utils.py

from pathlib import Path

import numpy as np
from stable_baselines3.common.callbacks import BaseCallback
from stable_baselines3.common.results_plotter import load_results, ts2xy
import json


class SaveOnBestTrainingRewardCallback(BaseCallback):
    """
    Callback for saving a model (the check is done every ``check_freq`` steps)
    based on the training reward (in practice, we recommend using ``EvalCallback``).

    :param check_freq: (int)
    :param log_dir: (str) Path to the folder where the model will be saved.
      It must contains the file created by the ``Monitor`` wrapper.
    :param verbose: (int)
    """

    def __init__(self, check_freq: int, log_dir: str, save_dir: str, verbose=1):
        super(SaveOnBestTrainingRewardCallback, self).__init__(verbose)
        self.check_freq = check_freq
        self.log_dir = log_dir
        self.best_save_path = Path(save_dir) / "best_model"
        self.all_save_path = Path(save_dir) / "current_model"
        self.save_dir = Path(save_dir)
        self.best_mean_reward = -np.inf

    def _init_callback(self) -> None:
        # Create folder if needed
        if self.best_save_path is not None:
            self.best_save_path.parent.mkdir(parents=True, exist_ok=True)
        if self.all_save_path is not None:
            self.all_save_path.parent.mkdir(parents=True, exist_ok=True)

    def _on_step(self) -> bool:
        if self.n_calls % self.check_freq == 0:

            # Retrieve training reward
            x, y = ts2xy(load_results(self.log_dir), "timesteps")
            if len(x) > 0:
                # Mean training reward over the last 100 episodes
                mean_reward = np.mean(y[-100:])
                if self.verbose > 0:
                    print(f"Num timesteps: {self.num_timesteps}")
                    print(
                        f"Best mean reward: {self.best_mean_reward:.2f} - Last mean reward per episode: {mean_reward:.2f}"
                    )

                # New best model, you could save the agent here
                if mean_reward > self.best_mean_reward:
                    self.best_mean_reward = mean_reward
                    # Example for saving best model
                    if self.verbose > 0:
                        print(f"Saving new best model to {self.best_save_path}")
                    self.model.save(self.best_save_path)
                
                # Save it anyway
                self.model.save(self.all_save_path)

            # Metadata
            metadata = {"step": self.n_calls}

            with open(self.save_dir / "metadata.json") as f:
                json.dump(metadata, f)

        return True