CUMTHenry commited on
Commit
fed1013
Β·
verified Β·
1 Parent(s): ef5f07e

Add model card documenting layout and contents

Browse files
Files changed (1) hide show
  1. README.md +116 -0
README.md ADDED
@@ -0,0 +1,116 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: mit
3
+ tags:
4
+ - reinforcement-learning
5
+ - offline-rl
6
+ - diffusion-models
7
+ - consistency-models
8
+ - ogbench
9
+ - compdiffuser
10
+ - pytorch
11
+ library_name: pytorch
12
+ ---
13
+
14
+ # CompDiffuser checkpoints
15
+
16
+ Model parameters for the Native Chunk Consistency Models codebase,
17
+ [Henry0132/comp_flow_map](https://github.com/Henry0132/comp_flow_map).
18
+
19
+ This repository holds weights and their saved configurations only. Rollout results,
20
+ videos and evaluation output are deliberately excluded, so it stays a checkpoint
21
+ store rather than an experiment archive.
22
+
23
+ ## Download
24
+
25
+ The repository root holds one directory per dataset plus `dql_results/`, while the
26
+ code expects dataset directories under `logs/`. The helper script in the code
27
+ repository restores both into the right places:
28
+
29
+ ```bash
30
+ bash download_artifacts_from_hf.sh --repo CUMTHenry/compdiffuser_checkpoints
31
+ ```
32
+
33
+ By hand, the equivalent is two calls:
34
+
35
+ ```bash
36
+ hf download CUMTHenry/compdiffuser_checkpoints --include "dql_results/*" --local-dir .
37
+ hf download CUMTHenry/compdiffuser_checkpoints --exclude "dql_results/*" --local-dir logs
38
+ ```
39
+
40
+ Set `HF_ENDPOINT=https://hf-mirror.com` for a faster download inside mainland China.
41
+
42
+ ## Layout
43
+
44
+ ```text
45
+ <dataset>/diffusion/<experiment>/ original CompDiffuser planner and inverse-dynamics weights
46
+ <dataset>/chunk_consistency/<experiment>/ chunk consistency student distilled from that teacher
47
+ dql_results/<cube-dataset>/ released Cube low-level controllers
48
+ ```
49
+
50
+ ## Contents
51
+
52
+ Steps are the saved checkpoint step in thousands. `β€”` means that role is not part
53
+ of this dataset.
54
+
55
+ | Dataset | Planner / denoiser | Inverse dynamics | Chunk consistency |
56
+ | --- | --- | --- | --- |
57
+ | `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 |
58
+ | `antmaze-medium-stitch-v0` | `og_antM_Me_o2d_Cd_Stgl_PadBuf_Ft64_ts512_T512` @ 1600k | `og_antM_Me_o29d_g2d_invdyn_h12` @ 800k | β€” |
59
+ | `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 | β€” |
60
+ | `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 | β€” |
61
+ | `cube-single-play-v0` | `og_cubeSingle_o28d_DiTd1024dp12_PadBuf_Ft64_fs4_h160_ovlp56MditD512_H160_T512` @ 1600k | β€” | β€” |
62
+ | `cube-double-play-v0` | `og_cubeDouble_o37d_DiTd1024dp12_PadBuf_Ft64_fs4_h160_ovlp56MditD512_H160_T512` @ 1600k | β€” | β€” |
63
+ | `cube-triple-play-v0` | `og_cubeTriple_o46d_DiTd1024dp12_PadBuf_Ft64_fs4_h160_ovlp56MditD512_H160_T512` @ 1600k | β€” | β€” |
64
+ | `cube-quadruple-play-v0` | `og_cubeQuadruple_o55d_DiTd1024dp12_PadBuf_Ft64_fs4_h160_ovlp56MditD512_H160_T512` @ 1600k | β€” | β€” |
65
+ | `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 | β€” |
66
+ | `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 | β€” |
67
+ | `pointmaze-giant-stitch-v0` | `og_pntM_Gi_o2d_Cd_Stgl_PadBuf_Ft64_ts512_T512` @ 1200k | β€” | β€” |
68
+ | `pointmaze-medium-stitch-v0` | `og_pntM_Me_o2d_Cd_Stgl_PadBuf_Ft64_ts512_T512` @ 1200k | β€” | β€” |
69
+
70
+ Cube low-level controllers, one directory per dataset under `dql_results/`:
71
+
72
+ | Dataset | Files |
73
+ | --- | --- |
74
+ | `cube-single-play-v0` | `actor_200.pth`, `critic_200.pth`, `critic_target_200.pth`, `variant.json` |
75
+ | `cube-double-play-v0` | same |
76
+ | `cube-triple-play-v0` | same |
77
+ | `cube-quadruple-play-v0` | same |
78
+
79
+ ## File types
80
+
81
+ | File | Contents |
82
+ | --- | --- |
83
+ | `state_<step>.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`. |
84
+ | `model_config.pkl` | Network class and constructor arguments. |
85
+ | `dfu_model.pkl` | Diffusion process class and arguments. |
86
+ | `dataset_config.pkl` | Dataset class, normalization and sequence settings. |
87
+ | `small_trainer_config.pkl`, `inv_trainer_config.pkl` | Trainer class and settings. |
88
+ | `render_config.pkl` | Renderer settings used for evaluation videos. |
89
+ | `model_config.txt` | Human-readable module summary. |
90
+ | `args.json` | Command-line arguments of the original run. |
91
+ | `train_metrics.jsonl` | Per-interval training metrics. |
92
+
93
+ The saved pickles are authoritative. Experiment directory names encode the task,
94
+ observation and goal dimensions, architecture and training settings, but the
95
+ `ts` tokens are not documented here; read `args.json` and `model_config.pkl` for
96
+ the exact values. A checkpoint filename step and its internal `step` field can
97
+ differ, because the original release named files by epoch.
98
+
99
+ ## Notes
100
+
101
+ - Only `antmaze-giant-stitch-v0` has a chunk consistency student so far. Its run
102
+ was still training when these weights were uploaded, so later checkpoints may
103
+ appear in a future revision.
104
+ - `train_metrics.jsonl` is append-only and grows while training runs, so its size
105
+ can differ from the local copy.
106
+ - Planning additionally needs the datasets and the per-task assets expected by
107
+ each environment; those are not part of this repository.
108
+
109
+ ## License and acknowledgements
110
+
111
+ Released under the [MIT License](https://github.com/Henry0132/comp_flow_map/blob/main/LICENSE).
112
+ Built on [RCD](https://refining-compositional-diffusion.github.io/),
113
+ [CompDiffuser](https://github.com/devinluo27/comp_diffuser_release),
114
+ [CDGS](https://github.com/UtkarshMishra04/CDGS_ogbench) and
115
+ [OGBench](https://github.com/seohongpark/ogbench). Preserve the upstream notices
116
+ and cite the corresponding works.