Add Motius DART HumanML3D checkpoint

.gitattributes

@@ -1,35 +1,2 @@
-*.7z filter=lfs diff=lfs merge=lfs -text
-*.arrow filter=lfs diff=lfs merge=lfs -text
-*.bin filter=lfs diff=lfs merge=lfs -text
-*.bz2 filter=lfs diff=lfs merge=lfs -text
-*.ckpt filter=lfs diff=lfs merge=lfs -text
-*.ftz filter=lfs diff=lfs merge=lfs -text
-*.gz filter=lfs diff=lfs merge=lfs -text
-*.h5 filter=lfs diff=lfs merge=lfs -text
-*.joblib filter=lfs diff=lfs merge=lfs -text
-*.lfs.* filter=lfs diff=lfs merge=lfs -text
-*.mlmodel filter=lfs diff=lfs merge=lfs -text
-*.model filter=lfs diff=lfs merge=lfs -text
-*.msgpack filter=lfs diff=lfs merge=lfs -text
-*.npy filter=lfs diff=lfs merge=lfs -text
-*.npz filter=lfs diff=lfs merge=lfs -text
-*.onnx filter=lfs diff=lfs merge=lfs -text
-*.ot filter=lfs diff=lfs merge=lfs -text
-*.parquet filter=lfs diff=lfs merge=lfs -text
-*.pb filter=lfs diff=lfs merge=lfs -text
-*.pickle filter=lfs diff=lfs merge=lfs -text
-*.pkl filter=lfs diff=lfs merge=lfs -text
 *.pt filter=lfs diff=lfs merge=lfs -text
-*.pth filter=lfs diff=lfs merge=lfs -text
-*.rar filter=lfs diff=lfs merge=lfs -text
-*.safetensors filter=lfs diff=lfs merge=lfs -text
-saved_model/**/* filter=lfs diff=lfs merge=lfs -text
-*.tar.* filter=lfs diff=lfs merge=lfs -text
-*.tar filter=lfs diff=lfs merge=lfs -text
-*.tflite filter=lfs diff=lfs merge=lfs -text
-*.tgz filter=lfs diff=lfs merge=lfs -text
-*.wasm filter=lfs diff=lfs merge=lfs -text
-*.xz filter=lfs diff=lfs merge=lfs -text
-*.zip filter=lfs diff=lfs merge=lfs -text
-*.zst filter=lfs diff=lfs merge=lfs -text
-*tfevents* filter=lfs diff=lfs merge=lfs -text
+*.pkl filter=lfs diff=lfs merge=lfs -text

README.md

@@ -0,0 +1,27 @@
+---
+library_name: motius
+tags:
+- text-to-motion
+- human-motion-generation
+- dart
+- dart276
+---
+
+# Motius DART HumanML3D Checkpoint
+
+This artifact contains the DART/DartControl HumanML3D denoiser, MVAE, runtime configuration, seed motion, and DART276 normalization/text-embedding assets used by the Motius DART pipeline.
+
+It intentionally does **not** include SMPL, SMPL-H, or SMPL-X body model files. Those files are license-controlled assets and must be installed locally by the user under `data/smplx_lockedhead_20230207/models_lockedhead` before running full SMPL rollout or mesh export.
+
+Load with:
+
+```python
+from motius.pipelines.dart import DARTPipeline
+
+pipe = DARTPipeline.from_pretrained(
+    "ZeyuLing/motius-dart-humanml3d",
+    device="cuda",
+)
+```
+
+Native motion representation: `DART276` at 20 fps. SMPL and `motion_135` exports are adapters for rendering and evaluation.

config_files/config_hydra/default.yaml

@@ -0,0 +1,3 @@
+name: default
+defaults:
+  - motion_primitive: mp_2_8

config_files/config_hydra/motion_primitive/hml_mp_h20_f1_r1.yaml

@@ -0,0 +1,4 @@
+fps: 20
+history_length: 20
+future_length: 1
+num_primitive: 1

config_files/config_hydra/motion_primitive/hml_mp_h20_f40_r1.yaml

@@ -0,0 +1,4 @@
+fps: 20
+history_length: 20
+future_length: 40
+num_primitive: 1

config_files/config_hydra/motion_primitive/hml_mp_h20_f40_r2.yaml

@@ -0,0 +1,4 @@
+fps: 20
+history_length: 20
+future_length: 40
+num_primitive: 2

config_files/config_hydra/motion_primitive/hml_mp_h20_f40_r4.yaml

@@ -0,0 +1,4 @@
+fps: 20
+history_length: 20
+future_length: 40
+num_primitive: 4

config_files/config_hydra/motion_primitive/hml_mp_h2_f8_r4.yaml

@@ -0,0 +1,4 @@
+fps: 20
+history_length: 2
+future_length: 8
+num_primitive: 4

config_files/config_hydra/motion_primitive/hml_mp_h2_f8_r8.yaml

@@ -0,0 +1,4 @@
+fps: 20
+history_length: 2
+future_length: 8
+num_primitive: 8

config_files/config_hydra/motion_primitive/mp_2_8.yaml

@@ -0,0 +1,4 @@
+fps: 30
+history_length: 2
+future_length: 8
+num_primitive: 8

config_files/config_hydra/motion_primitive/mp_h1_f1_r32.yaml

@@ -0,0 +1,4 @@
+fps: 30
+history_length: 1
+future_length: 1
+num_primitive: 32

config_files/config_hydra/motion_primitive/mp_h1_f1_r4.yaml

@@ -0,0 +1,4 @@
+fps: 30
+history_length: 1
+future_length: 1
+num_primitive: 4

config_files/config_hydra/motion_primitive/mp_h1_f1_r8.yaml

@@ -0,0 +1,4 @@
+fps: 30
+history_length: 1
+future_length: 1
+num_primitive: 8

config_files/config_hydra/motion_primitive/mp_h2_f16_r2.yaml

@@ -0,0 +1,4 @@
+fps: 30
+history_length: 2
+future_length: 16
+num_primitive: 2

config_files/config_hydra/motion_primitive/mp_h2_f16_r4.yaml

@@ -0,0 +1,4 @@
+fps: 30
+history_length: 2
+future_length: 16
+num_primitive: 4

config_files/config_hydra/motion_primitive/mp_h2_f16_r8.yaml

@@ -0,0 +1,4 @@
+fps: 30
+history_length: 2
+future_length: 16
+num_primitive: 8

config_files/config_hydra/motion_primitive/mp_h2_f4_r2.yaml

@@ -0,0 +1,4 @@
+fps: 30
+history_length: 2
+future_length: 4
+num_primitive: 2

config_files/config_hydra/motion_primitive/mp_h2_f4_r4.yaml

@@ -0,0 +1,4 @@
+fps: 30
+history_length: 2
+future_length: 4
+num_primitive: 4

config_files/config_hydra/motion_primitive/mp_h2_f4_r8.yaml

@@ -0,0 +1,4 @@
+fps: 30
+history_length: 2
+future_length: 4
+num_primitive: 8

config_files/config_hydra/motion_primitive/mp_h2_f8_r1.yaml

@@ -0,0 +1,4 @@
+fps: 30
+history_length: 2
+future_length: 8
+num_primitive: 1

config_files/config_hydra/motion_primitive/mp_h2_f8_r16.yaml

@@ -0,0 +1,4 @@
+fps: 30
+history_length: 2
+future_length: 8
+num_primitive: 16

config_files/config_hydra/motion_primitive/mp_h2_f8_r2.yaml

@@ -0,0 +1,4 @@
+fps: 30
+history_length: 2
+future_length: 8
+num_primitive: 2

config_files/config_hydra/motion_primitive/mp_h2_f8_r4.yaml

@@ -0,0 +1,4 @@
+fps: 30
+history_length: 2
+future_length: 8
+num_primitive: 4

config_files/config_hydra/motion_primitive/mp_h2_f8_r8.yaml

@@ -0,0 +1,4 @@
+fps: 30
+history_length: 2
+future_length: 8
+num_primitive: 8

config_files/config_hydra/motion_primitive/mp_h4_f16_r8.yaml

@@ -0,0 +1,4 @@
+fps: 30
+history_length: 4
+future_length: 16
+num_primitive: 8

config_files/config_hydra/motion_primitive/mp_h4_f8_r2.yaml

@@ -0,0 +1,4 @@
+fps: 30
+history_length: 4
+future_length: 8
+num_primitive: 2

config_files/config_hydra/motion_primitive/mp_h4_f8_r4.yaml

@@ -0,0 +1,4 @@
+fps: 30
+history_length: 4
+future_length: 8
+num_primitive: 4

config_files/config_hydra/motion_primitive/mp_h4_f8_r8.yaml

@@ -0,0 +1,4 @@
+fps: 30
+history_length: 4
+future_length: 8
+num_primitive: 8

config_files/config_hydra/motion_primitive/mp_h8_f8_r1.yaml

@@ -0,0 +1,4 @@
+fps: 30
+history_length: 8
+future_length: 8
+num_primitive: 1

config_files/config_hydra/motion_primitive/mp_h8_f8_r2.yaml

@@ -0,0 +1,4 @@
+fps: 30
+history_length: 8
+future_length: 8
+num_primitive: 2

config_files/config_hydra/motion_primitive/mp_h8_f8_r4.yaml

@@ -0,0 +1,4 @@
+fps: 30
+history_length: 8
+future_length: 8
+num_primitive: 4

config_files/config_hydra/motion_primitive/mp_h8_f8_r8.yaml

@@ -0,0 +1,4 @@
+fps: 30
+history_length: 8
+future_length: 8
+num_primitive: 8

config_files/data_paths.py

@@ -0,0 +1,7 @@
+from pathlib import Path
+
+
+dataset_root_dir = Path('data')
+body_model_dir = dataset_root_dir / 'smplx_lockedhead_20230207/models_lockedhead/'
+amass_dir = dataset_root_dir / 'amass'
+babel_dir = amass_dir / 'babel-teach'

data/stand.pkl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7d784ff362f7e2a2ae6ea5954d673953ca7c4980bcc3aca3a4b47204aa1031db
+size 12008

data/stand_20fps.pkl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e4601960acc9abba527b62326b14cfd4b8c20b31b47a8ea01454d4354d20e215
+size 12008

mld_denoiser/smplh_hml3d_2_8_4/args.yaml

@@ -0,0 +1,23 @@
+"# tyro YAML.\n!dataclass:MLDArgs\ndata_args: !dataclass:DataArgs\n  body_type: smplh\n\
+  \  cfg_path: ./config_files/config_hydra/motion_primitive/hml_mp_h2_f8_r4.yaml\n\
+  \  data_dir: ./data/hml3d_smplh/seq_data_zero_male/\n  dataset: hml3d\n  enforce_gender:\
+  \ male\n  enforce_zero_beta: 1\n  feature_dim: 276\n  future_length: 8\n  history_length:\
+  \ 2\n  num_primitive: 4\n  prob_static: 0.0\n  text_tolerance: 0.0\n  weight_scheme:\
+  \ uniform\ndenoiser_args: !dataclass:DenoiserArgs\n  diffusion_args: !dataclass:DiffusionArgs\n\
+  \    diffusion_steps: 10\n    noise_schedule: cosine\n    respacing: ''\n    sigma_small:\
+  \ true\n  model_args: !dataclass:DenoiserTransformerArgs\n    activation: gelu\n\
+  \    clip_dim: 512\n    cond_mask_prob: 0.1\n    dropout: 0.1\n    ff_size: 1024\n\
+  \    h_dim: 512\n    history_shape: !!python/tuple\n    - 2\n    - 276\n    noise_shape:\
+  \ !!python/tuple\n    - 1\n    - 256\n    num_heads: 4\n    num_layers: 8\n  model_type:\
+  \ transformer\n  mvae_path: ./mvae/mvae_smplh_hml3d_2_8_4/checkpoint_200000.pt\n\
+  \  rescale_latent: 1\n  train_rollout_history: rollout\n  train_rollout_type: full\n\
+  device: cuda\nexp_name: smplh_hml3d_2_8_4\nsave_dir: !!python/object/apply:pathlib.PosixPath\n\
+  - mld_denoiser\n- smplh_hml3d_2_8_4\nseed: 0\ntorch_deterministic: true\ntrack:\
+  \ 1\ntrain_args: !dataclass:TrainArgs\n  anneal_lr: 1\n  batch_size: 1024\n  ema_decay:\
+  \ 0.999\n  grad_clip: 1.0\n  learning_rate: 0.0001\n  log_interval: 1000\n  resume_checkpoint:\
+  \ null\n  save_interval: 100000\n  stage1_steps: 100000\n  stage2_steps: 100000\n\
+  \  stage3_steps: 100000\n  use_amp: 1\n  use_predicted_joints: 0\n  val_interval:\
+  \ 10000\n  weight_feature_rec: 1.0\n  weight_joints_consistency: 0.0\n  weight_joints_delta:\
+  \ 10000.0\n  weight_kl: 0.0001\n  weight_latent_rec: 1.0\n  weight_orient_delta:\
+  \ 10000.0\n  weight_rec: 1.0\n  weight_smpl_joints_rec: 0.0\n  weight_transl_delta:\
+  \ 10000.0\nwandb_entity: interaction\nwandb_project_name: mld_denoiser\n"

mld_denoiser/smplh_hml3d_2_8_4/args_read.yaml

@@ -0,0 +1,74 @@
+data_args:
+  body_type: smplh
+  cfg_path: ./config_files/config_hydra/motion_primitive/hml_mp_h2_f8_r4.yaml
+  data_dir: ./data/hml3d_smplh/seq_data_zero_male/
+  dataset: hml3d
+  enforce_gender: male
+  enforce_zero_beta: 1
+  feature_dim: 276
+  future_length: 8
+  history_length: 2
+  num_primitive: 4
+  prob_static: 0.0
+  text_tolerance: 0.0
+  weight_scheme: uniform
+denoiser_args:
+  diffusion_args:
+    diffusion_steps: 10
+    noise_schedule: cosine
+    respacing: ''
+    sigma_small: true
+  model_args:
+    activation: gelu
+    clip_dim: 512
+    cond_mask_prob: 0.1
+    dropout: 0.1
+    ff_size: 1024
+    h_dim: 512
+    history_shape: !!python/tuple
+    - 2
+    - 276
+    noise_shape: !!python/tuple
+    - 1
+    - 256
+    num_heads: 4
+    num_layers: 8
+  model_type: transformer
+  mvae_path: ./mvae/mvae_smplh_hml3d_2_8_4/checkpoint_200000.pt
+  rescale_latent: 1
+  train_rollout_history: rollout
+  train_rollout_type: full
+device: cuda
+exp_name: smplh_hml3d_2_8_4
+save_dir: !!python/object/apply:pathlib.PosixPath
+- mld_denoiser
+- smplh_hml3d_2_8_4
+seed: 0
+torch_deterministic: true
+track: 1
+train_args:
+  anneal_lr: 1
+  batch_size: 1024
+  ema_decay: 0.999
+  grad_clip: 1.0
+  learning_rate: 0.0001
+  log_interval: 1000
+  resume_checkpoint: null
+  save_interval: 100000
+  stage1_steps: 100000
+  stage2_steps: 100000
+  stage3_steps: 100000
+  use_amp: 1
+  use_predicted_joints: 0
+  val_interval: 10000
+  weight_feature_rec: 1.0
+  weight_joints_consistency: 0.0
+  weight_joints_delta: 10000.0
+  weight_kl: 0.0001
+  weight_latent_rec: 1.0
+  weight_orient_delta: 10000.0
+  weight_rec: 1.0
+  weight_smpl_joints_rec: 0.0
+  weight_transl_delta: 10000.0
+wandb_entity: interaction
+wandb_project_name: mld_denoiser

mld_denoiser/smplh_hml3d_2_8_4/checkpoint_300000.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9a3acd0a0fdadff64ff9fad41a94cc2911316169a7bb80c69a97e6934ae73528
+size 82342402

model_index.json

@@ -0,0 +1,19 @@
+{
+  "model_type": "dart",
+  "library_name": "motius",
+  "task": "text-to-motion",
+  "motion_representation": "dart276",
+  "export_adapters": [
+    "smpl_sequence",
+    "smplh_motion135"
+  ],
+  "denoiser_checkpoint": "mld_denoiser/smplh_hml3d_2_8_4/checkpoint_300000.pt",
+  "vae_checkpoint": "mvae/mvae_smplh_hml3d_2_8_4/checkpoint_200000.pt",
+  "seed_sequence": "data/stand_20fps.pkl",
+  "guidance_param": 5.0,
+  "coord_conversion": "mbench",
+  "translation_source": "floor_aligned_smpl_transl",
+  "initial_transform": "standard",
+  "requires_body_model": true,
+  "body_model_note": "Place licensed SMPL-H/SMPL-X body model files under data/smplx_lockedhead_20230207/models_lockedhead before full rollout."
+}

mvae/mvae_smplh_hml3d_2_8_4/args.yaml

@@ -0,0 +1,19 @@
+"# tyro YAML.\n!dataclass:Args\ndata_args: !dataclass:DataArgs\n  body_type: smplh\n\
+  \  cfg_path: ./config_files/config_hydra/motion_primitive/hml_mp_h2_f8_r4.yaml\n\
+  \  data_dir: ./data/hml3d_smplh/seq_data_zero_male/\n  dataset: hml3d\n  enforce_gender:\
+  \ male\n  enforce_zero_beta: 1\n  feature_dim: 276\n  future_length: 8\n  history_length:\
+  \ 2\n  num_primitive: 4\n  prob_static: 0.0\n  text_tolerance: 0.0\n  weight_scheme:\
+  \ uniform\ndevice: cuda\nexp_name: mvae_smplh_hml3d_2_8_4\nmodel_args: !dataclass:VAEArgs\n\
+  \  activation: gelu\n  arch: all_encoder\n  dropout: 0.1\n  ff_size: 1024\n  h_dim:\
+  \ 256\n  latent_dim: !!python/tuple\n  - 1\n  - 256\n  nfeats: 276\n  normalize_before:\
+  \ false\n  num_heads: 4\n  num_layers: 7\n  position_embedding: learned\nsave_dir:\
+  \ !!python/object/apply:pathlib.PosixPath\n- mvae\n- mvae_smplh_hml3d_2_8_4\nseed:\
+  \ 0\ntorch_deterministic: true\ntrack: 1\ntrain_args: !dataclass:TrainArgs\n  anneal_lr:\
+  \ 1\n  batch_size: 128\n  ema_decay: 0.999\n  grad_clip: 1.0\n  learning_rate: 0.0001\n\
+  \  log_interval: 1000\n  resume_checkpoint: null\n  save_interval: 50000\n  stage1_steps:\
+  \ 100000\n  stage2_steps: 50000\n  stage3_steps: 50000\n  use_amp: 0\n  use_predicted_joints:\
+  \ 0\n  val_interval: 10000\n  weight_feature_rec: 0.0\n  weight_joints_consistency:\
+  \ 10.0\n  weight_joints_delta: 100.0\n  weight_kl: 1.0e-06\n  weight_latent_rec:\
+  \ 1.0\n  weight_orient_delta: 100.0\n  weight_rec: 1.0\n  weight_smpl_joints_rec:\
+  \ 10.0\n  weight_transl_delta: 100.0\nwandb_entity: interaction\nwandb_project_name:\
+  \ mld_vae\n"

mvae/mvae_smplh_hml3d_2_8_4/args_read.yaml

@@ -0,0 +1,62 @@
+data_args:
+  body_type: smplh
+  cfg_path: ./config_files/config_hydra/motion_primitive/hml_mp_h2_f8_r4.yaml
+  data_dir: ./data/hml3d_smplh/seq_data_zero_male/
+  dataset: hml3d
+  enforce_gender: male
+  enforce_zero_beta: 1
+  feature_dim: 276
+  future_length: 8
+  history_length: 2
+  num_primitive: 4
+  prob_static: 0.0
+  text_tolerance: 0.0
+  weight_scheme: uniform
+device: cuda
+exp_name: mvae_smplh_hml3d_2_8_4
+model_args:
+  activation: gelu
+  arch: all_encoder
+  dropout: 0.1
+  ff_size: 1024
+  h_dim: 256
+  latent_dim: !!python/tuple
+  - 1
+  - 256
+  nfeats: 276
+  normalize_before: false
+  num_heads: 4
+  num_layers: 7
+  position_embedding: learned
+save_dir: !!python/object/apply:pathlib.PosixPath
+- mvae
+- mvae_smplh_hml3d_2_8_4
+seed: 0
+torch_deterministic: true
+track: 1
+train_args:
+  anneal_lr: 1
+  batch_size: 128
+  ema_decay: 0.999
+  grad_clip: 1.0
+  learning_rate: 0.0001
+  log_interval: 1000
+  resume_checkpoint: null
+  save_interval: 50000
+  stage1_steps: 100000
+  stage2_steps: 50000
+  stage3_steps: 50000
+  use_amp: 0
+  use_predicted_joints: 0
+  val_interval: 10000
+  weight_feature_rec: 0.0
+  weight_joints_consistency: 10.0
+  weight_joints_delta: 100.0
+  weight_kl: 1.0e-06
+  weight_latent_rec: 1.0
+  weight_orient_delta: 100.0
+  weight_rec: 1.0
+  weight_smpl_joints_rec: 10.0
+  weight_transl_delta: 100.0
+wandb_entity: interaction
+wandb_project_name: mld_vae

mvae/mvae_smplh_hml3d_2_8_4/checkpoint_200000.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3db8211c2f83007a242c8d75c63c8618a88354439bd55816bb5b376d5a0e7261
+size 49577410