defaults.yaml

defaults:
  # Job
  n_steps: 100_000_000
  n_env_steps: 1_000_000_000
  offline_data_dir:
  offline_prefill_dir:
  offline_test_dir:
  offline_eval_dir:
  log_interval: 100
  logbatch_interval: 1000
  save_interval: 500
  eval_interval: 2000
  data_workers: 4
  allow_mid_reset: True
  enable_profiler: False
  verbose: False

  # Features
  image_key: image
  image_size: 64
  image_channels: 3
  image_categorical: False
  action_dim: 0
  clip_rewards:
  
  # Probe features
  map_key:
  map_size: 0
  map_channels: 0
  map_categorical: True
  goals_size: 0
  
  # Training
  buffer_size: 10_000_000
  buffer_size_offline: 0
  reset_interval: 200
  iwae_samples: 1
  kl_balance: 0.8
  kl_weight: 1.0
  image_weight: 1.0
  vecobs_weight: 1.0
  reward_weight: 1.0
  terminal_weight: 1.0
  adam_lr: 3.0e-4
  adam_lr_actor: 1.0e-4
  adam_lr_critic: 1.0e-4
  adam_eps: 1.0e-5
  keep_state: True
  batch_length: 48
  batch_size: 32
  device: "cuda:0"
  grad_clip: 200
  grad_clip_ac: 200
  image_decoder_min_prob: 0
  amp: False  # AMP no longer helps on A100
  probe_gradients: False  # for baselines

  # Eval
  test_batches: 61  # For unbiased test needs to be enough to cover full episodes
  test_batch_size: 10  # Smaller for faster test
  test_save_size: 1
  eval_batches: 61  # Big enough to reach episode end (xN). +1 to log last episode
  eval_samples: 1  # no multi-sampling to have reproducible paper results
  eval_batch_size: 32  # Limited by GPU mem
  eval_save_size: 1  # How many envs to save in npz

  # Model
  model: dreamer
  deter_dim: 2048
  stoch_dim: 32
  stoch_discrete: 32
  hidden_dim: 1000
  gru_layers: 1
  gru_type: gru
  layer_norm: True
  vecobs_size: 0
  image_encoder: cnn
  cnn_depth: 48
  image_encoder_layers: 0
  image_decoder: cnn
  image_decoder_layers: 0
  reward_input: False
  reward_decoder_layers: 4
  reward_decoder_categorical:
  terminal_decoder_layers: 4
  
  # Probe
  probe_model: none
  map_decoder: dense
  map_hidden_layers: 4
  map_hidden_dim: 1024

  # Actor Critic
  gamma: 0.995
  lambda_gae: 0.95
  entropy: 0.003
  target_interval: 100
  imag_horizon: 15
  actor_grad: reinforce
  actor_dist: onehot

  # Auxiliary Actor critic
  aux_critic: False
  aux_critic_weight: 1.0
  gamma_aux: 0.99
  lambda_gae_aux: 0.95
  target_interval_aux: 1000

  # Generator
  generator_workers: 8        # train+eval generators
  generator_workers_train: 0  # train-only generators
  generator_workers_eval: 0   # eval-only generators
  generator_prefill_steps: 50_000
  generator_prefill_policy: random
  limit_step_ratio: 0
  env_id:
  env_id_eval:
  env_action_repeat: 1
  env_time_limit: 0
  env_no_terminal: False

minigrid:
  env_id: MiniGrid-MazeS11N-v0
  # Features
  image_key: image
  image_size: 7
  image_channels: 4
  image_categorical: True
  map_key: map
  map_size: 11
  map_channels: 4
  map_categorical: True
  action_dim: 7
  reward_input: True
  # Model
  image_encoder: dense
  image_encoder_layers: 3
  image_decoder: dense
  image_decoder_layers: 2
  probe_model: map
  imag_horizon: 1

miniworld:
  # Env
  env_id: MiniWorld-ScavengerHuntSmall-v0
  env_no_terminal: True
  map_size: 9
  map_channels: 14
  action_dim: 3
  # Features
  image_key: image
  image_size: 64
  image_channels: 3
  image_categorical: False
  map_key: map
  map_categorical: True
  reward_input: True
  # Model
  probe_model: goals
  cnn_depth: 32
  data_workers: 6

miniworld_offline:
  # Env
  env_id: MiniWorld-ScavengerHuntSmall-v0
  env_no_terminal: True
  map_size: 9
  map_channels: 14
  action_dim: 3
  # Features
  image_key: image
  image_size: 64
  image_channels: 3
  image_categorical: False
  map_key: map
  map_categorical: True
  reward_input: True
  # Model
  probe_model: goals
  cnn_depth: 32
  data_workers: 6
  # Offline
  allow_mid_reset: False
  reward_decoder_layers: 1  # faster
  terminal_decoder_layers: 1  # faster
  imag_horizon: 1  # faster

atari:
  n_env_steps: 200_000_000
  # Env
  env_id: Atari-Pong
  action_dim: 18
  env_action_repeat: 4
  env_time_limit: 27_000  # =108_000/action_repeat = 30 minutes of game play.
  generator_workers: 1
  clip_rewards: tanh
  # Model
  deter_dim: 1024
  kl_weight: 0.1
  gamma: 0.99
  entropy: 0.001

dmc:
  env_id: DMC-quadruped_run
  action_dim: 12
  env_time_limit: 500  # =1000/action_repeat
  env_action_repeat: 2
  generator_workers: 1
  generator_prefill_steps: 5000
  entropy: 1.0e-4
  actor_grad: dynamics
  actor_dist: tanh_normal
  clip_rewards: tanh

dmlab:
  env_id: DmLab-rooms_watermaze
  action_dim: 15
  env_action_repeat: 4
  env_no_terminal: True
  kl_weight: 3.0

dmmemory:
  env_id: DMM-spot_diff_passive_train
  action_dim: 9
  kl_weight: 3.0

dmlab_offline:
  env_id: DmLab-rooms_watermaze
  action_dim: 15
  env_action_repeat: 4
  env_no_terminal: True
  kl_weight: 1.0
  imag_horizon: 5
  n_steps: 225_000

vectorenv:
  env_id: CartPole-v0
  action_dim: 2
  vecobs_size: 4
  image_key:
  image_encoder:
  image_decoder:

minecraft:
  env_id: Embodied-minecraft_diamond
  action_dim: 29
  vecobs_size: 27
  clip_rewards: log1p

memmaze:
  env_id: memory_maze:MemoryMaze-9x9-v0
  action_dim: 6
  entropy: 0.001

procgen:
  env_id: procgen:procgen-coinrun-v0
  action_dim: 15
  kl_weight: 0.1

debug:
  device: cpu
  log_interval: 5
  save_interval: 10
  data_workers: 0
  generator_workers: 1
  generator_prefill_steps: 10_000
  batch_length: 15
  batch_size: 5
  imag_horizon: 3