Upload config.yaml with huggingface_hub

config.yaml

@@ -0,0 +1,72 @@
+datasets:
+  vla_data:
+    action_mode: abs
+    data_mix: robotwin_all_50
+    data_root_dir: playground/Datasets/RoboTwin
+    dataset_py: lerobot_datasets
+    obs_image_size:
+    - 224
+    - 224
+    per_device_batch_size: 8
+    video_backend: pyav
+framework:
+  action_model:
+    action_dim: 14
+    action_horizon: 50
+    add_pos_embed: true
+    diffusion_model_cfg:
+      action_dit_hidden_dim: 1024
+      attention_head_dim: 64
+      cross_attention_dim: 1024
+      dropout: 0.2
+      final_dropout: true
+      input_embedding_dim: 1024
+      interleave_self_attention: true
+      norm_type: ada_norm
+      num_attention_heads: 16
+      num_layers: 28
+      output_dim: 2560
+      positional_embeddings: null
+    max_seq_len: 1024
+    noise_beta_alpha: 1.5
+    noise_beta_beta: 1.0
+    noise_s: 0.999
+    num_inference_timesteps: 10
+    num_target_vision_tokens: 32
+    num_timestep_buckets: 1000
+    state_dim: 14
+  name: QwenPI_v3
+  qwenvl:
+    attn_implementation: flash_attention_2
+    base_vlm: playground/Pretrained_models/Qwen/Qwen3-VL-2B-Instruct
+    num_vl_layers: 28
+    vl_hidden_dim: 2048
+output_dir: ./results/Checkpoints/0601_robotwin_all_50_qwen3piv3_all
+run_id: 0601_robotwin_all_50_qwen3piv3_all
+run_root_dir: ./results/Checkpoints
+seed: 42
+trainer:
+  eval_batch_size: 8
+  eval_interval: 1000
+  eval_num_samples: 256
+  freeze_modules: true
+  gradient_clipping: 1.0
+  is_resume: true
+  learning_rate:
+    action_model: 0.0001
+    base: 1.0e-05
+    qwen_vl_interface: 1.0e-05
+  logging_frequency: 100
+  lr_scheduler_type: cosine_with_min_lr
+  max_train_steps: 150000
+  num_warmup_steps: 5000
+  optimizer:
+    betas:
+    - 0.9
+    - 0.95
+    eps: 1.0e-08
+    weight_decay: 1.0e-08
+  save_interval: 10000
+  scheduler_specific_kwargs:
+    min_lr: 5.0e-07
+wandb_project: starVLA_Robotwin