backup-0611: pref_stageb_b0_contact/config.yaml

pref_stageb_b0_contact/config.yaml

@@ -0,0 +1,68 @@
+datasets:
+  vla_data:
+    action_space: joint
+    cameras:
+    - head_camera
+    - left_camera
+    - right_camera
+    data_mix: contact_taskB_b0
+    data_root_dir: /mnt/localssd/kaiwenh/pref/data/0526/contact/taskB
+    dataset_py: pref_hdf5_stageb
+    future_action_window_size: 49
+    image_size:
+    - 224
+    - 224
+    include_state: false
+    num_workers: 4
+    past_action_window_size: 0
+    per_device_batch_size: 8
+    pref_category: contact
+    pref_keys:
+    - '25'
+    - '75'
+    split_val_fraction: 0.0
+    stats_json_path: examples/preference/dataset/stats_contact_0526_joint.json
+    task_groups:
+    - put_boxdrink3_plate
+    with_pref_suffix: false
+framework:
+  action_model:
+    action_dim: 14
+    action_hidden_dim: 2560
+    action_model_type: DiT-B
+    future_action_window_size: 49
+    past_action_window_size: 0
+  name: QwenOFT
+  qwenvl:
+    attn_implementation: sdpa
+    base_vlm: /mnt/localssd/kaiwenh/pref/Qwen3-VL-4B-Instruct
+output_dir: results/Checkpoints/pref_stageb_b0_contact
+run_id: pref_stageb_b0_contact
+run_root_dir: results/Checkpoints
+seed: 42
+trainer:
+  eval_interval: 100000
+  freeze_modules: ''
+  gradient_accumulation_steps: 1
+  gradient_clipping: 1.0
+  is_resume: false
+  learning_rate:
+    action_model: 1.0e-05
+    base: 1.0e-06
+    qwen_vl_interface: 1.0e-06
+  logging_frequency: 25
+  lr_scheduler_type: cosine_with_min_lr
+  max_train_steps: 2500
+  num_warmup_steps: 150
+  optimizer:
+    betas:
+    - 0.9
+    - 0.95
+    eps: 1.0e-08
+    weight_decay: 1.0e-08
+  pretrained_checkpoint: results/Checkpoints/pref_oft_baseline_contact_10k/checkpoints/steps_10000_pytorch_model.pt
+  save_interval: 250
+  scheduler_specific_kwargs:
+    min_lr: 1.0e-07
+wandb_entity: kaiwenh-17-uiuc
+wandb_project: pref-sim