config.yaml

seed: 42
output_dir: ${hydra:runtime.output_dir}
domains: mujoco_metaworld # domains to train on
wb_tag: "default" # wandb tag
log_interval: 10 # how many steps before logging to wandb
script_name: "" # log the running script
pretrained_dir: "" # pretrained model directory
parallel_eval: False # use ray to do parallel evaluation
slurm_job_id: "" # the slurm job id for logging purposes
user_id: "" # the machine user id for logging purposes

epoch_size: 10
total_num_traj: 0
train_time: 0
cross_validate: False
cross_validate_eps_num: 20
cross_validate_freq: 100
save_intermedia_models: False
comment: ""

defaults:
  - _self_
  - env: mujoco_metaworld

# dataset config
dataset:
  _target_: hpt.dataset.local_traj_dataset.LocalTrajDataset
  horizon: 1 # horizon for each dataset sample. not used
  val_ratio: 0.1 # the train-validation ratio
  pad_after: 0 # padding after the episode
  episode_cnt: 10000 # total episodes by default
  step_cnt: 100000 # total data transitions
  data_augmentation: False  # data augmentation
  use_disk: False # use disk instead of memory to store the data
  pad_before: 0 # padding before the episode
  data_ratio: 1 # only use a fraction of data
  action_horizon: 8 # observation: (observation + action) is action horizon
  observation_horizon: 1 # before observation horizon is observation
  dataset_postfix: "_traj${dataset.episode_cnt}" # postfix for the dataset
  precompute_feat: True # precompute features using pretrained models for stems
  image_encoder: 'resnet' # which encoder to use as the pretrained model
  dataset_encoder_postfix: "_${dataset.image_encoder}" # another postfix
  use_multiview: False # use multiple camera views
  normalize_state: True # whether to normalize the states in datasets
  regenerate: False # regenerate data
  action_multiple_horizon: True  # multiply action dimensions by horizon
  random_mask_obs: True # whether to randomize observation input length
  data_augment_ratio: 1 # add data augmentation to the images
  proprioception_expand: False # expand proprioception to use multiple tokens
  proprioception_expand_dim: 32 # expand proprioception dimensions

# trunk transformer config
network:
  _target_: hpt.models.policy.Policy
  embed_dim: 128
  num_blocks: 16 # num of blocks in the trunk transformer
  num_heads: 8 # num of heads in the trunk transformer
  drop_path: 0.1 # drop path in the trunk transformer
  use_modality_embedding: True # add trainable modality position tokens
  use_domain_embedding: False # whether to add domain-specific trainable parameters
  observation_horizon: ${dataset.horizon} # the observation history
  action_horizon: 1 # open loop action steps. <= the dataset action horizons
  token_postprocessing: "mean" # maxpool or meanpool the tokens
  cross_stem_attention: True # use cross attention to combine state and action
  weight_init_style: 'pytorch' # weight init
  no_trunk: False # ignore trunk
  finetune_encoder: False # whether to finetune encoders

# stem network for different modalities
stem:
  modalities: ['image', 'state'] # 'language'
  modality_embed_dim: ${network.embed_dim}
  normalize_state: ${dataset.normalize_state} # normalize state vectors
  state_embedding_dim: 1 # dimension of positional encoding for state
  cross_attention: True # whether to use cross attention or not
  precompute_feat: True # whether to use precomputed features. if not, will finetune.
  image_encoder: ${dataset.image_encoder} # what image encoder to use
  crossattn_dim_head: 64 # for cross attention modules
  crossattn_heads: 8 # number of heads in cross attention
  crossattn_modality_dropout: 0.1 # the dropout ratio for cross attention
  num_blocks: 1 # number of blocks for stem transformer's cross and self attention
  observation_horizon: ${dataset.observation_horizon} # observation horizon
  masked_autoencoding: False # random mask encoding and then reconstruction
  random_horizon_masking: True # whether to randomize observation input length
  add_pos_embedding_to_state: False # positional embedding for the state

  # standardize token sizes for each modality
  crossattn_latent:
    image: 16
    state: 16
    # language: 8

  image:
    _target_: hpt.models.policy_stem.MLP
    input_dim: 512
    output_dim: ${network.embed_dim}
    widths:  [128]
    num_of_copy: 1

  # each item is a token
  state:
    _target_: hpt.models.policy_stem.MLP
    input_dim: ${stem.state_embedding_dim} # ovewrite based on the dataset
    output_dim: ${network.embed_dim}
    widths: [128]

# head network
head:
    _target_: hpt.models.policy_head.MLP
    input_dim: ${network.embed_dim}
    tanh_end: True # normalized action output
    output_dim: -1 # overwrite based on dataset
    widths: [256, 128]
    normalize_action: ${head.tanh_end}
    dropout: True

# self-explanatory torch config
dataloader:
  batch_size: 256
  num_workers: 1
  pin_memory: True
  persistent_workers: True
  shuffle: True
  drop_last: False

val_dataloader:
  batch_size: 256
  num_workers: 1
  shuffle: False
  pin_memory: True
  persistent_workers: True
  drop_last: False

optimizer:
  _target_: torch.optim.AdamW
  lr: 1.0e-5 # 1e-4
  eps: 1.0e-7
  weight_decay: 1e-4

optimizer_misc:
  nontrunk_lr_scale: 1.

warmup_lr:
  lr: 1e-10
  step: 1000 # first 1000 iterations

lr_scheduler:
  _target_: torch.optim.lr_scheduler.CosineAnnealingLR
  T_max: ${train.total_epochs}
  eta_min: 1e-6

# training config
train:
  total_epochs: 500  # maximum training epochs before termination. usually set as maximum
  total_iters: 20000 # maximum training steps before termination
  epoch_iters: 1000  # training steps in each epoch
  validation_iters: 100 # maximum iterations for validation
  pretrained_dir: "" # pretrained model path for testing
  freeze_trunk: True # whether to freeze the trunk during finetuning
  wandb_pretrained_dir: "" # use models pretrained on wandb