config.yaml

# env
env: cartpole_swingup
env_index: 0 
token_index: 0 
envs: [
        ["walker_stand", "walker_walk"],
        ["finger_turn_easy","finger_turn_hard"],
        ["hopper_stand","hopper_hop"],
        ["cartpole_swingup","cartpole_swingup_sparse"],
        ["reacher_easy","reacher_hard"]
      ]
# IMPORTANT: if action_repeat is used the effective number of env steps needs to be
# multiplied by action_repeat in the result graphs.
# This is a common practice for a fair comparison.
# See the 2nd paragraph in Appendix C of SLAC: https://arxiv.org/pdf/1907.00953.pdf
# See Dreamer TF2's implementation: https://github.com/danijar/dreamer/blob/02f0210f5991c7710826ca7881f19c64a012290c/dreamer.py#L340
action_repeat: 2
# train
num_train_steps: 100000
num_train_iters: 1
num_seed_steps: 1000
replay_buffer_capacity: 100000
seed: 1
# eval
eval_frequency: 5000
num_eval_episodes: 10
# misc
log_frequency_step: 10000
log_save_tb: true
save_video: False
device: -1
# observation
image_size: 84
image_pad: 4
frame_stack: 3
# global params
lr: 1e-4
# IMPORTANT: please use a batch size of 512 to reproduce the results in the paper. Hovewer, with a smaller batch size it still works well.
batch_size: 512
encoder_conf: 0.0025
load_pretrain: True
scale: 0.0425
tag: trans_same


# agent configuration
agent:
  name: drq
  class: drq.DRQAgent
  params:
    obs_shape: ??? # to be specified later
    action_shape: ??? # to be specified later
    action_range: ??? # to be specified later
    device: ${device}
    encoder_cfg: ${encoder}
    critic_cfg: ${critic}
    actor_cfg: ${actor}
    discount: 0.99
    init_temperature: 0.1
    lr: ${lr}
    actor_update_frequency: 2
    critic_tau: 0.01
    critic_target_update_frequency: 2
    batch_size: ${batch_size}

critic:
  class: drq.Critic
  params:
    encoder_cfg: ${token_index}
    action_shape: ${agent.params.action_shape}
    hidden_dim: 1024
    hidden_depth: 3

actor:
  class: drq.Actor
  params:
    encoder_cfg: ${token_index}
    action_shape: ${agent.params.action_shape}
    hidden_depth: 3
    hidden_dim: 1024
    log_std_bounds: [-10, 2]

encoder:
  class: drq.Encoder
  params:
      obs_shape: ${agent.params.obs_shape}
      feature_dim: 50


# hydra configuration
hydra:
  name: ${env}
  run:
          #dir: /apdcephfs/share_1290939/shoufachen/EXP_DR/${now:%Y.%m.%d}/${now:%H%M%S}_${hydra.job.override_dirname}
    dir: ./runs/${now:%Y.%m.%d}/${now:%H%M%S}_${hydra.job.override_dirname}