full_train.yaml

defaults:
  - agent: ddpg
  - _self_ # this yaml file is overriding configs from the Defaults List (eg. {agent}.yaml)
  - resources: null

# mode
reward_free: false
# task settings
domain: walker # primal task will be infered in runtime
obs_type: states # [states, pixels]
frame_stack: 3 # only works if obs_type=pixels
action_repeat: 1 # set to 2 for pixels
discount: 0.99
# train settings
#num_train_frames: 2000010
num_train_frames: 2000000
num_seed_frames: 4000
# eval
eval_every_frames: 10000
num_eval_episodes: 10
# snapshot
save_snapshot_after_each_task: false
snapshot_dir: ./${mode}/models/${obs_type}/${domain}/${agent.name}/${seed}
snapshot_base_dir: ./full_train/models/ # for testing loading from saved models only
# replay buffer
reset_buffer_every_task: false
replay_buffer_size: 1000000
replay_buffer_num_workers: 2
batch_size: ${agent.batch_size}
nstep: ${agent.nstep}
update_encoder: true # should always be true for pre-training
# misc
seed: 1
#device: cuda
device: cpu
save_eval_video: true
save_train_video: false
use_tb: false
use_wandb: false
# experiment
experiment: exp
num_exposures: 2
log_freq: 10  # log every n episodes
#log_freq: 10000  # log every n steps
terminate_after_first_task: false
same_reward_for_all_tasks: false
diff_dynamics_for_all_tasks: false
diff_run_speed_for_all_tasks: false
diff_domains_same_reward: false
walk_run_tasks: false
stand_run_tasks: false
run_jump_tasks: false
single_task_run_fast: false
single_task_run_backward: false
env_type: dm_control

wandb_dir: ./wandb
mode: full_train

hydra:
  run:
    dir: ./exp_local/${now:%Y.%m.%d}/${now:%H%M%S}_${agent.name}
  sweep:
    dir: ./exp_sweep/${now:%Y.%m.%d}/${now:%H%M}_${agent.name}_${experiment}
    subdir: ${hydra.job.num}
  launcher:
    timeout_min: 4300
    cpus_per_task: 10
    gpus_per_node: 1
    tasks_per_node: 1
    mem_gb: 160
    nodes: 1
    submitit_folder: ./exp_sweep/${now:%Y.%m.%d}/${now:%H%M}_${agent.name}_${experiment}/.slurm