pretrain.yaml

defaults:
  - agent: ddpg
  - override hydra/launcher: submitit_local

# mode
reward_free: true
# task settings
domain: jaco
use_custom: true
obs_type: states 
frame_stack: 3 # only works if obs_type=pixels
action_repeat: 1 # set to 2 for pixels
discount: 0.99
time_limit: None
use_time_limit: false
random_reset: true
# train settings
num_train_frames: 2000010
num_seed_frames: 4000
# eval settings
eval_every_frames: 2000
num_eval_episodes: 10
# experiment
experiment_folder: test
experiment_name: test
# snapshot
snapshots: [100000, 500000, 1000000, 1500000, 1700000, 1800000, 1910000, 1920000, 1930000, 1940000, 1950000, 1960000, 1970000, 1980000, 1990000, 2000000, 5000000, 10000000, 15000000, 20000000]
snapshot_dir: ./pretrained_model
# restore
restore_snapshot_dir: None
restore_snapshot_ts: 0
# replay buffer
replay_buffer_size: 100000
replay_buffer_num_workers: 4
save_replay_buffer: None
batch_size: ${agent.batch_size}
nstep: ${agent.nstep}
# misc
seed: 1
device: cuda:0
save_video: true
save_train_video: false
use_tb: true
use_wandb: false
# IRM settings
z_id: random_skill
learnable_reward_scale: None
epic_gradient_descent_steps: 0
z_lr: None
num_cem_iterations: 0
num_cem_samples: 0
num_cem_elites: 0
grid_search_size: None


hydra:
  run:
    dir: ./icml_rebuttal_experiments/pretrain/${experiment_folder}/${experiment_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