Add HandX diffusion checkpoints (layers 4/8/12)

layers12/config.yaml

@@ -0,0 +1,88 @@
+data:
+  _target_: src.diffusion.data_loader.handx.HandXDataset
+  contact_label: false
+  data_dir: data/testdata
+  data_file_name: can_pos_all_wotextfeat.npz
+  difference: false
+  fixed_length: 60
+  normalize: true
+  num_workers: 8
+  ratio: 1.0
+  repr: joint_pos_w_scalar_rot
+model:
+  activation: gelu
+  arch: trans_dec_treble_residual
+  cond_mask_prob: 0.1
+  cond_mode: text
+  contact_prediction: false
+  diffusion:
+    contact_loss: false
+    diffusion_steps: 1000
+    lambda_acce: 0.1
+    lambda_contact: 1.0
+    lambda_contact_predict: 1.0
+    lambda_fc: 0.0
+    lambda_ig: 0.0
+    lambda_rcxyz: 0.0
+    lambda_repr: 0.5
+    lambda_vel: 0.2
+    lambda_w_ig: 0.0
+    noise_schedule: cosine
+    repr: joint_pos_w_scalar_rot
+    sigma_small: true
+  dropout: 0.1
+  ff_size: 1024
+  latent_dim: 512
+  layers: 12
+  max_text_length: null
+  num_heads: 4
+  repr: joint_pos_w_scalar_rot
+  text_model: t5-base
+  treble_mask_prob: 1.0
+seed: 42
+train:
+  dataloader:
+    batch_size: 4096
+    num_workers: 16
+    shuffle: true
+  eval_cfg:
+    dataloader:
+      batch_size: 512
+      num_workers: 8
+      shuffle: false
+    eval_interval: 50000
+    num_samples_on_train: 2048
+    num_samples_on_val: 512
+    num_samples_per_condition: 16
+  eval_during_training: false
+  log_interval: 16
+  num_steps: 100000000
+  optimizer:
+    lr: 0.0001
+    weight_decay: 1.0e-05
+  overwrite: true
+  resume_checkpoint: ''
+  sample:
+    guidance_param: 2.5
+  save_dir: exps/data100_layers12
+  save_interval: 5000
+  train_platform_type: WandbPlatform
+  val_cfg:
+    dataloader:
+      batch_size: 4096
+      num_workers: 16
+      shuffle: false
+    val_interval: 2048
+  val_during_training: true
+  viz_cfg:
+    denoising_steps:
+    - 999
+    - 500
+    - 250
+    - 100
+    - 50
+    - 10
+    - 5
+    - 0
+    samples_count: 1
+  viz_during_training: true

layers12/model.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9247465c2277d85c40ec6de649dd26f7a94eccfc3ab356c451cdc1c32208e335
+size 176343850

layers4/config.yaml

@@ -0,0 +1,88 @@
+data:
+  _target_: src.diffusion.data_loader.handx.HandXDataset
+  contact_label: true
+  data_dir: data/testdata
+  data_file_name: can_pos_all_wotextfeat.npz
+  difference: false
+  fixed_length: 60
+  normalize: true
+  num_workers: 8
+  ratio: 1.0
+  repr: joint_pos_w_scalar_rot
+model:
+  activation: gelu
+  arch: trans_dec_treble_residual
+  cond_mask_prob: 0.1
+  cond_mode: text
+  contact_prediction: false
+  diffusion:
+    contact_loss: false
+    diffusion_steps: 1000
+    lambda_acce: 0.1
+    lambda_contact: 1.0
+    lambda_contact_predict: 1.0
+    lambda_fc: 0.0
+    lambda_ig: 0.0
+    lambda_rcxyz: 0.0
+    lambda_repr: 0.5
+    lambda_vel: 0.2
+    lambda_w_ig: 0.0
+    noise_schedule: cosine
+    repr: joint_pos_w_scalar_rot
+    sigma_small: true
+  dropout: 0.1
+  ff_size: 1024
+  latent_dim: 256
+  layers: 4
+  max_text_length: null
+  num_heads: 4
+  repr: joint_pos_w_scalar_rot
+  text_model: t5-base
+  treble_mask_prob: 1.0
+seed: 42
+train:
+  dataloader:
+    batch_size: 6144
+    num_workers: 8
+    shuffle: true
+  eval_cfg:
+    dataloader:
+      batch_size: 512
+      num_workers: 8
+      shuffle: false
+    eval_interval: 50000
+    num_samples_on_train: 2048
+    num_samples_on_val: 512
+    num_samples_per_condition: 16
+  eval_during_training: false
+  log_interval: 16
+  num_steps: 100000000
+  optimizer:
+    lr: 0.0001
+    weight_decay: 1.0e-05
+  overwrite: true
+  resume_checkpoint: ''
+  sample:
+    guidance_param: 2.5
+  save_dir: exps/data100_layers4
+  save_interval: 5000
+  train_platform_type: WandbPlatform
+  val_cfg:
+    dataloader:
+      batch_size: 6144
+      num_workers: 8
+      shuffle: false
+    val_interval: 1000
+  val_during_training: true
+  viz_cfg:
+    denoising_steps:
+    - 999
+    - 500
+    - 250
+    - 100
+    - 50
+    - 10
+    - 5
+    - 0
+    samples_count: 1
+  viz_during_training: true

layers4/model.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:47d5fd2b55439ab6a57031951a31a95cc2243f4f3a9ad6aa5058b562a2682960
+size 28787534

layers8/config.yaml

@@ -0,0 +1,88 @@
+data:
+  _target_: src.diffusion.data_loader.handx.HandXDataset
+  contact_label: true
+  data_dir: data/testdata
+  data_file_name: can_pos_all_wotextfeat.npz
+  difference: false
+  fixed_length: 60
+  normalize: true
+  num_workers: 8
+  ratio: 1.0
+  repr: joint_pos_w_scalar_rot
+model:
+  activation: gelu
+  arch: trans_dec_treble_residual
+  cond_mask_prob: 0.1
+  cond_mode: text
+  contact_prediction: false
+  diffusion:
+    contact_loss: false
+    diffusion_steps: 1000
+    lambda_acce: 0.1
+    lambda_contact: 1.0
+    lambda_contact_predict: 1.0
+    lambda_fc: 0.0
+    lambda_ig: 0.0
+    lambda_rcxyz: 0.0
+    lambda_repr: 0.5
+    lambda_vel: 0.2
+    lambda_w_ig: 0.0
+    noise_schedule: cosine
+    repr: joint_pos_w_scalar_rot
+    sigma_small: true
+  dropout: 0.1
+  ff_size: 1024
+  latent_dim: 512
+  layers: 8
+  max_text_length: null
+  num_heads: 4
+  repr: joint_pos_w_scalar_rot
+  text_model: t5-base
+  treble_mask_prob: 1.0
+seed: 42
+train:
+  dataloader:
+    batch_size: 4096
+    num_workers: 8
+    shuffle: true
+  eval_cfg:
+    dataloader:
+      batch_size: 512
+      num_workers: 8
+      shuffle: false
+    eval_interval: 50000
+    num_samples_on_train: 2048
+    num_samples_on_val: 512
+    num_samples_per_condition: 16
+  eval_during_training: false
+  log_interval: 16
+  num_steps: 100000000
+  optimizer:
+    lr: 0.0001
+    weight_decay: 1.0e-05
+  overwrite: true
+  resume_checkpoint: ''
+  sample:
+    guidance_param: 2.5
+  save_dir: exps/data100_layers8
+  save_interval: 5000
+  train_platform_type: WandbPlatform
+  val_cfg:
+    dataloader:
+      batch_size: 4096
+      num_workers: 8
+      shuffle: false
+    val_interval: 1000
+  val_during_training: true
+  viz_cfg:
+    denoising_steps:
+    - 999
+    - 500
+    - 250
+    - 100
+    - 50
+    - 10
+    - 5
+    - 0
+    samples_count: 1
+  viz_during_training: true

layers8/model.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cde0968c91854a903dc88236cba809acfb3e8f860e3663db9a0e251ea1221c36
+size 125849906