datasets:
  vla_data:
    action_mode: abs
    data_mix: dynamic-329-v4
    data_root_dir: playground/Datasets/FastUMI
    dataset_py: lerobot_datasets
    image_size:
    - 224
    - 224
    per_device_batch_size: 8
    video_backend: torchvision_av
framework:
  action_model:
    action_dim: 10
    add_pos_embed: true
    decode_schedule: cosine
    diffusion_model_cfg:
      attention_head_dim: 64
      cross_attention_dim: 2048
      dropout: 0.2
      final_dropout: true
      input_embedding_dim: 2048
      interleave_self_attention: true
      norm_type: ada_norm
      num_attention_heads: 32
      num_layers: 36
      output_dim: 256
      positional_embeddings: sinusoidal
    future_action_window_size: 15
    l1_loss_weight: 0.1
    max_seq_len: 1024
    no_mask_token_prob: 0.0
    num_bins: 256
    num_inference_steps: 8
    num_target_vision_tokens: 32
    past_action_window_size: 0
    representation: bin
    state_dim: 10
    train_mask_schedule: cosine
    use_simple_max: false
  name: QwenDiscreteDiffusion
  qwenvl:
    attn_implementation: flash_attention_2
    base_vlm: playground/Pretrained_models/Qwen2.5-VL-3B-Instruct-Action
    num_vl_layers: 36
    vl_hidden_dim: 2048
output_dir: ./results/Checkpoints/fastumi_pickandplace_qwenDiscreteDiffusion_329v4
run_id: fastumi_pickandplace_qwenDiscreteDiffusion_329v4
run_root_dir: ./results/Checkpoints
seed: 42
trainer:
  eval_interval: 100
  freeze_modules: null
  gradient_accumulation_steps: 1
  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: 50
  lr_scheduler_type: cosine_with_min_lr
  max_train_steps: 20000
  num_warmup_steps: 5000
  optimizer:
    betas:
    - 0.9
    - 0.95
    eps: 1.0e-08
    weight_decay: 1.0e-08
  repeated_diffusion_steps: 4
  save_format: pt
  save_interval: 10000
  scheduler_specific_kwargs:
    min_lr: 5.0e-07
wandb_entity: 2200011093-peking-university
wandb_project: starVLA_FastUMI_dynamic_1