Training migration metadata 2026-10-12

.gitattributes

@@ -58,3 +58,8 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 # Video files - compressed
 *.mp4 filter=lfs diff=lfs merge=lfs -text
 *.webm filter=lfs diff=lfs merge=lfs -text
+home/wuyifan/a7_canonical_gt_rotationfix_20260930/cache/native_inventory.json filter=lfs diff=lfs merge=lfs -text
+home/wuyifan/a7_canonical_gt_rotationfix_20260930/label_inventory.json filter=lfs diff=lfs merge=lfs -text
+home/wuyifan/foundation_hand_clean_gt_dino_20260925/cache/adapter_trainable_t8/arctic_t8.jsonl filter=lfs diff=lfs merge=lfs -text
+home/wuyifan/foundation_hand_clean_gt_dino_20260925/cache/adapter_trainable_t8/hot3d_t8.jsonl filter=lfs diff=lfs merge=lfs -text
+home/wuyifan/foundation_hand_clean_gt_dino_20260925/cache/trainable_t8_index.sqlite filter=lfs diff=lfs merge=lfs -text

README.md

@@ -0,0 +1,16 @@
+# EgoNet training migration metadata
+
+This private repository preserves the code snapshot, configuration, fixed
+8-frame training manifests, SQLite sampling index, normalization statistics,
+and canonical-label provenance for the H2O, ARCTIC, HOT3D, and GRAB mixed run.
+Paths below `home/wuyifan/` mirror their original absolute paths on g2.
+
+This repository is not a complete training dataset. It does not contain source
+RGB/depth/VRS recordings, rendered GRAB videos, filtered GT NPZ files,
+canonical label NPZ files, personalized MANO templates, licensed MANO model
+files, ViTPose visual weights, or training checkpoints. Restore those through
+the appropriate authorized channels before running the training code.
+
+The manifests and inventory contain absolute source paths and content hashes.
+Preserve the original path layout on the destination, or explicitly rebuild
+the associated index, inventory, and provenance checks after changing paths.

home/wuyifan/EgoNet_train_20261008/.gitignore

@@ -0,0 +1,36 @@
+__pycache__/
+*.py[cod]
+.pytest_cache/
+.ruff_cache/
+.mypy_cache/
+.venv/
+*.egg-info/
+
+.env
+.env.*
+*.pem
+*.key
+id_rsa*
+id_ed25519*
+
+data/
+datasets/
+cache/
+checkpoints/
+runs/
+artifacts/
+third_party/
+*.pt
+*.pth
+*.ckpt
+*.safetensors
+*.npy
+*.npz
+*.vrs
+*.mp4
+*.avi
+*.png
+*.jpg
+*.jpeg
+*.lock
+*.log

home/wuyifan/EgoNet_train_20261008/README.md

@@ -0,0 +1,19 @@
+# EgoNet
+
+EgoNet is the review snapshot of the Foundation Hand bimanual reconstruction code and its current RGB training lineage. The Python package is still named `foundation_hand` so checkpoint keys and historical imports remain recognizable.
+
+This repository contains model and dataset code, training/evaluation scripts, configs, tests, and the experiment-specific source used for S4, A7, U0, N0, and N1. It does **not** contain datasets, fitted label caches, MANO assets, DINO weights, model checkpoints, videos, or third-party source trees. See [REVIEW_GUIDE.md](REVIEW_GUIDE.md) for the exact lineage, review entry points, and known limitations.
+
+The snapshot preserves the code that produced the reported experiments. In particular, the historical N0/N1 trainer still has a teacher term and the known bf16 GT-6D precision issue; do not use it for the new B0/B1 contract. The new no-teacher, MANO-pkl-mean-pose-delta model and loss path has passed two real updates per arm, but formal training and benchmark validation have not started.
+
+The matched RGB non-MoT/MoT smoke configuration is in `experiments/from_scratch_mean_pose/config.json`. Both arms use mean-pose delta and normalized orthogonal GT 6D. It remains smoke-only (`training_approved=false`) and is not consumed by the historical trainer. See `docs/PIONAI_NON2D_HANDOFF_20261005.md` for the PionAI transfer and readiness status.
+
+Primary materials:
+
+- `experiments/u0_normalized6d/train.py`: N0/N1 training entry point.
+- `experiments/u0_normalized6d/normalized_loss.py`: normalized direct-6D and geometry supervision.
+- `src/foundation_hand/`: core architecture, data readers, MANO geometry, and losses.
+- `docs/HK.md`: evaluation and hard-pose diagnostics.
+- `docs/FROM_SCRATCH_MEAN_POSE_REVIEW.md`: proposed next controlled experiment, not launched.
+
+Python 3.11 is required by `pyproject.toml`. The original environment also requires the external packages and pinned upstream repositories listed in [THIRD_PARTY.md](THIRD_PARTY.md), plus separately provided licensed MANO assets and datasets. Several historical scripts contain host-specific absolute paths; they are retained for audit, not presented as portable launch commands.

home/wuyifan/EgoNet_train_20261008/REVIEW_GUIDE.md

@@ -0,0 +1,40 @@
+# Code review guide
+
+## Scope and provenance
+
+The root `src/foundation_hand/`, `scripts/`, and training helper files are from the frozen `foundation_hand_clean_gt_dino_20260925/code` snapshot used by the current lineage. Missing general tests and historical configs come from the `/home/jiasen/foundation_hand` working tree; where files overlapped, the frozen training snapshot takes precedence. The original project was at commit `5d59528` before local changes. No source tree or checkpoint was modified to make this review snapshot.
+
+The experiment chain uses these directories, in dependency order:
+
+| Directory | Role |
+| --- | --- |
+| `experiments/clean_gt_dino` | Clean-GT data filtering and from-DINO training helpers |
+| `experiments/s4_modality_experts` | RGB/K/depth expert sampling and training contracts |
+| `experiments/a7_mot_residual` | MoT residual expert code |
+| `experiments/a7_canonical_gt_cache` | Corrected-left MANO label migration and audits |
+| `experiments/a7_rotation_fix` | Frozen A7 rotation-fix data/training scripts |
+| `experiments/a7_canonical_gt_train` | Canonical GT A7 trainer and geometry validation |
+| `experiments/u0_depthsafe` | U0 RGB parent model, loss, and trainer |
+| `experiments/u0_normalized6d` | Matched N0/N1 direct-6D continuation trainer |
+| `experiments/pose_delta_diagnostics` | Frozen-panel and hard-case diagnostic scripts |
+
+These are source snapshots, not a single importable package. Their original configs name absolute paths under `/home/wuyifan` and data roots under `/SSD*`; the review of what ran should use those values together with the SHA256 checkpoint and code records in the experiment reports. To rerun elsewhere, map those paths, obtain the datasets/weights/assets, and revalidate each hash-bound registration. Do not treat a path substitution alone as a reproduced experiment.
+
+## Start here for the current training review
+
+1. `experiments/u0_normalized6d/config.json` fixes a 60-clip update with H2O/ARCTIC/HOT3D quotas 21/15/24, AdamW learning rates, loss coefficients, the RGB teacher weight, and checkpoint interval.
+2. `experiments/u0_normalized6d/train.py` controls data coverage, source scaling, DDP, optimizer groups, clipping, checkpoints, and provenance.
+3. `experiments/u0_normalized6d/normalized_loss.py` defines the direct rotation-6D and geometry terms. `src/foundation_hand/models/bimanual.py` and `src/foundation_hand/models/decoder.py` define prediction and temporal fusion.
+4. `experiments/a7_canonical_gt_train/targets.py` and `common.py` apply corrected-left labels and check MANO reconstruction. `src/foundation_hand/data/{h2o,arctic,hot3d}_dataset.py` construct camera-space source targets.
+5. `experiments/pose_delta_diagnostics/pose_hard_case_monitor.py` defines GT-hard buckets and paired evaluation. The reports in `reports/` use source-native GT on fixed development panels; they are not official benchmark submissions.
+
+## Known review findings
+
+- N0/N1 both continue from the same U0 RGB checkpoint and use a teacher geometry anchor of weight 0.05. They do not test from-scratch or mean-pose-delta training.
+- The training loss compares Gram-Schmidt-normalized predicted 6D with the first two rows of GT rotations. GT axis-angle conversion is currently executed inside bf16 autocast, causing a small orthogonality error. This is recorded in `docs/HK.md`; it is not silently patched in this snapshot.
+- Valid 3D hand geometry and root pose are camera-space. The 15 MANO finger rotations are local to the kinematic chain. The cached label is axis-angle; 6D is generated during loss computation.
+- N1 improves some high-bend HOT3D hands but does not pass cross-source four-metric/no-harm gates. N0 was stopped after log step 19,247; its last complete checkpoint is 18K. See `docs/HK.md` for matched 18K numbers and the N1 20K curve point.
+
+## What is intentionally absent
+
+No raw or filtered dataset, per-frame label cache, MANO asset, DINO weight, training checkpoint, optimizer state, image/video, SSH credential, or access token is included. Runtime JSON manifests that can enumerate private data recordings are omitted; short config and aggregate reports are included. The code can be reviewed immediately but cannot reproduce training from this repository alone.

home/wuyifan/EgoNet_train_20261008/THIRD_PARTY.md

@@ -0,0 +1,12 @@
+# External dependencies
+
+`pyproject.toml` lists Python package dependencies. The original Foundation Hand repository also recorded these external repositories as Git submodules; their source is not redistributed in this review snapshot:
+
+| Upstream | Original revision |
+| --- | --- |
+| `https://github.com/taeinkwon/h2odataset.git` | `ebcb48996849f747580907dc22f8745633ae60b0` |
+| `https://github.com/facebookresearch/dinov3.git` | `346f38fee679c56a6888f91c51670fae61d364e0` |
+| `https://github.com/taeinkwon/h2oplayer.git` | `6f30a287f04e50bb96c6ba5e986ff72c75dd05e5` |
+| `https://github.com/hassony2/manopth.git` | `4f1dcad1201ff1bfca6e065a85f0e3456e1aa32b` |
+
+The HOT3D Python API was present in the original runtime as a separate third-party tree, but no pinned upstream revision was registered in the project's `.gitmodules`. Reviewers should obtain a version compatible with the original dataset API before attempting to rerun its reader. MANO model files and DINO pretrained weights require their own licenses and are not included.

home/wuyifan/EgoNet_train_20261008/configs/a_rgb_single_frame.yaml

@@ -0,0 +1,16 @@
+name: a_rgb_single_frame
+seed: 20260716
+data: {resolution_hw: [288, 512], T: 1, stride: 1, batch_size: 2}
+model:
+  encoder: dinov3
+  decoder_dim: 512
+  decoder_layers: 3
+  attention_heads: 8
+  decoder_mode: sequential
+  use_temporal: false
+  use_depth: false
+  use_rays: false
+  use_bimanual_interaction: true
+  require_mano: true
+assets: {dinov3_source: third_party/dinov3, dinov3_weights: /SSD3/jiasen/foundation_hand/pretrained/dinov3, mano_root: /SSD3/jiasen/foundation_hand/pretrained/mano}
+precision: bf16

home/wuyifan/EgoNet_train_20261008/configs/b_rgb_video_sequential.yaml

@@ -0,0 +1,16 @@
+name: b_rgb_video_sequential
+seed: 20260716
+data: {resolution_hw: [288, 512], T: 8, stride: 1, batch_size: 1}
+model:
+  encoder: dinov3
+  decoder_dim: 512
+  decoder_layers: 3
+  attention_heads: 8
+  decoder_mode: sequential
+  use_temporal: true
+  use_depth: false
+  use_rays: false
+  use_bimanual_interaction: true
+  require_mano: true
+assets: {dinov3_source: third_party/dinov3, dinov3_weights: /SSD3/jiasen/foundation_hand/pretrained/dinov3, mano_root: /SSD3/jiasen/foundation_hand/pretrained/mano}
+precision: bf16

home/wuyifan/EgoNet_train_20261008/configs/c_rgb_video_alternating.yaml

@@ -0,0 +1,16 @@
+name: c_rgb_video_alternating
+seed: 20260716
+data: {resolution_hw: [288, 512], T: 8, stride: 1, batch_size: 1}
+model:
+  encoder: dinov3
+  decoder_dim: 512
+  decoder_layers: 3
+  attention_heads: 8
+  decoder_mode: alternating
+  use_temporal: true
+  use_depth: false
+  use_rays: false
+  use_bimanual_interaction: true
+  require_mano: true
+assets: {dinov3_source: third_party/dinov3, dinov3_weights: /SSD3/jiasen/foundation_hand/pretrained/dinov3, mano_root: /SSD3/jiasen/foundation_hand/pretrained/mano}
+precision: bf16

home/wuyifan/EgoNet_train_20261008/configs/d_rgb_k_video_alternating.yaml

@@ -0,0 +1,16 @@
+name: d_rgb_k_video_alternating
+seed: 20260716
+data: {resolution_hw: [288, 512], T: 8, stride: 1, batch_size: 1}
+model:
+  encoder: dinov3
+  decoder_dim: 512
+  decoder_layers: 3
+  attention_heads: 8
+  decoder_mode: alternating
+  use_temporal: true
+  use_depth: false
+  use_rays: true
+  use_bimanual_interaction: true
+  require_mano: true
+assets: {dinov3_source: third_party/dinov3, dinov3_weights: /SSD3/jiasen/foundation_hand/pretrained/dinov3, mano_root: /SSD3/jiasen/foundation_hand/pretrained/mano}
+precision: bf16

home/wuyifan/EgoNet_train_20261008/configs/e_rgb_depth_video_alternating.yaml

@@ -0,0 +1,18 @@
+name: e_rgb_depth_video_alternating
+seed: 20260716
+data: {resolution_hw: [288, 512], T: 8, stride: 1, batch_size: 1}
+model:
+  encoder: dinov3
+  decoder_dim: 512
+  decoder_layers: 3
+  attention_heads: 8
+  decoder_mode: alternating
+  use_temporal: true
+  use_depth: true
+  depth_mode: log_depth
+  depth_gate_floor: 0.05
+  use_rays: false
+  use_bimanual_interaction: true
+  require_mano: true
+assets: {dinov3_source: third_party/dinov3, dinov3_weights: /SSD3/jiasen/foundation_hand/pretrained/dinov3, mano_root: /SSD3/jiasen/foundation_hand/pretrained/mano}
+precision: bf16

home/wuyifan/EgoNet_train_20261008/configs/e_rgb_depth_video_alternating_fixed_gate.yaml

@@ -0,0 +1,19 @@
+name: e_rgb_depth_video_alternating_fixed_gate
+seed: 20260716
+data: {resolution_hw: [288, 512], T: 8, stride: 1, batch_size: 1}
+model:
+  encoder: dinov3
+  decoder_dim: 512
+  decoder_layers: 3
+  attention_heads: 8
+  decoder_mode: alternating
+  use_temporal: true
+  use_depth: true
+  depth_mode: log_depth
+  depth_gate_mode: fixed
+  depth_gate_fixed_value: 0.5
+  use_rays: false
+  use_bimanual_interaction: true
+  require_mano: true
+assets: {dinov3_source: third_party/dinov3, dinov3_weights: /SSD3/jiasen/foundation_hand/pretrained/dinov3, mano_root: /SSD3/jiasen/foundation_hand/pretrained/mano}
+precision: bf16

home/wuyifan/EgoNet_train_20261008/configs/f_geometry_ablation.yaml

@@ -0,0 +1,18 @@
+name: f_geometry_ablation
+seed: 20260716
+data: {resolution_hw: [288, 512], T: 8, stride: 1, batch_size: 1}
+model:
+  encoder: dinov3
+  decoder_dim: 512
+  decoder_layers: 3
+  attention_heads: 8
+  decoder_mode: alternating
+  use_temporal: true
+  use_depth: true
+  depth_mode: geometry
+  depth_gate_floor: 0.05
+  use_rays: true
+  use_bimanual_interaction: true
+  require_mano: true
+assets: {dinov3_source: third_party/dinov3, dinov3_weights: /SSD3/jiasen/foundation_hand/pretrained/dinov3, mano_root: /SSD3/jiasen/foundation_hand/pretrained/mano}
+precision: bf16

home/wuyifan/EgoNet_train_20261008/configs/f_geometry_fixed_gate_ablation.yaml

@@ -0,0 +1,19 @@
+name: f_geometry_fixed_gate_ablation
+seed: 20260716
+data: {resolution_hw: [288, 512], T: 8, stride: 1, batch_size: 1}
+model:
+  encoder: dinov3
+  decoder_dim: 512
+  decoder_layers: 3
+  attention_heads: 8
+  decoder_mode: alternating
+  use_temporal: true
+  use_depth: true
+  depth_mode: geometry
+  depth_gate_mode: fixed
+  depth_gate_fixed_value: 0.5
+  use_rays: true
+  use_bimanual_interaction: true
+  require_mano: true
+assets: {dinov3_source: third_party/dinov3, dinov3_weights: /SSD3/jiasen/foundation_hand/pretrained/dinov3, mano_root: /SSD3/jiasen/foundation_hand/pretrained/mano}
+precision: bf16

home/wuyifan/EgoNet_train_20261008/configs/f_rgb_depth_k_video_alternating.yaml

@@ -0,0 +1,18 @@
+name: f_rgb_depth_k_video_alternating
+seed: 20260716
+data: {resolution_hw: [288, 512], T: 8, stride: 1, batch_size: 1}
+model:
+  encoder: dinov3
+  decoder_dim: 512
+  decoder_layers: 3
+  attention_heads: 8
+  decoder_mode: alternating
+  use_temporal: true
+  use_depth: true
+  depth_mode: log_depth
+  depth_gate_floor: 0.05
+  use_rays: true
+  use_bimanual_interaction: true
+  require_mano: true
+assets: {dinov3_source: third_party/dinov3, dinov3_weights: /SSD3/jiasen/foundation_hand/pretrained/dinov3, mano_root: /SSD3/jiasen/foundation_hand/pretrained/mano}
+precision: bf16

home/wuyifan/EgoNet_train_20261008/configs/f_rgb_depth_k_video_alternating_fixed_gate.yaml

@@ -0,0 +1,19 @@
+name: f_rgb_depth_k_video_alternating_fixed_gate
+seed: 20260716
+data: {resolution_hw: [288, 512], T: 8, stride: 1, batch_size: 1}
+model:
+  encoder: dinov3
+  decoder_dim: 512
+  decoder_layers: 3
+  attention_heads: 8
+  decoder_mode: alternating
+  use_temporal: true
+  use_depth: true
+  depth_mode: log_depth
+  depth_gate_mode: fixed
+  depth_gate_fixed_value: 0.5
+  use_rays: true
+  use_bimanual_interaction: true
+  require_mano: true
+assets: {dinov3_source: third_party/dinov3, dinov3_weights: /SSD3/jiasen/foundation_hand/pretrained/dinov3, mano_root: /SSD3/jiasen/foundation_hand/pretrained/mano}
+precision: bf16

home/wuyifan/EgoNet_train_20261008/configs/phase3/core/core_m0_l0_t0_s20260716.yaml

@@ -0,0 +1,63 @@
+name: core_m0_l0_t0_s20260716
+seed: 20260716
+experiment:
+  phase: 3
+  matrix: decoder_temporal_core
+  modality: M0
+  layout: L0
+  temporal: T0
+  registered_axes:
+  - layout
+  - temporal_attention
+  - modality
+data:
+  resolution_hw:
+  - 288
+  - 512
+  T: 8
+  stride: 1
+  batch_size: 1
+model:
+  encoder: dinov3
+  decoder_dim: 512
+  decoder_layers: 3
+  attention_heads: 8
+  decoder_mode: sequential
+  use_temporal: false
+  use_depth: false
+  depth_mode: log_depth
+  depth_gate_mode: fixed
+  depth_gate_fixed_value: 0.5
+  use_rays: false
+  use_bimanual_interaction: true
+  predict_presence: true
+  require_mano: true
+loss:
+  joint2d: 1.0
+  joint3d_camera: 10.0
+  vertices_camera: 1.0
+  local_pose: 1.0
+  global_orientation: 1.0
+  shape: 0.1
+  translation: 10.0
+  reprojection: 1.0
+  temporal_velocity: 0.1
+  temporal_acceleration: 0.05
+  bimanual_relative_wrist: 1.0
+  presence: 0.0
+assets:
+  dinov3_source: third_party/dinov3
+  dinov3_weights: /SSD3/jiasen/foundation_hand/pretrained/dinov3
+  mano_root: /SSD3/jiasen/foundation_hand/pretrained/mano
+precision: bf16
+protocol:
+  max_steps: 100000
+  num_workers: 4
+  checkpoint_interval: 25000
+  validation_interval: 5000
+  validation_batches: 256
+  learning_rate: 5.0e-05
+  min_learning_rate: 5.0e-06
+  warmup_steps: 500
+  early_stopping_patience: 0
+  validation_subset_manifest: artifacts/phase3/validation_subset/validation_subset_256.json

home/wuyifan/EgoNet_train_20261008/configs/phase3/core/core_m0_l0_t0_s20260717.yaml

@@ -0,0 +1,63 @@
+name: core_m0_l0_t0_s20260717
+seed: 20260717
+experiment:
+  phase: 3
+  matrix: decoder_temporal_core
+  modality: M0
+  layout: L0
+  temporal: T0
+  registered_axes:
+  - layout
+  - temporal_attention
+  - modality
+data:
+  resolution_hw:
+  - 288
+  - 512
+  T: 8
+  stride: 1
+  batch_size: 1
+model:
+  encoder: dinov3
+  decoder_dim: 512
+  decoder_layers: 3
+  attention_heads: 8
+  decoder_mode: sequential
+  use_temporal: false
+  use_depth: false
+  depth_mode: log_depth
+  depth_gate_mode: fixed
+  depth_gate_fixed_value: 0.5
+  use_rays: false
+  use_bimanual_interaction: true
+  predict_presence: true
+  require_mano: true
+loss:
+  joint2d: 1.0
+  joint3d_camera: 10.0
+  vertices_camera: 1.0
+  local_pose: 1.0
+  global_orientation: 1.0
+  shape: 0.1
+  translation: 10.0
+  reprojection: 1.0
+  temporal_velocity: 0.1
+  temporal_acceleration: 0.05
+  bimanual_relative_wrist: 1.0
+  presence: 0.0
+assets:
+  dinov3_source: third_party/dinov3
+  dinov3_weights: /SSD3/jiasen/foundation_hand/pretrained/dinov3
+  mano_root: /SSD3/jiasen/foundation_hand/pretrained/mano
+precision: bf16
+protocol:
+  max_steps: 100000
+  num_workers: 4
+  checkpoint_interval: 25000
+  validation_interval: 5000
+  validation_batches: 256
+  learning_rate: 5.0e-05
+  min_learning_rate: 5.0e-06
+  warmup_steps: 500
+  early_stopping_patience: 0
+  validation_subset_manifest: artifacts/phase3/validation_subset/validation_subset_256.json

home/wuyifan/EgoNet_train_20261008/configs/phase3/core/core_m0_l0_t0_s20260718.yaml

@@ -0,0 +1,63 @@
+name: core_m0_l0_t0_s20260718
+seed: 20260718
+experiment:
+  phase: 3
+  matrix: decoder_temporal_core
+  modality: M0
+  layout: L0
+  temporal: T0
+  registered_axes:
+  - layout
+  - temporal_attention
+  - modality
+data:
+  resolution_hw:
+  - 288
+  - 512
+  T: 8
+  stride: 1
+  batch_size: 1
+model:
+  encoder: dinov3
+  decoder_dim: 512
+  decoder_layers: 3
+  attention_heads: 8
+  decoder_mode: sequential
+  use_temporal: false
+  use_depth: false
+  depth_mode: log_depth
+  depth_gate_mode: fixed
+  depth_gate_fixed_value: 0.5
+  use_rays: false
+  use_bimanual_interaction: true
+  predict_presence: true
+  require_mano: true
+loss:
+  joint2d: 1.0
+  joint3d_camera: 10.0
+  vertices_camera: 1.0
+  local_pose: 1.0
+  global_orientation: 1.0
+  shape: 0.1
+  translation: 10.0
+  reprojection: 1.0
+  temporal_velocity: 0.1
+  temporal_acceleration: 0.05
+  bimanual_relative_wrist: 1.0
+  presence: 0.0
+assets:
+  dinov3_source: third_party/dinov3
+  dinov3_weights: /SSD3/jiasen/foundation_hand/pretrained/dinov3
+  mano_root: /SSD3/jiasen/foundation_hand/pretrained/mano
+precision: bf16
+protocol:
+  max_steps: 100000
+  num_workers: 4
+  checkpoint_interval: 25000
+  validation_interval: 5000
+  validation_batches: 256
+  learning_rate: 5.0e-05
+  min_learning_rate: 5.0e-06
+  warmup_steps: 500
+  early_stopping_patience: 0
+  validation_subset_manifest: artifacts/phase3/validation_subset/validation_subset_256.json

home/wuyifan/EgoNet_train_20261008/configs/phase3/core/core_m0_l0_t1_s20260716.yaml

@@ -0,0 +1,63 @@
+name: core_m0_l0_t1_s20260716
+seed: 20260716
+experiment:
+  phase: 3
+  matrix: decoder_temporal_core
+  modality: M0
+  layout: L0
+  temporal: T1
+  registered_axes:
+  - layout
+  - temporal_attention
+  - modality
+data:
+  resolution_hw:
+  - 288
+  - 512
+  T: 8
+  stride: 1
+  batch_size: 1
+model:
+  encoder: dinov3
+  decoder_dim: 512
+  decoder_layers: 3
+  attention_heads: 8
+  decoder_mode: sequential
+  use_temporal: true
+  use_depth: false
+  depth_mode: log_depth
+  depth_gate_mode: fixed
+  depth_gate_fixed_value: 0.5
+  use_rays: false
+  use_bimanual_interaction: true
+  predict_presence: true
+  require_mano: true
+loss:
+  joint2d: 1.0
+  joint3d_camera: 10.0
+  vertices_camera: 1.0
+  local_pose: 1.0
+  global_orientation: 1.0
+  shape: 0.1
+  translation: 10.0
+  reprojection: 1.0
+  temporal_velocity: 0.1
+  temporal_acceleration: 0.05
+  bimanual_relative_wrist: 1.0
+  presence: 0.0
+assets:
+  dinov3_source: third_party/dinov3
+  dinov3_weights: /SSD3/jiasen/foundation_hand/pretrained/dinov3
+  mano_root: /SSD3/jiasen/foundation_hand/pretrained/mano
+precision: bf16
+protocol:
+  max_steps: 100000
+  num_workers: 4
+  checkpoint_interval: 25000
+  validation_interval: 5000
+  validation_batches: 256
+  learning_rate: 5.0e-05
+  min_learning_rate: 5.0e-06
+  warmup_steps: 500
+  early_stopping_patience: 0
+  validation_subset_manifest: artifacts/phase3/validation_subset/validation_subset_256.json

home/wuyifan/EgoNet_train_20261008/configs/phase3/core/core_m0_l0_t1_s20260717.yaml

@@ -0,0 +1,63 @@
+name: core_m0_l0_t1_s20260717
+seed: 20260717
+experiment:
+  phase: 3
+  matrix: decoder_temporal_core
+  modality: M0
+  layout: L0
+  temporal: T1
+  registered_axes:
+  - layout
+  - temporal_attention
+  - modality
+data:
+  resolution_hw:
+  - 288
+  - 512
+  T: 8
+  stride: 1
+  batch_size: 1
+model:
+  encoder: dinov3
+  decoder_dim: 512
+  decoder_layers: 3
+  attention_heads: 8
+  decoder_mode: sequential
+  use_temporal: true
+  use_depth: false
+  depth_mode: log_depth
+  depth_gate_mode: fixed
+  depth_gate_fixed_value: 0.5
+  use_rays: false
+  use_bimanual_interaction: true
+  predict_presence: true
+  require_mano: true
+loss:
+  joint2d: 1.0
+  joint3d_camera: 10.0
+  vertices_camera: 1.0
+  local_pose: 1.0
+  global_orientation: 1.0
+  shape: 0.1
+  translation: 10.0
+  reprojection: 1.0
+  temporal_velocity: 0.1
+  temporal_acceleration: 0.05
+  bimanual_relative_wrist: 1.0
+  presence: 0.0
+assets:
+  dinov3_source: third_party/dinov3
+  dinov3_weights: /SSD3/jiasen/foundation_hand/pretrained/dinov3
+  mano_root: /SSD3/jiasen/foundation_hand/pretrained/mano
+precision: bf16
+protocol:
+  max_steps: 100000
+  num_workers: 4
+  checkpoint_interval: 25000
+  validation_interval: 5000
+  validation_batches: 256
+  learning_rate: 5.0e-05
+  min_learning_rate: 5.0e-06
+  warmup_steps: 500
+  early_stopping_patience: 0
+  validation_subset_manifest: artifacts/phase3/validation_subset/validation_subset_256.json

home/wuyifan/EgoNet_train_20261008/configs/phase3/core/core_m0_l0_t1_s20260718.yaml

@@ -0,0 +1,63 @@
+name: core_m0_l0_t1_s20260718
+seed: 20260718
+experiment:
+  phase: 3
+  matrix: decoder_temporal_core
+  modality: M0
+  layout: L0
+  temporal: T1
+  registered_axes:
+  - layout
+  - temporal_attention
+  - modality
+data:
+  resolution_hw:
+  - 288
+  - 512
+  T: 8
+  stride: 1
+  batch_size: 1
+model:
+  encoder: dinov3
+  decoder_dim: 512
+  decoder_layers: 3
+  attention_heads: 8
+  decoder_mode: sequential
+  use_temporal: true
+  use_depth: false
+  depth_mode: log_depth
+  depth_gate_mode: fixed
+  depth_gate_fixed_value: 0.5
+  use_rays: false
+  use_bimanual_interaction: true
+  predict_presence: true
+  require_mano: true
+loss:
+  joint2d: 1.0
+  joint3d_camera: 10.0
+  vertices_camera: 1.0
+  local_pose: 1.0
+  global_orientation: 1.0
+  shape: 0.1
+  translation: 10.0
+  reprojection: 1.0
+  temporal_velocity: 0.1
+  temporal_acceleration: 0.05
+  bimanual_relative_wrist: 1.0
+  presence: 0.0
+assets:
+  dinov3_source: third_party/dinov3
+  dinov3_weights: /SSD3/jiasen/foundation_hand/pretrained/dinov3
+  mano_root: /SSD3/jiasen/foundation_hand/pretrained/mano
+precision: bf16
+protocol:
+  max_steps: 100000
+  num_workers: 4
+  checkpoint_interval: 25000
+  validation_interval: 5000
+  validation_batches: 256
+  learning_rate: 5.0e-05
+  min_learning_rate: 5.0e-06
+  warmup_steps: 500
+  early_stopping_patience: 0
+  validation_subset_manifest: artifacts/phase3/validation_subset/validation_subset_256.json

home/wuyifan/EgoNet_train_20261008/configs/phase3/core/core_m0_l1_t0_s20260716.yaml

@@ -0,0 +1,63 @@
+name: core_m0_l1_t0_s20260716
+seed: 20260716
+experiment:
+  phase: 3
+  matrix: decoder_temporal_core
+  modality: M0
+  layout: L1
+  temporal: T0
+  registered_axes:
+  - layout
+  - temporal_attention
+  - modality
+data:
+  resolution_hw:
+  - 288
+  - 512
+  T: 8
+  stride: 1
+  batch_size: 1
+model:
+  encoder: dinov3
+  decoder_dim: 512
+  decoder_layers: 3
+  attention_heads: 8
+  decoder_mode: alternating
+  use_temporal: false
+  use_depth: false
+  depth_mode: log_depth
+  depth_gate_mode: fixed
+  depth_gate_fixed_value: 0.5
+  use_rays: false
+  use_bimanual_interaction: true
+  predict_presence: true
+  require_mano: true
+loss:
+  joint2d: 1.0
+  joint3d_camera: 10.0
+  vertices_camera: 1.0
+  local_pose: 1.0
+  global_orientation: 1.0
+  shape: 0.1
+  translation: 10.0
+  reprojection: 1.0
+  temporal_velocity: 0.1
+  temporal_acceleration: 0.05
+  bimanual_relative_wrist: 1.0
+  presence: 0.0
+assets:
+  dinov3_source: third_party/dinov3
+  dinov3_weights: /SSD3/jiasen/foundation_hand/pretrained/dinov3
+  mano_root: /SSD3/jiasen/foundation_hand/pretrained/mano
+precision: bf16
+protocol:
+  max_steps: 100000
+  num_workers: 4
+  checkpoint_interval: 25000
+  validation_interval: 5000
+  validation_batches: 256
+  learning_rate: 5.0e-05
+  min_learning_rate: 5.0e-06
+  warmup_steps: 500
+  early_stopping_patience: 0
+  validation_subset_manifest: artifacts/phase3/validation_subset/validation_subset_256.json

home/wuyifan/EgoNet_train_20261008/configs/phase3/core/core_m0_l1_t0_s20260717.yaml

@@ -0,0 +1,63 @@
+name: core_m0_l1_t0_s20260717
+seed: 20260717
+experiment:
+  phase: 3
+  matrix: decoder_temporal_core
+  modality: M0
+  layout: L1
+  temporal: T0
+  registered_axes:
+  - layout
+  - temporal_attention
+  - modality
+data:
+  resolution_hw:
+  - 288
+  - 512
+  T: 8
+  stride: 1
+  batch_size: 1
+model:
+  encoder: dinov3
+  decoder_dim: 512
+  decoder_layers: 3
+  attention_heads: 8
+  decoder_mode: alternating
+  use_temporal: false
+  use_depth: false
+  depth_mode: log_depth
+  depth_gate_mode: fixed
+  depth_gate_fixed_value: 0.5
+  use_rays: false
+  use_bimanual_interaction: true
+  predict_presence: true
+  require_mano: true
+loss:
+  joint2d: 1.0
+  joint3d_camera: 10.0
+  vertices_camera: 1.0
+  local_pose: 1.0
+  global_orientation: 1.0
+  shape: 0.1
+  translation: 10.0
+  reprojection: 1.0
+  temporal_velocity: 0.1
+  temporal_acceleration: 0.05
+  bimanual_relative_wrist: 1.0
+  presence: 0.0
+assets:
+  dinov3_source: third_party/dinov3
+  dinov3_weights: /SSD3/jiasen/foundation_hand/pretrained/dinov3
+  mano_root: /SSD3/jiasen/foundation_hand/pretrained/mano
+precision: bf16
+protocol:
+  max_steps: 100000
+  num_workers: 4
+  checkpoint_interval: 25000
+  validation_interval: 5000
+  validation_batches: 256
+  learning_rate: 5.0e-05
+  min_learning_rate: 5.0e-06
+  warmup_steps: 500
+  early_stopping_patience: 0
+  validation_subset_manifest: artifacts/phase3/validation_subset/validation_subset_256.json

home/wuyifan/EgoNet_train_20261008/configs/phase3/core/core_m0_l1_t0_s20260718.yaml

@@ -0,0 +1,63 @@
+name: core_m0_l1_t0_s20260718
+seed: 20260718
+experiment:
+  phase: 3
+  matrix: decoder_temporal_core
+  modality: M0
+  layout: L1
+  temporal: T0
+  registered_axes:
+  - layout
+  - temporal_attention
+  - modality
+data:
+  resolution_hw:
+  - 288
+  - 512
+  T: 8
+  stride: 1
+  batch_size: 1
+model:
+  encoder: dinov3
+  decoder_dim: 512
+  decoder_layers: 3
+  attention_heads: 8
+  decoder_mode: alternating
+  use_temporal: false
+  use_depth: false
+  depth_mode: log_depth
+  depth_gate_mode: fixed
+  depth_gate_fixed_value: 0.5
+  use_rays: false
+  use_bimanual_interaction: true
+  predict_presence: true
+  require_mano: true
+loss:
+  joint2d: 1.0
+  joint3d_camera: 10.0
+  vertices_camera: 1.0
+  local_pose: 1.0
+  global_orientation: 1.0
+  shape: 0.1
+  translation: 10.0
+  reprojection: 1.0
+  temporal_velocity: 0.1
+  temporal_acceleration: 0.05
+  bimanual_relative_wrist: 1.0
+  presence: 0.0
+assets:
+  dinov3_source: third_party/dinov3
+  dinov3_weights: /SSD3/jiasen/foundation_hand/pretrained/dinov3
+  mano_root: /SSD3/jiasen/foundation_hand/pretrained/mano
+precision: bf16
+protocol:
+  max_steps: 100000
+  num_workers: 4
+  checkpoint_interval: 25000
+  validation_interval: 5000
+  validation_batches: 256
+  learning_rate: 5.0e-05
+  min_learning_rate: 5.0e-06
+  warmup_steps: 500
+  early_stopping_patience: 0
+  validation_subset_manifest: artifacts/phase3/validation_subset/validation_subset_256.json

home/wuyifan/EgoNet_train_20261008/configs/phase3/core/core_m0_l1_t1_s20260716.yaml

@@ -0,0 +1,63 @@
+name: core_m0_l1_t1_s20260716
+seed: 20260716
+experiment:
+  phase: 3
+  matrix: decoder_temporal_core
+  modality: M0
+  layout: L1
+  temporal: T1
+  registered_axes:
+  - layout
+  - temporal_attention
+  - modality
+data:
+  resolution_hw:
+  - 288
+  - 512
+  T: 8
+  stride: 1
+  batch_size: 1
+model:
+  encoder: dinov3
+  decoder_dim: 512
+  decoder_layers: 3
+  attention_heads: 8
+  decoder_mode: alternating
+  use_temporal: true
+  use_depth: false
+  depth_mode: log_depth
+  depth_gate_mode: fixed
+  depth_gate_fixed_value: 0.5
+  use_rays: false
+  use_bimanual_interaction: true
+  predict_presence: true
+  require_mano: true
+loss:
+  joint2d: 1.0
+  joint3d_camera: 10.0
+  vertices_camera: 1.0
+  local_pose: 1.0
+  global_orientation: 1.0
+  shape: 0.1
+  translation: 10.0
+  reprojection: 1.0
+  temporal_velocity: 0.1
+  temporal_acceleration: 0.05
+  bimanual_relative_wrist: 1.0
+  presence: 0.0
+assets:
+  dinov3_source: third_party/dinov3
+  dinov3_weights: /SSD3/jiasen/foundation_hand/pretrained/dinov3
+  mano_root: /SSD3/jiasen/foundation_hand/pretrained/mano
+precision: bf16
+protocol:
+  max_steps: 100000
+  num_workers: 4
+  checkpoint_interval: 25000
+  validation_interval: 5000
+  validation_batches: 256
+  learning_rate: 5.0e-05
+  min_learning_rate: 5.0e-06
+  warmup_steps: 500
+  early_stopping_patience: 0
+  validation_subset_manifest: artifacts/phase3/validation_subset/validation_subset_256.json

home/wuyifan/EgoNet_train_20261008/configs/phase3/core/core_m0_l1_t1_s20260717.yaml

@@ -0,0 +1,63 @@
+name: core_m0_l1_t1_s20260717
+seed: 20260717
+experiment:
+  phase: 3
+  matrix: decoder_temporal_core
+  modality: M0
+  layout: L1
+  temporal: T1
+  registered_axes:
+  - layout
+  - temporal_attention
+  - modality
+data:
+  resolution_hw:
+  - 288
+  - 512
+  T: 8
+  stride: 1
+  batch_size: 1
+model:
+  encoder: dinov3
+  decoder_dim: 512
+  decoder_layers: 3
+  attention_heads: 8
+  decoder_mode: alternating
+  use_temporal: true
+  use_depth: false
+  depth_mode: log_depth
+  depth_gate_mode: fixed
+  depth_gate_fixed_value: 0.5
+  use_rays: false
+  use_bimanual_interaction: true
+  predict_presence: true
+  require_mano: true
+loss:
+  joint2d: 1.0
+  joint3d_camera: 10.0
+  vertices_camera: 1.0
+  local_pose: 1.0
+  global_orientation: 1.0
+  shape: 0.1
+  translation: 10.0
+  reprojection: 1.0
+  temporal_velocity: 0.1
+  temporal_acceleration: 0.05
+  bimanual_relative_wrist: 1.0
+  presence: 0.0
+assets:
+  dinov3_source: third_party/dinov3
+  dinov3_weights: /SSD3/jiasen/foundation_hand/pretrained/dinov3
+  mano_root: /SSD3/jiasen/foundation_hand/pretrained/mano
+precision: bf16
+protocol:
+  max_steps: 100000
+  num_workers: 4
+  checkpoint_interval: 25000
+  validation_interval: 5000
+  validation_batches: 256
+  learning_rate: 5.0e-05
+  min_learning_rate: 5.0e-06
+  warmup_steps: 500
+  early_stopping_patience: 0
+  validation_subset_manifest: artifacts/phase3/validation_subset/validation_subset_256.json

home/wuyifan/EgoNet_train_20261008/configs/phase3/core/core_m0_l1_t1_s20260718.yaml

@@ -0,0 +1,63 @@
+name: core_m0_l1_t1_s20260718
+seed: 20260718
+experiment:
+  phase: 3
+  matrix: decoder_temporal_core
+  modality: M0
+  layout: L1
+  temporal: T1
+  registered_axes:
+  - layout
+  - temporal_attention
+  - modality
+data:
+  resolution_hw:
+  - 288
+  - 512
+  T: 8
+  stride: 1
+  batch_size: 1
+model:
+  encoder: dinov3
+  decoder_dim: 512
+  decoder_layers: 3
+  attention_heads: 8
+  decoder_mode: alternating
+  use_temporal: true
+  use_depth: false
+  depth_mode: log_depth
+  depth_gate_mode: fixed
+  depth_gate_fixed_value: 0.5
+  use_rays: false
+  use_bimanual_interaction: true
+  predict_presence: true
+  require_mano: true
+loss:
+  joint2d: 1.0
+  joint3d_camera: 10.0
+  vertices_camera: 1.0
+  local_pose: 1.0
+  global_orientation: 1.0
+  shape: 0.1
+  translation: 10.0
+  reprojection: 1.0
+  temporal_velocity: 0.1
+  temporal_acceleration: 0.05
+  bimanual_relative_wrist: 1.0
+  presence: 0.0
+assets:
+  dinov3_source: third_party/dinov3
+  dinov3_weights: /SSD3/jiasen/foundation_hand/pretrained/dinov3
+  mano_root: /SSD3/jiasen/foundation_hand/pretrained/mano
+precision: bf16
+protocol:
+  max_steps: 100000
+  num_workers: 4
+  checkpoint_interval: 25000
+  validation_interval: 5000
+  validation_batches: 256
+  learning_rate: 5.0e-05
+  min_learning_rate: 5.0e-06
+  warmup_steps: 500
+  early_stopping_patience: 0
+  validation_subset_manifest: artifacts/phase3/validation_subset/validation_subset_256.json

home/wuyifan/EgoNet_train_20261008/configs/phase3/core/core_m0_l2_t1_s20260716.yaml

@@ -0,0 +1,63 @@
+name: core_m0_l2_t1_s20260716
+seed: 20260716
+experiment:
+  phase: 3
+  matrix: decoder_temporal_core
+  modality: M0
+  layout: L2
+  temporal: T1
+  registered_axes:
+  - layout
+  - temporal_attention
+  - modality
+data:
+  resolution_hw:
+  - 288
+  - 512
+  T: 8
+  stride: 1
+  batch_size: 1
+model:
+  encoder: dinov3
+  decoder_dim: 512
+  decoder_layers: 3
+  attention_heads: 8
+  decoder_mode: hybrid_reread
+  use_temporal: true
+  use_depth: false
+  depth_mode: log_depth
+  depth_gate_mode: fixed
+  depth_gate_fixed_value: 0.5
+  use_rays: false
+  use_bimanual_interaction: true
+  predict_presence: true
+  require_mano: true
+loss:
+  joint2d: 1.0
+  joint3d_camera: 10.0
+  vertices_camera: 1.0
+  local_pose: 1.0
+  global_orientation: 1.0
+  shape: 0.1
+  translation: 10.0
+  reprojection: 1.0
+  temporal_velocity: 0.1
+  temporal_acceleration: 0.05
+  bimanual_relative_wrist: 1.0
+  presence: 0.0
+assets:
+  dinov3_source: third_party/dinov3
+  dinov3_weights: /SSD3/jiasen/foundation_hand/pretrained/dinov3
+  mano_root: /SSD3/jiasen/foundation_hand/pretrained/mano
+precision: bf16
+protocol:
+  max_steps: 100000
+  num_workers: 4
+  checkpoint_interval: 25000
+  validation_interval: 5000
+  validation_batches: 256
+  learning_rate: 5.0e-05
+  min_learning_rate: 5.0e-06
+  warmup_steps: 500
+  early_stopping_patience: 0
+  validation_subset_manifest: artifacts/phase3/validation_subset/validation_subset_256.json

home/wuyifan/EgoNet_train_20261008/configs/phase3/core/core_m0_l2_t1_s20260717.yaml

@@ -0,0 +1,63 @@
+name: core_m0_l2_t1_s20260717
+seed: 20260717
+experiment:
+  phase: 3
+  matrix: decoder_temporal_core
+  modality: M0
+  layout: L2
+  temporal: T1
+  registered_axes:
+  - layout
+  - temporal_attention
+  - modality
+data:
+  resolution_hw:
+  - 288
+  - 512
+  T: 8
+  stride: 1
+  batch_size: 1
+model:
+  encoder: dinov3
+  decoder_dim: 512
+  decoder_layers: 3
+  attention_heads: 8
+  decoder_mode: hybrid_reread
+  use_temporal: true
+  use_depth: false
+  depth_mode: log_depth
+  depth_gate_mode: fixed
+  depth_gate_fixed_value: 0.5
+  use_rays: false
+  use_bimanual_interaction: true
+  predict_presence: true
+  require_mano: true
+loss:
+  joint2d: 1.0
+  joint3d_camera: 10.0
+  vertices_camera: 1.0
+  local_pose: 1.0
+  global_orientation: 1.0
+  shape: 0.1
+  translation: 10.0
+  reprojection: 1.0
+  temporal_velocity: 0.1
+  temporal_acceleration: 0.05
+  bimanual_relative_wrist: 1.0
+  presence: 0.0
+assets:
+  dinov3_source: third_party/dinov3
+  dinov3_weights: /SSD3/jiasen/foundation_hand/pretrained/dinov3
+  mano_root: /SSD3/jiasen/foundation_hand/pretrained/mano
+precision: bf16
+protocol:
+  max_steps: 100000
+  num_workers: 4
+  checkpoint_interval: 25000
+  validation_interval: 5000
+  validation_batches: 256
+  learning_rate: 5.0e-05
+  min_learning_rate: 5.0e-06
+  warmup_steps: 500
+  early_stopping_patience: 0
+  validation_subset_manifest: artifacts/phase3/validation_subset/validation_subset_256.json

home/wuyifan/EgoNet_train_20261008/configs/phase3/core/core_m0_l2_t1_s20260718.yaml

@@ -0,0 +1,63 @@
+name: core_m0_l2_t1_s20260718
+seed: 20260718
+experiment:
+  phase: 3
+  matrix: decoder_temporal_core
+  modality: M0
+  layout: L2
+  temporal: T1
+  registered_axes:
+  - layout
+  - temporal_attention
+  - modality
+data:
+  resolution_hw:
+  - 288
+  - 512
+  T: 8
+  stride: 1
+  batch_size: 1
+model:
+  encoder: dinov3
+  decoder_dim: 512
+  decoder_layers: 3
+  attention_heads: 8
+  decoder_mode: hybrid_reread
+  use_temporal: true
+  use_depth: false
+  depth_mode: log_depth
+  depth_gate_mode: fixed
+  depth_gate_fixed_value: 0.5
+  use_rays: false
+  use_bimanual_interaction: true
+  predict_presence: true
+  require_mano: true
+loss:
+  joint2d: 1.0
+  joint3d_camera: 10.0
+  vertices_camera: 1.0
+  local_pose: 1.0
+  global_orientation: 1.0
+  shape: 0.1
+  translation: 10.0
+  reprojection: 1.0
+  temporal_velocity: 0.1
+  temporal_acceleration: 0.05
+  bimanual_relative_wrist: 1.0
+  presence: 0.0
+assets:
+  dinov3_source: third_party/dinov3
+  dinov3_weights: /SSD3/jiasen/foundation_hand/pretrained/dinov3
+  mano_root: /SSD3/jiasen/foundation_hand/pretrained/mano
+precision: bf16
+protocol:
+  max_steps: 100000
+  num_workers: 4
+  checkpoint_interval: 25000
+  validation_interval: 5000
+  validation_batches: 256
+  learning_rate: 5.0e-05
+  min_learning_rate: 5.0e-06
+  warmup_steps: 500
+  early_stopping_patience: 0
+  validation_subset_manifest: artifacts/phase3/validation_subset/validation_subset_256.json

home/wuyifan/EgoNet_train_20261008/configs/phase3/core/core_m3_l0_t0_s20260716.yaml

@@ -0,0 +1,63 @@
+name: core_m3_l0_t0_s20260716
+seed: 20260716
+experiment:
+  phase: 3
+  matrix: decoder_temporal_core
+  modality: M3
+  layout: L0
+  temporal: T0
+  registered_axes:
+  - layout
+  - temporal_attention
+  - modality
+data:
+  resolution_hw:
+  - 288
+  - 512
+  T: 8
+  stride: 1
+  batch_size: 1
+model:
+  encoder: dinov3
+  decoder_dim: 512
+  decoder_layers: 3
+  attention_heads: 8
+  decoder_mode: sequential
+  use_temporal: false
+  use_depth: true
+  depth_mode: log_depth
+  depth_gate_mode: fixed
+  depth_gate_fixed_value: 0.5
+  use_rays: true
+  use_bimanual_interaction: true
+  predict_presence: true
+  require_mano: true
+loss:
+  joint2d: 1.0
+  joint3d_camera: 10.0
+  vertices_camera: 1.0
+  local_pose: 1.0
+  global_orientation: 1.0
+  shape: 0.1
+  translation: 10.0
+  reprojection: 1.0
+  temporal_velocity: 0.1
+  temporal_acceleration: 0.05
+  bimanual_relative_wrist: 1.0
+  presence: 0.0
+assets:
+  dinov3_source: third_party/dinov3
+  dinov3_weights: /SSD3/jiasen/foundation_hand/pretrained/dinov3
+  mano_root: /SSD3/jiasen/foundation_hand/pretrained/mano
+precision: bf16
+protocol:
+  max_steps: 100000
+  num_workers: 4
+  checkpoint_interval: 25000
+  validation_interval: 5000
+  validation_batches: 256
+  learning_rate: 5.0e-05
+  min_learning_rate: 5.0e-06
+  warmup_steps: 500
+  early_stopping_patience: 0
+  validation_subset_manifest: artifacts/phase3/validation_subset/validation_subset_256.json

home/wuyifan/EgoNet_train_20261008/configs/phase3/core/core_m3_l0_t0_s20260717.yaml

@@ -0,0 +1,63 @@
+name: core_m3_l0_t0_s20260717
+seed: 20260717
+experiment:
+  phase: 3
+  matrix: decoder_temporal_core
+  modality: M3
+  layout: L0
+  temporal: T0
+  registered_axes:
+  - layout
+  - temporal_attention
+  - modality
+data:
+  resolution_hw:
+  - 288
+  - 512
+  T: 8
+  stride: 1
+  batch_size: 1
+model:
+  encoder: dinov3
+  decoder_dim: 512
+  decoder_layers: 3
+  attention_heads: 8
+  decoder_mode: sequential
+  use_temporal: false
+  use_depth: true
+  depth_mode: log_depth
+  depth_gate_mode: fixed
+  depth_gate_fixed_value: 0.5
+  use_rays: true
+  use_bimanual_interaction: true
+  predict_presence: true
+  require_mano: true
+loss:
+  joint2d: 1.0
+  joint3d_camera: 10.0
+  vertices_camera: 1.0
+  local_pose: 1.0
+  global_orientation: 1.0
+  shape: 0.1
+  translation: 10.0
+  reprojection: 1.0
+  temporal_velocity: 0.1
+  temporal_acceleration: 0.05
+  bimanual_relative_wrist: 1.0
+  presence: 0.0
+assets:
+  dinov3_source: third_party/dinov3
+  dinov3_weights: /SSD3/jiasen/foundation_hand/pretrained/dinov3
+  mano_root: /SSD3/jiasen/foundation_hand/pretrained/mano
+precision: bf16
+protocol:
+  max_steps: 100000
+  num_workers: 4
+  checkpoint_interval: 25000
+  validation_interval: 5000
+  validation_batches: 256
+  learning_rate: 5.0e-05
+  min_learning_rate: 5.0e-06
+  warmup_steps: 500
+  early_stopping_patience: 0
+  validation_subset_manifest: artifacts/phase3/validation_subset/validation_subset_256.json

home/wuyifan/EgoNet_train_20261008/configs/phase3/core/core_m3_l0_t0_s20260718.yaml

@@ -0,0 +1,63 @@
+name: core_m3_l0_t0_s20260718
+seed: 20260718
+experiment:
+  phase: 3
+  matrix: decoder_temporal_core
+  modality: M3
+  layout: L0
+  temporal: T0
+  registered_axes:
+  - layout
+  - temporal_attention
+  - modality
+data:
+  resolution_hw:
+  - 288
+  - 512
+  T: 8
+  stride: 1
+  batch_size: 1
+model:
+  encoder: dinov3
+  decoder_dim: 512
+  decoder_layers: 3
+  attention_heads: 8
+  decoder_mode: sequential
+  use_temporal: false
+  use_depth: true
+  depth_mode: log_depth
+  depth_gate_mode: fixed
+  depth_gate_fixed_value: 0.5
+  use_rays: true
+  use_bimanual_interaction: true
+  predict_presence: true
+  require_mano: true
+loss:
+  joint2d: 1.0
+  joint3d_camera: 10.0
+  vertices_camera: 1.0
+  local_pose: 1.0
+  global_orientation: 1.0
+  shape: 0.1
+  translation: 10.0
+  reprojection: 1.0
+  temporal_velocity: 0.1
+  temporal_acceleration: 0.05
+  bimanual_relative_wrist: 1.0
+  presence: 0.0
+assets:
+  dinov3_source: third_party/dinov3
+  dinov3_weights: /SSD3/jiasen/foundation_hand/pretrained/dinov3
+  mano_root: /SSD3/jiasen/foundation_hand/pretrained/mano
+precision: bf16
+protocol:
+  max_steps: 100000
+  num_workers: 4
+  checkpoint_interval: 25000
+  validation_interval: 5000
+  validation_batches: 256
+  learning_rate: 5.0e-05
+  min_learning_rate: 5.0e-06
+  warmup_steps: 500
+  early_stopping_patience: 0
+  validation_subset_manifest: artifacts/phase3/validation_subset/validation_subset_256.json

home/wuyifan/EgoNet_train_20261008/configs/phase3/core/core_m3_l0_t1_s20260716.yaml

@@ -0,0 +1,63 @@
+name: core_m3_l0_t1_s20260716
+seed: 20260716
+experiment:
+  phase: 3
+  matrix: decoder_temporal_core
+  modality: M3
+  layout: L0
+  temporal: T1
+  registered_axes:
+  - layout
+  - temporal_attention
+  - modality
+data:
+  resolution_hw:
+  - 288
+  - 512
+  T: 8
+  stride: 1
+  batch_size: 1
+model:
+  encoder: dinov3
+  decoder_dim: 512
+  decoder_layers: 3
+  attention_heads: 8
+  decoder_mode: sequential
+  use_temporal: true
+  use_depth: true
+  depth_mode: log_depth
+  depth_gate_mode: fixed
+  depth_gate_fixed_value: 0.5
+  use_rays: true
+  use_bimanual_interaction: true
+  predict_presence: true
+  require_mano: true
+loss:
+  joint2d: 1.0
+  joint3d_camera: 10.0
+  vertices_camera: 1.0
+  local_pose: 1.0
+  global_orientation: 1.0
+  shape: 0.1
+  translation: 10.0
+  reprojection: 1.0
+  temporal_velocity: 0.1
+  temporal_acceleration: 0.05
+  bimanual_relative_wrist: 1.0
+  presence: 0.0
+assets:
+  dinov3_source: third_party/dinov3
+  dinov3_weights: /SSD3/jiasen/foundation_hand/pretrained/dinov3
+  mano_root: /SSD3/jiasen/foundation_hand/pretrained/mano
+precision: bf16
+protocol:
+  max_steps: 100000
+  num_workers: 4
+  checkpoint_interval: 25000
+  validation_interval: 5000
+  validation_batches: 256
+  learning_rate: 5.0e-05
+  min_learning_rate: 5.0e-06
+  warmup_steps: 500
+  early_stopping_patience: 0
+  validation_subset_manifest: artifacts/phase3/validation_subset/validation_subset_256.json

home/wuyifan/EgoNet_train_20261008/configs/phase3/core/core_m3_l0_t1_s20260717.yaml

@@ -0,0 +1,63 @@
+name: core_m3_l0_t1_s20260717
+seed: 20260717
+experiment:
+  phase: 3
+  matrix: decoder_temporal_core
+  modality: M3
+  layout: L0
+  temporal: T1
+  registered_axes:
+  - layout
+  - temporal_attention
+  - modality
+data:
+  resolution_hw:
+  - 288
+  - 512
+  T: 8
+  stride: 1
+  batch_size: 1
+model:
+  encoder: dinov3
+  decoder_dim: 512
+  decoder_layers: 3
+  attention_heads: 8
+  decoder_mode: sequential
+  use_temporal: true
+  use_depth: true
+  depth_mode: log_depth
+  depth_gate_mode: fixed
+  depth_gate_fixed_value: 0.5
+  use_rays: true
+  use_bimanual_interaction: true
+  predict_presence: true
+  require_mano: true
+loss:
+  joint2d: 1.0
+  joint3d_camera: 10.0
+  vertices_camera: 1.0
+  local_pose: 1.0
+  global_orientation: 1.0
+  shape: 0.1
+  translation: 10.0
+  reprojection: 1.0
+  temporal_velocity: 0.1
+  temporal_acceleration: 0.05
+  bimanual_relative_wrist: 1.0
+  presence: 0.0
+assets:
+  dinov3_source: third_party/dinov3
+  dinov3_weights: /SSD3/jiasen/foundation_hand/pretrained/dinov3
+  mano_root: /SSD3/jiasen/foundation_hand/pretrained/mano
+precision: bf16
+protocol:
+  max_steps: 100000
+  num_workers: 4
+  checkpoint_interval: 25000
+  validation_interval: 5000
+  validation_batches: 256
+  learning_rate: 5.0e-05
+  min_learning_rate: 5.0e-06
+  warmup_steps: 500
+  early_stopping_patience: 0
+  validation_subset_manifest: artifacts/phase3/validation_subset/validation_subset_256.json

home/wuyifan/EgoNet_train_20261008/configs/phase3/core/core_m3_l0_t1_s20260718.yaml

@@ -0,0 +1,63 @@
+name: core_m3_l0_t1_s20260718
+seed: 20260718
+experiment:
+  phase: 3
+  matrix: decoder_temporal_core
+  modality: M3
+  layout: L0
+  temporal: T1
+  registered_axes:
+  - layout
+  - temporal_attention
+  - modality
+data:
+  resolution_hw:
+  - 288
+  - 512
+  T: 8
+  stride: 1
+  batch_size: 1
+model:
+  encoder: dinov3
+  decoder_dim: 512
+  decoder_layers: 3
+  attention_heads: 8
+  decoder_mode: sequential
+  use_temporal: true
+  use_depth: true
+  depth_mode: log_depth
+  depth_gate_mode: fixed
+  depth_gate_fixed_value: 0.5
+  use_rays: true
+  use_bimanual_interaction: true
+  predict_presence: true
+  require_mano: true
+loss:
+  joint2d: 1.0
+  joint3d_camera: 10.0
+  vertices_camera: 1.0
+  local_pose: 1.0
+  global_orientation: 1.0
+  shape: 0.1
+  translation: 10.0
+  reprojection: 1.0
+  temporal_velocity: 0.1
+  temporal_acceleration: 0.05
+  bimanual_relative_wrist: 1.0
+  presence: 0.0
+assets:
+  dinov3_source: third_party/dinov3
+  dinov3_weights: /SSD3/jiasen/foundation_hand/pretrained/dinov3
+  mano_root: /SSD3/jiasen/foundation_hand/pretrained/mano
+precision: bf16
+protocol:
+  max_steps: 100000
+  num_workers: 4
+  checkpoint_interval: 25000
+  validation_interval: 5000
+  validation_batches: 256
+  learning_rate: 5.0e-05
+  min_learning_rate: 5.0e-06
+  warmup_steps: 500
+  early_stopping_patience: 0
+  validation_subset_manifest: artifacts/phase3/validation_subset/validation_subset_256.json

home/wuyifan/EgoNet_train_20261008/configs/phase3/core/core_m3_l1_t0_s20260716.yaml

@@ -0,0 +1,63 @@
+name: core_m3_l1_t0_s20260716
+seed: 20260716
+experiment:
+  phase: 3
+  matrix: decoder_temporal_core
+  modality: M3
+  layout: L1
+  temporal: T0
+  registered_axes:
+  - layout
+  - temporal_attention
+  - modality
+data:
+  resolution_hw:
+  - 288
+  - 512
+  T: 8
+  stride: 1
+  batch_size: 1
+model:
+  encoder: dinov3
+  decoder_dim: 512
+  decoder_layers: 3
+  attention_heads: 8
+  decoder_mode: alternating
+  use_temporal: false
+  use_depth: true
+  depth_mode: log_depth
+  depth_gate_mode: fixed
+  depth_gate_fixed_value: 0.5
+  use_rays: true
+  use_bimanual_interaction: true
+  predict_presence: true
+  require_mano: true
+loss:
+  joint2d: 1.0
+  joint3d_camera: 10.0
+  vertices_camera: 1.0
+  local_pose: 1.0
+  global_orientation: 1.0
+  shape: 0.1
+  translation: 10.0
+  reprojection: 1.0
+  temporal_velocity: 0.1
+  temporal_acceleration: 0.05
+  bimanual_relative_wrist: 1.0
+  presence: 0.0
+assets:
+  dinov3_source: third_party/dinov3
+  dinov3_weights: /SSD3/jiasen/foundation_hand/pretrained/dinov3
+  mano_root: /SSD3/jiasen/foundation_hand/pretrained/mano
+precision: bf16
+protocol:
+  max_steps: 100000
+  num_workers: 4
+  checkpoint_interval: 25000
+  validation_interval: 5000
+  validation_batches: 256
+  learning_rate: 5.0e-05
+  min_learning_rate: 5.0e-06
+  warmup_steps: 500
+  early_stopping_patience: 0
+  validation_subset_manifest: artifacts/phase3/validation_subset/validation_subset_256.json

home/wuyifan/EgoNet_train_20261008/configs/phase3/core/core_m3_l1_t0_s20260717.yaml

@@ -0,0 +1,63 @@
+name: core_m3_l1_t0_s20260717
+seed: 20260717
+experiment:
+  phase: 3
+  matrix: decoder_temporal_core
+  modality: M3
+  layout: L1
+  temporal: T0
+  registered_axes:
+  - layout
+  - temporal_attention
+  - modality
+data:
+  resolution_hw:
+  - 288
+  - 512
+  T: 8
+  stride: 1
+  batch_size: 1
+model:
+  encoder: dinov3
+  decoder_dim: 512
+  decoder_layers: 3
+  attention_heads: 8
+  decoder_mode: alternating
+  use_temporal: false
+  use_depth: true
+  depth_mode: log_depth
+  depth_gate_mode: fixed
+  depth_gate_fixed_value: 0.5
+  use_rays: true
+  use_bimanual_interaction: true
+  predict_presence: true
+  require_mano: true
+loss:
+  joint2d: 1.0
+  joint3d_camera: 10.0
+  vertices_camera: 1.0
+  local_pose: 1.0
+  global_orientation: 1.0
+  shape: 0.1
+  translation: 10.0
+  reprojection: 1.0
+  temporal_velocity: 0.1
+  temporal_acceleration: 0.05
+  bimanual_relative_wrist: 1.0
+  presence: 0.0
+assets:
+  dinov3_source: third_party/dinov3
+  dinov3_weights: /SSD3/jiasen/foundation_hand/pretrained/dinov3
+  mano_root: /SSD3/jiasen/foundation_hand/pretrained/mano
+precision: bf16
+protocol:
+  max_steps: 100000
+  num_workers: 4
+  checkpoint_interval: 25000
+  validation_interval: 5000
+  validation_batches: 256
+  learning_rate: 5.0e-05
+  min_learning_rate: 5.0e-06
+  warmup_steps: 500
+  early_stopping_patience: 0
+  validation_subset_manifest: artifacts/phase3/validation_subset/validation_subset_256.json

home/wuyifan/EgoNet_train_20261008/configs/phase3/core/core_m3_l1_t0_s20260718.yaml

@@ -0,0 +1,63 @@
+name: core_m3_l1_t0_s20260718
+seed: 20260718
+experiment:
+  phase: 3
+  matrix: decoder_temporal_core
+  modality: M3
+  layout: L1
+  temporal: T0
+  registered_axes:
+  - layout
+  - temporal_attention
+  - modality
+data:
+  resolution_hw:
+  - 288
+  - 512
+  T: 8
+  stride: 1
+  batch_size: 1
+model:
+  encoder: dinov3
+  decoder_dim: 512
+  decoder_layers: 3
+  attention_heads: 8
+  decoder_mode: alternating
+  use_temporal: false
+  use_depth: true
+  depth_mode: log_depth
+  depth_gate_mode: fixed
+  depth_gate_fixed_value: 0.5
+  use_rays: true
+  use_bimanual_interaction: true
+  predict_presence: true
+  require_mano: true
+loss:
+  joint2d: 1.0
+  joint3d_camera: 10.0
+  vertices_camera: 1.0
+  local_pose: 1.0
+  global_orientation: 1.0
+  shape: 0.1
+  translation: 10.0
+  reprojection: 1.0
+  temporal_velocity: 0.1
+  temporal_acceleration: 0.05
+  bimanual_relative_wrist: 1.0
+  presence: 0.0
+assets:
+  dinov3_source: third_party/dinov3
+  dinov3_weights: /SSD3/jiasen/foundation_hand/pretrained/dinov3
+  mano_root: /SSD3/jiasen/foundation_hand/pretrained/mano
+precision: bf16
+protocol:
+  max_steps: 100000
+  num_workers: 4
+  checkpoint_interval: 25000
+  validation_interval: 5000
+  validation_batches: 256
+  learning_rate: 5.0e-05
+  min_learning_rate: 5.0e-06
+  warmup_steps: 500
+  early_stopping_patience: 0
+  validation_subset_manifest: artifacts/phase3/validation_subset/validation_subset_256.json

home/wuyifan/EgoNet_train_20261008/configs/phase3/core/core_m3_l1_t1_s20260716.yaml

@@ -0,0 +1,63 @@
+name: core_m3_l1_t1_s20260716
+seed: 20260716
+experiment:
+  phase: 3
+  matrix: decoder_temporal_core
+  modality: M3
+  layout: L1
+  temporal: T1
+  registered_axes:
+  - layout
+  - temporal_attention
+  - modality
+data:
+  resolution_hw:
+  - 288
+  - 512
+  T: 8
+  stride: 1
+  batch_size: 1
+model:
+  encoder: dinov3
+  decoder_dim: 512
+  decoder_layers: 3
+  attention_heads: 8
+  decoder_mode: alternating
+  use_temporal: true
+  use_depth: true
+  depth_mode: log_depth
+  depth_gate_mode: fixed
+  depth_gate_fixed_value: 0.5
+  use_rays: true
+  use_bimanual_interaction: true
+  predict_presence: true
+  require_mano: true
+loss:
+  joint2d: 1.0
+  joint3d_camera: 10.0
+  vertices_camera: 1.0
+  local_pose: 1.0
+  global_orientation: 1.0
+  shape: 0.1
+  translation: 10.0
+  reprojection: 1.0
+  temporal_velocity: 0.1
+  temporal_acceleration: 0.05
+  bimanual_relative_wrist: 1.0
+  presence: 0.0
+assets:
+  dinov3_source: third_party/dinov3
+  dinov3_weights: /SSD3/jiasen/foundation_hand/pretrained/dinov3
+  mano_root: /SSD3/jiasen/foundation_hand/pretrained/mano
+precision: bf16
+protocol:
+  max_steps: 100000
+  num_workers: 4
+  checkpoint_interval: 25000
+  validation_interval: 5000
+  validation_batches: 256
+  learning_rate: 5.0e-05
+  min_learning_rate: 5.0e-06
+  warmup_steps: 500
+  early_stopping_patience: 0
+  validation_subset_manifest: artifacts/phase3/validation_subset/validation_subset_256.json

home/wuyifan/EgoNet_train_20261008/configs/phase3/core/core_m3_l1_t1_s20260717.yaml

@@ -0,0 +1,63 @@
+name: core_m3_l1_t1_s20260717
+seed: 20260717
+experiment:
+  phase: 3
+  matrix: decoder_temporal_core
+  modality: M3
+  layout: L1
+  temporal: T1
+  registered_axes:
+  - layout
+  - temporal_attention
+  - modality
+data:
+  resolution_hw:
+  - 288
+  - 512
+  T: 8
+  stride: 1
+  batch_size: 1
+model:
+  encoder: dinov3
+  decoder_dim: 512
+  decoder_layers: 3
+  attention_heads: 8
+  decoder_mode: alternating
+  use_temporal: true
+  use_depth: true
+  depth_mode: log_depth
+  depth_gate_mode: fixed
+  depth_gate_fixed_value: 0.5
+  use_rays: true
+  use_bimanual_interaction: true
+  predict_presence: true
+  require_mano: true
+loss:
+  joint2d: 1.0
+  joint3d_camera: 10.0
+  vertices_camera: 1.0
+  local_pose: 1.0
+  global_orientation: 1.0
+  shape: 0.1
+  translation: 10.0
+  reprojection: 1.0
+  temporal_velocity: 0.1
+  temporal_acceleration: 0.05
+  bimanual_relative_wrist: 1.0
+  presence: 0.0
+assets:
+  dinov3_source: third_party/dinov3
+  dinov3_weights: /SSD3/jiasen/foundation_hand/pretrained/dinov3
+  mano_root: /SSD3/jiasen/foundation_hand/pretrained/mano
+precision: bf16
+protocol:
+  max_steps: 100000
+  num_workers: 4
+  checkpoint_interval: 25000
+  validation_interval: 5000
+  validation_batches: 256
+  learning_rate: 5.0e-05
+  min_learning_rate: 5.0e-06
+  warmup_steps: 500
+  early_stopping_patience: 0
+  validation_subset_manifest: artifacts/phase3/validation_subset/validation_subset_256.json

home/wuyifan/EgoNet_train_20261008/configs/phase3/core/core_m3_l1_t1_s20260718.yaml

@@ -0,0 +1,63 @@
+name: core_m3_l1_t1_s20260718
+seed: 20260718
+experiment:
+  phase: 3
+  matrix: decoder_temporal_core
+  modality: M3
+  layout: L1
+  temporal: T1
+  registered_axes:
+  - layout
+  - temporal_attention
+  - modality
+data:
+  resolution_hw:
+  - 288
+  - 512
+  T: 8
+  stride: 1
+  batch_size: 1
+model:
+  encoder: dinov3
+  decoder_dim: 512
+  decoder_layers: 3
+  attention_heads: 8
+  decoder_mode: alternating
+  use_temporal: true
+  use_depth: true
+  depth_mode: log_depth
+  depth_gate_mode: fixed
+  depth_gate_fixed_value: 0.5
+  use_rays: true
+  use_bimanual_interaction: true
+  predict_presence: true
+  require_mano: true
+loss:
+  joint2d: 1.0
+  joint3d_camera: 10.0
+  vertices_camera: 1.0
+  local_pose: 1.0
+  global_orientation: 1.0
+  shape: 0.1
+  translation: 10.0
+  reprojection: 1.0
+  temporal_velocity: 0.1
+  temporal_acceleration: 0.05
+  bimanual_relative_wrist: 1.0
+  presence: 0.0
+assets:
+  dinov3_source: third_party/dinov3
+  dinov3_weights: /SSD3/jiasen/foundation_hand/pretrained/dinov3
+  mano_root: /SSD3/jiasen/foundation_hand/pretrained/mano
+precision: bf16
+protocol:
+  max_steps: 100000
+  num_workers: 4
+  checkpoint_interval: 25000
+  validation_interval: 5000
+  validation_batches: 256
+  learning_rate: 5.0e-05
+  min_learning_rate: 5.0e-06
+  warmup_steps: 500
+  early_stopping_patience: 0
+  validation_subset_manifest: artifacts/phase3/validation_subset/validation_subset_256.json

home/wuyifan/EgoNet_train_20261008/configs/phase3/core/core_m3_l2_t1_s20260716.yaml

@@ -0,0 +1,63 @@
+name: core_m3_l2_t1_s20260716
+seed: 20260716
+experiment:
+  phase: 3
+  matrix: decoder_temporal_core
+  modality: M3
+  layout: L2
+  temporal: T1
+  registered_axes:
+  - layout
+  - temporal_attention
+  - modality
+data:
+  resolution_hw:
+  - 288
+  - 512
+  T: 8
+  stride: 1
+  batch_size: 1
+model:
+  encoder: dinov3
+  decoder_dim: 512
+  decoder_layers: 3
+  attention_heads: 8
+  decoder_mode: hybrid_reread
+  use_temporal: true
+  use_depth: true
+  depth_mode: log_depth
+  depth_gate_mode: fixed
+  depth_gate_fixed_value: 0.5
+  use_rays: true
+  use_bimanual_interaction: true
+  predict_presence: true
+  require_mano: true
+loss:
+  joint2d: 1.0
+  joint3d_camera: 10.0
+  vertices_camera: 1.0
+  local_pose: 1.0
+  global_orientation: 1.0
+  shape: 0.1
+  translation: 10.0
+  reprojection: 1.0
+  temporal_velocity: 0.1
+  temporal_acceleration: 0.05
+  bimanual_relative_wrist: 1.0
+  presence: 0.0
+assets:
+  dinov3_source: third_party/dinov3
+  dinov3_weights: /SSD3/jiasen/foundation_hand/pretrained/dinov3
+  mano_root: /SSD3/jiasen/foundation_hand/pretrained/mano
+precision: bf16
+protocol:
+  max_steps: 100000
+  num_workers: 4
+  checkpoint_interval: 25000
+  validation_interval: 5000
+  validation_batches: 256
+  learning_rate: 5.0e-05
+  min_learning_rate: 5.0e-06
+  warmup_steps: 500
+  early_stopping_patience: 0
+  validation_subset_manifest: artifacts/phase3/validation_subset/validation_subset_256.json

home/wuyifan/EgoNet_train_20261008/configs/phase3/core/core_m3_l2_t1_s20260717.yaml

@@ -0,0 +1,63 @@
+name: core_m3_l2_t1_s20260717
+seed: 20260717
+experiment:
+  phase: 3
+  matrix: decoder_temporal_core
+  modality: M3
+  layout: L2
+  temporal: T1
+  registered_axes:
+  - layout
+  - temporal_attention
+  - modality
+data:
+  resolution_hw:
+  - 288
+  - 512
+  T: 8
+  stride: 1
+  batch_size: 1
+model:
+  encoder: dinov3
+  decoder_dim: 512
+  decoder_layers: 3
+  attention_heads: 8
+  decoder_mode: hybrid_reread
+  use_temporal: true
+  use_depth: true
+  depth_mode: log_depth
+  depth_gate_mode: fixed
+  depth_gate_fixed_value: 0.5
+  use_rays: true
+  use_bimanual_interaction: true
+  predict_presence: true
+  require_mano: true
+loss:
+  joint2d: 1.0
+  joint3d_camera: 10.0
+  vertices_camera: 1.0
+  local_pose: 1.0
+  global_orientation: 1.0
+  shape: 0.1
+  translation: 10.0
+  reprojection: 1.0
+  temporal_velocity: 0.1
+  temporal_acceleration: 0.05
+  bimanual_relative_wrist: 1.0
+  presence: 0.0
+assets:
+  dinov3_source: third_party/dinov3
+  dinov3_weights: /SSD3/jiasen/foundation_hand/pretrained/dinov3
+  mano_root: /SSD3/jiasen/foundation_hand/pretrained/mano
+precision: bf16
+protocol:
+  max_steps: 100000
+  num_workers: 4
+  checkpoint_interval: 25000
+  validation_interval: 5000
+  validation_batches: 256
+  learning_rate: 5.0e-05
+  min_learning_rate: 5.0e-06
+  warmup_steps: 500
+  early_stopping_patience: 0
+  validation_subset_manifest: artifacts/phase3/validation_subset/validation_subset_256.json

home/wuyifan/EgoNet_train_20261008/configs/phase3/core/core_m3_l2_t1_s20260718.yaml

@@ -0,0 +1,63 @@
+name: core_m3_l2_t1_s20260718
+seed: 20260718
+experiment:
+  phase: 3
+  matrix: decoder_temporal_core
+  modality: M3
+  layout: L2
+  temporal: T1
+  registered_axes:
+  - layout
+  - temporal_attention
+  - modality
+data:
+  resolution_hw:
+  - 288
+  - 512
+  T: 8
+  stride: 1
+  batch_size: 1
+model:
+  encoder: dinov3
+  decoder_dim: 512
+  decoder_layers: 3
+  attention_heads: 8
+  decoder_mode: hybrid_reread
+  use_temporal: true
+  use_depth: true
+  depth_mode: log_depth
+  depth_gate_mode: fixed
+  depth_gate_fixed_value: 0.5
+  use_rays: true
+  use_bimanual_interaction: true
+  predict_presence: true
+  require_mano: true
+loss:
+  joint2d: 1.0
+  joint3d_camera: 10.0
+  vertices_camera: 1.0
+  local_pose: 1.0
+  global_orientation: 1.0
+  shape: 0.1
+  translation: 10.0
+  reprojection: 1.0
+  temporal_velocity: 0.1
+  temporal_acceleration: 0.05
+  bimanual_relative_wrist: 1.0
+  presence: 0.0
+assets:
+  dinov3_source: third_party/dinov3
+  dinov3_weights: /SSD3/jiasen/foundation_hand/pretrained/dinov3
+  mano_root: /SSD3/jiasen/foundation_hand/pretrained/mano
+precision: bf16
+protocol:
+  max_steps: 100000
+  num_workers: 4
+  checkpoint_interval: 25000
+  validation_interval: 5000
+  validation_batches: 256
+  learning_rate: 5.0e-05
+  min_learning_rate: 5.0e-06
+  warmup_steps: 500
+  early_stopping_patience: 0
+  validation_subset_manifest: artifacts/phase3/validation_subset/validation_subset_256.json

home/wuyifan/EgoNet_train_20261008/configs/phase3/depth_fusion/depth_dropout005_geometry_m3_l1_t1_s20260716.yaml

@@ -0,0 +1,65 @@
+name: depth_dropout005_geometry_m3_l1_t1_s20260716
+seed: 20260716
+experiment:
+  phase: 3
+  matrix: depth_modality_dropout_dose
+  modality: M3
+  layout: L1
+  temporal: T1
+  variant: dropout005_geometry
+  registered_axes:
+  - modality_dropout_dose
+  reference: D2_dropout00_and_dropout02_endpoints
+data:
+  resolution_hw:
+  - 288
+  - 512
+  T: 8
+  stride: 1
+  batch_size: 1
+model:
+  encoder: dinov3
+  decoder_dim: 512
+  decoder_layers: 3
+  attention_heads: 8
+  decoder_mode: alternating
+  use_temporal: true
+  use_depth: true
+  depth_mode: geometry
+  depth_gate_mode: fixed
+  depth_gate_fixed_value: 0.5
+  use_rays: true
+  use_bimanual_interaction: true
+  predict_presence: true
+  require_mano: true
+  depth_gate_floor: 0.0
+  modality_dropout_prob: 0.05
+loss:
+  joint2d: 1.0
+  joint3d_camera: 10.0
+  vertices_camera: 1.0
+  local_pose: 1.0
+  global_orientation: 1.0
+  shape: 0.1
+  translation: 10.0
+  reprojection: 1.0
+  temporal_velocity: 0.1
+  temporal_acceleration: 0.05
+  bimanual_relative_wrist: 1.0
+  presence: 0.0
+assets:
+  dinov3_source: third_party/dinov3
+  dinov3_weights: /SSD3/jiasen/foundation_hand/pretrained/dinov3
+  mano_root: /SSD3/jiasen/foundation_hand/pretrained/mano
+precision: bf16
+protocol:
+  max_steps: 100000
+  num_workers: 4
+  checkpoint_interval: 25000
+  validation_interval: 5000
+  validation_batches: 256
+  learning_rate: 5.0e-05
+  min_learning_rate: 5.0e-06
+  warmup_steps: 500
+  early_stopping_patience: 0
+  validation_subset_manifest: artifacts/phase3/validation_subset/validation_subset_256.json

home/wuyifan/EgoNet_train_20261008/configs/phase3/depth_fusion/depth_dropout005_geometry_m3_l1_t1_s20260717.yaml

@@ -0,0 +1,65 @@
+name: depth_dropout005_geometry_m3_l1_t1_s20260717
+seed: 20260717
+experiment:
+  phase: 3
+  matrix: depth_modality_dropout_dose
+  modality: M3
+  layout: L1
+  temporal: T1
+  variant: dropout005_geometry
+  registered_axes:
+  - modality_dropout_dose
+  reference: D2_dropout00_and_dropout02_endpoints
+data:
+  resolution_hw:
+  - 288
+  - 512
+  T: 8
+  stride: 1
+  batch_size: 1
+model:
+  encoder: dinov3
+  decoder_dim: 512
+  decoder_layers: 3
+  attention_heads: 8
+  decoder_mode: alternating
+  use_temporal: true
+  use_depth: true
+  depth_mode: geometry
+  depth_gate_mode: fixed
+  depth_gate_fixed_value: 0.5
+  use_rays: true
+  use_bimanual_interaction: true
+  predict_presence: true
+  require_mano: true
+  depth_gate_floor: 0.0
+  modality_dropout_prob: 0.05
+loss:
+  joint2d: 1.0
+  joint3d_camera: 10.0
+  vertices_camera: 1.0
+  local_pose: 1.0
+  global_orientation: 1.0
+  shape: 0.1
+  translation: 10.0
+  reprojection: 1.0
+  temporal_velocity: 0.1
+  temporal_acceleration: 0.05
+  bimanual_relative_wrist: 1.0
+  presence: 0.0
+assets:
+  dinov3_source: third_party/dinov3
+  dinov3_weights: /SSD3/jiasen/foundation_hand/pretrained/dinov3
+  mano_root: /SSD3/jiasen/foundation_hand/pretrained/mano
+precision: bf16
+protocol:
+  max_steps: 100000
+  num_workers: 4
+  checkpoint_interval: 25000
+  validation_interval: 5000
+  validation_batches: 256
+  learning_rate: 5.0e-05
+  min_learning_rate: 5.0e-06
+  warmup_steps: 500
+  early_stopping_patience: 0
+  validation_subset_manifest: artifacts/phase3/validation_subset/validation_subset_256.json

home/wuyifan/EgoNet_train_20261008/configs/phase3/depth_fusion/depth_dropout005_geometry_m3_l1_t1_s20260718.yaml

@@ -0,0 +1,65 @@
+name: depth_dropout005_geometry_m3_l1_t1_s20260718
+seed: 20260718
+experiment:
+  phase: 3
+  matrix: depth_modality_dropout_dose
+  modality: M3
+  layout: L1
+  temporal: T1
+  variant: dropout005_geometry
+  registered_axes:
+  - modality_dropout_dose
+  reference: D2_dropout00_and_dropout02_endpoints
+data:
+  resolution_hw:
+  - 288
+  - 512
+  T: 8
+  stride: 1
+  batch_size: 1
+model:
+  encoder: dinov3
+  decoder_dim: 512
+  decoder_layers: 3
+  attention_heads: 8
+  decoder_mode: alternating
+  use_temporal: true
+  use_depth: true
+  depth_mode: geometry
+  depth_gate_mode: fixed
+  depth_gate_fixed_value: 0.5
+  use_rays: true
+  use_bimanual_interaction: true
+  predict_presence: true
+  require_mano: true
+  depth_gate_floor: 0.0
+  modality_dropout_prob: 0.05
+loss:
+  joint2d: 1.0
+  joint3d_camera: 10.0
+  vertices_camera: 1.0
+  local_pose: 1.0
+  global_orientation: 1.0
+  shape: 0.1
+  translation: 10.0
+  reprojection: 1.0
+  temporal_velocity: 0.1
+  temporal_acceleration: 0.05
+  bimanual_relative_wrist: 1.0
+  presence: 0.0
+assets:
+  dinov3_source: third_party/dinov3
+  dinov3_weights: /SSD3/jiasen/foundation_hand/pretrained/dinov3
+  mano_root: /SSD3/jiasen/foundation_hand/pretrained/mano
+precision: bf16
+protocol:
+  max_steps: 100000
+  num_workers: 4
+  checkpoint_interval: 25000
+  validation_interval: 5000
+  validation_batches: 256
+  learning_rate: 5.0e-05
+  min_learning_rate: 5.0e-06
+  warmup_steps: 500
+  early_stopping_patience: 0
+  validation_subset_manifest: artifacts/phase3/validation_subset/validation_subset_256.json

home/wuyifan/EgoNet_train_20261008/configs/phase3/depth_fusion/depth_dropout00_geometry_m3_l1_t1_s20260716.yaml

@@ -0,0 +1,65 @@
+name: depth_dropout00_geometry_m3_l1_t1_s20260716
+seed: 20260716
+experiment:
+  phase: 3
+  matrix: depth_modality_dropout
+  modality: M3
+  layout: L1
+  temporal: T1
+  variant: dropout00_geometry
+  registered_axes:
+  - modality_dropout
+  reference: dropout00_geometry_same_commit
+data:
+  resolution_hw:
+  - 288
+  - 512
+  T: 8
+  stride: 1
+  batch_size: 1
+model:
+  encoder: dinov3
+  decoder_dim: 512
+  decoder_layers: 3
+  attention_heads: 8
+  decoder_mode: alternating
+  use_temporal: true
+  use_depth: true
+  depth_mode: geometry
+  depth_gate_mode: fixed
+  depth_gate_fixed_value: 0.5
+  use_rays: true
+  use_bimanual_interaction: true
+  predict_presence: true
+  require_mano: true
+  depth_gate_floor: 0.0
+  modality_dropout_prob: 0.0
+loss:
+  joint2d: 1.0
+  joint3d_camera: 10.0
+  vertices_camera: 1.0
+  local_pose: 1.0
+  global_orientation: 1.0
+  shape: 0.1
+  translation: 10.0
+  reprojection: 1.0
+  temporal_velocity: 0.1
+  temporal_acceleration: 0.05
+  bimanual_relative_wrist: 1.0
+  presence: 0.0
+assets:
+  dinov3_source: third_party/dinov3
+  dinov3_weights: /SSD3/jiasen/foundation_hand/pretrained/dinov3
+  mano_root: /SSD3/jiasen/foundation_hand/pretrained/mano
+precision: bf16
+protocol:
+  max_steps: 100000
+  num_workers: 4
+  checkpoint_interval: 25000
+  validation_interval: 5000
+  validation_batches: 256
+  learning_rate: 5.0e-05
+  min_learning_rate: 5.0e-06
+  warmup_steps: 500
+  early_stopping_patience: 0
+  validation_subset_manifest: artifacts/phase3/validation_subset/validation_subset_256.json