defaults:
  - _self_
hydra:
  run:
    dir: ${logdir}
_target_: agent.finetune.train_calql_agent.TrainCalQLAgent

name: ${env_name}_calql_mlp_ta${horizon_steps}
logdir: ${oc.env:DPPO_LOG_DIR}/robomimic-finetune/${name}/${now:%Y-%m-%d}_${now:%H-%M-%S}_${seed}
base_policy_path: ${oc.env:DPPO_LOG_DIR}/robomimic-pretrain/square_calql_mlp_ta1/2024-10-25_22-44-12_42/checkpoint/state_999.pt
robomimic_env_cfg_path: cfg/robomimic/env_meta/${env_name}.json
normalization_path: ${oc.env:DPPO_DATA_DIR}/robomimic/${env_name}/normalization.npz
offline_dataset_path: ${oc.env:DPPO_DATA_DIR}/robomimic/${env_name}/train.npz

seed: 42
device: cuda:0
env_name: square
obs_dim: 23
action_dim: 7
cond_steps: 1
horizon_steps: 1
act_steps: 1

env:
  n_envs: 1
  name: ${env_name}
  best_reward_threshold_for_success: 1
  max_episode_steps: 400
  reset_at_iteration: False
  save_video: False
  wrappers:
    robomimic_lowdim:
      normalization_path: ${normalization_path}
      low_dim_keys: ['robot0_eef_pos',
                    'robot0_eef_quat',
                    'robot0_gripper_qpos',
                    'object'] # same order of preprocessed observations
    multi_step:
      n_obs_steps: ${cond_steps}
      n_action_steps: ${act_steps}
      max_episode_steps: ${env.max_episode_steps}
      reset_within_step: True

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

train:
  n_train_itr: 10000
  n_steps: 1  # not used
  n_episode_per_epoch: 1
  gamma: 0.99
  actor_lr: 1e-4
  actor_weight_decay: 0
  actor_lr_scheduler:
    first_cycle_steps: 1000
    warmup_steps: 10
    min_lr: 1e-4
  critic_lr: 3e-4
  critic_weight_decay: 0
  critic_lr_scheduler:
    first_cycle_steps: 1000
    warmup_steps: 10
    min_lr: 3e-4
  save_model_freq: 100
  val_freq: 10
  render:
    freq: 1
    num: 0
  log_freq: 1
  # CalQL specific
  train_online: True
  batch_size: 256
  n_random_actions: 4
  target_ema_rate: 0.005
  scale_reward_factor: 1.0
  num_update: 1000
  buffer_size: 1000000
  online_utd_ratio: 1
  n_eval_episode: 40
  n_explore_steps: 0
  target_entropy: ${eval:'- ${action_dim} * ${act_steps}'}
  init_temperature: 1
  automatic_entropy_tuning: True

model:
  _target_: model.rl.gaussian_calql.CalQL_Gaussian
  randn_clip_value: 3
  cql_min_q_weight: 5.0
  tanh_output: True
  network_path: ${base_policy_path}
  actor:
    _target_: model.common.mlp_gaussian.Gaussian_MLP
    mlp_dims: [512, 512, 512]
    activation_type: ReLU
    tanh_output: False  # squash after sampling instead
    cond_dim: ${eval:'${obs_dim} * ${cond_steps}'}
    horizon_steps: ${horizon_steps}
    action_dim: ${action_dim}    
    std_max: 7.3891
    std_min: 0.0067
  critic:
    _target_: model.common.critic.CriticObsAct
    mlp_dims: [256, 256, 256]
    activation_type: ReLU
    use_layernorm: True
    double_q: True
    cond_dim: ${eval:'${obs_dim} * ${cond_steps}'}
    action_dim: ${action_dim}
    action_steps: ${act_steps}
  horizon_steps: ${horizon_steps}
  device: ${device}

offline_dataset:
  _target_: agent.dataset.sequence.StitchedSequenceQLearningDataset
  dataset_path: ${offline_dataset_path}
  horizon_steps: ${horizon_steps}
  cond_steps: ${cond_steps}
  device: ${device}
  discount_factor: ${train.gamma}
  get_mc_return: True