policy:
  name: acp
  _target_: force_il.policies.acp.ACPPolicy
  action_dim: 19
  pred_horizon: 16
  obs_horizon: 2
  action_horizon: 16
  vision_encoder_name: vit_base_patch32_clip_224.openai
  vision_pretrained: true
  vision_output_dim: 768
  wrench_encoding: causalconv
  wrench_window_size: 256
  wrench_freq: 5000.0
  wrench_output_dim: 384
  fuse_mode: modality-attention
  obs_num_rgb_frames: 2
  include_low_dim: true
  low_dim_eef_pos_horizon: 3
  low_dim_eef_pos_dim: 3
  low_dim_eef_rot_horizon: 3
  low_dim_eef_rot_dim: 6
  num_diffusion_steps: 50
  noise_scheduler: ddim
  unet_down_dims:
  - 256
  - 512
  - 1024
  diffusion_step_embed_dim: 32
  unet_kernel_size: 5
  unet_n_groups: 8
  cond_predict_scale: true
  input_perturbation: 0.1
  unified_action: true
  unified_action_dim: 19
  pose_dim: 9
  vt_dim: 9
  stiffness_dim: 1
  predict_virtual_target: false
  predict_stiffness: false
  default_stiffness: 100.0
  num_inference_steps: 16
task:
  name: flip_up
  dataset_path: data/flip_up_230/flip_up_new_v5
  action_dim: 19
  obs_horizon: 2
  eef_obs_horizon: 3
  eef_down_sample_steps: 5
  pred_horizon: 16
  action_horizon: 16
  action_down_sample_steps: 50
  wrench_horizon: 32
  wrench_down_sample_steps: 4
  query_down_sample_steps: 8
  image_size: 224
  val_ratio: 0.05
data:
  _target_: force_il.data.acp_dataset.ACPEpisodeDataset
  dataset_path: data/flip-20e_acp
  pred_horizon: 16
  obs_horizon: 2
  rgb_down_sample_steps: 10
  eef_obs_horizon: 3
  eef_down_sample_steps: 5
  action_down_sample_steps: 50
  wrench_horizon: 32
  wrench_down_sample_steps: 4
  query_down_sample_steps: 8
  image_size: 224
  val_ratio: 0.05
  seed: 42
  is_val: false
  relative_actions: true
train:
  epochs: 300
  batch_size: 128
  gradient_accumulation: 1
  optimizer:
    _target_: torch.optim.AdamW
    lr: 0.0003
    weight_decay: 1.0e-06
    betas:
    - 0.95
    - 0.999
  scheduler:
    type: cosine
    warmup_steps: 2000
    min_lr: 1.0e-06
  encoder_lr_scale: 0.1
  mixed_precision: fp16
  use_ema: true
  ema_decay: 0.9999
  ema_power: 0.75
  ema_update_after_step: 0
  log_interval: 50
  use_wandb: true
  wandb_project: force_il
  wandb_name: null
  val_every: 5
  sample_every: 5
  val_interval: 500
  checkpoint_interval: 10
  save_best: true
  best_checkpoint_monitor: train_loss
  num_workers: 8
  pin_memory: true
  prefetch_factor: 2
  normalizer_use_full_dataset: false
  upload_to_hf: true
  hf_repo_id: tangPR/acp-flip-up
  hf_upload_interval: 100
  seed: 42
experiment_name: acp_flip_up
seed: 42
validate:
  checkpoint: null
  num_samples: 5
  output_dir: ./validation_plots
  use_train_split: false
output_dir: outputs/2026-03-23/acp_flip_up_205858
checkpoint_dir: outputs/2026-03-23/acp_flip_up_205858/checkpoints
log_dir: outputs/2026-03-23/acp_flip_up_205858/logs