Upload folder using huggingface_hub

pretrain_droid/checkpoints/weights/step_00005000.pt

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:c851c88acb19d27361973fbbb6e45549bc4db9089246e4de90a8c3d7895b17a2
+oid sha256:69bbd5234f1ea2943329bd28858c38662604ab13e778b4d81b7bf42753ff333a
 size 10894838421

pretrain_droid/config.yaml

@@ -1,4 +1,4 @@
-output_dir: ./runs/cotrain_droid_teleop/2026-07-04_01-48-06
+output_dir: ./runs/cotrain_droid_teleop/2026-07-05_01-26-59
 batch_size: 8
 num_workers: 4
 prefetch_factor: 4
@@ -6,7 +6,7 @@ lr_scheduler_type: constant
 learning_rate: 1.0e-05
 num_epochs: null
 log_every: 10
-save_every: 5000
+save_every: 2500
 save_accelerator_state: true
 eval_every: 1000
 eval_num_inference_samples: 8
@@ -17,7 +17,7 @@ mixed_precision: bf16
 seed: 42
 max_grad_norm: 1.0
 weight_decay: 0.001
-resume: runs/cotrain_droid_teleop/video_only/checkpoints/weights/step_00015000.pt
+resume: runs/cotrain_droid_teleop/video_only/checkpoints/weights/step_00015000_adim10.pt
 resume_reset: true
 train_mode: joint
 wandb:
@@ -30,135 +30,66 @@ wandb:
 data:
   train:
     _target_: fastwam.datasets.lerobot.lerobot_pretrain_dataset_v3.LerobotPretrainDatasetV3
-    data_mix:
-      teleop: 1.0
-    data_root_per_dataset:
-      droid: ./data/datasets/lerobot_v3/droid
-      teleop: data/datasets/teleop/pnp202/train
+    datasets:
+      teleop:
+        data_root: data/datasets/teleop/pnp202/train
+        weight: 1.0
+        fps: 15
+        processor:
+          _target_: fastwam.datasets.lerobot.processors.TeleopProcessor
+          required_keys:
+            video:
+              primary: observation.images.right_3pv_camera
+              wrist: observation.images.wrist_camera
+            state:
+            - state.joint_pos
+            - state.gripper_state
+            action:
+            - state.eef_xyz
+            - state.eef_rot6d
+            - state.gripper_state
+          frame_size:
+          - 256
+          - 320
+          norm_stats_path: data/norm_stats/pnp202_train_fps15.json
+          normalize: true
+          image_aug: true
+    mixture_norm_stats_path: null
+    mixture_normalize: false
     num_frames: 17
     action_video_freq_ratio: 2
-    frame_size:
-    - 256
-    - 320
-    camera_mapping_per_dataset:
-      droid:
-        primary: observation.images.exterior_image_1_left
-        wrist: observation.images.wrist_image_left
-      teleop:
-        primary: observation.images.right_3pv_camera
-        wrist: observation.images.wrist_camera
-    action_mapping_per_dataset:
-      droid:
-        skip: 10
-      teleop:
-        action.body_eef_xyz: 3
-        action.body_eef_rot6d: 6
-        action.gripper_state: 1
-    proprio_mapping_per_dataset:
-      droid:
-        skip: 8
-      teleop:
-        state.joint_pos: 7
-        state.gripper_state: 1
-    action_norm_mask_per_dataset:
-      droid:
-        action:
-        - true
-        - true
-        - true
-        - true
-        - true
-        - true
-        - false
-        - false
-      teleop:
-        action.body_eef_xyz:
-        - true
-        - true
-        - true
-        action.body_eef_rot6d:
-        - true
-        - true
-        - true
-        - true
-        - true
-        - true
-        action.gripper_state:
-        - false
-    proprio_norm_mask_per_dataset:
-      droid:
-        action:
-        - true
-        - true
-        - true
-        - true
-        - true
-        - true
-        - false
-        - false
-      teleop:
-        state.joint_pos:
-        - true
-        - true
-        - true
-        - true
-        - true
-        - true
-        - true
-        state.gripper_state:
-        - false
     emit_raw_video: true
     context_len: 128
-    image_aug: false
     sampler_type: hierarchical_chunk
     episode_weighting: uniform
     sampler_num_samples: 1000000000
   val:
     _target_: fastwam.datasets.lerobot.lerobot_dataset_v3.LerobotDatasetV3
     data_root: data/datasets/teleop/pnp202/val
+    fps: 15
+    processor:
+      _target_: fastwam.datasets.lerobot.processors.TeleopProcessor
+      required_keys:
+        video:
+          primary: observation.images.right_3pv_camera
+          wrist: observation.images.wrist_camera
+        state:
+        - state.joint_pos
+        - state.gripper_state
+        action:
+        - state.eef_xyz
+        - state.eef_rot6d
+        - state.gripper_state
+      frame_size:
+      - 256
+      - 320
+      norm_stats_path: data/norm_stats/pnp202_train_fps15.json
+      normalize: true
+      image_aug: false
     num_frames: 17
     action_video_freq_ratio: 2
-    frame_size:
-    - 256
-    - 320
-    camera_mapping:
-      primary: observation.images.right_3pv_camera
-      wrist: observation.images.wrist_camera
-    action_mapping:
-      action.body_eef_xyz: 3
-      action.body_eef_rot6d: 6
-      action.gripper_state: 1
-    proprio_mapping:
-      state.joint_pos: 7
-      state.gripper_state: 1
-    action_norm_mask:
-      action.body_eef_xyz:
-      - true
-      - true
-      - true
-      action.body_eef_rot6d:
-      - true
-      - true
-      - true
-      - true
-      - true
-      - true
-      action.gripper_state:
-      - false
-    proprio_norm_mask:
-      state.joint_pos:
-      - true
-      - true
-      - true
-      - true
-      - true
-      - true
-      - true
-      state.gripper_state:
-      - false
     emit_video: true
     context_len: 128
-    image_aug: false
     sampler_type: resumable_epoch
 model:
   _target_: fastwam.runtime.create_fastwam_joint
@@ -167,10 +98,9 @@ model:
   tokenizer_max_len: 128
   load_text_encoder: false
   redirect_common_files: false
-  mot_checkpoint_mixed_attn: true
+  mot_checkpoint_mixed_attn: false
   action_dit_pretrained_path: checkpoints/ActionDiT_linear_interp_Wan22_alphascale_512hdim_adim10.pt
   skip_dit_load_from_pretrain: false
-  checkpoint_video_dit_only: true
   num_frames: 17
   video_dit_config:
     patch_size:
@@ -187,7 +117,7 @@ model:
     attn_head_dim: 128
     num_layers: 30
     eps: 1.0e-06
-    use_gradient_checkpointing: true
+    use_gradient_checkpointing: false
   action_dit_config:
     proprio_dim: 8
     action_dim: 10
@@ -199,7 +129,7 @@ model:
     text_dim: 4096
     freq_dim: 256
     eps: 1.0e-06
-    use_gradient_checkpointing: true
+    use_gradient_checkpointing: false
   video_scheduler:
     train_shift: 3.0
     infer_shift: null