Skip to content

Latest commit

 

History

History
47 lines (37 loc) · 2.98 KB

README.md

File metadata and controls

47 lines (37 loc) · 2.98 KB
Raw

How to customize config

Config file management rules

  • The config file provided by default is mainly managed in the form of config/[agent]/[env].py.
  • For a specific environment group that shares parameters, manage it in the form of config/[agent]/[env_group], and specify the environment name with --env.name in the run command.

reference: dqn/cartpole.py, dqn/atari.py

Config setting

  • The config file is managed with a total of four dictionary variables: agent, env, optim, and train.

    agent

    • The agent dictionary manages input parameters used by the agent class.
      • name: The key of the agent class you want to use.
      • others: You can check it in the agent class.

    env

    • The env dictionary manages input parameters used by the env class.
      • name: The key of the env class you want to use.
      • others: You can check it in the env class.

    optim

    • The optim dictionary manages input parameters used by the optimizer class. Since the optimizer of pytorch is used as it is, any optimizer supported by pytorch can be used.
      • name: The key of the optimizer class you want to use.
      • others: You can check it in the optimizer class supported by pytorch.

    train

    • The optim dictionary manages parameters used in the main script.

      • training: It means whether to learn. Set to False in the eval.py script and True otherwise.
      • load_path: It means the path to load the model. If you want to load the model or in the eval.py script, you need to set it. If not, set it None.
      • run_step: It determines the total number of interactions to proceed.
      • print_period: It means the cycle(unit=step) to print the progress.
      • save_period: It means the cycle(unit=step) to save the model.
      • eval_iteration: It means how many episodes will be run in total to get the evaluation score.
      • record: It means whether to record the simulation as the evaluation proceeds. If you set it True, simulation is saved as a gif file in save_path. If you set it True and env is recordable, simulation is saved as a gif file in save_path. (Note that this does not work for non-recordable environments.)
      • record_period: It means the cycle(unit=step) to record.
      • id: If set, log to logs/[env]/id path. (default: agent.name)
      • experiment: If set, log to logs/experiment/[env]/id. Otherwise, log to logs/[env]/id.
      • distributed_batch_size: In distributed script, uses distributed_batch_size instead of agent.batch_size.
      • update_period: It means the cycle(unit=step) in which actors pass transition data to learner.
      • num_workers: Total number of distributed actors which interact with env.
      • eval_time_limit: Time limit(unit=seconds) given per episode when evaluating the model. (default: No limit).

      distributed_batch_size, update_period and num_workers are only used in distributed scripts.

reference: ppo/atari.py