callbacks.yaml

wandb_entity: openrl-lab
experiment_name: test_callbacks
run_dir: ./wandb_run
callbacks:
  - id: "ProgressBarCallback"
  - id: "StopTrainingOnMaxEpisodes"
    args: {
      "max_episodes": 25, # the max number of episodes to run
      "verbose": 1,
    }
  - id: "CheckpointCallback"
    args: {
        "save_freq": 500, # how often to save the model
        "save_path": "./results/checkpoints/",  # where to save the model
        "name_prefix": "ppo", # the prefix of the saved model
        "save_replay_buffer": True, # not work yet
         "verbose": 2,
    }
  - id: "EvalCallback"
    args: {
      "eval_env": { "id": "CartPole-v1","env_num": 5 }, # how many envs to set up for evaluation
      "n_eval_episodes": 5, # how many episodes to run for each evaluation
      "eval_freq": 500, # how often to run evaluation
      "log_path": "./results/eval_log_path", # where to save the evaluation results
      "best_model_save_path": "./results/best_model/", # where to save the best model
      "deterministic": True, # whether to use deterministic action
      "render": False, # whether to render the env
      "asynchronous": True, # whether to run evaluation asynchronously
      "stop_logic": "OR", # the logic to stop training, OR means training stops when any one of the conditions is met, AND means training stops when all conditions are met
      "callbacks_on_new_best": [
        {
          id: "StopTrainingOnRewardThreshold",
          args: {
            "reward_threshold": 500, # the reward threshold to stop training
            "verbose": 1,
          }
        } ],
      "callbacks_after_eval": [
        {
          id: "StopTrainingOnNoModelImprovement",
          args: {
            "max_no_improvement_evals": 10, # Maximum number of consecutive evaluations without a new best model.
            "min_evals": 2, # Number of evaluations before start to count evaluations without improvements.
          }
        },
      ],
    }
  - id: "EveryNTimesteps" # This is same to "CheckpointCallback"
    args: {
      "n_steps": 5000,
      "callbacks":[
        {
          "id": "CheckpointCallback",
          args: {
            "save_freq": 1,
            "save_path": "./results/checkpoints_with_EveryNTimesteps/",  # where to save the model
            "name_prefix": "ppo", # the prefix of the saved model
            "save_replay_buffer": True, # not work yet
            "verbose": 2,
          }
        }
      ]
    }