Update hftrainer metadata

README.md

@@ -1,104 +1,3 @@
----
-license: mit
-tags:
-  - text-to-motion
-  - human-motion-generation
-  - motion-generation
-  - momask
-  - humanml3d
-library_name: hftrainer
-pipeline_tag: text-to-3d
----
-
-# MoMask — HumanML3D (hftrainer artifact)
-
-Self-contained hftrainer re-export of the official **MoMask** text-to-motion
-model (Guo et al., CVPR 2024) on **HumanML3D-263**. The RVQ-VAE tokenizer,
-masked generative transformer, residual transformer and length estimator are
-vendored and **numerically parity-equal to the released checkpoints**
-(`convert_momask_checkpoint.py --verify` reports max-abs-diff `0.0` on
-generation). The CLIP ViT-B/32 text encoder is **not** stored here; it is
-reloaded by name at load time.
-
-- **Task**: Text-to-Motion (T2M)
-- **Representation**: HumanML3D-263 (263-dim, 20 fps, 22-joint SMPL skeleton)
-- **Tokenizer**: RVQ-VAE, 6 residual quantizers, codebook 512×512
-- **Generator**: MaskTransformer (masked iterative decoding) + ResidualTransformer
-- **Text encoder**: CLIP ViT-B/32 (frozen, reloaded by name)
-- **Paper**: [MoMask: Generative Masked Modeling of 3D Human Motions](https://arxiv.org/abs/2312.00063)
-- **Original code**: https://github.com/EricGuo5513/momask-codes
-
-## Files
-
-| File | Contents |
-|---|---|
-| `momask_config.json` | architecture config for all sub-modules |
-| `vq.safetensors` | RVQ-VAE weights |
-| `t2m_trans.safetensors` | MaskTransformer (no CLIP) |
-| `res_trans.safetensors` | ResidualTransformer (no CLIP) |
-| `length_est.safetensors` | LengthEstimator |
-| `Mean.npy` / `Std.npy` | 263-dim de-normalization stats |
-
-## Usage
-
-```python
-from huggingface_hub import snapshot_download
-from hftrainer.models.motion.momask import MoMaskBundle
-from hftrainer.pipelines.momask import MoMaskPipeline
-
-path = snapshot_download("ZeyuLing/momask-humanml3d")
-bundle = MoMaskBundle.from_pretrained(path)
-pipe = MoMaskPipeline(bundle, device="cuda")
-
-# length picked by the length estimator:
-motions = pipe.infer_t2m(["a person walks forward then sits down"])  # list of (T, 263)
-# or a fixed length (frames @ 20 fps):
-motions = pipe.infer_t2m(["a person walks forward then sits down"], [120])
-```
-
-## Evaluation (HumanML3D-263, Guo evaluator)
-
-20 repeats, n = 3970. Generation under the official HumanML3D protocol
-(standard test split, native 263-dim @ 20 fps).
-
-| Metric | this artifact | MoMask paper |
-|---|---|---|
-| FID ↓ | **0.097** | 0.045 |
-| R-Precision Top-1 / 2 / 3 ↑ | **0.516 / 0.709 / 0.804** | 0.521 / 0.713 / 0.807 |
-| MM-Dist ↓ | **2.990** | 2.958 |
-| Diversity → | **9.460** | 9.620 |
-
-R-Precision, MM-Dist and Diversity match the paper essentially exactly. The
-small residual FID gap is a data-processing / evaluation-population difference,
-not a generation-quality gap (the decode path is verified parity-equal to the
-released MoMask inference).
-
-## Citation
-
-```bibtex
-@inproceedings{guo2024momask,
-  title     = {MoMask: Generative Masked Modeling of 3D Human Motions},
-  author    = {Guo, Chuan and Mu, Yuxuan and Javed, Muhammad Gohar and Wang, Sen and Cheng, Li},
-  booktitle = {CVPR},
-  year      = {2024}
-}
-```
-
-## Evaluation (MotionStreamer-272, SMPL retarget path)
-
-For cross-model comparison with the MotionStreamer / HYMotion-M2M evaluator,
-native HumanML3D-263 predictions are retargeted through the validated MDM-style
-chain: HML263 -> SMPL `motion_135` (IK refine-80, 20 -> 30 fps) ->
-MotionStreamer-272 -> `MotionStreamer272Evaluator`.
-
-| Metric | MoMask (this artifact) | MS-272 GT/Real |
-|---|---:|---:|
-| FID ↓ | **114.869** | 0.000 |
-| R-Precision Top-1 ↑ | **0.485** | 0.706 |
-| R-Precision Top-2 ↑ | **0.650** | 0.857 |
-| R-Precision Top-3 ↑ | **0.731** | 0.911 |
-| MM-Dist ↓ | **19.411** | 15.007 |
-| Diversity → | **25.427** | 27.281 |
-
-Run details: `n_repeats = 20`, `n_samples_used = 7392`, `skipped_no_pred = 0`.
+# MoMask hftrainer Artifact
 
+Load with `MoMaskBundle.from_pretrained(...)`. CLIP ViT-B/32 is an external runtime dependency and is not stored here.