defaults:
  - _self_
hydra:
  run:  
    dir: ${logdir}
_target_: agent.finetune.train_ppo_gaussian_img_agent.TrainPPOImgGaussianAgent

name: ${env_name}_ft_gaussian_mlp_img_ta${horizon_steps}
logdir: ${oc.env:DPPO_LOG_DIR}/robomimic-finetune/${name}/${now:%Y-%m-%d}_${now:%H-%M-%S}_${seed}
base_policy_path: ${oc.env:DPPO_LOG_DIR}/robomimic-pretrain/can/can_pre_gaussian_mlp_img_ta4/2024-07-28_21-54-40/checkpoint/state_1000.pt
robomimic_env_cfg_path: cfg/robomimic/env_meta/${env_name}-img.json
normalization_path: ${oc.env:DPPO_DATA_DIR}/robomimic/${env_name}-img/normalization.npz

seed: 42
device: cuda:0
env_name: can
obs_dim: 9
action_dim: 7
transition_dim: ${action_dim}
cond_steps: 1
img_cond_steps: 1
horizon_steps: 4
act_steps: 4

env:
  n_envs: 50
  name: ${env_name}
  best_reward_threshold_for_success: 1
  max_episode_steps: 300
  save_video: False
  use_image_obs: True
  wrappers:
    robomimic_image:
      normalization_path: ${normalization_path}
      low_dim_keys: ['robot0_eef_pos',
                     'robot0_eef_quat',
                     'robot0_gripper_qpos']
      image_keys: ['robot0_eye_in_hand_image']
      shape_meta: ${shape_meta}
    multi_step:
      n_obs_steps: ${cond_steps}
      n_action_steps: ${act_steps}
      max_episode_steps: ${env.max_episode_steps}
      reset_within_step: True

shape_meta:
  obs:
    rgb:
      shape: [3, 96, 96]
    state:
      shape: [9]
  action: 
    shape: [7]

wandb:
  entity: ${oc.env:DPPO_WANDB_ENTITY}
  project: robomimic-${env_name}-finetune
  run: ${now:%H-%M-%S}_${name}

train:
  n_train_itr: 200
  n_critic_warmup_itr: 2
  n_steps: 300
  gamma: 0.999
  augment: True
  grad_accumulate: 5
  actor_lr: 1e-5
  actor_weight_decay: 0
  actor_lr_scheduler:
    first_cycle_steps: 200
    warmup_steps: 10
    min_lr: 1e-5
  critic_lr: 1e-3
  critic_weight_decay: 0
  critic_lr_scheduler:
    first_cycle_steps: 200
    warmup_steps: 10
    min_lr: 1e-3
  save_model_freq: 100
  val_freq: 10
  render:
    freq: 1
    num: 0
  # PPO specific
  reward_scale_running: True
  reward_scale_const: 1.0
  gae_lambda: 0.95
  batch_size: 1500
  logprob_batch_size: 1000
  update_epochs: 10
  vf_coef: 0.5
  target_kl: 1

model:
  _target_: model.rl.gaussian_ppo.PPO_Gaussian
  clip_ploss_coef: 0.01
  randn_clip_value: 3
  network_path: ${base_policy_path}
  actor:
    _target_: model.common.mlp_gaussian.Gaussian_VisionMLP
    backbone:
      _target_: model.common.vit.VitEncoder
      obs_shape: ${shape_meta.obs.rgb.shape}
      num_channel: ${eval:'${shape_meta.obs.rgb.shape[0]} * ${img_cond_steps}'} # each image patch is history concatenated
      cfg:
        patch_size: 8
        depth: 1
        embed_dim: 128
        num_heads: 4
        embed_style: embed2
        embed_norm: 0
    augment: False
    spatial_emb: 128
    mlp_dims: [512, 512, 512]
    residual_style: True
    fixed_std: 0.1
    learn_fixed_std: True
    std_min: 0.01
    std_max: 0.2
    img_cond_steps: ${img_cond_steps}
    cond_dim: ${eval:'${obs_dim} * ${cond_steps}'}
    horizon_steps: ${horizon_steps}
    transition_dim: ${transition_dim}
  critic:
    _target_: model.common.critic.ViTCritic
    spatial_emb: 128
    augment: False
    backbone:
      _target_: model.common.vit.VitEncoder
      obs_shape: ${shape_meta.obs.rgb.shape}
      num_channel: ${eval:'${shape_meta.obs.rgb.shape[0]} * ${img_cond_steps}'} # each image patch is history concatenated
      cfg:
        patch_size: 8
        depth: 1
        embed_dim: 128
        num_heads: 4
        embed_style: embed2
        embed_norm: 0
    img_cond_steps: ${img_cond_steps}
    cond_dim: ${eval:'${obs_dim} * ${cond_steps}'}
    mlp_dims: [256, 256, 256]
    activation_type: Mish
    residual_style: True
  horizon_steps: ${horizon_steps}
  device: ${device}