real_mutil_diffusion_policy_2.yaml

defaults:
  - _self_
  - task: real_hand_image

name: train_diffusion_unet_image
_target_: diffusion_policy.workspace.train_diffusion_real_mutil_workspace_2.TrainDiffusionUnetImageWorkspace

task_name: ${task.name}
shape_meta: ${task.shape_meta}
exp_name: "default"

horizon: 8
n_obs_steps: 8
n_action_steps: 6
n_latency_steps: 0
dataset_obs_steps: ${n_obs_steps}
past_action_visible: False
keypoint_visible_rate: 1.0
obs_as_global_cond: True

policy:
  _target_: diffusion_policy.policy.diffusion_real_mutil_policy_2.DiffusionUnetImagePolicy

  shape_meta: ${shape_meta}
  
  noise_scheduler:
    _target_: diffusers.schedulers.scheduling_ddim.DDIMScheduler
    num_train_timesteps: 100
    beta_start: 0.0001
    beta_end: 0.02
    # beta_schedule is important
    # this is the best we found
    beta_schedule: squaredcos_cap_v2
    clip_sample: True
    set_alpha_to_one: True
    steps_offset: 0
    prediction_type: sample #epsilon # or sample

  obs_encoder:
    _target_: diffusion_policy.model.vision.multi_image_obs_encoder.MultiImageObsEncoder
    shape_meta: ${shape_meta}
    rgb_model:
      _target_: diffusion_policy.model.vision.model_getter.get_resnet
      name: resnet18
      weights: null
    resize_shape: [240, 320]
    crop_shape: [216, 288] # ch, cw 240x320 90%
    random_crop: True
    use_group_norm: True
    share_rgb_model: False
    imagenet_norm: True

  image_encoder:
    _target_: diffusion_policy.model.mutil.diffusion_mutil_encoders.ImageEncoder
    out_dim: 108
    input_channels: 3

  image_decoder:
    _target_: diffusion_policy.model.mutil.diffusion_mutil_encoders.ImageDecoder
    input_dim: 64
    # output_size: (96, 96)
    output_size: 64
    # channels: [512, 256, 128, 64, 32]
    # attn_layers: [2, 3]
    # use_skip: True
    output_channels: 3

  torque_encoder:
    _target_: diffusion_policy.model.mutil.diffusion_mutil_encoders.TorqueEncoderWithLSTM
    input_dim: 6
    hidden_dim: 64
    output_channels: 3
    output_height: 64
    output_width: 64

  angle_encoder:
    _target_: diffusion_policy.model.mutil.diffusion_mutil_encoders.AngleEncoderWithLSTM
    input_dim: 6
    hidden_dim: 64
    output_channels: 3
    output_height: 64
    output_width: 64

  torque_decoder:
    _target_: diffusion_policy.model.mutil.diffusion_mutil_encoders.torqueMomentumDecoder
    in_channels: 9
    out_joints: 6

  angle_decoder:
    _target_: diffusion_policy.model.mutil.diffusion_mutil_encoders.angleMomentumDecoder
    in_channels: 9
    out_joints: 6

  horizon: ${horizon}
  n_action_steps: ${eval:'${n_action_steps}+${n_latency_steps}'}
  n_obs_steps: ${n_obs_steps}
  num_inference_steps: 100
  obs_as_global_cond: ${obs_as_global_cond}
  # crop_shape: null
  diffusion_step_embed_dim: 128
  down_dims: [512, 1024, 2048]
  kernel_size: 5
  n_groups: 8
  cond_predict_scale: True
  encoder_dim: 64
  # scheduler.step params
  # predict_epsilon: True

ema:
  _target_: diffusion_policy.model.diffusion.ema_model.EMAModel
  update_after_step: 0
  inv_gamma: 1.0
  power: 0.75
  min_value: 0.0
  max_value: 0.9999

dataloader:
  batch_size: 2
  num_workers: 4
  shuffle: True
  pin_memory: True
  persistent_workers: True

val_dataloader:
  batch_size: 16
  num_workers: 8
  shuffle: False
  pin_memory: True
  persistent_workers: True

optimizer:
  _target_: torch.optim.AdamW
  lr: 1.0e-4
  betas: [0.95, 0.999]
  eps: 1.0e-8
  weight_decay: 1.0e-6

task:
  dataset:
    _target_: diffusion_policy.dataset.hand_dataset.handDataset
    horizon: 8
    max_train_episodes: 90
    pad_after: 7
    pad_before: 1
    seed: 42
    val_ratio: 0.02
    log_file: data/data_csv_1/train_2.csv
    # val_csv: data/data_csv_1/val_1.csv
    # test_csv: data/data_csv_1/test_1.csv
    # data_folder: data/test_recordings
    data_folder: /home/yons/11_project/hand_dapg-master/see_hear_feel-master/data/test_recordings
    # data_folder: /home/nubot-11/Data/code/diffusion_policy-main/data/test_recordings
    # num_episode: 

training:
  device: "cuda:0"
  seed: 42
  debug: False
  resume: True
  # optimization
  lr_scheduler: cosine
  lr_warmup_steps: 500
  num_epochs: 2005
  gradient_accumulate_every: 1
  # EMA destroys performance when used with BatchNorm
  # replace BatchNorm with GroupNorm.
  use_ema: False
  freeze_encoder: False
  # training loop control
  # in epochs
  rollout_every: 50
  checkpoint_every: 50
  val_every: 1
  sample_every: 5
  # steps per epoch
  max_train_steps: null
  max_val_steps: null
  # misc
  tqdm_interval_sec: 1.0

logging:
  project: diffusion_policy_debug
  resume: True
  mode: online
  name: ${now:%Y.%m.%d-%H.%M.%S}_${name}_${task_name}
  tags: ["${name}", "${task_name}", "${exp_name}"]
  id: null
  group: null

checkpoint:
  topk:
    monitor_key: train_loss
    mode: min
    k: 5
    format_str: 'epoch={epoch:04d}-train_loss={train_loss:.3f}.ckpt'
  save_last_ckpt: True
  save_last_snapshot: False

multi_run:
  run_dir: data/outputs/${now:%Y.%m.%d}/${now:%H.%M.%S}_${name}_${task_name}
  wandb_name_base: ${now:%Y.%m.%d-%H.%M.%S}_${name}_${task_name}

hydra:
  job:
    override_dirname: ${name}
  run:
    dir: data/outputs/${now:%Y.%m.%d}/${now:%H.%M.%S}_${name}_${task_name}
  sweep:
    dir: data/outputs/${now:%Y.%m.%d}/${now:%H.%M.%S}_${name}_${task_name}
    subdir: ${hydra.job.num}