config.yml

# env
env_name: PourWater
# 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: 4 # not actually used
# train
num_train_steps: 1000000
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: true
device: cuda
# observation
image_size: 128 # 84
image_pad: 6 # 4
frame_stack: 3 # not actually used
# global params
# 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: 128

# agent configuration
agent:
  # obs_shape: ??? # to be specified later
  # action_shape: ??? # to be specified later
  # action_range: ??? # to be specified later
  device: cuda
  discount: 0.99
  init_temperature: 0.1
  lr: 0.001
  actor_update_frequency: 2
  critic_tau: 0.01
  critic_target_update_frequency: 2
  batch_size: 128
    
critic:
  # class: drq.Critic
  # params:
  # action_shape: ${agent.action_shape}
  hidden_dim: 1024
  hidden_depth: 2
    
actor:
  # class: drq.Actor
  # params:
  # action_shape: ${agent.action_shape}
  hidden_depth: 2
  hidden_dim: 1024
  log_std_bounds: [-10, 2]
    
encoder:
  # class: drq.Encoder
  # params:
  # obs_shape: ${agent.obs_shape}
  feature_dim: 50


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