# @package _global_

# Defaults for training for the PushT dataset as per https://github.com/real-stanford/diffusion_policy.
# Note: We do not track EMA model weights as we discovered it does not improve the results. See
#       https://github.com/huggingface/lerobot/pull/134 for more details.

seed: 100000
dataset_repo_id: lerobot/pusht

training:
  offline_steps: 200000
  online_steps: 0
  eval_freq: 5000
  save_freq: 5000
  log_freq: 250
  save_model: true

  batch_size: 64
  grad_clip_norm: 10
  lr: 1.0e-4
  lr_scheduler: cosine
  lr_warmup_steps: 500
  adam_betas: [0.95, 0.999]
  adam_eps: 1.0e-8
  adam_weight_decay: 1.0e-6
  online_steps_between_rollouts: 1

  delta_timestamps:
    observation.image: "[i / ${fps} for i in range(1 - ${policy.n_obs_steps}, 1)]"
    observation.state: "[i / ${fps} for i in range(1 - ${policy.n_obs_steps}, 1)]"
    action: "[i / ${fps} for i in range(1 - ${policy.n_obs_steps}, 1 - ${policy.n_obs_steps} + ${policy.horizon})]"

eval:
  n_episodes: 50
  batch_size: 50

override_dataset_stats:
  # TODO(rcadene, alexander-soare): should we remove image stats as well? do we use a pretrained vision model?
  observation.image:
    mean: [[[0.5]], [[0.5]], [[0.5]]]  # (c,1,1)
    std: [[[0.5]], [[0.5]], [[0.5]]]  # (c,1,1)
  # TODO(rcadene, alexander-soare): we override state and action stats to use the same as the pretrained model
  # from the original codebase, but we should remove these and train our own pretrained model
  observation.state:
    min: [13.456424, 32.938293]
    max: [496.14618, 510.9579]
  action:
    min: [12.0, 25.0]
    max: [511.0, 511.0]

policy:
  name: diffusion

  # Input / output structure.
  n_obs_steps: 2
  horizon: 16
  n_action_steps: 8

  input_shapes:
    # TODO(rcadene, alexander-soare): add variables for height and width from the dataset/env?
    observation.image: [3, 96, 96]
    observation.state: ["${env.state_dim}"]
  output_shapes:
    action: ["${env.action_dim}"]

  # Normalization / Unnormalization
  input_normalization_modes:
    observation.image: mean_std
    observation.state: min_max
  output_normalization_modes:
    action: min_max

  # Architecture / modeling.
  # Vision backbone.
  vision_backbone: resnet18
  crop_shape: [84, 84]
  crop_is_random: True
  pretrained_backbone_weights: null
  use_group_norm: True
  spatial_softmax_num_keypoints: 32
  # Unet.
  down_dims: [512, 1024, 2048]
  kernel_size: 5
  n_groups: 8
  diffusion_step_embed_dim: 128
  use_film_scale_modulation: True
  # Noise scheduler.
  num_train_timesteps: 100
  beta_schedule: squaredcos_cap_v2
  beta_start: 0.0001
  beta_end: 0.02
  prediction_type: epsilon # epsilon / sample
  clip_sample: True
  clip_sample_range: 1.0

  # Inference
  num_inference_steps: 100

  # Loss computation
  do_mask_loss_for_padding: false