configs.yaml

name: Dreamer
env: quadruped_run
seed: 42

ckpt_callback:
  save_top_k: 2
  mode: min
  monitor: loss
  save_on_train_epoch_end: True
  save_last: True

  trainer_params:
  default_root_dir: None
  gpus: 1
  gradient_clip_val: 100.0
  val_check_interval: 5
  max_epochs: 1000

dreamer:
  td_model_lr: 5.0e-4
  actor_lr: 8.0e-5
  critic_lr: 8.0e-5
  default_lr: 5.0e-4
  weight_decay: 1.0e-6
  
  batch_size: 50
  batch_length: 50
  length: 50
  prefill_timesteps: 5000
  explore_noise: 0.3
  max_episode_length: 1000
  collect_interval: 100
  max_experience_size: 1000
  save_episodes: False
  
  discount: 0.99
  lambda: 0.95
  clip_actions: False
  horizon: 1000
  imagine_horizon: 15
  free_nats: 3.0
  kl_coeff: 1.0

  dreamer_model:
    obs_space: [3, 64, 64]
    num_outputs: 1
    custom_model: DreamerModel
    deter_size: 200
    stoch_size: 30
    depth_size: 32
    hidden_size: 400
    action_init_std: 5.0
  
  env_config:
    action_repeat: 2