "raw_bytes": 275797929, + "archive_bytes": 284825600, + "sha256": "3417d9dd1f46b6527cfc521a6e5c9fab659cd43a73b9ca6467fc1d7795e869f2", + "source_signature": "1a90f902b867a5dcd727a58ccbab73fa0aec55640d0b0d0f5bc04fc61dfa04a6", + "index_sha256": "290249570c9b7b53d5ceefd80fecfc1213a55f781ef2365ce8fe6051a3825571" + }, + { + "file": "shards/seed_motion/00072.tar", + "files": 8833, + "samples": 1105, + "raw_bytes": 417682195, + "archive_bytes": 426577920, + "sha256": "2a35377335b065b34692ab3c6e18e338eec04e0bc50bf5322cdc1349d9089d2d", + "source_signature": "7959ca632154c5489e43f7518910674e97953733341cce1f4e88de0f5e3befea", + "index_sha256": "9e654649865c89e94f64a28a252ab0f6f76c2f4e395f8efe8ed88f04fea00cc9" + }, + { + "file": "shards/seed_motion/00073.tar", + "files": 8592, + "samples": 1074, + "raw_bytes": 259277525, + "archive_bytes": 267919360, + "sha256": "8df6d302a6a57c181e50e14893825b7007a77fa2beede30e3a0c0f06331a5b79", + "source_signature": "f62bb8fc851dd7699d73bd7861145d53659c702efab248c88f66c7632b00b38c", + "index_sha256": "41448d347e94bddae260a88a51b9824bee4ca7081345df8fcbc7ee077c2bcd0e" + }, + { + "file": "shards/seed_motion/00074.tar", + "files": 8896, + "samples": 1112, + "raw_bytes": 271954518, + "archive_bytes": 280944640, + "sha256": "747a989cc3ebc7cf4bf5e18552740ab8d0eb557a3ab2facdac1103f0dc83c74f", + "source_signature": "60afdb2a2de637f007d0606fefbcd5b9b187574101bcbf03dc8352ad5d199a64", + "index_sha256": "4212229579996b9511c522c4ec7f4ab4558dd784d2fac4ba26becda5cb60eef2" + }, + { + "file": "shards/seed_motion/00075.tar", + "files": 8840, + "samples": 1105, + "raw_bytes": 272992099, + "archive_bytes": 281896960, + "sha256": "3fadf1856b9d2d727eeb92de45b00d07aea6620fce6813e58a5eec72e27da682", + "source_signature": "74a7897e76f4434016c999fd2c58bc0e612d2d3195d26fb576945cefa8515fb6", + "index_sha256": "9678c0c17f9485419e3b5b2591fc65a076df2109a6da8869caec9ea2b7a367c1" + }, + { + "file": "shards/seed_motion/00076.tar", + "files": 8856, + "samples": 1107, + "raw_bytes": 265029040, + "archive_bytes": 273960960, + "sha256": "8456893d6256276bc84787008e5abc91b141b98c366b62cc38c9c73f3f8a93e4", + "source_signature": "4193600375c47da4a380536640c783550f4fcb480abf0ef76b1e501aaa9dacca", + "index_sha256": "f12e04a795861e0fb30fa2d60c92644140f5e87d9f09cc56290e567c568d6309" + }, + { + "file": "shards/seed_motion/00077.tar", + "files": 9016, + "samples": 1127, + "raw_bytes": 269455513, + "archive_bytes": 278548480, + "sha256": "47b6184143faf42db503780035d4b4c0d742a5f74793cebe2e8de4ed3a69710d", + "source_signature": "dde661299151586583fa1ddb95481e225af189e292f6d70c325aab2635ab08c5", + "index_sha256": "819f35394d8d2560ea67a1d7c735736109354c07ca71bee397531cf8a709bc97" + }, + { + "file": "shards/seed_motion/00078.tar", + "files": 9048, + "samples": 1131, + "raw_bytes": 276972467, + "archive_bytes": 286115840, + "sha256": "e4cbf15f753b52e51cf0835f3ea75941cface5008dd1bc25f52aff55b000ecc0", + "source_signature": "f2aed3e5493abe99a4d9ed0eafe5232843ae0edcde3630af0649ac37b9b63d67", + "index_sha256": "bd16f7a2f670c58066ae6da73602b9ea610ee096f5e2caa4d5cd4d34aba63a46" + }, + { + "file": "shards/seed_motion/00079.tar", + "files": 8960, + "samples": 1120, + "raw_bytes": 288596348, + "archive_bytes": 297646080, + "sha256": "552e819b96e3e7822beba76d1c8040b94d94f13e46048b175c04c99c1a14734f", + "source_signature": "5e790fb3086c4bf02b7162ef17f1060285851c3086b790407a853341b0114cc3", + "index_sha256": "3c7d9af1d8b2cc0d2851e1c797eaf9d146f25a66161d9a3654b273680814a3b2" + }, + { + "file": "shards/seed_motion/00080.tar", + "files": 8536, + "samples": 1067, + "raw_bytes": 275683417, + "archive_bytes": 284293120, + "sha256": "214faedf4b15dc2194df4695d367224a6848527ea52a81756cb7378f20518000", + "source_signature": "86048eb7e6bca0aaa44d8f016301d7757412c7ab3f2ea8fe4ae249bec4204a2c", + "index_sha256": "0eef9973e26536653586c1530ad813b5ee1574c2d4f175d9f4d238dda0fafe24" + }, + { + "file": "shards/seed_motion/00081.tar", + "files": 9008, + "samples": 1126, + "raw_bytes": 276830061, + "archive_bytes": 285921280, + "sha256": "9dee1932d4e7b6c5794ae2308411f88e9e2bc1537ecf6c7dd6168ee4cb1a3677", + "source_signature": "2fcb9b278d1c36b6d51a37ddf98e4c057dfca8172a73ced1f4145a1667c2199d", + "index_sha256": "a30d51a1b88fa6131e5874bafa9d13043a532dffb4c4ac29d00cf4a67fe3d047" + }, + { + "file": "shards/seed_motion/00082.tar", + "files": 8929, + "samples": 1116, + "raw_bytes": 275196457, + "archive_bytes": 284221440, + "sha256": "759003e820bf273d2a9ebbece38dfab02ba1f5ca5bf5454b63a651d36a273548", + "source_signature": "9128798cbcf9327e1ec4e6a3e9b19e1838cc60eb94ab9ae625381fbfb7bf0b96", + "index_sha256": "d3c245a357ab07f71c1e09ce0e1461edfd374a5dd44ffd357e3ca36b0b95ecda" + }, + { + "file": "shards/seed_motion/00083.tar", + "files": 8920, + "samples": 1115, + "raw_bytes": 258801548, + "archive_bytes": 267776000, + "sha256": "bcfb704c377aa9459c4b38b1cf1746c1b78a12ce43cb2572f87cacb7563ab8c1", + "source_signature": "88f23c390e260a375cb5525d7ee29fc70da27eb71f58009ae0fc4e921717bc8e", + "index_sha256": "4ad1989a0b0430e07e521e65d63ff6d4b46fd43d6b6e8f45802a31a6ddd5c4f8" + }, + { + "file": "shards/seed_motion/00084.tar", + "files": 8296, + "samples": 1037, + "raw_bytes": 246851436, + "archive_bytes": 255211520, + "sha256": "3a1f9d376dd1009d764c8cbb6fe452310fe5d9372c3281726b1c54a154e236bb", + "source_signature": "fc3fedb8879b3595c4cab6bfd2630d1a017424ed74a6bb71a2e638b07b227553", + "index_sha256": "503d5cb4193af679fdd3b8abaf6aac46ff7c4e4b9aecc2e2749faca9d84d8766" + }, + { + "file": "shards/seed_motion/00085.tar", + "files": 8872, + "samples": 1109, + "raw_bytes": 275893349, + "archive_bytes": 284856320, + "sha256": "4e08faf876c4db20c598581815515e795da8dcbf89510a8bac7b735450209bde", + "source_signature": "1e44fec7faaf3e4818b3867c7710fee26444a7c9b56e46d50e4cb3866320bd7b", + "index_sha256": "9ab0c0a4a01118085350c050d41edb044df8cc9b3ccc4af20d983b241587de1c" + }, + { + "file": "shards/seed_motion/00086.tar", + "files": 9096, + "samples": 1137, + "raw_bytes": 267152544, + "archive_bytes": 276316160, + "sha256": "18055ff38d394c94b60eded15ebcff6ceef46bce30f9109d05087a55e6a1f4ef", + "source_signature": "5e32f3952ac65198f9decc9b90edf04dcf6a7179ff30a3e02bce55a6c01f5581", + "index_sha256": "2a448e8cf8389490c3a4df13d43277a4ae6d68319242e6eef9c0f50d572db570" + }, + { + "file": "shards/seed_motion/00087.tar", + "files": 9208, + "samples": 1151, + "raw_bytes": 283112991, + "archive_bytes": 292392960, + "sha256": "061208830ce73da2cc69cb1877a1bc791a10e02faaee122ebcfb09e6fcd0c57a", + "source_signature": "280e86f744630028a9486fb764772f09ddc96f921a2d6d0d9d353d8cd94c6082", + "index_sha256": "3f3d8c913cd1e53584601b67ec5aa08c71c04fde280e33fc2db426e9bd1db483" + }, + { + "file": "shards/seed_motion/00088.tar", + "files": 8664, + "samples": 1083, + "raw_bytes": 257706487, + "archive_bytes": 266444800, + "sha256": "2eccc99b93d8874e27df46ce4d174778bbcef5bb96774cd3edc833ef02d42df2", + "source_signature": "e079716943483977701622e0c47fab48cfa9e3d9d994e9a9e8bd90b74f4e5d0a", + "index_sha256": "94cf5b183c6f5bd6b9518c33fc835f4c07bf4b49addf75ae27eed81bdb524f97" + }, + { + "file": "shards/seed_motion/00089.tar", + "files": 8784, + "samples": 1098, + "raw_bytes": 268672117, + "archive_bytes": 277524480, + "sha256": "bdc3c563069611df03a246221174ab70404a9d755b5233985bed9c0a5bd70bc4", + "source_signature": "3780ecd409fac9d8805aae86d22a713b5f42a69bef489e57a0c7188adabe9ab2", + "index_sha256": "a51a0426f31daa91324a693b4848dabc659f31b623bcce4cd3f4f565513b2185" + }, + { + "file": "shards/seed_motion/00090.tar", + "files": 8472, + "samples": 1059, + "raw_bytes": 268534613, + "archive_bytes": 277053440, + "sha256": "32641649a60488324da3746a36af41ae7abb9fb4fff84caaaee3f307e8f5ac3d", + "source_signature": "f9441516b79f7f4b94e88a29b02e0355e1a4637f0d4b3be8b45001298878f320", + "index_sha256": "f8874dd56d6262a526094e3b39fe1f66b68af8e7dfdb7c81d1590f37f5b07e49" + }, + { + "file": "shards/seed_motion/00091.tar", + "files": 8608, + "samples": 1076, + "raw_bytes": 275246209, + "archive_bytes": 283934720, + "sha256": "25db63f66f0cde752eb5884327186e3123c9f34716f15857047adaa79f61485c", + "source_signature": "16af8e79cc541c7c9649458de45cf5d1304497884b1e466454897345a1b0c0e7", + "index_sha256": "67570ec308acf423968d6d9b52e437daa4525b09c0764db1e4b6493cca8e95c1" + }, + { + "file": "shards/seed_motion/00092.tar", + "files": 8808, + "samples": 1101, + "raw_bytes": 280631028, + "archive_bytes": 289515520, + "sha256": "d678fb00ae2098c83480c14aeaab525cecf21bd93b75a5beacf687566a28bb7a", + "source_signature": "5eff823eea46e41bfb92fac96be95aa712e2cd332729d6f8ccc8bb6dff162523", + "index_sha256": "4cc6f840ed4f2b87c5739f1c8b56f133ef6c192e32702539b92f860377965984" + }, + { + "file": "shards/seed_motion/00093.tar", + "files": 8728, + "samples": 1091, + "raw_bytes": 256053988, + "archive_bytes": 264857600, + "sha256": "0073dee08e866ee6f4318a8ca17cbdc8111e11f8650984e31f2cfdfa09bf663e", + "source_signature": "4b7d71309d47528d7362c27da739923bfe94a2a164d5a30579af8587a038dc28", + "index_sha256": "f5aecb3efa1905fad63418bdf7b16b0fc104457cdce9c0e1f67a160f5403c259" + }, + { + "file": "shards/seed_motion/00094.tar", + "files": 8608, + "samples": 1076, + "raw_bytes": 267803171, + "archive_bytes": 276490240, + "sha256": "21129b891d9b7d82b931baadbdcf406ed3ef820084e921af75e8154e06fd9971", + "source_signature": "0f1038a6d63ad6e6fed4cf254bfda1ef4dde04328641d1b5d86d72c69d2b9d31", + "index_sha256": "286d8bf43cb82eecc99d00bc4fd72eff73d1773975ea80499e5f7c0fcff84505" + }, + { + "file": "shards/seed_motion/00095.tar", + "files": 9225, + "samples": 1153, + "raw_bytes": 282181152, + "archive_bytes": 291491840, + "sha256": "4feaf812d1e1d06af170252a3a012082e048e9cd5b849ae8d223a73d6b06acb9", + "source_signature": "ca59236c991eedb48abce481a5e0637a0786bcb6e562e91533e6ebb4ecbc558d", + "index_sha256": "add90e50d6db8a44ae798a42e3c98fc692ede615fa4983d940bf4e3f9aad0578" + }, + { + "file": "shards/seed_motion/00096.tar", + "files": 9273, + "samples": 1159, + "raw_bytes": 280845658, + "archive_bytes": 290201600, + "sha256": "07ca36c2ce6a4a24a82d176c3160ada90e33978f395f869cadc1af399cf94f9f", + "source_signature": "7480a62fb4a96691e50973004be53ec58b93d338232b8605bcf97023056624ae", + "index_sha256": "379bda768d81197175d40dd2975132b326f5eed38b1ee9ca86145e8e888069f9" + }, + { + "file": "shards/seed_motion/00097.tar", + "files": 8600, + "samples": 1075, + "raw_bytes": 272573044, + "archive_bytes": 281241600, + "sha256": "89b04682a3790cd53cf2775eeaf972a08ee3e145c263c30e4fa5232af58666d9", + "source_signature": "3112a862f62a8f81036047d3e0dcb14f4d045ee6569a6dd575ecd5ae0ad90218", + "index_sha256": "2c12247e5c745816f6a3b0099f04f7dbc79cf11ae7c312c91101b4b5682d1697" + }, + { + "file": "shards/seed_motion/00098.tar", + "files": 8904, + "samples": 1113, + "raw_bytes": 271798682, + "archive_bytes": 280760320, + "sha256": "0c41cd207a51ca9880aa940f7f12a688b05c5566978b9994f88f9b39faf6ad84", + "source_signature": "50dd5f53d3ce1f7cd0b6504294b4b65cb601963729428122abdf9c40a17cbfe6", + "index_sha256": "cdb6fa17b30228505078c931201254b4482fccf33844f7dc34e6bc613d009b74" + }, + { + "file": "shards/seed_motion/00099.tar", + "files": 9024, + "samples": 1128, + "raw_bytes": 279853406, + "archive_bytes": 288962560, + "sha256": "f99448d2ba325b9c32fca8b6382d1851c0b0d75afcf242eeed4c3adcc2841c3c", + "source_signature": "2f68419ee8ccc0a23f55c3633983171519f8935e86b7815ddc1cfa25450aa63f", + "index_sha256": "d5ae132f92d621274e03ced62595c3e3ef0da58dfa048f9240667e82e2d03941" + }, + { + "file": "shards/seed_motion/00100.tar", + "files": 8712, + "samples": 1089, + "raw_bytes": 270518551, + "archive_bytes": 279296000, + "sha256": "e4fa467331a282a3a09094a01683affe417aeee0705b05503b81451c49391fb8", + "source_signature": "40cbce93441291f5fabad3fc75c5711b7fe0ede2c7dd7a705ad7f8788a454813", + "index_sha256": "8ceb9731ed80b721462c6107e5df858664069dc98233ea8214c05923ff334e86" + }, + { + "file": "shards/seed_motion/00101.tar", + "files": 8608, + "samples": 1076, + "raw_bytes": 277834599, + "archive_bytes": 286515200, + "sha256": "9f1f136e765761834f832a2cc18e554f1d35a2d10b513d06f03ab76f14d97b0f", + "source_signature": "0b0e6a40db03e3a1a9af478e87086df1b94b01eea7da437bf306a1410e3f6819", + "index_sha256": "735b9b2d45d91b3134eb7dea4cb23fd260f9034e4043b5c46036a246130b815f" + }, + { + "file": "shards/seed_motion/00102.tar", + "files": 9336, + "samples": 1167, + "raw_bytes": 289902164, + "archive_bytes": 299304960, + "sha256": "60c51effd15ccdf0bb1b283e5e38494b12c3a2d834425ba110b64e2c4a5c2dc7", + "source_signature": "c6b7e493a3489124b054f877d30fe3d181d5d917b97b92c2ba08c8928cb3aac8", + "index_sha256": "04207e541e13a4808d787a747701a79224f040e7962a45077b3480c26b33dc5d" + }, + { + "file": "shards/seed_motion/00103.tar", + "files": 8664, + "samples": 1083, + "raw_bytes": 271214464, + "archive_bytes": 279971840, + "sha256": "5a98ce312e7bd88625963004686f168fa805d9956f7b8c216ff0957e7817aa1e", + "source_signature": "b2009a6237e3b506a58ed15fd526b359afa11a55f9b98f1a6b27585f39697b76", + "index_sha256": "107a25d6996ef3fb0579d5657a9f5d79497d520d966f503ed0f27555ef860764" + }, + { + "file": "shards/seed_motion/00104.tar", + "files": 9232, + "samples": 1154, + "raw_bytes": 292114186, + "archive_bytes": 301414400, + "sha256": "9e2348ce5e373082a42a8b9d51012eb8a6a625bc1e23a3ee347d53e8357a394c", + "source_signature": "a110b658f2cb86ff9f9f32017abcd79c7e4e710493b78aa678a44a810a214ade", + "index_sha256": "3b1b27134a6beb2fa46e75a7eb1307e880befe3fa624cc0df2fd5005716bf727" + }, + { + "file": "shards/seed_motion/00105.tar", + "files": 8544, + "samples": 1068, + "raw_bytes": 274091552, + "archive_bytes": 282716160, + "sha256": "9405ee21016251e2a211a1014a7698c76b5be7e296fa5b76edd061470a6d459b", + "source_signature": "d61133402f6cc56f40d600efe4f1eda78a2eed3acaeee3e5aced8746dad582f7", + "index_sha256": "3827f5b97ff49cc3ac14fdb8550b383db6beddb94d7bbd9d583f93428a722c35" + }, + { + "file": "shards/seed_motion/00106.tar", + "files": 8576, + "samples": 1072, + "raw_bytes": 271895673, + "archive_bytes": 280524800, + "sha256": "42a95cadb60db431ac757742443ca4de1b10b56b9a8c746d9adb79e1afd20b5a", + "source_signature": "712284e6791c9abbde66ab9b3b8105b80fe0b7524a047588e3a0157fffbdd47f", + "index_sha256": "f1d925cdbf8b4d4395378a041ae0368c846cd279ba9026df5d1d87458de869cd" + }, + { + "file": "shards/seed_motion/00107.tar", + "files": 9120, + "samples": 1140, + "raw_bytes": 273800148, + "archive_bytes": 282992640, + "sha256": "45f04df408c9a3c9c2daa92bdc753efbb692462af7167e8b7056428d6af7a69c", + "source_signature": "b5befee98e7259e1708a009c0560f8844ed52d4bc0290f8310da8b9c7b20c003", + "index_sha256": "2b232f29aca23f6863b59fd5dea80898360e872fece38da46a8dd7b6530bf6c6" + }, + { + "file": "shards/seed_motion/00108.tar", + "files": 8088, + "samples": 1011, + "raw_bytes": 252410474, + "archive_bytes": 260546560, + "sha256": "6c7f805b6f2f92765f6af232b58029d63feb95f9fcb2c16399e29bc32744b775", + "source_signature": "1a01259273058666a2538ab5e8772741cb66e8c99d31eeb6647e262f88ad3102", + "index_sha256": "d3667ea705503122c41698470c06c7afc3e3fc0c5179022df95acb27d26c3d70" + }, + { + "file": "shards/seed_motion/00109.tar", + "files": 8944, + "samples": 1118, + "raw_bytes": 281245313, + "archive_bytes": 290252800, + "sha256": "9d4fe83efe03d3374306cf19de33aaebde79dc98ccef491ab22b922583ea9f89", + "source_signature": "ad8f4b2729820beed3c9cd4e3679d2c6c18e412b74aeb040ca379d8c03ecf29f", + "index_sha256": "4554cbaf10102923450c660e3295ea272a5154b250eac8e24977d50b8721c94d" + }, + { + "file": "shards/seed_motion/00110.tar", + "files": 8872, + "samples": 1109, + "raw_bytes": 283058989, + "archive_bytes": 292003840, + "sha256": "fb59c9d2a792d52c1be4204e1a7857d87913b5eb3b563a190dd2254f4e65d95e", + "source_signature": "46df67c4f2fa647e3779bd7b781b01045aaa720caaeaefcfa3b2280dd3728980", + "index_sha256": "8c006864afc76e31274bb617c54df9f3ce1c9b67558a7d29efb579693efa72fe" + }, + { + "file": "shards/seed_motion/00111.tar", + "files": 8649, + "samples": 1081, + "raw_bytes": 269221384, + "archive_bytes": 277944320, + "sha256": "1582c4da1ecd79d7918524be9ebf857c15428ddef17b8869d7011f778d4fe667", + "source_signature": "5d642f05a63a34c7333c4ffae043ec13598f982ce93a8db5c5c9557160ca4254", + "index_sha256": "34b7fd919667371fc01b666bcacfe0bca027324c2531c2cdd99f01b59b940598" + }, + { + "file": "shards/seed_motion/00112.tar", + "files": 8592, + "samples": 1074, + "raw_bytes": 270895884, + "archive_bytes": 279562240, + "sha256": "71f4397556ab9ab0bbfb0cd740817fc10e1870c263700b67db0922355c6552f0", + "source_signature": "2a11877d412767d67ca23b590d3d5e13202b938eef027dac9fa521d6d408892f", + "index_sha256": "eaff5d3e2fa20aef742f9468aae3fd94433e58ffae2587b5d3bec41b7989c9f6" + }, + { + "file": "shards/seed_motion/00113.tar", + "files": 8576, + "samples": 1072, + "raw_bytes": 261060660, + "archive_bytes": 269701120, + "sha256": "28c05dd161033300276641dd3b65403f33d0c871d9d17a7f6f0dd30e4ea7beb8", + "source_signature": "c91759493a5693b6576f02ac7b736edfba0e3b9b0ac6f1739ce8d2a662bbf34f", + "index_sha256": "2618182230cd521c2b1c953bc682d1085c14577b72eec2b0ca01e68205fbd75a" + }, + { + "file": "shards/seed_motion/00114.tar", + "files": 8864, + "samples": 1108, + "raw_bytes": 260706845, + "archive_bytes": 269649920, + "sha256": "5aae80b15aa72a65da102d3597edda2b02a268878afbf0b108a7cfa353ad683a", + "source_signature": "7eda13817f4d79f6d0aab4268d63c87558969d593b453af3280e11971b96c6e0", + "index_sha256": "f7e56850d8986d1e9d00783950e5a6c508f8ccc982516746ffb4e81b458cdaf6" + }, + { + "file": "shards/seed_motion/00115.tar", + "files": 9024, + "samples": 1128, + "raw_bytes": 275508509, + "archive_bytes": 284631040, + "sha256": "bdcfa534afb3f20e37b00ef501dcf73a0afddd1c52503b5ee43e6a125dc001a9", + "source_signature": "af26e5af981ed6ae6167d4910330cb0df1c5878f6c5e0b1bee16326ed9cbd764", + "index_sha256": "3a9d969263e2beac1d04a818e3d33f1140f53a87a9edff6039bf09f12735fd3c" + }, + { + "file": "shards/seed_motion/00116.tar", + "files": 8840, + "samples": 1105, + "raw_bytes": 279428000, + "archive_bytes": 288358400, + "sha256": "c7e2e8afea2614dfa323da6da31fc233a43b7e39cc9d463f3d61827f2c2f2ce3", + "source_signature": "1bc80d30e37e40d7b27da730bbacc86d9001c086670f89780261cbc44310fa6d", + "index_sha256": "de02a54a28b44041c79e0fdff0c8ee9c2601431a46f4929cc2aa2a0cfce9b441" + }, + { + "file": "shards/seed_motion/00117.tar", + "files": 8664, + "samples": 1083, + "raw_bytes": 266605339, + "archive_bytes": 275343360, + "sha256": "cc8976cdb007f204260881f20027f81a445eb88723f9d8ffe9271515217b52cd", + "source_signature": "1672aebfcb160907f031e4fffdf53920d4c06a29ee7babf88c1dc5301d5bf728", + "index_sha256": "805d3433f8d83ea27aaaa81300309a78aa8985041126e5ebb4b0a457569e31a6" + }, + { + "file": "shards/seed_motion/00118.tar", + "files": 8864, + "samples": 1108, + "raw_bytes": 269324379, + "archive_bytes": 278272000, + "sha256": "ab5d146518709809ab81d521e3e41d2077a412416e78b4b48e5261ff0780897b", + "source_signature": "e6d5c4125b47af8ccacb27bb93d78fee4c5ff1826f9eaba9142bf29671d6fd2d", + "index_sha256": "21ae58816f1c3d4b6fb8c7afcb49db1f43655daebe84a072a6010af5e14e8623" + }, + { + "file": "shards/seed_motion/00119.tar", + "files": 8928, + "samples": 1116, + "raw_bytes": 285826265, + "archive_bytes": 294819840, + "sha256": "6477bb669bd5a6afc7834a517605bfc2f20b8a7b5ebc29860b9f2834e86430f6", + "source_signature": "f1776a2169632549fd5938a1512100aa4b2de36e99b173f55b370337484b5a46", + "index_sha256": "4eee0f017295b464f252160bc3c87dcbb802f49401210f276e12ba1014f52f42" + }, + { + "file": "shards/seed_motion/00120.tar", + "files": 9008, + "samples": 1126, + "raw_bytes": 282142971, + "archive_bytes": 291215360, + "sha256": "7deb87c945fcd2c5314586bea797cc77a8b4a5aa6cc1a26079f33ec3f2a1eec0", + "source_signature": "c1879838c93bc8d97b316b377a1de35f98c1e4655858f8da71d3a990e60827f8", + "index_sha256": "6ac3fc591af7577791c01575d0e1e04ef7759b6131168507d174bae853ee2e83" + }, + { + "file": "shards/seed_motion/00121.tar", + "files": 8800, + "samples": 1100, + "raw_bytes": 250751612, + "archive_bytes": 259624960, + "sha256": "1329f2774b66193d1b8dd8b33657aacbfa34da7f7afcf5222545078e6f67fa26", + "source_signature": "555dad26e5bec8b2a33fa7ffa3429305d4666be6845e125473d600cb8a2f0320", + "index_sha256": "1d7fc1126c75b861a8c3775a19a02c25ef7ddf43fd030a97905a0b21fa9eeb0a" + }, + { + "file": "shards/seed_motion/00122.tar", + "files": 9025, + "samples": 1128, + "raw_bytes": 273627365, + "archive_bytes": 282726400, + "sha256": "3302eb76b038b71b8bab8b88528069cec30eaa0b8598fc199bd5e30e8c8756ff", + "source_signature": "b20f4a1d04a01a17e0220311a5bc3e7ceb768068d535168bff752548c5150b18", + "index_sha256": "3637dd7962d0a06a8e8356cb97a1e56bf8269a1b06678e1aaf940a8498bbff7e" + }, + { + "file": "shards/seed_motion/00123.tar", + "files": 8888, + "samples": 1111, + "raw_bytes": 279537666, + "archive_bytes": 288491520, + "sha256": "4e51262565f40137604bf0ac3161600b75ae01e859f714143227afa2a0ab46b2", + "source_signature": "cc0e5bd276d84774d2a8a89b6773e8bb92054d9a37f867db7e2ce1b0a8ca758f", + "index_sha256": "3bfdcbc54c87612b785304f3b11ed342015345a3b0affe217407c1af1b91f587" + }, + { + "file": "shards/seed_motion/00124.tar", + "files": 8544, + "samples": 1068, + "raw_bytes": 278916409, + "archive_bytes": 287518720, + "sha256": "66b3bf90c06e2fb4a47856c5c49ba226961b2d19ee65d02037009435f8a2e9de", + "source_signature": "101cb7c3dbaac39b738cefa08f8eed3ee894b35059c25d1bb9349d0ac2e22988", + "index_sha256": "86a4095bdf7d7265f978d4c854683d34681a8ee6163a669820a8225a9b3e0c0c" + }, + { + "file": "shards/seed_motion/00125.tar", + "files": 8904, + "samples": 1113, + "raw_bytes": 278379776, + "archive_bytes": 287324160, + "sha256": "b97d2865d431bcbc222731c639ef14c263d7a157ed8c104c6428ac1688a427b1", + "source_signature": "8e4e45ab90abf3041b5b51eef4e3672a97f866d8e046e87fe5a2a2614768e3b2", + "index_sha256": "72fd90c46cf8f32babfe4e3177a4a91fab632d7573afdb85ba05a9d33349a08b" + }, + { + "file": "shards/seed_motion/00126.tar", + "files": 8984, + "samples": 1123, + "raw_bytes": 285003804, + "archive_bytes": 294072320, + "sha256": "74bd78fbc027156d222692bbeed8f935c5f9e039bdfc20fe3cea2eb8601a2b5d", + "source_signature": "a0458a9113c99a12ccccbc788c65d3ba52d9023b6d09aba12c73d04a1f0b92ba", + "index_sha256": "558a9e7078f418d612c3ed42e70dc57fded90dcfdcb96b1253506890392fcab0" + }, + { + "file": "shards/seed_motion/00127.tar", + "files": 9056, + "samples": 1132, + "raw_bytes": 277608268, + "archive_bytes": 286750720, + "sha256": "51c9b70a50f9b614650488db532bbf6b206f2c1b8652ecdce072894417aa7f1a", + "source_signature": "fad3ae2e07d7c570b7b38f3925ec0a494af155ebcb1798b6569139f9ebb33a33", + "index_sha256": "4bad557547aa9c4c0ad8f0c54a8c4ba68d6476e6adf97120184a131a1d065162" + }, + { + "file": "shards/seed_tts/00000.tar", + "files": 370, + "samples": 370, + "raw_bytes": 58330520, + "archive_bytes": 59033600, + "sha256": "74dd45dcee765df943a308021ca170b9d9eb31fcf85fddebe085158b426c72c8", + "source_signature": "4f987f86d2e645d2ef6ff093ad0ce8c76357834511387ce215b65de858fc6c43", + "index_sha256": "173bedcb1fa48ed8bbf9e30884b5720e27c83719a31a720ec261c0a98a4a3277" + }, + { + "file": "shards/seed_tts/00001.tar", + "files": 294, + "samples": 294, + "raw_bytes": 46979976, + "archive_bytes": 47534080, + "sha256": "9fb9a346cc7e0db05721d047e00c2407b827951180106f0435e7ae1dfb8bad17", + "source_signature": "7fc58ab33eb33e7aa3beca91ed36508eaa38c428212fa9dbbb498b2ff9fdcd7d", + "index_sha256": "3b63f9f4b2c1176a8947ccd0d72dd290df27c192ed137517f4650f3d808e366f" + }, + { + "file": "shards/seed_tts/00002.tar", + "files": 329, + "samples": 329, + "raw_bytes": 51378316, + "archive_bytes": 51998720, + "sha256": "e54afc9948f05075654b4560f9c79c16393509ec7bac376ccdfbac1e64659985", + "source_signature": "fa0be160c7b6a38964a0cf71a2b0bf0661adc3c8a8c321ddae15ee6676105e2e", + "index_sha256": "fac7895050c5454197e54741aa4f7ae730c681377d42716cc61ecff1f1b7d670" + }, + { + "file": "shards/seed_tts/00003.tar", + "files": 289, + "samples": 289, + "raw_bytes": 46565036, + "archive_bytes": 47114240, + "sha256": "ddbca480026e5eec491ee2c2610c3d78e6787ac9b2b36b9f02c9a1dc2a223380", + "source_signature": "0d00d7d18c3df90b3a081281efaeef552e2459179b093eb98a52e38028544182", + "index_sha256": "14279f15f9a1905907c0488984f1efb0513e808464e449e6e880abeb997dcdd6" + }, + { + "file": "shards/seed_tts/00004.tar", + "files": 313, + "samples": 313, + "raw_bytes": 50540492, + "archive_bytes": 51138560, + "sha256": "cff147bddc1dbe96722e1161409c14cd8dcf6a2e2954495f8d97c0c217460f40", + "source_signature": "42f2e0a23ea2ad5a9948da814a4d1dc96fa398a51d1aabd192a8ddbfcfd52c7c", + "index_sha256": "d409246cfb9944386f71348ce870c2f515f4cff324234240e49563ba7af32d7e" + }, + { + "file": "shards/seed_tts/00005.tar", + "files": 349, + "samples": 349, + "raw_bytes": 54670076, + "archive_bytes": 55336960, + "sha256": "4b7f54ee0b8edcd4ef71a8043239b5d4b00ec55183ae9f83129e0a9e267ac130", + "source_signature": "ad8c77398c803d70068ffa182a5301ff28edffaf17c7ea645c9f9c3c5289f41c", + "index_sha256": "3cd714cf027587de3255724cc6b0c8724362ae0073d83cc0f799a6b7413380f9" + }, + { + "file": "shards/seed_tts/00006.tar", + "files": 304, + "samples": 304, + "raw_bytes": 48335936, + "archive_bytes": 48916480, + "sha256": "dc4fa2b9ff03f3ed1e0a4cf43dd5b3516e769b84e25aedba7622067ce5d0ea50", + "source_signature": "04edbb8695db0ccacbf9a24921551d2bedd423cc79a0e403cc1f4b5ad20949d3", + "index_sha256": "9a656a644e41a7931edfec43eef518145fdafcf3cce79111e9021939430b51e9" + }, + { + "file": "shards/seed_tts/00007.tar", + "files": 334, + "samples": 334, + "raw_bytes": 52484456, + "archive_bytes": 53114880, + "sha256": "90324f53da52c3217f5fde695b9984be6e96707a0786ae04a8c6f7c498e5aa75", + "source_signature": "de28b09d27110834de53f5bdb9d5f62f5a7519c0b71034c3235aa87d096723e1", + "index_sha256": "c860516117354016ced281e7adfaa52e04a8cd327f048c51895794bd86f9b749" + }, + { + "file": "shards/seed_tts/00008.tar", + "files": 335, + "samples": 335, + "raw_bytes": 55514260, + "archive_bytes": 56145920, + "sha256": "d5154d53ffd2d1b02ac008db5b80e5c3829ed7ac755cf71650f3f852a96bcd4d", + "source_signature": "65b9caec8126848394e605b432a31ad5a3fae82ac54eede8236b281f80feabfa", + "index_sha256": "d6b23aff5f7caaf0d235d350dd780ccf74e227e90d6589ffeda2ed4401c536d5" + }, + { + "file": "shards/seed_tts/00009.tar", + "files": 350, + "samples": 350, + "raw_bytes": 55614760, + "archive_bytes": 56279040, + "sha256": "e7630ba7f5face89f34c43bcfd4cb16d2208019c99568f113e4a3d89bdb58983", + "source_signature": "ffca0121b5945c7623e632790f2dae8970090282c581a3a04a92126ce0502a69", + "index_sha256": "82be21b3ea4d385d1faab2565f6ca91e20cbb6e5047f2953611bf523b456d617" + }, + { + "file": "shards/seed_tts/00010.tar", + "files": 307, + "samples": 307, + "raw_bytes": 48551108, + "archive_bytes": 49131520, + "sha256": "4c48029287a97ca3a1f616c88c8867ca6fe2d794b550ad7bfbd2320d1ee25a77", + "source_signature": "4eec96aabbc77226f6d1f91b80c67ec5e6344bec1bce800e4d0a979ba32fdfc5", + "index_sha256": "6cd5beabc67d77a96d9462a4cf86c85bb8e4482cdceb2fabeed4cf466fd51005" + }, + { + "file": "shards/seed_tts/00011.tar", + "files": 338, + "samples": 338, + "raw_bytes": 53974552, + "archive_bytes": 54609920, + "sha256": "a731bce23d5a8d18ea94abcf98af0dbc5276e473301ce76f53a3c7c8a4b4464b", + "source_signature": "4bd8478089390bc1e2a9f1abcb598eb0ce8f784db55cdcd8688ca4d75cf3644d", + "index_sha256": "c425e28e81dac39e658c97506ddcadae36a2255b776bbf5eeaaa046ace55f1d6" + }, + { + "file": "shards/seed_tts/00012.tar", + "files": 342, + "samples": 342, + "raw_bytes": 53947848, + "archive_bytes": 54599680, + "sha256": "5d31941eefe07192883f3736264a4ef8792c2b4f0e1485c6cb2a1918ff5e616f", + "source_signature": "419ce5952d53b816835de2e19733de29dd83e0cdbfdfe229525a184b468be66d", + "index_sha256": "c293f2a51252f5ae3ce506de623ffc1ebbe2d4322b9aed840dc3f8b15ad9e132" + }, + { + "file": "shards/seed_tts/00013.tar", + "files": 316, + "samples": 316, + "raw_bytes": 85694894, + "archive_bytes": 86292480, + "sha256": "72010e76e654f6859153dd715bed3b1e8fec8f37d76f5b0efe43a406318280ea", + "source_signature": "0d23ed1bc6c1c0e764814ca2164cd6e53d911edb7da49e09deae67e51bd85698", + "index_sha256": "61133c2e061b91a6b010d85405d2244cc18b985b74d103ffa54d21ec0080775c" + }, + { + "file": "shards/seed_tts/00014.tar", + "files": 313, + "samples": 313, + "raw_bytes": 52157132, + "archive_bytes": 52756480, + "sha256": "9b675bc2a3e3b1c21543d3c568efe73e74a585d9fa3e24a9e1d2d317ff8e1a9c", + "source_signature": "ab909b46a7ebd05569d4052da779114a8f410385731f66bffb11a5c7bdf85ab3", + "index_sha256": "41798780141aef3b5bddef8f17bc57cfa4ce2e8eb3a744b3de55401ce8b619e2" + }, + { + "file": "shards/seed_tts/00015.tar", + "files": 351, + "samples": 351, + "raw_bytes": 56762964, + "archive_bytes": 57425920, + "sha256": "8c17135a60cd30632d4f80e5ed07b33bf1d421665f3b501b2d5e3979a1765b1c", + "source_signature": "fc2905b5d21fa5f764ac9d83064780adf36b1bd76324cf952a3e64ec525e574e", + "index_sha256": "944e27841484096d5d557d70355350ad4462b3be025f09477a313057ab926bfe" + }, + { + "file": "shards/seed_tts/00016.tar", + "files": 340, + "samples": 340, + "raw_bytes": 55495280, + "archive_bytes": 56145920, + "sha256": "42d95f56c4c01112327fe6552780c2322b076a35ddca30bee708e3e2f053d703", + "source_signature": "8df0fed41ce2b2164f4c13a20483d76e6d8a3f26497941fee0b6751278d40d6e", + "index_sha256": "ab7f966c54609d3895be285a9ac47788165f0f78b677b93d6e3f7109c09b3411" + }, + { + "file": "shards/seed_tts/00017.tar", + "files": 338, + "samples": 338, + "raw_bytes": 54020632, + "archive_bytes": 54661120, + "sha256": "ae0f52be58fece7479d5926e29643b106f000457252a4380effcb7f14b5e21e8", + "source_signature": "86edc925dbe439f0c5ecbd8eea9693316576e19fe21dc32d40be828e00d59871", + "index_sha256": "fdb10a5329fb48533719255148b2b2f59ced40ce9295c73535b329a950e68e9e" + }, + { + "file": "shards/seed_tts/00018.tar", + "files": 315, + "samples": 315, + "raw_bytes": 49484580, + "archive_bytes": 50083840, + "sha256": "1273391423362fe33d9719d9ca8c3e544e6177d3420c741614708fce77c43c91", + "source_signature": "6937b5507b304a48ff55df481d042d5d2416874909e8c308994f08dd5e605823", + "index_sha256": "d93a3e19e934cb5d38d6e641a5c7c133e2978234f0b24e8a39434d05e91feda5" + }, + { + "file": "shards/seed_tts/00019.tar", + "files": 284, + "samples": 284, + "raw_bytes": 44921296, + "archive_bytes": 45465600, + "sha256": "1abff838fcc753fc90cee6b4717b021f3bcac23be6da99e0f521a3d09fce354f", + "source_signature": "43688ecad7b1deddc3988d03bd34da8ef9a296ce3451127db4d9ecadf689388d", + "index_sha256": "f6e7e717aaefd37d2c92b491600314821aeec63a7ceb69357465ec8013dd5b87" + }, + { + "file": "shards/seed_tts/00020.tar", + "files": 327, + "samples": 327, + "raw_bytes": 53716788, + "archive_bytes": 54333440, + "sha256": "5f1993de90570a4699f9c4a9fd5e440640cb6ee0e5dbd21003aa99271cdd22d4", + "source_signature": "e77470934bd7ef2617cf7322d6c01849c0e97c475884c5447d9d9b8c4f3b4ac3", + "index_sha256": "573c9f3603fd4152c2debad5ff2a123346d24efa996333262a3486bb12b5f18d" + }, + { + "file": "shards/seed_tts/00021.tar", + "files": 298, + "samples": 298, + "raw_bytes": 47690552, + "archive_bytes": 48250880, + "sha256": "0456e0597d990329dfd102d19075d7203355e5a821a1e7a3f4d60cd5eda5af18", + "source_signature": "fe3325914c7658faf7755eb0b2469a20a38b0b30f588127a979146cd8e4453ee", + "index_sha256": "9fc876f161caad9335c81a290e0c452da181582fc4dc6abc006dab8238c3d100" + }, + { + "file": "shards/seed_tts/00022.tar", + "files": 337, + "samples": 337, + "raw_bytes": 53321708, + "archive_bytes": 53954560, + "sha256": "05b017325742e44dfefaad2903ae2ab3583621362a2856861e37b33772af018f", + "source_signature": "c732a86d7ee2c951b7f6e9654be66dca644f45169f70f1032eadef797f9392be", + "index_sha256": "690e29df4c50fa4d64131db2c47c654b4e6dd2e31b04fa5cd2ad3ad85c59122a" + }, + { + "file": "shards/seed_tts/00023.tar", + "files": 290, + "samples": 290, + "raw_bytes": 46561240, + "archive_bytes": 47114240, + "sha256": "6e5ccefaede09e1b141aa60e87aa8a8b6694cc9abdd299ac7d1984832dd2ab1b", + "source_signature": "cd11a62ad50e062267f10a13aedd30b34da3e6f6948372f3fc6bd24308878cc6", + "index_sha256": "82677279f267638a18a468df446ed3f89b6af6e1bf6bef5c10c17c27ae6fca73" + }, + { + "file": "shards/seed_tts/00024.tar", + "files": 327, + "samples": 327, + "raw_bytes": 49101108, + "archive_bytes": 49725440, + "sha256": "35a362f1ae5b632794123d8e7c4886f68ed95b5ba459b30e749060101ee9ab9f", + "source_signature": "d9c5010a418aba3fb17e397ec8355cbb93e40079d08675df58e3f1f397b6a867", + "index_sha256": "66ac4ea74c701f6f8e0850579997b32892911edd0f6ed40e280b7d8bfe0d5b89" + }, + { + "file": "shards/seed_tts/00025.tar", + "files": 333, + "samples": 333, + "raw_bytes": 53717052, + "archive_bytes": 54343680, + "sha256": "1652b35a9ae15ca6cc0c73ece227ec4bc512c166c506f167c67e1aa39a226b80", + "source_signature": "f7f8296283427c8b66ca1b305c5297b22aee86a3a0f7030bd5ae0f62a14a6f37", + "index_sha256": "048cf62e45fbef2e5627bcfbcb101c08220dea63614109605023a9a75f59c46a" + }, + { + "file": "shards/seed_tts/00026.tar", + "files": 369, + "samples": 369, + "raw_bytes": 57627756, + "archive_bytes": 58327040, + "sha256": "b305e28ac72abc4d6801d8770645162d82bc4e83811df69a3e935813e1a1e92b", + "source_signature": "2a5e4091b6651cfa232bcbf6ab8fa431e3148daae079c99f55f52820c9a3f73c", + "index_sha256": "9f22dae8a177f5e3b7ae6e1b8f5251f81b51ff07a06cd993eda44227f37c0fac" + }, + { + "file": "shards/seed_tts/00027.tar", + "files": 351, + "samples": 351, + "raw_bytes": 55898964, + "archive_bytes": 56565760, + "sha256": "b73fca2780f218c806b19551288526eae127d5ffc9537c6c7c8bf6d5d00de409", + "source_signature": "c9724b29f4e165af322b67715c537638872f0be7a26dcb237aef98dfe38093f1", + "index_sha256": "c091807d5fa102e1b1858fc42a3468d9befd11117cf20369618c2bebb6dec4dc" + }, + { + "file": "shards/seed_tts/00028.tar", + "files": 343, + "samples": 343, + "raw_bytes": 53602292, + "archive_bytes": 54251520, + "sha256": "63ec5f57c6207a27d948367d5714a97773c04941b7ce6199466a3a0008a6e585", + "source_signature": "ef697a5d18b9cc758183fa6e518acea1f68fdddfa0c1f2808001f7378ac2955a", + "index_sha256": "f7a96ace116c40b88a119ad8d4c397612024d8baf6e94b31ddeeaf61d3620b97" + }, + { + "file": "shards/seed_tts/00029.tar", + "files": 321, + "samples": 321, + "raw_bytes": 48809004, + "archive_bytes": 49418240, + "sha256": "77b29b13385df8f2c28f773f812b2fe3b9d147d9f4459a7c0506c07a60643abb", + "source_signature": "6539da8050de42f1d54414b1d03553690130df50a63173512255d4aa2e983445", + "index_sha256": "6411fd9e8b1610880d5bb0a2b604f2bc27a224c176a55ec6f30792b51a6937ee" + }, + { + "file": "shards/seed_tts/00030.tar", + "files": 311, + "samples": 311, + "raw_bytes": 48320884, + "archive_bytes": 48906240, + "sha256": "0ef894eb7429443b79c2743a8dad8609038c88e0247df3a7068ac7848cffa428", + "source_signature": "9320d89c1036bdd866bc1366a7cfcd835eae17132de0894803a511f05679ab22", + "index_sha256": "5fe6d1f6b98586a1b9fbd31356b8c0a4a2225319d450309425df9bdd7b0b105a" + }, + { + "file": "shards/seed_tts/00031.tar", + "files": 288, + "samples": 288, + "raw_bytes": 44760192, + "archive_bytes": 45301760, + "sha256": "57347291516a8e276680b4fd695cf19a57b962bbc99c84f2b2d52d75ebd16448", + "source_signature": "f1e55ae920b588c18028a00fd62c0db51d62d7ad8f9cca3eee6de187d6b8119c", + "index_sha256": "bdf719e8e7910fa83202dba80c40ff153534de3f49ac2ec989b6cf46b3d97e19" + }, + { + "file": "shards/seed_tts/00032.tar", + "files": 329, + "samples": 329, + "raw_bytes": 51228556, + "archive_bytes": 51855360, + "sha256": "e54459237d214b7298473f0156e76137ea586e210e5fb2cc5b700ed47b3e7b35", + "source_signature": "88c3ff2ff2d94122744b201146c9f7cb91dba23b6c43d945f912f2a58bce80ea", + "index_sha256": "be093357c2a611364278c59fde00eedacde1b911ffd53d77e506e8e262e7169c" + }, + { + "file": "shards/seed_tts/00033.tar", + "files": 346, + "samples": 346, + "raw_bytes": 55902584, + "archive_bytes": 56555520, + "sha256": "76875967d8d6713cdbdb24b872a3df9f269f5fe686108ecfc718ea2be297784f", + "source_signature": "63ded01d7fa39957ba42b405ab587a3c99e735cfa9e4a98bc0513d67f9ae7f64", + "index_sha256": "27ac3dfa734b265f2322c55bcb36620d2ab9a8260b71a7b933b8afe1a6a10a25" + }, + { + "file": "shards/seed_tts/00034.tar", + "files": 311, + "samples": 311, + "raw_bytes": 49657204, + "archive_bytes": 50247680, + "sha256": "ad954ab2591d97f8e8c0932ea2479014788451ee8e96be4697743f84b2314241", + "source_signature": "fce3dbdf468e1ccfc15ab1d404653944341b2e8b13d9d63e630d9ddb1e8e4cfa", + "index_sha256": "25fe94b8421d19a094bb0831e342c70086343d4426c0fee19d461e8b248251c5" + }, + { + "file": "shards/seed_tts/00035.tar", + "files": 325, + "samples": 325, + "raw_bytes": 50583260, + "archive_bytes": 51200000, + "sha256": "341b0c29f04124578b13e24691d3daebf11ad771f3958b77297cde6508bc0b13", + "source_signature": "6e5cea6c79ce83a28a3078b085bf124ba924b7b44dc6d0a451c74de78e47c187", + "index_sha256": "3d60ce5a3bfc89b00dce209be81a70e8f0a1c76cc12e1dc081ef0f2b46f66554" + }, + { + "file": "shards/seed_tts/00036.tar", + "files": 349, + "samples": 349, + "raw_bytes": 57235196, + "archive_bytes": 57896960, + "sha256": "f7c9fa30c22e30b6902b47536e3dcb74c891ba43642e487bebe211e4948682f2", + "source_signature": "a860d1e2f995ffd4a9a7447f7eb9fc56cf4931fa729bc5751d88537bb67eacae", + "index_sha256": "24c4e24864a01e1409ac617c0498d41507afdef0f6a163a0f7c3034433237def" + }, + { + "file": "shards/seed_tts/00037.tar", + "files": 326, + "samples": 326, + "raw_bytes": 51685384, + "archive_bytes": 52305920, + "sha256": "0859ebf47f55922cbef5c501f7dd918b7a5b658a9d08c0145b70570c3bef6f62", + "source_signature": "6613db3d2b93f4096bfde0b6cfa9cfd5057b1bf8dff52837b1702cef681dd678", + "index_sha256": "dfe9fdae2e28d2f37d7cf2a81a7f1047a2e4741f33bebea03a5b06e2cf0309be" + }, + { + "file": "shards/seed_tts/00038.tar", + "files": 349, + "samples": 349, + "raw_bytes": 58079996, + "archive_bytes": 58736640, + "sha256": "23faa3b19f5b886def1d1c65693b079aa9fd4a5a80bc1e31bc57bd6009cd9dfc", + "source_signature": "723770108c232356760fd87a971d280f940e0db67100c5658265b32398cd7a76", + "index_sha256": "c0594ab00963444589049350c9f0005d869f9751ab0964837308b530a76f4fb4" + }, + { + "file": "shards/seed_tts/00039.tar", + "files": 323, + "samples": 323, + "raw_bytes": 53467012, + "archive_bytes": 54077440, + "sha256": "ab0dbcf1f6dc9ab1b3c7ae72cb7c6b152ffd749cf1c41d9324ca36231167f364", + "source_signature": "a2a96c2e8a8c88004303d44b51d64db85c38829ce8e04239614f5f4b4b6503ab", + "index_sha256": "b950735923cd8222ab200ef574315e3929b61ea8ef5055d093fa3d6e3979c454" + }, + { + "file": "shards/seed_tts/00040.tar", + "files": 333, + "samples": 333, + "raw_bytes": 53471292, + "archive_bytes": 54108160, + "sha256": "4eb4e163ee87039e6e8a8048f67b545b20f10351cfa48729b2d38915c9944203", + "source_signature": "c7f0ffd7d1f473d0649663b2095a673457d0a12c116375fe72e04e4eb47f8e2b", + "index_sha256": "721b90f1aaa9c22efbdeb76c18f7bd65e7c00f3b2a4f150c9c050bb66485e30e" + }, + { + "file": "shards/seed_tts/00041.tar", + "files": 334, + "samples": 334, + "raw_bytes": 56965736, + "archive_bytes": 57600000, + "sha256": "e93710e5908776b8d2dbcae4e9e82b72c89f90c37633b75a260de5212ab7e8eb", + "source_signature": "77666c5c16fe69ef1ee8780793204a1a2f1794483241486909557faf93601cbf", + "index_sha256": "f3609d3577ae8b160e0d0cfc33bed30210cc0f465d23c5868b3d346344142a47" + }, + { + "file": "shards/seed_tts/00042.tar", + "files": 310, + "samples": 310, + "raw_bytes": 48770120, + "archive_bytes": 49356800, + "sha256": "bbe4cbe5c1bc30914b6bb1bd7d55c267ec8fcfd2e5e2a97dc623cb974496f9f9", + "source_signature": "6cacf301e139303f1474a00e2e9476f64a449101a607262fffeaab6f4137b95b", + "index_sha256": "2555620e4d8f41fcadfa0299b4cf14771d0deedcaa8dc78331110be1a6a011df" + }, + { + "file": "shards/seed_tts/00043.tar", + "files": 310, + "samples": 310, + "raw_bytes": 48612680, + "archive_bytes": 49203200, + "sha256": "700d187b6ea9c009c6254b42bfaa9cfe85a93d9c33ae5050221fba5841d4019c", + "source_signature": "d4632b44ff07234b84cc3ded3dc1940b05bc7f14ab0f15f6c278316003aaac80", + "index_sha256": "ab4ab0fa8a53c3ff03d1d884039b56051878f1f3be44b0c11b047a6d4ce46b2a" + }, + { + "file": "shards/seed_tts/00044.tar", + "files": 341, + "samples": 341, + "raw_bytes": 55956124, + "archive_bytes": 56606720, + "sha256": "55eb29f0057cb3fce6519f3baa163ec3ca7fd5fc09d7bec125a7c7b8a205eca5", + "source_signature": "9959204a5753024725cb6f7dd1cebb30a656146462b28b7ac6fcbd471e4fad13", + "index_sha256": "42dcca63484bb5fa506fdfb7503bb6ce17bea70ce3ba0ebe3e0e4ce79755bedf" + }, + { + "file": "shards/seed_tts/00045.tar", + "files": 323, + "samples": 323, + "raw_bytes": 49031812, + "archive_bytes": 49643520, + "sha256": "303d7eb90da7a292e9b913fe04cb2a9e6ee59bc0c85168f52d0d5245ec7b588f", + "source_signature": "7704e6af155835020d966aaa8915849c7f0ce4e1b919e3ed54a1779024b75500", + "index_sha256": "8588a5dedaffd54fc3625d0decbb25b5d4ac74b23530f6741ffb193251de6fd7" + }, + { + "file": "shards/seed_tts/00046.tar", + "files": 309, + "samples": 309, + "raw_bytes": 47226396, + "archive_bytes": 47820800, + "sha256": "9afe2f2b3bf5225a36731a27abbc6b357af749cb7f7af5fb5a0b1bd9f038282f", + "source_signature": "99c57b4e27678e046b4b8a4567532b70461170d61c567efc9f6c4df6e48d59eb", + "index_sha256": "cda7d2f7d9c10a6d11db4a42a7aef5e681aaf89d1e15b8b41767ba3b6ec40ac1" + }, + { + "file": "shards/seed_tts/00047.tar", + "files": 340, + "samples": 340, + "raw_bytes": 53901680, + "archive_bytes": 54548480, + "sha256": "55813aa2fd12b05e4078179ca411ee2127f8aa028033607708cd8a711a46ab11", + "source_signature": "d65a311d8d86de9165102cf4798ac4bd2ce84cf2ebe8c92df8852d655a6a4236", + "index_sha256": "bd718566aa37fd29aa9013543eed4a7fc48580394c1f2ff261dbabbd10172b38" + }, + { + "file": "shards/seed_tts/00048.tar", + "files": 295, + "samples": 295, + "raw_bytes": 48055220, + "archive_bytes": 48619520, + "sha256": "ff0d7cac73a2ca7a7473da2d1ba3d64d8a9916664e8023debcbf51b3e451461a", + "source_signature": "2f21ab758f452e48bca8e37dbbcf8864deda708256fbaa0bb64c7d773d76e514", + "index_sha256": "03d0265b78de5a3127adbaedd8868b96e13f96757dcf68aa1612d43a0fc2139f" + }, + { + "file": "shards/seed_tts/00049.tar", + "files": 326, + "samples": 326, + "raw_bytes": 51474184, + "archive_bytes": 52090880, + "sha256": "fea504532379bdd32a87fdf131e280bc7850387c1307b532866c4b5c1d25d288", + "source_signature": "4ad01c1e849fcd625d6a3bd034c02f888dd33240ee7b56ef84022a11e7a7bd6e", + "index_sha256": "934af6637e649c783e1abd59e412cece1702ca648762d65c1b63726d9a30764b" + }, + { + "file": "shards/seed_tts/00050.tar", + "files": 324, + "samples": 324, + "raw_bytes": 51259056, + "archive_bytes": 51875840, + "sha256": "8815ebd38609722c851595a479c502c1e75bf055801121d705c8e01797b7a12b", + "source_signature": "6f3f8ea0e742ab3e4d11226dd2d54ecbbdebc3263b530e9498b73e3ff0f128b7", + "index_sha256": "b4dafcfd028c4d6237d4b18d53022c0faaa8d4c74090d79a8d1c75126ab961b1" + }, + { + "file": "shards/seed_tts/00051.tar", + "files": 327, + "samples": 327, + "raw_bytes": 51055668, + "archive_bytes": 51671040, + "sha256": "3650a8b24beb6a623652d083d522fbf78035084e5dec33440bdb4b298267d702", + "source_signature": "dd6cf7aa1a6f8f8a84e0607ded37c1d52218f9625da6f63c43240eb4186c5e34", + "index_sha256": "b3ceb269173746a535d9de31ecfaf6db563b91bbb9770bdab7abfece97aa131f" + }, + { + "file": "shards/seed_tts/00052.tar", + "files": 339, + "samples": 339, + "raw_bytes": 53882436, + "archive_bytes": 54517760, + "sha256": "517338a96ab01522cf1129002e41993082e68f4c6847c7ea31f04bf5d091e0e1", + "source_signature": "8d058b3109c7599c722fb749e93c644f4c9ec3c03605641a14ddec9f3b1631dd", + "index_sha256": "70c8e1d95a5bae8af1a69daed809d559455cd2a1e01c103648ad8e1273435f57" + }, + { + "file": "shards/seed_tts/00053.tar", + "files": 343, + "samples": 343, + "raw_bytes": 54343412, + "archive_bytes": 54988800, + "sha256": "a5c836dd69583145d576065ee450c0a388a1038bfb5e0e51e54e146bea132adb", + "source_signature": "ceba272a850ffc0ff34374142c0f449dac607c7b68edcce0f01ebf54391f09c9", + "index_sha256": "4f0696f70bea99c06921bdc2e538b47b523290576d67403827cf3ef0845a07b8" + }, + { + "file": "shards/seed_tts/00054.tar", + "files": 316, + "samples": 316, + "raw_bytes": 50348624, + "archive_bytes": 50944000, + "sha256": "117fc6882c647aa27df805684fe2970ed42669d47f123b54bcd62776bea1483b", + "source_signature": "4388f5d9be3d675719e1e2f8fb06753c06ed37b32682c27b3ef68bd942668e42", + "index_sha256": "2bf70033790874171e449a576024869f648c3f667f9ff2b57b5571905c564aa7" + }, + { + "file": "shards/seed_tts/00055.tar", + "files": 327, + "samples": 327, + "raw_bytes": 51374388, + "archive_bytes": 51988480, + "sha256": "57d7af449e2dd846d238e7d0159e1ec01f627415f1fbba848ac7b0fde5c1fc29", + "source_signature": "2ae429374575d88fe62ef4b2481b7f4651aae7725bd501e8cfc6e871b7786ea4", + "index_sha256": "f07c1e0fddd3eccbfeb3497a64f649c91a7435237a2627d5c3dbcb6c59d23364" + }, + { + "file": "shards/seed_tts/00056.tar", + "files": 322, + "samples": 322, + "raw_bytes": 51923288, + "archive_bytes": 52541440, + "sha256": "7b8fc71a69fea209648085a35f75671c9941472f5b3ab6a373ed3aff45e7d69d", + "source_signature": "5072c814bf30c3fdabc5530f9c4268a31214d379a4f18a38dd14974938378f13", + "index_sha256": "915151c221b091e93088cb20570dcb138f095487ba3ed1f1cf93bd41465f672e" + }, + { + "file": "shards/seed_tts/00057.tar", + "files": 311, + "samples": 311, + "raw_bytes": 51385204, + "archive_bytes": 51978240, + "sha256": "a47e1b676044274dc5125658acbaf20056be7836cc75f45deb4c5d3afe2c59d3", + "source_signature": "de7c00c0c957e60b756bdac31d2caac8471445f3ec6418de1dc0038ed36c002b", + "index_sha256": "1344c517442a275387c220884ff5a03473808cc78dbd84911f14d9fa9b0754e7" + }, + { + "file": "shards/seed_tts/00058.tar", + "files": 323, + "samples": 323, + "raw_bytes": 53747332, + "archive_bytes": 54364160, + "sha256": "973f577f709b9a901b6e5943a792757b8be2631e68b9abef9baf53237ce667c1", + "source_signature": "6fa3396f4f5527d4bb87d38ffeabc6e10cccdf2b36db9461b1c30c8439be8d15", + "index_sha256": "abced9417af2271347f317009e2af10130913c0bd7e7e71e70a6e4804e46be34" + }, + { + "file": "shards/seed_tts/00059.tar", + "files": 337, + "samples": 337, + "raw_bytes": 55890668, + "archive_bytes": 56524800, + "sha256": "9eab4a7b4d3bd7dad796b5698c8338b5dffb7c30d240e70dce5c4e20ccd48feb", + "source_signature": "1ef7b65f93d90ff1f3a1c306a211965511418a86d115a36583951653f89b3f44", + "index_sha256": "37dfbc12eb28447d045970fbddbbcad79ed41c5cfdd783f131784e781bfdf328" + }, + { + "file": "shards/seed_tts/00060.tar", + "files": 345, + "samples": 345, + "raw_bytes": 54420300, + "archive_bytes": 55080960, + "sha256": "b54606ba8fe9c134a984a3411705b88eaa0e593e385425bd998160c0c164668d", + "source_signature": "b86615e305c9e9dbc548933d8712d07ef01bc645135d4b2dfdc064b734d9f0ac", + "index_sha256": "9f20d45d9ecaebf282b71af8da1b65619f1b635eb9a588a951fe989cfe405654" + }, + { + "file": "shards/seed_tts/00061.tar", + "files": 324, + "samples": 324, + "raw_bytes": 51466416, + "archive_bytes": 52080640, + "sha256": "c7347031a33fb75f1a1a3afcff6fa4abeabed75256986a157d0b20074bab591f", + "source_signature": "a09438d6d955cf4ee02e59b68b39f20a1b87e191f335c937db7685c6a10c4ba9", + "index_sha256": "c3002d61edccaf38d89122175fb0ef692ee96cbf765da5d1507fb4623beb950f" + }, + { + "file": "shards/seed_tts/00062.tar", + "files": 329, + "samples": 329, + "raw_bytes": 53954956, + "archive_bytes": 54579200, + "sha256": "25c219d466bb484797a78081a8d1bad3b710b12d803445a89ea1106e24a07a30", + "source_signature": "30b888d176edc46799878d4e438748effd8f7defc65694acd76b993dafa76352", + "index_sha256": "8a72a38e2d973233efc6002f0d83865bcf81e3922b5182c6eb70561199ec4d81" + }, + { + "file": "shards/seed_tts/00063.tar", + "files": 337, + "samples": 337, + "raw_bytes": 53690348, + "archive_bytes": 54323200, + "sha256": "74b3b80f0a1bd4f7de92318faab0f7cda761bb71a4b3e68bc1d29168f725d0df", + "source_signature": "737142e58dc67078e5988bc788179ebb16c2c70ca9345cbdd4d73d0bfc403411", + "index_sha256": "7414603d0860040dc9533ef78d0307a2016eee817b6775b733bee762da76ce46" + }, + { + "file": "shards/seed_tts/00064.tar", + "files": 309, + "samples": 309, + "raw_bytes": 49342236, + "archive_bytes": 49930240, + "sha256": "bc94ad73c4602fa6f27796db386fd0a52b811066a81f6bd29b83093d38dc2c79", + "source_signature": "b34f237fe095b5d5e87011d9f08cd24b7757cb5c1c3e125606f53849351aaef5", + "index_sha256": "d42ba8fe5b3b5e37450f9ee09330f2ac6e4e20a38f75356db26bd7c2edafbbd3" + }, + { + "file": "shards/seed_tts/00065.tar", + "files": 322, + "samples": 322, + "raw_bytes": 53317208, + "archive_bytes": 53923840, + "sha256": "0a62f269c76b7348b42bcfd4b1436c686a0f4484329d13440a852569691c5058", + "source_signature": "2d785af8c96b88491769f45348a515eb1f40b41d35f255297d5db503452bc864", + "index_sha256": "1498059018657d59f0593ca3e1a2cfbe0d4820d82c99a3edab23e7313b720692" + }, + { + "file": "shards/seed_tts/00066.tar", + "files": 335, + "samples": 335, + "raw_bytes": 55326100, + "archive_bytes": 55961600, + "sha256": "16a6d1564af649c31b1f1639fa841d1f32381de72b19ba3b1bbff571abe2f110", + "source_signature": "e3818fa80bfe6b9071c1df554c38c870d507f83b19d969af87747c79640bf5be", + "index_sha256": "ba31145759d1da60fc46653e6e2681e704210b826287d2f8a0c53370c9b591cd" + }, + { + "file": "shards/seed_tts/00067.tar", + "files": 314, + "samples": 314, + "raw_bytes": 52806136, + "archive_bytes": 53401600, + "sha256": "54ec30153c1ebb815cafb05d10391cb1553fd723c8a7d98818f42c4e96cd4604", + "source_signature": "aabfdf5399d13de4e5ae0950f5316bddfa7b16b5af131aa4de6c8c42ae167840", + "index_sha256": "4eb0176b42af2058209e8ca3ca55c8fd55adccb1985c0bb03fb714c1a0562c3b" + }, + { + "file": "shards/seed_tts/00068.tar", + "files": 324, + "samples": 324, + "raw_bytes": 49999536, + "archive_bytes": 50616320, + "sha256": "271e0e676f23377701c89fad30328118bf9921d46de46ee0f5f2da32fc0b6752", + "source_signature": "4e32f366bc0c6d8a4f8e82994f29334f2b4cb8a76c1e4e68b102096acfae0117", + "index_sha256": "2eb95caca764438d034d84d5c31ec818d62b031350e5c444f6fd46eb4ae77892" + }, + { + "file": "shards/seed_tts/00069.tar", + "files": 337, + "samples": 337, + "raw_bytes": 53329388, + "archive_bytes": 53975040, + "sha256": "12f5d299ac387e56fa02b2841c8c3fb4b6d0b585f2408e15f871030c0badb208", + "source_signature": "4bf556226e1674feefaeaa5d5386edc3c6fca92af02c257d89cba106ed68a2ea", + "index_sha256": "45e91b77897fe39557d821cdc562a81b36e9aa8ff27bbe5ee9d109011194867e" + }, + { + "file": "shards/seed_tts/00070.tar", + "files": 354, + "samples": 354, + "raw_bytes": 56928216, + "archive_bytes": 57600000, + "sha256": "c328736a3fae0d369d2bba60fd7dec1ad975473657899bdb5533dd108981338b", + "source_signature": "5d430f98175d9bf7002e3dd2c71c076cfdbb57519a435bc58834b4c234b28f15", + "index_sha256": "e8039ac678db98f5972318b104653dfd4ccfd80f42d885217e3741a960779a3d" + }, + { + "file": "shards/seed_tts/00071.tar", + "files": 360, + "samples": 360, + "raw_bytes": 60588000, + "archive_bytes": 61265920, + "sha256": "45398199c8bcfd051ef038ba8dd75c5ed9e5ae11256f3e32daaac6fb07ae6ee3", + "source_signature": "83ba57ebfa9432df83345939d8bbf77d0c5191a527d773a8cfcd2c7068483942", + "index_sha256": "167a7859ce273b888f9a676cce3e60bc729c4d7f899a22f4a8ce2ba7a691dbf6" + }, + { + "file": "shards/seed_tts/00072.tar", + "files": 327, + "samples": 327, + "raw_bytes": 50448948, + "archive_bytes": 51066880, + "sha256": "b66f520eb88a4be2942ce7cf765323963109fe4b1a7b57ba228c5c142a85f343", + "source_signature": "65e22959aee9097408c561177b6e0b270dc746de335398feee8870c81bc68888", + "index_sha256": "a29a2ad96fcde2dd9592bfc166411af5a8d303a867f9617dd5c3bdf5aeab07a1" + }, + { + "file": "shards/seed_tts/00073.tar", + "files": 338, + "samples": 338, + "raw_bytes": 49965592, + "archive_bytes": 50606080, + "sha256": "c338ab11e33e3691528dc5b7382058f73743bec119c0b97a8aa2e5ba459048d1", + "source_signature": "41b0e36623ff39422dd9f83b993b1ff5dacbd0446b838830480e4961790b1c61", + "index_sha256": "ad6088b21c1acdf0f0d77f8b1792c2bed409988dbfab6e2635906532f3498cc9" + }, + { + "file": "shards/seed_tts/00074.tar", + "files": 293, + "samples": 293, + "raw_bytes": 51418972, + "archive_bytes": 51968000, + "sha256": "5e1b5897177c61683b9fb82dec4762cc68b50c802b3f53e9a24e2e27a8271fc3", + "source_signature": "27d35516b8a653e402507360bca8b3ee021cfed80dc072bc55af6330ede44578", + "index_sha256": "84698aff6068e89f3d2e53247110c9f014d4beb1d0393823c62174de0b04db2c" + }, + { + "file": "shards/seed_tts/00075.tar", + "files": 308, + "samples": 308, + "raw_bytes": 47687152, + "archive_bytes": 48271360, + "sha256": "1b0e1504b68d7548854dfb3ef6971f435c9aecafde5875e669c684215ab917d8", + "source_signature": "23683027a5b8c607ad38c76283e24c7ee29c83c36e09f707e658f7e431c9159c", + "index_sha256": "24a5a2428e0b588747b008979f372b3cdadd4f6281d18fdb6cd3d34bd5b82a3d" + }, + { + "file": "shards/seed_tts/00076.tar", + "files": 318, + "samples": 318, + "raw_bytes": 50325672, + "archive_bytes": 50933760, + "sha256": "c3845668760e7f8107aa2b83cbc302be09c8cbf4d89c92db85292b17d04d329b", + "source_signature": "86b02675bb4f6af53a6217469a4bdae0764207c9eabf47033cb872074b3d85ae", + "index_sha256": "180af87f0732d2a4b66c422b8fd5df6b67cd183d6939fdde5c684d2213e1f8e4" + }, + { + "file": "shards/seed_tts/00077.tar", + "files": 318, + "samples": 318, + "raw_bytes": 49914792, + "archive_bytes": 50513920, + "sha256": "c1a0e9bc3233ee56493d228867512c2d8babbca4e7c48903a848c3fd67fcc217", + "source_signature": "ea99759713faf4eff3e78dbf1b915972c8def7a0ed80cea00bc8963c0e998734", + "index_sha256": "78349e5c20a83ef754ce6bdd176b14cab9ac67647cb9cfd85e02d02ba860b46a" + }, + { + "file": "shards/seed_tts/00078.tar", + "files": 357, + "samples": 357, + "raw_bytes": 56843868, + "archive_bytes": 57518080, + "sha256": "eabb0dd8651cade80aba3d1b3c79271a30ee78a8bf4fe363cdbb1b8f6dee6380", + "source_signature": "4ad4e5c3247fefe40f7fd9c2c3e6507c9b108c373211029a862d39475e9732c9", + "index_sha256": "9b658b98be78b09076741077ac45d0a52f4e940c506fc9fb48b41e317676edd7" + }, + { + "file": "shards/seed_tts/00079.tar", + "files": 319, + "samples": 319, + "raw_bytes": 52049876, + "archive_bytes": 52654080, + "sha256": "abf502a2f0f535e8b9f7df374b14e825fe4e8d70df5aef84354602dfe213eb35", + "source_signature": "3086431ed112e61eefed9a646ee48647b38a4f6983fd65e451779e23b7e0fd3a", + "index_sha256": "1adfd685a1b47c1e2fb9dd3bf381a39f83792c14f3be8bb9788ac29e4d942875" + }, + { + "file": "shards/seed_tts/00080.tar", + "files": 333, + "samples": 333, + "raw_bytes": 54139452, + "archive_bytes": 54763520, + "sha256": "37eddc1d861b7edeac0b4393534e0fb6ae910654f01c88833691456509fd25e7", + "source_signature": "5c0ee37fe5da15191885049315225499df8b74dd3e355886976aa110ff837074", + "index_sha256": "5378f6ae8d063485f11dd13d50c25a5998a3cc25c58030df456b0f1658840d29" + }, + { + "file": "shards/seed_tts/00081.tar", + "files": 314, + "samples": 314, + "raw_bytes": 50770936, + "archive_bytes": 51363840, + "sha256": "53c06931d68dc82136069f3d70096d02dd13c1991118d169f2bfb3b3d6d9ffe2", + "source_signature": "5c94d44c7971e969c79b06c38f82e352e0a3b34e28d7811b50afdf070228f0f2", + "index_sha256": "bd43ae88ec8fd4ea9d839d03f24bbd704ebe610849bc5b7f4b7c010fdc6eef0c" + }, + { + "file": "shards/seed_tts/00082.tar", + "files": 328, + "samples": 328, + "raw_bytes": 52683872, + "archive_bytes": 53299200, + "sha256": "cdcf6665ed6735daed36ae962309deb92e8de208ce09bd76e473b69230753e83", + "source_signature": "2c33c90b7b8f568eb7a1ef9d8e895155e072d17499b6f053635e6237ce062b67", + "index_sha256": "49dfb2fdb0948f00c08169b2f38eac09d15be5ca1ea1824e1a56782122c73356" + }, + { + "file": "shards/seed_tts/00083.tar", + "files": 353, + "samples": 353, + "raw_bytes": 58110892, + "archive_bytes": 58777600, + "sha256": "1bde8b65bf2778eae16d09c41c9ac020135e75e5bfdd8ed740f6140ee9a11d87", + "source_signature": "5adab9003e2e9439d81cffc1618556cd2aa57c612a36d439ea936b0a7f666cdd", + "index_sha256": "c713d70a14c8fa1e8a30aa87555d15ddc2df03c76ec6b4838d0953d6a297e2a3" + }, + { + "file": "shards/seed_tts/00084.tar", + "files": 317, + "samples": 317, + "raw_bytes": 49695868, + "archive_bytes": 50298880, + "sha256": "aaee475cc31194d726b4fd6655f0ef988315a06f5aa0340bbe09209e7d9305d4", + "source_signature": "238259707316f27e1f954b2f6ce95e3a0f57c6d1bf89fd84bece8b029d2f8757", + "index_sha256": "80a746115ad811611ad78923486c9a8c1a14df6ad3f95b7e5ef28dd51dca0962" + }, + { + "file": "shards/seed_tts/00085.tar", + "files": 330, + "samples": 330, + "raw_bytes": 53167800, + "archive_bytes": 53790720, + "sha256": "5bd3bf51ef632fdc02cabd2bfd0992813c6fa0493fa930872ad4616c04dd3956", + "source_signature": "40019a5b0d5bb895628f5a1e6670d192b5d53916088763224149429492aab784", + "index_sha256": "dc8607b447397aff0fdc2ce1a41b13714f51b78da01f5de95d46a9ae1d9cd03f" + }, + { + "file": "shards/seed_tts/00086.tar", + "files": 308, + "samples": 308, + "raw_bytes": 49849072, + "archive_bytes": 50432000, + "sha256": "95be5b1284663aa422c8193baf276820fc091284b66c58a95d67258b316f80ac", + "source_signature": "7143cc83ec81eaa176825665fc4abc3ef34372378cedbe8f3853cf4a3929a780", + "index_sha256": "1d71defd41854b6d47439fcffc35ae6632001cb1e0d1cf7ec19c0b5199abef64" + }, + { + "file": "shards/seed_tts/00087.tar", + "files": 305, + "samples": 305, + "raw_bytes": 47744620, + "archive_bytes": 48322560, + "sha256": "20f5940a37b7fd7f5a731c0f99cde67f9ce75dd3fb780f3048a16a2e735853d6", + "source_signature": "24648ab46943f8d58736013683dc3278ed34461483e652598384e821f92e9cab", + "index_sha256": "cc55485db4304b7c6af79e2ae9a7e88c2d5315943c87b253ec5dd6dabc8c5e8f" + }, + { + "file": "shards/seed_tts/00088.tar", + "files": 307, + "samples": 307, + "raw_bytes": 47552708, + "archive_bytes": 48138240, + "sha256": "430c0cbfc548263bd463320665a6adef100ac824fd9b82b3cca097bdf04aee48", + "source_signature": "f5c324bda5a8727a54dd879d1a6b20142105650715b712c3b80d728cbde005b1", + "index_sha256": "da980ee2ffdf28bfd8dcccc45baab8737e39dfe011f9b97a33824dc76645f4f5" + }, + { + "file": "shards/seed_tts/00089.tar", + "files": 333, + "samples": 333, + "raw_bytes": 51159612, + "archive_bytes": 51783680, + "sha256": "e833d0b04b08681c6f5d91c9a298a0da9962e7cfd7b9b55488404e42bab58279", + "source_signature": "9c9fce514d364196ef729ab64b9796f530759f7d3913d6ffd14e0aa0ed254e68", + "index_sha256": "5c97fbee4096bde4c3774cba90651c39c76197d9d8dbe138ef420f1e125f8fee" + }, + { + "file": "shards/seed_tts/00090.tar", + "files": 315, + "samples": 315, + "raw_bytes": 51538980, + "archive_bytes": 52142080, + "sha256": "13ccd7c33064b3e613ccc827efc6b2860b8111cfec5422111e4ce293e112f4a7", + "source_signature": "cdd60fafaa06fc34616b4d657fa08770344703592039513fab92e2ccc167dcd6", + "index_sha256": "2cafd2739d27f336433694c065116ea0de60295e84d8610e9754065d638e1643" + }, + { + "file": "shards/seed_tts/00091.tar", + "files": 342, + "samples": 342, + "raw_bytes": 55967688, + "archive_bytes": 56616960, + "sha256": "5cb1c96075f8a08f23655334bb3ab4b4f8c615964f0e5648d13bc5cd19282a33", + "source_signature": "6ef68a52496ccf33cfd61d4a057ca78019c040c5dfb226f256be7ea0b2d886f8", + "index_sha256": "1aa996df5b43af3b5837dd7f122ff22c049493694eded0aab37d0aa2656e6da9" + }, + { + "file": "shards/seed_tts/00092.tar", + "files": 337, + "samples": 337, + "raw_bytes": 52903148, + "archive_bytes": 53544960, + "sha256": "d1b40bf6df367f862c4be7f2bda2879d6b796e6a2bfe39550e09aad2e31bed53", + "source_signature": "dd33fef4e993865d0b99aedc2baa49b0459daa4533be9878653057d450bf8d1d", + "index_sha256": "d2742c77867cedf2e12ef35e152a9f5ee0d65947d125a84892a4a9721d196b4b" + }, + { + "file": "shards/seed_tts/00093.tar", + "files": 319, + "samples": 319, + "raw_bytes": 48543956, + "archive_bytes": 49141760, + "sha256": "b6c248443df1a550419f04d00b96cbbce87cb20c9294628df969913de86b0ba2", + "source_signature": "04362387471e173842b73c385f24cbbdd032079dd49e40874c0631154173e219", + "index_sha256": "70d890d53d9a508a7857f1f3d20dddef34b9c9566080aecb28f0a097daa9a3d2" + }, + { + "file": "shards/seed_tts/00094.tar", + "files": 321, + "samples": 321, + "raw_bytes": 51082284, + "archive_bytes": 51691520, + "sha256": "fb4fee6022520bbe8f0145425d113bc8f18255e3d4d5ed5dfbca902119c15883", + "source_signature": "66a9a33eff4793ce99453c3edafe7f55b8d62225ccc0d0cce78ba3ce64d3e88e", + "index_sha256": "12ec75f37c90dad3602e0056afe68bcce1415c3d98a50c9fd21559722d14e06e" + }, + { + "file": "shards/seed_tts/00095.tar", + "files": 318, + "samples": 318, + "raw_bytes": 51907752, + "archive_bytes": 52510720, + "sha256": "efd14bc486beda15fb1c77765f1c92d978a480e2992dd949c2b16e07a9d40699", + "source_signature": "5f1d9d04646c2aeeb1daf579166ee3898e8be707422376bb28310ea2640c71f8", + "index_sha256": "593cf5ba8813ea23f2ddc3332f34741cfbc9baf549970b78b2bb6398e3324bf2" + }, + { + "file": "shards/seed_tts/00096.tar", + "files": 323, + "samples": 323, + "raw_bytes": 52011652, + "archive_bytes": 52623360, + "sha256": "30036122cad284bf7965868005d41dca854d2d28c68ee03382b30dda8cc94b18", + "source_signature": "466e80c2ff97cef5d85215ac36c4502106e38c71ec3eaff12b1ad0664241de99", + "index_sha256": "1b12c164ef1efab57ec08de55167452d9380f1d6a02193cc33e2e9beb320fb23" + }, + { + "file": "shards/seed_tts/00097.tar", + "files": 324, + "samples": 324, + "raw_bytes": 51539376, + "archive_bytes": 52152320, + "sha256": "6ab8b2f0818f7fc01c375ef19765d4e98b2666cb9f00b1eae55c4b49e1a015d9", + "source_signature": "98d9fd4932b78eb925c1fdf8ad8fe424d056a7ecf052080e20832183940a8a99", + "index_sha256": "a907b0a9de7ae4975d2e4d85c119b99378e7b72d0416abbf959bd1fd4bc51c4d" + }, + { + "file": "shards/seed_tts/00098.tar", + "files": 324, + "samples": 324, + "raw_bytes": 53298096, + "archive_bytes": 53913600, + "sha256": "4e8c6a8c387ab98be19e4f8513be9cfabeac8ec38d6b26f6f0c66f739fa0631c", + "source_signature": "9e277260af6ba579b20124172dfdf62ec7bcab4e60ffd1adf94b2ef6b35ee3ec", + "index_sha256": "80b383b71a33c08c5462057d0d9c8ed53c9826b305c431f9a390a22844d04058" + }, + { + "file": "shards/seed_tts/00099.tar", + "files": 343, + "samples": 343, + "raw_bytes": 57265652, + "archive_bytes": 57917440, + "sha256": "6b8dc5fa0488506e6a5dacebb5cab16b9fffa86caf4365667f270794c397bc68", + "source_signature": "8474c645208576b606e78307821d436e89b64974b9ed4d1402bf21218875cfd7", + "index_sha256": "0e2fb8a1812e47bc8d3ebd505519f5379c3998a7bcfd4e70026c711ddea1b430" + }, + { + "file": "shards/seed_tts/00100.tar", + "files": 350, + "samples": 350, + "raw_bytes": 56290600, + "archive_bytes": 56954880, + "sha256": "d7541f23e3c5e016de03d1665ae87a344fe04d2cc6340761b235688c5e6da2ef", + "source_signature": "cc064a1d52500b0bd0cee5c59cd810e2bf16f43b13099555da988ec1c99f3d9e", + "index_sha256": "c7c438d37be1a680569a89f8dd487ba94f1bd8c4163cbc55950b8a99e75ff0a5" + }, + { + "file": "shards/seed_tts/00101.tar", + "files": 328, + "samples": 328, + "raw_bytes": 52276832, + "archive_bytes": 52899840, + "sha256": "1934fbf8572edd6fb28373221e107f5089a503e62c05bde8d77c9017a4f0666c", + "source_signature": "4e6a60d2d44b5eced59c8d39d6a274d1131b07828dafa3cf71b54a256508e1a8", + "index_sha256": "9b3a3f5eafaed3ac40bc7fd83563f96a4863fe1447c66470e28dfdb24622c3bf" + }, + { + "file": "shards/seed_tts/00102.tar", + "files": 322, + "samples": 322, + "raw_bytes": 50387288, + "archive_bytes": 50995200, + "sha256": "09bcc199b0bd579b7a92240d2c8ef10a6a8c9b60c3ce0b15548dc42e4ef3e107", + "source_signature": "bd83e06e87606624a7a9146265a2849e5000475bdfe6760e2d6df08e04779f5f", + "index_sha256": "9fb7fc5901fff4df72ed54ca237f55ec3adfebe1d19c91ca9637fbacd831d06a" + }, + { + "file": "shards/seed_tts/00103.tar", + "files": 317, + "samples": 317, + "raw_bytes": 48263548, + "archive_bytes": 48865280, + "sha256": "5f77b7f72523a1e9a13f8714a16ba8080dc0352b9c148d8d697a0dd78f2c3093", + "source_signature": "1723a9aaaaa4d2e0fafd2a54cf1584fdaaaf7889685105fdc76e99bb8afe2e44", + "index_sha256": "46f269c0cc7cd472cba765e7bad6d019f755317d03d32ccb4e42be6729324479" + }, + { + "file": "shards/seed_tts/00104.tar", + "files": 350, + "samples": 350, + "raw_bytes": 52707880, + "archive_bytes": 53370880, + "sha256": "caffaaa3fd5e2452801208bdc088a149a11b38e6303f6f5dcdf4d5bad081a18f", + "source_signature": "409a39a6d6c5a9539c2e9e52735a46a5a81cb5e31159e3ff256498d8863b00f0", + "index_sha256": "4a7312591d03bae8827e611f481319b2e0ae4d44e16dddb1760c404cc057bc10" + }, + { + "file": "shards/seed_tts/00105.tar", + "files": 312, + "samples": 312, + "raw_bytes": 50932128, + "archive_bytes": 51517440, + "sha256": "40d2aee06909709452424b3b31814492d5f24daf8dd5f61123559af1633b8035", + "source_signature": "30f44bc55c54ea82be6eb366b8922cb54214708c3107623bfd060391992c3330", + "index_sha256": "52a4bbae28ecf4fa7d3c84af4264feedfc8f5aff6e2a4dd4949d891b3c871275" + }, + { + "file": "shards/seed_tts/00106.tar", + "files": 314, + "samples": 314, + "raw_bytes": 51423736, + "archive_bytes": 52019200, + "sha256": "5a4272f4f99b42de1b285fc85d817db7273793eb3e611a8c0ad17deeaaa097eb", + "source_signature": "9eb8151a569db5f7a501a1024a07579ff01f8459d95d2bd38ac2cb411a77f205", + "index_sha256": "94e3d3f80acb184e0861730137f463510b9b135022f81bf6ca39ed839d426304" + }, + { + "file": "shards/seed_tts/00107.tar", + "files": 342, + "samples": 342, + "raw_bytes": 53506248, + "archive_bytes": 54149120, + "sha256": "c11a32e3b61082b744597849efa94501005dbcc18ecc32ecbd9203ae181d8a32", + "source_signature": "ff4af221a6dfb5018dee41d6dc4c4fafe0104fb21f7db0591bbdb9ca8ada190d", + "index_sha256": "91289f0482977af2f4c787be8fe4b113e0aec10acb21a00a912ea981ded9b719" + }, + { + "file": "shards/seed_tts/00108.tar", + "files": 321, + "samples": 321, + "raw_bytes": 52253484, + "archive_bytes": 52858880, + "sha256": "bcf22acf9cd623ce271b561a250a51ee8ec8043b464617361a01f76491bc76cf", + "source_signature": "7fffe21240db85ee8e6f347e2b2cfde5aeb068a2e3f241243fe9acd222ffde18", + "index_sha256": "5b2a53116b2cc81590ffca5604caf06ef0401d6753083585ab5e1cce084f6b4e" + }, + { + "file": "shards/seed_tts/00109.tar", + "files": 346, + "samples": 346, + "raw_bytes": 54316664, + "archive_bytes": 54968320, + "sha256": "661c0cc86be3a5f9fee54d97af660bff3835b20d7dff8229efb7f3b1467588d2", + "source_signature": "2b30ca31a65bc2f4fd35fb87c7c679d120882acfe68d564c00c36ed02c06e308", + "index_sha256": "29a678e383753786eb8bf2899531b4a2ed752e0b97142646c92ca714a7894f13" + }, + { + "file": "shards/seed_tts/00110.tar", + "files": 331, + "samples": 331, + "raw_bytes": 53613284, + "archive_bytes": 54241280, + "sha256": "17e35fdf88848cf067f0b7d678dba8aedea7e6a1c71a29128f2216c647407507", + "source_signature": "20ba3fa001f372c5bdadd6191ae40699c08620cb63e493a3305af9c54920a7da", + "index_sha256": "d131e038e305f4a5230f87e88477de4233b7db319f434b774ad0d4205cabd014" + }, + { + "file": "shards/seed_tts/00111.tar", + "files": 332, + "samples": 332, + "raw_bytes": 51413008, + "archive_bytes": 52049920, + "sha256": "95d4bb4a06707049b9740f805a7b13b60ac72e572b10a8d489bec40994a2e612", + "source_signature": "8ea5e9bfdf544da33ad97a9bd090533ec19ec565e1502a688e83d62abc743b6b", + "index_sha256": "4809d34564c54756ecacb91d4517a7dc5f6f55c2a6fac9e834aab305411436bb" + }, + { + "file": "shards/seed_tts/00112.tar", + "files": 344, + "samples": 344, + "raw_bytes": 56286496, + "archive_bytes": 56934400, + "sha256": "05d7ae03abb5016bc9e68e8d555697f0b04e635d05463de79c3e34023780e5c9", + "source_signature": "b8a117698cfe37c5fbc3e724996493daaa1b885c86ec276194f7dab19806623f", + "index_sha256": "0b8091ae92f9892524ae8204617dd6df31abb91984e0948f0f5b060d17f92730" + }, + { + "file": "shards/seed_tts/00113.tar", + "files": 311, + "samples": 311, + "raw_bytes": 47917684, + "archive_bytes": 48506880, + "sha256": "f54834b11ba4263cb5c0529b48595aa3f6d2f71137e0df3c6b2151387bbea594", + "source_signature": "751abd28dcd169ccde3fcafccd332cfa32697e05cf24b2aeacf4ca68291d4e89", + "index_sha256": "a388e06d83743304ffe1af3e355627f47928f8a86f6be63c46f4914a351f90d5" + }, + { + "file": "shards/seed_tts/00114.tar", + "files": 317, + "samples": 317, + "raw_bytes": 49284988, + "archive_bytes": 49879040, + "sha256": "aee4562e26d24a4eecbd0bf28e264c8f9546a93b71a69a4bfe94ccb2be840521", + "source_signature": "986e4c18c73d519d9fd8c9cf1428521596602cf13b71b0117645c14073c4dfcc", + "index_sha256": "5e2b58b6d3488d7dc430952ecd51e7f9ae09035a78cfd27c6eeedcf0291a2365" + }, + { + "file": "shards/seed_tts/00115.tar", + "files": 326, + "samples": 326, + "raw_bytes": 51332104, + "archive_bytes": 51947520, + "sha256": "a451683b4a61eb3c72d4c7159c2241e8d44bb9440abe50fe02db4c24de9aa950", + "source_signature": "e0113bb008890c7bb3a84bbbed51741a5780b52200c1f4df848735cad1300b47", + "index_sha256": "654018b69885db8b3a50ab24a4ccd2ef73c6f2d0a238a61b0c1eb8eed4bf8ee1" + }, + { + "file": "shards/seed_tts/00116.tar", + "files": 335, + "samples": 335, + "raw_bytes": 50391700, + "archive_bytes": 51025920, + "sha256": "b7a27fcfa54cc14bfb8ed9cfeddbd9eda88d6032adb0479db06b5df27cead7b9", + "source_signature": "fea30705b0c2c58cca530c586d1024df1f0ef70291b36e633a84b8489d71cdb1", + "index_sha256": "d6b1ac58ff94b4dd322679db79d75a739ae8a8c8d8b982a51adb9ff4a9c60035" + }, + { + "file": "shards/seed_tts/00117.tar", + "files": 307, + "samples": 307, + "raw_bytes": 49342148, + "archive_bytes": 49930240, + "sha256": "189a8e62e1950bc76004f6970dbbb38659078c9473a29a5aa8f4cba6b1280987", + "source_signature": "36ff1d9a206c22b5c2c9e51ceab439c3ec8a11698b6c1c08f06fe58fdc25a3a1", + "index_sha256": "d09ccf62ce450009661d1aeeb23e1816c2811b9f847cadb875d292a8055a7a6a" + }, + { + "file": "shards/seed_tts/00118.tar", + "files": 336, + "samples": 336, + "raw_bytes": 53505984, + "archive_bytes": 54138880, + "sha256": "67ccfce28084682261652e73445116befcf8436420bd899f64e32e36d863b05a", + "source_signature": "e842f5098fe098c4ac488d96c77c0e0c4ed198043d18beb268274744c6892a5c", + "index_sha256": "652748920215fd6b276b097ecf06c95dcf0f3293ab100e1e68a95db921e586ba" + }, + { + "file": "shards/seed_tts/00119.tar", + "files": 334, + "samples": 334, + "raw_bytes": 51992936, + "archive_bytes": 52623360, + "sha256": "fb9c3b896b837c764868bd2145e3ae4bdee148a5473286dac5595cbf63283f7c", + "source_signature": "a6d913709c4ab793cc59acdc1ef467a3b87ec1e78459a908b1d2edf9e9b0835b", + "index_sha256": "01bce24668e9d65642ac3b1850d8e7961345864059b3b552b5a06c0e860a7b49" + }, + { + "file": "shards/seed_tts/00120.tar", + "files": 306, + "samples": 306, + "raw_bytes": 48743064, + "archive_bytes": 49326080, + "sha256": "497b12786cd502dbdad32433f30be0c7f3bb97fa929e55b24526d3ea3f2242ca", + "source_signature": "d7eb57e7e8273982d58f7d7dae78d44daf3402aac7eaa5a3ff6ff205a7d2f864", + "index_sha256": "37179bbbaf353e242e7e1560d1f06a26c363a22ba4a034693f00bae2a17b435f" + }, + { + "file": "shards/seed_tts/00121.tar", + "files": 327, + "samples": 327, + "raw_bytes": 51531828, + "archive_bytes": 52152320, + "sha256": "ad8c005fe6c5f0c188af25efd6298fa7b1b20d21a54c9fd66566acfddaba8927", + "source_signature": "82a885cbecd89eabf34db4df645fd986b8c3cf1020c172cb070d13c2b4edcb3c", + "index_sha256": "6502eb8428620390df8dba97757e366b54469f7dec63146d5a7b943b10ab8b92" + }, + { + "file": "shards/seed_tts/00122.tar", + "files": 340, + "samples": 340, + "raw_bytes": 54308720, + "archive_bytes": 54958080, + "sha256": "b7368164f305ab3d69bc98b911644a2542aa539199a75428064c3a076b6cc9ca", + "source_signature": "77d175909f818ce895fa05e4f7bc3af97286132f82b1ad79ce83705e89b1a908", + "index_sha256": "4e85660a6c3f93d46aa390f48722a3ab35ba2d5177b014d8ae4c7a70aa1b88be" + }, + { + "file": "shards/seed_tts/00123.tar", + "files": 314, + "samples": 314, + "raw_bytes": 50578936, + "archive_bytes": 51169280, + "sha256": "9fe8d886250c5ce599aac286498140c89a90f81ab0cdfee64a73dc550e340ea2", + "source_signature": "9451c33fb8f7caa14f6d3410fa32f00d63e9316a9b01bad5e5f561767151e9aa", + "index_sha256": "93ae21127778c167d6d9a673dd0aa8c38d7a93c399f98551f1a0335da71abf65" + }, + { + "file": "shards/seed_tts/00124.tar", + "files": 323, + "samples": 323, + "raw_bytes": 51439492, + "archive_bytes": 52060160, + "sha256": "614e1057c643aef914434950aabeb2d06c5b06827e88284fc3718cf1dafc44b8", + "source_signature": "370ed5ab937e42b9a5553521e9d9dc88047345d1540f5d5cd72ea7b1978c6b09", + "index_sha256": "5a7d217cff3319efa076914eba6f0555ada23a5be8050e0127bef743f6298f67" + }, + { + "file": "shards/seed_tts/00125.tar", + "files": 337, + "samples": 337, + "raw_bytes": 54554348, + "archive_bytes": 55193600, + "sha256": "a92539bba0a22b5cde0ee72181b4dad030eb4aae1402460f078dd29c756ee6be", + "source_signature": "92b66bf762752c0c0818083e17b041e9b72fef63dd91b768a704a4d2780769c3", + "index_sha256": "8a107bfc35c2dd5ef96979d0c9e5fe0f73c83231838a1b7583ebfda33d98d1e2" + }, + { + "file": "shards/seed_tts/00126.tar", + "files": 295, + "samples": 295, + "raw_bytes": 46788020, + "archive_bytes": 47349760, + "sha256": "7bf8cf95dbffe74087354cde8c860f758c1bd4427cf23cde942d6f67124f4cea", + "source_signature": "f20c1303e1a0734acb21c4d21020899347bc51928f81e268e662bbdbcb776777", + "index_sha256": "1c5e4375684aab8aff148768d9c135a3246324c8eb36950a4008528bf3026d49" + }, + { + "file": "shards/seed_tts/00127.tar", + "files": 328, + "samples": 328, + "raw_bytes": 49085792, + "archive_bytes": 49704960, + "sha256": "17761d5f32d19285e74703d59bb3df1ecfeaae235cb069a742f2db90fbf7b067", + "source_signature": "52ca57bc32f3b959bfd0301701155af336d0ff741fc2056abc780e2559428c4d", + "index_sha256": "cc3db6eb28547f30fa3b3f56eff3f96ffa88003cf0c2db9924b0e8a866921f94" + }, + { + "file": "shards/seed_smplx/00000.tar", + "files": 1158, + "samples": 1158, + "raw_bytes": 162217018, + "archive_bytes": 164290560, + "sha256": "c77225960a99253727a5ddd7e8f107c20193a97c869efe13cac988d0754f0049", + "source_signature": "1fa50a481c6f8168db0476eab40a4749423c8e0f17c61413a1f714eecac4291d", + "index_sha256": "f0f03fb9116714ee73c78e0841a9ac7b56acff3615a60db2da7338cb3516dcd7" + }, + { + "file": "shards/seed_smplx/00001.tar", + "files": 1072, + "samples": 1072, + "raw_bytes": 148316621, + "archive_bytes": 150241280, + "sha256": "960665290e76038d2332fbe5c51791a5afe3ffa6fede75955fb5e0cd51fef310", + "source_signature": "8a803b30f17d43083bcceb8e3c371ee336a7f25dd6a87198ab0796ccd1be8c9f", + "index_sha256": "6dda913fc288ef3e0c41b544d68eac244ee9f72d06d9859515da8f6f66f050dd" + }, + { + "file": "shards/seed_smplx/00002.tar", + "files": 1140, + "samples": 1140, + "raw_bytes": 165530203, + "archive_bytes": 167567360, + "sha256": "f17d5e7fc901d933bef30c20d2a471287a0fc5f72c378af12e4f20632634d684", + "source_signature": "f22188ad39016e216da03db8783da8070de3ce8912eb2bc375d2e38a87ddb8d1", + "index_sha256": "f4210a164f9ece91179b70ac37641d1cb7884204d2940c0d9ad119e81d98243d" + }, + { + "file": "shards/seed_smplx/00003.tar", + "files": 1153, + "samples": 1153, + "raw_bytes": 163608068, + "archive_bytes": 165683200, + "sha256": "6ffb6166b1b645d56978641e8c47769aaa700d0402503a9cd71bdee61a67beb5", + "source_signature": "7d0045d11086a98327f973037fdfb577e576427ffac0a50cdbd33cdd9191fa52", + "index_sha256": "74449286de26ff06f7868c0019e29c45583f016edca1213305991584974a851e" + }, + { + "file": "shards/seed_smplx/00004.tar", + "files": 1107, + "samples": 1107, + "raw_bytes": 160933371, + "archive_bytes": 162918400, + "sha256": "4656b1450fe8d88be36b42c0ce303e5747090769fb67ee190fb7dab7b07b62e7", + "source_signature": "c71483051648916164925513e2382fccd75151bad21c12a49aa943066e14c02a", + "index_sha256": "1e7954664cb5f8eb21daa9a56193923c609e1db393a54fb750dd26470d96db5c" + }, + { + "file": "shards/seed_smplx/00005.tar", + "files": 1112, + "samples": 1112, + "raw_bytes": 153644940, + "archive_bytes": 155648000, + "sha256": "df665aa23f1c93fd0b901bff5168ab410e638fe440066620ab151aff18dbfebf", + "source_signature": "88761d516b6e3f6e7afc1a2559fab28df44b123c2490dc4a80c5a6e760067a8e", + "index_sha256": "06ba132141ec09615e3914a574ba3fc86ff5c47cd2f78ec0e2acaa38adf8b36a" + }, + { + "file": "shards/seed_smplx/00006.tar", + "files": 1064, + "samples": 1064, + "raw_bytes": 148612332, + "archive_bytes": 150528000, + "sha256": "c48246cdbf9dfaa2519dd96b1dd6fde6314261da181425ccc60a6145fb6bdfb4", + "source_signature": "e46dbe368864c1e7555d18b41838b8f111494d6f846bfd83b2f0959b340eb91e", + "index_sha256": "8308e383437bccfd998140c43a3d085f509dff8681a1798755201a6e6ef73913" + }, + { + "file": "shards/seed_smplx/00007.tar", + "files": 1056, + "samples": 1056, + "raw_bytes": 169847045, + "archive_bytes": 171745280, + "sha256": "2ed728c78fd0ed1a2e7af4a11fcd45c1d96504c54d8ba96ad90dd23a01cf4902", + "source_signature": "a101307b67c322c22a139af8110fd10678ce1f039033d0e667c3a6d56c226f7d", + "index_sha256": "679568674e68a6188e8d9370bed89bdd177288fa732548f8d27a106115129681" + }, + { + "file": "shards/seed_smplx/00008.tar", + "files": 1130, + "samples": 1130, + "raw_bytes": 159186179, + "archive_bytes": 161208320, + "sha256": "fca0c18e9a0e898aca139baebfb2fdc72d7a87ba1da57be400beeb75adeec852", + "source_signature": "b889cabae326843cf80eee27204c10301a151449c3915c3e8bee84c52a999c8b", + "index_sha256": "965ccd310dd1f2e950a5f5a44ea1b5016bb2a8f45b0b352a2ba7508dd62027af" + }, + { + "file": "shards/seed_smplx/00009.tar", + "files": 1116, + "samples": 1116, + "raw_bytes": 155356089, + "archive_bytes": 157368320, + "sha256": "33beb447513d524f1a48865304a2b0af6aae2a81b903c206b5922ceb487a5229", + "source_signature": "348db60385f7c5b93a9a6bd56ebe4e6501ec0be8f424c8e67df3aff2f7a87d74", + "index_sha256": "bd819b5f1266f2a99bb1ce1ab834cad5f067a28f00e1c2f8ffa391ec71bad140" + }, + { + "file": "shards/seed_smplx/00010.tar", + "files": 1093, + "samples": 1093, + "raw_bytes": 155108789, + "archive_bytes": 157071360, + "sha256": "82597e5016c3ce7c533f7669ee2c77f3f7a01e71c7a6bdf3651f0e0c364f231f", + "source_signature": "e654193fddafaaa51421e227e832717fed87efc2e3e47d39e0b39f33b4886269", + "index_sha256": "4677f1c43d1cd87609173af3c945328322c4b5a56232562898699194f2404b82" + }, + { + "file": "shards/seed_smplx/00011.tar", + "files": 1104, + "samples": 1104, + "raw_bytes": 156949478, + "archive_bytes": 158935040, + "sha256": "09e566d83072ff391c169306f5495095ed4e4bcf24140e02e830a6f1ac09ba3f", + "source_signature": "85ad66bdabd671dbfff8c1dfbbdb772bc535d7659783a7a3f97c47c160443aa3", + "index_sha256": "71d67e65aa920970ffba9bf5a8d03d65fba29cc8c43e76bd757c205a06d980b9" + }, + { + "file": "shards/seed_smplx/00012.tar", + "files": 1127, + "samples": 1127, + "raw_bytes": 162399249, + "archive_bytes": 164433920, + "sha256": "d170c787805c095addf204a88270e5189c02b17aa1be971ec291a4397cba3f40", + "source_signature": "06fc3d2be0295e668bf87b8c8f7fad6c81ee94d946b24dbfe9ef314c98c29aec", + "index_sha256": "667e51548d781ca951a1163e5a9089b2f7e3e835c2748cfaa7bc16b9ac568135" + }, + { + "file": "shards/seed_smplx/00013.tar", + "files": 1106, + "samples": 1106, + "raw_bytes": 161859827, + "archive_bytes": 163850240, + "sha256": "3737a7912ddee59210f17da6e8cf71f131c036cad41f83166de29b3668ae281b", + "source_signature": "05032ba2722c3c43dd08f1f94d39ec2d32663b5b2e8acda9a60fef9ee66fdba6", + "index_sha256": "a29be76f71739bdb561527c0a907cc6f03966c5536868afba8c04bd6bfe92e4b" + }, + { + "file": "shards/seed_smplx/00014.tar", + "files": 1143, + "samples": 1143, + "raw_bytes": 169108008, + "archive_bytes": 171161600, + "sha256": "d60e45433d3fc0205678a2b89d58db415f69137c47ed9cf240e1a4749048f95a", + "source_signature": "063202dde3771a338cdfe8601e97397571aa3f9e09ca70d4c2cd8e30733a99c5", + "index_sha256": "9315d25acdac6f7ab4100f17e3fd6ce9d65be9bbbdef61891320e9cd347e7972" + }, + { + "file": "shards/seed_smplx/00015.tar", + "files": 1080, + "samples": 1080, + "raw_bytes": 148609485, + "archive_bytes": 150548480, + "sha256": "d9c736c22c2808f20bf33b7d436da4b23dd7584ee2cb0b2a7d468fbd0853c989", + "source_signature": "a74be8e5935b9d98c515208d4d3f689959aca33acb51c6ea04e882c683bb10f8", + "index_sha256": "117080b182ad17cf950b31bc49a0c7e02f01a08561d025f9d2f0482b1eb95d2a" + }, + { + "file": "shards/seed_smplx/00016.tar", + "files": 1107, + "samples": 1107, + "raw_bytes": 157764481, + "archive_bytes": 159754240, + "sha256": "0c37436f68827291643ff722738608f0929b6b491e8c43b32dae48ba139868bf", + "source_signature": "b4a452cd0cf46df188a42df9381c3cc681153605fa471310ee6465bcba4048d3", + "index_sha256": "870d0328d7f9fb176e9b38e256c2caed73aeb0c0f776e58570dcb74b84211baf" + }, + { + "file": "shards/seed_smplx/00017.tar", + "files": 1091, + "samples": 1091, + "raw_bytes": 155246863, + "archive_bytes": 157214720, + "sha256": "bc95fe2f4cb88b1edd3a1d80c75a41fe0afad7f6054df2d7d7d80ea1b52b522a", + "source_signature": "57b8c6342ca3a1ebfc0945269ade7f152c7a534944cba9e44876901ea020abac", + "index_sha256": "9fb49a3c76708c400fc718897866db82b0dd130d9f6d49642f8381d96070cd9e" + }, + { + "file": "shards/seed_smplx/00018.tar", + "files": 1060, + "samples": 1060, + "raw_bytes": 159959292, + "archive_bytes": 161863680, + "sha256": "ed37807c0fec627f8259120357808fa9b2c0c91ca8994c29f3865aa5e440e79c", + "source_signature": "09f2b88dff7d89c0e0ffcc13e34c3f98da233c868c3ee8473f4c4dd0ab3b80cd", + "index_sha256": "d61521893a743afd93aba792dfd42109ead6558b541c5231ca2c7fa58c357595" + }, + { + "file": "shards/seed_smplx/00019.tar", + "files": 1119, + "samples": 1119, + "raw_bytes": 161972428, + "archive_bytes": 163983360, + "sha256": "f38813b8a5b3236b4de71942777bbd7112f9d0b37cf93a0205e71a8115f49c17", + "source_signature": "4db94c4ee0fbb98024f9f53875f8cf35859f7843dfcceff06a273103ae9f7738", + "index_sha256": "617d93fd87684f4a996575af047cec02a184eaa9a68bd6c51770bdf2aa21c793" + }, + { + "file": "shards/seed_smplx/00020.tar", + "files": 1085, + "samples": 1085, + "raw_bytes": 157105684, + "archive_bytes": 159047680, + "sha256": "ce2f61e2f62575e4961e59b23de82c1bd05a3563f7b4b090f510e3636e543ad3", + "source_signature": "a34bfad3859aebd09b5d9dc1d798f8d359b8103dd83636507c5a555b3f2e951b", + "index_sha256": "46087674e0ff92b157fc3a873e03612f6f4b9689daf64b9b5844c50e1c6bb98b" + }, + { + "file": "shards/seed_smplx/00021.tar", + "files": 1132, + "samples": 1132, + "raw_bytes": 162171042, + "archive_bytes": 164198400, + "sha256": "3def28eb4b3aae483f3f4905436c2c0dd4dcf5fc64fce722953fd375aa3b7582", + "source_signature": "d271e4491629b0204f06b8d7c7cb33ed547fa717567905f3a63ecc65300c3005", + "index_sha256": "12f26d24012eab463df93f7818d661027ef752eb6334c361296249ab12d7da5f" + }, + { + "file": "shards/seed_smplx/00022.tar", + "files": 1138, + "samples": 1138, + "raw_bytes": 170025612, + "archive_bytes": 172072960, + "sha256": "0050b53039d71629e4379ca2cf24e5293f412a117f45a0ae857521afc5e7d740", + "source_signature": "d2c1b4b77a22203f143cb18108bd3718764c7644d8a98c62dc1b18356d24b0ad", + "index_sha256": "7343f8c95197396450806a3d7868d5ce413a8338428156972ec9e728a9f3b495" + }, + { + "file": "shards/seed_smplx/00023.tar", + "files": 1131, + "samples": 1131, + "raw_bytes": 159730962, + "archive_bytes": 161751040, + "sha256": "7fcf9bb96cb4cd49774be2d88e2c407b23e26faddb225147d308a68cee24a361", + "source_signature": "25cbff9c71960475a4cc1dd0846d054a545accffac48f87d9f47c30a18efee57", + "index_sha256": "a9d6b55ee1e86b88ba11a09ec3cb643e612af22fc49a121f62e333cf0959bf4d" + }, + { + "file": "shards/seed_smplx/00024.tar", + "files": 1060, + "samples": 1060, + "raw_bytes": 157031816, + "archive_bytes": 158924800, + "sha256": "0e92301b73128a8cfddf0ee7fe38f426dffe569521af6e792ea02ae06154d4b9", + "source_signature": "6b06662a0508fd8ddf61b7ba834b6ed00c1f702a3575f897d9b57cdc818b9143", + "index_sha256": "be1de944e90c386c4f781cc38fcf7cfb7d3577616cf361c4e8d2e3860543c6d1" + }, + { + "file": "shards/seed_smplx/00025.tar", + "files": 1084, + "samples": 1084, + "raw_bytes": 152189554, + "archive_bytes": 154142720, + "sha256": "65e4d5933ced1d2fdf303329e497deb8cf0c4a6745f4bfc6d32f52d373747bfa", + "source_signature": "9be7ff869f331b6a1916da16bd933db9dd6e52d450887306000aeec3bfc9395a", + "index_sha256": "bab3ca95a9aec8df5720c0b7aed6d61524b64c54484a9625d39f5ddf1c073b6e" + }, + { + "file": "shards/seed_smplx/00026.tar", + "files": 1066, + "samples": 1066, + "raw_bytes": 157450958, + "archive_bytes": 159375360, + "sha256": "0de868ecad5293e0e4300baf7c53265fe0fa2071ac0abfe31519a56931842118", + "source_signature": "863a97b4877190e3d338ccc634e158829213961eb9d7040914ed22dc7c53ca1c", + "index_sha256": "539388c860b339ab98739079a2d64234229c12cc644008a2dc2df7a26bd5b100" + }, + { + "file": "shards/seed_smplx/00027.tar", + "files": 1093, + "samples": 1093, + "raw_bytes": 158199809, + "archive_bytes": 160174080, + "sha256": "fabe11361ca368ec0f9c211561cf574adad08c8e23e0b16d431844d402f541f8", + "source_signature": "5a7572400682a773eefb6c7a077979658f759e18950ecf14577d56a200afd650", + "index_sha256": "475f4c12a360db553d581dc8e5a45b9a9bf7ba99b799153100917fdc209703f2" + }, + { + "file": "shards/seed_smplx/00028.tar", + "files": 1060, + "samples": 1060, + "raw_bytes": 145208592, + "archive_bytes": 147107840, + "sha256": "472ca80c586fe59df26c1f2b3f836143fbdbac90507dcd81205da30b0c963b12", + "source_signature": "b76887c9e7d4976b94b3958f3d8fa12d60ac7915ebbc087784c66a9afb0c367a", + "index_sha256": "6f12f88399512497cfd18eef7da4156e631dcfa4f5f98de890749122b5edd1f4" + }, + { + "file": "shards/seed_smplx/00029.tar", + "files": 1077, + "samples": 1077, + "raw_bytes": 150963828, + "archive_bytes": 152893440, + "sha256": "82d940ab1e86a8aa92901ea85473a0a2bd7e0512a62b8c7673dc5719582cecc5", + "source_signature": "fe548dc96e2078a3700c1f61455a23d893298c45adcb913b72e422733488315b", + "index_sha256": "756ed399dde0d72e95054b3b06c6e6ee1062533d9bf4fe5f4519afa3b4637522" + }, + { + "file": "shards/seed_smplx/00030.tar", + "files": 1159, + "samples": 1159, + "raw_bytes": 169234645, + "archive_bytes": 171315200, + "sha256": "10740e1af2cbab75b085253e1c5e4b73090eaafccea745792c2309d1e04c00d3", + "source_signature": "3fa4f6584c3d322ffaf15f666da856689cb7ac85cc5945245a590d92fee44a71", + "index_sha256": "6a74ad0daa4cdd4b44d6a5c52db957e89eda7342f620602006489758d82590c7" + }, + { + "file": "shards/seed_smplx/00031.tar", + "files": 1100, + "samples": 1100, + "raw_bytes": 154651978, + "archive_bytes": 156631040, + "sha256": "cfb780400cefbf2f4f92ed5a7978bfaf58083d85b8487dc3e237147171a80bbc", + "source_signature": "6b879abbd1d1a3d10d796cdccee92e8ce80596c7fdd34b0c7f43535d7e363079", + "index_sha256": "0cbee7df0b39dd83535dcf7d69679970b96cda763b5919ad25ee5c3f90f48301" + }, + { + "file": "shards/seed_smplx/00032.tar", + "files": 1083, + "samples": 1083, + "raw_bytes": 156002004, + "archive_bytes": 157952000, + "sha256": "10c4f9c2efebe694ac19cbe99ee67a890088e63e46dcc49e2c9f3781b4c0e273", + "source_signature": "df3d080302efd4580a359c23ce64f2f7287aae4c6ecf00f7c958e8c550ba3927", + "index_sha256": "f5f6c9d9f17f59c6126db41190dc45723c5df9204434b92a43f524e4eb5c046a" + }, + { + "file": "shards/seed_smplx/00033.tar", + "files": 1112, + "samples": 1112, + "raw_bytes": 155876490, + "archive_bytes": 157870080, + "sha256": "5c0a1159d1a522ab2a0cd7be5434a35418d92ea73b459cc475db99962c44799c", + "source_signature": "ee44005e9a9cd32afc7f26ab4b17a39194e72dc49f3406569d83134257bb815d", + "index_sha256": "29cd3494c46733a245187503615fe75bd123ec07c3df85677b0840763b357fbf" + }, + { + "file": "shards/seed_smplx/00034.tar", + "files": 1018, + "samples": 1018, + "raw_bytes": 142895301, + "archive_bytes": 144732160, + "sha256": "d6b81f265f94657d75a20c3dca94fb4d2a9127cee0ac6cad24d332c1440da995", + "source_signature": "ef3562400aecebd1da1ed4819b34931b776325e327c742d3ca49816f0f193980", + "index_sha256": "103a3fd89306c90f7664e732c8a1164ebae9f4f43808f8de7aa773b26505b636" + }, + { + "file": "shards/seed_smplx/00035.tar", + "files": 1155, + "samples": 1155, + "raw_bytes": 165631524, + "archive_bytes": 167700480, + "sha256": "11e260f2b74c21e6403e7c666451ed0f4cf3ced3a73a0d62bed5693b524bb29c", + "source_signature": "fc9564ea218189c8de9093ac210334c487097e9b8133a6e4a1640df69a977112", + "index_sha256": "58d48fdbe95d068f0f9f70903aebec3f5d1b64ee018139f771e3db5108876129" + }, + { + "file": "shards/seed_smplx/00036.tar", + "files": 1160, + "samples": 1160, + "raw_bytes": 164011748, + "archive_bytes": 166092800, + "sha256": "a03520957193aaa5757761b270b96fdbbb77d7537e1a75efb8686f11a39a7a41", + "source_signature": "3dca1a8032e8386b2207ce9d4e19d42a6b2381e997e63c7caa25275b9f36753f", + "index_sha256": "af2ca3d81ac9bbed6b1e0eb60618b506b6b876c4b0fdbc1f382c41d2a27dcbf1" + }, + { + "file": "shards/seed_smplx/00037.tar", + "files": 1097, + "samples": 1097, + "raw_bytes": 162240052, + "archive_bytes": 164208640, + "sha256": "1967cf13849cc87c2481c3de38af8bfb8d8f40bc80cf4ab6b051e07a1e10b74c", + "source_signature": "1c8a8c3f2fa53c8b4b061da7f817b57936ddf44e1b68f8240c8a08aa42f2b9fe", + "index_sha256": "0e19f237d1bb48783302d6cfb1e782389db8e3ae1f79e8af3e3375c5aabc3c71" + }, + { + "file": "shards/seed_smplx/00038.tar", + "files": 1026, + "samples": 1026, + "raw_bytes": 145392959, + "archive_bytes": 147240960, + "sha256": "0028fae70cf71c15a14b72116ad55f62fd84750f5725bcfac4b384c6dd225100", + "source_signature": "e891876b388f62a7acd1b59af655d92759a0febae2636a72e22b1977ea70abca", + "index_sha256": "cd5867bfebe66cab062e9707a1c3782c0743d08c2c36686e58d0154b69f172e0" + }, + { + "file": "shards/seed_smplx/00039.tar", + "files": 1129, + "samples": 1129, + "raw_bytes": 161797361, + "archive_bytes": 163819520, + "sha256": "9ddfd341bebf2c0e29870f7939fd76ddec4583729355994686189f504ad412e9", + "source_signature": "d392b4316ff7e1fde23e2ce4ca0ff70cb8be2e32a616d8489782b714a240eb0f", + "index_sha256": "7021e41908a8695e123666842b4b29277b715255a2af7a7f0369899274369abd" + }, + { + "file": "shards/seed_smplx/00040.tar", + "files": 1110, + "samples": 1110, + "raw_bytes": 158733069, + "archive_bytes": 160737280, + "sha256": "85e3b584987c1ff99ac219d065e279d0f42ab9cd7f803df72214655c7fd07c8c", + "source_signature": "f54c95b8159b2b77af5313f096c038dcb819b763af00e3011bd4ffd33719d98f", + "index_sha256": "6799906798c403b0b3df6eaf4bc562fcdf7cdd88a3ae2cf415689cf4d48786d8" + }, + { + "file": "shards/seed_smplx/00041.tar", + "files": 1134, + "samples": 1134, + "raw_bytes": 164130937, + "archive_bytes": 166164480, + "sha256": "92483966167a025ae5dbd419e5518930c9a85c25fd72680200e08b932f64f5ea", + "source_signature": "4c99463ce4a3a9c34c528c0a131e9a9efa97e772b953d7d98e836e8e85acc8b2", + "index_sha256": "9dd7e878cb8b59676e44761f68a7430a5ab3de825004cfb8b1da86ba35ab0832" + }, + { + "file": "shards/seed_smplx/00042.tar", + "files": 1080, + "samples": 1080, + "raw_bytes": 155108751, + "archive_bytes": 157050880, + "sha256": "88e41948d7e4687ec8602042263fdb0c6f0562328847e94d9af83f8fa2c27e0d", + "source_signature": "063b8a91081c4b75a6c8edfd4453cd67cf0b006cdb9741f7f5412a0e032bb59d", + "index_sha256": "0503bb588df3bd34065d351b11a63e8eb3c4dd85e9839a44c6bb49d0fbd0698f" + }, + { + "file": "shards/seed_smplx/00043.tar", + "files": 1145, + "samples": 1145, + "raw_bytes": 160043199, + "archive_bytes": 162099200, + "sha256": "83ed7365600362b947e179ce6b2f0002a0f5a20398ed65a61e61331c3cc63b39", + "source_signature": "2a456607c414fe297030b5fea86780c549f99ac8ea7a51cb1fbaca2a1bf18e58", + "index_sha256": "312a6116de09904bdc2bc74f9218954a417bbdb980c9b5deb05db8633e897fe9" + }, + { + "file": "shards/seed_smplx/00044.tar", + "files": 1122, + "samples": 1122, + "raw_bytes": 161881137, + "archive_bytes": 163901440, + "sha256": "b21e34892ff735418aec1cc9d4197bc4026ec18f09b7b9ab23d9e969e59a7b4b", + "source_signature": "507006c99443dea74befb4996e908b93c0a6bb74859871a24a248d25847c3f79", + "index_sha256": "98ecedba40dd2ba26771499c4cee0cb51322516cad4bf69afa118a86985205b7" + }, + { + "file": "shards/seed_smplx/00045.tar", + "files": 1085, + "samples": 1085, + "raw_bytes": 153549528, + "archive_bytes": 155504640, + "sha256": "81ba40d03fe239a51d88d7335c3cbf059227b6b79dadf6cb80192b20e204195b", + "source_signature": "d9ce59363efa67d288313973155fadc25469e4d4546ad501be013ec5e3cb86cc", + "index_sha256": "574de8d31cc545b3b0ccba461ea66d9e02978a5cbcb1efedd2715a398e117b4e" + }, + { + "file": "shards/seed_smplx/00046.tar", + "files": 1135, + "samples": 1135, + "raw_bytes": 165721190, + "archive_bytes": 167761920, + "sha256": "4f9de85ae5aac1bc7a043e7b61beaeab0434ac31d0a2b2242151dca50f18d33a", + "source_signature": "93dc571f4a10325ce3df50f3da761917d03ceb22d97f1b2ea9609a91d83ef03c", + "index_sha256": "1ec0d485d062c2b2d5e880968f0eb8409fecc746cce712f5bd6f53b0d25cc563" + }, + { + "file": "shards/seed_smplx/00047.tar", + "files": 1102, + "samples": 1102, + "raw_bytes": 165128395, + "archive_bytes": 167116800, + "sha256": "026e8d87642ea5a89481f64d28eb6a07c66cc4c018664b7e5a2a48f3281f7ff9", + "source_signature": "410a3d920eb2ba09f1acb13762cfffdc27ae8e1f28fd1540aa078fc37f5c6d2a", + "index_sha256": "8e87cd469c064e8f3a87f2268e0d6e89b0c23189d6fca4d709c6e044c4b8302c" + }, + { + "file": "shards/seed_smplx/00048.tar", + "files": 1136, + "samples": 1136, + "raw_bytes": 161304817, + "archive_bytes": 163348480, + "sha256": "cf8a9058fb02c9ee3ecbe620a4968d8a617b422f66e9cfd9ece864fd4ee96a01", + "source_signature": "2d50661898df44895704932766736be9c007c02924746a9645b03f4ccf6cc30b", + "index_sha256": "98472d4db1b655b942fc977b633fb2e5e80781988279b2eac3353c810feae10e" + }, + { + "file": "shards/seed_smplx/00049.tar", + "files": 1113, + "samples": 1113, + "raw_bytes": 160001208, + "archive_bytes": 161996800, + "sha256": "708cdc87d3b99eafe57c73c2d3604848f482d7a2946efaf2668210462527b126", + "source_signature": "0f418f1647c36fa0d5b638a0ae1c2f66b9ef9c9b76bc37ab648a491b82aa08e2", + "index_sha256": "1b7b823bcdb8fd06554f50edc3ed09a67757a26ca8f1c35d8dd2f3f94be46104" + }, + { + "file": "shards/seed_smplx/00050.tar", + "files": 1113, + "samples": 1113, + "raw_bytes": 156452793, + "archive_bytes": 158453760, + "sha256": "e56f829fe85b7ede4c88cb2a4c7d4955fefaba31a7bf0d30ee288f57bbdfb978", + "source_signature": "bd8d0d32ee86dea3f3bbbeb00bda30b7ddcb112343f42f9d690c6cd95ea36a63", + "index_sha256": "785ecdc6ad9d416b3adfd0cb8fb49d992c5715dd35320e6446ec2a63f183f78c" + }, + { + "file": "shards/seed_smplx/00051.tar", + "files": 1130, + "samples": 1130, + "raw_bytes": 163207181, + "archive_bytes": 165242880, + "sha256": "85d1d8662f80b0eb46e5e56f49c20feefc44fd3c7d3dd259492888138ee20ac2", + "source_signature": "901d58e99b61f3ddf1c4c916c081be9e9a2257dd08e5195376ce1510be15fbe9", + "index_sha256": "cf51c31573628452d18e67988e51c182c821c0caa0252d1fd747687cf852497e" + }, + { + "file": "shards/seed_smplx/00052.tar", + "files": 1041, + "samples": 1041, + "raw_bytes": 146789789, + "archive_bytes": 148664320, + "sha256": "82440c6f49e7aaeb3c8ec6cead9e3045dd03ef6de5d214a8f9214fda5dcb5565", + "source_signature": "9cfca3fa28e00c6bd49abcdbe6b095c499a835f7a79d4e6eb8c3cb8eadf31e28", + "index_sha256": "2e909ada1ef62fb6595d18b5f700bb8696501af5762ec429c4744a60e07cb22f" + }, + { + "file": "shards/seed_smplx/00053.tar", + "files": 1140, + "samples": 1140, + "raw_bytes": 165238009, + "archive_bytes": 167290880, + "sha256": "f06bfe33fd661d251c2b29e2b671f9dd0aba941f0e930ef92da7a3f21dd2c256", + "source_signature": "9ba4725d64d934e8d35c92b236c79ad8b5aa4b5b625905f2810b24d8a0571706", + "index_sha256": "1dbe241437aa72e25eaeb1a226c5b14b5559c8e43734f598c20fb7b76f0e7a0c" + }, + { + "file": "shards/seed_smplx/00054.tar", + "files": 1150, + "samples": 1150, + "raw_bytes": 160011681, + "archive_bytes": 162078720, + "sha256": "5c90166e4a803747a4c7277299b34bb91a4f2e71af6aa766d518350ea0bd89da", + "source_signature": "774900f301fbeb3f2a0ca6301a30270c15ddbc2ae43cd2237c6d46c041d3fa07", + "index_sha256": "14d53e6b18941b406d9cf43db666b98b29a3dbf01270d85d44bc0330e76d550b" + }, + { + "file": "shards/seed_smplx/00055.tar", + "files": 1058, + "samples": 1058, + "raw_bytes": 144173325, + "archive_bytes": 146063360, + "sha256": "4b5004e73ea05115c2cbb3597827484dd647531f94b5e6400f30ae9c21b2c4ff", + "source_signature": "452a2477774fb37890a003241a5c2d2402a63d7565e148d2300bca8cf6b6a98d", + "index_sha256": "6d18001ce514dca29fc1b827a50dbe173c7a7e93a6fe870196af57525347f595" + }, + { + "file": "shards/seed_smplx/00056.tar", + "files": 1119, + "samples": 1119, + "raw_bytes": 170236168, + "archive_bytes": 172247040, + "sha256": "9c005b42393d3569affd484d86b8f9034e5401248ccf15f578453043f7880d75", + "source_signature": "fac54a609c30e2ee11246b099e94e4c0eec7632de6ceeedab7f99c45dd74eae0", + "index_sha256": "990e7a0b507b2fca2795097df586554ea2747f3f25b7d186a040621ee997b533" + }, + { + "file": "shards/seed_smplx/00057.tar", + "files": 1131, + "samples": 1131, + "raw_bytes": 165813611, + "archive_bytes": 167854080, + "sha256": "7a791a918b32a7f2ea30d77847a3449e7b8d90fb17014ad9fe4e06911f5ea6e8", + "source_signature": "82cdbdd051433d03f02cd9643380d7cbec064ff4b12ffe5bbace00eba73f2569", + "index_sha256": "f7a9e8acc51417db8f0483fc4fffacd5c24ba5b815c75a0f785935a83bc336f7" + }, + { + "file": "shards/seed_smplx/00058.tar", + "files": 1179, + "samples": 1179, + "raw_bytes": 172190440, + "archive_bytes": 174305280, + "sha256": "6e8c4ba6e4c4187dad42594e71194c6d734b49e40d92445329c6e90e5b264bcd", + "source_signature": "b3399c163ae8ecc35256f20230b668f0a0a74a3c5c35c273be846b1f4736a10c", + "index_sha256": "0d4bce1ec585099c32185ee175977aeacded62045443ca8b5eb2cfde1cd6a49e" + }, + { + "file": "shards/seed_smplx/00059.tar", + "files": 1112, + "samples": 1112, + "raw_bytes": 157292696, + "archive_bytes": 159283200, + "sha256": "3ebea086764145ae33bb48f6e8861e8ab083e05bf270afecf7bf85539e625e70", + "source_signature": "ba9c143b92b494199d7a8cc6abbf3cab06945ec0fdbf27292ba4f3e57909e04f", + "index_sha256": "24eec12a363da8efbb0a58cec95ec457364978ca606741a4acb1ca508320d380" + }, + { + "file": "shards/seed_smplx/00060.tar", + "files": 1090, + "samples": 1090, + "raw_bytes": 155475052, + "archive_bytes": 157429760, + "sha256": "beeab8f7ec52274f42e0b78df40b28dc96b8e9987ba4123ebfcf6d862e20437e", + "source_signature": "65f988527455048990605527b3818e0aad1c8f2046edcbcc50ec50255c0693ff", + "index_sha256": "9144456ab91aa820e90bf54b2d296972850aafd89247c72f6a78d0da0785496e" + }, + { + "file": "shards/seed_smplx/00061.tar", + "files": 1090, + "samples": 1090, + "raw_bytes": 152057452, + "archive_bytes": 154019840, + "sha256": "2fb05a24c70b2d0b5c4c48133702b703c445af7d21696e341acdbbb23ccac4bf", + "source_signature": "89dd8bfb53aa8e6bb8f89492efb541e8ae12942381600c3b92c108f12bc91992", + "index_sha256": "81f98e6588e8e0c9a5d9c52bf23019f80cc82971133867fcca68958a81ff33c9" + }, + { + "file": "shards/seed_smplx/00062.tar", + "files": 1189, + "samples": 1189, + "raw_bytes": 171506389, + "archive_bytes": 173639680, + "sha256": "9645e137db730ab098f096af1db2cce8a60a8f4da905cc05f250f5e65d694b1c", + "source_signature": "5102fa2d4023fd4fb59fe17fb0b8b8aac1edea1f437f2dfcb6d5b98004d5fee9", + "index_sha256": "09eb6913de835d9d23e27a716606efc9ff6ae4dc5e48dce3c22fc08452d17426" + }, + { + "file": "shards/seed_smplx/00063.tar", + "files": 1094, + "samples": 1094, + "raw_bytes": 153912118, + "archive_bytes": 155883520, + "sha256": "d8f320627e7269ae17197a758b722bda249325f3414fcc9777bc3b2a2c9a4b63", + "source_signature": "f7bb0898aa5115fcbf6ed4f957de5b69189466dbdfa59f5be24caa87fafe0e05", + "index_sha256": "9b5e86556f42373457bb96b34b54ab9bf40247b67d49bb10a491db94a677f626" + }, + { + "file": "shards/seed_smplx/00064.tar", + "files": 1117, + "samples": 1117, + "raw_bytes": 155042964, + "archive_bytes": 157050880, + "sha256": "43bb73a9c7ab85b2846ae44dc5fe42c222b922cbbbea31de4eace1fbad28e108", + "source_signature": "6adb7b8f1476239be9d0f8c28ad846d3d9ae5c620e962c47e80cbb83c7da2dd0", + "index_sha256": "70274730a45ab328054013c8c4d8dbc4d979108748daf0cfba90a6d7ecbe7799" + }, + { + "file": "shards/seed_smplx/00065.tar", + "files": 1120, + "samples": 1120, + "raw_bytes": 160761627, + "archive_bytes": 162775040, + "sha256": "35c579ccb9b575669f077e7b488c0b53960adcec60fe02bbf9ce2cd4b1c90c2d", + "source_signature": "23b2425b5d09b7bea5150e52c1794fbd18874bc82831eab91cebdd863d3d54a1", + "index_sha256": "2d26dd67b49eae4c6a24d7311e48e35ed994c5c4e5147c06391a650f645ae140" + }, + { + "file": "shards/seed_smplx/00066.tar", + "files": 1084, + "samples": 1084, + "raw_bytes": 156119871, + "archive_bytes": 158064640, + "sha256": "c64d585c47ad7f184219e777aeff69b5386487c72cd0b4656343226d8b4b6437", + "source_signature": "a1360c8fa4eb7542c90958937901f2a707b25b84cd70fe8ad49b13c9951a523f", + "index_sha256": "bf77b8d5a5d7a13a5b871602bf585e15ec53a3118560ae1b9e2cab4696408b41" + }, + { + "file": "shards/seed_smplx/00067.tar", + "files": 1052, + "samples": 1052, + "raw_bytes": 140191840, + "archive_bytes": 142080000, + "sha256": "79dcb210eed238622e7e07c1f900069d2c1c0a8d651499bb4ee0358f312eb294", + "source_signature": "5aadca7bfd9b049d08e200065b7173f559ab84b335ca11cee28ac432efc3f804", + "index_sha256": "f5c0a845d09a7dfb4c8c08a54c6dbf522df754cb2ebd3e89ea9f73fa55a04d2a" + }, + { + "file": "shards/seed_smplx/00068.tar", + "files": 1170, + "samples": 1170, + "raw_bytes": 176064241, + "archive_bytes": 178155520, + "sha256": "26ab96af081305870964e097c816ca5a9e8b673543fa5fd3d91a212bb44fef28", + "source_signature": "58366e0ee14ad00003d6cfef01afd1bd2aca515860b7fa412294497bb0c42431", + "index_sha256": "083711ac1d4313182ecccb56829f0deefb64c14fe06735bc4b980c9d0c67b2f8" + }, + { + "file": "shards/seed_smplx/00069.tar", + "files": 1086, + "samples": 1086, + "raw_bytes": 161247627, + "archive_bytes": 163205120, + "sha256": "affea53d462505153b58a66b4ced7cf23f075f454444405147e0beb5ae7f3e92", + "source_signature": "e11b9d950bfdb990cf78d6218247a7cbf15e1a89b7460140a4e69f8675443f28", + "index_sha256": "8c06ff9f3265c585346e744f65d468d09e664e173db7d58ee4f617ae2a672de5" + }, + { + "file": "shards/seed_smplx/00070.tar", + "files": 1112, + "samples": 1112, + "raw_bytes": 163474331, + "archive_bytes": 165478400, + "sha256": "dc07436d8a91687082585af00363ee79ce5f43339f5a018a88b4584736072dbe", + "source_signature": "c45f1560da2d313a702625d35b41c1f9d708fd02556fdb93eb9ba3bf36805cf7", + "index_sha256": "1fcbb0e12a6e9cb77fc0b292d9ca84c4fdb838cd544eb9a4ee392dbfaa892fc0" + }, + { + "file": "shards/seed_smplx/00071.tar", + "files": 1060, + "samples": 1060, + "raw_bytes": 147881998, + "archive_bytes": 149780480, + "sha256": "56b5ecd195c2d53849c09274d54acdecb1d1327761c73f5f9ed84281897ac34e", + "source_signature": "1f1488a7a149915785f478785b9bda93c7f83882e4bdb241f378a9b1d6aff587", + "index_sha256": "f0c717e32af784eb20397c8be4634e747a1ae28330e3c6d73d220c60541f1e4a" + }, + { + "file": "shards/seed_smplx/00072.tar", + "files": 1125, + "samples": 1125, + "raw_bytes": 166431600, + "archive_bytes": 168448000, + "sha256": "7e8885e9625f362f90d98366afea21f48bc97124cf02ac459525c10ea31c2f8b", + "source_signature": "fca43b26446b574c7bd073a815f11ab92e1077dcdf2cd0d7511429cab46336b6", + "index_sha256": "abd0fc065ddda82727c7fd72f156c3a229738a070b5a369deefad2d0923d021f" + }, + { + "file": "shards/seed_smplx/00073.tar", + "files": 1107, + "samples": 1107, + "raw_bytes": 162578363, + "archive_bytes": 164567040, + "sha256": "8bde5e66bd9434f6f7e86c62f706b553a73b6a09aa76cc78320a9455c5469fcd", + "source_signature": "008994a74b6396278a62e79f8c76d1f13c9563f44b139ee573d8006176816919", + "index_sha256": "0a80c21e640eba3da03febaedf344e5a7b2bc48d1bd04d683efd80c41ba88506" + }, + { + "file": "shards/seed_smplx/00074.tar", + "files": 1139, + "samples": 1139, + "raw_bytes": 165220321, + "archive_bytes": 167270400, + "sha256": "7bd167ed7793d0b4ca8953bd175ca1c542e8ee4511387b7f679a3794fcdf1d49", + "source_signature": "f2a23d33a50d72316d148b9e6e0007c07a50b73d6063506b6811080f185e9b51", + "index_sha256": "3343b70e5fd2e0c2a2f0f935b864de43915587518b73f02c564780d882d71da8" + }, + { + "file": "shards/seed_smplx/00075.tar", + "files": 1118, + "samples": 1118, + "raw_bytes": 155707934, + "archive_bytes": 157716480, + "sha256": "74ccb501400795638eac256e483a1303f6cea1f769a15e8d343fefaf06828867", + "source_signature": "963d237670979269eda570df8a7e885669068f4a0ec9f5dc73ad28d8e7a4a879", + "index_sha256": "09f2da9514d925495efce451762434f46bd8f57567a7610e6268b57b24e3f710" + }, + { + "file": "shards/seed_smplx/00076.tar", + "files": 1084, + "samples": 1084, + "raw_bytes": 150374195, + "archive_bytes": 152330240, + "sha256": "8a516b737aa571e9adc0c3fee766cbd96252b0dfa090545d2cd0c1ca7f8102a6", + "source_signature": "489eb51c5f47843e7baee1afe692eaf8ac3038e0a143b32a07533d1317649da9", + "index_sha256": "91fb1bc1dc9210af744b14bda62e8f7801f5a107ae8e113eac86c22a6c5b08c9" + }, + { + "file": "shards/seed_smplx/00077.tar", + "files": 1090, + "samples": 1090, + "raw_bytes": 161194922, + "archive_bytes": 163164160, + "sha256": "445822226c2dd04597de0c231fad4de6a279bc58efa2fdcd6aee48ef017431dd", + "source_signature": "e0c6237eb64281a06b4423e42235545051216ddd3a9cf882efcc844f4a600018", + "index_sha256": "d7c4e6fc25ae59c6dd7ecffb477667cd34b75c28177186da1ee29877f50aa075" + }, + { + "file": "shards/seed_smplx/00078.tar", + "files": 1139, + "samples": 1139, + "raw_bytes": 160252711, + "archive_bytes": 162304000, + "sha256": "52439b22771d2f2fe1546afaba22b34f7080aec3a935c1343c13bd8a04c92cc6", + "source_signature": "ff899cfc5ccf2253298566979033650da2185dddb90c59425dc09cd7bb99e9fb", + "index_sha256": "36f0b8fd12a824a44e7b0fc50bd8610eb4716ce0e7ee9635b3b374314890df79" + }, + { + "file": "shards/seed_smplx/00079.tar", + "files": 1081, + "samples": 1081, + "raw_bytes": 150099222, + "archive_bytes": 152033280, + "sha256": "a88eb8cb09c3dd3128dfacf8e29202af21abeed9ebb283974dc4921632b5e18b", + "source_signature": "de71a5dd1984a1b9318b3bc86d9d6454d4b65785102c059d7fcc908be6ce7463", + "index_sha256": "bf4774433afdb42e5798958bbaf434dbd6614777f8293a7aa5966e979edec558" + }, + { + "file": "shards/seed_smplx/00080.tar", + "files": 1103, + "samples": 1103, + "raw_bytes": 166490097, + "archive_bytes": 168478720, + "sha256": "22ff13ad98d3db2cfc79f3c306aa4f3dc915e2baeddfe76d76c57e1ecdf13bf9", + "source_signature": "15d6987efdbf7bb13343434a8b274e326d592a71a4f160c817c3cead79301b43", + "index_sha256": "bd5db1fee9c8c3aa82c1f55728a83731473afd6603e6af5219c6ee7554a4c6bb" + }, + { + "file": "shards/seed_smplx/00081.tar", + "files": 1069, + "samples": 1069, + "raw_bytes": 156807422, + "archive_bytes": 158740480, + "sha256": "1d7d8c2d75cdaf62d4e652deb3a840ae6187a8b29f5e9d3d1280afa5fc276f0f", + "source_signature": "8ca95f010e600f56d556c296bb21fd16e7b96219731295addfbdc85a54aa8fb2", + "index_sha256": "7f11d1adbfe029c8fe47a3e3be328f7c01cdde81a5922abf25c7706d493164cf" + }, + { + "file": "shards/seed_smplx/00082.tar", + "files": 1102, + "samples": 1102, + "raw_bytes": 152589917, + "archive_bytes": 154572800, + "sha256": "75d76380e58a10e5698dedea5f524868b18be5272e63db8665fde9dba62ce961", + "source_signature": "7e955a1b36d5645c8bbd0201c4e01dce82235fceca250a36b00f8fb4fdfadb99", + "index_sha256": "9944e6f55a1de611cc96e6eb854b9b3b273da10a647484242e6c2cb753b6c95f" + }, + { + "file": "shards/seed_smplx/00083.tar", + "files": 1067, + "samples": 1067, + "raw_bytes": 153406921, + "archive_bytes": 155320320, + "sha256": "9f1fde4275b33d94e0bd826aa7cdcd2a980f28b0786ae213b67a7b0a757f1d5d", + "source_signature": "74e44897271d8056c3860c4b6f6ef5c8db264144aa5aad47273ff523fe184fd5", + "index_sha256": "4e394838f644122ff472caba227c79e49d8534b072913e4194b0d88e01d50eb0" + }, + { + "file": "shards/seed_smplx/00084.tar", + "files": 1140, + "samples": 1140, + "raw_bytes": 167930591, + "archive_bytes": 169984000, + "sha256": "0bf9b2320f4caecf91c5d3774e936ef5cec363f80947516d82eda50c4f9800bc", + "source_signature": "df0c4b4a4278dfecec6dbeee04014697e55340f6c4e2d1de72e38ffa48b3e3fe", + "index_sha256": "77ca4966f96acaa281572fbc69ff693c8499487d1b5e4c63162bf7802e4e0344" + }, + { + "file": "shards/seed_smplx/00085.tar", + "files": 1157, + "samples": 1157, + "raw_bytes": 158891910, + "archive_bytes": 160972800, + "sha256": "cb383310fdcc234b9080685498c56ca2c510186ae0716ea81e97e11631fdf162", + "source_signature": "f9185f2dba565daedbbfff018244f2d609ea677d3f0196af1fac8c1afd47dec9", + "index_sha256": "2c4f4895a83d3acad120936d11c0a64a67aad2118a54b345e068d12c88b89e26" + }, + { + "file": "shards/seed_smplx/00086.tar", + "files": 1129, + "samples": 1129, + "raw_bytes": 158316865, + "archive_bytes": 160337920, + "sha256": "add1d09b7dd494b77ff1fc24c36892369c6955bdcc5dedb8d4255e9313156510", + "source_signature": "c0a74b7d7f780c54e0f9c59369fdc42987dfec0db6449ad4202bd3b341c01c49", + "index_sha256": "9ae7811b0cdaef88ddd85236debc28c2b9e5492a1e383c05e9257d5e8d7cdd06" + }, + { + "file": "shards/seed_smplx/00087.tar", + "files": 1029, + "samples": 1029, + "raw_bytes": 138009116, + "archive_bytes": 139857920, + "sha256": "c512036cae87dbfc26558bcd40009639e12ad2797826933c368030e8b42c1470", + "source_signature": "76069f3e98076322906becd4d289c920a5be682f49aff6a8b391c7aa19a42e9d", + "index_sha256": "bcf07fcb56d1288bdad652a198d83bca9cba39cf80187df9850d57e7f5c86bd9" + }, + { + "file": "shards/seed_smplx/00088.tar", + "files": 1116, + "samples": 1116, + "raw_bytes": 157766563, + "archive_bytes": 159764480, + "sha256": "7d8123f241eaa5f8ba3784f1fb119126e006fb8d2a014747c7c3f7e9215d98f0", + "source_signature": "a8246a7cffa80ec4da8766d615a386997c3d88d91bed8222365e300d0c65c770", + "index_sha256": "873faf2d1c75bef75f5625f5795ffa5f1b6691d85e91e7e76f5afb05c351e96c" + }, + { + "file": "shards/seed_smplx/00089.tar", + "files": 1112, + "samples": 1112, + "raw_bytes": 161883707, + "archive_bytes": 163870720, + "sha256": "3423b19b76baf6297c4e411fa952548213c021c13fd2106833e3caa3e1770b9c", + "source_signature": "60d753a86911d2967325b31bee99efc6782b9e07cfcfe2fd4f38104cedac43e4", + "index_sha256": "ede37c1877f77a04ce137ec73b6153bf840a315212fece563304da86b8d91acb" + }, + { + "file": "shards/seed_smplx/00090.tar", + "files": 1112, + "samples": 1112, + "raw_bytes": 153975139, + "archive_bytes": 155975680, + "sha256": "a30489f2fd1b8d3c3af87e1df341a5fdb623f3fe41e48b67c52537a9aa76f1c7", + "source_signature": "48706d305ef8f7d411fee7d85782c311f3ce2914f11715d5079ead87a8a61076", + "index_sha256": "57dddef0e0dc18fca4d89e66e6333338c0b0ae96804f87d8ee344fe9576a6c70" + }, + { + "file": "shards/seed_smplx/00091.tar", + "files": 1131, + "samples": 1131, + "raw_bytes": 173458652, + "archive_bytes": 175493120, + "sha256": "7ee8197f163e96a15958dbc963b545ace07c0496c43847f74838ef20fea2542b", + "source_signature": "f601728c63f363e009853cd632f90cc3917670289d10454a826cf8ac477deefa", + "index_sha256": "a281b9213fb194c00f3ed2cde9d44c167ce0e878f2464874e9c2ef52e9bbf3f5" + }, + { + "file": "shards/seed_smplx/00092.tar", + "files": 1088, + "samples": 1088, + "raw_bytes": 151510145, + "archive_bytes": 153466880, + "sha256": "de8defa5a90437c6302c5db4116ccc53acdb319819deb5b0ccaad087ef706aee", + "source_signature": "9d28105227f15229ff6d1daa99cca4809a05e42c43e791a7f90b28726358c6bd", + "index_sha256": "5a97a33cc46d22e2073b647aeaeee8c6ce1b559344ac9d1533220eb5c3585198" + }, + { + "file": "shards/seed_smplx/00093.tar", + "files": 1123, + "samples": 1123, + "raw_bytes": 152662816, + "archive_bytes": 154664960, + "sha256": "3bb4ce0c9d802fc3a053cc04b0d2b94752e146c4acc928485c0543890d8e912b", + "source_signature": "3edda319b0a581f61db3c047f507162002b247aaf4f3bf2c0ec8d54fd5c950c9", + "index_sha256": "95d62e2b80368c04216735b8618d471eafcf52cb13ad990769f24a01bf2056fb" + }, + { + "file": "shards/seed_smplx/00094.tar", + "files": 1089, + "samples": 1089, + "raw_bytes": 158716142, + "archive_bytes": 160665600, + "sha256": "796bf0b686988ce703cbd88d727a1f36239afcdc3efea713c6381bf9d03459d3", + "source_signature": "ad14affd897e38e422ecf6935ad936f6ed5a3d8835f0058d441d361b0dc4e6f0", + "index_sha256": "cfa972507890e731f9e71f56300cb887db99bf3b19d96fedf99c57e874313535" + }, + { + "file": "shards/seed_smplx/00095.tar", + "files": 1065, + "samples": 1065, + "raw_bytes": 148812058, + "archive_bytes": 150722560, + "sha256": "ec2f5ea4f70050877465aa739f7e6c55afb4682f37c54c3cd035022b2173f95d", + "source_signature": "0be47cefca60fff2ed84430d84ce373edc2fee9da58de05bf02a29204c7bc3ac", + "index_sha256": "c669efb68efa55fe3a57663517164b5cad82c2bdec50c78accb66f44e6814bfa" + }, + { + "file": "shards/seed_smplx/00096.tar", + "files": 1108, + "samples": 1108, + "raw_bytes": 167900885, + "archive_bytes": 169891840, + "sha256": "ff768278abdc0b4264213d373ecf48d53411b412383fcfdebaec2ac7a9b1e631", + "source_signature": "d4d83239bd14ec4c50ff1ac50893599c075b299a985e4677087048c211b06b65", + "index_sha256": "7ad6e80393659f00bd3ae44deca74d3cb17bcd8fa4f6bc5effb798d11e0b2e59" + }, + { + "file": "shards/seed_smplx/00097.tar", + "files": 1115, + "samples": 1115, + "raw_bytes": 159064425, + "archive_bytes": 161064960, + "sha256": "b30aba94f12abe0c33289b553b68cf9768bfbe2cf523dded564e13dcb676ea41", + "source_signature": "bfd725ec7a16a6b7f6417d564c9de6fbe63546a924164e9231ca9a96f3e33ecd", + "index_sha256": "f8eed96ed1f4644ca225dcdc28786dd79183b12b79a6e522c98ce6d60af9290b" + }, + { + "file": "shards/seed_smplx/00098.tar", + "files": 1188, + "samples": 1188, + "raw_bytes": 167446286, + "archive_bytes": 169584640, + "sha256": "0ebffb2371584be72009a17a2d2d862a0925e12a574b1b78a2005fe2ddfc7271", + "source_signature": "155fc173cbf5d70fdbbedc28aff94588f813beb8d6c04e0add5cb5dbd3b4619b", + "index_sha256": "122b5ae6c5166b5c7cf7e00c2157d074c806b63c41ef6ffb919eb2a676760bee" + }, + { + "file": "shards/seed_smplx/00099.tar", + "files": 1094, + "samples": 1094, + "raw_bytes": 149489618, + "archive_bytes": 151449600, + "sha256": "7ccab2fd5a83b2f27d03dc42b53c41265ff92c935cf1124bbc7636cf7092da5d", + "source_signature": "a04a95e43ff7633f8b14fe6dce575e87183eb6b3136d100378a1358a39f1a027", + "index_sha256": "a97a754726edfbee7ea6a901d767e47cb2759c3ee1157ce02ca3f41632db197c" + }, + { + "file": "shards/seed_smplx/00100.tar", + "files": 1134, + "samples": 1134, + "raw_bytes": 169402665, + "archive_bytes": 171427840, + "sha256": "80b3846c2d36f6d553046ed754c917c3a439d20cc03f96e70b625d356bfc1700", + "source_signature": "b1e44020439c16d74b82b84fa5e01f11ce2892cdb81ec6ab86bb018ebea026ef", + "index_sha256": "c19124a1557c62bd028d78303ffefb75c4546c4b555fa7b36dfc74858efa9518" + }, + { + "file": "shards/seed_smplx/00101.tar", + "files": 1099, + "samples": 1099, + "raw_bytes": 159755539, + "archive_bytes": 161730560, + "sha256": "44fbd0d33d6677ac32e7733d9fe85bab1a1721c49e0cb821217a1ec05d00d9c0", + "source_signature": "d5d2cfee3b48343766a72bf6b7ca6db543041a648f769ac8d603c03fa0ff9345", + "index_sha256": "326be9a56238e231b2102ffd20f600cf1d21a6788486e0588b5f93ddd13dd8ae" + }, + { + "file": "shards/seed_smplx/00102.tar", + "files": 1120, + "samples": 1120, + "raw_bytes": 164181470, + "archive_bytes": 166195200, + "sha256": "d2bde9bcf2be9dec45acb0b9832fce6ce0cd110a404d9163dbc133eeca4b163d", + "source_signature": "7e65bf0ca0286ec2ec9b4f39bf3a5e9ecfd5d5efe6816e7796761b7f8aa85556", + "index_sha256": "8665f0e3621133f9ffd9840be4127a354820c6eb53216f9ea9018d078f17335e" + }, + { + "file": "shards/seed_smplx/00103.tar", + "files": 1105, + "samples": 1105, + "raw_bytes": 158015319, + "archive_bytes": 160000000, + "sha256": "3bbd14328ebc00579004247032b9db6a8ca291a389ae5800a76a027b0f3f2ef3", + "source_signature": "afa7a50bc17d8c30737d2908c7f6caacaa49eeb7e5039490983b3129b9b105c8", + "index_sha256": "88af818ce259df37a183ed1a03f578defce1fac2c20902b174d2813d9e86eb14" + }, + { + "file": "shards/seed_smplx/00104.tar", + "files": 1129, + "samples": 1129, + "raw_bytes": 164760734, + "archive_bytes": 166789120, + "sha256": "2b09632d4e6ddf06b007f14d5b38ef14c50ae17c49f960e7d12527bfcf1607eb", + "source_signature": "8aaddde9a8f99206bc5eebd1d4de59532c57a555c295255f21f177bbf2e8b09b", + "index_sha256": "51b46733c92c631fd91ebbc16c06db84a354e9775ce5c535b48bd8862e5bcdea" + }, + { + "file": "shards/seed_smplx/00105.tar", + "files": 1125, + "samples": 1125, + "raw_bytes": 166888142, + "archive_bytes": 168908800, + "sha256": "03d9900c9197b641fc63df498f8a0f4bc083aa393d2f46e2fa8d16a532a53f12", + "source_signature": "1652ea715c589bb91497c47bc38f7109e9d4e6be63aed7130c6d026af3b78fc8", + "index_sha256": "699b7b3d7cc906d372d697eb4a7b6c4825db315292d516953120b3173927f3e7" + }, + { + "file": "shards/seed_smplx/00106.tar", + "files": 1187, + "samples": 1187, + "raw_bytes": 173912570, + "archive_bytes": 176056320, + "sha256": "2024bb5757a9c0517a387a5eb638f8e2a54fa2628e4ff7bbb3bbf4b43beef1d1", + "source_signature": "35054e56c1d2a33a3c1b5c058a50cf239cfcac76c646e0f954a194086a78d706", + "index_sha256": "cecc0d1e8c96a9d8eed16e78ef8f1b010120ea8adac1ad75156a891020e10ec6" + }, + { + "file": "shards/seed_smplx/00107.tar", + "files": 1113, + "samples": 1113, + "raw_bytes": 168590177, + "archive_bytes": 170577920, + "sha256": "adac2af366efd39aaa1adf21abf4a86e199b2fb1cf61f7ad6548b5919b4354e2", + "source_signature": "de47b6698f61375c154a7d7efe9c82368600e7810a81b090b0934a116e2fbf81", + "index_sha256": "fe9f9e6afa98f745f3daa07c7c09a3b4ddfc77062906e858eee5b908292f0f0d" + }, + { + "file": "shards/seed_smplx/00108.tar", + "files": 1084, + "samples": 1084, + "raw_bytes": 151749621, + "archive_bytes": 153702400, + "sha256": "d97e07f76068e0c4f50c06db41a0100ef72c15bd391ce8c8f388310c5c1c8498", + "source_signature": "c57b45ab313445f029ad5d8476ac37cf5b5529b7d7867e36b341c3ed23e7bcc2", + "index_sha256": "3734a927cf3315682cc78abd26c263993e372a57ac225668d56f4da9e5d7e7bb" + }, + { + "file": "shards/seed_smplx/00109.tar", + "files": 1114, + "samples": 1114, + "raw_bytes": 164219011, + "archive_bytes": 166205440, + "sha256": "b8572df590f8846e4499cf4c163ad72a7a7d8c8a6bd6227e074135b3082ff19a", + "source_signature": "ee4c5357ae3b5a8951b47e449c625904d5e4172915c6f35703a1331a6d3c1e78", + "index_sha256": "24c2c5357ea697ef3c78a9fbf6802d695fa9ebca5f4359e3dcff5cd4f79722a5" + }, + { + "file": "shards/seed_smplx/00110.tar", + "files": 1172, + "samples": 1172, + "raw_bytes": 167759736, + "archive_bytes": 169861120, + "sha256": "24d9748947b6c0d68cf626c448fd2570a913d504b2b479bcca4b9fec1ef83d2b", + "source_signature": "447f19b3c9104058dae1c6e7d2368e9a5a32eac8ceec230cbb8ed91e2e1b25b0", + "index_sha256": "087c00eeba2d539cf76a1e707bac965e051ffbff731f9525a6d161071249b2d4" + }, + { + "file": "shards/seed_smplx/00111.tar", + "files": 1120, + "samples": 1120, + "raw_bytes": 154490440, + "archive_bytes": 156497920, + "sha256": "de034737374beab7141048d5a6d7bca589a3e965b40307a59c62d2bfca077f35", + "source_signature": "c9a5fdf9b583da8bb17e7ed43ef6a86633f5f24900f2817e872b63282b517697", + "index_sha256": "573e17bb6e1426cdde3b35c04f473b12c4b486187e94723b97ecd4f908289f67" + }, + { + "file": "shards/seed_smplx/00112.tar", + "files": 1161, + "samples": 1161, + "raw_bytes": 167205541, + "archive_bytes": 169297920, + "sha256": "342f203aeecac3ef6e6824e5bcf6d3157633cb6b1e99e2a4ed1a12ca6409c039", + "source_signature": "748926e6ba7f4fea212bbf5fba1d5afe9bc56f6f7b18086408b86796e4130285", + "index_sha256": "98498aa3d02aec357edc516457389c504e563ffdbbfcea7ac4d64b735919dae1" + }, + { + "file": "shards/seed_smplx/00113.tar", + "files": 1129, + "samples": 1129, + "raw_bytes": 156959639, + "archive_bytes": 158986240, + "sha256": "411cd69678a545e3522dcb8e93e9fe1f12cbd7852edcc63b08b5dad5f90fefa6", + "source_signature": "20024e92d89a17ee06ddbaa36fda9a7f8db41fbcfbdb54dca8b5f06a7cc929ea", + "index_sha256": "730d2cf54fb4c5321649cb0822c2e0bf56fbb08e1af465faea584a3989ee3ae9" + }, + { + "file": "shards/seed_smplx/00114.tar", + "files": 1099, + "samples": 1099, + "raw_bytes": 156931048, + "archive_bytes": 158904320, + "sha256": "82ab9f50d717692125f013679d568684dd7d900533c7168d65d56f2af5de5b1c", + "source_signature": "8f764836450e95973fcc448217bc955cc9aa9fea8c82c078622a867caf9afd81", + "index_sha256": "26926b24f3fb9bbd9d74f9fe84e2c2106467f757f305776fe99d910867813674" + }, + { + "file": "shards/seed_smplx/00115.tar", + "files": 1128, + "samples": 1128, + "raw_bytes": 165496746, + "archive_bytes": 167526400, + "sha256": "d94bc16522683787acc2b1cc201f262be611ccfa5a57a0e03f2178e099f4f9c3", + "source_signature": "518331ab3c8cfb022f51c1be926b854f5c4bed782c639b3761f25ad451954efe", + "index_sha256": "34276f44245226e7034f18b13f719547bf5fa034e77c88a0c23f30cf159988a0" + }, + { + "file": "shards/seed_smplx/00116.tar", + "files": 1184, + "samples": 1184, + "raw_bytes": 161442045, + "archive_bytes": 163553280, + "sha256": "30131142e712ca3d36562bbf20db1a353e081ce6279a59d3dcf5cf6f3a600205", + "source_signature": "61d778fa86bc0fd175241234a2713fc8f3221eee88835288c2e2aa8d994979be", + "index_sha256": "df958916ab5f979087b5bfafc2dcedf7ddb02931ee00488266ea2dfcd81f8ed2" + }, + { + "file": "shards/seed_smplx/00117.tar", + "files": 1081, + "samples": 1081, + "raw_bytes": 156347220, + "archive_bytes": 158289920, + "sha256": "b0566cd2fac2aa22fe3f2ac986c944cc2fae4330e04a357e1cb6d0ca2cfbf1ef", + "source_signature": "8e5efba92bb8dcd7f56e2b062dc180d9118dc2aaa40e29630f83f82df3ad7a46", + "index_sha256": "dda4559493e22145d59fbd9945f55c353d9069db605c81746c35d3a16a7148e6" + }, + { + "file": "shards/seed_smplx/00118.tar", + "files": 1093, + "samples": 1093, + "raw_bytes": 158411940, + "archive_bytes": 160368640, + "sha256": "d0005dcb1ed8618e9bed50820f947e0e5cf361342ef28d3fcbb95e373e66d3d7", + "source_signature": "bc4482a871c4c4cfd78c519f2b87959789d686b59493d359078691b8dae342bf", + "index_sha256": "88796cd62651b43b48198bf4604e46cd121987368c0ec6346c53608f3c3ae19c" + }, + { + "file": "shards/seed_smplx/00119.tar", + "files": 1085, + "samples": 1085, + "raw_bytes": 161619858, + "archive_bytes": 163563520, + "sha256": "4a066e2ff705a63af50185ee15f6988089aaf22529481b56cb91d3526cd7ad2e", + "source_signature": "3ef8099344efb286f555ef80ae3259d2451669a801d2a1c1b7a51c907e1a7ff4", + "index_sha256": "edae1800e6cb25836e3a2d90e741ea947dd344b35f9ed68e3d306423a99743c6" + }, + { + "file": "shards/seed_smplx/00120.tar", + "files": 1071, + "samples": 1071, + "raw_bytes": 158197568, + "archive_bytes": 160133120, + "sha256": "302f21eb04a633d6ad4247e69301ecae3b809e3809406e2fb81576d7ddf039f6", + "source_signature": "adfa9555bdac49043a7fb9db3c7acb4df6add8a18541f731ae0b969a5a460bf7", + "index_sha256": "e264d0e122f63715a20787121c80e7784b58c1d9df708164e875f7512a0dd076" + }, + { + "file": "shards/seed_smplx/00121.tar", + "files": 1129, + "samples": 1129, + "raw_bytes": 165471880, + "archive_bytes": 167505920, + "sha256": "e385f363a274e36708b4410fccc95fce239164fc11267222fefd2ad0fe675423", + "source_signature": "e8a023f891fe7cffd767fc6080642fa95bff3a035cac2f276fd1a8df5bc3564f", + "index_sha256": "967a49d688890a9ec6d3042159369c4a94c73ae46c6d6fe621f352c6afbca7c0" + }, + { + "file": "shards/seed_smplx/00122.tar", + "files": 1166, + "samples": 1166, + "raw_bytes": 164925685, + "archive_bytes": 167014400, + "sha256": "e79d43633fac92569bb35a84f58a8c40cb4b4ed1e91f7697a1f03175fbb9c18e", + "source_signature": "3f552b2e4ee827235cb5137fae41dc4ada1e24dd14f39ad68e59a8ba67050eb5", + "index_sha256": "fae46148d32c3e157b468aab723a67b2194da97e8b410c689c93bfd33a6e443a" + }, + { + "file": "shards/seed_smplx/00123.tar", + "files": 1140, + "samples": 1140, + "raw_bytes": 169472391, + "archive_bytes": 171509760, + "sha256": "ed5c7bee0e4c1660025bd84038d5ddeddd3962a92984d94af78a0dd8012a720a", + "source_signature": "a65b2b5f193110e0358eff8a33493451b7b8119107375cf224eb2def58266f4e", + "index_sha256": "4445f0707048ae7eb110fd14797c472972f863b78ec617a48c7ed7e0324c484a" + }, + { + "file": "shards/seed_smplx/00124.tar", + "files": 1140, + "samples": 1140, + "raw_bytes": 160146598, + "archive_bytes": 162191360, + "sha256": "973accb6920a8d95328c654e739c7877996e57b37c3ca9494be98b9e90962f59", + "source_signature": "663d8218e831169776f86450d80f30b993769228667e049a01e0776305cb1a06", + "index_sha256": "f0d313ca1a4b549fb594523de13d530a36df905b904f7ba6cdf910f65ac12a13" + }, + { + "file": "shards/seed_smplx/00125.tar", + "files": 1132, + "samples": 1132, + "raw_bytes": 169878665, + "archive_bytes": 171919360, + "sha256": "c19aeddc48d32aef1ba0c5617a73e8a236dd26645c226e50fec6f2443c487d44", + "source_signature": "eb9b7debd0793d90d3df30b51a7f6e9b4d8724aca68d31c34438cb98263c36ab", + "index_sha256": "97419714b22b07d7bc992f8971d7865ddea01dac3a698ca9912847931ed75491" + }, + { + "file": "shards/seed_smplx/00126.tar", + "files": 1189, + "samples": 1189, + "raw_bytes": 166561299, + "archive_bytes": 168693760, + "sha256": "44193111236847a56910119f53d30909c7f16ee37be777ef180c789c9db0d7ff", + "source_signature": "ee8ce232e0eb7b6c0cadd1441940bc05ae86664241e23179f738f8e801a40828", + "index_sha256": "3b977ad47e88ec0158c8a884bdb1740cf85d525e50a7b39e237fddaebb9bb37b" + }, + { + "file": "shards/seed_smplx/00127.tar", + "files": 1075, + "samples": 1075, + "raw_bytes": 153824417, + "archive_bytes": 155770880, + "sha256": "08838d1b699443124f18283f627f8fc9c7479af2a187a54831c81963840ecc3a", + "source_signature": "0c209e8a54fb60a2d9cf83d6addf7e4012c13cad04684e45e1a2e983a16cfcc8", + "index_sha256": "98ddc59941a9c83bba68452825d669f1c7e7a0d2d5a30bf09c84ec796e1c8e65" + } +] diff --git a/handoff-v1/resources/RAW_REQUIRED_FILES.json b/handoff-v1/resources/RAW_REQUIRED_FILES.json new file mode 100644 index 0000000000000000000000000000000000000000..9bf45ce5f56b8c646c315609de51410c50a1ffa7 --- /dev/null +++ b/handoff-v1/resources/RAW_REQUIRED_FILES.json @@ -0,0 +1,4057 @@ +{ + "repo_id": "Tinghong-Ye22/hfd_data", + "repo_type": "dataset", + "revision": "93ba4a7a9c66b3e1b7041e5f4ea038ef2418b948", + "files": [ + { + "path": "shards/hfd_audio_command_continuous_office_v1/00000.tar", + "bytes": 41472000, + "sha256": "edc572e1618859c3bfc0222779798067814e48d656d7c0e8f41b2e32be6f159d", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00001.tar", + "bytes": 38277120, + "sha256": "48adba5d312b2294c84dff371f6035765843c80691eb7e3f37ae289e18d1b717", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00002.tar", + "bytes": 44943360, + "sha256": "ffbb4f34de637beced4f9d77fb461397a92073137b1b8988def1eda2c335b6fc", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00003.tar", + "bytes": 42844160, + "sha256": "8505a4adbf2c097fe9623ee02208948f00fa70111a4b4054860c54cc4493b9d9", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00004.tar", + "bytes": 41994240, + "sha256": "7cbc76258f5c97f4f9521675c702e9f62e6dda8a6d6269b14d1c44e014c8a120", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00005.tar", + "bytes": 40919040, + "sha256": "73db027a323025f5ab1fa692a2b49c8c4a56fbb04ed9767b66688095523bd9c4", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00006.tar", + "bytes": 41953280, + "sha256": "637182c81d39415b614beb5f51c652747f0ec27a5080f6d1421bd9c7a74aa3e5", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00007.tar", + "bytes": 40417280, + "sha256": "40e304f756efdbc92c6a7529ee1a0ee25946c882ddc0fb42c96815f9bff273cf", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00008.tar", + "bytes": 42936320, + "sha256": "37ff9d9fc933151d2488fe89e2e128381deb70a18c1362aa1e6bcd635d841e48", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00009.tar", + "bytes": 43202560, + "sha256": "0e331a96c0319595445874e497c0251e47f17a9e937d8ce19bde4e159d845199", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00010.tar", + "bytes": 38768640, + "sha256": "05340ebc28eecb9edbf72e811969f5e1ac5b371cae6846121b97d40c577b712c", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00011.tar", + "bytes": 41902080, + "sha256": "ee7eccf466d7359ad848fb6900c596793c69e5275b9282939ea03013fc228681", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00012.tar", + "bytes": 40704000, + "sha256": "a6c4f86dac1bc08371ea25b61fd0f2af3258156d180400c7bcdb5a19fc265ee0", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00013.tar", + "bytes": 41779200, + "sha256": "be5e8a931aacbd1a3ff45739484d1583f65e3debce50c7b4dcdbba27f9566b3e", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00014.tar", + "bytes": 41216000, + "sha256": "e9b0c8fafb1a632d43efd8547f7e9007c06d2a259e14b0a310b795892c2918ee", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00015.tar", + "bytes": 43622400, + "sha256": "bc43e1804a64e13b30856d1ece422122176fae4070eb97de7f0d3be137d8dbd0", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00016.tar", + "bytes": 37580800, + "sha256": "f8f11a7fb8b5ecac915fa7c56b3d1000ba85736dad30b64e3512d37d2cefa8e0", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00017.tar", + "bytes": 40550400, + "sha256": "778473972e4c65dac29f6aa4ab4f3178acad333e56bccc31903bf898dfb0532c", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00018.tar", + "bytes": 42516480, + "sha256": "08421ed611e54982a7019caa09dad9c3a0ee73f2bc7e93828f18027a56c91d96", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00019.tar", + "bytes": 43427840, + "sha256": "351c532d3fa29c7580e1389295a51009d67a7102efb089dc5127daae1f878609", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00020.tar", + "bytes": 40038400, + "sha256": "d91e183aaa8643ead7dbae2d8225f05579f2d5ab4386d315a25a5235b1eebe2d", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00021.tar", + "bytes": 41871360, + "sha256": "87ae2ed5a7df09e1439fe0c6016f440cf6388f19030af300f4b95e7521089153", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00022.tar", + "bytes": 35420160, + "sha256": "8813ecde7e8d1d2a76d47c94155b8aa99319e1aed985b14803ba4bcc58a05055", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00023.tar", + "bytes": 35491840, + "sha256": "2baf666f0d99ace2c70b993a3d86fa58a2e7167971060fd22450cf7601c8cfcd", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00024.tar", + "bytes": 46438400, + "sha256": "bf91d285b5d7caa4939fd7aad56758be2163c1501238f8dab56dad87e7bc9ad8", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00025.tar", + "bytes": 37519360, + "sha256": "cee03cda337df0d03b086d27d890221ef54b99957269eb4d8d35c14218422dae", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00026.tar", + "bytes": 42117120, + "sha256": "e914a78d22e8b9f1b87cbd0e8295f88fcab53c5d3727edbc874735ca7184056a", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00027.tar", + "bytes": 43171840, + "sha256": "6b9dbe8e9c05480799e806cffb727416d76767a6bc1674b5218a6231264e257b", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00028.tar", + "bytes": 39833600, + "sha256": "e6a113ba70b6d4eb0c756808db92cda84e22edebe2a2e901ba8e11e47b78ad10", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00029.tar", + "bytes": 40448000, + "sha256": "8426fcd42016303d4326b3baa4968318a311c09ba9779d528d97c68dc4969c70", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00030.tar", + "bytes": 40294400, + "sha256": "22ec6bf5bd0ccccb93b11199d821601681a13473b246ca83d4e825abc5bdae6a", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00031.tar", + "bytes": 43591680, + "sha256": "72bf55d2c6b2d983aee13b13db10ff85e277cb6c8e158ec2823351cf5e06ba8a", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00032.tar", + "bytes": 41738240, + "sha256": "4f283ae1a4d187da989098dd9913fcc8296d28fbabb37d4b1f702abc39b9b519", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00033.tar", + "bytes": 43683840, + "sha256": "97e2c3783d67edf47ea1fc2a05416709727df703540322f0932dfead1841b692", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00034.tar", + "bytes": 44001280, + "sha256": "945539f498ddb717abec6d29b8705024c86eff8e5dce0baeaa0011a5782bb2cd", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00035.tar", + "bytes": 63549440, + "sha256": "4e4146a507f665b867f7ab0834e6057f54f34a9f29a794ee5dc99f6d1c2a11a7", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00036.tar", + "bytes": 40693760, + "sha256": "45fdb33aeece62165daeb1a642ec232577f93bca909be39a3a73fcf9007e13ba", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00037.tar", + "bytes": 39249920, + "sha256": "3cd66bd03656b3a12233e4cdc1c72e0012d5cdf6122cd441389d4b688b5669f1", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00038.tar", + "bytes": 43878400, + "sha256": "e3d82343e93766f15a91d814a29635af5fb043935b6551ff2bd7888e4efcec21", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00039.tar", + "bytes": 40079360, + "sha256": "b8439316f65c8c9eeac2c006447887dc2f688fe6a9a81bfd40dcc7574302eb00", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00040.tar", + "bytes": 40355840, + "sha256": "93715b6b0417beec03ab3f262ef21ab94fadd5e8bc20051cf9e91be2c4bda3d3", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00041.tar", + "bytes": 40888320, + "sha256": "e2ee393fec34f04412b8f042a0672936759dc8e05d77f95c37ff028cbfce80f4", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00042.tar", + "bytes": 41789440, + "sha256": "da97f8de1a9e03bacb919b1a8a0a6ef08a93276d4678d7b85f44356c6ee60855", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00043.tar", + "bytes": 44001280, + "sha256": "fc6239a33b3de72d7493735536b3b56e992a1a46ced8896ff69937bcb51111d3", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00044.tar", + "bytes": 47267840, + "sha256": "8b52ae33f153c1e151c89879a721d56ddff7f8cd3a68b7479dee2e3485e9b3c7", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00045.tar", + "bytes": 37027840, + "sha256": "0f8467a200b004acd9611000866e5405e1e35c0ab28a331e1f1556fd79f3c3cb", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00046.tar", + "bytes": 44738560, + "sha256": "12c02c76fe1f4f4a01c142adcd7527c099d92597af940162ce035abf1a1bba36", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00047.tar", + "bytes": 41728000, + "sha256": "602e4c90756feb899a46ea47185b94eb4c6eb8ab71f0e1229683d335d1304532", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00048.tar", + "bytes": 41349120, + "sha256": "978ff0520d79d4a710286de4de5362fd48d6142c59f7d679459c4507f5e29fec", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00049.tar", + "bytes": 44943360, + "sha256": "45ce369858b176730b69e5a3764bcfc09cd155c37cf07721dbd3a5b0d3bb9939", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00050.tar", + "bytes": 38819840, + "sha256": "2c7e6761a7a5986ecb3aa53d58ead04f7311e5c2dc15f72adaa67dbef13bb560", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00051.tar", + "bytes": 38748160, + "sha256": "5639f7fe8f0ac4f6b412ad180327f9e1aac58a5f560e6a72eec416edc0543c90", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00052.tar", + "bytes": 40837120, + "sha256": "a6b80039645a4e5f973e3617e5997885a2463897a7529240e0a158891558d27b", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00053.tar", + "bytes": 39249920, + "sha256": "eb9646a3a6e8295b74a482c5937a07e2854ca740d9b4efd38d82b7cb3ba72ae0", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00054.tar", + "bytes": 58787840, + "sha256": "43cb1b9bad8d93c75f55ee64f497dd4524650b4e3f5f10a304b31fc55262da1c", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00055.tar", + "bytes": 41543680, + "sha256": "594293e2b454e020f2241ef35dc4d54083c30e1e68be7e36168e851bda4b7bb6", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00056.tar", + "bytes": 39557120, + "sha256": "c0184deee534d40d20c498e9a54740046ded3a977e0c888883b285de25c51231", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00057.tar", + "bytes": 40601600, + "sha256": "2bb72b2f5a79632475c7e91294d8192a0f101ff6c4a3f7813e64ebbf902076e4", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00058.tar", + "bytes": 70123520, + "sha256": "969c7520b93ad5e36cbc169fa0b0fa699245c62d09d7ee82705fff436bffdd8f", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00059.tar", + "bytes": 39434240, + "sha256": "bffb34ec454ca4dd677264b2da5d53976c0db61862b027ee2b7c1e66895c57df", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00060.tar", + "bytes": 40919040, + "sha256": "3b192f005ff5cb00e235a4e0a7f2eaa8f1617ee259589516985932f327b698ef", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00061.tar", + "bytes": 39884800, + "sha256": "20f155fadb145c348311681cf82acdc57f478628986fc42f44295866a3b7eb3c", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00062.tar", + "bytes": 37160960, + "sha256": "32f51e566fbbebc7a89ed10f37cc4162913362cc4ba1b8427c3aac99f837f8e8", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00063.tar", + "bytes": 39608320, + "sha256": "b6136c399c8b46a1e14f84341c3fa37c53747df052713139f4519ea3fe10a4fb", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00064.tar", + "bytes": 47595520, + "sha256": "71fcb02436ac976d21d07022132df2951852b4eb9a7fbb0f8b2263a991968dea", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00065.tar", + "bytes": 39147520, + "sha256": "18f884d8b5a21d7b960650ae238900897beeafae85918431ea7dbdf8e07cb463", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00066.tar", + "bytes": 46438400, + "sha256": "e01646ec62dfafe4d711a8e0c6b49814a984e05b31ca6853cff5db3e03f39144", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00067.tar", + "bytes": 49387520, + "sha256": "f16fac3abd419f287b53249993a0f0c1637f7df24e8bcf428cdd8dacd16309be", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00068.tar", + "bytes": 42106880, + "sha256": "6beadcbf216402f3b14c3e2770f00b1a07a0923d1ee6fd0f853f83f494bbd335", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00069.tar", + "bytes": 38553600, + "sha256": "abaace22b7c89b32060f802742fe03e895849204efbb1ffcdcaa55096d8d79de", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00070.tar", + "bytes": 43991040, + "sha256": "223e4d79664172045ac043581c4d7c76b2c23250134749cbac63b13a07da36ed", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00071.tar", + "bytes": 45916160, + "sha256": "01649b88a4883a71207105df06d4fd18acb5a7109a4085ceb643921fb42b151a", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00072.tar", + "bytes": 43151360, + "sha256": "448343d454f4e05f61cbe1161623c7e1ad45c38c00361bab71f356b0d561cad5", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00073.tar", + "bytes": 39311360, + "sha256": "d098cc02497e16396f28283a7d3586a345be5175fa7ed644bc865cef662c2316", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00074.tar", + "bytes": 75888640, + "sha256": "d071f984dc13e707aaf2684f0db7d6b19f8f9be39c831aa7fe7143fae446d18b", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00075.tar", + "bytes": 41420800, + "sha256": "b4092c8b245b8000ddc54595626c97f4f1461bc3aaa4038cd15081363f62963e", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00076.tar", + "bytes": 43038720, + "sha256": "a19e9986aa349d9afa65ef740e4cea5e8d8b7e5c4cc42e1bde3cdd434d2ea0d5", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00077.tar", + "bytes": 42700800, + "sha256": "9239b7655b44b17bf87cbb04aa455df6684765b82552ec80dba6bf0113c7d74d", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00078.tar", + "bytes": 44308480, + "sha256": "f2d3f57be0d1c3ae2881496d54e38ab176fb4c57630a149a5a883f5d6b3c1898", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00079.tar", + "bytes": 41041920, + "sha256": "82bac1648b4b439109310ae93392ffedf25b028be1f5d3970441692724493c77", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00080.tar", + "bytes": 40140800, + "sha256": "d07687b31500b57ff57082ff4363c24b63f811a379da6fe894f4b3ccae52a458", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00081.tar", + "bytes": 46131200, + "sha256": "fa201091f166f5ab939f633edc3e01858f99f07a4811a49525edcc02ea31eb44", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00082.tar", + "bytes": 41277440, + "sha256": "d29bd533254574772d4f7a7243f0bca5ea3e8b9046eb7dbad942ac5da22c90c6", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00083.tar", + "bytes": 62863360, + "sha256": "923f03cba37138f88709d92ad6ed98d2d056dc1e566602b54a02106288d289eb", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00084.tar", + "bytes": 40816640, + "sha256": "ed803810e3e7601bfb4fabe2e48ca43b75d0335d16eb6491aacf3e888f756c82", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00085.tar", + "bytes": 40079360, + "sha256": "14b31ab861b555c49570309a242df0e18493ccebce09877105167420d69ef7f9", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00086.tar", + "bytes": 46776320, + "sha256": "05231c8bcad79c362e9a6ae1ff0dc2fde89c7dc7a106880e97a8d9b1df6e2c64", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00087.tar", + "bytes": 42465280, + "sha256": "d53a7f427f9c3dc2effe79789be216bbf6311258a9947bef2e341a75699adc3d", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00088.tar", + "bytes": 43560960, + "sha256": "98a82dabf77bcebbc01fbc81894654058f67e680af05ed709bea3a0dfaf2839b", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00089.tar", + "bytes": 42229760, + "sha256": "4152bc085746365052fe54dc3346c3b9780320de6d42ee1028c404a81c7df14b", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00090.tar", + "bytes": 42496000, + "sha256": "23923f495694833ebc9bcec4805846f3295a2add1dcdfe3b38f7014891840c3b", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00091.tar", + "bytes": 38656000, + "sha256": "6723d8f0dcde70cd25451836309d135edf9e481a8a8c827bfd8c66bf9f0d8d8f", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00092.tar", + "bytes": 41123840, + "sha256": "49559f0c42628020a5f0d96ad9296321d740ed3c20c3ccec68b091956a5bbf41", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00093.tar", + "bytes": 41154560, + "sha256": "a48ad37706aeb963e9849535d2c86c46bb6dc20e5afb4571e0a755488da1e950", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00094.tar", + "bytes": 42147840, + "sha256": "c9da41d6a27f32a8ea4a93e890cc7f0f175b89e8941e5f1fd1e7a2ff5a9d61df", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00095.tar", + "bytes": 43100160, + "sha256": "067f2fd10ea207b0fcf6532518f9340142f0d2ffe6655adc48ea913273f58277", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00096.tar", + "bytes": 41297920, + "sha256": "89db4237b791793c89857aaf2d4f3383bf60a1f117d0e4f7923606c2c35adb41", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00097.tar", + "bytes": 37662720, + "sha256": "b2b31d169c73febb3bc57f4edcbaa1ae151ad3789fb0ba90f460102893e87ec8", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00098.tar", + "bytes": 55040000, + "sha256": "3d2519e4566a715186efa5eca08556c3c46e592e8be5c64df8da44aca188261f", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00099.tar", + "bytes": 42035200, + "sha256": "ece552defec9edce614cc97b2da44edf9118a029e566c8c9af48fd796e68ba9a", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00100.tar", + "bytes": 39680000, + "sha256": "43232dde4494ea3aa9bac35824c19a65a2784599292a601e0c4ad0cacfa89faf", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00101.tar", + "bytes": 42803200, + "sha256": "1990375066bac9ff97a52db8eb52f11f83bf2e210482a2749e29af3232b5410b", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00102.tar", + "bytes": 40632320, + "sha256": "789ef419af768dafd3a278d2c22576144206da2d039151c41a9af2a30af379e7", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00103.tar", + "bytes": 57989120, + "sha256": "eeb72aa4f1f304d45e4a5d12509ec92441ee86af984fe68d6397084921d88157", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00104.tar", + "bytes": 40007680, + "sha256": "4e726608043f83c504f8fb4f26672d3db1f6c65bb1b02048ebf407b9ccf30de8", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00105.tar", + "bytes": 40847360, + "sha256": "3556fe5e90aad3a199297564e49f252645c2141d60945e724a9bbea9a375007c", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00106.tar", + "bytes": 40601600, + "sha256": "997a200db4a74be3c2074b938f30018499c26f7b6042b8a54557d933f4f78293", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00107.tar", + "bytes": 43407360, + "sha256": "4eb564a14ff017642b519d6c28d939a8ee516a06b46d235a55142028844b5dc0", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00108.tar", + "bytes": 37048320, + "sha256": "76f443b78b67068156e8d67f4910d787733e49b70ed2d1ac58f9193b3fb3b1e5", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00109.tar", + "bytes": 42301440, + "sha256": "ce06f2e54a0c327a265f17c40d54153e05782713abc839cecd3876b22d51e8da", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00110.tar", + "bytes": 42782720, + "sha256": "0470ed9e94514ca7d8bdc28a65588fbb7ce5b0902d59f37ed1ece4cf332e483b", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00111.tar", + "bytes": 44052480, + "sha256": "46f810f6d5ad672c2f29711be6e6676febd036f453b0084caf2ae1f110aeff07", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00112.tar", + "bytes": 43284480, + "sha256": "e0bcc570e0abc5af3bb1296565188f0f27fbaa703e9d32cb5e7118a5e861676b", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00113.tar", + "bytes": 40345600, + "sha256": "aeca80e0e283c7af5198c52ce88d2d7439a517f279f0f218c5a9eefc104da0f7", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00114.tar", + "bytes": 44871680, + "sha256": "1a2b8a51cf49a512c4355875add743c403b1944336e66a8a5db69b6421ce734a", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00115.tar", + "bytes": 40263680, + "sha256": "3f1bdd4d33c18269283cb5a9a9eb9f1a727abd41348b3d2120e92e15f54b567a", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00116.tar", + "bytes": 40775680, + "sha256": "229db215622b0b1880abdec66bdac1b8a996ed2f9b7889e5b93df34f74dc53e5", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00117.tar", + "bytes": 41902080, + "sha256": "e86b54618e621b61bf8d32740fcdb81cba22e83bfd6a41f75fecda09021e54d4", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00118.tar", + "bytes": 42987520, + "sha256": "499d6c22010c4c9c01fba3fea7e2cfe96bd9cf7ea14ba4906700ea4cc4e754bf", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00119.tar", + "bytes": 43141120, + "sha256": "5fffb4e6348fec257a194e5a5aadf46f83693bb808b92a41dfd472cad9b084e2", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00120.tar", + "bytes": 43345920, + "sha256": "9d7ff1263e7e2db30beb53e09f91fdaebf0e60f060e1a2e16c05a56126dbf94a", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00121.tar", + "bytes": 41379840, + "sha256": "1e8e690796fac27f42d7912610fbc5a9e1f1e8c745766f15ea4d1f10fd0dffc4", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00122.tar", + "bytes": 39792640, + "sha256": "ecbbeeb16f1d1a6adcc46e142d93c1a5cb932924ca7b3a537643223d065a25b5", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00123.tar", + "bytes": 46233600, + "sha256": "d6c8cd91debbe72a09084e1486da3dc6c35c5064eeb20dffc1934beffba9734f", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00124.tar", + "bytes": 43048960, + "sha256": "592e96be49810c79254232a94cac250bd799da6950109a5db96125069d2e729c", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00125.tar", + "bytes": 37437440, + "sha256": "87497e9131a30052fab22487c6c902a9b1f8c01c8047b276601696324b9386c8", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00126.tar", + "bytes": 38184960, + "sha256": "f882065bf4f6d11fc6c369b5d6b1b448fee89904eee177556ad6d843263cacbe", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_audio_command_continuous_office_v1/00127.tar", + "bytes": 60928000, + "sha256": "9871cab79cd56608dc454b80806f820444ba67655193fa0177dc248a4f489d2e", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_backchannel_commands_v2_short/00000.tar", + "bytes": 2560000, + "sha256": "e87a6cf29bd174f62239fba65ec71c06aaa28f145c519fc9b4dc55ff3219b886", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_control_commands_v1/00000.tar", + "bytes": 3686400, + "sha256": "931ed66438de3bb950af8b0b950e0515b9eeaaccd334d78088bfb1c51c09de32", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_hesitation_pause_v1/00000.tar", + "bytes": 145408000, + "sha256": "011993688d55f5b0f0c54f8d402abd82957b382d94e94fd788737142df117d4b", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_hesitation_pause_v1/00001.tar", + "bytes": 144261120, + "sha256": "77e21dd06fac3dab3fe80d4b36b8a648d36deb222dc5e33f86be55abdf2353de", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_hesitation_pause_v1/00002.tar", + "bytes": 142284800, + "sha256": "bc461a0b4a156301109af850d9fc90b802a8f2f4dd777d43f01127fb9d93846e", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_hesitation_pause_v1/00003.tar", + "bytes": 142551040, + "sha256": "22453fb9b1268170eef0f518d4c3ae9cd2ef441589493f3b4ec3fad05c9eb66b", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_hesitation_pause_v1/00004.tar", + "bytes": 145909760, + "sha256": "6509f7416d93de7bceb2cbc89053da36342e8bef1d43e9a7eda3d4534866ad4a", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_hesitation_pause_v1/00005.tar", + "bytes": 149227520, + "sha256": "ffc3758689d1e9ac16249910838f1dfe79b6f0d022e8b888130de62676c8a028", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_hesitation_pause_v1/00006.tar", + "bytes": 215091200, + "sha256": "26fef958c65f1f44634a44a0f38b00e2505af44f1b6988f2206f42b4d6359464", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_hesitation_pause_v1/00007.tar", + "bytes": 147374080, + "sha256": "1b5271e59827d30fc06e5628b0aad017bf6c400f0bbcff8459690d6044dc144e", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_hesitation_pause_v1/00008.tar", + "bytes": 148039680, + "sha256": "dd70aed4ccbd59699f99ba10dcaac7f63dc9ae4cb79323239f71f41879605b26", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_hesitation_pause_v1/00009.tar", + "bytes": 147189760, + "sha256": "3efab2954680ebbe3447efd7e71dcaa632790bb30eb32904778d53f227047912", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_hesitation_pause_v1/00010.tar", + "bytes": 144496640, + "sha256": "e5d5a9a238b9bb6f0740524ec90522730e2fe87d047aa7858b21d0581b941f69", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_hesitation_pause_v1/00011.tar", + "bytes": 142622720, + "sha256": "19015558a10a31329dd01f7c82c8513cebcadc8870c49b3d66880fd98ef61210", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_hesitation_pause_v1/00012.tar", + "bytes": 145285120, + "sha256": "7f769b427299d26c54bc04cb0caae40e7d1e671774ade2ac64d547c026a855f7", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_hesitation_pause_v1/00013.tar", + "bytes": 146882560, + "sha256": "02f79dcc1d83a06712499c84628b688f54e10a269bce733550a1cf628766be36", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_hesitation_pause_v1/00014.tar", + "bytes": 144967680, + "sha256": "86d3636ffac3b47cac59b07b3bc923beebaa25f17f3188db786bd39e62115d40", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_hesitation_pause_v1/00015.tar", + "bytes": 148101120, + "sha256": "4fd8bdfa4f853e78dd4c7b06d4269d1789e76c3af6f6a6ec5a67df76f8e23e5e", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_hesitation_pause_v1/00016.tar", + "bytes": 147619840, + "sha256": "f2aa704fad58a877eb32cf309ff0cc0cb9fbc47c5a56a65c5b97a1522fe2d832", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_hesitation_pause_v1/00017.tar", + "bytes": 146708480, + "sha256": "18efd732851496c53f75f5d44a2da76a8e3d062f4abf214c836b244de7e7238d", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_hesitation_pause_v1/00018.tar", + "bytes": 145438720, + "sha256": "84a4be2df65fed861892fefe25c7c5ad441ff0f0d87d4ada379649742e393c21", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_hesitation_pause_v1/00019.tar", + "bytes": 145448960, + "sha256": "89ef29d2f8569b6738e1eff723e36218c73112f1ac7c35992a164a8870f44a3c", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_hesitation_pause_v1/00020.tar", + "bytes": 144199680, + "sha256": "3cafec0fa7c33d67440c4c44ec6743373953b3c65547a3eaf264412eb3e5f40a", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_hesitation_pause_v1/00021.tar", + "bytes": 142612480, + "sha256": "6a29792a46cedc7e63104d080e7621457dfbb6af022b1f7063d7556a5c2deec3", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_hesitation_pause_v1/00022.tar", + "bytes": 146278400, + "sha256": "81f8deb18d47404583fe461bda3d5fdad6800dd8311b180ea091f45ba1d74687", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_hesitation_pause_v1/00023.tar", + "bytes": 149114880, + "sha256": "f667666f5b7e8abd71c6019d1f886e905296c6ac2711af8d821cbd469c546ece", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_hesitation_pause_v1/00024.tar", + "bytes": 144844800, + "sha256": "f6d04acb28c1d065d840ea66f1b60794adef947905efb5a53f31a9d1744c9afa", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_hesitation_pause_v1/00025.tar", + "bytes": 144916480, + "sha256": "be51f5d3e190ee72d273feb21e516bf5471eeb5996dff4954796c561c2928d42", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_hesitation_pause_v1/00026.tar", + "bytes": 145172480, + "sha256": "94a043687cde1ddf9885f064482c476cba515ca464c11d64d0dc34d660dd2149", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_hesitation_pause_v1/00027.tar", + "bytes": 144158720, + "sha256": "445cf676beb1e7260b737aa56b76ed894e636db7056aafaeaf96a197e03286c1", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_hesitation_pause_v1/00028.tar", + "bytes": 145479680, + "sha256": "2269c653264e9f7e03ed161f257e354b97bcb82e2c2610760caf6e8d0fee971f", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_hesitation_pause_v1/00029.tar", + "bytes": 136335360, + "sha256": "631ab8736d60fbb3e69a7dc9f52dd1f870cf049938000bd8db521f61aa9a009e", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_hesitation_pause_v1/00030.tar", + "bytes": 141629440, + "sha256": "8a52a90ca65d319a09ddb7f37bc94b95dfc313a8072a9fbca283c09eee668650", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_hesitation_pause_v1/00031.tar", + "bytes": 141352960, + "sha256": "8dc406b85228bc7d715674a804cb47f5a23f2794cb856a06e3914bd0288fa54a", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_hesitation_pause_v2/00000.tar", + "bytes": 74752000, + "sha256": "61fb5be780e093a3909c11afa72cefcfb1e37cce7c14cee21c76e89a2c202bc5", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00000.tar", + "bytes": 293949440, + "sha256": "1c9ad74856af5e2cae5983021f8b3111740c8916bf54c0ee0e9cdb79782e1253", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00001.tar", + "bytes": 314675200, + "sha256": "55f9bae2a707115bfc3350348598ed588361526f59e9ee80781d28ed06f34c16", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00002.tar", + "bytes": 307712000, + "sha256": "308b66356270aebb88db350f2e116f26692b13b20a3ee789e623da6fda0f29df", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00003.tar", + "bytes": 312176640, + "sha256": "a47267a9787e6a1ef5c05487ee7cedf5334d7fec2f5067f45d4df27bfd618483", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00004.tar", + "bytes": 308664320, + "sha256": "1e81c204064fa2415cda1b0647dacd302ab510890580c9c28888e5291f05f7b5", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00005.tar", + "bytes": 321116160, + "sha256": "96310618b1003049fc9287c53ced265ee94bf59538da0d3d0a0e546b4748dc8c", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00006.tar", + "bytes": 338053120, + "sha256": "1f18db221a4daabfbfc0c452181f87ee9baacc59b11a04a16617ff0fe31aac57", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00007.tar", + "bytes": 344504320, + "sha256": "9aa75097318f7c978838541223cfd3c3ec28c8365f39ee7da03646636d9b748e", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00008.tar", + "bytes": 357457920, + "sha256": "4496597f9aecde5d258fe8f82400eccc58cf64c4e89aa79976a96c3a2455e799", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00009.tar", + "bytes": 371159040, + "sha256": "734e52471a065b147c774fa3765b5908ad368500ee85e7ed5e82d9d02ca027ee", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00010.tar", + "bytes": 351733760, + "sha256": "ba9f9dbce8a2963e5241f15f9653ff2a5bab71af45f8d29efead466d02a41e17", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00011.tar", + "bytes": 320174080, + "sha256": "1f08984f9c747716bd791ee66818bd26a1c0f5cdc21e8e2905acd2279e0a0f39", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00012.tar", + "bytes": 354334720, + "sha256": "d001d40c0361e7bcb58793c769230706442a77c45a859853c8c94d2ae3568752", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00013.tar", + "bytes": 337080320, + "sha256": "36708e2e45744e223b269f5f29631f3d8bb38c91b241854de58185731c3e608c", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00014.tar", + "bytes": 333844480, + "sha256": "aee98fc516e10edc93590adcc265381c07f51277d31498c7a54fd62b1d62d994", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00015.tar", + "bytes": 317808640, + "sha256": "f5ac09d3a98e1ec3c86943f9774462ea07ed01d4a165787d32ecce7e96065a18", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00016.tar", + "bytes": 338350080, + "sha256": "53cb60eef6e7edf98f2729e8dd94f01889a5a812cef7c79e7bdfac8e4db4dc58", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00017.tar", + "bytes": 308049920, + "sha256": "5efaa10e12feb6e8268de9af55da43903efa3cd47f0ca717e6aa40db69235c75", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00018.tar", + "bytes": 289638400, + "sha256": "72e3ad262204413634f2975d1019e7531ba4ff0f8893d5495011cf8b9a17606f", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00019.tar", + "bytes": 351928320, + "sha256": "75660261661d19ef82c7ae7f3719aa14939f440a8777818adc0bd38e856c85cb", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00020.tar", + "bytes": 341565440, + "sha256": "d2303a912d7e4f20eb7b372e1298e6222410934fdec4ba9908695887567dc856", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00021.tar", + "bytes": 342476800, + "sha256": "122f8b711d6d18e89178b3a8688707749b4581b6b2dfc702f529bd9121555ec4", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00022.tar", + "bytes": 294154240, + "sha256": "5fdc4c60bf8c79b22d782d6ebc3e977b5ee40b268249f3dbd754641d342986b8", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00023.tar", + "bytes": 338053120, + "sha256": "ccb2a0bd081c4e8309e8b4a470e586b1b37c2485d5ada709e75af81602f90285", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00024.tar", + "bytes": 317757440, + "sha256": "60b0e1f3ad4a99b000b147d122e0ec677b1a3475cb9975cf61e455f7c9672694", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00025.tar", + "bytes": 339353600, + "sha256": "a2ba05e545d85b88bc5e2480c212f929c8440fea0834d7e1b5f381ab726ccc9b", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00026.tar", + "bytes": 337264640, + "sha256": "9da5009d95e3d5c3077d3abaddda61218356b04f36c1f27b73709ffef4eea133", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00027.tar", + "bytes": 335585280, + "sha256": "0dfc7b642f4a8831e9f1929b37cc2a204f8e1f8e536e8f3d3f0a94e4695bae84", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00028.tar", + "bytes": 324126720, + "sha256": "d3cbaadbec108a5957a74b6c1ad502ed6468a77ff2d72418282d6f057b49c193", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00029.tar", + "bytes": 319744000, + "sha256": "408b5420b884930ae570aa47e5df155de535f0c31c57f8f62f507455a83911e0", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00030.tar", + "bytes": 314787840, + "sha256": "530c047ffa98f2eb76b57fa66f03d2c669ee26c37cb1fbf0bc89a4ed5e243820", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00031.tar", + "bytes": 308684800, + "sha256": "da57045f971935630a5f9d3f0daf808aa2763f2664cb7e1979ee40ba25d2c451", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00032.tar", + "bytes": 341217280, + "sha256": "1dd72089903e036e029c048b453d7124f1cd60b08ed6a86b3938b24685212a01", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00033.tar", + "bytes": 326963200, + "sha256": "451d605759f71271e0225db4c70f8849e4eeeb21b175a9c922c0fb2825d797d2", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00034.tar", + "bytes": 313743360, + "sha256": "ec3abadae348565c90eff18823fc404291625ef0d51a869ccd7d357c9e9b9a30", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00035.tar", + "bytes": 288727040, + "sha256": "870bc04d5bc0d04937e17e8aa19538edb8492d2ce8dbf05db5b0798294a2d46e", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00036.tar", + "bytes": 354242560, + "sha256": "b1dfb1580954f1a698408d75ed9ce2fa40ffa80bb98bab94059d741b13cb9a44", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00037.tar", + "bytes": 304035840, + "sha256": "c9baa10f19d5aa40204dadb00f883bbb3996fca24caa8a1140ccadb18e1352fe", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00038.tar", + "bytes": 350924800, + "sha256": "1ee34cb4229c289ff96ea8f8a4325ff51763d9778a4c6545ca2713a8c8ac7862", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00039.tar", + "bytes": 319293440, + "sha256": "de6d29e96367a594b43acaaa9a71463194ad7a7af26d1cc089a7daafccff39ca", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00040.tar", + "bytes": 304742400, + "sha256": "2b5b14fd83f8572de3589a19f8983753acc65afbc88b9954419df3867b8c03cc", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00041.tar", + "bytes": 314132480, + "sha256": "783c3781403f651364e7f5e2de2624d06cd2d9bd15cb1406f45886a97963bda3", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00042.tar", + "bytes": 324915200, + "sha256": "fc887574c022d4f94780d5739b87a7322a464f6e1a2192cf05442d43bd399327", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00043.tar", + "bytes": 344832000, + "sha256": "ea90500972f46d0962a71198ffc49b8926b73772bac4f6b13e850cc554a5c499", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00044.tar", + "bytes": 300441600, + "sha256": "c8ec7991acac016961e2ae2520b7bbb5006b4ad4718589b86085eaa75dde754d", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00045.tar", + "bytes": 320501760, + "sha256": "d2f9de075f06d47128dc48d751a4d2584cc73f329ea9c628bd1151af9a222714", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00046.tar", + "bytes": 383119360, + "sha256": "06237f7b6491e03b5d9fb5311487cfe8e9a63e35ae22fd47242d3ea2c4b75e28", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00047.tar", + "bytes": 332912640, + "sha256": "b8079593abe90a4c79c0c5a01699e07d35092d2f7916cc26fb258618b3632418", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00048.tar", + "bytes": 311695360, + "sha256": "3b964d2ee4868d8cfa81ddd0f068abe64f0906da3d4708db24d629cfb88d2684", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00049.tar", + "bytes": 343040000, + "sha256": "ee60f36ccf5b309d2b9c96d52179daead4412c9924b459bc9f7dababb50defc6", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00050.tar", + "bytes": 313088000, + "sha256": "319c0fa4bb6be2130a063ba93fd7eef3017dc45470028b7ed76ff473511c52a3", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00051.tar", + "bytes": 301598720, + "sha256": "d0ed32b0982ad69fbffcb853daf69091f4756c8de3dd61d093402a9a298075dc", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00052.tar", + "bytes": 306524160, + "sha256": "f6c66439c86803c1566bf49945fa0dd2f584d494696f6bfba3eec37f186e844f", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00053.tar", + "bytes": 317532160, + "sha256": "0cafd8695f4916638cafbd7484edeb90c7bb7660bf0c4694d3803df4643d4ef1", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00054.tar", + "bytes": 334213120, + "sha256": "53979336f63d5738efa0af5d1a944d6fbe57943fc23b329bbb834696d479c282", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00055.tar", + "bytes": 337254400, + "sha256": "1c60239d1744f4c93122a50a3cd39352384af7da4bf4b1fec666ac05b476eb2e", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00056.tar", + "bytes": 331233280, + "sha256": "2826f6946920258f62dc915c47805496fcc2f30c81b2afa4a3dfacefc9e97fe0", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00057.tar", + "bytes": 330209280, + "sha256": "bb0d119a6130e9b296881e74c6eb61edc451eb8b14ad0c0e77c1cb320c32cbba", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00058.tar", + "bytes": 294481920, + "sha256": "7faf38b8b8a802cf807680d5a167038d5b208b4fd824e910cdb9e7f4d789e951", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00059.tar", + "bytes": 290775040, + "sha256": "0c9cf7b2d4e1ddd13cb8ffc027fd74e6307d2c2e88908be4087ff1284684769e", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00060.tar", + "bytes": 362045440, + "sha256": "df506e0c05a43939acd6380b7d71e2b525fecdd6cf72718c13a1e8b97af52957", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00061.tar", + "bytes": 378224640, + "sha256": "c910aceb128d873d4b9dee0e77a5e2916929ec79ab0619d1df005132ba2098a6", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00062.tar", + "bytes": 294768640, + "sha256": "8af54d0b5d34ced1943381260adb34101b5e2b19e0da7a571cafbc427627ae86", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00063.tar", + "bytes": 331581440, + "sha256": "6736c26a668ac8ee870aa7217d558bbdb34dcdcd1157ef21fb148aca5588054d", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00064.tar", + "bytes": 320686080, + "sha256": "0718f9c61bcb321f5aaa43239102428d887b5ebc832c43f4cab0b89576b33dc3", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00065.tar", + "bytes": 375091200, + "sha256": "6ee30fce9882e0adb037981dd92179d304c0946dcf841ee78aa160388bb8a716", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00066.tar", + "bytes": 320317440, + "sha256": "12bc796f01af4f920c22c8f6fff10e85d59b5e06b08c096abf24496eb40f0a14", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00067.tar", + "bytes": 337940480, + "sha256": "86f1ffb18c7e1ca2b057d8aa8877d559e9eac90c544bb5d96730d354ef7f3172", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00068.tar", + "bytes": 333066240, + "sha256": "4718392b80bc55fc6c2dd74ccab2ad271e4fe2b302fb2305ebcc8d2d692e6c3f", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00069.tar", + "bytes": 340162560, + "sha256": "1c4948bb65c546ae42405dfb6199c2e1326e9b6c98e4af3e1a173b712eb27e1a", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00070.tar", + "bytes": 305827840, + "sha256": "0bf61911897478814f197295b5cd6ccf43ce3676ee1b185a1d0bac87164ea871", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00071.tar", + "bytes": 365025280, + "sha256": "20abc8ea01326cf3470d6a2a64bda1c7bf54b06094075488a98c588e541fe80d", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00072.tar", + "bytes": 319457280, + "sha256": "d64fb6fc7821b61a13de248ce42c1ca2de18bc29474e319c9fd90477cdeecd16", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00073.tar", + "bytes": 342691840, + "sha256": "d93bc4e91644241e2ab4e5aa14d0cb24a23e46c515a22ebd1a7a72afc05ef8dd", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00074.tar", + "bytes": 336773120, + "sha256": "1aa0611545419f89622248db0edab678c1d03212149d89519fc92d3d0546c80f", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00075.tar", + "bytes": 880414720, + "sha256": "4c50161a37d5bc8a7f0f25ac2bb68c19bc42329eeb0d46c663cb7cdfbe8b0eeb", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00076.tar", + "bytes": 660930560, + "sha256": "d2cfd39fbc700e851bc3fc394b2a018bc7c29f5bc0aa46b3c3e2a8332554d85b", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00077.tar", + "bytes": 338493440, + "sha256": "943aad8ccf298dc51076d26e8923e5fa0ef1bc305a0321765818e5e7c8a4e483", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00078.tar", + "bytes": 376145920, + "sha256": "73833e6967801ce6b67a8fabb378f38c88b78b81425086425f47c9f6de713746", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00079.tar", + "bytes": 319713280, + "sha256": "46a26496bbf1eb956fa24669f4a20becbea17213b10fe34a60a93769b0cdd83b", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00080.tar", + "bytes": 305571840, + "sha256": "16caf6267f03c79dc0f1f1d4fa0c62efabe02e2e7656d52459ea33f8701872da", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00081.tar", + "bytes": 317419520, + "sha256": "29b2af650a82e214da22a53a458f4876d7c3dc08051c8b0784e20b6bd4fb4be8", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00082.tar", + "bytes": 338739200, + "sha256": "07ee3554badbb246ef26d4eb6143e8b5900c1aac3106629a629402b054155249", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00083.tar", + "bytes": 282920960, + "sha256": "14dd731408b707bc481303754614584052daee43c2c5299cc287fd0c192aa240", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00084.tar", + "bytes": 322467840, + "sha256": "a8058395f87919342e195cb5eb3cdbba9eb76d9a2a0d75cd245b9cfaefde1bfa", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00085.tar", + "bytes": 349050880, + "sha256": "bb83da45305d5f5e9a5609f14f4eaae54fc831387ded8ded934efceefc64b3a2", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00086.tar", + "bytes": 334735360, + "sha256": "6d254737bf80078b32feaa76804868603fe3f3bc20f2a298dd3fd92376137084", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00087.tar", + "bytes": 333813760, + "sha256": "b2f33b342fa65cf5b2eb53d71a46856255ab5c79959985a23ed73d60a1e89e06", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00088.tar", + "bytes": 306196480, + "sha256": "324dcb11c27ba0db6a49629115ac2d6909055880e245fc48afb2affdf3d19d2a", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00089.tar", + "bytes": 326840320, + "sha256": "3e8807a87cff8036d6cbb20ab87951485641e3184e8d652b13391e14d0d577d5", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00090.tar", + "bytes": 318504960, + "sha256": "5a435445d686904cd601fbff9c3115d34919757da07a7d7bd19e727d23cb4c42", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00091.tar", + "bytes": 296458240, + "sha256": "7e16396c8e81d4eb41fb0103e013cfc8ce1cf8980c5332615996e5538adee7e4", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00092.tar", + "bytes": 329963520, + "sha256": "34367e7d484fc3aefe4215173c1101785e9ebb65da566c351c99d53d8f028871", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00093.tar", + "bytes": 323573760, + "sha256": "92c1c77f7f05e2e929089d635006feb045fe45c7edf2010c7c75e8177e8e0c7f", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00094.tar", + "bytes": 294850560, + "sha256": "455336914fb1e3821534106f941f2546e13cf07f8454f812e1a3fed0ec1e4843", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00095.tar", + "bytes": 332247040, + "sha256": "172856a0749ad3f406812faa3878097632622c7869d5a14e76e33505104b27ec", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00096.tar", + "bytes": 304885760, + "sha256": "8c7df12f3bc98e70660e95a689dc974114f27db74e8e61052dcac8bd8f68995a", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00097.tar", + "bytes": 352378880, + "sha256": "5689e37f949062daa7268258b07e796c45060c12dfa2dc05a392f228a66743c7", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00098.tar", + "bytes": 343408640, + "sha256": "f130d0cb5661873cc4734a8d6ac3ed18bd0d950e77e43544a0d72c341ef2a86e", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00099.tar", + "bytes": 353740800, + "sha256": "fa0b85e837fd126d41337607570ecf164123f9871923f07650f4ae5bd22e1fdc", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00100.tar", + "bytes": 277217280, + "sha256": "41084fab083711d71a75c0e55e18135efd9f3f1e6208954e0929223925c36ccf", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00101.tar", + "bytes": 319426560, + "sha256": "2d7eb1e76cb19cc4785a7d6abbeaac401e0a9c79077e0c9d347beedfb5bc8b17", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00102.tar", + "bytes": 323481600, + "sha256": "9009b46e01572019721fffc81f75f96e9acfd1e7dfd291483e2129500db2740e", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00103.tar", + "bytes": 321402880, + "sha256": "0e278403b429cde12926b37b2bc03b19e0bfa96574e2fcdacd95b754f744d936", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00104.tar", + "bytes": 321433600, + "sha256": "e4c44beae821cd55db0dc195bbfad4bbb64577000f90eddfa054bc5b97994414", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00105.tar", + "bytes": 350105600, + "sha256": "e43a251edd9fc1425189dfb4658f1b1e9f07e79b7a6259da3b84a21b1408ed07", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00106.tar", + "bytes": 352686080, + "sha256": "77fe843e912f1c1682f075a9feaaeb85aec3ee2b5bacabe337d9d92dafc31303", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00107.tar", + "bytes": 319692800, + "sha256": "432ab61c66614615c4545999f15f67b5e63f81f2fc885e24b923db4809efe176", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00108.tar", + "bytes": 331735040, + "sha256": "478dedfb846ad42e5df1da3560538882882e2040c97824ad940ddf6d818476c3", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00109.tar", + "bytes": 339681280, + "sha256": "1bd91bc9e20e28284ab80688416e5f4810e2018e4b2a89bfe54af5a9d49ab9f1", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00110.tar", + "bytes": 312688640, + "sha256": "00565ea5b764375bc987c764b0d685bbedadf88df84abab5148724dba52fc839", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00111.tar", + "bytes": 337377280, + "sha256": "03870790c50d943cc37487ad5cfeca0bb37290b3d730ca19c65fe38239607276", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00112.tar", + "bytes": 356147200, + "sha256": "91e52fbb379e9e7e967b1f2b68f9b234a0d8abc93990e7523c4a2de37a7712e0", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00113.tar", + "bytes": 326318080, + "sha256": "0a3b9de5e18ff311689c1c8589fe4cebb348548edd5149274f4c61466a9beada", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00114.tar", + "bytes": 320614400, + "sha256": "d2753ebb97ecdb76c38b0aa6e740b987d14314e9abd46c21172e54595a113482", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00115.tar", + "bytes": 295997440, + "sha256": "ffd07ee7cedd3c50f9a3603aa90f236db6e6c649d2f561ef2a5f532104171aa3", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00116.tar", + "bytes": 330936320, + "sha256": "d1d12233bdc20e3690664474456c2c041d3118b6a914f065f512b2cb8f7356da", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00117.tar", + "bytes": 311736320, + "sha256": "54660cad25086216a0169743b8d16f39d8b34ab0baed1ebd4b369cf8b5ac1f63", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00118.tar", + "bytes": 326236160, + "sha256": "b6d42e0d5cb852d4e47072df4ebee01437934c4a75d3f09868af5de7bb2b6cf0", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00119.tar", + "bytes": 323205120, + "sha256": "7217213264c4094adf23e18566abc44e3e98fc1844749877d3e72d11577d3293", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00120.tar", + "bytes": 350853120, + "sha256": "c0e4e0b4e809c2e8147aff4122eb5c281497f5da7af081696b8e3a884484c095", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00121.tar", + "bytes": 337561600, + "sha256": "770a17dda0904f80c0df41774b12a161d9c8ddcb2040518efb7564b182134e2d", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00122.tar", + "bytes": 322641920, + "sha256": "9fc9130e5696b47e228de04ee2b8d64a5bb459c0ca313bccb5cd209ee734535c", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00123.tar", + "bytes": 311173120, + "sha256": "92a88fe95bf95b266d0864c87d1116cd8987b5d5d1ded312eaaa779632bfe747", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00124.tar", + "bytes": 318638080, + "sha256": "e7483ec65364dffa4011e214c2a2c96b06e21dabb95c555c6a9b2b111cdbabc6", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00125.tar", + "bytes": 329594880, + "sha256": "5eacf25f9fccaff30b85173cb48e424b8b125a35164adbaeab2450f83f5037d0", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00126.tar", + "bytes": 329779200, + "sha256": "8bdbd02eed859696fb3457744ed1a4753c3fb8630a5e4063576289c3c2cb9099", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/hfd_third_party_multi_v5/00127.tar", + "bytes": 308776960, + "sha256": "00b5da5ef62f52c5f6238164c65f43b45c5bae057ed4fdc56eadac9b1f49278a", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00000.tar", + "bytes": 303155200, + "sha256": "21cf81d19a18a6c492b9c7ee5a5dd6614f7ff0256aae19ba76f9a9e667bf3f80", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00001.tar", + "bytes": 282050560, + "sha256": "5338b56a86b2b3ee40900ce03d614c049158ad610f7de47432fd6188ce0ae86f", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00002.tar", + "bytes": 289925120, + "sha256": "9f0e540a37ff741b10ea27ad70162454da2fc4cb65008ff2f485c6259e826ca1", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00003.tar", + "bytes": 293601280, + "sha256": "3747593145025aae1eb3c96e0a122f24628d7ac790a0f642357def7d59cf3d2e", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00004.tar", + "bytes": 281292800, + "sha256": "e71c6a11bd9255d1e11d865b30d7e97a84055baf786aa0f91f4c9e72ac1ce5f6", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00005.tar", + "bytes": 263249920, + "sha256": "7d5e4314cf486e077b3e83a0dc52be13a9e9704efd30cbf62f98502adeeeb3e2", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00006.tar", + "bytes": 287989760, + "sha256": "09aa3d009dd1a9c07772499233e6275139de29d72ecaf3911069c8df28b651c2", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00007.tar", + "bytes": 286976000, + "sha256": "f7f397b4d88d1b584b0918d86e70534206ced6d46788d6dac914a1794de5f22e", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00008.tar", + "bytes": 289433600, + "sha256": "a6fbb12e25f74b4070ceb9f087025712c49f547893e0a2c3ccc40e840e626a82", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00009.tar", + "bytes": 281425920, + "sha256": "ceec1b6144f47a47eca74377eac2f51db1bc33524c2ba7b26c4b0794fdf7fb1b", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00010.tar", + "bytes": 288655360, + "sha256": "2b179da6ee30f38018500ff77dbdedae956d68bb66d82104303d2880ee1a2083", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00011.tar", + "bytes": 285655040, + "sha256": "d247c6349aaa75b24a14878506026422c8d65a30b886e88230033908d4c0d350", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00012.tar", + "bytes": 288348160, + "sha256": "3a92184192bf798e061d27d168120e09ce907c2f47dbb38a2d2a7b7c17b80916", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00013.tar", + "bytes": 264785920, + "sha256": "24277149065d929068e2a4e4036e23f1349a98c935e53e882d12b97c3f56a11a", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00014.tar", + "bytes": 266004480, + "sha256": "3b14e0d4a3d5b36494a57d10a6163394b73e8e86801b44d2b028143754476475", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00015.tar", + "bytes": 278824960, + "sha256": "306ca7f9d1e2d5648cbb38a7bbe1344298eaa4c68a8bc58e24c3965352e1c26c", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00016.tar", + "bytes": 278958080, + "sha256": "fab5f20b1d342303f50463df765574886c5e8f611abdadde1a5ccddbbbdeae3c", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00017.tar", + "bytes": 292720640, + "sha256": "f43cd15b0c9954a6e94185281c3746c9ef13b8d14ef7cfe7eda5120d7ba55778", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00018.tar", + "bytes": 281886720, + "sha256": "715abd22cedc53d023f2c04a740bf965fcd4bfb0439f7a0a1eac5c25e43f8582", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00019.tar", + "bytes": 280944640, + "sha256": "2dd11c497b5459fd71cd56517f10ca4f1555867edf60ff40c816307777e7a755", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00020.tar", + "bytes": 279367680, + "sha256": "a242daceb0f6d4a68cc1082a28f2a44c5e366464c5a0d366107e49ebf0f03189", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00021.tar", + "bytes": 269240320, + "sha256": "063135bbe0033f926b8ced9b574ddd28525a0365097831c13613302bbf3fb815", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00022.tar", + "bytes": 279511040, + "sha256": "e84eefaa4b6d7cc8e9440c3dda402fdc0546c0c6ddd4dd0dbfb6cc2dccc1280a", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00023.tar", + "bytes": 284108800, + "sha256": "f33b60722621eaf17db88364c9b8e0da2400018dfcc2451cc6c1f3a6bdb94a36", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00024.tar", + "bytes": 290508800, + "sha256": "99f54ede1f37785557c575fc3db068abd8ed1ef976af19147e8654680f313a85", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00025.tar", + "bytes": 281856000, + "sha256": "5d10d202a0908ebc8f50b2e601a05fd8c7ae55e9ba508b7cf2ff23691f530cae", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00026.tar", + "bytes": 288399360, + "sha256": "df9a36ff3abce1683c0d39a7a7ff2eaa0aed75a60692071f3faa3f7d4d68a653", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00027.tar", + "bytes": 308244480, + "sha256": "c9d307b00141b84d26f81d56fa67ca621f5253a1309f3076b221dc7a95395d20", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00028.tar", + "bytes": 281456640, + "sha256": "88771cc284eca82e4af16d3513345f0fbe85c75bc858bd7dfb6c23d0b1bb16a5", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00029.tar", + "bytes": 275998720, + "sha256": "06114bc60070ed7190cfa49738fc0c7d1e65e8143d01251cf105920d7d63bdae", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00030.tar", + "bytes": 269056000, + "sha256": "be64e5cafb3e0dab83ee66f4efaa0d835378650ae20c149d50cfb218321cd9f3", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00031.tar", + "bytes": 290160640, + "sha256": "da6a6564e9bec49a371653a6b6954646caf29e9afb8f986a33976131c62a318a", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00032.tar", + "bytes": 296110080, + "sha256": "b1edae9a0e2a4c79c3aa5e6b8e69d704dd6a2e18f281676803c5395267c4fb5e", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00033.tar", + "bytes": 280504320, + "sha256": "be88e19574c38cab69dc8e5bda988b83654bd4a0761f9bca4ee0950e9a97b527", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00034.tar", + "bytes": 309084160, + "sha256": "3f9c354420959e58736ad1d7e720993936d2cff7ee3a52b9e4f868d73a094b2e", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00035.tar", + "bytes": 281118720, + "sha256": "7ba66f29f78494a4b6b78525fbdfe6ed6fdea05cd8c77d97a2fd557209a6746f", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00036.tar", + "bytes": 281374720, + "sha256": "145852f101081648c524136ac8a72d96996000ed3804b79d753e1bfd0745b6e5", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00037.tar", + "bytes": 293099520, + "sha256": "68bb4966d8385a156f16d3c20ef92cd45e84cafd405d5f7b549716cb6ed495e4", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00038.tar", + "bytes": 279244800, + "sha256": "53f19ecddd03c8f5042c67f1ad0be76d5c5d7d7aa70b9dc489e698c9b2e27b44", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00039.tar", + "bytes": 299151360, + "sha256": "ab84d94fccd047b6354650b3d05e5eddc650219e5c1a8c9495d86a1f13aed381", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00040.tar", + "bytes": 289730560, + "sha256": "ab8c0448548c96fada47ffe292b2be144d600657c1f3b8cef8b63e0d41322a3d", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00041.tar", + "bytes": 277739520, + "sha256": "f66eb6e2ae4c4a6cc6216260770ad5cb0149e5cda302d0ab669daef63775b9cb", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00042.tar", + "bytes": 287559680, + "sha256": "9cea05baffb1355ea46b4ac7d7266d1f93aede04a03e3404bd5e4bad79ba50da", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00043.tar", + "bytes": 301363200, + "sha256": "1d6db1ef9b5cfcb33cc8d1be69a6f682a47fdf32eb30b016172efd807ed872a1", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00044.tar", + "bytes": 283883520, + "sha256": "5a97edae9a88532f813d4fc2e8773dc617f8d975a13d621f68c2e6521c35cda8", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00045.tar", + "bytes": 265871360, + "sha256": "a7cd4e8ff1f888bd04d328df206328214bc58281704d55ac0f8f532cf7fca51f", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00046.tar", + "bytes": 279040000, + "sha256": "fd878c0fd0ccb0ebc48f35852d427e9fe1774260f8b185b8086793b5ddc9f7c0", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00047.tar", + "bytes": 290037760, + "sha256": "3380b93ad872e9b00f159051ab722af7bc5d976cd054ec6e3966fc04427495ed", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00048.tar", + "bytes": 284938240, + "sha256": "58741f5c277ebd6d79e8f33e66f8afb8a4347e917963448208340ee607306f8d", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00049.tar", + "bytes": 265912320, + "sha256": "81881903048e4591e5aa3f5faf9c45a56fa43d5e4f4f494077e22ed4f1016ded", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00050.tar", + "bytes": 274780160, + "sha256": "482a2441cca75764a7aa3b3d4f6053dfc94efe4a0ca36c67df461aa19a7ce320", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00051.tar", + "bytes": 269373440, + "sha256": "f4977cb655e30b396d0c8516958db7a23c31ea9da48e24e7e3c74ea097583397", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00052.tar", + "bytes": 278661120, + "sha256": "aa61e749a048f4a2f83daf577aa8e5097b1a7300bdc2a83aacba871c48b93968", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00053.tar", + "bytes": 292423680, + "sha256": "da7f67fa49b28faf9efaf0882765cbe094cc08347e8f9c3bea20ae16824574dd", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00054.tar", + "bytes": 285870080, + "sha256": "61c082536825c0ce7f44a4ae1670d9be233794cca9e8aa43a60c3f371a6aa22c", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00055.tar", + "bytes": 299755520, + "sha256": "d92577fcc1803f49f7d8fcd6a83d7a0b517659b2ab50cb80bd6e32f403148579", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00056.tar", + "bytes": 295546880, + "sha256": "97c5efa2e49927cede60611c78a9fdeb9c57dab769c1417f9786d4123d2d2e02", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00057.tar", + "bytes": 284610560, + "sha256": "f5167ae8b4c79aa65c2317212ead0932fbc1d085a54b8e48e2a1fe0b1907db35", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00058.tar", + "bytes": 298711040, + "sha256": "b839affc1abfb61815417badae9f801f7bbb69678f0ebbe769ab8c75f88ae91a", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00059.tar", + "bytes": 279152640, + "sha256": "7387a08d744e0d8be6cb6e1d01786e5c8b5a3f79e282c37765655b857de1e33c", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00060.tar", + "bytes": 275394560, + "sha256": "31307130775d7adbe3bcec12260b59f69bf6a43bc3fe8c8c671a3f95833747d2", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00061.tar", + "bytes": 297246720, + "sha256": "4540c3b9e4dc7bb59cdab4e3ddf5ade0bca5b889b6d590cb1fb63d0d06e37a78", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00062.tar", + "bytes": 289597440, + "sha256": "df296fbff59b26ebe47ffa2b61a62f9ba90465adee1ad8988c9f99d7b9db6792", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00063.tar", + "bytes": 285521920, + "sha256": "0f1ae643872ecdccf412c490974868d07ad9d95b02874694a4686b87f7b902ed", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00064.tar", + "bytes": 278896640, + "sha256": "8e41f4719ee115be17eeb7b02554014bfc084012a5d24fa75d563d2a065b4598", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00065.tar", + "bytes": 315054080, + "sha256": "c08987a76ca1e6b3de56465fcfa73d7a8cd20070dc46cb56b27aad51df7c1d2d", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00066.tar", + "bytes": 286760960, + "sha256": "d530920fafa72e8a607252efa844ae1881f1fe41be6a1d8014476edafb01f68d", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00067.tar", + "bytes": 302520320, + "sha256": "bd0a4f6b8a6c7b83cebd9c6283478e7e80b210ea1ec5ad1a2fcd74ca051727f8", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00068.tar", + "bytes": 265533440, + "sha256": "503dd2afa88eedaf9d18022b4bc01aa42cfd0a46e2df87034cd4c74fa3f7fa1b", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00069.tar", + "bytes": 292792320, + "sha256": "890980c5d504999f2c8dc49c199d211b693385f89c48cbbc74b455285e2b63f6", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00070.tar", + "bytes": 267673600, + "sha256": "1c4d78b7e3a82fbe6b76d64653a7b3adb2f51ec9f9b3766bbfc9598ff256a7c6", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00071.tar", + "bytes": 284825600, + "sha256": "3417d9dd1f46b6527cfc521a6e5c9fab659cd43a73b9ca6467fc1d7795e869f2", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00072.tar", + "bytes": 426577920, + "sha256": "2a35377335b065b34692ab3c6e18e338eec04e0bc50bf5322cdc1349d9089d2d", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00073.tar", + "bytes": 267919360, + "sha256": "8df6d302a6a57c181e50e14893825b7007a77fa2beede30e3a0c0f06331a5b79", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00074.tar", + "bytes": 280944640, + "sha256": "747a989cc3ebc7cf4bf5e18552740ab8d0eb557a3ab2facdac1103f0dc83c74f", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00075.tar", + "bytes": 281896960, + "sha256": "3fadf1856b9d2d727eeb92de45b00d07aea6620fce6813e58a5eec72e27da682", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00076.tar", + "bytes": 273960960, + "sha256": "8456893d6256276bc84787008e5abc91b141b98c366b62cc38c9c73f3f8a93e4", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00077.tar", + "bytes": 278548480, + "sha256": "47b6184143faf42db503780035d4b4c0d742a5f74793cebe2e8de4ed3a69710d", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00078.tar", + "bytes": 286115840, + "sha256": "e4cbf15f753b52e51cf0835f3ea75941cface5008dd1bc25f52aff55b000ecc0", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00079.tar", + "bytes": 297646080, + "sha256": "552e819b96e3e7822beba76d1c8040b94d94f13e46048b175c04c99c1a14734f", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00080.tar", + "bytes": 284293120, + "sha256": "214faedf4b15dc2194df4695d367224a6848527ea52a81756cb7378f20518000", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00081.tar", + "bytes": 285921280, + "sha256": "9dee1932d4e7b6c5794ae2308411f88e9e2bc1537ecf6c7dd6168ee4cb1a3677", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00082.tar", + "bytes": 284221440, + "sha256": "759003e820bf273d2a9ebbece38dfab02ba1f5ca5bf5454b63a651d36a273548", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00083.tar", + "bytes": 267776000, + "sha256": "bcfb704c377aa9459c4b38b1cf1746c1b78a12ce43cb2572f87cacb7563ab8c1", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00084.tar", + "bytes": 255211520, + "sha256": "3a1f9d376dd1009d764c8cbb6fe452310fe5d9372c3281726b1c54a154e236bb", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00085.tar", + "bytes": 284856320, + "sha256": "4e08faf876c4db20c598581815515e795da8dcbf89510a8bac7b735450209bde", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00086.tar", + "bytes": 276316160, + "sha256": "18055ff38d394c94b60eded15ebcff6ceef46bce30f9109d05087a55e6a1f4ef", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00087.tar", + "bytes": 292392960, + "sha256": "061208830ce73da2cc69cb1877a1bc791a10e02faaee122ebcfb09e6fcd0c57a", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00088.tar", + "bytes": 266444800, + "sha256": "2eccc99b93d8874e27df46ce4d174778bbcef5bb96774cd3edc833ef02d42df2", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00089.tar", + "bytes": 277524480, + "sha256": "bdc3c563069611df03a246221174ab70404a9d755b5233985bed9c0a5bd70bc4", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00090.tar", + "bytes": 277053440, + "sha256": "32641649a60488324da3746a36af41ae7abb9fb4fff84caaaee3f307e8f5ac3d", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00091.tar", + "bytes": 283934720, + "sha256": "25db63f66f0cde752eb5884327186e3123c9f34716f15857047adaa79f61485c", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00092.tar", + "bytes": 289515520, + "sha256": "d678fb00ae2098c83480c14aeaab525cecf21bd93b75a5beacf687566a28bb7a", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00093.tar", + "bytes": 264857600, + "sha256": "0073dee08e866ee6f4318a8ca17cbdc8111e11f8650984e31f2cfdfa09bf663e", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00094.tar", + "bytes": 276490240, + "sha256": "21129b891d9b7d82b931baadbdcf406ed3ef820084e921af75e8154e06fd9971", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00095.tar", + "bytes": 291491840, + "sha256": "4feaf812d1e1d06af170252a3a012082e048e9cd5b849ae8d223a73d6b06acb9", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00096.tar", + "bytes": 290201600, + "sha256": "07ca36c2ce6a4a24a82d176c3160ada90e33978f395f869cadc1af399cf94f9f", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00097.tar", + "bytes": 281241600, + "sha256": "89b04682a3790cd53cf2775eeaf972a08ee3e145c263c30e4fa5232af58666d9", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00098.tar", + "bytes": 280760320, + "sha256": "0c41cd207a51ca9880aa940f7f12a688b05c5566978b9994f88f9b39faf6ad84", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00099.tar", + "bytes": 288962560, + "sha256": "f99448d2ba325b9c32fca8b6382d1851c0b0d75afcf242eeed4c3adcc2841c3c", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00100.tar", + "bytes": 279296000, + "sha256": "e4fa467331a282a3a09094a01683affe417aeee0705b05503b81451c49391fb8", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00101.tar", + "bytes": 286515200, + "sha256": "9f1f136e765761834f832a2cc18e554f1d35a2d10b513d06f03ab76f14d97b0f", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00102.tar", + "bytes": 299304960, + "sha256": "60c51effd15ccdf0bb1b283e5e38494b12c3a2d834425ba110b64e2c4a5c2dc7", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00103.tar", + "bytes": 279971840, + "sha256": "5a98ce312e7bd88625963004686f168fa805d9956f7b8c216ff0957e7817aa1e", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00104.tar", + "bytes": 301414400, + "sha256": "9e2348ce5e373082a42a8b9d51012eb8a6a625bc1e23a3ee347d53e8357a394c", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00105.tar", + "bytes": 282716160, + "sha256": "9405ee21016251e2a211a1014a7698c76b5be7e296fa5b76edd061470a6d459b", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00106.tar", + "bytes": 280524800, + "sha256": "42a95cadb60db431ac757742443ca4de1b10b56b9a8c746d9adb79e1afd20b5a", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00107.tar", + "bytes": 282992640, + "sha256": "45f04df408c9a3c9c2daa92bdc753efbb692462af7167e8b7056428d6af7a69c", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00108.tar", + "bytes": 260546560, + "sha256": "6c7f805b6f2f92765f6af232b58029d63feb95f9fcb2c16399e29bc32744b775", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00109.tar", + "bytes": 290252800, + "sha256": "9d4fe83efe03d3374306cf19de33aaebde79dc98ccef491ab22b922583ea9f89", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00110.tar", + "bytes": 292003840, + "sha256": "fb59c9d2a792d52c1be4204e1a7857d87913b5eb3b563a190dd2254f4e65d95e", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00111.tar", + "bytes": 277944320, + "sha256": "1582c4da1ecd79d7918524be9ebf857c15428ddef17b8869d7011f778d4fe667", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00112.tar", + "bytes": 279562240, + "sha256": "71f4397556ab9ab0bbfb0cd740817fc10e1870c263700b67db0922355c6552f0", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00113.tar", + "bytes": 269701120, + "sha256": "28c05dd161033300276641dd3b65403f33d0c871d9d17a7f6f0dd30e4ea7beb8", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00114.tar", + "bytes": 269649920, + "sha256": "5aae80b15aa72a65da102d3597edda2b02a268878afbf0b108a7cfa353ad683a", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00115.tar", + "bytes": 284631040, + "sha256": "bdcfa534afb3f20e37b00ef501dcf73a0afddd1c52503b5ee43e6a125dc001a9", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00116.tar", + "bytes": 288358400, + "sha256": "c7e2e8afea2614dfa323da6da31fc233a43b7e39cc9d463f3d61827f2c2f2ce3", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00117.tar", + "bytes": 275343360, + "sha256": "cc8976cdb007f204260881f20027f81a445eb88723f9d8ffe9271515217b52cd", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00118.tar", + "bytes": 278272000, + "sha256": "ab5d146518709809ab81d521e3e41d2077a412416e78b4b48e5261ff0780897b", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00119.tar", + "bytes": 294819840, + "sha256": "6477bb669bd5a6afc7834a517605bfc2f20b8a7b5ebc29860b9f2834e86430f6", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00120.tar", + "bytes": 291215360, + "sha256": "7deb87c945fcd2c5314586bea797cc77a8b4a5aa6cc1a26079f33ec3f2a1eec0", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00121.tar", + "bytes": 259624960, + "sha256": "1329f2774b66193d1b8dd8b33657aacbfa34da7f7afcf5222545078e6f67fa26", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00122.tar", + "bytes": 282726400, + "sha256": "3302eb76b038b71b8bab8b88528069cec30eaa0b8598fc199bd5e30e8c8756ff", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00123.tar", + "bytes": 288491520, + "sha256": "4e51262565f40137604bf0ac3161600b75ae01e859f714143227afa2a0ab46b2", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00124.tar", + "bytes": 287518720, + "sha256": "66b3bf90c06e2fb4a47856c5c49ba226961b2d19ee65d02037009435f8a2e9de", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00125.tar", + "bytes": 287324160, + "sha256": "b97d2865d431bcbc222731c639ef14c263d7a157ed8c104c6428ac1688a427b1", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00126.tar", + "bytes": 294072320, + "sha256": "74bd78fbc027156d222692bbeed8f935c5f9e039bdfc20fe3cea2eb8601a2b5d", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_motion/00127.tar", + "bytes": 286750720, + "sha256": "51c9b70a50f9b614650488db532bbf6b206f2c1b8652ecdce072894417aa7f1a", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00000.tar", + "bytes": 164290560, + "sha256": "c77225960a99253727a5ddd7e8f107c20193a97c869efe13cac988d0754f0049", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00001.tar", + "bytes": 150241280, + "sha256": "960665290e76038d2332fbe5c51791a5afe3ffa6fede75955fb5e0cd51fef310", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00002.tar", + "bytes": 167567360, + "sha256": "f17d5e7fc901d933bef30c20d2a471287a0fc5f72c378af12e4f20632634d684", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00003.tar", + "bytes": 165683200, + "sha256": "6ffb6166b1b645d56978641e8c47769aaa700d0402503a9cd71bdee61a67beb5", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00004.tar", + "bytes": 162918400, + "sha256": "4656b1450fe8d88be36b42c0ce303e5747090769fb67ee190fb7dab7b07b62e7", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00005.tar", + "bytes": 155648000, + "sha256": "df665aa23f1c93fd0b901bff5168ab410e638fe440066620ab151aff18dbfebf", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00006.tar", + "bytes": 150528000, + "sha256": "c48246cdbf9dfaa2519dd96b1dd6fde6314261da181425ccc60a6145fb6bdfb4", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00007.tar", + "bytes": 171745280, + "sha256": "2ed728c78fd0ed1a2e7af4a11fcd45c1d96504c54d8ba96ad90dd23a01cf4902", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00008.tar", + "bytes": 161208320, + "sha256": "fca0c18e9a0e898aca139baebfb2fdc72d7a87ba1da57be400beeb75adeec852", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00009.tar", + "bytes": 157368320, + "sha256": "33beb447513d524f1a48865304a2b0af6aae2a81b903c206b5922ceb487a5229", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00010.tar", + "bytes": 157071360, + "sha256": "82597e5016c3ce7c533f7669ee2c77f3f7a01e71c7a6bdf3651f0e0c364f231f", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00011.tar", + "bytes": 158935040, + "sha256": "09e566d83072ff391c169306f5495095ed4e4bcf24140e02e830a6f1ac09ba3f", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00012.tar", + "bytes": 164433920, + "sha256": "d170c787805c095addf204a88270e5189c02b17aa1be971ec291a4397cba3f40", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00013.tar", + "bytes": 163850240, + "sha256": "3737a7912ddee59210f17da6e8cf71f131c036cad41f83166de29b3668ae281b", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00014.tar", + "bytes": 171161600, + "sha256": "d60e45433d3fc0205678a2b89d58db415f69137c47ed9cf240e1a4749048f95a", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00015.tar", + "bytes": 150548480, + "sha256": "d9c736c22c2808f20bf33b7d436da4b23dd7584ee2cb0b2a7d468fbd0853c989", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00016.tar", + "bytes": 159754240, + "sha256": "0c37436f68827291643ff722738608f0929b6b491e8c43b32dae48ba139868bf", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00017.tar", + "bytes": 157214720, + "sha256": "bc95fe2f4cb88b1edd3a1d80c75a41fe0afad7f6054df2d7d7d80ea1b52b522a", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00018.tar", + "bytes": 161863680, + "sha256": "ed37807c0fec627f8259120357808fa9b2c0c91ca8994c29f3865aa5e440e79c", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00019.tar", + "bytes": 163983360, + "sha256": "f38813b8a5b3236b4de71942777bbd7112f9d0b37cf93a0205e71a8115f49c17", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00020.tar", + "bytes": 159047680, + "sha256": "ce2f61e2f62575e4961e59b23de82c1bd05a3563f7b4b090f510e3636e543ad3", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00021.tar", + "bytes": 164198400, + "sha256": "3def28eb4b3aae483f3f4905436c2c0dd4dcf5fc64fce722953fd375aa3b7582", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00022.tar", + "bytes": 172072960, + "sha256": "0050b53039d71629e4379ca2cf24e5293f412a117f45a0ae857521afc5e7d740", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00023.tar", + "bytes": 161751040, + "sha256": "7fcf9bb96cb4cd49774be2d88e2c407b23e26faddb225147d308a68cee24a361", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00024.tar", + "bytes": 158924800, + "sha256": "0e92301b73128a8cfddf0ee7fe38f426dffe569521af6e792ea02ae06154d4b9", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00025.tar", + "bytes": 154142720, + "sha256": "65e4d5933ced1d2fdf303329e497deb8cf0c4a6745f4bfc6d32f52d373747bfa", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00026.tar", + "bytes": 159375360, + "sha256": "0de868ecad5293e0e4300baf7c53265fe0fa2071ac0abfe31519a56931842118", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00027.tar", + "bytes": 160174080, + "sha256": "fabe11361ca368ec0f9c211561cf574adad08c8e23e0b16d431844d402f541f8", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00028.tar", + "bytes": 147107840, + "sha256": "472ca80c586fe59df26c1f2b3f836143fbdbac90507dcd81205da30b0c963b12", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00029.tar", + "bytes": 152893440, + "sha256": "82d940ab1e86a8aa92901ea85473a0a2bd7e0512a62b8c7673dc5719582cecc5", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00030.tar", + "bytes": 171315200, + "sha256": "10740e1af2cbab75b085253e1c5e4b73090eaafccea745792c2309d1e04c00d3", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00031.tar", + "bytes": 156631040, + "sha256": "cfb780400cefbf2f4f92ed5a7978bfaf58083d85b8487dc3e237147171a80bbc", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00032.tar", + "bytes": 157952000, + "sha256": "10c4f9c2efebe694ac19cbe99ee67a890088e63e46dcc49e2c9f3781b4c0e273", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00033.tar", + "bytes": 157870080, + "sha256": "5c0a1159d1a522ab2a0cd7be5434a35418d92ea73b459cc475db99962c44799c", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00034.tar", + "bytes": 144732160, + "sha256": "d6b81f265f94657d75a20c3dca94fb4d2a9127cee0ac6cad24d332c1440da995", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00035.tar", + "bytes": 167700480, + "sha256": "11e260f2b74c21e6403e7c666451ed0f4cf3ced3a73a0d62bed5693b524bb29c", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00036.tar", + "bytes": 166092800, + "sha256": "a03520957193aaa5757761b270b96fdbbb77d7537e1a75efb8686f11a39a7a41", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00037.tar", + "bytes": 164208640, + "sha256": "1967cf13849cc87c2481c3de38af8bfb8d8f40bc80cf4ab6b051e07a1e10b74c", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00038.tar", + "bytes": 147240960, + "sha256": "0028fae70cf71c15a14b72116ad55f62fd84750f5725bcfac4b384c6dd225100", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00039.tar", + "bytes": 163819520, + "sha256": "9ddfd341bebf2c0e29870f7939fd76ddec4583729355994686189f504ad412e9", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00040.tar", + "bytes": 160737280, + "sha256": "85e3b584987c1ff99ac219d065e279d0f42ab9cd7f803df72214655c7fd07c8c", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00041.tar", + "bytes": 166164480, + "sha256": "92483966167a025ae5dbd419e5518930c9a85c25fd72680200e08b932f64f5ea", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00042.tar", + "bytes": 157050880, + "sha256": "88e41948d7e4687ec8602042263fdb0c6f0562328847e94d9af83f8fa2c27e0d", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00043.tar", + "bytes": 162099200, + "sha256": "83ed7365600362b947e179ce6b2f0002a0f5a20398ed65a61e61331c3cc63b39", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00044.tar", + "bytes": 163901440, + "sha256": "b21e34892ff735418aec1cc9d4197bc4026ec18f09b7b9ab23d9e969e59a7b4b", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00045.tar", + "bytes": 155504640, + "sha256": "81ba40d03fe239a51d88d7335c3cbf059227b6b79dadf6cb80192b20e204195b", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00046.tar", + "bytes": 167761920, + "sha256": "4f9de85ae5aac1bc7a043e7b61beaeab0434ac31d0a2b2242151dca50f18d33a", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00047.tar", + "bytes": 167116800, + "sha256": "026e8d87642ea5a89481f64d28eb6a07c66cc4c018664b7e5a2a48f3281f7ff9", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00048.tar", + "bytes": 163348480, + "sha256": "cf8a9058fb02c9ee3ecbe620a4968d8a617b422f66e9cfd9ece864fd4ee96a01", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00049.tar", + "bytes": 161996800, + "sha256": "708cdc87d3b99eafe57c73c2d3604848f482d7a2946efaf2668210462527b126", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00050.tar", + "bytes": 158453760, + "sha256": "e56f829fe85b7ede4c88cb2a4c7d4955fefaba31a7bf0d30ee288f57bbdfb978", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00051.tar", + "bytes": 165242880, + "sha256": "85d1d8662f80b0eb46e5e56f49c20feefc44fd3c7d3dd259492888138ee20ac2", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00052.tar", + "bytes": 148664320, + "sha256": "82440c6f49e7aaeb3c8ec6cead9e3045dd03ef6de5d214a8f9214fda5dcb5565", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00053.tar", + "bytes": 167290880, + "sha256": "f06bfe33fd661d251c2b29e2b671f9dd0aba941f0e930ef92da7a3f21dd2c256", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00054.tar", + "bytes": 162078720, + "sha256": "5c90166e4a803747a4c7277299b34bb91a4f2e71af6aa766d518350ea0bd89da", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00055.tar", + "bytes": 146063360, + "sha256": "4b5004e73ea05115c2cbb3597827484dd647531f94b5e6400f30ae9c21b2c4ff", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00056.tar", + "bytes": 172247040, + "sha256": "9c005b42393d3569affd484d86b8f9034e5401248ccf15f578453043f7880d75", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00057.tar", + "bytes": 167854080, + "sha256": "7a791a918b32a7f2ea30d77847a3449e7b8d90fb17014ad9fe4e06911f5ea6e8", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00058.tar", + "bytes": 174305280, + "sha256": "6e8c4ba6e4c4187dad42594e71194c6d734b49e40d92445329c6e90e5b264bcd", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00059.tar", + "bytes": 159283200, + "sha256": "3ebea086764145ae33bb48f6e8861e8ab083e05bf270afecf7bf85539e625e70", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00060.tar", + "bytes": 157429760, + "sha256": "beeab8f7ec52274f42e0b78df40b28dc96b8e9987ba4123ebfcf6d862e20437e", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00061.tar", + "bytes": 154019840, + "sha256": "2fb05a24c70b2d0b5c4c48133702b703c445af7d21696e341acdbbb23ccac4bf", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00062.tar", + "bytes": 173639680, + "sha256": "9645e137db730ab098f096af1db2cce8a60a8f4da905cc05f250f5e65d694b1c", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00063.tar", + "bytes": 155883520, + "sha256": "d8f320627e7269ae17197a758b722bda249325f3414fcc9777bc3b2a2c9a4b63", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00064.tar", + "bytes": 157050880, + "sha256": "43bb73a9c7ab85b2846ae44dc5fe42c222b922cbbbea31de4eace1fbad28e108", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00065.tar", + "bytes": 162775040, + "sha256": "35c579ccb9b575669f077e7b488c0b53960adcec60fe02bbf9ce2cd4b1c90c2d", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00066.tar", + "bytes": 158064640, + "sha256": "c64d585c47ad7f184219e777aeff69b5386487c72cd0b4656343226d8b4b6437", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00067.tar", + "bytes": 142080000, + "sha256": "79dcb210eed238622e7e07c1f900069d2c1c0a8d651499bb4ee0358f312eb294", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00068.tar", + "bytes": 178155520, + "sha256": "26ab96af081305870964e097c816ca5a9e8b673543fa5fd3d91a212bb44fef28", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00069.tar", + "bytes": 163205120, + "sha256": "affea53d462505153b58a66b4ced7cf23f075f454444405147e0beb5ae7f3e92", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00070.tar", + "bytes": 165478400, + "sha256": "dc07436d8a91687082585af00363ee79ce5f43339f5a018a88b4584736072dbe", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00071.tar", + "bytes": 149780480, + "sha256": "56b5ecd195c2d53849c09274d54acdecb1d1327761c73f5f9ed84281897ac34e", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00072.tar", + "bytes": 168448000, + "sha256": "7e8885e9625f362f90d98366afea21f48bc97124cf02ac459525c10ea31c2f8b", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00073.tar", + "bytes": 164567040, + "sha256": "8bde5e66bd9434f6f7e86c62f706b553a73b6a09aa76cc78320a9455c5469fcd", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00074.tar", + "bytes": 167270400, + "sha256": "7bd167ed7793d0b4ca8953bd175ca1c542e8ee4511387b7f679a3794fcdf1d49", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00075.tar", + "bytes": 157716480, + "sha256": "74ccb501400795638eac256e483a1303f6cea1f769a15e8d343fefaf06828867", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00076.tar", + "bytes": 152330240, + "sha256": "8a516b737aa571e9adc0c3fee766cbd96252b0dfa090545d2cd0c1ca7f8102a6", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00077.tar", + "bytes": 163164160, + "sha256": "445822226c2dd04597de0c231fad4de6a279bc58efa2fdcd6aee48ef017431dd", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00078.tar", + "bytes": 162304000, + "sha256": "52439b22771d2f2fe1546afaba22b34f7080aec3a935c1343c13bd8a04c92cc6", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00079.tar", + "bytes": 152033280, + "sha256": "a88eb8cb09c3dd3128dfacf8e29202af21abeed9ebb283974dc4921632b5e18b", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00080.tar", + "bytes": 168478720, + "sha256": "22ff13ad98d3db2cfc79f3c306aa4f3dc915e2baeddfe76d76c57e1ecdf13bf9", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00081.tar", + "bytes": 158740480, + "sha256": "1d7d8c2d75cdaf62d4e652deb3a840ae6187a8b29f5e9d3d1280afa5fc276f0f", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00082.tar", + "bytes": 154572800, + "sha256": "75d76380e58a10e5698dedea5f524868b18be5272e63db8665fde9dba62ce961", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00083.tar", + "bytes": 155320320, + "sha256": "9f1fde4275b33d94e0bd826aa7cdcd2a980f28b0786ae213b67a7b0a757f1d5d", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00084.tar", + "bytes": 169984000, + "sha256": "0bf9b2320f4caecf91c5d3774e936ef5cec363f80947516d82eda50c4f9800bc", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00085.tar", + "bytes": 160972800, + "sha256": "cb383310fdcc234b9080685498c56ca2c510186ae0716ea81e97e11631fdf162", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00086.tar", + "bytes": 160337920, + "sha256": "add1d09b7dd494b77ff1fc24c36892369c6955bdcc5dedb8d4255e9313156510", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00087.tar", + "bytes": 139857920, + "sha256": "c512036cae87dbfc26558bcd40009639e12ad2797826933c368030e8b42c1470", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00088.tar", + "bytes": 159764480, + "sha256": "7d8123f241eaa5f8ba3784f1fb119126e006fb8d2a014747c7c3f7e9215d98f0", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00089.tar", + "bytes": 163870720, + "sha256": "3423b19b76baf6297c4e411fa952548213c021c13fd2106833e3caa3e1770b9c", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00090.tar", + "bytes": 155975680, + "sha256": "a30489f2fd1b8d3c3af87e1df341a5fdb623f3fe41e48b67c52537a9aa76f1c7", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00091.tar", + "bytes": 175493120, + "sha256": "7ee8197f163e96a15958dbc963b545ace07c0496c43847f74838ef20fea2542b", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00092.tar", + "bytes": 153466880, + "sha256": "de8defa5a90437c6302c5db4116ccc53acdb319819deb5b0ccaad087ef706aee", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00093.tar", + "bytes": 154664960, + "sha256": "3bb4ce0c9d802fc3a053cc04b0d2b94752e146c4acc928485c0543890d8e912b", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00094.tar", + "bytes": 160665600, + "sha256": "796bf0b686988ce703cbd88d727a1f36239afcdc3efea713c6381bf9d03459d3", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00095.tar", + "bytes": 150722560, + "sha256": "ec2f5ea4f70050877465aa739f7e6c55afb4682f37c54c3cd035022b2173f95d", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00096.tar", + "bytes": 169891840, + "sha256": "ff768278abdc0b4264213d373ecf48d53411b412383fcfdebaec2ac7a9b1e631", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00097.tar", + "bytes": 161064960, + "sha256": "b30aba94f12abe0c33289b553b68cf9768bfbe2cf523dded564e13dcb676ea41", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00098.tar", + "bytes": 169584640, + "sha256": "0ebffb2371584be72009a17a2d2d862a0925e12a574b1b78a2005fe2ddfc7271", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00099.tar", + "bytes": 151449600, + "sha256": "7ccab2fd5a83b2f27d03dc42b53c41265ff92c935cf1124bbc7636cf7092da5d", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00100.tar", + "bytes": 171427840, + "sha256": "80b3846c2d36f6d553046ed754c917c3a439d20cc03f96e70b625d356bfc1700", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00101.tar", + "bytes": 161730560, + "sha256": "44fbd0d33d6677ac32e7733d9fe85bab1a1721c49e0cb821217a1ec05d00d9c0", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00102.tar", + "bytes": 166195200, + "sha256": "d2bde9bcf2be9dec45acb0b9832fce6ce0cd110a404d9163dbc133eeca4b163d", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00103.tar", + "bytes": 160000000, + "sha256": "3bbd14328ebc00579004247032b9db6a8ca291a389ae5800a76a027b0f3f2ef3", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00104.tar", + "bytes": 166789120, + "sha256": "2b09632d4e6ddf06b007f14d5b38ef14c50ae17c49f960e7d12527bfcf1607eb", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00105.tar", + "bytes": 168908800, + "sha256": "03d9900c9197b641fc63df498f8a0f4bc083aa393d2f46e2fa8d16a532a53f12", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00106.tar", + "bytes": 176056320, + "sha256": "2024bb5757a9c0517a387a5eb638f8e2a54fa2628e4ff7bbb3bbf4b43beef1d1", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00107.tar", + "bytes": 170577920, + "sha256": "adac2af366efd39aaa1adf21abf4a86e199b2fb1cf61f7ad6548b5919b4354e2", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00108.tar", + "bytes": 153702400, + "sha256": "d97e07f76068e0c4f50c06db41a0100ef72c15bd391ce8c8f388310c5c1c8498", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00109.tar", + "bytes": 166205440, + "sha256": "b8572df590f8846e4499cf4c163ad72a7a7d8c8a6bd6227e074135b3082ff19a", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00110.tar", + "bytes": 169861120, + "sha256": "24d9748947b6c0d68cf626c448fd2570a913d504b2b479bcca4b9fec1ef83d2b", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00111.tar", + "bytes": 156497920, + "sha256": "de034737374beab7141048d5a6d7bca589a3e965b40307a59c62d2bfca077f35", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00112.tar", + "bytes": 169297920, + "sha256": "342f203aeecac3ef6e6824e5bcf6d3157633cb6b1e99e2a4ed1a12ca6409c039", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00113.tar", + "bytes": 158986240, + "sha256": "411cd69678a545e3522dcb8e93e9fe1f12cbd7852edcc63b08b5dad5f90fefa6", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00114.tar", + "bytes": 158904320, + "sha256": "82ab9f50d717692125f013679d568684dd7d900533c7168d65d56f2af5de5b1c", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00115.tar", + "bytes": 167526400, + "sha256": "d94bc16522683787acc2b1cc201f262be611ccfa5a57a0e03f2178e099f4f9c3", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00116.tar", + "bytes": 163553280, + "sha256": "30131142e712ca3d36562bbf20db1a353e081ce6279a59d3dcf5cf6f3a600205", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00117.tar", + "bytes": 158289920, + "sha256": "b0566cd2fac2aa22fe3f2ac986c944cc2fae4330e04a357e1cb6d0ca2cfbf1ef", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00118.tar", + "bytes": 160368640, + "sha256": "d0005dcb1ed8618e9bed50820f947e0e5cf361342ef28d3fcbb95e373e66d3d7", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00119.tar", + "bytes": 163563520, + "sha256": "4a066e2ff705a63af50185ee15f6988089aaf22529481b56cb91d3526cd7ad2e", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00120.tar", + "bytes": 160133120, + "sha256": "302f21eb04a633d6ad4247e69301ecae3b809e3809406e2fb81576d7ddf039f6", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00121.tar", + "bytes": 167505920, + "sha256": "e385f363a274e36708b4410fccc95fce239164fc11267222fefd2ad0fe675423", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00122.tar", + "bytes": 167014400, + "sha256": "e79d43633fac92569bb35a84f58a8c40cb4b4ed1e91f7697a1f03175fbb9c18e", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00123.tar", + "bytes": 171509760, + "sha256": "ed5c7bee0e4c1660025bd84038d5ddeddd3962a92984d94af78a0dd8012a720a", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00124.tar", + "bytes": 162191360, + "sha256": "973accb6920a8d95328c654e739c7877996e57b37c3ca9494be98b9e90962f59", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00125.tar", + "bytes": 171919360, + "sha256": "c19aeddc48d32aef1ba0c5617a73e8a236dd26645c226e50fec6f2443c487d44", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00126.tar", + "bytes": 168693760, + "sha256": "44193111236847a56910119f53d30909c7f16ee37be777ef180c789c9db0d7ff", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_smplx/00127.tar", + "bytes": 155770880, + "sha256": "08838d1b699443124f18283f627f8fc9c7479af2a187a54831c81963840ecc3a", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00000.tar", + "bytes": 59033600, + "sha256": "74dd45dcee765df943a308021ca170b9d9eb31fcf85fddebe085158b426c72c8", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00001.tar", + "bytes": 47534080, + "sha256": "9fb9a346cc7e0db05721d047e00c2407b827951180106f0435e7ae1dfb8bad17", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00002.tar", + "bytes": 51998720, + "sha256": "e54afc9948f05075654b4560f9c79c16393509ec7bac376ccdfbac1e64659985", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00003.tar", + "bytes": 47114240, + "sha256": "ddbca480026e5eec491ee2c2610c3d78e6787ac9b2b36b9f02c9a1dc2a223380", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00004.tar", + "bytes": 51138560, + "sha256": "cff147bddc1dbe96722e1161409c14cd8dcf6a2e2954495f8d97c0c217460f40", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00005.tar", + "bytes": 55336960, + "sha256": "4b7f54ee0b8edcd4ef71a8043239b5d4b00ec55183ae9f83129e0a9e267ac130", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00006.tar", + "bytes": 48916480, + "sha256": "dc4fa2b9ff03f3ed1e0a4cf43dd5b3516e769b84e25aedba7622067ce5d0ea50", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00007.tar", + "bytes": 53114880, + "sha256": "90324f53da52c3217f5fde695b9984be6e96707a0786ae04a8c6f7c498e5aa75", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00008.tar", + "bytes": 56145920, + "sha256": "d5154d53ffd2d1b02ac008db5b80e5c3829ed7ac755cf71650f3f852a96bcd4d", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00009.tar", + "bytes": 56279040, + "sha256": "e7630ba7f5face89f34c43bcfd4cb16d2208019c99568f113e4a3d89bdb58983", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00010.tar", + "bytes": 49131520, + "sha256": "4c48029287a97ca3a1f616c88c8867ca6fe2d794b550ad7bfbd2320d1ee25a77", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00011.tar", + "bytes": 54609920, + "sha256": "a731bce23d5a8d18ea94abcf98af0dbc5276e473301ce76f53a3c7c8a4b4464b", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00012.tar", + "bytes": 54599680, + "sha256": "5d31941eefe07192883f3736264a4ef8792c2b4f0e1485c6cb2a1918ff5e616f", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00013.tar", + "bytes": 86292480, + "sha256": "72010e76e654f6859153dd715bed3b1e8fec8f37d76f5b0efe43a406318280ea", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00014.tar", + "bytes": 52756480, + "sha256": "9b675bc2a3e3b1c21543d3c568efe73e74a585d9fa3e24a9e1d2d317ff8e1a9c", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00015.tar", + "bytes": 57425920, + "sha256": "8c17135a60cd30632d4f80e5ed07b33bf1d421665f3b501b2d5e3979a1765b1c", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00016.tar", + "bytes": 56145920, + "sha256": "42d95f56c4c01112327fe6552780c2322b076a35ddca30bee708e3e2f053d703", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00017.tar", + "bytes": 54661120, + "sha256": "ae0f52be58fece7479d5926e29643b106f000457252a4380effcb7f14b5e21e8", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00018.tar", + "bytes": 50083840, + "sha256": "1273391423362fe33d9719d9ca8c3e544e6177d3420c741614708fce77c43c91", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00019.tar", + "bytes": 45465600, + "sha256": "1abff838fcc753fc90cee6b4717b021f3bcac23be6da99e0f521a3d09fce354f", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00020.tar", + "bytes": 54333440, + "sha256": "5f1993de90570a4699f9c4a9fd5e440640cb6ee0e5dbd21003aa99271cdd22d4", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00021.tar", + "bytes": 48250880, + "sha256": "0456e0597d990329dfd102d19075d7203355e5a821a1e7a3f4d60cd5eda5af18", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00022.tar", + "bytes": 53954560, + "sha256": "05b017325742e44dfefaad2903ae2ab3583621362a2856861e37b33772af018f", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00023.tar", + "bytes": 47114240, + "sha256": "6e5ccefaede09e1b141aa60e87aa8a8b6694cc9abdd299ac7d1984832dd2ab1b", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00024.tar", + "bytes": 49725440, + "sha256": "35a362f1ae5b632794123d8e7c4886f68ed95b5ba459b30e749060101ee9ab9f", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00025.tar", + "bytes": 54343680, + "sha256": "1652b35a9ae15ca6cc0c73ece227ec4bc512c166c506f167c67e1aa39a226b80", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00026.tar", + "bytes": 58327040, + "sha256": "b305e28ac72abc4d6801d8770645162d82bc4e83811df69a3e935813e1a1e92b", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00027.tar", + "bytes": 56565760, + "sha256": "b73fca2780f218c806b19551288526eae127d5ffc9537c6c7c8bf6d5d00de409", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00028.tar", + "bytes": 54251520, + "sha256": "63ec5f57c6207a27d948367d5714a97773c04941b7ce6199466a3a0008a6e585", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00029.tar", + "bytes": 49418240, + "sha256": "77b29b13385df8f2c28f773f812b2fe3b9d147d9f4459a7c0506c07a60643abb", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00030.tar", + "bytes": 48906240, + "sha256": "0ef894eb7429443b79c2743a8dad8609038c88e0247df3a7068ac7848cffa428", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00031.tar", + "bytes": 45301760, + "sha256": "57347291516a8e276680b4fd695cf19a57b962bbc99c84f2b2d52d75ebd16448", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00032.tar", + "bytes": 51855360, + "sha256": "e54459237d214b7298473f0156e76137ea586e210e5fb2cc5b700ed47b3e7b35", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00033.tar", + "bytes": 56555520, + "sha256": "76875967d8d6713cdbdb24b872a3df9f269f5fe686108ecfc718ea2be297784f", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00034.tar", + "bytes": 50247680, + "sha256": "ad954ab2591d97f8e8c0932ea2479014788451ee8e96be4697743f84b2314241", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00035.tar", + "bytes": 51200000, + "sha256": "341b0c29f04124578b13e24691d3daebf11ad771f3958b77297cde6508bc0b13", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00036.tar", + "bytes": 57896960, + "sha256": "f7c9fa30c22e30b6902b47536e3dcb74c891ba43642e487bebe211e4948682f2", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00037.tar", + "bytes": 52305920, + "sha256": "0859ebf47f55922cbef5c501f7dd918b7a5b658a9d08c0145b70570c3bef6f62", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00038.tar", + "bytes": 58736640, + "sha256": "23faa3b19f5b886def1d1c65693b079aa9fd4a5a80bc1e31bc57bd6009cd9dfc", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00039.tar", + "bytes": 54077440, + "sha256": "ab0dbcf1f6dc9ab1b3c7ae72cb7c6b152ffd749cf1c41d9324ca36231167f364", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00040.tar", + "bytes": 54108160, + "sha256": "4eb4e163ee87039e6e8a8048f67b545b20f10351cfa48729b2d38915c9944203", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00041.tar", + "bytes": 57600000, + "sha256": "e93710e5908776b8d2dbcae4e9e82b72c89f90c37633b75a260de5212ab7e8eb", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00042.tar", + "bytes": 49356800, + "sha256": "bbe4cbe5c1bc30914b6bb1bd7d55c267ec8fcfd2e5e2a97dc623cb974496f9f9", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00043.tar", + "bytes": 49203200, + "sha256": "700d187b6ea9c009c6254b42bfaa9cfe85a93d9c33ae5050221fba5841d4019c", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00044.tar", + "bytes": 56606720, + "sha256": "55eb29f0057cb3fce6519f3baa163ec3ca7fd5fc09d7bec125a7c7b8a205eca5", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00045.tar", + "bytes": 49643520, + "sha256": "303d7eb90da7a292e9b913fe04cb2a9e6ee59bc0c85168f52d0d5245ec7b588f", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00046.tar", + "bytes": 47820800, + "sha256": "9afe2f2b3bf5225a36731a27abbc6b357af749cb7f7af5fb5a0b1bd9f038282f", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00047.tar", + "bytes": 54548480, + "sha256": "55813aa2fd12b05e4078179ca411ee2127f8aa028033607708cd8a711a46ab11", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00048.tar", + "bytes": 48619520, + "sha256": "ff0d7cac73a2ca7a7473da2d1ba3d64d8a9916664e8023debcbf51b3e451461a", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00049.tar", + "bytes": 52090880, + "sha256": "fea504532379bdd32a87fdf131e280bc7850387c1307b532866c4b5c1d25d288", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00050.tar", + "bytes": 51875840, + "sha256": "8815ebd38609722c851595a479c502c1e75bf055801121d705c8e01797b7a12b", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00051.tar", + "bytes": 51671040, + "sha256": "3650a8b24beb6a623652d083d522fbf78035084e5dec33440bdb4b298267d702", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00052.tar", + "bytes": 54517760, + "sha256": "517338a96ab01522cf1129002e41993082e68f4c6847c7ea31f04bf5d091e0e1", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00053.tar", + "bytes": 54988800, + "sha256": "a5c836dd69583145d576065ee450c0a388a1038bfb5e0e51e54e146bea132adb", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00054.tar", + "bytes": 50944000, + "sha256": "117fc6882c647aa27df805684fe2970ed42669d47f123b54bcd62776bea1483b", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00055.tar", + "bytes": 51988480, + "sha256": "57d7af449e2dd846d238e7d0159e1ec01f627415f1fbba848ac7b0fde5c1fc29", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00056.tar", + "bytes": 52541440, + "sha256": "7b8fc71a69fea209648085a35f75671c9941472f5b3ab6a373ed3aff45e7d69d", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00057.tar", + "bytes": 51978240, + "sha256": "a47e1b676044274dc5125658acbaf20056be7836cc75f45deb4c5d3afe2c59d3", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00058.tar", + "bytes": 54364160, + "sha256": "973f577f709b9a901b6e5943a792757b8be2631e68b9abef9baf53237ce667c1", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00059.tar", + "bytes": 56524800, + "sha256": "9eab4a7b4d3bd7dad796b5698c8338b5dffb7c30d240e70dce5c4e20ccd48feb", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00060.tar", + "bytes": 55080960, + "sha256": "b54606ba8fe9c134a984a3411705b88eaa0e593e385425bd998160c0c164668d", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00061.tar", + "bytes": 52080640, + "sha256": "c7347031a33fb75f1a1a3afcff6fa4abeabed75256986a157d0b20074bab591f", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00062.tar", + "bytes": 54579200, + "sha256": "25c219d466bb484797a78081a8d1bad3b710b12d803445a89ea1106e24a07a30", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00063.tar", + "bytes": 54323200, + "sha256": "74b3b80f0a1bd4f7de92318faab0f7cda761bb71a4b3e68bc1d29168f725d0df", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00064.tar", + "bytes": 49930240, + "sha256": "bc94ad73c4602fa6f27796db386fd0a52b811066a81f6bd29b83093d38dc2c79", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00065.tar", + "bytes": 53923840, + "sha256": "0a62f269c76b7348b42bcfd4b1436c686a0f4484329d13440a852569691c5058", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00066.tar", + "bytes": 55961600, + "sha256": "16a6d1564af649c31b1f1639fa841d1f32381de72b19ba3b1bbff571abe2f110", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00067.tar", + "bytes": 53401600, + "sha256": "54ec30153c1ebb815cafb05d10391cb1553fd723c8a7d98818f42c4e96cd4604", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00068.tar", + "bytes": 50616320, + "sha256": "271e0e676f23377701c89fad30328118bf9921d46de46ee0f5f2da32fc0b6752", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00069.tar", + "bytes": 53975040, + "sha256": "12f5d299ac387e56fa02b2841c8c3fb4b6d0b585f2408e15f871030c0badb208", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00070.tar", + "bytes": 57600000, + "sha256": "c328736a3fae0d369d2bba60fd7dec1ad975473657899bdb5533dd108981338b", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00071.tar", + "bytes": 61265920, + "sha256": "45398199c8bcfd051ef038ba8dd75c5ed9e5ae11256f3e32daaac6fb07ae6ee3", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00072.tar", + "bytes": 51066880, + "sha256": "b66f520eb88a4be2942ce7cf765323963109fe4b1a7b57ba228c5c142a85f343", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00073.tar", + "bytes": 50606080, + "sha256": "c338ab11e33e3691528dc5b7382058f73743bec119c0b97a8aa2e5ba459048d1", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00074.tar", + "bytes": 51968000, + "sha256": "5e1b5897177c61683b9fb82dec4762cc68b50c802b3f53e9a24e2e27a8271fc3", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00075.tar", + "bytes": 48271360, + "sha256": "1b0e1504b68d7548854dfb3ef6971f435c9aecafde5875e669c684215ab917d8", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00076.tar", + "bytes": 50933760, + "sha256": "c3845668760e7f8107aa2b83cbc302be09c8cbf4d89c92db85292b17d04d329b", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00077.tar", + "bytes": 50513920, + "sha256": "c1a0e9bc3233ee56493d228867512c2d8babbca4e7c48903a848c3fd67fcc217", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00078.tar", + "bytes": 57518080, + "sha256": "eabb0dd8651cade80aba3d1b3c79271a30ee78a8bf4fe363cdbb1b8f6dee6380", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00079.tar", + "bytes": 52654080, + "sha256": "abf502a2f0f535e8b9f7df374b14e825fe4e8d70df5aef84354602dfe213eb35", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00080.tar", + "bytes": 54763520, + "sha256": "37eddc1d861b7edeac0b4393534e0fb6ae910654f01c88833691456509fd25e7", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00081.tar", + "bytes": 51363840, + "sha256": "53c06931d68dc82136069f3d70096d02dd13c1991118d169f2bfb3b3d6d9ffe2", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00082.tar", + "bytes": 53299200, + "sha256": "cdcf6665ed6735daed36ae962309deb92e8de208ce09bd76e473b69230753e83", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00083.tar", + "bytes": 58777600, + "sha256": "1bde8b65bf2778eae16d09c41c9ac020135e75e5bfdd8ed740f6140ee9a11d87", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00084.tar", + "bytes": 50298880, + "sha256": "aaee475cc31194d726b4fd6655f0ef988315a06f5aa0340bbe09209e7d9305d4", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00085.tar", + "bytes": 53790720, + "sha256": "5bd3bf51ef632fdc02cabd2bfd0992813c6fa0493fa930872ad4616c04dd3956", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00086.tar", + "bytes": 50432000, + "sha256": "95be5b1284663aa422c8193baf276820fc091284b66c58a95d67258b316f80ac", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00087.tar", + "bytes": 48322560, + "sha256": "20f5940a37b7fd7f5a731c0f99cde67f9ce75dd3fb780f3048a16a2e735853d6", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00088.tar", + "bytes": 48138240, + "sha256": "430c0cbfc548263bd463320665a6adef100ac824fd9b82b3cca097bdf04aee48", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00089.tar", + "bytes": 51783680, + "sha256": "e833d0b04b08681c6f5d91c9a298a0da9962e7cfd7b9b55488404e42bab58279", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00090.tar", + "bytes": 52142080, + "sha256": "13ccd7c33064b3e613ccc827efc6b2860b8111cfec5422111e4ce293e112f4a7", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00091.tar", + "bytes": 56616960, + "sha256": "5cb1c96075f8a08f23655334bb3ab4b4f8c615964f0e5648d13bc5cd19282a33", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00092.tar", + "bytes": 53544960, + "sha256": "d1b40bf6df367f862c4be7f2bda2879d6b796e6a2bfe39550e09aad2e31bed53", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00093.tar", + "bytes": 49141760, + "sha256": "b6c248443df1a550419f04d00b96cbbce87cb20c9294628df969913de86b0ba2", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00094.tar", + "bytes": 51691520, + "sha256": "fb4fee6022520bbe8f0145425d113bc8f18255e3d4d5ed5dfbca902119c15883", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00095.tar", + "bytes": 52510720, + "sha256": "efd14bc486beda15fb1c77765f1c92d978a480e2992dd949c2b16e07a9d40699", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00096.tar", + "bytes": 52623360, + "sha256": "30036122cad284bf7965868005d41dca854d2d28c68ee03382b30dda8cc94b18", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00097.tar", + "bytes": 52152320, + "sha256": "6ab8b2f0818f7fc01c375ef19765d4e98b2666cb9f00b1eae55c4b49e1a015d9", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00098.tar", + "bytes": 53913600, + "sha256": "4e8c6a8c387ab98be19e4f8513be9cfabeac8ec38d6b26f6f0c66f739fa0631c", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00099.tar", + "bytes": 57917440, + "sha256": "6b8dc5fa0488506e6a5dacebb5cab16b9fffa86caf4365667f270794c397bc68", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00100.tar", + "bytes": 56954880, + "sha256": "d7541f23e3c5e016de03d1665ae87a344fe04d2cc6340761b235688c5e6da2ef", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00101.tar", + "bytes": 52899840, + "sha256": "1934fbf8572edd6fb28373221e107f5089a503e62c05bde8d77c9017a4f0666c", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00102.tar", + "bytes": 50995200, + "sha256": "09bcc199b0bd579b7a92240d2c8ef10a6a8c9b60c3ce0b15548dc42e4ef3e107", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00103.tar", + "bytes": 48865280, + "sha256": "5f77b7f72523a1e9a13f8714a16ba8080dc0352b9c148d8d697a0dd78f2c3093", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00104.tar", + "bytes": 53370880, + "sha256": "caffaaa3fd5e2452801208bdc088a149a11b38e6303f6f5dcdf4d5bad081a18f", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00105.tar", + "bytes": 51517440, + "sha256": "40d2aee06909709452424b3b31814492d5f24daf8dd5f61123559af1633b8035", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00106.tar", + "bytes": 52019200, + "sha256": "5a4272f4f99b42de1b285fc85d817db7273793eb3e611a8c0ad17deeaaa097eb", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00107.tar", + "bytes": 54149120, + "sha256": "c11a32e3b61082b744597849efa94501005dbcc18ecc32ecbd9203ae181d8a32", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00108.tar", + "bytes": 52858880, + "sha256": "bcf22acf9cd623ce271b561a250a51ee8ec8043b464617361a01f76491bc76cf", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00109.tar", + "bytes": 54968320, + "sha256": "661c0cc86be3a5f9fee54d97af660bff3835b20d7dff8229efb7f3b1467588d2", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00110.tar", + "bytes": 54241280, + "sha256": "17e35fdf88848cf067f0b7d678dba8aedea7e6a1c71a29128f2216c647407507", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00111.tar", + "bytes": 52049920, + "sha256": "95d4bb4a06707049b9740f805a7b13b60ac72e572b10a8d489bec40994a2e612", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00112.tar", + "bytes": 56934400, + "sha256": "05d7ae03abb5016bc9e68e8d555697f0b04e635d05463de79c3e34023780e5c9", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00113.tar", + "bytes": 48506880, + "sha256": "f54834b11ba4263cb5c0529b48595aa3f6d2f71137e0df3c6b2151387bbea594", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00114.tar", + "bytes": 49879040, + "sha256": "aee4562e26d24a4eecbd0bf28e264c8f9546a93b71a69a4bfe94ccb2be840521", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00115.tar", + "bytes": 51947520, + "sha256": "a451683b4a61eb3c72d4c7159c2241e8d44bb9440abe50fe02db4c24de9aa950", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00116.tar", + "bytes": 51025920, + "sha256": "b7a27fcfa54cc14bfb8ed9cfeddbd9eda88d6032adb0479db06b5df27cead7b9", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00117.tar", + "bytes": 49930240, + "sha256": "189a8e62e1950bc76004f6970dbbb38659078c9473a29a5aa8f4cba6b1280987", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00118.tar", + "bytes": 54138880, + "sha256": "67ccfce28084682261652e73445116befcf8436420bd899f64e32e36d863b05a", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00119.tar", + "bytes": 52623360, + "sha256": "fb9c3b896b837c764868bd2145e3ae4bdee148a5473286dac5595cbf63283f7c", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00120.tar", + "bytes": 49326080, + "sha256": "497b12786cd502dbdad32433f30be0c7f3bb97fa929e55b24526d3ea3f2242ca", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00121.tar", + "bytes": 52152320, + "sha256": "ad8c005fe6c5f0c188af25efd6298fa7b1b20d21a54c9fd66566acfddaba8927", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00122.tar", + "bytes": 54958080, + "sha256": "b7368164f305ab3d69bc98b911644a2542aa539199a75428064c3a076b6cc9ca", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00123.tar", + "bytes": 51169280, + "sha256": "9fe8d886250c5ce599aac286498140c89a90f81ab0cdfee64a73dc550e340ea2", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00124.tar", + "bytes": 52060160, + "sha256": "614e1057c643aef914434950aabeb2d06c5b06827e88284fc3718cf1dafc44b8", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00125.tar", + "bytes": 55193600, + "sha256": "a92539bba0a22b5cde0ee72181b4dad030eb4aae1402460f078dd29c756ee6be", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00126.tar", + "bytes": 47349760, + "sha256": "7bf8cf95dbffe74087354cde8c860f758c1bd4427cf23cde942d6f67124f4cea", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + }, + { + "path": "shards/seed_tts/00127.tar", + "bytes": 49704960, + "sha256": "17761d5f32d19285e74703d59bb3df1ecfeaae235cb069a742f2db90fbf7b067", + "hash_source": "upstream SHA256SUMS; large raw shard not rehashed for this delivery" + } + ] +} \ No newline at end of file diff --git a/handoff-v1/resources/RAW_SHA256SUMS b/handoff-v1/resources/RAW_SHA256SUMS new file mode 100644 index 0000000000000000000000000000000000000000..4f60e936dd9b79897b80ef2e57c5a4443fc980ae --- /dev/null +++ b/handoff-v1/resources/RAW_SHA256SUMS @@ -0,0 +1,684 @@ +edc572e1618859c3bfc0222779798067814e48d656d7c0e8f41b2e32be6f159d shards/hfd_audio_command_continuous_office_v1/00000.tar +48adba5d312b2294c84dff371f6035765843c80691eb7e3f37ae289e18d1b717 shards/hfd_audio_command_continuous_office_v1/00001.tar +ffbb4f34de637beced4f9d77fb461397a92073137b1b8988def1eda2c335b6fc shards/hfd_audio_command_continuous_office_v1/00002.tar +8505a4adbf2c097fe9623ee02208948f00fa70111a4b4054860c54cc4493b9d9 shards/hfd_audio_command_continuous_office_v1/00003.tar +7cbc76258f5c97f4f9521675c702e9f62e6dda8a6d6269b14d1c44e014c8a120 shards/hfd_audio_command_continuous_office_v1/00004.tar +73db027a323025f5ab1fa692a2b49c8c4a56fbb04ed9767b66688095523bd9c4 shards/hfd_audio_command_continuous_office_v1/00005.tar +637182c81d39415b614beb5f51c652747f0ec27a5080f6d1421bd9c7a74aa3e5 shards/hfd_audio_command_continuous_office_v1/00006.tar +40e304f756efdbc92c6a7529ee1a0ee25946c882ddc0fb42c96815f9bff273cf shards/hfd_audio_command_continuous_office_v1/00007.tar +37ff9d9fc933151d2488fe89e2e128381deb70a18c1362aa1e6bcd635d841e48 shards/hfd_audio_command_continuous_office_v1/00008.tar +0e331a96c0319595445874e497c0251e47f17a9e937d8ce19bde4e159d845199 shards/hfd_audio_command_continuous_office_v1/00009.tar +05340ebc28eecb9edbf72e811969f5e1ac5b371cae6846121b97d40c577b712c shards/hfd_audio_command_continuous_office_v1/00010.tar +ee7eccf466d7359ad848fb6900c596793c69e5275b9282939ea03013fc228681 shards/hfd_audio_command_continuous_office_v1/00011.tar +a6c4f86dac1bc08371ea25b61fd0f2af3258156d180400c7bcdb5a19fc265ee0 shards/hfd_audio_command_continuous_office_v1/00012.tar +be5e8a931aacbd1a3ff45739484d1583f65e3debce50c7b4dcdbba27f9566b3e shards/hfd_audio_command_continuous_office_v1/00013.tar +e9b0c8fafb1a632d43efd8547f7e9007c06d2a259e14b0a310b795892c2918ee shards/hfd_audio_command_continuous_office_v1/00014.tar +bc43e1804a64e13b30856d1ece422122176fae4070eb97de7f0d3be137d8dbd0 shards/hfd_audio_command_continuous_office_v1/00015.tar +f8f11a7fb8b5ecac915fa7c56b3d1000ba85736dad30b64e3512d37d2cefa8e0 shards/hfd_audio_command_continuous_office_v1/00016.tar +778473972e4c65dac29f6aa4ab4f3178acad333e56bccc31903bf898dfb0532c shards/hfd_audio_command_continuous_office_v1/00017.tar +08421ed611e54982a7019caa09dad9c3a0ee73f2bc7e93828f18027a56c91d96 shards/hfd_audio_command_continuous_office_v1/00018.tar +351c532d3fa29c7580e1389295a51009d67a7102efb089dc5127daae1f878609 shards/hfd_audio_command_continuous_office_v1/00019.tar +d91e183aaa8643ead7dbae2d8225f05579f2d5ab4386d315a25a5235b1eebe2d shards/hfd_audio_command_continuous_office_v1/00020.tar +87ae2ed5a7df09e1439fe0c6016f440cf6388f19030af300f4b95e7521089153 shards/hfd_audio_command_continuous_office_v1/00021.tar +8813ecde7e8d1d2a76d47c94155b8aa99319e1aed985b14803ba4bcc58a05055 shards/hfd_audio_command_continuous_office_v1/00022.tar +2baf666f0d99ace2c70b993a3d86fa58a2e7167971060fd22450cf7601c8cfcd shards/hfd_audio_command_continuous_office_v1/00023.tar +bf91d285b5d7caa4939fd7aad56758be2163c1501238f8dab56dad87e7bc9ad8 shards/hfd_audio_command_continuous_office_v1/00024.tar +cee03cda337df0d03b086d27d890221ef54b99957269eb4d8d35c14218422dae shards/hfd_audio_command_continuous_office_v1/00025.tar +e914a78d22e8b9f1b87cbd0e8295f88fcab53c5d3727edbc874735ca7184056a shards/hfd_audio_command_continuous_office_v1/00026.tar +6b9dbe8e9c05480799e806cffb727416d76767a6bc1674b5218a6231264e257b shards/hfd_audio_command_continuous_office_v1/00027.tar +e6a113ba70b6d4eb0c756808db92cda84e22edebe2a2e901ba8e11e47b78ad10 shards/hfd_audio_command_continuous_office_v1/00028.tar +8426fcd42016303d4326b3baa4968318a311c09ba9779d528d97c68dc4969c70 shards/hfd_audio_command_continuous_office_v1/00029.tar +22ec6bf5bd0ccccb93b11199d821601681a13473b246ca83d4e825abc5bdae6a shards/hfd_audio_command_continuous_office_v1/00030.tar +72bf55d2c6b2d983aee13b13db10ff85e277cb6c8e158ec2823351cf5e06ba8a shards/hfd_audio_command_continuous_office_v1/00031.tar +4f283ae1a4d187da989098dd9913fcc8296d28fbabb37d4b1f702abc39b9b519 shards/hfd_audio_command_continuous_office_v1/00032.tar +97e2c3783d67edf47ea1fc2a05416709727df703540322f0932dfead1841b692 shards/hfd_audio_command_continuous_office_v1/00033.tar +945539f498ddb717abec6d29b8705024c86eff8e5dce0baeaa0011a5782bb2cd shards/hfd_audio_command_continuous_office_v1/00034.tar +4e4146a507f665b867f7ab0834e6057f54f34a9f29a794ee5dc99f6d1c2a11a7 shards/hfd_audio_command_continuous_office_v1/00035.tar +45fdb33aeece62165daeb1a642ec232577f93bca909be39a3a73fcf9007e13ba shards/hfd_audio_command_continuous_office_v1/00036.tar +3cd66bd03656b3a12233e4cdc1c72e0012d5cdf6122cd441389d4b688b5669f1 shards/hfd_audio_command_continuous_office_v1/00037.tar +e3d82343e93766f15a91d814a29635af5fb043935b6551ff2bd7888e4efcec21 shards/hfd_audio_command_continuous_office_v1/00038.tar +b8439316f65c8c9eeac2c006447887dc2f688fe6a9a81bfd40dcc7574302eb00 shards/hfd_audio_command_continuous_office_v1/00039.tar +93715b6b0417beec03ab3f262ef21ab94fadd5e8bc20051cf9e91be2c4bda3d3 shards/hfd_audio_command_continuous_office_v1/00040.tar +e2ee393fec34f04412b8f042a0672936759dc8e05d77f95c37ff028cbfce80f4 shards/hfd_audio_command_continuous_office_v1/00041.tar +da97f8de1a9e03bacb919b1a8a0a6ef08a93276d4678d7b85f44356c6ee60855 shards/hfd_audio_command_continuous_office_v1/00042.tar +fc6239a33b3de72d7493735536b3b56e992a1a46ced8896ff69937bcb51111d3 shards/hfd_audio_command_continuous_office_v1/00043.tar +8b52ae33f153c1e151c89879a721d56ddff7f8cd3a68b7479dee2e3485e9b3c7 shards/hfd_audio_command_continuous_office_v1/00044.tar +0f8467a200b004acd9611000866e5405e1e35c0ab28a331e1f1556fd79f3c3cb shards/hfd_audio_command_continuous_office_v1/00045.tar +12c02c76fe1f4f4a01c142adcd7527c099d92597af940162ce035abf1a1bba36 shards/hfd_audio_command_continuous_office_v1/00046.tar +602e4c90756feb899a46ea47185b94eb4c6eb8ab71f0e1229683d335d1304532 shards/hfd_audio_command_continuous_office_v1/00047.tar +978ff0520d79d4a710286de4de5362fd48d6142c59f7d679459c4507f5e29fec shards/hfd_audio_command_continuous_office_v1/00048.tar +45ce369858b176730b69e5a3764bcfc09cd155c37cf07721dbd3a5b0d3bb9939 shards/hfd_audio_command_continuous_office_v1/00049.tar +2c7e6761a7a5986ecb3aa53d58ead04f7311e5c2dc15f72adaa67dbef13bb560 shards/hfd_audio_command_continuous_office_v1/00050.tar +5639f7fe8f0ac4f6b412ad180327f9e1aac58a5f560e6a72eec416edc0543c90 shards/hfd_audio_command_continuous_office_v1/00051.tar +a6b80039645a4e5f973e3617e5997885a2463897a7529240e0a158891558d27b shards/hfd_audio_command_continuous_office_v1/00052.tar +eb9646a3a6e8295b74a482c5937a07e2854ca740d9b4efd38d82b7cb3ba72ae0 shards/hfd_audio_command_continuous_office_v1/00053.tar +43cb1b9bad8d93c75f55ee64f497dd4524650b4e3f5f10a304b31fc55262da1c shards/hfd_audio_command_continuous_office_v1/00054.tar +594293e2b454e020f2241ef35dc4d54083c30e1e68be7e36168e851bda4b7bb6 shards/hfd_audio_command_continuous_office_v1/00055.tar +c0184deee534d40d20c498e9a54740046ded3a977e0c888883b285de25c51231 shards/hfd_audio_command_continuous_office_v1/00056.tar +2bb72b2f5a79632475c7e91294d8192a0f101ff6c4a3f7813e64ebbf902076e4 shards/hfd_audio_command_continuous_office_v1/00057.tar +969c7520b93ad5e36cbc169fa0b0fa699245c62d09d7ee82705fff436bffdd8f shards/hfd_audio_command_continuous_office_v1/00058.tar +bffb34ec454ca4dd677264b2da5d53976c0db61862b027ee2b7c1e66895c57df shards/hfd_audio_command_continuous_office_v1/00059.tar +3b192f005ff5cb00e235a4e0a7f2eaa8f1617ee259589516985932f327b698ef shards/hfd_audio_command_continuous_office_v1/00060.tar +20f155fadb145c348311681cf82acdc57f478628986fc42f44295866a3b7eb3c shards/hfd_audio_command_continuous_office_v1/00061.tar +32f51e566fbbebc7a89ed10f37cc4162913362cc4ba1b8427c3aac99f837f8e8 shards/hfd_audio_command_continuous_office_v1/00062.tar +b6136c399c8b46a1e14f84341c3fa37c53747df052713139f4519ea3fe10a4fb shards/hfd_audio_command_continuous_office_v1/00063.tar +71fcb02436ac976d21d07022132df2951852b4eb9a7fbb0f8b2263a991968dea shards/hfd_audio_command_continuous_office_v1/00064.tar +18f884d8b5a21d7b960650ae238900897beeafae85918431ea7dbdf8e07cb463 shards/hfd_audio_command_continuous_office_v1/00065.tar +e01646ec62dfafe4d711a8e0c6b49814a984e05b31ca6853cff5db3e03f39144 shards/hfd_audio_command_continuous_office_v1/00066.tar +f16fac3abd419f287b53249993a0f0c1637f7df24e8bcf428cdd8dacd16309be shards/hfd_audio_command_continuous_office_v1/00067.tar +6beadcbf216402f3b14c3e2770f00b1a07a0923d1ee6fd0f853f83f494bbd335 shards/hfd_audio_command_continuous_office_v1/00068.tar +abaace22b7c89b32060f802742fe03e895849204efbb1ffcdcaa55096d8d79de shards/hfd_audio_command_continuous_office_v1/00069.tar +223e4d79664172045ac043581c4d7c76b2c23250134749cbac63b13a07da36ed shards/hfd_audio_command_continuous_office_v1/00070.tar +01649b88a4883a71207105df06d4fd18acb5a7109a4085ceb643921fb42b151a shards/hfd_audio_command_continuous_office_v1/00071.tar +448343d454f4e05f61cbe1161623c7e1ad45c38c00361bab71f356b0d561cad5 shards/hfd_audio_command_continuous_office_v1/00072.tar +d098cc02497e16396f28283a7d3586a345be5175fa7ed644bc865cef662c2316 shards/hfd_audio_command_continuous_office_v1/00073.tar +d071f984dc13e707aaf2684f0db7d6b19f8f9be39c831aa7fe7143fae446d18b shards/hfd_audio_command_continuous_office_v1/00074.tar +b4092c8b245b8000ddc54595626c97f4f1461bc3aaa4038cd15081363f62963e shards/hfd_audio_command_continuous_office_v1/00075.tar +a19e9986aa349d9afa65ef740e4cea5e8d8b7e5c4cc42e1bde3cdd434d2ea0d5 shards/hfd_audio_command_continuous_office_v1/00076.tar +9239b7655b44b17bf87cbb04aa455df6684765b82552ec80dba6bf0113c7d74d shards/hfd_audio_command_continuous_office_v1/00077.tar +f2d3f57be0d1c3ae2881496d54e38ab176fb4c57630a149a5a883f5d6b3c1898 shards/hfd_audio_command_continuous_office_v1/00078.tar +82bac1648b4b439109310ae93392ffedf25b028be1f5d3970441692724493c77 shards/hfd_audio_command_continuous_office_v1/00079.tar +d07687b31500b57ff57082ff4363c24b63f811a379da6fe894f4b3ccae52a458 shards/hfd_audio_command_continuous_office_v1/00080.tar +fa201091f166f5ab939f633edc3e01858f99f07a4811a49525edcc02ea31eb44 shards/hfd_audio_command_continuous_office_v1/00081.tar +d29bd533254574772d4f7a7243f0bca5ea3e8b9046eb7dbad942ac5da22c90c6 shards/hfd_audio_command_continuous_office_v1/00082.tar +923f03cba37138f88709d92ad6ed98d2d056dc1e566602b54a02106288d289eb shards/hfd_audio_command_continuous_office_v1/00083.tar +ed803810e3e7601bfb4fabe2e48ca43b75d0335d16eb6491aacf3e888f756c82 shards/hfd_audio_command_continuous_office_v1/00084.tar +14b31ab861b555c49570309a242df0e18493ccebce09877105167420d69ef7f9 shards/hfd_audio_command_continuous_office_v1/00085.tar +05231c8bcad79c362e9a6ae1ff0dc2fde89c7dc7a106880e97a8d9b1df6e2c64 shards/hfd_audio_command_continuous_office_v1/00086.tar +d53a7f427f9c3dc2effe79789be216bbf6311258a9947bef2e341a75699adc3d shards/hfd_audio_command_continuous_office_v1/00087.tar +98a82dabf77bcebbc01fbc81894654058f67e680af05ed709bea3a0dfaf2839b shards/hfd_audio_command_continuous_office_v1/00088.tar +4152bc085746365052fe54dc3346c3b9780320de6d42ee1028c404a81c7df14b shards/hfd_audio_command_continuous_office_v1/00089.tar +23923f495694833ebc9bcec4805846f3295a2add1dcdfe3b38f7014891840c3b shards/hfd_audio_command_continuous_office_v1/00090.tar +6723d8f0dcde70cd25451836309d135edf9e481a8a8c827bfd8c66bf9f0d8d8f shards/hfd_audio_command_continuous_office_v1/00091.tar +49559f0c42628020a5f0d96ad9296321d740ed3c20c3ccec68b091956a5bbf41 shards/hfd_audio_command_continuous_office_v1/00092.tar +a48ad37706aeb963e9849535d2c86c46bb6dc20e5afb4571e0a755488da1e950 shards/hfd_audio_command_continuous_office_v1/00093.tar +c9da41d6a27f32a8ea4a93e890cc7f0f175b89e8941e5f1fd1e7a2ff5a9d61df shards/hfd_audio_command_continuous_office_v1/00094.tar +067f2fd10ea207b0fcf6532518f9340142f0d2ffe6655adc48ea913273f58277 shards/hfd_audio_command_continuous_office_v1/00095.tar +89db4237b791793c89857aaf2d4f3383bf60a1f117d0e4f7923606c2c35adb41 shards/hfd_audio_command_continuous_office_v1/00096.tar +b2b31d169c73febb3bc57f4edcbaa1ae151ad3789fb0ba90f460102893e87ec8 shards/hfd_audio_command_continuous_office_v1/00097.tar +3d2519e4566a715186efa5eca08556c3c46e592e8be5c64df8da44aca188261f shards/hfd_audio_command_continuous_office_v1/00098.tar +ece552defec9edce614cc97b2da44edf9118a029e566c8c9af48fd796e68ba9a shards/hfd_audio_command_continuous_office_v1/00099.tar +43232dde4494ea3aa9bac35824c19a65a2784599292a601e0c4ad0cacfa89faf shards/hfd_audio_command_continuous_office_v1/00100.tar +1990375066bac9ff97a52db8eb52f11f83bf2e210482a2749e29af3232b5410b shards/hfd_audio_command_continuous_office_v1/00101.tar +789ef419af768dafd3a278d2c22576144206da2d039151c41a9af2a30af379e7 shards/hfd_audio_command_continuous_office_v1/00102.tar +eeb72aa4f1f304d45e4a5d12509ec92441ee86af984fe68d6397084921d88157 shards/hfd_audio_command_continuous_office_v1/00103.tar +4e726608043f83c504f8fb4f26672d3db1f6c65bb1b02048ebf407b9ccf30de8 shards/hfd_audio_command_continuous_office_v1/00104.tar +3556fe5e90aad3a199297564e49f252645c2141d60945e724a9bbea9a375007c shards/hfd_audio_command_continuous_office_v1/00105.tar +997a200db4a74be3c2074b938f30018499c26f7b6042b8a54557d933f4f78293 shards/hfd_audio_command_continuous_office_v1/00106.tar +4eb564a14ff017642b519d6c28d939a8ee516a06b46d235a55142028844b5dc0 shards/hfd_audio_command_continuous_office_v1/00107.tar +76f443b78b67068156e8d67f4910d787733e49b70ed2d1ac58f9193b3fb3b1e5 shards/hfd_audio_command_continuous_office_v1/00108.tar +ce06f2e54a0c327a265f17c40d54153e05782713abc839cecd3876b22d51e8da shards/hfd_audio_command_continuous_office_v1/00109.tar +0470ed9e94514ca7d8bdc28a65588fbb7ce5b0902d59f37ed1ece4cf332e483b shards/hfd_audio_command_continuous_office_v1/00110.tar +46f810f6d5ad672c2f29711be6e6676febd036f453b0084caf2ae1f110aeff07 shards/hfd_audio_command_continuous_office_v1/00111.tar +e0bcc570e0abc5af3bb1296565188f0f27fbaa703e9d32cb5e7118a5e861676b shards/hfd_audio_command_continuous_office_v1/00112.tar +aeca80e0e283c7af5198c52ce88d2d7439a517f279f0f218c5a9eefc104da0f7 shards/hfd_audio_command_continuous_office_v1/00113.tar +1a2b8a51cf49a512c4355875add743c403b1944336e66a8a5db69b6421ce734a shards/hfd_audio_command_continuous_office_v1/00114.tar +3f1bdd4d33c18269283cb5a9a9eb9f1a727abd41348b3d2120e92e15f54b567a shards/hfd_audio_command_continuous_office_v1/00115.tar +229db215622b0b1880abdec66bdac1b8a996ed2f9b7889e5b93df34f74dc53e5 shards/hfd_audio_command_continuous_office_v1/00116.tar +e86b54618e621b61bf8d32740fcdb81cba22e83bfd6a41f75fecda09021e54d4 shards/hfd_audio_command_continuous_office_v1/00117.tar +499d6c22010c4c9c01fba3fea7e2cfe96bd9cf7ea14ba4906700ea4cc4e754bf shards/hfd_audio_command_continuous_office_v1/00118.tar +5fffb4e6348fec257a194e5a5aadf46f83693bb808b92a41dfd472cad9b084e2 shards/hfd_audio_command_continuous_office_v1/00119.tar +9d7ff1263e7e2db30beb53e09f91fdaebf0e60f060e1a2e16c05a56126dbf94a shards/hfd_audio_command_continuous_office_v1/00120.tar +1e8e690796fac27f42d7912610fbc5a9e1f1e8c745766f15ea4d1f10fd0dffc4 shards/hfd_audio_command_continuous_office_v1/00121.tar +ecbbeeb16f1d1a6adcc46e142d93c1a5cb932924ca7b3a537643223d065a25b5 shards/hfd_audio_command_continuous_office_v1/00122.tar +d6c8cd91debbe72a09084e1486da3dc6c35c5064eeb20dffc1934beffba9734f shards/hfd_audio_command_continuous_office_v1/00123.tar +592e96be49810c79254232a94cac250bd799da6950109a5db96125069d2e729c shards/hfd_audio_command_continuous_office_v1/00124.tar +87497e9131a30052fab22487c6c902a9b1f8c01c8047b276601696324b9386c8 shards/hfd_audio_command_continuous_office_v1/00125.tar +f882065bf4f6d11fc6c369b5d6b1b448fee89904eee177556ad6d843263cacbe shards/hfd_audio_command_continuous_office_v1/00126.tar +9871cab79cd56608dc454b80806f820444ba67655193fa0177dc248a4f489d2e shards/hfd_audio_command_continuous_office_v1/00127.tar +e87a6cf29bd174f62239fba65ec71c06aaa28f145c519fc9b4dc55ff3219b886 shards/hfd_backchannel_commands_v2_short/00000.tar +931ed66438de3bb950af8b0b950e0515b9eeaaccd334d78088bfb1c51c09de32 shards/hfd_control_commands_v1/00000.tar +011993688d55f5b0f0c54f8d402abd82957b382d94e94fd788737142df117d4b shards/hfd_hesitation_pause_v1/00000.tar +77e21dd06fac3dab3fe80d4b36b8a648d36deb222dc5e33f86be55abdf2353de shards/hfd_hesitation_pause_v1/00001.tar +bc461a0b4a156301109af850d9fc90b802a8f2f4dd777d43f01127fb9d93846e shards/hfd_hesitation_pause_v1/00002.tar +22453fb9b1268170eef0f518d4c3ae9cd2ef441589493f3b4ec3fad05c9eb66b shards/hfd_hesitation_pause_v1/00003.tar +6509f7416d93de7bceb2cbc89053da36342e8bef1d43e9a7eda3d4534866ad4a shards/hfd_hesitation_pause_v1/00004.tar +ffc3758689d1e9ac16249910838f1dfe79b6f0d022e8b888130de62676c8a028 shards/hfd_hesitation_pause_v1/00005.tar +26fef958c65f1f44634a44a0f38b00e2505af44f1b6988f2206f42b4d6359464 shards/hfd_hesitation_pause_v1/00006.tar +1b5271e59827d30fc06e5628b0aad017bf6c400f0bbcff8459690d6044dc144e shards/hfd_hesitation_pause_v1/00007.tar +dd70aed4ccbd59699f99ba10dcaac7f63dc9ae4cb79323239f71f41879605b26 shards/hfd_hesitation_pause_v1/00008.tar +3efab2954680ebbe3447efd7e71dcaa632790bb30eb32904778d53f227047912 shards/hfd_hesitation_pause_v1/00009.tar +e5d5a9a238b9bb6f0740524ec90522730e2fe87d047aa7858b21d0581b941f69 shards/hfd_hesitation_pause_v1/00010.tar +19015558a10a31329dd01f7c82c8513cebcadc8870c49b3d66880fd98ef61210 shards/hfd_hesitation_pause_v1/00011.tar +7f769b427299d26c54bc04cb0caae40e7d1e671774ade2ac64d547c026a855f7 shards/hfd_hesitation_pause_v1/00012.tar +02f79dcc1d83a06712499c84628b688f54e10a269bce733550a1cf628766be36 shards/hfd_hesitation_pause_v1/00013.tar +86d3636ffac3b47cac59b07b3bc923beebaa25f17f3188db786bd39e62115d40 shards/hfd_hesitation_pause_v1/00014.tar +4fd8bdfa4f853e78dd4c7b06d4269d1789e76c3af6f6a6ec5a67df76f8e23e5e shards/hfd_hesitation_pause_v1/00015.tar +f2aa704fad58a877eb32cf309ff0cc0cb9fbc47c5a56a65c5b97a1522fe2d832 shards/hfd_hesitation_pause_v1/00016.tar +18efd732851496c53f75f5d44a2da76a8e3d062f4abf214c836b244de7e7238d shards/hfd_hesitation_pause_v1/00017.tar +84a4be2df65fed861892fefe25c7c5ad441ff0f0d87d4ada379649742e393c21 shards/hfd_hesitation_pause_v1/00018.tar +89ef29d2f8569b6738e1eff723e36218c73112f1ac7c35992a164a8870f44a3c shards/hfd_hesitation_pause_v1/00019.tar +3cafec0fa7c33d67440c4c44ec6743373953b3c65547a3eaf264412eb3e5f40a shards/hfd_hesitation_pause_v1/00020.tar +6a29792a46cedc7e63104d080e7621457dfbb6af022b1f7063d7556a5c2deec3 shards/hfd_hesitation_pause_v1/00021.tar +81f8deb18d47404583fe461bda3d5fdad6800dd8311b180ea091f45ba1d74687 shards/hfd_hesitation_pause_v1/00022.tar +f667666f5b7e8abd71c6019d1f886e905296c6ac2711af8d821cbd469c546ece shards/hfd_hesitation_pause_v1/00023.tar +f6d04acb28c1d065d840ea66f1b60794adef947905efb5a53f31a9d1744c9afa shards/hfd_hesitation_pause_v1/00024.tar +be51f5d3e190ee72d273feb21e516bf5471eeb5996dff4954796c561c2928d42 shards/hfd_hesitation_pause_v1/00025.tar +94a043687cde1ddf9885f064482c476cba515ca464c11d64d0dc34d660dd2149 shards/hfd_hesitation_pause_v1/00026.tar +445cf676beb1e7260b737aa56b76ed894e636db7056aafaeaf96a197e03286c1 shards/hfd_hesitation_pause_v1/00027.tar +2269c653264e9f7e03ed161f257e354b97bcb82e2c2610760caf6e8d0fee971f shards/hfd_hesitation_pause_v1/00028.tar +631ab8736d60fbb3e69a7dc9f52dd1f870cf049938000bd8db521f61aa9a009e shards/hfd_hesitation_pause_v1/00029.tar +8a52a90ca65d319a09ddb7f37bc94b95dfc313a8072a9fbca283c09eee668650 shards/hfd_hesitation_pause_v1/00030.tar +8dc406b85228bc7d715674a804cb47f5a23f2794cb856a06e3914bd0288fa54a shards/hfd_hesitation_pause_v1/00031.tar +61fb5be780e093a3909c11afa72cefcfb1e37cce7c14cee21c76e89a2c202bc5 shards/hfd_hesitation_pause_v2/00000.tar +1c9ad74856af5e2cae5983021f8b3111740c8916bf54c0ee0e9cdb79782e1253 shards/hfd_third_party_multi_v5/00000.tar +55f9bae2a707115bfc3350348598ed588361526f59e9ee80781d28ed06f34c16 shards/hfd_third_party_multi_v5/00001.tar +308b66356270aebb88db350f2e116f26692b13b20a3ee789e623da6fda0f29df shards/hfd_third_party_multi_v5/00002.tar +a47267a9787e6a1ef5c05487ee7cedf5334d7fec2f5067f45d4df27bfd618483 shards/hfd_third_party_multi_v5/00003.tar +1e81c204064fa2415cda1b0647dacd302ab510890580c9c28888e5291f05f7b5 shards/hfd_third_party_multi_v5/00004.tar +96310618b1003049fc9287c53ced265ee94bf59538da0d3d0a0e546b4748dc8c shards/hfd_third_party_multi_v5/00005.tar +1f18db221a4daabfbfc0c452181f87ee9baacc59b11a04a16617ff0fe31aac57 shards/hfd_third_party_multi_v5/00006.tar +9aa75097318f7c978838541223cfd3c3ec28c8365f39ee7da03646636d9b748e shards/hfd_third_party_multi_v5/00007.tar +4496597f9aecde5d258fe8f82400eccc58cf64c4e89aa79976a96c3a2455e799 shards/hfd_third_party_multi_v5/00008.tar +734e52471a065b147c774fa3765b5908ad368500ee85e7ed5e82d9d02ca027ee shards/hfd_third_party_multi_v5/00009.tar +ba9f9dbce8a2963e5241f15f9653ff2a5bab71af45f8d29efead466d02a41e17 shards/hfd_third_party_multi_v5/00010.tar +1f08984f9c747716bd791ee66818bd26a1c0f5cdc21e8e2905acd2279e0a0f39 shards/hfd_third_party_multi_v5/00011.tar +d001d40c0361e7bcb58793c769230706442a77c45a859853c8c94d2ae3568752 shards/hfd_third_party_multi_v5/00012.tar +36708e2e45744e223b269f5f29631f3d8bb38c91b241854de58185731c3e608c shards/hfd_third_party_multi_v5/00013.tar +aee98fc516e10edc93590adcc265381c07f51277d31498c7a54fd62b1d62d994 shards/hfd_third_party_multi_v5/00014.tar +f5ac09d3a98e1ec3c86943f9774462ea07ed01d4a165787d32ecce7e96065a18 shards/hfd_third_party_multi_v5/00015.tar +53cb60eef6e7edf98f2729e8dd94f01889a5a812cef7c79e7bdfac8e4db4dc58 shards/hfd_third_party_multi_v5/00016.tar +5efaa10e12feb6e8268de9af55da43903efa3cd47f0ca717e6aa40db69235c75 shards/hfd_third_party_multi_v5/00017.tar +72e3ad262204413634f2975d1019e7531ba4ff0f8893d5495011cf8b9a17606f shards/hfd_third_party_multi_v5/00018.tar +75660261661d19ef82c7ae7f3719aa14939f440a8777818adc0bd38e856c85cb shards/hfd_third_party_multi_v5/00019.tar +d2303a912d7e4f20eb7b372e1298e6222410934fdec4ba9908695887567dc856 shards/hfd_third_party_multi_v5/00020.tar +122f8b711d6d18e89178b3a8688707749b4581b6b2dfc702f529bd9121555ec4 shards/hfd_third_party_multi_v5/00021.tar +5fdc4c60bf8c79b22d782d6ebc3e977b5ee40b268249f3dbd754641d342986b8 shards/hfd_third_party_multi_v5/00022.tar +ccb2a0bd081c4e8309e8b4a470e586b1b37c2485d5ada709e75af81602f90285 shards/hfd_third_party_multi_v5/00023.tar +60b0e1f3ad4a99b000b147d122e0ec677b1a3475cb9975cf61e455f7c9672694 shards/hfd_third_party_multi_v5/00024.tar +a2ba05e545d85b88bc5e2480c212f929c8440fea0834d7e1b5f381ab726ccc9b shards/hfd_third_party_multi_v5/00025.tar +9da5009d95e3d5c3077d3abaddda61218356b04f36c1f27b73709ffef4eea133 shards/hfd_third_party_multi_v5/00026.tar +0dfc7b642f4a8831e9f1929b37cc2a204f8e1f8e536e8f3d3f0a94e4695bae84 shards/hfd_third_party_multi_v5/00027.tar +d3cbaadbec108a5957a74b6c1ad502ed6468a77ff2d72418282d6f057b49c193 shards/hfd_third_party_multi_v5/00028.tar +408b5420b884930ae570aa47e5df155de535f0c31c57f8f62f507455a83911e0 shards/hfd_third_party_multi_v5/00029.tar +530c047ffa98f2eb76b57fa66f03d2c669ee26c37cb1fbf0bc89a4ed5e243820 shards/hfd_third_party_multi_v5/00030.tar +da57045f971935630a5f9d3f0daf808aa2763f2664cb7e1979ee40ba25d2c451 shards/hfd_third_party_multi_v5/00031.tar +1dd72089903e036e029c048b453d7124f1cd60b08ed6a86b3938b24685212a01 shards/hfd_third_party_multi_v5/00032.tar +451d605759f71271e0225db4c70f8849e4eeeb21b175a9c922c0fb2825d797d2 shards/hfd_third_party_multi_v5/00033.tar +ec3abadae348565c90eff18823fc404291625ef0d51a869ccd7d357c9e9b9a30 shards/hfd_third_party_multi_v5/00034.tar +870bc04d5bc0d04937e17e8aa19538edb8492d2ce8dbf05db5b0798294a2d46e shards/hfd_third_party_multi_v5/00035.tar +b1dfb1580954f1a698408d75ed9ce2fa40ffa80bb98bab94059d741b13cb9a44 shards/hfd_third_party_multi_v5/00036.tar +c9baa10f19d5aa40204dadb00f883bbb3996fca24caa8a1140ccadb18e1352fe shards/hfd_third_party_multi_v5/00037.tar +1ee34cb4229c289ff96ea8f8a4325ff51763d9778a4c6545ca2713a8c8ac7862 shards/hfd_third_party_multi_v5/00038.tar +de6d29e96367a594b43acaaa9a71463194ad7a7af26d1cc089a7daafccff39ca shards/hfd_third_party_multi_v5/00039.tar +2b5b14fd83f8572de3589a19f8983753acc65afbc88b9954419df3867b8c03cc shards/hfd_third_party_multi_v5/00040.tar +783c3781403f651364e7f5e2de2624d06cd2d9bd15cb1406f45886a97963bda3 shards/hfd_third_party_multi_v5/00041.tar +fc887574c022d4f94780d5739b87a7322a464f6e1a2192cf05442d43bd399327 shards/hfd_third_party_multi_v5/00042.tar +ea90500972f46d0962a71198ffc49b8926b73772bac4f6b13e850cc554a5c499 shards/hfd_third_party_multi_v5/00043.tar +c8ec7991acac016961e2ae2520b7bbb5006b4ad4718589b86085eaa75dde754d shards/hfd_third_party_multi_v5/00044.tar +d2f9de075f06d47128dc48d751a4d2584cc73f329ea9c628bd1151af9a222714 shards/hfd_third_party_multi_v5/00045.tar +06237f7b6491e03b5d9fb5311487cfe8e9a63e35ae22fd47242d3ea2c4b75e28 shards/hfd_third_party_multi_v5/00046.tar +b8079593abe90a4c79c0c5a01699e07d35092d2f7916cc26fb258618b3632418 shards/hfd_third_party_multi_v5/00047.tar +3b964d2ee4868d8cfa81ddd0f068abe64f0906da3d4708db24d629cfb88d2684 shards/hfd_third_party_multi_v5/00048.tar +ee60f36ccf5b309d2b9c96d52179daead4412c9924b459bc9f7dababb50defc6 shards/hfd_third_party_multi_v5/00049.tar +319c0fa4bb6be2130a063ba93fd7eef3017dc45470028b7ed76ff473511c52a3 shards/hfd_third_party_multi_v5/00050.tar +d0ed32b0982ad69fbffcb853daf69091f4756c8de3dd61d093402a9a298075dc shards/hfd_third_party_multi_v5/00051.tar +f6c66439c86803c1566bf49945fa0dd2f584d494696f6bfba3eec37f186e844f shards/hfd_third_party_multi_v5/00052.tar +0cafd8695f4916638cafbd7484edeb90c7bb7660bf0c4694d3803df4643d4ef1 shards/hfd_third_party_multi_v5/00053.tar +53979336f63d5738efa0af5d1a944d6fbe57943fc23b329bbb834696d479c282 shards/hfd_third_party_multi_v5/00054.tar +1c60239d1744f4c93122a50a3cd39352384af7da4bf4b1fec666ac05b476eb2e shards/hfd_third_party_multi_v5/00055.tar +2826f6946920258f62dc915c47805496fcc2f30c81b2afa4a3dfacefc9e97fe0 shards/hfd_third_party_multi_v5/00056.tar +bb0d119a6130e9b296881e74c6eb61edc451eb8b14ad0c0e77c1cb320c32cbba shards/hfd_third_party_multi_v5/00057.tar +7faf38b8b8a802cf807680d5a167038d5b208b4fd824e910cdb9e7f4d789e951 shards/hfd_third_party_multi_v5/00058.tar +0c9cf7b2d4e1ddd13cb8ffc027fd74e6307d2c2e88908be4087ff1284684769e shards/hfd_third_party_multi_v5/00059.tar +df506e0c05a43939acd6380b7d71e2b525fecdd6cf72718c13a1e8b97af52957 shards/hfd_third_party_multi_v5/00060.tar +c910aceb128d873d4b9dee0e77a5e2916929ec79ab0619d1df005132ba2098a6 shards/hfd_third_party_multi_v5/00061.tar +8af54d0b5d34ced1943381260adb34101b5e2b19e0da7a571cafbc427627ae86 shards/hfd_third_party_multi_v5/00062.tar +6736c26a668ac8ee870aa7217d558bbdb34dcdcd1157ef21fb148aca5588054d shards/hfd_third_party_multi_v5/00063.tar +0718f9c61bcb321f5aaa43239102428d887b5ebc832c43f4cab0b89576b33dc3 shards/hfd_third_party_multi_v5/00064.tar +6ee30fce9882e0adb037981dd92179d304c0946dcf841ee78aa160388bb8a716 shards/hfd_third_party_multi_v5/00065.tar +12bc796f01af4f920c22c8f6fff10e85d59b5e06b08c096abf24496eb40f0a14 shards/hfd_third_party_multi_v5/00066.tar +86f1ffb18c7e1ca2b057d8aa8877d559e9eac90c544bb5d96730d354ef7f3172 shards/hfd_third_party_multi_v5/00067.tar +4718392b80bc55fc6c2dd74ccab2ad271e4fe2b302fb2305ebcc8d2d692e6c3f shards/hfd_third_party_multi_v5/00068.tar +1c4948bb65c546ae42405dfb6199c2e1326e9b6c98e4af3e1a173b712eb27e1a shards/hfd_third_party_multi_v5/00069.tar +0bf61911897478814f197295b5cd6ccf43ce3676ee1b185a1d0bac87164ea871 shards/hfd_third_party_multi_v5/00070.tar +20abc8ea01326cf3470d6a2a64bda1c7bf54b06094075488a98c588e541fe80d shards/hfd_third_party_multi_v5/00071.tar +d64fb6fc7821b61a13de248ce42c1ca2de18bc29474e319c9fd90477cdeecd16 shards/hfd_third_party_multi_v5/00072.tar +d93bc4e91644241e2ab4e5aa14d0cb24a23e46c515a22ebd1a7a72afc05ef8dd shards/hfd_third_party_multi_v5/00073.tar +1aa0611545419f89622248db0edab678c1d03212149d89519fc92d3d0546c80f shards/hfd_third_party_multi_v5/00074.tar +4c50161a37d5bc8a7f0f25ac2bb68c19bc42329eeb0d46c663cb7cdfbe8b0eeb shards/hfd_third_party_multi_v5/00075.tar +d2cfd39fbc700e851bc3fc394b2a018bc7c29f5bc0aa46b3c3e2a8332554d85b shards/hfd_third_party_multi_v5/00076.tar +943aad8ccf298dc51076d26e8923e5fa0ef1bc305a0321765818e5e7c8a4e483 shards/hfd_third_party_multi_v5/00077.tar +73833e6967801ce6b67a8fabb378f38c88b78b81425086425f47c9f6de713746 shards/hfd_third_party_multi_v5/00078.tar +46a26496bbf1eb956fa24669f4a20becbea17213b10fe34a60a93769b0cdd83b shards/hfd_third_party_multi_v5/00079.tar +16caf6267f03c79dc0f1f1d4fa0c62efabe02e2e7656d52459ea33f8701872da shards/hfd_third_party_multi_v5/00080.tar +29b2af650a82e214da22a53a458f4876d7c3dc08051c8b0784e20b6bd4fb4be8 shards/hfd_third_party_multi_v5/00081.tar +07ee3554badbb246ef26d4eb6143e8b5900c1aac3106629a629402b054155249 shards/hfd_third_party_multi_v5/00082.tar +14dd731408b707bc481303754614584052daee43c2c5299cc287fd0c192aa240 shards/hfd_third_party_multi_v5/00083.tar +a8058395f87919342e195cb5eb3cdbba9eb76d9a2a0d75cd245b9cfaefde1bfa shards/hfd_third_party_multi_v5/00084.tar +bb83da45305d5f5e9a5609f14f4eaae54fc831387ded8ded934efceefc64b3a2 shards/hfd_third_party_multi_v5/00085.tar +6d254737bf80078b32feaa76804868603fe3f3bc20f2a298dd3fd92376137084 shards/hfd_third_party_multi_v5/00086.tar +b2f33b342fa65cf5b2eb53d71a46856255ab5c79959985a23ed73d60a1e89e06 shards/hfd_third_party_multi_v5/00087.tar +324dcb11c27ba0db6a49629115ac2d6909055880e245fc48afb2affdf3d19d2a shards/hfd_third_party_multi_v5/00088.tar +3e8807a87cff8036d6cbb20ab87951485641e3184e8d652b13391e14d0d577d5 shards/hfd_third_party_multi_v5/00089.tar +5a435445d686904cd601fbff9c3115d34919757da07a7d7bd19e727d23cb4c42 shards/hfd_third_party_multi_v5/00090.tar +7e16396c8e81d4eb41fb0103e013cfc8ce1cf8980c5332615996e5538adee7e4 shards/hfd_third_party_multi_v5/00091.tar +34367e7d484fc3aefe4215173c1101785e9ebb65da566c351c99d53d8f028871 shards/hfd_third_party_multi_v5/00092.tar +92c1c77f7f05e2e929089d635006feb045fe45c7edf2010c7c75e8177e8e0c7f shards/hfd_third_party_multi_v5/00093.tar +455336914fb1e3821534106f941f2546e13cf07f8454f812e1a3fed0ec1e4843 shards/hfd_third_party_multi_v5/00094.tar +172856a0749ad3f406812faa3878097632622c7869d5a14e76e33505104b27ec shards/hfd_third_party_multi_v5/00095.tar +8c7df12f3bc98e70660e95a689dc974114f27db74e8e61052dcac8bd8f68995a shards/hfd_third_party_multi_v5/00096.tar +5689e37f949062daa7268258b07e796c45060c12dfa2dc05a392f228a66743c7 shards/hfd_third_party_multi_v5/00097.tar +f130d0cb5661873cc4734a8d6ac3ed18bd0d950e77e43544a0d72c341ef2a86e shards/hfd_third_party_multi_v5/00098.tar +fa0b85e837fd126d41337607570ecf164123f9871923f07650f4ae5bd22e1fdc shards/hfd_third_party_multi_v5/00099.tar +41084fab083711d71a75c0e55e18135efd9f3f1e6208954e0929223925c36ccf shards/hfd_third_party_multi_v5/00100.tar +2d7eb1e76cb19cc4785a7d6abbeaac401e0a9c79077e0c9d347beedfb5bc8b17 shards/hfd_third_party_multi_v5/00101.tar +9009b46e01572019721fffc81f75f96e9acfd1e7dfd291483e2129500db2740e shards/hfd_third_party_multi_v5/00102.tar +0e278403b429cde12926b37b2bc03b19e0bfa96574e2fcdacd95b754f744d936 shards/hfd_third_party_multi_v5/00103.tar +e4c44beae821cd55db0dc195bbfad4bbb64577000f90eddfa054bc5b97994414 shards/hfd_third_party_multi_v5/00104.tar +e43a251edd9fc1425189dfb4658f1b1e9f07e79b7a6259da3b84a21b1408ed07 shards/hfd_third_party_multi_v5/00105.tar +77fe843e912f1c1682f075a9feaaeb85aec3ee2b5bacabe337d9d92dafc31303 shards/hfd_third_party_multi_v5/00106.tar +432ab61c66614615c4545999f15f67b5e63f81f2fc885e24b923db4809efe176 shards/hfd_third_party_multi_v5/00107.tar +478dedfb846ad42e5df1da3560538882882e2040c97824ad940ddf6d818476c3 shards/hfd_third_party_multi_v5/00108.tar +1bd91bc9e20e28284ab80688416e5f4810e2018e4b2a89bfe54af5a9d49ab9f1 shards/hfd_third_party_multi_v5/00109.tar +00565ea5b764375bc987c764b0d685bbedadf88df84abab5148724dba52fc839 shards/hfd_third_party_multi_v5/00110.tar +03870790c50d943cc37487ad5cfeca0bb37290b3d730ca19c65fe38239607276 shards/hfd_third_party_multi_v5/00111.tar +91e52fbb379e9e7e967b1f2b68f9b234a0d8abc93990e7523c4a2de37a7712e0 shards/hfd_third_party_multi_v5/00112.tar +0a3b9de5e18ff311689c1c8589fe4cebb348548edd5149274f4c61466a9beada shards/hfd_third_party_multi_v5/00113.tar +d2753ebb97ecdb76c38b0aa6e740b987d14314e9abd46c21172e54595a113482 shards/hfd_third_party_multi_v5/00114.tar +ffd07ee7cedd3c50f9a3603aa90f236db6e6c649d2f561ef2a5f532104171aa3 shards/hfd_third_party_multi_v5/00115.tar +d1d12233bdc20e3690664474456c2c041d3118b6a914f065f512b2cb8f7356da shards/hfd_third_party_multi_v5/00116.tar +54660cad25086216a0169743b8d16f39d8b34ab0baed1ebd4b369cf8b5ac1f63 shards/hfd_third_party_multi_v5/00117.tar +b6d42e0d5cb852d4e47072df4ebee01437934c4a75d3f09868af5de7bb2b6cf0 shards/hfd_third_party_multi_v5/00118.tar +7217213264c4094adf23e18566abc44e3e98fc1844749877d3e72d11577d3293 shards/hfd_third_party_multi_v5/00119.tar +c0e4e0b4e809c2e8147aff4122eb5c281497f5da7af081696b8e3a884484c095 shards/hfd_third_party_multi_v5/00120.tar +770a17dda0904f80c0df41774b12a161d9c8ddcb2040518efb7564b182134e2d shards/hfd_third_party_multi_v5/00121.tar +9fc9130e5696b47e228de04ee2b8d64a5bb459c0ca313bccb5cd209ee734535c shards/hfd_third_party_multi_v5/00122.tar +92a88fe95bf95b266d0864c87d1116cd8987b5d5d1ded312eaaa779632bfe747 shards/hfd_third_party_multi_v5/00123.tar +e7483ec65364dffa4011e214c2a2c96b06e21dabb95c555c6a9b2b111cdbabc6 shards/hfd_third_party_multi_v5/00124.tar +5eacf25f9fccaff30b85173cb48e424b8b125a35164adbaeab2450f83f5037d0 shards/hfd_third_party_multi_v5/00125.tar +8bdbd02eed859696fb3457744ed1a4753c3fb8630a5e4063576289c3c2cb9099 shards/hfd_third_party_multi_v5/00126.tar +00b5da5ef62f52c5f6238164c65f43b45c5bae057ed4fdc56eadac9b1f49278a shards/hfd_third_party_multi_v5/00127.tar +21cf81d19a18a6c492b9c7ee5a5dd6614f7ff0256aae19ba76f9a9e667bf3f80 shards/seed_motion/00000.tar +5338b56a86b2b3ee40900ce03d614c049158ad610f7de47432fd6188ce0ae86f shards/seed_motion/00001.tar +9f0e540a37ff741b10ea27ad70162454da2fc4cb65008ff2f485c6259e826ca1 shards/seed_motion/00002.tar +3747593145025aae1eb3c96e0a122f24628d7ac790a0f642357def7d59cf3d2e shards/seed_motion/00003.tar +e71c6a11bd9255d1e11d865b30d7e97a84055baf786aa0f91f4c9e72ac1ce5f6 shards/seed_motion/00004.tar +7d5e4314cf486e077b3e83a0dc52be13a9e9704efd30cbf62f98502adeeeb3e2 shards/seed_motion/00005.tar +09aa3d009dd1a9c07772499233e6275139de29d72ecaf3911069c8df28b651c2 shards/seed_motion/00006.tar +f7f397b4d88d1b584b0918d86e70534206ced6d46788d6dac914a1794de5f22e shards/seed_motion/00007.tar +a6fbb12e25f74b4070ceb9f087025712c49f547893e0a2c3ccc40e840e626a82 shards/seed_motion/00008.tar +ceec1b6144f47a47eca74377eac2f51db1bc33524c2ba7b26c4b0794fdf7fb1b shards/seed_motion/00009.tar +2b179da6ee30f38018500ff77dbdedae956d68bb66d82104303d2880ee1a2083 shards/seed_motion/00010.tar +d247c6349aaa75b24a14878506026422c8d65a30b886e88230033908d4c0d350 shards/seed_motion/00011.tar +3a92184192bf798e061d27d168120e09ce907c2f47dbb38a2d2a7b7c17b80916 shards/seed_motion/00012.tar +24277149065d929068e2a4e4036e23f1349a98c935e53e882d12b97c3f56a11a shards/seed_motion/00013.tar +3b14e0d4a3d5b36494a57d10a6163394b73e8e86801b44d2b028143754476475 shards/seed_motion/00014.tar +306ca7f9d1e2d5648cbb38a7bbe1344298eaa4c68a8bc58e24c3965352e1c26c shards/seed_motion/00015.tar +fab5f20b1d342303f50463df765574886c5e8f611abdadde1a5ccddbbbdeae3c shards/seed_motion/00016.tar +f43cd15b0c9954a6e94185281c3746c9ef13b8d14ef7cfe7eda5120d7ba55778 shards/seed_motion/00017.tar +715abd22cedc53d023f2c04a740bf965fcd4bfb0439f7a0a1eac5c25e43f8582 shards/seed_motion/00018.tar +2dd11c497b5459fd71cd56517f10ca4f1555867edf60ff40c816307777e7a755 shards/seed_motion/00019.tar +a242daceb0f6d4a68cc1082a28f2a44c5e366464c5a0d366107e49ebf0f03189 shards/seed_motion/00020.tar +063135bbe0033f926b8ced9b574ddd28525a0365097831c13613302bbf3fb815 shards/seed_motion/00021.tar +e84eefaa4b6d7cc8e9440c3dda402fdc0546c0c6ddd4dd0dbfb6cc2dccc1280a shards/seed_motion/00022.tar +f33b60722621eaf17db88364c9b8e0da2400018dfcc2451cc6c1f3a6bdb94a36 shards/seed_motion/00023.tar +99f54ede1f37785557c575fc3db068abd8ed1ef976af19147e8654680f313a85 shards/seed_motion/00024.tar +5d10d202a0908ebc8f50b2e601a05fd8c7ae55e9ba508b7cf2ff23691f530cae shards/seed_motion/00025.tar +df9a36ff3abce1683c0d39a7a7ff2eaa0aed75a60692071f3faa3f7d4d68a653 shards/seed_motion/00026.tar +c9d307b00141b84d26f81d56fa67ca621f5253a1309f3076b221dc7a95395d20 shards/seed_motion/00027.tar +88771cc284eca82e4af16d3513345f0fbe85c75bc858bd7dfb6c23d0b1bb16a5 shards/seed_motion/00028.tar +06114bc60070ed7190cfa49738fc0c7d1e65e8143d01251cf105920d7d63bdae shards/seed_motion/00029.tar +be64e5cafb3e0dab83ee66f4efaa0d835378650ae20c149d50cfb218321cd9f3 shards/seed_motion/00030.tar +da6a6564e9bec49a371653a6b6954646caf29e9afb8f986a33976131c62a318a shards/seed_motion/00031.tar +b1edae9a0e2a4c79c3aa5e6b8e69d704dd6a2e18f281676803c5395267c4fb5e shards/seed_motion/00032.tar +be88e19574c38cab69dc8e5bda988b83654bd4a0761f9bca4ee0950e9a97b527 shards/seed_motion/00033.tar +3f9c354420959e58736ad1d7e720993936d2cff7ee3a52b9e4f868d73a094b2e shards/seed_motion/00034.tar +7ba66f29f78494a4b6b78525fbdfe6ed6fdea05cd8c77d97a2fd557209a6746f shards/seed_motion/00035.tar +145852f101081648c524136ac8a72d96996000ed3804b79d753e1bfd0745b6e5 shards/seed_motion/00036.tar +68bb4966d8385a156f16d3c20ef92cd45e84cafd405d5f7b549716cb6ed495e4 shards/seed_motion/00037.tar +53f19ecddd03c8f5042c67f1ad0be76d5c5d7d7aa70b9dc489e698c9b2e27b44 shards/seed_motion/00038.tar +ab84d94fccd047b6354650b3d05e5eddc650219e5c1a8c9495d86a1f13aed381 shards/seed_motion/00039.tar +ab8c0448548c96fada47ffe292b2be144d600657c1f3b8cef8b63e0d41322a3d shards/seed_motion/00040.tar +f66eb6e2ae4c4a6cc6216260770ad5cb0149e5cda302d0ab669daef63775b9cb shards/seed_motion/00041.tar +9cea05baffb1355ea46b4ac7d7266d1f93aede04a03e3404bd5e4bad79ba50da shards/seed_motion/00042.tar +1d6db1ef9b5cfcb33cc8d1be69a6f682a47fdf32eb30b016172efd807ed872a1 shards/seed_motion/00043.tar +5a97edae9a88532f813d4fc2e8773dc617f8d975a13d621f68c2e6521c35cda8 shards/seed_motion/00044.tar +a7cd4e8ff1f888bd04d328df206328214bc58281704d55ac0f8f532cf7fca51f shards/seed_motion/00045.tar +fd878c0fd0ccb0ebc48f35852d427e9fe1774260f8b185b8086793b5ddc9f7c0 shards/seed_motion/00046.tar +3380b93ad872e9b00f159051ab722af7bc5d976cd054ec6e3966fc04427495ed shards/seed_motion/00047.tar +58741f5c277ebd6d79e8f33e66f8afb8a4347e917963448208340ee607306f8d shards/seed_motion/00048.tar +81881903048e4591e5aa3f5faf9c45a56fa43d5e4f4f494077e22ed4f1016ded shards/seed_motion/00049.tar +482a2441cca75764a7aa3b3d4f6053dfc94efe4a0ca36c67df461aa19a7ce320 shards/seed_motion/00050.tar +f4977cb655e30b396d0c8516958db7a23c31ea9da48e24e7e3c74ea097583397 shards/seed_motion/00051.tar +aa61e749a048f4a2f83daf577aa8e5097b1a7300bdc2a83aacba871c48b93968 shards/seed_motion/00052.tar +da7f67fa49b28faf9efaf0882765cbe094cc08347e8f9c3bea20ae16824574dd shards/seed_motion/00053.tar +61c082536825c0ce7f44a4ae1670d9be233794cca9e8aa43a60c3f371a6aa22c shards/seed_motion/00054.tar +d92577fcc1803f49f7d8fcd6a83d7a0b517659b2ab50cb80bd6e32f403148579 shards/seed_motion/00055.tar +97c5efa2e49927cede60611c78a9fdeb9c57dab769c1417f9786d4123d2d2e02 shards/seed_motion/00056.tar +f5167ae8b4c79aa65c2317212ead0932fbc1d085a54b8e48e2a1fe0b1907db35 shards/seed_motion/00057.tar +b839affc1abfb61815417badae9f801f7bbb69678f0ebbe769ab8c75f88ae91a shards/seed_motion/00058.tar +7387a08d744e0d8be6cb6e1d01786e5c8b5a3f79e282c37765655b857de1e33c shards/seed_motion/00059.tar +31307130775d7adbe3bcec12260b59f69bf6a43bc3fe8c8c671a3f95833747d2 shards/seed_motion/00060.tar +4540c3b9e4dc7bb59cdab4e3ddf5ade0bca5b889b6d590cb1fb63d0d06e37a78 shards/seed_motion/00061.tar +df296fbff59b26ebe47ffa2b61a62f9ba90465adee1ad8988c9f99d7b9db6792 shards/seed_motion/00062.tar +0f1ae643872ecdccf412c490974868d07ad9d95b02874694a4686b87f7b902ed shards/seed_motion/00063.tar +8e41f4719ee115be17eeb7b02554014bfc084012a5d24fa75d563d2a065b4598 shards/seed_motion/00064.tar +c08987a76ca1e6b3de56465fcfa73d7a8cd20070dc46cb56b27aad51df7c1d2d shards/seed_motion/00065.tar +d530920fafa72e8a607252efa844ae1881f1fe41be6a1d8014476edafb01f68d shards/seed_motion/00066.tar +bd0a4f6b8a6c7b83cebd9c6283478e7e80b210ea1ec5ad1a2fcd74ca051727f8 shards/seed_motion/00067.tar +503dd2afa88eedaf9d18022b4bc01aa42cfd0a46e2df87034cd4c74fa3f7fa1b shards/seed_motion/00068.tar +890980c5d504999f2c8dc49c199d211b693385f89c48cbbc74b455285e2b63f6 shards/seed_motion/00069.tar +1c4d78b7e3a82fbe6b76d64653a7b3adb2f51ec9f9b3766bbfc9598ff256a7c6 shards/seed_motion/00070.tar +3417d9dd1f46b6527cfc521a6e5c9fab659cd43a73b9ca6467fc1d7795e869f2 shards/seed_motion/00071.tar +2a35377335b065b34692ab3c6e18e338eec04e0bc50bf5322cdc1349d9089d2d shards/seed_motion/00072.tar +8df6d302a6a57c181e50e14893825b7007a77fa2beede30e3a0c0f06331a5b79 shards/seed_motion/00073.tar +747a989cc3ebc7cf4bf5e18552740ab8d0eb557a3ab2facdac1103f0dc83c74f shards/seed_motion/00074.tar +3fadf1856b9d2d727eeb92de45b00d07aea6620fce6813e58a5eec72e27da682 shards/seed_motion/00075.tar +8456893d6256276bc84787008e5abc91b141b98c366b62cc38c9c73f3f8a93e4 shards/seed_motion/00076.tar +47b6184143faf42db503780035d4b4c0d742a5f74793cebe2e8de4ed3a69710d shards/seed_motion/00077.tar +e4cbf15f753b52e51cf0835f3ea75941cface5008dd1bc25f52aff55b000ecc0 shards/seed_motion/00078.tar +552e819b96e3e7822beba76d1c8040b94d94f13e46048b175c04c99c1a14734f shards/seed_motion/00079.tar +214faedf4b15dc2194df4695d367224a6848527ea52a81756cb7378f20518000 shards/seed_motion/00080.tar +9dee1932d4e7b6c5794ae2308411f88e9e2bc1537ecf6c7dd6168ee4cb1a3677 shards/seed_motion/00081.tar +759003e820bf273d2a9ebbece38dfab02ba1f5ca5bf5454b63a651d36a273548 shards/seed_motion/00082.tar +bcfb704c377aa9459c4b38b1cf1746c1b78a12ce43cb2572f87cacb7563ab8c1 shards/seed_motion/00083.tar +3a1f9d376dd1009d764c8cbb6fe452310fe5d9372c3281726b1c54a154e236bb shards/seed_motion/00084.tar +4e08faf876c4db20c598581815515e795da8dcbf89510a8bac7b735450209bde shards/seed_motion/00085.tar +18055ff38d394c94b60eded15ebcff6ceef46bce30f9109d05087a55e6a1f4ef shards/seed_motion/00086.tar +061208830ce73da2cc69cb1877a1bc791a10e02faaee122ebcfb09e6fcd0c57a shards/seed_motion/00087.tar +2eccc99b93d8874e27df46ce4d174778bbcef5bb96774cd3edc833ef02d42df2 shards/seed_motion/00088.tar +bdc3c563069611df03a246221174ab70404a9d755b5233985bed9c0a5bd70bc4 shards/seed_motion/00089.tar +32641649a60488324da3746a36af41ae7abb9fb4fff84caaaee3f307e8f5ac3d shards/seed_motion/00090.tar +25db63f66f0cde752eb5884327186e3123c9f34716f15857047adaa79f61485c shards/seed_motion/00091.tar +d678fb00ae2098c83480c14aeaab525cecf21bd93b75a5beacf687566a28bb7a shards/seed_motion/00092.tar +0073dee08e866ee6f4318a8ca17cbdc8111e11f8650984e31f2cfdfa09bf663e shards/seed_motion/00093.tar +21129b891d9b7d82b931baadbdcf406ed3ef820084e921af75e8154e06fd9971 shards/seed_motion/00094.tar +4feaf812d1e1d06af170252a3a012082e048e9cd5b849ae8d223a73d6b06acb9 shards/seed_motion/00095.tar +07ca36c2ce6a4a24a82d176c3160ada90e33978f395f869cadc1af399cf94f9f shards/seed_motion/00096.tar +89b04682a3790cd53cf2775eeaf972a08ee3e145c263c30e4fa5232af58666d9 shards/seed_motion/00097.tar +0c41cd207a51ca9880aa940f7f12a688b05c5566978b9994f88f9b39faf6ad84 shards/seed_motion/00098.tar +f99448d2ba325b9c32fca8b6382d1851c0b0d75afcf242eeed4c3adcc2841c3c shards/seed_motion/00099.tar +e4fa467331a282a3a09094a01683affe417aeee0705b05503b81451c49391fb8 shards/seed_motion/00100.tar +9f1f136e765761834f832a2cc18e554f1d35a2d10b513d06f03ab76f14d97b0f shards/seed_motion/00101.tar +60c51effd15ccdf0bb1b283e5e38494b12c3a2d834425ba110b64e2c4a5c2dc7 shards/seed_motion/00102.tar +5a98ce312e7bd88625963004686f168fa805d9956f7b8c216ff0957e7817aa1e shards/seed_motion/00103.tar +9e2348ce5e373082a42a8b9d51012eb8a6a625bc1e23a3ee347d53e8357a394c shards/seed_motion/00104.tar +9405ee21016251e2a211a1014a7698c76b5be7e296fa5b76edd061470a6d459b shards/seed_motion/00105.tar +42a95cadb60db431ac757742443ca4de1b10b56b9a8c746d9adb79e1afd20b5a shards/seed_motion/00106.tar +45f04df408c9a3c9c2daa92bdc753efbb692462af7167e8b7056428d6af7a69c shards/seed_motion/00107.tar +6c7f805b6f2f92765f6af232b58029d63feb95f9fcb2c16399e29bc32744b775 shards/seed_motion/00108.tar +9d4fe83efe03d3374306cf19de33aaebde79dc98ccef491ab22b922583ea9f89 shards/seed_motion/00109.tar +fb59c9d2a792d52c1be4204e1a7857d87913b5eb3b563a190dd2254f4e65d95e shards/seed_motion/00110.tar +1582c4da1ecd79d7918524be9ebf857c15428ddef17b8869d7011f778d4fe667 shards/seed_motion/00111.tar +71f4397556ab9ab0bbfb0cd740817fc10e1870c263700b67db0922355c6552f0 shards/seed_motion/00112.tar +28c05dd161033300276641dd3b65403f33d0c871d9d17a7f6f0dd30e4ea7beb8 shards/seed_motion/00113.tar +5aae80b15aa72a65da102d3597edda2b02a268878afbf0b108a7cfa353ad683a shards/seed_motion/00114.tar +bdcfa534afb3f20e37b00ef501dcf73a0afddd1c52503b5ee43e6a125dc001a9 shards/seed_motion/00115.tar +c7e2e8afea2614dfa323da6da31fc233a43b7e39cc9d463f3d61827f2c2f2ce3 shards/seed_motion/00116.tar +cc8976cdb007f204260881f20027f81a445eb88723f9d8ffe9271515217b52cd shards/seed_motion/00117.tar +ab5d146518709809ab81d521e3e41d2077a412416e78b4b48e5261ff0780897b shards/seed_motion/00118.tar +6477bb669bd5a6afc7834a517605bfc2f20b8a7b5ebc29860b9f2834e86430f6 shards/seed_motion/00119.tar +7deb87c945fcd2c5314586bea797cc77a8b4a5aa6cc1a26079f33ec3f2a1eec0 shards/seed_motion/00120.tar +1329f2774b66193d1b8dd8b33657aacbfa34da7f7afcf5222545078e6f67fa26 shards/seed_motion/00121.tar +3302eb76b038b71b8bab8b88528069cec30eaa0b8598fc199bd5e30e8c8756ff shards/seed_motion/00122.tar +4e51262565f40137604bf0ac3161600b75ae01e859f714143227afa2a0ab46b2 shards/seed_motion/00123.tar +66b3bf90c06e2fb4a47856c5c49ba226961b2d19ee65d02037009435f8a2e9de shards/seed_motion/00124.tar +b97d2865d431bcbc222731c639ef14c263d7a157ed8c104c6428ac1688a427b1 shards/seed_motion/00125.tar +74bd78fbc027156d222692bbeed8f935c5f9e039bdfc20fe3cea2eb8601a2b5d shards/seed_motion/00126.tar +51c9b70a50f9b614650488db532bbf6b206f2c1b8652ecdce072894417aa7f1a shards/seed_motion/00127.tar +74dd45dcee765df943a308021ca170b9d9eb31fcf85fddebe085158b426c72c8 shards/seed_tts/00000.tar +9fb9a346cc7e0db05721d047e00c2407b827951180106f0435e7ae1dfb8bad17 shards/seed_tts/00001.tar +e54afc9948f05075654b4560f9c79c16393509ec7bac376ccdfbac1e64659985 shards/seed_tts/00002.tar +ddbca480026e5eec491ee2c2610c3d78e6787ac9b2b36b9f02c9a1dc2a223380 shards/seed_tts/00003.tar +cff147bddc1dbe96722e1161409c14cd8dcf6a2e2954495f8d97c0c217460f40 shards/seed_tts/00004.tar +4b7f54ee0b8edcd4ef71a8043239b5d4b00ec55183ae9f83129e0a9e267ac130 shards/seed_tts/00005.tar +dc4fa2b9ff03f3ed1e0a4cf43dd5b3516e769b84e25aedba7622067ce5d0ea50 shards/seed_tts/00006.tar +90324f53da52c3217f5fde695b9984be6e96707a0786ae04a8c6f7c498e5aa75 shards/seed_tts/00007.tar +d5154d53ffd2d1b02ac008db5b80e5c3829ed7ac755cf71650f3f852a96bcd4d shards/seed_tts/00008.tar +e7630ba7f5face89f34c43bcfd4cb16d2208019c99568f113e4a3d89bdb58983 shards/seed_tts/00009.tar +4c48029287a97ca3a1f616c88c8867ca6fe2d794b550ad7bfbd2320d1ee25a77 shards/seed_tts/00010.tar +a731bce23d5a8d18ea94abcf98af0dbc5276e473301ce76f53a3c7c8a4b4464b shards/seed_tts/00011.tar +5d31941eefe07192883f3736264a4ef8792c2b4f0e1485c6cb2a1918ff5e616f shards/seed_tts/00012.tar +72010e76e654f6859153dd715bed3b1e8fec8f37d76f5b0efe43a406318280ea shards/seed_tts/00013.tar +9b675bc2a3e3b1c21543d3c568efe73e74a585d9fa3e24a9e1d2d317ff8e1a9c shards/seed_tts/00014.tar +8c17135a60cd30632d4f80e5ed07b33bf1d421665f3b501b2d5e3979a1765b1c shards/seed_tts/00015.tar +42d95f56c4c01112327fe6552780c2322b076a35ddca30bee708e3e2f053d703 shards/seed_tts/00016.tar +ae0f52be58fece7479d5926e29643b106f000457252a4380effcb7f14b5e21e8 shards/seed_tts/00017.tar +1273391423362fe33d9719d9ca8c3e544e6177d3420c741614708fce77c43c91 shards/seed_tts/00018.tar +1abff838fcc753fc90cee6b4717b021f3bcac23be6da99e0f521a3d09fce354f shards/seed_tts/00019.tar +5f1993de90570a4699f9c4a9fd5e440640cb6ee0e5dbd21003aa99271cdd22d4 shards/seed_tts/00020.tar +0456e0597d990329dfd102d19075d7203355e5a821a1e7a3f4d60cd5eda5af18 shards/seed_tts/00021.tar +05b017325742e44dfefaad2903ae2ab3583621362a2856861e37b33772af018f shards/seed_tts/00022.tar +6e5ccefaede09e1b141aa60e87aa8a8b6694cc9abdd299ac7d1984832dd2ab1b shards/seed_tts/00023.tar +35a362f1ae5b632794123d8e7c4886f68ed95b5ba459b30e749060101ee9ab9f shards/seed_tts/00024.tar +1652b35a9ae15ca6cc0c73ece227ec4bc512c166c506f167c67e1aa39a226b80 shards/seed_tts/00025.tar +b305e28ac72abc4d6801d8770645162d82bc4e83811df69a3e935813e1a1e92b shards/seed_tts/00026.tar +b73fca2780f218c806b19551288526eae127d5ffc9537c6c7c8bf6d5d00de409 shards/seed_tts/00027.tar +63ec5f57c6207a27d948367d5714a97773c04941b7ce6199466a3a0008a6e585 shards/seed_tts/00028.tar +77b29b13385df8f2c28f773f812b2fe3b9d147d9f4459a7c0506c07a60643abb shards/seed_tts/00029.tar +0ef894eb7429443b79c2743a8dad8609038c88e0247df3a7068ac7848cffa428 shards/seed_tts/00030.tar +57347291516a8e276680b4fd695cf19a57b962bbc99c84f2b2d52d75ebd16448 shards/seed_tts/00031.tar +e54459237d214b7298473f0156e76137ea586e210e5fb2cc5b700ed47b3e7b35 shards/seed_tts/00032.tar +76875967d8d6713cdbdb24b872a3df9f269f5fe686108ecfc718ea2be297784f shards/seed_tts/00033.tar +ad954ab2591d97f8e8c0932ea2479014788451ee8e96be4697743f84b2314241 shards/seed_tts/00034.tar +341b0c29f04124578b13e24691d3daebf11ad771f3958b77297cde6508bc0b13 shards/seed_tts/00035.tar +f7c9fa30c22e30b6902b47536e3dcb74c891ba43642e487bebe211e4948682f2 shards/seed_tts/00036.tar +0859ebf47f55922cbef5c501f7dd918b7a5b658a9d08c0145b70570c3bef6f62 shards/seed_tts/00037.tar +23faa3b19f5b886def1d1c65693b079aa9fd4a5a80bc1e31bc57bd6009cd9dfc shards/seed_tts/00038.tar +ab0dbcf1f6dc9ab1b3c7ae72cb7c6b152ffd749cf1c41d9324ca36231167f364 shards/seed_tts/00039.tar +4eb4e163ee87039e6e8a8048f67b545b20f10351cfa48729b2d38915c9944203 shards/seed_tts/00040.tar +e93710e5908776b8d2dbcae4e9e82b72c89f90c37633b75a260de5212ab7e8eb shards/seed_tts/00041.tar +bbe4cbe5c1bc30914b6bb1bd7d55c267ec8fcfd2e5e2a97dc623cb974496f9f9 shards/seed_tts/00042.tar +700d187b6ea9c009c6254b42bfaa9cfe85a93d9c33ae5050221fba5841d4019c shards/seed_tts/00043.tar +55eb29f0057cb3fce6519f3baa163ec3ca7fd5fc09d7bec125a7c7b8a205eca5 shards/seed_tts/00044.tar +303d7eb90da7a292e9b913fe04cb2a9e6ee59bc0c85168f52d0d5245ec7b588f shards/seed_tts/00045.tar +9afe2f2b3bf5225a36731a27abbc6b357af749cb7f7af5fb5a0b1bd9f038282f shards/seed_tts/00046.tar +55813aa2fd12b05e4078179ca411ee2127f8aa028033607708cd8a711a46ab11 shards/seed_tts/00047.tar +ff0d7cac73a2ca7a7473da2d1ba3d64d8a9916664e8023debcbf51b3e451461a shards/seed_tts/00048.tar +fea504532379bdd32a87fdf131e280bc7850387c1307b532866c4b5c1d25d288 shards/seed_tts/00049.tar +8815ebd38609722c851595a479c502c1e75bf055801121d705c8e01797b7a12b shards/seed_tts/00050.tar +3650a8b24beb6a623652d083d522fbf78035084e5dec33440bdb4b298267d702 shards/seed_tts/00051.tar +517338a96ab01522cf1129002e41993082e68f4c6847c7ea31f04bf5d091e0e1 shards/seed_tts/00052.tar +a5c836dd69583145d576065ee450c0a388a1038bfb5e0e51e54e146bea132adb shards/seed_tts/00053.tar +117fc6882c647aa27df805684fe2970ed42669d47f123b54bcd62776bea1483b shards/seed_tts/00054.tar +57d7af449e2dd846d238e7d0159e1ec01f627415f1fbba848ac7b0fde5c1fc29 shards/seed_tts/00055.tar +7b8fc71a69fea209648085a35f75671c9941472f5b3ab6a373ed3aff45e7d69d shards/seed_tts/00056.tar +a47e1b676044274dc5125658acbaf20056be7836cc75f45deb4c5d3afe2c59d3 shards/seed_tts/00057.tar +973f577f709b9a901b6e5943a792757b8be2631e68b9abef9baf53237ce667c1 shards/seed_tts/00058.tar +9eab4a7b4d3bd7dad796b5698c8338b5dffb7c30d240e70dce5c4e20ccd48feb shards/seed_tts/00059.tar +b54606ba8fe9c134a984a3411705b88eaa0e593e385425bd998160c0c164668d shards/seed_tts/00060.tar +c7347031a33fb75f1a1a3afcff6fa4abeabed75256986a157d0b20074bab591f shards/seed_tts/00061.tar +25c219d466bb484797a78081a8d1bad3b710b12d803445a89ea1106e24a07a30 shards/seed_tts/00062.tar +74b3b80f0a1bd4f7de92318faab0f7cda761bb71a4b3e68bc1d29168f725d0df shards/seed_tts/00063.tar +bc94ad73c4602fa6f27796db386fd0a52b811066a81f6bd29b83093d38dc2c79 shards/seed_tts/00064.tar +0a62f269c76b7348b42bcfd4b1436c686a0f4484329d13440a852569691c5058 shards/seed_tts/00065.tar +16a6d1564af649c31b1f1639fa841d1f32381de72b19ba3b1bbff571abe2f110 shards/seed_tts/00066.tar +54ec30153c1ebb815cafb05d10391cb1553fd723c8a7d98818f42c4e96cd4604 shards/seed_tts/00067.tar +271e0e676f23377701c89fad30328118bf9921d46de46ee0f5f2da32fc0b6752 shards/seed_tts/00068.tar +12f5d299ac387e56fa02b2841c8c3fb4b6d0b585f2408e15f871030c0badb208 shards/seed_tts/00069.tar +c328736a3fae0d369d2bba60fd7dec1ad975473657899bdb5533dd108981338b shards/seed_tts/00070.tar +45398199c8bcfd051ef038ba8dd75c5ed9e5ae11256f3e32daaac6fb07ae6ee3 shards/seed_tts/00071.tar +b66f520eb88a4be2942ce7cf765323963109fe4b1a7b57ba228c5c142a85f343 shards/seed_tts/00072.tar +c338ab11e33e3691528dc5b7382058f73743bec119c0b97a8aa2e5ba459048d1 shards/seed_tts/00073.tar +5e1b5897177c61683b9fb82dec4762cc68b50c802b3f53e9a24e2e27a8271fc3 shards/seed_tts/00074.tar +1b0e1504b68d7548854dfb3ef6971f435c9aecafde5875e669c684215ab917d8 shards/seed_tts/00075.tar +c3845668760e7f8107aa2b83cbc302be09c8cbf4d89c92db85292b17d04d329b shards/seed_tts/00076.tar +c1a0e9bc3233ee56493d228867512c2d8babbca4e7c48903a848c3fd67fcc217 shards/seed_tts/00077.tar +eabb0dd8651cade80aba3d1b3c79271a30ee78a8bf4fe363cdbb1b8f6dee6380 shards/seed_tts/00078.tar +abf502a2f0f535e8b9f7df374b14e825fe4e8d70df5aef84354602dfe213eb35 shards/seed_tts/00079.tar +37eddc1d861b7edeac0b4393534e0fb6ae910654f01c88833691456509fd25e7 shards/seed_tts/00080.tar +53c06931d68dc82136069f3d70096d02dd13c1991118d169f2bfb3b3d6d9ffe2 shards/seed_tts/00081.tar +cdcf6665ed6735daed36ae962309deb92e8de208ce09bd76e473b69230753e83 shards/seed_tts/00082.tar +1bde8b65bf2778eae16d09c41c9ac020135e75e5bfdd8ed740f6140ee9a11d87 shards/seed_tts/00083.tar +aaee475cc31194d726b4fd6655f0ef988315a06f5aa0340bbe09209e7d9305d4 shards/seed_tts/00084.tar +5bd3bf51ef632fdc02cabd2bfd0992813c6fa0493fa930872ad4616c04dd3956 shards/seed_tts/00085.tar +95be5b1284663aa422c8193baf276820fc091284b66c58a95d67258b316f80ac shards/seed_tts/00086.tar +20f5940a37b7fd7f5a731c0f99cde67f9ce75dd3fb780f3048a16a2e735853d6 shards/seed_tts/00087.tar +430c0cbfc548263bd463320665a6adef100ac824fd9b82b3cca097bdf04aee48 shards/seed_tts/00088.tar +e833d0b04b08681c6f5d91c9a298a0da9962e7cfd7b9b55488404e42bab58279 shards/seed_tts/00089.tar +13ccd7c33064b3e613ccc827efc6b2860b8111cfec5422111e4ce293e112f4a7 shards/seed_tts/00090.tar +5cb1c96075f8a08f23655334bb3ab4b4f8c615964f0e5648d13bc5cd19282a33 shards/seed_tts/00091.tar +d1b40bf6df367f862c4be7f2bda2879d6b796e6a2bfe39550e09aad2e31bed53 shards/seed_tts/00092.tar +b6c248443df1a550419f04d00b96cbbce87cb20c9294628df969913de86b0ba2 shards/seed_tts/00093.tar +fb4fee6022520bbe8f0145425d113bc8f18255e3d4d5ed5dfbca902119c15883 shards/seed_tts/00094.tar +efd14bc486beda15fb1c77765f1c92d978a480e2992dd949c2b16e07a9d40699 shards/seed_tts/00095.tar +30036122cad284bf7965868005d41dca854d2d28c68ee03382b30dda8cc94b18 shards/seed_tts/00096.tar +6ab8b2f0818f7fc01c375ef19765d4e98b2666cb9f00b1eae55c4b49e1a015d9 shards/seed_tts/00097.tar +4e8c6a8c387ab98be19e4f8513be9cfabeac8ec38d6b26f6f0c66f739fa0631c shards/seed_tts/00098.tar +6b8dc5fa0488506e6a5dacebb5cab16b9fffa86caf4365667f270794c397bc68 shards/seed_tts/00099.tar +d7541f23e3c5e016de03d1665ae87a344fe04d2cc6340761b235688c5e6da2ef shards/seed_tts/00100.tar +1934fbf8572edd6fb28373221e107f5089a503e62c05bde8d77c9017a4f0666c shards/seed_tts/00101.tar +09bcc199b0bd579b7a92240d2c8ef10a6a8c9b60c3ce0b15548dc42e4ef3e107 shards/seed_tts/00102.tar +5f77b7f72523a1e9a13f8714a16ba8080dc0352b9c148d8d697a0dd78f2c3093 shards/seed_tts/00103.tar +caffaaa3fd5e2452801208bdc088a149a11b38e6303f6f5dcdf4d5bad081a18f shards/seed_tts/00104.tar +40d2aee06909709452424b3b31814492d5f24daf8dd5f61123559af1633b8035 shards/seed_tts/00105.tar +5a4272f4f99b42de1b285fc85d817db7273793eb3e611a8c0ad17deeaaa097eb shards/seed_tts/00106.tar +c11a32e3b61082b744597849efa94501005dbcc18ecc32ecbd9203ae181d8a32 shards/seed_tts/00107.tar +bcf22acf9cd623ce271b561a250a51ee8ec8043b464617361a01f76491bc76cf shards/seed_tts/00108.tar +661c0cc86be3a5f9fee54d97af660bff3835b20d7dff8229efb7f3b1467588d2 shards/seed_tts/00109.tar +17e35fdf88848cf067f0b7d678dba8aedea7e6a1c71a29128f2216c647407507 shards/seed_tts/00110.tar +95d4bb4a06707049b9740f805a7b13b60ac72e572b10a8d489bec40994a2e612 shards/seed_tts/00111.tar +05d7ae03abb5016bc9e68e8d555697f0b04e635d05463de79c3e34023780e5c9 shards/seed_tts/00112.tar +f54834b11ba4263cb5c0529b48595aa3f6d2f71137e0df3c6b2151387bbea594 shards/seed_tts/00113.tar +aee4562e26d24a4eecbd0bf28e264c8f9546a93b71a69a4bfe94ccb2be840521 shards/seed_tts/00114.tar +a451683b4a61eb3c72d4c7159c2241e8d44bb9440abe50fe02db4c24de9aa950 shards/seed_tts/00115.tar +b7a27fcfa54cc14bfb8ed9cfeddbd9eda88d6032adb0479db06b5df27cead7b9 shards/seed_tts/00116.tar +189a8e62e1950bc76004f6970dbbb38659078c9473a29a5aa8f4cba6b1280987 shards/seed_tts/00117.tar +67ccfce28084682261652e73445116befcf8436420bd899f64e32e36d863b05a shards/seed_tts/00118.tar +fb9c3b896b837c764868bd2145e3ae4bdee148a5473286dac5595cbf63283f7c shards/seed_tts/00119.tar +497b12786cd502dbdad32433f30be0c7f3bb97fa929e55b24526d3ea3f2242ca shards/seed_tts/00120.tar +ad8c005fe6c5f0c188af25efd6298fa7b1b20d21a54c9fd66566acfddaba8927 shards/seed_tts/00121.tar +b7368164f305ab3d69bc98b911644a2542aa539199a75428064c3a076b6cc9ca shards/seed_tts/00122.tar +9fe8d886250c5ce599aac286498140c89a90f81ab0cdfee64a73dc550e340ea2 shards/seed_tts/00123.tar +614e1057c643aef914434950aabeb2d06c5b06827e88284fc3718cf1dafc44b8 shards/seed_tts/00124.tar +a92539bba0a22b5cde0ee72181b4dad030eb4aae1402460f078dd29c756ee6be shards/seed_tts/00125.tar +7bf8cf95dbffe74087354cde8c860f758c1bd4427cf23cde942d6f67124f4cea shards/seed_tts/00126.tar +17761d5f32d19285e74703d59bb3df1ecfeaae235cb069a742f2db90fbf7b067 shards/seed_tts/00127.tar +eac457a588a6635785c6a03caa3032e2d1eaae619202dc7fcb98668c792aaca0 normalization/seed_all_g1_motion39_textop_stats.json +a7d91a52dabd142d73a8e93ed86403a5763385b3115ff2217777bf82d2bb492e MANIFEST.json +0e89066ad6758bfbd2811ba497e541aaf2d8f51aeeda3959379b765930f53098 PATH_MAP.json +51295fcb16eb6f81cc1a2a676ae5d3b176d0a4de13c1e763b9329cadd35c1359 README.md +54dfcc3f068174295bf2293dfa5db13aebdd0626cbe17a92230fe5a0473730c1 SMPLX_HANDOFF.md +e41b7ba92d5ccd57092b7c3a190828cc0e8a545c79a8bce575633a68f2bb19bf index.sqlite +63fcd6eb1deb1c270d93a7944f9f2fe95a4433c3bf557cb4ca8671bd18197b10 metadata/seed_g1_smplx_pairs.parquet +b895786ae78f0e6ed58e4fe8bfd7c00c39a33e2fa4365441098b16a8788cdf5d metadata/seed_smplx_summary.json +3ebc253141b89c2dd2d3c6eb4e0470e606d20280ad62ec9aaacc4d205fb276d7 read_hf_audio_motion.py +c77225960a99253727a5ddd7e8f107c20193a97c869efe13cac988d0754f0049 shards/seed_smplx/00000.tar +960665290e76038d2332fbe5c51791a5afe3ffa6fede75955fb5e0cd51fef310 shards/seed_smplx/00001.tar +f17d5e7fc901d933bef30c20d2a471287a0fc5f72c378af12e4f20632634d684 shards/seed_smplx/00002.tar +6ffb6166b1b645d56978641e8c47769aaa700d0402503a9cd71bdee61a67beb5 shards/seed_smplx/00003.tar +4656b1450fe8d88be36b42c0ce303e5747090769fb67ee190fb7dab7b07b62e7 shards/seed_smplx/00004.tar +df665aa23f1c93fd0b901bff5168ab410e638fe440066620ab151aff18dbfebf shards/seed_smplx/00005.tar +c48246cdbf9dfaa2519dd96b1dd6fde6314261da181425ccc60a6145fb6bdfb4 shards/seed_smplx/00006.tar +2ed728c78fd0ed1a2e7af4a11fcd45c1d96504c54d8ba96ad90dd23a01cf4902 shards/seed_smplx/00007.tar +fca0c18e9a0e898aca139baebfb2fdc72d7a87ba1da57be400beeb75adeec852 shards/seed_smplx/00008.tar +33beb447513d524f1a48865304a2b0af6aae2a81b903c206b5922ceb487a5229 shards/seed_smplx/00009.tar +82597e5016c3ce7c533f7669ee2c77f3f7a01e71c7a6bdf3651f0e0c364f231f shards/seed_smplx/00010.tar +09e566d83072ff391c169306f5495095ed4e4bcf24140e02e830a6f1ac09ba3f shards/seed_smplx/00011.tar +d170c787805c095addf204a88270e5189c02b17aa1be971ec291a4397cba3f40 shards/seed_smplx/00012.tar +3737a7912ddee59210f17da6e8cf71f131c036cad41f83166de29b3668ae281b shards/seed_smplx/00013.tar +d60e45433d3fc0205678a2b89d58db415f69137c47ed9cf240e1a4749048f95a shards/seed_smplx/00014.tar +d9c736c22c2808f20bf33b7d436da4b23dd7584ee2cb0b2a7d468fbd0853c989 shards/seed_smplx/00015.tar +0c37436f68827291643ff722738608f0929b6b491e8c43b32dae48ba139868bf shards/seed_smplx/00016.tar +bc95fe2f4cb88b1edd3a1d80c75a41fe0afad7f6054df2d7d7d80ea1b52b522a shards/seed_smplx/00017.tar +ed37807c0fec627f8259120357808fa9b2c0c91ca8994c29f3865aa5e440e79c shards/seed_smplx/00018.tar +f38813b8a5b3236b4de71942777bbd7112f9d0b37cf93a0205e71a8115f49c17 shards/seed_smplx/00019.tar +ce2f61e2f62575e4961e59b23de82c1bd05a3563f7b4b090f510e3636e543ad3 shards/seed_smplx/00020.tar +3def28eb4b3aae483f3f4905436c2c0dd4dcf5fc64fce722953fd375aa3b7582 shards/seed_smplx/00021.tar +0050b53039d71629e4379ca2cf24e5293f412a117f45a0ae857521afc5e7d740 shards/seed_smplx/00022.tar +7fcf9bb96cb4cd49774be2d88e2c407b23e26faddb225147d308a68cee24a361 shards/seed_smplx/00023.tar +0e92301b73128a8cfddf0ee7fe38f426dffe569521af6e792ea02ae06154d4b9 shards/seed_smplx/00024.tar +65e4d5933ced1d2fdf303329e497deb8cf0c4a6745f4bfc6d32f52d373747bfa shards/seed_smplx/00025.tar +0de868ecad5293e0e4300baf7c53265fe0fa2071ac0abfe31519a56931842118 shards/seed_smplx/00026.tar +fabe11361ca368ec0f9c211561cf574adad08c8e23e0b16d431844d402f541f8 shards/seed_smplx/00027.tar +472ca80c586fe59df26c1f2b3f836143fbdbac90507dcd81205da30b0c963b12 shards/seed_smplx/00028.tar +82d940ab1e86a8aa92901ea85473a0a2bd7e0512a62b8c7673dc5719582cecc5 shards/seed_smplx/00029.tar +10740e1af2cbab75b085253e1c5e4b73090eaafccea745792c2309d1e04c00d3 shards/seed_smplx/00030.tar +cfb780400cefbf2f4f92ed5a7978bfaf58083d85b8487dc3e237147171a80bbc shards/seed_smplx/00031.tar +10c4f9c2efebe694ac19cbe99ee67a890088e63e46dcc49e2c9f3781b4c0e273 shards/seed_smplx/00032.tar +5c0a1159d1a522ab2a0cd7be5434a35418d92ea73b459cc475db99962c44799c shards/seed_smplx/00033.tar +d6b81f265f94657d75a20c3dca94fb4d2a9127cee0ac6cad24d332c1440da995 shards/seed_smplx/00034.tar +11e260f2b74c21e6403e7c666451ed0f4cf3ced3a73a0d62bed5693b524bb29c shards/seed_smplx/00035.tar +a03520957193aaa5757761b270b96fdbbb77d7537e1a75efb8686f11a39a7a41 shards/seed_smplx/00036.tar +1967cf13849cc87c2481c3de38af8bfb8d8f40bc80cf4ab6b051e07a1e10b74c shards/seed_smplx/00037.tar +0028fae70cf71c15a14b72116ad55f62fd84750f5725bcfac4b384c6dd225100 shards/seed_smplx/00038.tar +9ddfd341bebf2c0e29870f7939fd76ddec4583729355994686189f504ad412e9 shards/seed_smplx/00039.tar +85e3b584987c1ff99ac219d065e279d0f42ab9cd7f803df72214655c7fd07c8c shards/seed_smplx/00040.tar +92483966167a025ae5dbd419e5518930c9a85c25fd72680200e08b932f64f5ea shards/seed_smplx/00041.tar +88e41948d7e4687ec8602042263fdb0c6f0562328847e94d9af83f8fa2c27e0d shards/seed_smplx/00042.tar +83ed7365600362b947e179ce6b2f0002a0f5a20398ed65a61e61331c3cc63b39 shards/seed_smplx/00043.tar +b21e34892ff735418aec1cc9d4197bc4026ec18f09b7b9ab23d9e969e59a7b4b shards/seed_smplx/00044.tar +81ba40d03fe239a51d88d7335c3cbf059227b6b79dadf6cb80192b20e204195b shards/seed_smplx/00045.tar +4f9de85ae5aac1bc7a043e7b61beaeab0434ac31d0a2b2242151dca50f18d33a shards/seed_smplx/00046.tar +026e8d87642ea5a89481f64d28eb6a07c66cc4c018664b7e5a2a48f3281f7ff9 shards/seed_smplx/00047.tar +cf8a9058fb02c9ee3ecbe620a4968d8a617b422f66e9cfd9ece864fd4ee96a01 shards/seed_smplx/00048.tar +708cdc87d3b99eafe57c73c2d3604848f482d7a2946efaf2668210462527b126 shards/seed_smplx/00049.tar +e56f829fe85b7ede4c88cb2a4c7d4955fefaba31a7bf0d30ee288f57bbdfb978 shards/seed_smplx/00050.tar +85d1d8662f80b0eb46e5e56f49c20feefc44fd3c7d3dd259492888138ee20ac2 shards/seed_smplx/00051.tar +82440c6f49e7aaeb3c8ec6cead9e3045dd03ef6de5d214a8f9214fda5dcb5565 shards/seed_smplx/00052.tar +f06bfe33fd661d251c2b29e2b671f9dd0aba941f0e930ef92da7a3f21dd2c256 shards/seed_smplx/00053.tar +5c90166e4a803747a4c7277299b34bb91a4f2e71af6aa766d518350ea0bd89da shards/seed_smplx/00054.tar +4b5004e73ea05115c2cbb3597827484dd647531f94b5e6400f30ae9c21b2c4ff shards/seed_smplx/00055.tar +9c005b42393d3569affd484d86b8f9034e5401248ccf15f578453043f7880d75 shards/seed_smplx/00056.tar +7a791a918b32a7f2ea30d77847a3449e7b8d90fb17014ad9fe4e06911f5ea6e8 shards/seed_smplx/00057.tar +6e8c4ba6e4c4187dad42594e71194c6d734b49e40d92445329c6e90e5b264bcd shards/seed_smplx/00058.tar +3ebea086764145ae33bb48f6e8861e8ab083e05bf270afecf7bf85539e625e70 shards/seed_smplx/00059.tar +beeab8f7ec52274f42e0b78df40b28dc96b8e9987ba4123ebfcf6d862e20437e shards/seed_smplx/00060.tar +2fb05a24c70b2d0b5c4c48133702b703c445af7d21696e341acdbbb23ccac4bf shards/seed_smplx/00061.tar +9645e137db730ab098f096af1db2cce8a60a8f4da905cc05f250f5e65d694b1c shards/seed_smplx/00062.tar +d8f320627e7269ae17197a758b722bda249325f3414fcc9777bc3b2a2c9a4b63 shards/seed_smplx/00063.tar +43bb73a9c7ab85b2846ae44dc5fe42c222b922cbbbea31de4eace1fbad28e108 shards/seed_smplx/00064.tar +35c579ccb9b575669f077e7b488c0b53960adcec60fe02bbf9ce2cd4b1c90c2d shards/seed_smplx/00065.tar +c64d585c47ad7f184219e777aeff69b5386487c72cd0b4656343226d8b4b6437 shards/seed_smplx/00066.tar +79dcb210eed238622e7e07c1f900069d2c1c0a8d651499bb4ee0358f312eb294 shards/seed_smplx/00067.tar +26ab96af081305870964e097c816ca5a9e8b673543fa5fd3d91a212bb44fef28 shards/seed_smplx/00068.tar +affea53d462505153b58a66b4ced7cf23f075f454444405147e0beb5ae7f3e92 shards/seed_smplx/00069.tar +dc07436d8a91687082585af00363ee79ce5f43339f5a018a88b4584736072dbe shards/seed_smplx/00070.tar +56b5ecd195c2d53849c09274d54acdecb1d1327761c73f5f9ed84281897ac34e shards/seed_smplx/00071.tar +7e8885e9625f362f90d98366afea21f48bc97124cf02ac459525c10ea31c2f8b shards/seed_smplx/00072.tar +8bde5e66bd9434f6f7e86c62f706b553a73b6a09aa76cc78320a9455c5469fcd shards/seed_smplx/00073.tar +7bd167ed7793d0b4ca8953bd175ca1c542e8ee4511387b7f679a3794fcdf1d49 shards/seed_smplx/00074.tar +74ccb501400795638eac256e483a1303f6cea1f769a15e8d343fefaf06828867 shards/seed_smplx/00075.tar +8a516b737aa571e9adc0c3fee766cbd96252b0dfa090545d2cd0c1ca7f8102a6 shards/seed_smplx/00076.tar +445822226c2dd04597de0c231fad4de6a279bc58efa2fdcd6aee48ef017431dd shards/seed_smplx/00077.tar +52439b22771d2f2fe1546afaba22b34f7080aec3a935c1343c13bd8a04c92cc6 shards/seed_smplx/00078.tar +a88eb8cb09c3dd3128dfacf8e29202af21abeed9ebb283974dc4921632b5e18b shards/seed_smplx/00079.tar +22ff13ad98d3db2cfc79f3c306aa4f3dc915e2baeddfe76d76c57e1ecdf13bf9 shards/seed_smplx/00080.tar +1d7d8c2d75cdaf62d4e652deb3a840ae6187a8b29f5e9d3d1280afa5fc276f0f shards/seed_smplx/00081.tar +75d76380e58a10e5698dedea5f524868b18be5272e63db8665fde9dba62ce961 shards/seed_smplx/00082.tar +9f1fde4275b33d94e0bd826aa7cdcd2a980f28b0786ae213b67a7b0a757f1d5d shards/seed_smplx/00083.tar +0bf9b2320f4caecf91c5d3774e936ef5cec363f80947516d82eda50c4f9800bc shards/seed_smplx/00084.tar +cb383310fdcc234b9080685498c56ca2c510186ae0716ea81e97e11631fdf162 shards/seed_smplx/00085.tar +add1d09b7dd494b77ff1fc24c36892369c6955bdcc5dedb8d4255e9313156510 shards/seed_smplx/00086.tar +c512036cae87dbfc26558bcd40009639e12ad2797826933c368030e8b42c1470 shards/seed_smplx/00087.tar +7d8123f241eaa5f8ba3784f1fb119126e006fb8d2a014747c7c3f7e9215d98f0 shards/seed_smplx/00088.tar +3423b19b76baf6297c4e411fa952548213c021c13fd2106833e3caa3e1770b9c shards/seed_smplx/00089.tar +a30489f2fd1b8d3c3af87e1df341a5fdb623f3fe41e48b67c52537a9aa76f1c7 shards/seed_smplx/00090.tar +7ee8197f163e96a15958dbc963b545ace07c0496c43847f74838ef20fea2542b shards/seed_smplx/00091.tar +de8defa5a90437c6302c5db4116ccc53acdb319819deb5b0ccaad087ef706aee shards/seed_smplx/00092.tar +3bb4ce0c9d802fc3a053cc04b0d2b94752e146c4acc928485c0543890d8e912b shards/seed_smplx/00093.tar +796bf0b686988ce703cbd88d727a1f36239afcdc3efea713c6381bf9d03459d3 shards/seed_smplx/00094.tar +ec2f5ea4f70050877465aa739f7e6c55afb4682f37c54c3cd035022b2173f95d shards/seed_smplx/00095.tar +ff768278abdc0b4264213d373ecf48d53411b412383fcfdebaec2ac7a9b1e631 shards/seed_smplx/00096.tar +b30aba94f12abe0c33289b553b68cf9768bfbe2cf523dded564e13dcb676ea41 shards/seed_smplx/00097.tar +0ebffb2371584be72009a17a2d2d862a0925e12a574b1b78a2005fe2ddfc7271 shards/seed_smplx/00098.tar +7ccab2fd5a83b2f27d03dc42b53c41265ff92c935cf1124bbc7636cf7092da5d shards/seed_smplx/00099.tar +80b3846c2d36f6d553046ed754c917c3a439d20cc03f96e70b625d356bfc1700 shards/seed_smplx/00100.tar +44fbd0d33d6677ac32e7733d9fe85bab1a1721c49e0cb821217a1ec05d00d9c0 shards/seed_smplx/00101.tar +d2bde9bcf2be9dec45acb0b9832fce6ce0cd110a404d9163dbc133eeca4b163d shards/seed_smplx/00102.tar +3bbd14328ebc00579004247032b9db6a8ca291a389ae5800a76a027b0f3f2ef3 shards/seed_smplx/00103.tar +2b09632d4e6ddf06b007f14d5b38ef14c50ae17c49f960e7d12527bfcf1607eb shards/seed_smplx/00104.tar +03d9900c9197b641fc63df498f8a0f4bc083aa393d2f46e2fa8d16a532a53f12 shards/seed_smplx/00105.tar +2024bb5757a9c0517a387a5eb638f8e2a54fa2628e4ff7bbb3bbf4b43beef1d1 shards/seed_smplx/00106.tar +adac2af366efd39aaa1adf21abf4a86e199b2fb1cf61f7ad6548b5919b4354e2 shards/seed_smplx/00107.tar +d97e07f76068e0c4f50c06db41a0100ef72c15bd391ce8c8f388310c5c1c8498 shards/seed_smplx/00108.tar +b8572df590f8846e4499cf4c163ad72a7a7d8c8a6bd6227e074135b3082ff19a shards/seed_smplx/00109.tar +24d9748947b6c0d68cf626c448fd2570a913d504b2b479bcca4b9fec1ef83d2b shards/seed_smplx/00110.tar +de034737374beab7141048d5a6d7bca589a3e965b40307a59c62d2bfca077f35 shards/seed_smplx/00111.tar +342f203aeecac3ef6e6824e5bcf6d3157633cb6b1e99e2a4ed1a12ca6409c039 shards/seed_smplx/00112.tar +411cd69678a545e3522dcb8e93e9fe1f12cbd7852edcc63b08b5dad5f90fefa6 shards/seed_smplx/00113.tar +82ab9f50d717692125f013679d568684dd7d900533c7168d65d56f2af5de5b1c shards/seed_smplx/00114.tar +d94bc16522683787acc2b1cc201f262be611ccfa5a57a0e03f2178e099f4f9c3 shards/seed_smplx/00115.tar +30131142e712ca3d36562bbf20db1a353e081ce6279a59d3dcf5cf6f3a600205 shards/seed_smplx/00116.tar +b0566cd2fac2aa22fe3f2ac986c944cc2fae4330e04a357e1cb6d0ca2cfbf1ef shards/seed_smplx/00117.tar +d0005dcb1ed8618e9bed50820f947e0e5cf361342ef28d3fcbb95e373e66d3d7 shards/seed_smplx/00118.tar +4a066e2ff705a63af50185ee15f6988089aaf22529481b56cb91d3526cd7ad2e shards/seed_smplx/00119.tar +302f21eb04a633d6ad4247e69301ecae3b809e3809406e2fb81576d7ddf039f6 shards/seed_smplx/00120.tar +e385f363a274e36708b4410fccc95fce239164fc11267222fefd2ad0fe675423 shards/seed_smplx/00121.tar +e79d43633fac92569bb35a84f58a8c40cb4b4ed1e91f7697a1f03175fbb9c18e shards/seed_smplx/00122.tar +ed5c7bee0e4c1660025bd84038d5ddeddd3962a92984d94af78a0dd8012a720a shards/seed_smplx/00123.tar +973accb6920a8d95328c654e739c7877996e57b37c3ca9494be98b9e90962f59 shards/seed_smplx/00124.tar +c19aeddc48d32aef1ba0c5617a73e8a236dd26645c226e50fec6f2443c487d44 shards/seed_smplx/00125.tar +44193111236847a56910119f53d30909c7f16ee37be777ef180c789c9db0d7ff shards/seed_smplx/00126.tar +08838d1b699443124f18283f627f8fc9c7479af2a187a54831c81963840ecc3a shards/seed_smplx/00127.tar diff --git a/handoff-v1/resources/all_baselines_acceptance_entry.py b/handoff-v1/resources/all_baselines_acceptance_entry.py new file mode 100644 index 0000000000000000000000000000000000000000..f0d1f8d5a8ae521ab313cc6b24f1c7e61d707a11 --- /dev/null +++ b/handoff-v1/resources/all_baselines_acceptance_entry.py @@ -0,0 +1,30 @@ +"""Observe original native loop; TextOp stops only after a complete 4-primitive chain.""" +import sys,os,json,runpy +from pathlib import Path +W='/root/autodl-tmp/hfd-baseline-portable';sys.path.insert(0,W) +R=Path(os.environ['HANDOFF_ACCEPTANCE_DIR']);target=int(os.environ['HANDOFF_STOP_STEP']) +from hfd_bridge.ondemand import WindowPool +original=WindowPool.next +def observed(self,*a,**kw): + rows=original(self,*a,**kw) + with (R/'sample_ids.jsonl').open('a') as f:f.write(json.dumps({'pool_step':self.step,'split':getattr(self.provider,'split',None),'rows':rows,'note':'Fetched logical batch; prefetch may include unconsumed next batch'})+'\n') + return rows +WindowPool.next=observed +if os.environ['HANDOFF_MODEL']=='textop': + sys.path.insert(0,W+'/upstream/textop/TextOpRobotMDAR') + from hfd_bridge.textop_checks import PipelineDARManager + old=PipelineDARManager.post_step + class AcceptedStop(Exception):pass + def post(self,is_eval=False,**kw): + old(self,is_eval=is_eval,**kw) + if not is_eval: + import torch + row={'step':self.step,'loss':{k:float(v.detach().cpu()) if torch.is_tensor(v) else float(v) for k,v in kw.get('loss_dict',{}).items()},'lr':[g['lr'] for g in self.optimizer.param_groups],'optimizer_steps':sorted(set(int(s['step']) for s in self.optimizer.state.values() if 'step'in s))} + with (R/'paired_steps.jsonl').open('a') as f:f.write(json.dumps(row)+'\n') + if self.step>=target: + assert self.step%self.resume_num_primitive==0 + self.save_model();raise AcceptedStop() + PipelineDARManager.post_step=post + try:runpy.run_path(W+'/train.py',run_name='__main__') + except AcceptedStop:print('HANDOFF_ACCEPTANCE_COMPLETE',target,flush=True) +else:runpy.run_path(W+'/train.py',run_name='__main__') diff --git a/handoff-v1/resources/all_baselines_acceptance_supervisor.py b/handoff-v1/resources/all_baselines_acceptance_supervisor.py new file mode 100644 index 0000000000000000000000000000000000000000..1a543b44aab90ceef82cd6817df115f81cdf6035 --- /dev/null +++ b/handoff-v1/resources/all_baselines_acceptance_supervisor.py @@ -0,0 +1,31 @@ +from pathlib import Path +import os,sys,json,subprocess,time,signal,traceback +R=Path('/root/autodl-fs/pangxinlei/handoffs/all-baselines-20261111');W='/root/autodl-tmp/hfd-baseline-portable' +sys.path.insert(0,'/root/autodl-tmp/hfd-motion-evaluation/HFD_bench/scripts/recovery_training_v54_kimodo') +from storage_guard import Reservation,atomic_json +jobs=[('textop','hfd-textop-english-split20261026','ckpt_299776.pth',299780,3),('omg-l','hfd-omg-l-english-split20261026','checkpoints/step-00098000.ckpt',98002,24),('omg-xl','hfd-omg-xl-english-split20261026','checkpoints/step-00090000.ckpt',90002,36)] +for m,run,ck,stop,gib in jobs: + root=R/'acceptance'/m;root.mkdir(parents=True,exist_ok=True);g=Reservation(root,gib*2**30);p=None + try: + assert int(subprocess.check_output(['nvidia-smi','-i','1','--query-gpu=memory.used','--format=csv,noheader,nounits'],text=True).strip())<100,'GPU1 busy' + atomic_json(root/'capacity.json',g.replenish(physical_probe=True));g.release() + family='textop' if m=='textop' else 'omg';source=Path('/root/autodl-tmp/runs')/run + cmd=['/root/autodl-tmp/envs/'+family+'/bin/python','-u','/tmp/all_baselines_acceptance_entry.py','--config',str(source/'resolved_config.yaml'),'--resume',str(source/ck),'--set','runtime.gpu_ids=[1]','paths.output_dir='+str(root/'run')] + if family=='omg':cmd+=['paths.checkpoint_root='+str(root/'weights'),'training.stop_after_steps='+str(stop)] + env=os.environ.copy();over=dict(HANDOFF_MODEL=family,HANDOFF_ACCEPTANCE_DIR=str(root),HANDOFF_STOP_STEP=str(stop),HF_HUB_OFFLINE='1',TRANSFORMERS_OFFLINE='1',HFD_ALLOW_RANGE_DOWNLOADS='0',PYTHONDONTWRITEBYTECODE='1',OMP_NUM_THREADS='4',MKL_NUM_THREADS='4');env.update(over) + atomic_json(root/'command.json',{'argv':cmd,'cwd':W,'environment_overrides':over}) + with (root/'stdout.log').open('w') as f: + p=subprocess.Popen(cmd,cwd=W,env=env,stdout=f,stderr=subprocess.STDOUT,start_new_session=True);start=time.time();atomic_json(root/'process.json',{'pid':p.pid}) + while p.poll() is None: + g.check() + if time.time()-start>1800:raise RuntimeError('30min acceptance timeout') + time.sleep(10) + atomic_json(root/'exit.json',{'returncode':p.returncode,'seconds':time.time()-start}) + print(m,'exit',p.returncode,flush=True) + except BaseException as e: + if p and p.poll() is None: + os.killpg(p.pid,signal.SIGTERM) + try:p.wait(60) + except subprocess.TimeoutExpired:os.killpg(p.pid,signal.SIGKILL);p.wait() + atomic_json(root/'error.json',{'error':str(e),'traceback':traceback.format_exc()});print(m,'ERROR',str(e),flush=True) + finally:g.release() diff --git a/handoff-v1/resources/checkpoint-layout.json b/handoff-v1/resources/checkpoint-layout.json new file mode 100644 index 0000000000000000000000000000000000000000..68ac920f38cb901c39e0fb75672857aac018a6b2 --- /dev/null +++ b/handoff-v1/resources/checkpoint-layout.json @@ -0,0 +1,14 @@ +{ + "run": "/root/autodl-tmp/runs/hfd-omg-english-split20261026", + "files": [ + { + "repo_path": "handoff-v1/checkpoints/last.ckpt", + "original_path": "/root/autodl-tmp/runs/hfd-omg-english-split20261026/last.ckpt" + }, + { + "repo_path": "handoff-v1/checkpoints/best.ckpt", + "original_path": "/root/autodl-tmp/runs/hfd-omg-english-split20261026/best/best.ckpt" + } + ], + "checkpoint_root": null +} \ No newline at end of file diff --git a/handoff-v1/resources/download_raw.py b/handoff-v1/resources/download_raw.py new file mode 100644 index 0000000000000000000000000000000000000000..90b154c8b78799beff7ed30c0db8f068fc0412d0 --- /dev/null +++ b/handoff-v1/resources/download_raw.py @@ -0,0 +1,6 @@ +import json,sys +from pathlib import Path +from huggingface_hub import hf_hub_download +a=json.loads(Path(sys.argv[1]).read_text()) +for x in a['files']: + hf_hub_download('Tinghong-Ye22/hfd_data',x['path'],repo_type='dataset',revision='93ba4a7a9c66b3e1b7041e5f4ea038ef2418b948',local_dir=sys.argv[2]) diff --git a/handoff-v1/resources/environment-lock.txt b/handoff-v1/resources/environment-lock.txt new file mode 100644 index 0000000000000000000000000000000000000000..16a94c8f01340df124fbabfc1c7bb3e13524c38e --- /dev/null +++ b/handoff-v1/resources/environment-lock.txt @@ -0,0 +1,119 @@ +absl-py==2.5.0 +aiohappyeyeballs==2.7.1 +aiohttp==3.14.3 +aiosignal==1.4.0 +annotated-types==0.8.0 +antlr4-python3-runtime==4.9.3 +anyio==4.15.1 +async-timeout==5.0.1 +attrs==26.1.0 +certifi==2026.7.22 +charset-normalizer==3.5.2 +click==8.5.0 +cloudpickle==3.1.2 +coloredlogs==15.0.1 +contourpy==1.3.2 +cycler==0.12.1 +datasets==4.8.5 +dill==0.4.1 +etils==1.13.0 +exceptiongroup==1.3.1 +filelock==4.0.7 +flatbuffers==25.12.19 +fonttools==4.65.0 +frozenlist==1.8.0 +fsspec==2026.2.0 +gitdb==4.0.12 +GitPython==3.2.0 +glfw==2.10.2 +h11==0.16.0 +hf-xet==1.6.0 +httpcore==1.0.9 +httpx==0.28.1 +huggingface_hub==0.36.2 +humanfriendly==10.0 +hydra-core==1.3.0 +idna==3.20 +ImageIO==2.38.0 +imageio-ffmpeg==0.6.0 +importlib_resources==7.1.0 +iniconfig==2.3.0 +Jinja2==3.1.6 +joblib==1.6.0 +kiwisolver==1.5.1 +lightning-utilities==0.15.3 +MarkupSafe==3.0.3 +matplotlib==3.10.9 +ml_dtypes==0.5.4 +mpmath==1.3.0 +mujoco==3.7.0 +multidict==6.9.1 +multiprocess==0.70.19 +networkx==3.4.2 +numpy==1.26.4 +nvidia-cublas-cu12==12.8.3.14 +nvidia-cuda-cupti-cu12==12.8.57 +nvidia-cuda-nvrtc-cu12==12.8.61 +nvidia-cuda-runtime-cu12==12.8.57 +nvidia-cudnn-cu12==9.7.1.26 +nvidia-cufft-cu12==11.3.3.41 +nvidia-cufile-cu12==1.13.0.11 +nvidia-curand-cu12==10.3.9.55 +nvidia-cusolver-cu12==11.7.2.55 +nvidia-cusparse-cu12==12.5.7.53 +nvidia-cusparselt-cu12==0.6.3 +nvidia-nccl-cu12==2.26.2 +nvidia-nvjitlink-cu12==12.8.61 +nvidia-nvtx-cu12==12.8.55 +omegaconf==2.3.0 +# Editable install with no version control (omg==0.1.0) +-e /root/autodl-tmp/models/OMG/OMG +onnx==1.23.1 +onnx-ir==1.0.0 +onnxruntime-gpu==1.23.2 +onnxscript==0.7.2 +opencv-python==4.11.0.86 +packaging==26.3 +pandas==2.3.3 +pillow==12.3.0 +platformdirs==4.12.2 +pluggy==1.6.0 +propcache==0.5.4 +protobuf==7.36.2 +pyarrow==21.0.0 +pydantic==2.13.5 +pydantic_core==2.46.5 +Pygments==2.21.0 +PyOpenGL==3.1.10 +pyparsing==3.3.3 +pytest==9.1.1 +python-dateutil==2.9.0.post0 +pytorch-lightning==2.6.6 +pytz==2026.4 +PyYAML==6.0.3 +pyzmq==26.4.0 +regex==2026.9.29 +requests==2.34.2 +sacremoses==0.2.0 +safetensors==0.8.0 +scipy==1.15.3 +sentencepiece==0.2.2 +sentry-sdk==2.71.0 +six==1.17.0 +smmap==5.0.3 +sympy==1.14.0 +tokenizers==0.22.2 +tomli==2.4.1 +torch==2.7.1+cu128 +torchmetrics==1.9.0 +tqdm==4.70.1 +transformers==4.57.6 +triton==3.3.1 +typing-inspection==0.4.4 +typing_extensions==4.16.0 +tzdata==2026.4 +urllib3==2.8.0 +wandb==0.26.1 +xxhash==4.0.1 +yarl==1.25.1 +zipp==4.1.0 diff --git a/handoff-v1/resources/final-resume-historical-source/hfd_bridge/genmo_prefetch.py b/handoff-v1/resources/final-resume-historical-source/hfd_bridge/genmo_prefetch.py new file mode 100644 index 0000000000000000000000000000000000000000..e9445d10ed1f77c9ada3ef9cce0a20a5cc3d85e6 --- /dev/null +++ b/handoff-v1/resources/final-resume-historical-source/hfd_bridge/genmo_prefetch.py @@ -0,0 +1,29 @@ +"""One CPU producer with bounded lookahead and consumed-batch resume state.""" +import copy +import os +import torch +from torch.utils.data import IterableDataset + + +class CPUBatches(IterableDataset): + def __init__(self,source): + # Serialize only configuration and sampler state, never SQLite/CUDA. + self.options=copy.deepcopy(source.provider.options) + self.state=source.pool.state_dict();self.batch_size=source.pool.batch_size + self.max_batches=source.max_batches;self.refresh=source.pool.refresh + self.output=str(source.output);self.recovery_fraction=source.recovery_fraction + + def __iter__(self): + os.environ['CUDA_VISIBLE_DEVICES']='' + torch.set_num_threads(2) + from .genmo_data import PairedNativeProvider,GenmoBatches + provider=PairedNativeProvider(**self.options) + data=GenmoBatches(provider,None,self.batch_size,self.max_batches,self.refresh,self.output, + recovery_fraction=self.recovery_fraction) + data.pool.load_state_dict(self.state) + for batch in data.cpu_batches(): + # T5 runs in the GPU process; don't transport zero placeholders. + batch.pop('text_embed',None) + previous=batch.get('_hfd_recovery_previous') + if previous:previous.pop('text_embed',None) + yield batch diff --git a/handoff-v1/resources/final-resume-historical-source/hfd_bridge/genmo_training.py b/handoff-v1/resources/final-resume-historical-source/hfd_bridge/genmo_training.py new file mode 100644 index 0000000000000000000000000000000000000000..3f094ed463c6a769c18570856dac7f2132073b27 --- /dev/null +++ b/handoff-v1/resources/final-resume-historical-source/hfd_bridge/genmo_training.py @@ -0,0 +1,132 @@ +"""Execution-weighted GEM training and causal, detached predicted-history recovery.""" +from contextlib import contextmanager +import torch +import torch.nn.functional as F + + +@contextmanager +def evaluation_mode(model): + states={module:module.training for module in model.modules()} + model.eval() + try:yield + finally: + for module,state in states.items():module.training=state + + +def execution_mask(mask,history,execute,weight): + """Preserve each feature's total loss mass while emphasizing executed frames.""" + result=mask.float().clone() + result[:,history:history+execute]*=float(weight) + result[:,history-1,148:151]*=float(weight) + total=mask.sum(dim=1,keepdim=True) + return result*(total/result.sum(dim=1,keepdim=True).clamp_min(1)) + + +def decoded_positions(model,x,origin): + from .genmo import rollout_encoded_displacements + with torch.autocast(device_type=x.device.type,enabled=False): + decoded=model.endecoder.decode(x.float()) + positions=rollout_encoded_displacements(decoded['local_transl_vel'],decoded['global_orient'],origin.float()) + return decoded,positions + + +def supervision_params(batch,history): + """Keep GT future motion, reanchoring its translation to the observed end. + + GEM observes velocities and rotations, not absolute root position. Asking + it to undo an invisible world-position offset would contradict its inputs. + Joint poses, root rotations, future displacements and timing stay unchanged. + """ + target=batch['smpl_params_w'] + observed=batch.get('_hfd_history_params') + if observed is None:return target + result={key:torch.cat([observed[key],value[:,history:]],dim=1) for key,value in target.items()} + offset=observed['transl'][:,[-1]]-target['transl'][:,[history-1]] + result['transl'][:,history:]+=offset + return result + + +def consistent_translation_loss(inputs,outputs,pipeline): + """World-root supervision uses the same predicted rotations as inference.""" + origin=inputs.get('_hfd_supervision_params',inputs['smpl_params_w'])['transl'][:,[0]] + _,positions=decoded_positions(pipeline,outputs['model_output']['pred_x'],origin) + mask=inputs['mask']['valid'].clone() + mask[inputs['mask']['spv_incam_only']]=False + mask[inputs['mask']['2d_only']]=False + target=inputs.get('_hfd_supervision_params',inputs['smpl_params_w'])['transl'] + return ((positions-target).abs()*mask[...,None]).mean() + + +@torch.no_grad() +def generate_parameters(model,batch): + """Generate from past parameters only. Future placeholders contain no GT.""" + from .genmo import HISTORY,FUTURE + with evaluation_mode(model): + model.prepare_batch(batch,'regression') + model.create_condition_mask(batch,{},'diffusion',train=False) + result=model.pipeline.denoiser3d.forward_test(batch) + x=result['pred_x'].float() + x=x*(1-batch['motion_mask_3d'])+batch['observed_motion_3d']*batch['motion_mask_3d'] + decoded,positions=decoded_positions(model,x,batch['smpl_params_w']['transl'][:,[0]]) + # Start at the last known position. Prefix integration must never shift + # the root because its last forward displacement is unknown. + shift=batch['smpl_params_w']['transl'][:,[HISTORY-1]]-positions[:,[HISTORY-1]] + return dict(body_pose=decoded['body_pose'][:,HISTORY:].detach(), + global_orient=decoded['global_orient'][:,HISTORY:].detach(), + transl=(positions+shift)[:,HISTORY:].detach(), + betas=batch['smpl_params_w']['betas'][:,[-1]].expand(-1,FUTURE,-1).detach()) + + +def recover_history(model,batch): + from .genmo import HISTORY,EXECUTE + config=model.hfd_strategy.get('recovery',{}) + previous=batch.pop('_hfd_recovery_previous',None) + indices=batch.pop('_hfd_recovery_indices',None) + if previous is None or indices is None:return 0 + step=int(model.global_step);warmup=int(config.get('warmup_steps',1000));ramp=int(config.get('ramp_steps',4000)) + probability=min(1.,max(0.,(step-warmup)/max(1,ramp))) + if not model.training or probability<=0:return 0 + selected=torch.rand(len(indices),device=indices.device)=self.start_step: + super().on_validation_end(trainer,module) + + +class OMGTrainingObserver(pl.Callback): + def __init__(self,output,check_pipeline=False,audio_policy='hfd_past_2s',use_asr_text=False,input_text_source='asr'): + self.output=Path(output);self.check_pipeline=check_pipeline;self.optimizer_updates=0;self.validation_batches=0 + self.audio_policy=audio_policy + self.use_asr_text=use_asr_text + self.input_text_source=input_text_source + def on_fit_start(self,trainer,module): + self.started=time.monotonic();self.previous_batch_end=None + if self.check_pipeline: + torch.cuda.reset_peak_memory_stats(module.device) + self.initial={n:p.detach().cpu().clone() for n,p in module.named_parameters() if p.requires_grad} + trainer.optimizers[0].register_step_post_hook(lambda *unused:self.count_update()) + def count_update(self): + self.optimizer_updates+=1 + def on_train_start(self,trainer,module): + self.initial_step=trainer.global_step + def on_before_optimizer_step(self,trainer,module,optimizer): + if self.check_pipeline: + grads=[p.grad for p in module.parameters() if p.grad is not None] + if not grads or not all(torch.isfinite(g).all() for g in grads): + raise RuntimeError('Missing/nonfinite OMG gradients') + def on_train_batch_end(self,trainer,module,outputs,batch,batch_idx): + loss=outputs['loss'] if isinstance(outputs,dict) else outputs + if not torch.isfinite(loss).all():raise RuntimeError('Nonfinite OMG training loss') + if trainer.is_global_zero: + with (self.output/'metrics.jsonl').open('a') as f: + f.write(json.dumps(dict(step=trainer.global_step,loss=float(loss.detach()), + elapsed_seconds=time.monotonic()-self.started, + data_wait_seconds=self.data_wait_seconds, + compute_seconds=time.monotonic()-self.batch_started))+'\n') + self.previous_batch_end=time.monotonic() + def on_train_batch_start(self,trainer,module,batch,batch_idx): + self.batch_started=time.monotonic() + self.data_wait_seconds=None if self.previous_batch_end is None else self.batch_started-self.previous_batch_end + def on_validation_batch_end(self,trainer,module,outputs,batch,batch_idx,dataloader_idx=0): + self.validation_batches+=1 + if outputs is not None: + loss=outputs['loss'] if isinstance(outputs,dict) else outputs + if not torch.isfinite(loss).all():raise RuntimeError('Nonfinite OMG validation loss') + def on_validation_end(self,trainer,module): + # Do not count validation/checkpoint time as next-batch data waiting. + self.previous_batch_end=None + if trainer.is_global_zero: + metrics={k:float(v.detach()) for k,v in trainer.callback_metrics.items() if k.startswith('val/') and torch.is_tensor(v) and v.numel()==1} + with (self.output/'validation.jsonl').open('a') as f: + f.write(json.dumps(dict(step=trainer.global_step,validation_batches=self.validation_batches,metrics=metrics))+'\n') + def on_train_end(self,trainer,module): + if trainer.is_global_zero: + best=[dict(path=c.best_model_path,score=float(c.best_model_score)) for c in trainer.checkpoint_callbacks + if c.monitor=='val/loss' and c.best_model_score is not None] + early=[str(getattr(c,'stopping_reason_message',None)) for c in trainer.callbacks if isinstance(c,EarlyStopping)] + (self.output/'training_summary.json').write_text(json.dumps(dict(global_step=trainer.global_step, + elapsed_seconds=time.monotonic()-self.started,best_checkpoints=best,early_stopping=early),indent=2)+'\n') + if not self.check_pipeline:return + changed=any(not torch.equal(self.initial[n],p.detach().cpu()) for n,p in module.named_parameters() if n in self.initial) + file=self.output/'pipeline.ckpt' + trainer.save_checkpoint(str(file)) + trainer.strategy.barrier() + ckpt=torch.load(file,map_location='cpu',weights_only=False) + module.load_state_dict(ckpt['state_dict'],strict=True) + expected_updates=trainer.global_step-getattr(self,'initial_step',0) + passed=changed and self.optimizer_updates==expected_updates + rank_report=dict(rank=trainer.global_rank,device=str(module.device),optimizer_updates=self.optimizer_updates, + peak_allocated_mib=torch.cuda.max_memory_allocated(module.device)/2**20, + peak_reserved_mib=torch.cuda.max_memory_reserved(module.device)/2**20) + (self.output/f'pipeline_rank{trainer.global_rank}.json').write_text(json.dumps(rank_report,indent=2)+'\n') + if trainer.is_global_zero: + report=dict(status='passed' if passed else 'failed',model='omg',framework='pytorch_lightning.Trainer/MotionGenerator.training_step', + finite_losses_and_gradients=True,parameters_changed=changed,optimizer_updates=self.optimizer_updates, + expected_updates=expected_updates,resumed_from_step=getattr(self,'initial_step',0),world_size=trainer.world_size,validation_batches=self.validation_batches,native_checkpoint=str(file),strict_checkpoint_reload=True, + input=(self.audio_policy+' audio + released oracle foreground transcript; no control labels') if self.input_text_source=='oracle_foreground_transcript' else self.audio_policy+(' audio + causal ASR text' if self.use_asr_text else ' audio')+'; no GT text or control labels',audio_policy=self.audio_policy,use_asr_text=self.use_asr_text,purpose='training pipeline health, not duplex ability') + (self.output/'pipeline_check.json').write_text(json.dumps(report,indent=2)+'\n') + if not passed:raise RuntimeError('OMG optimizer did not complete expected updates') diff --git a/handoff-v1/resources/final-resume-historical-source/hfd_bridge/resume.py b/handoff-v1/resources/final-resume-historical-source/hfd_bridge/resume.py new file mode 100644 index 0000000000000000000000000000000000000000..eeb0943a4f30710e4cc2db9b39c3e9d1910e958e --- /dev/null +++ b/handoff-v1/resources/final-resume-historical-source/hfd_bridge/resume.py @@ -0,0 +1,121 @@ +"""Checkpoint selection, run compatibility, RNG state and atomic persistence.""" +import copy +from datetime import datetime, timezone +import json +from pathlib import Path +import random + +import numpy as np +from omegaconf import OmegaConf + + +def rng_state(): + import torch + n=np.random.get_state() + numpy_state=(n[0],n[1].tolist(),int(n[2]),int(n[3]),float(n[4])) + return dict(python=random.getstate(), numpy=numpy_state, torch=torch.get_rng_state(), + cuda=torch.cuda.get_rng_state_all() if torch.cuda.is_initialized() else []) + + +def restore_rng(state): + import torch + random.setstate(state['python']) + n=state['numpy'] + np.random.set_state((n[0],np.asarray(n[1],dtype=np.uint32),n[2],n[3],n[4])) + torch.set_rng_state(state['torch'].cpu()) + if state['cuda']: + torch.cuda.set_rng_state_all([v.cpu() for v in state['cuda']]) + + +def atomic_save(payload, destination): + import torch + destination=Path(destination) + temporary=destination.with_name(destination.name+'.partial') + torch.save(payload, temporary) + temporary.replace(destination) + + +def contract(config): + """Keep optimizer/data semantics fixed; allow output paths and logging cadence.""" + c=OmegaConf.to_container(config, resolve=True) + training=copy.deepcopy(c['training']) + for key in ['resume','save_every','eval_every','keep_last_checkpoints','check_pipeline', + 'validation_batches','validation_batch_size','save_best','early_stopping','num_workers']: + training.pop(key,None) + data=copy.deepcopy(c['data']) + # These do not alter the selected data, text or training batch stream. + for key in ['validation_batch_size','prefetch_batches','cache_sessions']: + data.pop(key,None) + if isinstance(data.get('native'),dict):data['native'].pop('memory_mib',None) + return dict(model=c['model'],framework=c['framework'],training=training,data=data, + weights=c['weights'],native_overrides=c.get('native_overrides',{})) + + +def source_run(checkpoint): + checkpoint=Path(checkpoint) + for folder in [checkpoint.parent, checkpoint.parent.parent]: + if (folder/'resolved_config.yaml').is_file(): + return folder + raise ValueError('Resume requires an HFD training run with resolved_config.yaml') + + +def select_checkpoint(config, request, resolve): + if request in (None, False, ''): + return None + if int(config.runtime.num_gpus)!=1: + raise ValueError('Full HFD resume currently requires single-GPU training per model') + output=resolve(config.paths.output_dir) + if request=='auto': + if config.model=='omg': + candidates=[output/'last.ckpt', output/'checkpoints/last.ckpt'] + candidates += sorted((output/'checkpoints').glob('step-*.ckpt'), reverse=True) + else: + candidates=sorted(output.glob('ckpt_*.pth'),key=lambda p:int(p.stem.split('_')[-1]),reverse=True) + candidates=list({p.resolve():p for p in candidates if p.is_file()}.values()) + if not candidates: + raise FileNotFoundError(f'No resumable checkpoint in {output}; auto never starts a new run') + import torch + def progress(candidate): + state=torch.load(candidate,map_location='cpu',weights_only=False) + return int(state.get('global_step',state.get('step',-1))),candidate.stat().st_mtime_ns + checkpoint=max(candidates,key=progress) + else: + checkpoint=resolve(request) + if not checkpoint.is_file(): + raise FileNotFoundError(checkpoint) + source=source_run(checkpoint) + previous=OmegaConf.load(source/'resolved_config.yaml') + if contract(previous)!=contract(config): + raise ValueError('Resume configuration differs in training/data/weights. Use the original run config; max_steps is the original total target, not additional steps.') + import torch + state=torch.load(checkpoint,map_location='cpu',weights_only=False) + step=int(state.get('global_step',state.get('step',-1))) + optimizer=state.get('optimizer_states') if config.model=='omg' else state.get('optimizer') + if not optimizer or not 01 and c.data.cache_sessions= training.batch_size to retain the reused session pool') + assets=[path(recipe.raw_root)/'index.sqlite',path(recipe.raw_root)/'MANIFEST.json'] + assets.append(path(native.reference_cache) if reference else path(native.asr_model)) + for item in assets: + if not item.exists():raise FileNotFoundError(item) + if reference: + import math + guard=float(native.get('reference_guard_seconds',.1)) + if not math.isfinite(guard) or guard<0:raise ValueError('Invalid reference_guard_seconds') + from hfd_bridge.reference_runtime import ReferenceIndex + language=native.get('language','all') + if language not in ('all','en','zh'):raise ValueError('Unknown native foreground language') + if language=='en' and c.data.get('text_conditioning','none')!='none':raise ValueError('English-only reference training must disable translation') + index=ReferenceIndex(path(native.reference_cache),required_languages=['en','zh'] if language=='all' else [language]) + if recipe.get('include_third_party',False) and not index.db.execute('SELECT COUNT(*) FROM reference_recipes').fetchone()[0]: + raise ValueError('Third-party reference recipe index is empty') + if c.training.get('validation_batches',0)<=0:raise ValueError('On-demand requires bounded validation_batches') + train,val=[],[] + else: + train = session_paths(c.data.train_sessions) + val = session_paths(c.data.val_sessions) if c.data.val_sessions else [] + if c.data.get('frame_range') is not None: + span=list(c.data.frame_range) + if c.mode != 'smoke' or len(train) != 1 or len(span) != 2 or any(type(v) is not int for v in span) or not 0 <= span[0] < span[1]: + raise ValueError('data.frame_range requires smoke mode, one session and valid [start,end) frames') + if c.mode == 'train' and not val and not ondemand: + raise ValueError('Formal training requires separate validation sessions') + if c.mode == 'train' and set(train) & set(val): + raise ValueError('Train/validation session paths overlap') + for split, sessions in [('train',train),('val',val)]: + for p in sessions: + meta = json.loads(p.with_suffix('.json').read_text()) + if meta.get('audio_sr') != 16000 or meta.get('motion_fps') != 50: + raise ValueError(f'Unexpected HFD clocks: {p}') + if c.mode == 'train' and meta.get('split') != split: + raise ValueError(f'Expected {split} metadata for {p}') + if c.model == 'textop' or c.data.get('use_asr_text',False): + suffix='.reference.json' if c.data.get('caption_policy')=='aligned_reference_context' else '.asr.json' + sidecar = path(c.data.asr_dir)/(p.stem+suffix) + if not sidecar.is_file(): + raise FileNotFoundError(sidecar) + required = [path(c.paths.upstream_root),path(c.weights.init_checkpoint)] + if c.data.get('text_conditioning','none')=='asr_to_english': + translator=Path(os.environ.get('HFD_TRANSLATOR_ROOT',str(DEFAULT_TRANSLATOR))) + required += [translator/name for name in ['config.json','pytorch_model.bin','source.spm','target.spm','download_identity.json']] + if c.data.get('text_conditioning')=='nllb_to_english': + from hfd_bridge.semantics import DEFAULT_NLLB + translator=Path(os.environ.get('HFD_NLLB_ROOT',str(DEFAULT_NLLB))) + required += [translator/name for name in ['config.json','pytorch_model.bin','sentencepiece.bpe.model','download_identity.json']] + if c.model == 'textop': + required += [path(c.weights.native_config),path(c.weights.vae),path(c.weights.clip),path(c.weights.stats),path(c.weights.robot_assets)] + else: + required += [path(c.weights.models_root)/'t5-base-local'] + for p in required: + if not p.exists(): + raise FileNotFoundError(p) + return train,val + + +def setup(c): + for key,value in c.runtime.get('environment',{}).items(): + os.environ[str(key)] = str(value) + os.environ['CUDA_VISIBLE_DEVICES'] = ','.join(map(str,c.runtime.gpu_ids)) + os.environ['HF_HOME'] = str(path(c.paths.cache_root)/'huggingface') + os.environ['TORCH_HOME'] = str(path(c.paths.cache_root)/'torch') + os.environ['HF_HUB_OFFLINE'] = '1' if c.runtime.offline else '0' + os.environ['TRANSFORMERS_OFFLINE'] = '1' if c.runtime.offline else '0' + os.environ['TOKENIZERS_PARALLELISM'] = 'false' + os.environ['HFD_DATA_DISK'] = str(path(c.paths.data_disk)) + upstream = path(c.paths.upstream_root) + if c.model == 'textop': + sys.path.insert(0,str(upstream/'TextOpRobotMDAR')) + sys.path.insert(0,str(upstream/'deps/isaac_utils')) + sys.path.insert(0,str(upstream/'deps/CLIP')) + else: + sys.path.insert(0,str(upstream/'src')) + os.environ['OMG_MODELS_ROOT'] = str(path(c.weights.models_root)) + + +def textop(c, train, val, output): + from robotmdar.train.train_dar import main as native_main + from hfd_bridge.textop_checks import PipelineDARManager + cfg = OmegaConf.load(path(c.weights.native_config)) + cfg = OmegaConf.merge(cfg,c.native_overrides) + cfg.device='cuda:0'; cfg.seed=c.training.seed + cfg.experiment_dir=str(output); cfg.expname='hfd_baseline_'+c.mode + cfg.ckpt.vae=str(path(c.weights.vae)); cfg.ckpt.dar=c.training.get('resume') + cfg.data.num_primitive=c.training.num_primitive; cfg.data.batch_size=c.training.batch_size + robot=OmegaConf.to_container(cfg.data.skeleton,resolve=True) + robot['asset']['assetRoot']=str(path(c.weights.robot_assets)) + def data(sessions): + return OmegaConf.create(dict(_target_='hfd_bridge.configured_data.TextOpSessionData', + sessions=[str(p) for p in sessions],asr_dir=str(path(c.data.asr_dir)),robot_cfg=robot, + stats_path=str(path(c.weights.stats)),clip_path=str(path(c.weights.clip)), + batch_size=c.training.batch_size,num_primitive=c.training.num_primitive, + max_windows=c.data.max_windows,window_policy=c.data.window_policy, + device=cfg.device,cache_sessions=c.data.cache_sessions,asr_latency=c.data.asr_latency_seconds, + frame_range=OmegaConf.to_container(c.data.frame_range) if c.data.get('frame_range') is not None else None,caption_policy=c.data.get('caption_policy',DEFAULT_CAPTION_POLICY),clock_policy=c.data.get('clock_policy','legacy_hfd'),text_conditioning=c.data.get('text_conditioning','none'))) + cfg.data.train=data(train); cfg.data.val=data(val or train) + if c.data.get('backend')=='native_ondemand': + def native_data(split): + return OmegaConf.create(dict(_target_='hfd_bridge.ondemand_data.OnDemandTextOpData', + native_options=native_options(c),split=split,robot_cfg=robot,stats_path=str(path(c.weights.stats)), + clip_path=str(path(c.weights.clip)),batch_size=int(c.data.get('validation_batch_size',c.training.batch_size)) if split=='val' else c.training.batch_size,num_primitive=c.training.num_primitive, + device=cfg.device,cache_sessions=c.data.cache_sessions,output=str(output))) + cfg.data.train=native_data('train');cfg.data.val=native_data('val') + architecture=OmegaConf.to_container(cfg.denoiser,resolve=True) + cfg.denoiser=OmegaConf.create(dict(_target_='hfd_bridge.native.pretrained_denoiser',_recursive_=False, + architecture=architecture,checkpoint=str(path(c.weights.init_checkpoint)))) + m=cfg.train.manager + m._target_='hfd_bridge.textop_checks.PipelineDARManager'; m.stages=c.training.stages + m.device=cfg.device; m.save_dir=str(output); m.learning_rate=c.training.learning_rate + m.save_every=c.training.save_every; m.eval_every=c.training.eval_every; m.eval_steps=c.training.eval_steps + m.ckpt=cfg.ckpt; m.max_grad_norm=c.training.gradient_clip_val; m.check_pipeline=c.training.get('check_pipeline',c.mode=='smoke') + m.keep_last_checkpoints=c.training.get('keep_last_checkpoints',0) + m.input_text_source='oracle_foreground_transcript' if c.data.get('caption_policy')=='aligned_reference_context' else 'asr' + m.resume_num_primitive=int(c.training.num_primitive) + m.platform=OmegaConf.create(dict(_target_='hfd_bridge.native.ScalarLog',save_dir=str(output))) + OmegaConf.save(cfg,output/'native_config.yaml') + native_main(cfg) + + +def omg(c, train, val, output): + import torch + import pytorch_lightning as pl + from hydra import initialize_config_dir,compose + from torch.utils.data import DataLoader + from pytorch_lightning.callbacks import ModelCheckpoint + from pytorch_lightning.strategies import DDPStrategy + from datetime import timedelta + from hfd_bridge.configured_data import OMGAudioSessionDataset + from hfd_bridge.omg_checks import OMGTrainingObserver,FixedValidationNoise,WarmupEarlyStopping + from hfd_bridge.omg_model_loader import _load_model + pl.seed_everything(c.training.seed,workers=True) + with initialize_config_dir(config_dir=str(path(c.paths.upstream_root)/'configs/generation'),version_base='1.3'): + cfg=compose(config_name='train',overrides=['exp='+c.training.experiment,'logger=none','trainer=1gpu']) + cfg=OmegaConf.merge(cfg,c.native_overrides) + cfg.model.optimizer.lr=c.training.learning_rate + cfg.model.scheduler=OmegaConf.merge(cfg.model.scheduler,c.training.scheduler) + cfg.output_dir=str(output) + model=_load_model(cfg,str(path(c.weights.init_checkpoint)),torch.device('cpu')) + def data(sessions): + return OMGAudioSessionDataset(sessions,model.representation,c.data.window_policy,c.data.max_windows,c.data.cache_sessions,c.data.get('frame_range'),c.data.audio_policy, + use_asr_text=c.data.get('use_asr_text',False),asr_dir=str(path(c.data.asr_dir)),asr_latency=c.data.asr_latency_seconds,caption_policy=c.data.get('caption_policy',DEFAULT_CAPTION_POLICY),clock_policy=c.data.get('clock_policy','legacy_hfd'),text_conditioning=c.data.get('text_conditioning','none')) + workers=c.training.num_workers + loader_options=dict(batch_size=c.training.batch_size,num_workers=workers,pin_memory=True) + if workers: loader_options['persistent_workers']=True + ondemand=c.data.get('backend')=='native_ondemand' + if ondemand: + from hfd_bridge.ondemand_data import OnDemandOMGBatches + def batches(split): + return OnDemandOMGBatches(native_options(c),split,model.representation,int(c.data.get('validation_batch_size',c.training.batch_size)) if split=='val' else c.training.batch_size,c.data.cache_sessions,str(output), + fixed_validation=c.training.get('fixed_validation',False),batched_motion=c.data.get('batched_motion',True), + prefetch_batches=c.data.get('prefetch_batches',0), + max_batches=int(c.training.max_steps*c.training.accumulate_grad_batches) if split=='train' else int(c.training.validation_batches)) + train_loader=DataLoader(batches('train'),batch_size=None,num_workers=0) + val_loader=DataLoader(batches('val'),batch_size=None,num_workers=0) + else: + train_loader=DataLoader(data(train),shuffle=True,**loader_options) + val_loader=DataLoader(data(val),shuffle=False,**loader_options) if val else None + keep=int(c.training.get('keep_last_checkpoints',0)) + checkpoint=ModelCheckpoint(dirpath=str(output/'checkpoints'),every_n_train_steps=c.training.save_every, + save_top_k=keep if keep else -1,monitor='step' if keep else None,mode='max', + save_last='link' if keep else True,filename='step-{step:08d}',auto_insert_metric_name=False) + from hfd_bridge.resume_lightning import HFDResumeState + callbacks=[HFDResumeState(train_loader.dataset if ondemand else None,val_loader.dataset if ondemand else None), + checkpoint,OMGTrainingObserver(str(output),c.training.get('check_pipeline',c.mode=='smoke'),audio_policy=c.data.audio_policy,use_asr_text=c.data.get('use_asr_text',False),input_text_source='oracle_foreground_transcript' if c.data.get('caption_policy')=='aligned_reference_context' else 'asr')] + if ondemand or val: + if c.training.get('save_best',False): + callbacks.append(ModelCheckpoint(dirpath=str(output/'best'),monitor='val/loss',mode='min',save_top_k=1, + filename='best',auto_insert_metric_name=False,enable_version_counter=False, + every_n_train_steps=0,every_n_epochs=1,save_on_train_epoch_end=False)) + if c.training.get('fixed_validation',False):callbacks.append(FixedValidationNoise(c.training.seed+1000003)) + early=c.training.get('early_stopping',{}) + if early.get('enabled',False): + callbacks.append(WarmupEarlyStopping(monitor='val/loss',mode='min',start_step=early.get('start_step',30000), + patience=early.get('patience',10),min_delta=early.get('min_delta',0.0002),verbose=True,log_rank_zero_only=True)) + trainer_options=OmegaConf.to_container(cfg.trainer,resolve=True) + trainer_options.update(accelerator='gpu',devices=c.runtime.num_gpus, + strategy=DDPStrategy(find_unused_parameters=True,process_group_backend=c.runtime.get('distributed_backend','nccl'),timeout=timedelta(seconds=int(c.runtime.get('distributed_timeout_seconds',120)))) if c.runtime.num_gpus>1 else 'auto', + max_steps=c.training.max_steps,precision=c.training.precision, + accumulate_grad_batches=c.training.accumulate_grad_batches, + gradient_clip_val=c.training.gradient_clip_val,logger=False,enable_checkpointing=True, + enable_progress_bar=True,num_sanity_val_steps=0,default_root_dir=str(output), + callbacks=callbacks, + use_distributed_sampler=c.runtime.num_gpus>1 and not ondemand, + val_check_interval=c.training.eval_every,check_val_every_n_epoch=None) + if ondemand: + trainer_options['limit_val_batches']=int(c.training.validation_batches) + elif not val: + trainer_options['limit_val_batches']=0 + if os.environ.get('LOCAL_RANK','0')=='0': + OmegaConf.save(cfg,output/'native_config.yaml') + trainer=pl.Trainer(**trainer_options) + model.train() + if model.text_encoder is not None:model.text_encoder.encoder.eval() + trainer.fit(model,train_dataloaders=train_loader,val_dataloaders=val_loader,ckpt_path=c.training.get('resume')) + trainer.save_checkpoint(str(output/'last.ckpt')) + + +def native_options(c): + options=OmegaConf.to_container(c.data.native,resolve=True) + for key in ('config','asr_cache','asr_model','reference_cache'): + if key in options:options[key]=str(path(options[key])) + local_rank=int(os.environ.get('LOCAL_RANK','0')) + options.update(seed=c.training.seed,asr_device=f'cuda:{local_rank}',caption_policy=c.data.caption_policy,asr_latency=c.data.asr_latency_seconds,clock_policy=c.data.get('clock_policy','legacy_hfd'),text_conditioning=c.data.get('text_conditioning','none')) + recipe=OmegaConf.load(path(options['config']));OmegaConf.resolve(recipe) + options['include_third_party']=bool(recipe.get('include_third_party',False)) + return options + + +def main(): + p=argparse.ArgumentParser(description=__doc__) + p.add_argument('--config',required=True,type=Path) + p.add_argument('--set',nargs='*',default=[],help='OmegaConf dotted overrides, e.g. training.batch_size=16') + p.add_argument('--dry-run',action='store_true',help='Resolve and validate configuration without training') + p.add_argument('--resume',nargs='?',const='auto',help='Continue full training state from a checkpoint; omit path to select latest in output_dir') + a=p.parse_args() + c=OmegaConf.load(a.config) + c=OmegaConf.merge(c,OmegaConf.from_dotlist(a.set)); OmegaConf.resolve(c) + if c.runtime.python: + interpreter=Path(str(c.runtime.python)).expanduser() + interpreter=(interpreter if interpreter.is_absolute() else ROOT/interpreter).absolute() + if Path(sys.executable).absolute() != interpreter: + os.execv(str(interpreter),[str(interpreter),str(Path(__file__).resolve()),*sys.argv[1:]]) + train,val=prepare(c) + from hfd_bridge.resume import select_checkpoint,record_resume + resume=select_checkpoint(c,a.resume if a.resume is not None else c.training.get('resume'),path) + c.training.resume=str(resume) if resume else None + if a.dry_run: + print(OmegaConf.to_yaml(c)) + print(json.dumps(dict(status='validated_configuration',backend=c.data.get('backend','exported_npz'), + train_sessions=len(train) if train else None,val_sessions=len(val) if val else None, + note='Native index and on-demand reads are checked at runtime' if c.data.get('backend')=='native_ondemand' else 'Exported session paths checked'))) + return + setup(c) + if c.model=='omg' and c.runtime.num_gpus>1 and 'LOCAL_RANK' not in os.environ: + os.execv(sys.executable,[sys.executable,'-m','torch.distributed.run','--standalone', + '--nproc_per_node='+str(c.runtime.num_gpus),str(Path(__file__).resolve()),*sys.argv[1:]]) + import torch + torch.set_num_threads(c.runtime.cpu_threads);torch.manual_seed(c.training.seed) + if torch.cuda.device_count()!=c.runtime.num_gpus: + raise RuntimeError('Requested GPU visibility does not match available devices') + output=path(c.paths.output_dir) + rank=int(os.environ.get('LOCAL_RANK','0')) + torch.cuda.set_device(rank) + if rank==0: + existing=output.exists() + output.mkdir(parents=True,exist_ok=resume is not None) + if not existing:OmegaConf.save(c,output/'resolved_config.yaml') + audit_output=record_resume(output,c,resume) if resume else output + from hfd_bridge.semantics import conditioning_identity + manifest=dict(text_frontend=conditioning_identity(c.data.get('text_conditioning','none')), + clock_policy=c.data.get('clock_policy','legacy_hfd'),backend=c.data.get('backend','exported_npz'),train=list(map(str,train)),val=list(map(str,val))) + if c.data.get('caption_policy')=='aligned_reference_context': + manifest.update(text_source='oracle_foreground_transcript',caption_policy='aligned_reference_context', + reference_sidecars=[str(path(c.data.asr_dir)/(p.stem+'.reference.json')) for p in train+val]) + if c.data.get('backend')=='native_ondemand': + manifest.update(native_options=native_options(c),split_manifests=['train_native_manifest.json','val_native_manifest.json']) + (audit_output/'data_manifest.json').write_text(json.dumps(manifest,indent=2)+'\n') + provenance={} + for sub in [ROOT/'hfd_bridge',path(c.paths.upstream_root)]: + for f in sub.rglob('*.py'): + provenance[str(f.relative_to(ROOT)) if f.is_relative_to(ROOT) else str(f)]=hashlib.sha256(f.read_bytes()).hexdigest() + provenance['train.py']=hashlib.sha256(Path(__file__).read_bytes()).hexdigest() + (audit_output/'code_sha256.json').write_text(json.dumps(provenance,indent=2)+'\n') + (textop if c.model=='textop' else omg)(c,train,val,output) + if rank==0:print('COMPLETED',output,flush=True) + +if __name__=='__main__': + main() diff --git a/handoff-v1/resources/historical-source/hfd_bridge/genmo_prefetch.py b/handoff-v1/resources/historical-source/hfd_bridge/genmo_prefetch.py new file mode 100644 index 0000000000000000000000000000000000000000..e9445d10ed1f77c9ada3ef9cce0a20a5cc3d85e6 --- /dev/null +++ b/handoff-v1/resources/historical-source/hfd_bridge/genmo_prefetch.py @@ -0,0 +1,29 @@ +"""One CPU producer with bounded lookahead and consumed-batch resume state.""" +import copy +import os +import torch +from torch.utils.data import IterableDataset + + +class CPUBatches(IterableDataset): + def __init__(self,source): + # Serialize only configuration and sampler state, never SQLite/CUDA. + self.options=copy.deepcopy(source.provider.options) + self.state=source.pool.state_dict();self.batch_size=source.pool.batch_size + self.max_batches=source.max_batches;self.refresh=source.pool.refresh + self.output=str(source.output);self.recovery_fraction=source.recovery_fraction + + def __iter__(self): + os.environ['CUDA_VISIBLE_DEVICES']='' + torch.set_num_threads(2) + from .genmo_data import PairedNativeProvider,GenmoBatches + provider=PairedNativeProvider(**self.options) + data=GenmoBatches(provider,None,self.batch_size,self.max_batches,self.refresh,self.output, + recovery_fraction=self.recovery_fraction) + data.pool.load_state_dict(self.state) + for batch in data.cpu_batches(): + # T5 runs in the GPU process; don't transport zero placeholders. + batch.pop('text_embed',None) + previous=batch.get('_hfd_recovery_previous') + if previous:previous.pop('text_embed',None) + yield batch diff --git a/handoff-v1/resources/historical-source/hfd_bridge/genmo_training.py b/handoff-v1/resources/historical-source/hfd_bridge/genmo_training.py new file mode 100644 index 0000000000000000000000000000000000000000..3f094ed463c6a769c18570856dac7f2132073b27 --- /dev/null +++ b/handoff-v1/resources/historical-source/hfd_bridge/genmo_training.py @@ -0,0 +1,132 @@ +"""Execution-weighted GEM training and causal, detached predicted-history recovery.""" +from contextlib import contextmanager +import torch +import torch.nn.functional as F + + +@contextmanager +def evaluation_mode(model): + states={module:module.training for module in model.modules()} + model.eval() + try:yield + finally: + for module,state in states.items():module.training=state + + +def execution_mask(mask,history,execute,weight): + """Preserve each feature's total loss mass while emphasizing executed frames.""" + result=mask.float().clone() + result[:,history:history+execute]*=float(weight) + result[:,history-1,148:151]*=float(weight) + total=mask.sum(dim=1,keepdim=True) + return result*(total/result.sum(dim=1,keepdim=True).clamp_min(1)) + + +def decoded_positions(model,x,origin): + from .genmo import rollout_encoded_displacements + with torch.autocast(device_type=x.device.type,enabled=False): + decoded=model.endecoder.decode(x.float()) + positions=rollout_encoded_displacements(decoded['local_transl_vel'],decoded['global_orient'],origin.float()) + return decoded,positions + + +def supervision_params(batch,history): + """Keep GT future motion, reanchoring its translation to the observed end. + + GEM observes velocities and rotations, not absolute root position. Asking + it to undo an invisible world-position offset would contradict its inputs. + Joint poses, root rotations, future displacements and timing stay unchanged. + """ + target=batch['smpl_params_w'] + observed=batch.get('_hfd_history_params') + if observed is None:return target + result={key:torch.cat([observed[key],value[:,history:]],dim=1) for key,value in target.items()} + offset=observed['transl'][:,[-1]]-target['transl'][:,[history-1]] + result['transl'][:,history:]+=offset + return result + + +def consistent_translation_loss(inputs,outputs,pipeline): + """World-root supervision uses the same predicted rotations as inference.""" + origin=inputs.get('_hfd_supervision_params',inputs['smpl_params_w'])['transl'][:,[0]] + _,positions=decoded_positions(pipeline,outputs['model_output']['pred_x'],origin) + mask=inputs['mask']['valid'].clone() + mask[inputs['mask']['spv_incam_only']]=False + mask[inputs['mask']['2d_only']]=False + target=inputs.get('_hfd_supervision_params',inputs['smpl_params_w'])['transl'] + return ((positions-target).abs()*mask[...,None]).mean() + + +@torch.no_grad() +def generate_parameters(model,batch): + """Generate from past parameters only. Future placeholders contain no GT.""" + from .genmo import HISTORY,FUTURE + with evaluation_mode(model): + model.prepare_batch(batch,'regression') + model.create_condition_mask(batch,{},'diffusion',train=False) + result=model.pipeline.denoiser3d.forward_test(batch) + x=result['pred_x'].float() + x=x*(1-batch['motion_mask_3d'])+batch['observed_motion_3d']*batch['motion_mask_3d'] + decoded,positions=decoded_positions(model,x,batch['smpl_params_w']['transl'][:,[0]]) + # Start at the last known position. Prefix integration must never shift + # the root because its last forward displacement is unknown. + shift=batch['smpl_params_w']['transl'][:,[HISTORY-1]]-positions[:,[HISTORY-1]] + return dict(body_pose=decoded['body_pose'][:,HISTORY:].detach(), + global_orient=decoded['global_orient'][:,HISTORY:].detach(), + transl=(positions+shift)[:,HISTORY:].detach(), + betas=batch['smpl_params_w']['betas'][:,[-1]].expand(-1,FUTURE,-1).detach()) + + +def recover_history(model,batch): + from .genmo import HISTORY,EXECUTE + config=model.hfd_strategy.get('recovery',{}) + previous=batch.pop('_hfd_recovery_previous',None) + indices=batch.pop('_hfd_recovery_indices',None) + if previous is None or indices is None:return 0 + step=int(model.global_step);warmup=int(config.get('warmup_steps',1000));ramp=int(config.get('ramp_steps',4000)) + probability=min(1.,max(0.,(step-warmup)/max(1,ramp))) + if not model.training or probability<=0:return 0 + selected=torch.rand(len(indices),device=indices.device) times[-1]+1e-7: + raise ValueError('Source timestamps do not cover requested human motion') + query = np.clip(query, times[0], times[-1]) + out = np.empty((len(query), rows.shape[1]), np.float64) + for j in range(3): + out[:, j] = np.interp(query, times, rows[:, j]) + for j in range(3, rows.shape[1], 3): + out[:, j:j+3] = Slerp(times, Rotation.from_rotvec(rows[:, j:j+3]))(query).as_rotvec() + return out + + +def blend(anchor, target, weights): + anchor = np.asarray(anchor).reshape(-1) + target = np.asarray(target) + weights = np.broadcast_to(np.asarray(weights), (len(target),)) + out = target.copy() + out[:, :3] = anchor[:3] + weights[:, None]*(target[:, :3]-anchor[:3]) + for j in range(3, out.shape[1], 3): + a = Rotation.from_rotvec(np.repeat(anchor[None, j:j+3], len(out), axis=0)) + b = Rotation.from_rotvec(target[:, j:j+3]) + out[:, j:j+3] = (a * Rotation.from_rotvec((a.inv()*b).as_rotvec()*weights[:, None])).as_rotvec() + return out + + +def heading(row): + r = Rotation.from_rotvec(row[3:6]).as_matrix() + # Select a stable horizontal local axis; the same axis is used at both ends. + axis = r[:, 0] + return np.arctan2(axis[1], axis[0]) + + +def align(rows, anchor, pelvis): + rows = rows.copy() + yaw = Rotation.from_euler('z', heading(anchor)-heading(rows[0])) + p0 = rows[0, :3].copy()+pelvis + rows[:, :3] = yaw.apply(rows[:, :3]+pelvis-p0)+p0-pelvis + rows[:, :2] += anchor[:2]-rows[0, :2] + rows[:, 3:6] = (yaw*Rotation.from_rotvec(rows[:, 3:6])).as_rotvec() + return rows + + +class NativeHumanObserver: + def __init__(self, ds, module, root): + # Share the bridge's verified local metadata mirror. Raw SMPL-X assets + # remain indexed in their original shared-storage tar shards. + from .storage import ArchivePath + class IndexedReader: + def read(self, path):return ArchivePath(path).read_bytes() + self.store = IndexedReader() + pairs = pd.read_parquet(Path(root)/'metadata/seed_g1_smplx_pairs.parquet') + self.pairs = {r.motion_id:r for r in pairs.itertuples()} + self.ds, self.module = ds, module + self.rows = {} + self.raw = {} + self.ops = [] + self.parts = [] + self.scratch = {} + self.betas = None + self.pelvis = None + self.interpolate = interpolate + self.blend = blend + + def key(self, row): + return np.round(np.asarray(row, dtype=np.float64), 5).tobytes() + + def register(self, robot, human): + robot = np.asarray(robot).reshape(-1, 36) + human = np.asarray(human).reshape(-1, 159) + assert len(robot) == len(human) + for r, h in zip(robot, human): + self.rows[self.key(r)] = h.copy() + + def lookup(self, robot): + robot = np.asarray(robot) + try: + out = np.stack([self.rows[self.key(row)] for row in robot.reshape(-1, 36)]) + except KeyError: + raise RuntimeError('Untraced G1 pose reached human composition') from None + return out[0] if robot.ndim == 1 else out + + def source(self, path, robot): + name = str(path).rstrip('/').split('/')[-1] + if name not in self.raw: + pair = self.pairs[name] + blob = self.store.read(pair.smplx_path) + assert hashlib.sha256(blob).hexdigest() == pair.smplx_sha256 + with np.load(io.BytesIO(blob), allow_pickle=False) as z: + beta = z['betas'].reshape(-1)[:16].copy() + if self.betas is None: + self.betas = beta + # SMPL-X root offset is required for rigid yaw transforms. + asset = Path(os.environ.get('HFD_BODY_MODELS_ROOT','/root/autodl-fs/pangxinlei/models/body_models'))/'smplx/SMPLX_NEUTRAL.npz' + with np.load(asset, allow_pickle=True) as b: + verts = b['v_template'] + np.einsum('vck,k->vc', b['shapedirs'][:, :, :16], beta) + self.pelvis = b['J_regressor'][0] @ verts + assert np.allclose(beta, self.betas, atol=1e-6), 'Session shape changed' + data = np.concatenate([z['trans'], z['root_orient'], z['pose_body'], z['pose_hand']], -1) + times = z['sample_times'].astype(np.float64) + query = np.arange(len(robot))/50 + assert query[-1]-times[-1] <= 1/50+1e-6, (name, query[-1], times[-1]) + endpoint_hold = max(0., float(query[-1]-times[-1])) + human = self.interpolate(data, times, np.minimum(query,times[-1])) + self.raw[name] = (robot.copy(), human) + self.ops.append(dict(op='source', motion_id=name, smplx_path=pair.smplx_path, + sha256=pair.smplx_sha256, endpoint_hold_seconds=endpoint_hold, query_seconds=np.minimum(query,times[-1]).tolist())) + self.register(robot, self.raw[name][1]) + + def trace(self, frame, event, value): + if event != 'return' or frame.f_globals.get('__name__') != self.module.__name__: + return + name, d = frame.f_code.co_name, frame.f_locals + if name == '_load_motion': + self.source(d['path'], value) + elif name == '_nearest_safe_stationary_anchor': + selected = d['anchors'][d['selected']] + # Anchors are audited source frames, with only XY overwritten. + best = None + for motion_id, (robot, human) in self.raw.items(): + distance = np.max(np.abs(robot[:, :29]-selected[:29]), axis=1) + index = int(np.argmin(distance)) + if best is None or distance[index] < best[0]: + best = (distance[index], motion_id, index, human[index].copy()) + assert best and best[0] < 1e-5, 'Anchor source not traced' + h = best[3] + current = self.lookup(d['current_pose']) + h[:2] = current[:2] + self.register(value[0], h) + self.ops.append(dict(op='anchor', motion_id=best[1], frame=best[2])) + elif name == '_sample_stationary_anchor': + # Quaternion normalization may change float rounding at the boundary. + current = self.lookup(d['task_start_pose']) + # nearest anchor hook registered the pre-normalized returned row. + try: + h = self.lookup(value) + except RuntimeError: + anchors = d['self']._stationary_anchor_poses + k = np.argmin(np.max(np.abs(anchors[:, :29]-value[:29]), axis=1)) + raw = min(((np.max(np.abs(r[:, :29]-anchors[k, :29]),axis=1).min(), r, h) for r,h in self.raw.values()),key=lambda item:item[0]) + assert raw[0]<1e-5, 'Untraced normalized anchor' + idx=np.argmin(np.max(np.abs(raw[1][:,:29]-anchors[k,:29]),axis=1));h=raw[2][idx].copy();h[:2]=current[:2] + self.register(value, h) + elif name == '_crop_or_resample_motion': + base = self.lookup(d['motion']) + if len(base) >= d['frame_count']: + h = base[d.get('start', 0):d.get('start', 0)+d['frame_count']] + else: + h = self.interpolate(base, np.linspace(0,1,len(base)), np.linspace(0,1,d['frame_count'])) + self.register(value, h) + self.scratch[id(frame.f_back)] = h.copy() + self.ops.append(dict(op='stationary_crop', start=d.get('start'), frames=d['frame_count'], source_frames=len(base))) + elif name == '_listen_hold_target': + h = self.lookup(d['motion'])[int(d['selected'])] + self.register(value, h) + self.scratch[id(frame.f_back)] = np.repeat(h[None], frame.f_back.f_locals['frame_count'], 0) + self.ops.append(dict(op='listen_hold', source_frame=int(d['selected']))) + elif name == '_align_task_motion_to_anchor_heading': + # The helper is also called on freshly repeated listen rows. + try: + h = align(self.lookup(d['motion']), self.lookup(d['stationary_anchor']), self.pelvis) + self.register(value, h) + except (RuntimeError, KeyError): + pass # enclosing stationary/task hook records its complete result + elif name == '_sample_stationary_sequence' and d['frame_count'] > 0: + anchor = self.lookup(d['anchor_pose']) + h = align(self.scratch.pop(id(frame)), anchor, self.pelvis) + if d['phase'] == 'idle': + first = h[0].copy() + h[:, :3] = anchor[:3]+h[:, :3]-first[:3] + for j in range(6,159,3): + h[:,j:j+3]=(Rotation.from_rotvec(anchor[j:j+3])*Rotation.from_rotvec(first[j:j+3]).inv()*Rotation.from_rotvec(h[:,j:j+3])).as_rotvec() + h[:,3:6] = anchor[3:6] + start = self.lookup(d['entry_pose']) if d.get('entry_pose') is not None else anchor + h = self.blend(start, h, d['motion_weight']) + self.register(value[0], h) + self.ops.append(dict(op='stationary', phase=d['phase'], motion_id=d['selected_name'], frames=len(h), weights=d['motion_weight'].tolist())) + elif name == '_build_task_sequence': + anchor = self.lookup(d['stationary_anchor']) + h = align(self.lookup(d['motion']), anchor, self.pelvis) + assert d['self'].dynamic_task_entry_blend + n = int(d['blend_count']) + h[:n] = self.blend(anchor, h[:n], d['phase']) + self.register(value[0], h) + self.ops.append(dict(op='task_entry', frames=len(h), blend_frames=n)) + elif name == '_build_pose_transition': + h = self.blend(self.lookup(d['start_pose']), np.repeat(self.lookup(d['end_pose'])[None],d['count'],0), d['weight']) + self.register(value,h) + self.ops.append(dict(op='recovery', frames=len(h), weights=d['weight'].tolist())) + elif name == 'append_part' and d.get('count', 0) > 0: + h = self.lookup(d['array']) + self.parts.append(h) + self.ops.append(dict(op='append', frames=len(h), label=d['control_label'], turn_index=d['turn_index'], source_frames=d['original_frames'].tolist())) diff --git a/handoff-v1/resources/historical-source/hfd_bridge/kimodo.py b/handoff-v1/resources/historical-source/hfd_bridge/kimodo.py new file mode 100644 index 0000000000000000000000000000000000000000..67cb18d21e8b9c6dc067012f5d2c80a04bb8dfae --- /dev/null +++ b/handoff-v1/resources/historical-source/hfd_bridge/kimodo.py @@ -0,0 +1,208 @@ +"""Causal SMPL-X bridge to pinned Kimodo-SMPLX-RP-v1; no upstream edits.""" +import os +import sys +from pathlib import Path +from functools import lru_cache +import numpy as np +import torch +import torch.nn.functional as F + +ROOT=Path(os.environ.get('KIMODO_ROOT','/root/autodl-tmp/hfd-human-models/kimodo')) +sys.path.insert(0,str(ROOT)) +from kimodo.geometry import axis_angle_to_matrix,matrix_to_axis_angle +from kimodo.exports.smplx import kimodo_y_up_to_amass_coord_rotation_matrix +from kimodo.skeleton.kinematics import fk +from kimodo.motion_rep.feature_utils import compute_heading_angle + +HISTORY=60 +FUTURE=60 +EXECUTE=12 + + +def setup(gpu=0): + os.environ.update(CUDA_VISIBLE_DEVICES=str(gpu),HF_HUB_OFFLINE='1',TRANSFORMERS_OFFLINE='1', + TOKENIZERS_PARALLELISM='false',HFD_LOCAL_INDEX_CACHE='1', + HFD_BODY_MODELS_ROOT=os.environ.get('HFD_BODY_MODELS_ROOT','/root/autodl-fs/pangxinlei/models/body_models'), + TEXT_ENCODERS_DIR=os.environ.get('TEXT_ENCODERS_DIR',str(Path(os.environ.get('KIMODO_MODELS_ROOT','/root/autodl-fs/pangxinlei/models/kimodo'))/'runtime-text-encoders')), + TEXT_ENCODER_DEVICE='cuda:0',OMP_NUM_THREADS='4',MKL_NUM_THREADS='4') + torch.set_num_threads(4) + + +def representation(model_dir): + from hydra.utils import instantiate + from omegaconf import OmegaConf + c=OmegaConf.load(Path(model_dir)/'config.yaml') + c=OmegaConf.merge(c,dict(checkpoint_dir=str(model_dir))) + return instantiate(c.denoiser.motion_rep) + + +def load_network(model_dir,device='cuda',checkpoint=None): + from hydra.utils import instantiate + from omegaconf import OmegaConf + c=OmegaConf.merge(OmegaConf.load(Path(model_dir)/'config.yaml'),dict(checkpoint_dir=str(model_dir))) + net=instantiate(c.denoiser) + if checkpoint is not None: + state=torch.load(checkpoint,map_location='cpu',weights_only=False,mmap=True) + net.load_state_dict(state['denoiser'],strict=True) + if state.get('runtime'): + from .kimodo_finetuning import configure_runtime + configure_runtime(net,state['runtime']) + return net.to(device) + + +@lru_cache(maxsize=16) +def source_pelvis(beta): + path=Path(os.environ.get('HFD_BODY_MODELS_ROOT','/root/autodl-fs/pangxinlei/models/body_models'))/'smplx/SMPLX_NEUTRAL.npz' + with np.load(path,allow_pickle=True) as z: + vertices=z['v_template']+np.einsum('vck,k->vc',z['shapedirs'][:,:,:16],np.asarray(beta)) + return (z['J_regressor'][0]@vertices).astype(np.float32) + + +def to_native(rows,betas,skeleton): + """Preserve physical pelvis path while using the checkpoint's fixed skeleton.""" + rows=torch.as_tensor(rows,dtype=torch.float32) + Y=torch.tensor(kimodo_y_up_to_amass_coord_rotation_matrix()) + offset=torch.tensor(source_pelvis(tuple(np.asarray(betas,dtype=np.float32).tolist()))) + root=(rows[:,:3]+offset)@Y + local=axis_angle_to_matrix(rows[:,3:69].reshape(-1,22,3)) + local[:,0]=Y.T@local[:,0] + return local,root + + +def from_native(local,root,betas): + """Inverse conversion, restoring the known session's shape/translation gauge.""" + local=local.detach().float().cpu().clone();root=root.detach().float().cpu() + Y=torch.tensor(kimodo_y_up_to_amass_coord_rotation_matrix()) + local[:,0]=Y@local[:,0] + poses=matrix_to_axis_angle(local).reshape(-1,66) + offset=torch.tensor(source_pelvis(tuple(np.asarray(betas,dtype=np.float32).tolist()))) + return torch.cat([root@Y.T-offset,poses,torch.zeros(len(root),90)],-1).numpy() + + +def canonical_motion(rows,betas,rep): + """Only first known pose chooses heading/origin; no future-fitted frame.""" + local,root=to_native(rows,betas,rep.skeleton) + _,positions,_=fk(local[:1],root[:1],rep.skeleton) + angle=compute_heading_angle(positions,rep.skeleton).reshape(-1)[0] + # Kimodo heading increases clockwise when viewed from above. + rotation=axis_angle_to_matrix(torch.tensor([0.,float(-angle),0.])) + origin=root[0].clone();origin[1]=0 + local[:,0]=rotation@local[:,0] + root=(root-origin)@rotation.T + return local,root,dict(rotation=rotation,origin=origin) + + +def make_window(history,future,betas,rep,valid=None): + rows=np.concatenate([history,future]) + local,root,anchor=canonical_motion(rows,betas,rep) + if getattr(rep,'_hfd_causal_root',False): + from .kimodo_causal_rep import encode + x=encode(rep,local,root) + prefix=encode(rep,local[:HISTORY],root[:HISTORY]) + else: + x=rep(local,root,to_normalize=True) + prefix=rep(local[:HISTORY],root[:HISTORY],to_normalize=True) + # Independent prefix encoding is essential: smooth_root uses the entire clip. + observed=torch.zeros_like(x);mask=torch.zeros_like(x) + for name in ['smooth_root_pos','global_root_heading','local_joints_positions','global_rot_data']: + sl=rep.slice_dict[name];observed[:HISTORY,sl]=prefix[:,sl];mask[:HISTORY,sl]=1 + score=torch.zeros(len(x),dtype=torch.bool) + score[HISTORY:]=True if valid is None else torch.as_tensor(valid,dtype=torch.bool) + next_valid=torch.cat([score[1:],torch.zeros(1,dtype=torch.bool)]) + return dict(target=x,observed=observed,motion_mask=mask,score=score, + velocity_score=score&next_valid,heading=rep.get_first_heading_angle(prefix,normalized=True),anchor=anchor) + + +def window_at(rows,betas,decision,rep): + if decision%EXECUTE:raise ValueError('Decision must lie on .4s grid') + history=rows[np.maximum(0,decision+np.arange(-HISTORY+1,1))] + indices=decision+np.arange(1,getattr(rep,'_hfd_future_frames',FUTURE)+1) + future=rows[np.minimum(indices,len(rows)-1)] + return make_window(history,future,betas,rep,indices=self.start_step: + super().on_validation_end(trainer,module) + + +class OMGTrainingObserver(pl.Callback): + def __init__(self,output,check_pipeline=False,audio_policy='hfd_past_2s',use_asr_text=False,input_text_source='asr'): + self.output=Path(output);self.check_pipeline=check_pipeline;self.optimizer_updates=0;self.validation_batches=0 + self.audio_policy=audio_policy + self.use_asr_text=use_asr_text + self.input_text_source=input_text_source + def on_fit_start(self,trainer,module): + self.started=time.monotonic();self.previous_batch_end=None + if self.check_pipeline: + torch.cuda.reset_peak_memory_stats(module.device) + self.initial={n:p.detach().cpu().clone() for n,p in module.named_parameters() if p.requires_grad} + trainer.optimizers[0].register_step_post_hook(lambda *unused:self.count_update()) + def count_update(self): + self.optimizer_updates+=1 + def on_train_start(self,trainer,module): + self.initial_step=trainer.global_step + def on_before_optimizer_step(self,trainer,module,optimizer): + if self.check_pipeline: + grads=[p.grad for p in module.parameters() if p.grad is not None] + if not grads or not all(torch.isfinite(g).all() for g in grads): + raise RuntimeError('Missing/nonfinite OMG gradients') + def on_train_batch_end(self,trainer,module,outputs,batch,batch_idx): + loss=outputs['loss'] if isinstance(outputs,dict) else outputs + if not torch.isfinite(loss).all():raise RuntimeError('Nonfinite OMG training loss') + if trainer.is_global_zero: + with (self.output/'metrics.jsonl').open('a') as f: + f.write(json.dumps(dict(step=trainer.global_step,loss=float(loss.detach()), + elapsed_seconds=time.monotonic()-self.started, + data_wait_seconds=self.data_wait_seconds, + compute_seconds=time.monotonic()-self.batch_started))+'\n') + self.previous_batch_end=time.monotonic() + def on_train_batch_start(self,trainer,module,batch,batch_idx): + self.batch_started=time.monotonic() + self.data_wait_seconds=None if self.previous_batch_end is None else self.batch_started-self.previous_batch_end + def on_validation_batch_end(self,trainer,module,outputs,batch,batch_idx,dataloader_idx=0): + self.validation_batches+=1 + if outputs is not None: + loss=outputs['loss'] if isinstance(outputs,dict) else outputs + if not torch.isfinite(loss).all():raise RuntimeError('Nonfinite OMG validation loss') + def on_validation_end(self,trainer,module): + # Do not count validation/checkpoint time as next-batch data waiting. + self.previous_batch_end=None + if trainer.is_global_zero: + metrics={k:float(v.detach()) for k,v in trainer.callback_metrics.items() if k.startswith('val/') and torch.is_tensor(v) and v.numel()==1} + with (self.output/'validation.jsonl').open('a') as f: + f.write(json.dumps(dict(step=trainer.global_step,validation_batches=self.validation_batches,metrics=metrics))+'\n') + def on_train_end(self,trainer,module): + if trainer.is_global_zero: + best=[dict(path=c.best_model_path,score=float(c.best_model_score)) for c in trainer.checkpoint_callbacks + if c.monitor=='val/loss' and c.best_model_score is not None] + early=[str(getattr(c,'stopping_reason_message',None)) for c in trainer.callbacks if isinstance(c,EarlyStopping)] + (self.output/'training_summary.json').write_text(json.dumps(dict(global_step=trainer.global_step, + elapsed_seconds=time.monotonic()-self.started,best_checkpoints=best,early_stopping=early),indent=2)+'\n') + if not self.check_pipeline:return + changed=any(not torch.equal(self.initial[n],p.detach().cpu()) for n,p in module.named_parameters() if n in self.initial) + file=self.output/'pipeline.ckpt' + trainer.save_checkpoint(str(file)) + trainer.strategy.barrier() + ckpt=torch.load(file,map_location='cpu',weights_only=False) + module.load_state_dict(ckpt['state_dict'],strict=True) + expected_updates=trainer.global_step-getattr(self,'initial_step',0) + passed=changed and self.optimizer_updates==expected_updates + rank_report=dict(rank=trainer.global_rank,device=str(module.device),optimizer_updates=self.optimizer_updates, + peak_allocated_mib=torch.cuda.max_memory_allocated(module.device)/2**20, + peak_reserved_mib=torch.cuda.max_memory_reserved(module.device)/2**20) + (self.output/f'pipeline_rank{trainer.global_rank}.json').write_text(json.dumps(rank_report,indent=2)+'\n') + if trainer.is_global_zero: + report=dict(status='passed' if passed else 'failed',model='omg',framework='pytorch_lightning.Trainer/MotionGenerator.training_step', + finite_losses_and_gradients=True,parameters_changed=changed,optimizer_updates=self.optimizer_updates, + expected_updates=expected_updates,resumed_from_step=getattr(self,'initial_step',0),world_size=trainer.world_size,validation_batches=self.validation_batches,native_checkpoint=str(file),strict_checkpoint_reload=True, + input=(self.audio_policy+' audio + released oracle foreground transcript; no control labels') if self.input_text_source=='oracle_foreground_transcript' else self.audio_policy+(' audio + causal ASR text' if self.use_asr_text else ' audio')+'; no GT text or control labels',audio_policy=self.audio_policy,use_asr_text=self.use_asr_text,purpose='training pipeline health, not duplex ability') + (self.output/'pipeline_check.json').write_text(json.dumps(report,indent=2)+'\n') + if not passed:raise RuntimeError('OMG optimizer did not complete expected updates') diff --git a/handoff-v1/resources/historical-source/hfd_bridge/resume.py b/handoff-v1/resources/historical-source/hfd_bridge/resume.py new file mode 100644 index 0000000000000000000000000000000000000000..7d28f7abfa26c91a6ff3f963afb48f74045d6e9e --- /dev/null +++ b/handoff-v1/resources/historical-source/hfd_bridge/resume.py @@ -0,0 +1,120 @@ +"""Checkpoint selection, run compatibility, RNG state and atomic persistence.""" +import copy +from datetime import datetime, timezone +import json +from pathlib import Path +import random + +import numpy as np +from omegaconf import OmegaConf + + +def rng_state(): + import torch + n=np.random.get_state() + numpy_state=(n[0],n[1].tolist(),int(n[2]),int(n[3]),float(n[4])) + return dict(python=random.getstate(), numpy=numpy_state, torch=torch.get_rng_state(), + cuda=torch.cuda.get_rng_state_all() if torch.cuda.is_initialized() else []) + + +def restore_rng(state): + import torch + random.setstate(state['python']) + n=state['numpy'] + np.random.set_state((n[0],np.asarray(n[1],dtype=np.uint32),n[2],n[3],n[4])) + torch.set_rng_state(state['torch'].cpu()) + if state['cuda']: + torch.cuda.set_rng_state_all([v.cpu() for v in state['cuda']]) + + +def atomic_save(payload, destination): + import torch + destination=Path(destination) + temporary=destination.with_name(destination.name+'.partial') + torch.save(payload, temporary) + temporary.replace(destination) + + +def contract(config): + """Keep optimizer/data semantics fixed; allow output paths and logging cadence.""" + c=OmegaConf.to_container(config, resolve=True) + training=copy.deepcopy(c['training']) + for key in ['resume','save_every','eval_every','keep_last_checkpoints','check_pipeline', + 'validation_batches','validation_batch_size','save_best','early_stopping','num_workers']: + training.pop(key,None) + data=copy.deepcopy(c['data']) + # These do not alter the selected data, text or training batch stream. + for key in ['validation_batch_size','prefetch_batches']: + data.pop(key,None) + return dict(model=c['model'],framework=c['framework'],training=training,data=data, + weights=c['weights'],native_overrides=c.get('native_overrides',{})) + + +def source_run(checkpoint): + checkpoint=Path(checkpoint) + for folder in [checkpoint.parent, checkpoint.parent.parent]: + if (folder/'resolved_config.yaml').is_file(): + return folder + raise ValueError('Resume requires an HFD training run with resolved_config.yaml') + + +def select_checkpoint(config, request, resolve): + if request in (None, False, ''): + return None + if int(config.runtime.num_gpus)!=1: + raise ValueError('Full HFD resume currently requires single-GPU training per model') + output=resolve(config.paths.output_dir) + if request=='auto': + if config.model=='omg': + candidates=[output/'last.ckpt', output/'checkpoints/last.ckpt'] + candidates += sorted((output/'checkpoints').glob('step-*.ckpt'), reverse=True) + else: + candidates=sorted(output.glob('ckpt_*.pth'),key=lambda p:int(p.stem.split('_')[-1]),reverse=True) + candidates=list({p.resolve():p for p in candidates if p.is_file()}.values()) + if not candidates: + raise FileNotFoundError(f'No resumable checkpoint in {output}; auto never starts a new run') + import torch + def progress(candidate): + state=torch.load(candidate,map_location='cpu',weights_only=False) + return int(state.get('global_step',state.get('step',-1))),candidate.stat().st_mtime_ns + checkpoint=max(candidates,key=progress) + else: + checkpoint=resolve(request) + if not checkpoint.is_file(): + raise FileNotFoundError(checkpoint) + source=source_run(checkpoint) + previous=OmegaConf.load(source/'resolved_config.yaml') + if contract(previous)!=contract(config): + raise ValueError('Resume configuration differs in training/data/weights. Use the original run config; max_steps is the original total target, not additional steps.') + import torch + state=torch.load(checkpoint,map_location='cpu',weights_only=False) + step=int(state.get('global_step',state.get('step',-1))) + optimizer=state.get('optimizer_states') if config.model=='omg' else state.get('optimizer') + if not optimizer or not 01 and c.data.cache_sessions= training.batch_size to retain the reused session pool') + assets=[path(recipe.raw_root)/'index.sqlite',path(recipe.raw_root)/'MANIFEST.json'] + assets.append(path(native.reference_cache) if reference else path(native.asr_model)) + for item in assets: + if not item.exists():raise FileNotFoundError(item) + if reference: + import math + guard=float(native.get('reference_guard_seconds',.1)) + if not math.isfinite(guard) or guard<0:raise ValueError('Invalid reference_guard_seconds') + from hfd_bridge.reference_runtime import ReferenceIndex + language=native.get('language','all') + if language not in ('all','en','zh'):raise ValueError('Unknown native foreground language') + if language=='en' and c.data.get('text_conditioning','none')!='none':raise ValueError('English-only reference training must disable translation') + index=ReferenceIndex(path(native.reference_cache),required_languages=['en','zh'] if language=='all' else [language]) + if recipe.get('include_third_party',False) and not index.db.execute('SELECT COUNT(*) FROM reference_recipes').fetchone()[0]: + raise ValueError('Third-party reference recipe index is empty') + if c.training.get('validation_batches',0)<=0:raise ValueError('On-demand requires bounded validation_batches') + train,val=[],[] + else: + train = session_paths(c.data.train_sessions) + val = session_paths(c.data.val_sessions) if c.data.val_sessions else [] + if c.data.get('frame_range') is not None: + span=list(c.data.frame_range) + if c.mode != 'smoke' or len(train) != 1 or len(span) != 2 or any(type(v) is not int for v in span) or not 0 <= span[0] < span[1]: + raise ValueError('data.frame_range requires smoke mode, one session and valid [start,end) frames') + if c.mode == 'train' and not val and not ondemand: + raise ValueError('Formal training requires separate validation sessions') + if c.mode == 'train' and set(train) & set(val): + raise ValueError('Train/validation session paths overlap') + for split, sessions in [('train',train),('val',val)]: + for p in sessions: + meta = json.loads(p.with_suffix('.json').read_text()) + if meta.get('audio_sr') != 16000 or meta.get('motion_fps') != 50: + raise ValueError(f'Unexpected HFD clocks: {p}') + if c.mode == 'train' and meta.get('split') != split: + raise ValueError(f'Expected {split} metadata for {p}') + if c.model == 'textop' or c.data.get('use_asr_text',False): + suffix='.reference.json' if c.data.get('caption_policy')=='aligned_reference_context' else '.asr.json' + sidecar = path(c.data.asr_dir)/(p.stem+suffix) + if not sidecar.is_file(): + raise FileNotFoundError(sidecar) + required = [path(c.paths.upstream_root),path(c.weights.init_checkpoint)] + if c.data.get('text_conditioning','none')=='asr_to_english': + translator=Path(os.environ.get('HFD_TRANSLATOR_ROOT',str(DEFAULT_TRANSLATOR))) + required += [translator/name for name in ['config.json','pytorch_model.bin','source.spm','target.spm','download_identity.json']] + if c.data.get('text_conditioning')=='nllb_to_english': + from hfd_bridge.semantics import DEFAULT_NLLB + translator=Path(os.environ.get('HFD_NLLB_ROOT',str(DEFAULT_NLLB))) + required += [translator/name for name in ['config.json','pytorch_model.bin','sentencepiece.bpe.model','download_identity.json']] + if c.model == 'textop': + required += [path(c.weights.native_config),path(c.weights.vae),path(c.weights.clip),path(c.weights.stats),path(c.weights.robot_assets)] + else: + required += [path(c.weights.models_root)/'t5-base-local'] + for p in required: + if not p.exists(): + raise FileNotFoundError(p) + return train,val + + +def setup(c): + for key,value in c.runtime.get('environment',{}).items(): + os.environ[str(key)] = str(value) + os.environ['CUDA_VISIBLE_DEVICES'] = ','.join(map(str,c.runtime.gpu_ids)) + os.environ['HF_HOME'] = str(path(c.paths.cache_root)/'huggingface') + os.environ['TORCH_HOME'] = str(path(c.paths.cache_root)/'torch') + os.environ['HF_HUB_OFFLINE'] = '1' if c.runtime.offline else '0' + os.environ['TRANSFORMERS_OFFLINE'] = '1' if c.runtime.offline else '0' + os.environ['TOKENIZERS_PARALLELISM'] = 'false' + os.environ['HFD_DATA_DISK'] = str(path(c.paths.data_disk)) + upstream = path(c.paths.upstream_root) + if c.model == 'textop': + sys.path.insert(0,str(upstream/'TextOpRobotMDAR')) + sys.path.insert(0,str(upstream/'deps/isaac_utils')) + sys.path.insert(0,str(upstream/'deps/CLIP')) + else: + sys.path.insert(0,str(upstream/'src')) + os.environ['OMG_MODELS_ROOT'] = str(path(c.weights.models_root)) + + +def textop(c, train, val, output): + from robotmdar.train.train_dar import main as native_main + from hfd_bridge.textop_checks import PipelineDARManager + cfg = OmegaConf.load(path(c.weights.native_config)) + cfg = OmegaConf.merge(cfg,c.native_overrides) + cfg.device='cuda:0'; cfg.seed=c.training.seed + cfg.experiment_dir=str(output); cfg.expname='hfd_baseline_'+c.mode + cfg.ckpt.vae=str(path(c.weights.vae)); cfg.ckpt.dar=c.training.get('resume') + cfg.data.num_primitive=c.training.num_primitive; cfg.data.batch_size=c.training.batch_size + robot=OmegaConf.to_container(cfg.data.skeleton,resolve=True) + robot['asset']['assetRoot']=str(path(c.weights.robot_assets)) + def data(sessions): + return OmegaConf.create(dict(_target_='hfd_bridge.configured_data.TextOpSessionData', + sessions=[str(p) for p in sessions],asr_dir=str(path(c.data.asr_dir)),robot_cfg=robot, + stats_path=str(path(c.weights.stats)),clip_path=str(path(c.weights.clip)), + batch_size=c.training.batch_size,num_primitive=c.training.num_primitive, + max_windows=c.data.max_windows,window_policy=c.data.window_policy, + device=cfg.device,cache_sessions=c.data.cache_sessions,asr_latency=c.data.asr_latency_seconds, + frame_range=OmegaConf.to_container(c.data.frame_range) if c.data.get('frame_range') is not None else None,caption_policy=c.data.get('caption_policy',DEFAULT_CAPTION_POLICY),clock_policy=c.data.get('clock_policy','legacy_hfd'),text_conditioning=c.data.get('text_conditioning','none'))) + cfg.data.train=data(train); cfg.data.val=data(val or train) + if c.data.get('backend')=='native_ondemand': + def native_data(split): + return OmegaConf.create(dict(_target_='hfd_bridge.ondemand_data.OnDemandTextOpData', + native_options=native_options(c),split=split,robot_cfg=robot,stats_path=str(path(c.weights.stats)), + clip_path=str(path(c.weights.clip)),batch_size=int(c.data.get('validation_batch_size',c.training.batch_size)) if split=='val' else c.training.batch_size,num_primitive=c.training.num_primitive, + device=cfg.device,cache_sessions=c.data.cache_sessions,output=str(output))) + cfg.data.train=native_data('train');cfg.data.val=native_data('val') + architecture=OmegaConf.to_container(cfg.denoiser,resolve=True) + cfg.denoiser=OmegaConf.create(dict(_target_='hfd_bridge.native.pretrained_denoiser',_recursive_=False, + architecture=architecture,checkpoint=str(path(c.weights.init_checkpoint)))) + m=cfg.train.manager + m._target_='hfd_bridge.textop_checks.PipelineDARManager'; m.stages=c.training.stages + m.device=cfg.device; m.save_dir=str(output); m.learning_rate=c.training.learning_rate + m.save_every=c.training.save_every; m.eval_every=c.training.eval_every; m.eval_steps=c.training.eval_steps + m.ckpt=cfg.ckpt; m.max_grad_norm=c.training.gradient_clip_val; m.check_pipeline=c.training.get('check_pipeline',c.mode=='smoke') + m.keep_last_checkpoints=c.training.get('keep_last_checkpoints',0) + m.input_text_source='oracle_foreground_transcript' if c.data.get('caption_policy')=='aligned_reference_context' else 'asr' + m.resume_num_primitive=int(c.training.num_primitive) + m.platform=OmegaConf.create(dict(_target_='hfd_bridge.native.ScalarLog',save_dir=str(output))) + OmegaConf.save(cfg,output/'native_config.yaml') + native_main(cfg) + + +def omg(c, train, val, output): + import torch + import pytorch_lightning as pl + from hydra import initialize_config_dir,compose + from torch.utils.data import DataLoader + from pytorch_lightning.callbacks import ModelCheckpoint + from pytorch_lightning.strategies import DDPStrategy + from datetime import timedelta + from hfd_bridge.configured_data import OMGAudioSessionDataset + from hfd_bridge.omg_checks import OMGTrainingObserver,FixedValidationNoise,WarmupEarlyStopping + from hfd_bridge.omg_model_loader import _load_model + pl.seed_everything(c.training.seed,workers=True) + with initialize_config_dir(config_dir=str(path(c.paths.upstream_root)/'configs/generation'),version_base='1.3'): + cfg=compose(config_name='train',overrides=['exp='+c.training.experiment,'logger=none','trainer=1gpu']) + cfg=OmegaConf.merge(cfg,c.native_overrides) + cfg.model.optimizer.lr=c.training.learning_rate + cfg.model.scheduler=OmegaConf.merge(cfg.model.scheduler,c.training.scheduler) + cfg.output_dir=str(output) + model=_load_model(cfg,str(path(c.weights.init_checkpoint)),torch.device('cpu')) + def data(sessions): + return OMGAudioSessionDataset(sessions,model.representation,c.data.window_policy,c.data.max_windows,c.data.cache_sessions,c.data.get('frame_range'),c.data.audio_policy, + use_asr_text=c.data.get('use_asr_text',False),asr_dir=str(path(c.data.asr_dir)),asr_latency=c.data.asr_latency_seconds,caption_policy=c.data.get('caption_policy',DEFAULT_CAPTION_POLICY),clock_policy=c.data.get('clock_policy','legacy_hfd'),text_conditioning=c.data.get('text_conditioning','none')) + workers=c.training.num_workers + loader_options=dict(batch_size=c.training.batch_size,num_workers=workers,pin_memory=True) + if workers: loader_options['persistent_workers']=True + ondemand=c.data.get('backend')=='native_ondemand' + if ondemand: + from hfd_bridge.ondemand_data import OnDemandOMGBatches + def batches(split): + return OnDemandOMGBatches(native_options(c),split,model.representation,int(c.data.get('validation_batch_size',c.training.batch_size)) if split=='val' else c.training.batch_size,c.data.cache_sessions,str(output), + fixed_validation=c.training.get('fixed_validation',False),batched_motion=c.data.get('batched_motion',True), + prefetch_batches=c.data.get('prefetch_batches',0), + max_batches=int(c.training.max_steps*c.training.accumulate_grad_batches) if split=='train' else int(c.training.validation_batches)) + train_loader=DataLoader(batches('train'),batch_size=None,num_workers=0) + val_loader=DataLoader(batches('val'),batch_size=None,num_workers=0) + else: + train_loader=DataLoader(data(train),shuffle=True,**loader_options) + val_loader=DataLoader(data(val),shuffle=False,**loader_options) if val else None + keep=int(c.training.get('keep_last_checkpoints',0)) + checkpoint=ModelCheckpoint(dirpath=str(output/'checkpoints'),every_n_train_steps=c.training.save_every, + save_top_k=keep if keep else -1,monitor='step' if keep else None,mode='max', + save_last='link' if keep else True,filename='step-{step:08d}',auto_insert_metric_name=False) + from hfd_bridge.resume_lightning import HFDResumeState + callbacks=[HFDResumeState(train_loader.dataset if ondemand else None,val_loader.dataset if ondemand else None), + checkpoint,OMGTrainingObserver(str(output),c.training.get('check_pipeline',c.mode=='smoke'),audio_policy=c.data.audio_policy,use_asr_text=c.data.get('use_asr_text',False),input_text_source='oracle_foreground_transcript' if c.data.get('caption_policy')=='aligned_reference_context' else 'asr')] + if ondemand or val: + if c.training.get('save_best',False): + callbacks.append(ModelCheckpoint(dirpath=str(output/'best'),monitor='val/loss',mode='min',save_top_k=1, + filename='best',auto_insert_metric_name=False,enable_version_counter=False, + every_n_train_steps=0,every_n_epochs=1,save_on_train_epoch_end=False)) + if c.training.get('fixed_validation',False):callbacks.append(FixedValidationNoise(c.training.seed+1000003)) + early=c.training.get('early_stopping',{}) + if early.get('enabled',False): + callbacks.append(WarmupEarlyStopping(monitor='val/loss',mode='min',start_step=early.get('start_step',30000), + patience=early.get('patience',10),min_delta=early.get('min_delta',0.0002),verbose=True,log_rank_zero_only=True)) + trainer_options=OmegaConf.to_container(cfg.trainer,resolve=True) + trainer_options.update(accelerator='gpu',devices=c.runtime.num_gpus, + strategy=DDPStrategy(find_unused_parameters=True,process_group_backend=c.runtime.get('distributed_backend','nccl'),timeout=timedelta(seconds=int(c.runtime.get('distributed_timeout_seconds',120)))) if c.runtime.num_gpus>1 else 'auto', + max_steps=c.training.max_steps,precision=c.training.precision, + accumulate_grad_batches=c.training.accumulate_grad_batches, + gradient_clip_val=c.training.gradient_clip_val,logger=False,enable_checkpointing=True, + enable_progress_bar=True,num_sanity_val_steps=0,default_root_dir=str(output), + callbacks=callbacks, + use_distributed_sampler=c.runtime.num_gpus>1 and not ondemand, + val_check_interval=c.training.eval_every,check_val_every_n_epoch=None) + if ondemand: + trainer_options['limit_val_batches']=int(c.training.validation_batches) + elif not val: + trainer_options['limit_val_batches']=0 + if os.environ.get('LOCAL_RANK','0')=='0': + OmegaConf.save(cfg,output/'native_config.yaml') + trainer=pl.Trainer(**trainer_options) + model.train() + if model.text_encoder is not None:model.text_encoder.encoder.eval() + trainer.fit(model,train_dataloaders=train_loader,val_dataloaders=val_loader,ckpt_path=c.training.get('resume')) + trainer.save_checkpoint(str(output/'last.ckpt')) + + +def native_options(c): + options=OmegaConf.to_container(c.data.native,resolve=True) + for key in ('config','asr_cache','asr_model','reference_cache'): + if key in options:options[key]=str(path(options[key])) + local_rank=int(os.environ.get('LOCAL_RANK','0')) + options.update(seed=c.training.seed,asr_device=f'cuda:{local_rank}',caption_policy=c.data.caption_policy,asr_latency=c.data.asr_latency_seconds,clock_policy=c.data.get('clock_policy','legacy_hfd'),text_conditioning=c.data.get('text_conditioning','none')) + recipe=OmegaConf.load(path(options['config']));OmegaConf.resolve(recipe) + options['include_third_party']=bool(recipe.get('include_third_party',False)) + return options + + +def main(): + p=argparse.ArgumentParser(description=__doc__) + p.add_argument('--config',required=True,type=Path) + p.add_argument('--set',nargs='*',default=[],help='OmegaConf dotted overrides, e.g. training.batch_size=16') + p.add_argument('--dry-run',action='store_true',help='Resolve and validate configuration without training') + p.add_argument('--resume',nargs='?',const='auto',help='Continue full training state from a checkpoint; omit path to select latest in output_dir') + a=p.parse_args() + c=OmegaConf.load(a.config) + c=OmegaConf.merge(c,OmegaConf.from_dotlist(a.set)); OmegaConf.resolve(c) + if c.runtime.python: + interpreter=Path(str(c.runtime.python)).expanduser() + interpreter=(interpreter if interpreter.is_absolute() else ROOT/interpreter).absolute() + if Path(sys.executable).absolute() != interpreter: + os.execv(str(interpreter),[str(interpreter),str(Path(__file__).resolve()),*sys.argv[1:]]) + train,val=prepare(c) + from hfd_bridge.resume import select_checkpoint,record_resume + resume=select_checkpoint(c,a.resume if a.resume is not None else c.training.get('resume'),path) + c.training.resume=str(resume) if resume else None + if a.dry_run: + print(OmegaConf.to_yaml(c)) + print(json.dumps(dict(status='validated_configuration',backend=c.data.get('backend','exported_npz'), + train_sessions=len(train) if train else None,val_sessions=len(val) if val else None, + note='Native index and on-demand reads are checked at runtime' if c.data.get('backend')=='native_ondemand' else 'Exported session paths checked'))) + return + setup(c) + if c.model=='omg' and c.runtime.num_gpus>1 and 'LOCAL_RANK' not in os.environ: + os.execv(sys.executable,[sys.executable,'-m','torch.distributed.run','--standalone', + '--nproc_per_node='+str(c.runtime.num_gpus),str(Path(__file__).resolve()),*sys.argv[1:]]) + import torch + torch.set_num_threads(c.runtime.cpu_threads);torch.manual_seed(c.training.seed) + if torch.cuda.device_count()!=c.runtime.num_gpus: + raise RuntimeError('Requested GPU visibility does not match available devices') + output=path(c.paths.output_dir) + rank=int(os.environ.get('LOCAL_RANK','0')) + torch.cuda.set_device(rank) + if rank==0: + existing=output.exists() + output.mkdir(parents=True,exist_ok=resume is not None) + if not existing:OmegaConf.save(c,output/'resolved_config.yaml') + audit_output=record_resume(output,c,resume) if resume else output + from hfd_bridge.semantics import conditioning_identity + manifest=dict(text_frontend=conditioning_identity(c.data.get('text_conditioning','none')), + clock_policy=c.data.get('clock_policy','legacy_hfd'),backend=c.data.get('backend','exported_npz'),train=list(map(str,train)),val=list(map(str,val))) + if c.data.get('caption_policy')=='aligned_reference_context': + manifest.update(text_source='oracle_foreground_transcript',caption_policy='aligned_reference_context', + reference_sidecars=[str(path(c.data.asr_dir)/(p.stem+'.reference.json')) for p in train+val]) + if c.data.get('backend')=='native_ondemand': + manifest.update(native_options=native_options(c),split_manifests=['train_native_manifest.json','val_native_manifest.json']) + (audit_output/'data_manifest.json').write_text(json.dumps(manifest,indent=2)+'\n') + provenance={} + for sub in [ROOT/'hfd_bridge',path(c.paths.upstream_root)]: + for f in sub.rglob('*.py'): + provenance[str(f.relative_to(ROOT)) if f.is_relative_to(ROOT) else str(f)]=hashlib.sha256(f.read_bytes()).hexdigest() + provenance['train.py']=hashlib.sha256(Path(__file__).read_bytes()).hexdigest() + (audit_output/'code_sha256.json').write_text(json.dumps(provenance,indent=2)+'\n') + (textop if c.model=='textop' else omg)(c,train,val,output) + if rank==0:print('COMPLETED',output,flush=True) + +if __name__=='__main__': + main() diff --git a/handoff-v1/resources/report_all_baselines.py b/handoff-v1/resources/report_all_baselines.py new file mode 100644 index 0000000000000000000000000000000000000000..d33f700073335b9da892c1a9bdd2d73cffd40ed8 --- /dev/null +++ b/handoff-v1/resources/report_all_baselines.py @@ -0,0 +1,31 @@ +from pathlib import Path +import json,torch,hashlib +R=Path('/root/autodl-fs/pangxinlei/handoffs/all-baselines-20261111');models=json.loads((R/'inventory/models.json').read_text());results={} +for m in ['textop','omg-l','omg-xl']: + r=R/'acceptance'/m + if not (r/'exit.json').exists():continue + e=json.loads((r/'exit.json').read_text());target=299780 if m=='textop' else 98002 if m=='omg-l' else 90002 + ck=r/('run/ckpt_299780.pth' if m=='textop' else 'weights/last.ckpt') + if not ck.exists():results[m]={'exit':e,'status':'failed_no_output_checkpoint'};continue + d=torch.load(ck,map_location='cpu',weights_only=False,mmap=True);params=d.get('state_dict',d.get('denoiser'));optim=d['optimizer'] if m=='textop' else d['optimizer_states'][0] + tensor_rows=[] + for k,t in params.items(): + if not torch.is_tensor(t):continue + tensor_rows.append({'name':k,'shape':list(t.shape),'dtype':str(t.dtype),'finite':bool(torch.isfinite(t).all()),'sha256':hashlib.sha256(t.contiguous().reshape(-1).view(torch.uint8).numpy().tobytes()).hexdigest()}) + (r/'updated-parameter-sha256.json').write_text(json.dumps(tensor_rows,indent=2));step=d.get('global_step',d.get('step')) + old={} + with (Path(models[m]['run'])/'metrics.jsonl').open() as f: + for l in f: + x=json.loads(l) + if target-4