AAAcon commited on
Commit
2bb548b
·
verified ·
1 Parent(s): 13a8423

Add dataset card

Browse files
Files changed (1) hide show
  1. README.md +86 -0
README.md ADDED
@@ -0,0 +1,86 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: mit
3
+ task_categories:
4
+ - other
5
+ tags:
6
+ - motion-generation
7
+ - diffusion-models
8
+ - cross-skeleton
9
+ - retargeting
10
+ - anonymous-submission
11
+ size_categories:
12
+ - 1K<n<10K
13
+ ---
14
+
15
+ # anytop-checkpoints
16
+
17
+ Pre-trained checkpoints accompanying the anonymous NeurIPS 2026 submission
18
+ **"Why Cross-Skeleton Retargeting Is Non-Identifiable: Structural Limits of
19
+ Generative Motion Models"**.
20
+
21
+ This repository contains a single tarball, `anytop-checkpoints.tar.gz`
22
+ (6.5 GB compressed, 7.2 GB extracted), bundling the trained model weights
23
+ for the 14-method comparison reported in the paper.
24
+
25
+ ## Contents
26
+
27
+ After extraction (`tar -xzvf anytop-checkpoints.tar.gz`):
28
+
29
+ ```
30
+ anytop-checkpoints/
31
+ README.md (per-method usage notes)
32
+ anytop_v5/ AnyTop transductive (model000175000.pt, 34 MB)
33
+ ace_primary_70/ ACE-T transductive (ckpt_final.pt, 83 MB)
34
+ ace_inductive_60/ ACE-I inductive (ckpt_final.pt, 83 MB)
35
+ moreflow_primary_70/ MoReFlow-T (ckpt_final.pt, 79 MB)
36
+ moreflow_inductive_60/ MoReFlow-I (ckpt_final.pt, 79 MB)
37
+ anchor_label_flow/ AL-Flow (ckpt_final.pt, 86 MB)
38
+ anchor_label_flow_src/ AL-Flow-Src (ckpt_final.pt, 88 MB)
39
+ anchor_label_flow_src_g/ AL-Flow-Src-G (ckpt_final.pt, 87 MB)
40
+ dpg_sb_v3/ DPG-SB-v3 (final.pt 62 MB + z_stats.pt)
41
+ moreflow_vqvae/ 71 per-skeleton VQ-VAE encoders (~91 MB each)
42
+ Required for ACE / MoReFlow / DPG-SB-v3 inference
43
+ moreflow_caches/ Pre-extracted MoReFlow latents
44
+ (only required to retrain DPG-SB-v3 from scratch)
45
+ ```
46
+
47
+ ## Verification
48
+
49
+ ```
50
+ SHA-256: fc15f536510014a7bb953aa62cd24358de6af11f0d92f0f8360e2ec44b8251e2
51
+ ```
52
+
53
+ ```bash
54
+ sha256sum -c <<< "fc15f536510014a7bb953aa62cd24358de6af11f0d92f0f8360e2ec44b8251e2 anytop-checkpoints.tar.gz"
55
+ ```
56
+
57
+ ## How to use
58
+
59
+ The companion code repository contains training scripts, evaluation runners,
60
+ and the SIF metric implementation. After downloading and extracting this
61
+ tarball into the code repo's `save/` directory, every per-method invocation
62
+ in the code repo's `REPRODUCE.md` § 4 will work without retraining.
63
+
64
+ ```bash
65
+ # From the code repo root:
66
+ mkdir -p save
67
+ tar -xzvf anytop-checkpoints.tar.gz
68
+ mv anytop-checkpoints/* save/
69
+ rmdir anytop-checkpoints
70
+
71
+ # Verify with one method:
72
+ python -m eval.baselines.run_anytop_v5 \
73
+ --ckpt save/anytop_v5/model000175000.pt \
74
+ --manifest eval/benchmark_v3/queries_sif/manifest.json \
75
+ --output_dir results/anytop_sif
76
+ ```
77
+
78
+ ## License
79
+
80
+ MIT, matching the upstream AnyTop release.
81
+
82
+ ## Anonymity
83
+
84
+ Every `args.json` file in the package has been swept for personal paths,
85
+ author names, and W&B entity strings. Checkpoint binaries (`*.pt`) contain
86
+ only PyTorch model and optimiser state.