--- license: mit tags: - reinforcement-learning - offline-rl - diffusion-models - consistency-models - ogbench - compdiffuser - pytorch library_name: pytorch --- # CompDiffuser checkpoints Model parameters for the Native Chunk Consistency Models codebase, [Henry0132/comp_flow_map](https://github.com/Henry0132/comp_flow_map). This repository holds weights and their saved configurations only. Rollout results, videos and evaluation output are deliberately excluded, so it stays a checkpoint store rather than an experiment archive. ## Download The repository root holds one directory per dataset plus `dql_results/`, while the code expects dataset directories under `logs/`. The helper script in the code repository restores both into the right places: ```bash bash download_artifacts_from_hf.sh --repo CUMTHenry/compdiffuser_checkpoints ``` By hand, the equivalent is two calls: ```bash hf download CUMTHenry/compdiffuser_checkpoints --include "dql_results/*" --local-dir . hf download CUMTHenry/compdiffuser_checkpoints --exclude "dql_results/*" --local-dir logs ``` Set `HF_ENDPOINT=https://hf-mirror.com` for a faster download inside mainland China. ## Layout ```text /diffusion// original CompDiffuser planner and inverse-dynamics weights /chunk_consistency// chunk consistency student distilled from that teacher dql_results// released Cube low-level controllers ``` ## Contents Steps are the saved checkpoint step in thousands. `—` means that role is not part of this dataset. | Dataset | Planner / denoiser | Inverse dynamics | Chunk consistency | | --- | --- | --- | --- | | `antmaze-giant-stitch-v0` | `og_antM_Gi_o2d_Cd_Stgl_PadBuf_Ft64_ts512_T512` @ 1600k | `og_antM_Gi_o29d_g2d_invdyn_h12` @ 800k | `..._ChunkConsistency_v1_restart_s200k` @ 200k, 400k, 600k | | `antmaze-medium-stitch-v0` | `og_antM_Me_o2d_Cd_Stgl_PadBuf_Ft64_ts512_T512` @ 1600k | `og_antM_Me_o29d_g2d_invdyn_h12` @ 800k | — | | `antsoccer-arena-stitch-v0` | `og_antSoc_Ar_o17d_DiTd768_PadBuf_Ft64_ts512_fs4_h160_ovlp56MditD384_T512` @ 1800k | `og_antSoc_Ar_o42d_g17d_invdyn_h120_dout02` @ 1800k | — | | `antsoccer-medium-stitch-v0` | `og_antSoc_Me_o17d_DiTd768_PadBuf_Ft64_MaxOriPlan200_ts512_fs4_h160_ovlp56MditD384_T512` @ 1800k | `og_antSoc_Me_o42d_g17d_invdyn_h100_dm4_dout02` @ 1800k | — | | `cube-single-play-v0` | `og_cubeSingle_o28d_DiTd1024dp12_PadBuf_Ft64_fs4_h160_ovlp56MditD512_H160_T512` @ 1600k | — | — | | `cube-double-play-v0` | `og_cubeDouble_o37d_DiTd1024dp12_PadBuf_Ft64_fs4_h160_ovlp56MditD512_H160_T512` @ 1600k | — | — | | `cube-triple-play-v0` | `og_cubeTriple_o46d_DiTd1024dp12_PadBuf_Ft64_fs4_h160_ovlp56MditD512_H160_T512` @ 1600k | — | — | | `cube-quadruple-play-v0` | `og_cubeQuadruple_o55d_DiTd1024dp12_PadBuf_Ft64_fs4_h160_ovlp56MditD512_H160_T512` @ 1600k | — | — | | `humanoidmaze-giant-stitch-v0` | `og_humM_Gi_o2d_Cd_Stgl_PadBuf_Ft64_ts1k_h336_ovlp128_ovdm5_ts512_bs192_bd128_td96_drop02_T1000` @ 1800k | `og_humM_Gi_o69d_g2d_invdyn_h80_dm5_dout02` @ 1600k | — | | `humanoidmaze-medium-stitch-v0` | `og_humM_Me_o2d_Cd_Stgl_PadBuf_Ft64_ts1k_h336_ovlp128_ovdm5_ts512_bs192_bd128_td96_drop02_T1000` @ 1600k, 1800k | `og_humM_Me_o69d_g2d_invdyn_h80_dm5_dout02` @ 800k, 1600k | — | | `pointmaze-giant-stitch-v0` | `og_pntM_Gi_o2d_Cd_Stgl_PadBuf_Ft64_ts512_T512` @ 1200k | — | — | | `pointmaze-medium-stitch-v0` | `og_pntM_Me_o2d_Cd_Stgl_PadBuf_Ft64_ts512_T512` @ 1200k | — | — | Cube low-level controllers, one directory per dataset under `dql_results/`: | Dataset | Files | | --- | --- | | `cube-single-play-v0` | `actor_200.pth`, `critic_200.pth`, `critic_target_200.pth`, `variant.json` | | `cube-double-play-v0` | same | | `cube-triple-play-v0` | same | | `cube-quadruple-play-v0` | same | ## File types | File | Contents | | --- | --- | | `state_.pt` | Checkpoint dict. Planner and inverse-dynamics weights hold `step`, `model` and `ema`. Chunk consistency weights additionally hold `optimizer`, `consistency`, `teacher`, `checkpoint_schema_version` and `artifact_type`. | | `model_config.pkl` | Network class and constructor arguments. | | `dfu_model.pkl` | Diffusion process class and arguments. | | `dataset_config.pkl` | Dataset class, normalization and sequence settings. | | `small_trainer_config.pkl`, `inv_trainer_config.pkl` | Trainer class and settings. | | `render_config.pkl` | Renderer settings used for evaluation videos. | | `model_config.txt` | Human-readable module summary. | | `args.json` | Command-line arguments of the original run. | | `train_metrics.jsonl` | Per-interval training metrics. | The saved pickles are authoritative. Experiment directory names encode the task, observation and goal dimensions, architecture and training settings, but the `ts` tokens are not documented here; read `args.json` and `model_config.pkl` for the exact values. A checkpoint filename step and its internal `step` field can differ, because the original release named files by epoch. ## Notes - Only `antmaze-giant-stitch-v0` has a chunk consistency student so far. Its run was still training when these weights were uploaded, so later checkpoints may appear in a future revision. - `train_metrics.jsonl` is append-only and grows while training runs, so its size can differ from the local copy. - Planning additionally needs the datasets and the per-task assets expected by each environment; those are not part of this repository. ## License and acknowledgements Released under the [MIT License](https://github.com/Henry0132/comp_flow_map/blob/main/LICENSE). Built on [RCD](https://refining-compositional-diffusion.github.io/), [CompDiffuser](https://github.com/devinluo27/comp_diffuser_release), [CDGS](https://github.com/UtkarshMishra04/CDGS_ogbench) and [OGBench](https://github.com/seohongpark/ogbench). Preserve the upstream notices and cite the corresponding works.