inference:
    action_horizon: 32
    num_inference_steps: 10
    robotwin_camera_layout: four_grid_288x384
    image_crop_scale: 0.95
    processor:
      norm_default_mode: z-score
      use_stepwise_action_norm: false
      val_transforms:
      - _target_: pixelwam.datasets.lerobot.transforms.image.ToTensor
      - _target_: torchvision.transforms.Resize
        size:
        - 240
        - 320
model:
  mask_scheme: default
  proprio_dim: 14
  model_dtype: bf16
  video_scheduler:
    infer_shift: 3.0
    num_train_timesteps: 1000
  action_scheduler:
    infer_shift: 5.0
    num_train_timesteps: 1000
  model_path: ${oc.env:PIXELWAM_MODEL_PATH}
  source_path: ${oc.env:PIXELWAM_SOURCE_PATH}
  numerical_mode: legacy_bf16
  joint_layout: legacy
  semantic_image_mode: legacy_384
  target_image_mode: mosaic
  action_backbone: shared
  sdpa_backend: auto
  action_dim: 14
  max_action_horizon: 64
  patch_size: 16
  patch_weight_pooling: target_only_global_norm
  use_gradient_checkpointing: false
  action_prediction_mode: velocity
  reference_image_mode: mosaic
  copy_gate:
    enabled: false
  forward_chunk_size: 64
  portable_inference_rope: true
  vision_attention_mode: grouped_cached
  action_expert_stability_fp32: true
  _target_: pixelwam.inference.create_model