configs.yaml

defaults:
  # Dubins car environment parameters
  speed: 1.
  turnRate: 1.25
  x_min: -1.5
  x_max: 1.5
  y_min: -1.5
  y_max: 1.5
  buffer: 0.1
  dt: 0.05
  obs_x: 0
  obs_y: 0
  obs_r: 0.5

  # offline data parameters
  data_length: 100
  num_trajs: 4000 # 2000 in paper
  num_train_trajs: 3800 #1900 in paper
  size: [128, 128]
  dataset_path: 'wm_demos_dubins_sc_F_arrow_0.15_128.pkl' # '128' should be the same as the size above
  
  # offline dreamer parameters, these are from https://github.com/NM512/dreamerv3_torch
  wm_name: ''
  logdir: 'logs/dreamer_dubins' 
  rssm_ckpt_path: logs/dreamer_dubins/rssm_ckpt.pt # change this to whatever the path to your rssm checkpoint is
  traindir: null
  evaldir: null
  offline_traindir: ''
  offline_evaldir: ''
  seed: 0
  deterministic_run: False
  steps: 1e6
  parallel: False
  eval_every: 5e2
  eval_episode_num: 10
  log_every: 5e2
  reset_every: 0
  device: 'cuda:0'
  compile: True
  precision: 32
  debug: False
  video_pred_log: True

  rssm_train_steps: 10000 # 100000 in paper
  
  # Environment
  task: 'dubins-wm'
  envs: 1
  action_repeat: 1
  time_limit: 100
  grayscale: False
  prefill: 5000
  reward_EMA: True

  # Model
  dyn_hidden: 512
  dyn_deter: 512
  dyn_stoch: 32
  dyn_discrete: 0 # 0 for continuous latent
  dyn_rec_depth: 1
  dyn_mean_act: 'none'
  dyn_std_act: 'sigmoid2'
  dyn_min_std: 0.1
  grad_heads: ['decoder', 'cont'] # ,  'margin']
  units: 512
  act: 'SiLU'
  norm: True
  encoder:
    {mlp_keys: 'obs_state', cnn_keys: 'image', act: 'SiLU', norm: True, cnn_depth: 32, kernel_size: 4, minres: 4, mlp_layers: 5, mlp_units: 1024, symlog_inputs: True}
    # {mlp_keys: '', cnn_keys: 'image', act: 'SiLU', norm: True, cnn_depth: 32, kernel_size: 4, minres: 4, mlp_layers: 5, mlp_units: 1024, symlog_inputs: True}
  decoder:
    {mlp_keys: 'obs_state', cnn_keys: 'image', act: 'SiLU', norm: True, cnn_depth: 32, kernel_size: 4, minres: 4, mlp_layers: 5, mlp_units: 1024, cnn_sigmoid: False, image_dist: mse, vector_dist: symlog_mse, outscale: 1.0}
    # {mlp_keys: '', cnn_keys: 'image', act: 'SiLU', norm: True, cnn_depth: 32, kernel_size: 4, minres: 4, mlp_layers: 5, mlp_units: 1024, cnn_sigmoid: False, image_dist: mse, vector_dist: symlog_mse, outscale: 1.0}
  actor:
    {layers: 2, dist: 'normal', entropy: 3e-4, unimix_ratio: 0.01, std: 'learned', min_std: 0.1, max_std: 1.0, temp: 0.1, lr: 3e-5, eps: 1e-5, grad_clip: 100.0, outscale: 1.0}
  critic:
    {layers: 2, dist: 'symlog_disc', slow_target: True, slow_target_update: 1, slow_target_fraction: 0.02, lr: 3e-5, eps: 1e-5, grad_clip: 100.0, outscale: 0.0}
  #reward_head:
  #  {layers: 2, dist: 'symlog_disc', loss_scale: 1.0, outscale: 0.0}
  cont_head:
    {layers: 2, loss_scale: 1.0, outscale: 1.0}
  # new
  margin_head:
    {layers: 2, loss_scale: 10.0}
  gamma_lx: 0.75 
  dyn_scale: 0.5
  rep_scale: 0.1
  kl_free: 1.0
  weight_decay: 0.0
  unimix_ratio: 0.01
  initial: 'learned'

  # Dataset generation
  show_constraint: False
  constraint_color: 'red'
  arrow_size: 0.15 # Smaller is bigger here

  # Training
  batch_size: 16
  batch_length: 64
  train_ratio: 512
  pretrain: 100
  model_lr: 1e-4
  obs_lr: 1e-3
  lx_lr: 1e-4
  opt_eps: 1e-8
  grad_clip: 1000
  dataset_size: 1000000
  opt: 'adam'

  # Behavior.
  discount: 0.997
  discount_lambda: 0.95
  imag_horizon: 15
  imag_gradient: 'dynamics'
  imag_gradient_mix: 0.0
  eval_state_mean: False

  # Exploration
  expl_behavior: 'greedy'
  expl_until: 0
  expl_extr_scale: 0.0
  expl_intr_scale: 1.0
  disag_target: 'stoch'
  disag_log: True
  disag_models: 10
  disag_offset: 1
  disag_layers: 4
  disag_units: 400
  disag_action_cond: False

  
  # LCRL
  reward-threshold: null 
  #seed: 0 
  buffer-size: 1e6 # type=float, default=40000)
  actor-lr: 1e-4
  critic-lr: 1e-3
  gamma-pyhj: 0.9999 # type=float, default=0.95)
  tau: 0.005 # type=float, default=0.005)
  exploration-noise: 0.1 # type=float, default=0.1)
  epoch: 1 # type=int, default=10)
  warmup: 1 # number of epochs for warm up
  total-episodes: 20 # type=int, default=160)
  step-per-epoch: 40000 # type=int, default=40000)
  step-per-collect: 16 # type=int, default=8)
  update-per-step: 0.125 # type=float, default=0.125)
  batch_size-pyhj: 512 # type=int, default=512)
  control-net: [512, 512, 512, 512] # type=int, nargs='*', default=None) # for control policy
  critic-net: [512, 512, 512, 512]  # type=int, nargs='*', default=None) # for critic net
  constraint-embedding-dim: 512
  pass-semantic-constraint: True
  pass-semantic-state: False
  control-net-const: [] # type=int, nargs='*', default=(128, 128)
  training-num: 16 # type=int, default=8)
  test-num: 1 # type=int, default=100)
  render: 0. # type=float, default=0.)
  rew-norm: False # action='store_true', default=False)
  n-step: 1 # type=int, default=1)
  continue-training-logdir: null # type=str, default=None)
  continue-training-epoch: null # type=int, default=None)
  actor-gradient-steps: 1 # type=int, default=1)
  is-game-baseline: False # type=bool, default=False) # it will be set automatically
  target-update-freq: 400 # type=int, default=400)
  auto-alpha: 1
  alpha-lr: 3e-4
  alpha: 0.2 
  weight-decay-pyhj: 0.001

  env-dist-type: ds # 'v' - vanilla, 'fc' - four corners, 'fcfe' - four corners four edges, 'rh' - right half, 'br' - big radii, 'uni' - uniform, 'rhfe' - right half four edges, 'rhbr' - right half big radii
  safety-margin-type: 'cos_sim' # cos_sim, learned
  safety-margin-threshold: -0.64 # type=float, default=0.5)
  safety-margin-hard-threshold: False # type=bool, default=False)
  pass-prototype: False # type=bool, default=False) # whether to pass a prototype to the policy net, not the raw constraint

  safety_filter_eps: 2.0
  
  actor-activation: 'ReLU' #type=str, default='ReLU')
  critic-activation: 'ReLU' # type=str, default='ReLU')
  kwargs: {} # type=str, default='{}')
  warm-start-path: null # type=str, default=None) # e.g., log/ra_droneracing_Game-v6/epoch_id_10/policy.pth

  nx: 51
  ny: 51
  nt: 51

  # Semantic encoder training
  sem:
    sz_embedding: 512 # Size of embedding that is appended to backbone model.
    sz_batch: 150 # Number of samples per batch.
    nb_epochs: 60 # Number of training epochs.
    gpu_id: 0 # ID of GPU that is used for training.
    nb_workers: 4 # Number of workers for dataloader.
    loss: 'priv' # Criterion for training
    optimizer: 'adamw' # Optimizer setting
    lr: 1e-4 # Learning rate setting
    weight_decay: 1e-4 # Weight decay setting
    lr_decay_step: 10 # Learning decay step setting
    lr_decay_gamma: 0.5 # Learning decay gamma setting
    save_model: True # Don't save model