exp_manager.py

import argparse
import importlib
import os
import pickle as pkl
import time
import warnings
from collections import OrderedDict
from pathlib import Path
from pprint import pprint
from typing import Any, Callable, Dict, List, Optional, Tuple, Union

import gymnasium as gym
import numpy as np
import optuna
import torch as th
import yaml
from gymnasium import spaces
from huggingface_sb3 import EnvironmentName
from optuna.pruners import BasePruner, MedianPruner, NopPruner, SuccessiveHalvingPruner
from optuna.samplers import BaseSampler, RandomSampler, TPESampler
from optuna.study import MaxTrialsCallback
from optuna.trial import TrialState
from optuna.visualization import plot_optimization_history, plot_param_importances
from sb3_contrib.common.vec_env import AsyncEval

# For using HER with GoalEnv
from stable_baselines3 import HerReplayBuffer
from stable_baselines3.common.base_class import BaseAlgorithm
from stable_baselines3.common.callbacks import BaseCallback, CheckpointCallback, EvalCallback, ProgressBarCallback
from stable_baselines3.common.env_util import make_vec_env
from stable_baselines3.common.noise import NormalActionNoise, OrnsteinUhlenbeckActionNoise
from stable_baselines3.common.preprocessing import is_image_space, is_image_space_channels_first
from stable_baselines3.common.sb2_compat.rmsprop_tf_like import RMSpropTFLike  # noqa: F401
from stable_baselines3.common.utils import constant_fn
from stable_baselines3.common.vec_env import (
    DummyVecEnv,
    SubprocVecEnv,
    VecEnv,
    VecFrameStack,
    VecNormalize,
    VecTransposeImage,
    is_vecenv_wrapped,
)

# For custom activation fn
from torch import nn as nn

# Register custom envs
import rl_zoo3.import_envs  # noqa: F401
from rl_zoo3.callbacks import SaveVecNormalizeCallback, TrialEvalCallback
from rl_zoo3.hyperparams_opt import HYPERPARAMS_SAMPLER
from rl_zoo3.utils import ALGOS, get_callback_list, get_class_by_name, get_latest_run_id, get_wrapper_class, linear_schedule


class ExperimentManager:
    """
    Experiment manager: read the hyperparameters,
    preprocess them, create the environment and the RL model.

    Please take a look at `train.py` to have the details for each argument.
    """

    def __init__(
        self,
        args: argparse.Namespace,
        algo: str,
        env_id: str,
        log_folder: str,
        tensorboard_log: str = "",
        n_timesteps: int = 0,
        eval_freq: int = 10000,
        n_eval_episodes: int = 5,
        save_freq: int = -1,
        hyperparams: Optional[Dict[str, Any]] = None,
        env_kwargs: Optional[Dict[str, Any]] = None,
        eval_env_kwargs: Optional[Dict[str, Any]] = None,
        trained_agent: str = "",
        optimize_hyperparameters: bool = False,
        storage: Optional[str] = None,
        study_name: Optional[str] = None,
        n_trials: int = 1,
        max_total_trials: Optional[int] = None,
        n_jobs: int = 1,
        sampler: str = "tpe",
        pruner: str = "median",
        optimization_log_path: Optional[str] = None,
        n_startup_trials: int = 0,
        n_evaluations: int = 1,
        truncate_last_trajectory: bool = False,
        uuid_str: str = "",
        seed: int = 0,
        log_interval: int = 0,
        save_replay_buffer: bool = False,
        verbose: int = 1,
        vec_env_type: str = "dummy",
        n_eval_envs: int = 1,
        no_optim_plots: bool = False,
        device: Union[th.device, str] = "auto",
        config: Optional[str] = None,
        show_progress: bool = False,
    ):
        super().__init__()
        self.algo = algo
        self.env_name = EnvironmentName(env_id)
        # Custom params
        self.custom_hyperparams = hyperparams
        if (Path(__file__).parent / "hyperparams").is_dir():
            # Package version
            default_path = Path(__file__).parent
        else:
            # Take the root folder
            default_path = Path(__file__).parent.parent

        self.config = config or str(default_path / f"hyperparams/{self.algo}.yml")
        self.env_kwargs: Dict[str, Any] = env_kwargs or {}
        self.n_timesteps = n_timesteps
        self.normalize = False
        self.normalize_kwargs: Dict[str, Any] = {}
        self.env_wrapper: Optional[Callable] = None
        self.frame_stack = None
        self.seed = seed
        self.optimization_log_path = optimization_log_path

        self.vec_env_class = {"dummy": DummyVecEnv, "subproc": SubprocVecEnv}[vec_env_type]
        self.vec_env_wrapper: Optional[Callable] = None

        self.vec_env_kwargs: Dict[str, Any] = {}
        # self.vec_env_kwargs = {} if vec_env_type == "dummy" else {"start_method": "fork"}

        # Callbacks
        self.specified_callbacks: List = []
        self.callbacks: List[BaseCallback] = []
        # Use env-kwargs if eval_env_kwargs was not specified
        self.eval_env_kwargs: Dict[str, Any] = eval_env_kwargs or self.env_kwargs
        self.save_freq = save_freq
        self.eval_freq = eval_freq
        self.n_eval_episodes = n_eval_episodes
        self.n_eval_envs = n_eval_envs

        self.n_envs = 1  # it will be updated when reading hyperparams
        self.n_actions = 0  # For DDPG/TD3 action noise objects
        self._hyperparams: Dict[str, Any] = {}
        self.monitor_kwargs: Dict[str, Any] = {}

        self.trained_agent = trained_agent
        self.continue_training = trained_agent.endswith(".zip") and os.path.isfile(trained_agent)
        self.truncate_last_trajectory = truncate_last_trajectory

        self._is_atari = self.is_atari(env_id)
        # Hyperparameter optimization config
        self.optimize_hyperparameters = optimize_hyperparameters
        self.storage = storage
        self.study_name = study_name
        self.no_optim_plots = no_optim_plots
        # maximum number of trials for finding the best hyperparams
        self.n_trials = n_trials
        self.max_total_trials = max_total_trials
        # number of parallel jobs when doing hyperparameter search
        self.n_jobs = n_jobs
        self.sampler = sampler
        self.pruner = pruner
        self.n_startup_trials = n_startup_trials
        self.n_evaluations = n_evaluations
        self.deterministic_eval = not (self.is_atari(env_id) or self.is_minigrid(env_id))
        self.device = device

        # Logging
        self.log_folder = log_folder
        self.tensorboard_log = None if tensorboard_log == "" else os.path.join(tensorboard_log, self.env_name)
        self.verbose = verbose
        self.args = args
        self.log_interval = log_interval
        self.save_replay_buffer = save_replay_buffer
        self.show_progress = show_progress

        self.log_path = f"{log_folder}/{self.algo}/"
        self.save_path = os.path.join(
            self.log_path, f"{self.env_name}_{get_latest_run_id(self.log_path, self.env_name) + 1}{uuid_str}"
        )
        self.params_path = f"{self.save_path}/{self.env_name}"

    def setup_experiment(self) -> Optional[Tuple[BaseAlgorithm, Dict[str, Any]]]:
        """
        Read hyperparameters, pre-process them (create schedules, wrappers, callbacks, action noise objects)
        create the environment and possibly the model.

        :return: the initialized RL model
        """
        unprocessed_hyperparams, saved_hyperparams = self.read_hyperparameters()
        hyperparams, self.env_wrapper, self.callbacks, self.vec_env_wrapper = self._preprocess_hyperparams(
            unprocessed_hyperparams
        )

        self.create_log_folder()
        self.create_callbacks()

        # Create env to have access to action space for action noise
        n_envs = 1 if self.algo == "ars" or self.optimize_hyperparameters else self.n_envs
        env = self.create_envs(n_envs, no_log=False)

        self._hyperparams = self._preprocess_action_noise(hyperparams, saved_hyperparams, env)

        if self.continue_training:
            model = self._load_pretrained_agent(self._hyperparams, env)
        elif self.optimize_hyperparameters:
            env.close()
            return None
        else:
            # Train an agent from scratch
            model = ALGOS[self.algo](
                env=env,
                tensorboard_log=self.tensorboard_log,
                seed=self.seed,
                verbose=self.verbose,
                device=self.device,
                **self._hyperparams,
            )

        self._save_config(saved_hyperparams)
        return model, saved_hyperparams

    def learn(self, model: BaseAlgorithm) -> None:
        """
        :param model: an initialized RL model
        """
        kwargs: Dict[str, Any] = {}
        if self.log_interval > -1:
            kwargs = {"log_interval": self.log_interval}

        if len(self.callbacks) > 0:
            kwargs["callback"] = self.callbacks

        # Special case for ARS
        if self.algo == "ars" and self.n_envs > 1:
            kwargs["async_eval"] = AsyncEval(
                [lambda: self.create_envs(n_envs=1, no_log=True) for _ in range(self.n_envs)], model.policy
            )

        try:
            model.learn(self.n_timesteps, **kwargs)
        except KeyboardInterrupt:
            # this allows to save the model when interrupting training
            pass
        finally:
            # Clean progress bar
            if len(self.callbacks) > 0:
                self.callbacks[0].on_training_end()
            # Release resources
            try:
                assert model.env is not None
                model.env.close()
            except EOFError:
                pass

    def save_trained_model(self, model: BaseAlgorithm) -> None:
        """
        Save trained model optionally with its replay buffer
        and ``VecNormalize`` statistics

        :param model:
        """
        print(f"Saving to {self.save_path}")
        model.save(f"{self.save_path}/{self.env_name}")

        if hasattr(model, "save_replay_buffer") and self.save_replay_buffer:
            print("Saving replay buffer")
            model.save_replay_buffer(os.path.join(self.save_path, "replay_buffer.pkl"))

        if self.normalize:
            # Important: save the running average, for testing the agent we need that normalization
            vec_normalize = model.get_vec_normalize_env()
            assert vec_normalize is not None
            vec_normalize.save(os.path.join(self.params_path, "vecnormalize.pkl"))

    def _save_config(self, saved_hyperparams: Dict[str, Any]) -> None:
        """
        Save unprocessed hyperparameters, this can be use later
        to reproduce an experiment.

        :param saved_hyperparams:
        """
        # Save hyperparams
        with open(os.path.join(self.params_path, "config.yml"), "w") as f:
            yaml.dump(saved_hyperparams, f)

        # save command line arguments
        with open(os.path.join(self.params_path, "args.yml"), "w") as f:
            ordered_args = OrderedDict([(key, vars(self.args)[key]) for key in sorted(vars(self.args).keys())])
            yaml.dump(ordered_args, f)

        print(f"Log path: {self.save_path}")

    def read_hyperparameters(self) -> Tuple[Dict[str, Any], Dict[str, Any]]:
        print(f"Loading hyperparameters from: {self.config}")

        if self.config.endswith(".yml") or self.config.endswith(".yaml"):
            # Load hyperparameters from yaml file
            with open(self.config) as f:
                hyperparams_dict = yaml.safe_load(f)
        elif self.config.endswith(".py"):
            global_variables: Dict = {}
            # Load hyperparameters from python file
            exec(Path(self.config).read_text(), global_variables)
            hyperparams_dict = global_variables["hyperparams"]
        else:
            # Load hyperparameters from python package
            hyperparams_dict = importlib.import_module(self.config).hyperparams
            # raise ValueError(f"Unsupported config file format: {self.config}")

        if self.env_name.gym_id in list(hyperparams_dict.keys()):
            hyperparams = hyperparams_dict[self.env_name.gym_id]
        elif self._is_atari:
            hyperparams = hyperparams_dict["atari"]
        else:
            raise ValueError(f"Hyperparameters not found for {self.algo}-{self.env_name.gym_id} in {self.config}")

        if self.custom_hyperparams is not None:
            # Overwrite hyperparams if needed
            hyperparams.update(self.custom_hyperparams)
        # Sort hyperparams that will be saved
        saved_hyperparams = OrderedDict([(key, hyperparams[key]) for key in sorted(hyperparams.keys())])

        # Always print used hyperparameters
        print("Default hyperparameters for environment (ones being tuned will be overridden):")
        pprint(saved_hyperparams)

        return hyperparams, saved_hyperparams

    @staticmethod
    def _preprocess_schedules(hyperparams: Dict[str, Any]) -> Dict[str, Any]:
        # Create schedules
        for key in ["learning_rate", "clip_range", "clip_range_vf", "delta_std"]:
            if key not in hyperparams:
                continue
            if isinstance(hyperparams[key], str):
                schedule, initial_value = hyperparams[key].split("_")
                initial_value = float(initial_value)
                hyperparams[key] = linear_schedule(initial_value)
            elif isinstance(hyperparams[key], (float, int)):
                # Negative value: ignore (ex: for clipping)
                if hyperparams[key] < 0:
                    continue
                hyperparams[key] = constant_fn(float(hyperparams[key]))
            else:
                raise ValueError(f"Invalid value for {key}: {hyperparams[key]}")
        return hyperparams

    def _preprocess_normalization(self, hyperparams: Dict[str, Any]) -> Dict[str, Any]:
        if "normalize" in hyperparams.keys():
            self.normalize = hyperparams["normalize"]

            # Special case, instead of both normalizing
            # both observation and reward, we can normalize one of the two.
            # in that case `hyperparams["normalize"]` is a string
            # that can be evaluated as python,
            # ex: "dict(norm_obs=False, norm_reward=True)"
            if isinstance(self.normalize, str):
                self.normalize_kwargs = eval(self.normalize)
                self.normalize = True

            if isinstance(self.normalize, dict):
                self.normalize_kwargs = self.normalize
                self.normalize = True

            # Use the same discount factor as for the algorithm
            if "gamma" in hyperparams:
                self.normalize_kwargs["gamma"] = hyperparams["gamma"]

            del hyperparams["normalize"]
        return hyperparams

    def _preprocess_hyperparams(  # noqa: C901
        self, hyperparams: Dict[str, Any]
    ) -> Tuple[Dict[str, Any], Optional[Callable], List[BaseCallback], Optional[Callable]]:
        self.n_envs = hyperparams.get("n_envs", 1)

        if self.verbose > 0:
            print(f"Using {self.n_envs} environments")

        # Convert schedule strings to objects
        hyperparams = self._preprocess_schedules(hyperparams)

        # Pre-process train_freq
        if "train_freq" in hyperparams and isinstance(hyperparams["train_freq"], list):
            hyperparams["train_freq"] = tuple(hyperparams["train_freq"])

        # Should we overwrite the number of timesteps?
        if self.n_timesteps > 0:
            if self.verbose:
                print(f"Overwriting n_timesteps with n={self.n_timesteps}")
        else:
            self.n_timesteps = int(hyperparams["n_timesteps"])

        # Derive n_evaluations from number of timesteps if needed
        if self.n_evaluations is None and self.optimize_hyperparameters:
            self.n_evaluations = max(1, self.n_timesteps // int(1e5))
            print(
                f"Doing {self.n_evaluations} intermediate evaluations for pruning based on the number of timesteps."
                " (1 evaluation every 100k timesteps)"
            )

        # Pre-process normalize config
        hyperparams = self._preprocess_normalization(hyperparams)

        # Pre-process policy/buffer keyword arguments
        # Convert to python object if needed
        for kwargs_key in {"policy_kwargs", "replay_buffer_class", "replay_buffer_kwargs"}:
            if kwargs_key in hyperparams.keys() and isinstance(hyperparams[kwargs_key], str):
                hyperparams[kwargs_key] = eval(hyperparams[kwargs_key])

        # Preprocess monitor kwargs
        if "monitor_kwargs" in hyperparams.keys():
            self.monitor_kwargs = hyperparams["monitor_kwargs"]
            # Convert str to python code
            if isinstance(self.monitor_kwargs, str):
                self.monitor_kwargs = eval(self.monitor_kwargs)
            del hyperparams["monitor_kwargs"]

        # Delete keys so the dict can be pass to the model constructor
        if "n_envs" in hyperparams.keys():
            del hyperparams["n_envs"]
        del hyperparams["n_timesteps"]

        if "frame_stack" in hyperparams.keys():
            self.frame_stack = hyperparams["frame_stack"]
            del hyperparams["frame_stack"]

        # import the policy when using a custom policy
        if "policy" in hyperparams and "." in hyperparams["policy"]:
            hyperparams["policy"] = get_class_by_name(hyperparams["policy"])

        # obtain a class object from a wrapper name string in hyperparams
        # and delete the entry
        env_wrapper = get_wrapper_class(hyperparams)
        if "env_wrapper" in hyperparams.keys():
            del hyperparams["env_wrapper"]

        # Same for VecEnvWrapper
        vec_env_wrapper = get_wrapper_class(hyperparams, "vec_env_wrapper")
        if "vec_env_wrapper" in hyperparams.keys():
            del hyperparams["vec_env_wrapper"]

        callbacks = get_callback_list(hyperparams)
        if "callback" in hyperparams.keys():
            self.specified_callbacks = hyperparams["callback"]
            del hyperparams["callback"]

        return hyperparams, env_wrapper, callbacks, vec_env_wrapper

    def _preprocess_action_noise(
        self, hyperparams: Dict[str, Any], saved_hyperparams: Dict[str, Any], env: VecEnv
    ) -> Dict[str, Any]:
        # Parse noise string
        # Note: only off-policy algorithms are supported
        if hyperparams.get("noise_type") is not None:
            noise_type = hyperparams["noise_type"].strip()
            noise_std = hyperparams["noise_std"]

            # Save for later (hyperparameter optimization)
            assert isinstance(
                env.action_space, spaces.Box
            ), f"Action noise can only be used with Box action space, not {env.action_space}"
            self.n_actions = env.action_space.shape[0]

            if "normal" in noise_type:
                hyperparams["action_noise"] = NormalActionNoise(
                    mean=np.zeros(self.n_actions),
                    sigma=noise_std * np.ones(self.n_actions),
                )
            elif "ornstein-uhlenbeck" in noise_type:
                hyperparams["action_noise"] = OrnsteinUhlenbeckActionNoise(
                    mean=np.zeros(self.n_actions),
                    sigma=noise_std * np.ones(self.n_actions),
                )
            else:
                raise RuntimeError(f'Unknown noise type "{noise_type}"')

            print(f"Applying {noise_type} noise with std {noise_std}")

            del hyperparams["noise_type"]
            del hyperparams["noise_std"]

        return hyperparams

    def create_log_folder(self):
        os.makedirs(self.params_path, exist_ok=True)

    def create_callbacks(self):
        if self.show_progress:
            self.callbacks.append(ProgressBarCallback())

        if self.save_freq > 0:
            # Account for the number of parallel environments
            self.save_freq = max(self.save_freq // self.n_envs, 1)
            self.callbacks.append(
                CheckpointCallback(
                    save_freq=self.save_freq,
                    save_path=self.save_path,
                    name_prefix="rl_model",
                    verbose=1,
                )
            )

        # Create test env if needed, do not normalize reward
        if self.eval_freq > 0 and not self.optimize_hyperparameters:
            # Account for the number of parallel environments
            self.eval_freq = max(self.eval_freq // self.n_envs, 1)

            if self.verbose > 0:
                print("Creating test environment")

            save_vec_normalize = SaveVecNormalizeCallback(save_freq=1, save_path=self.params_path)
            eval_callback = EvalCallback(
                self.create_envs(self.n_eval_envs, eval_env=True),
                callback_on_new_best=save_vec_normalize,
                best_model_save_path=self.save_path,
                n_eval_episodes=self.n_eval_episodes,
                log_path=self.save_path,
                eval_freq=self.eval_freq,
                deterministic=self.deterministic_eval,
            )

            self.callbacks.append(eval_callback)

    @staticmethod
    def entry_point(env_id: str) -> str:
        return str(gym.envs.registry[env_id].entry_point)

    @staticmethod
    def is_atari(env_id: str) -> bool:
        return "AtariEnv" in ExperimentManager.entry_point(env_id)

    @staticmethod
    def is_minigrid(env_id: str) -> bool:
        return "minigrid" in ExperimentManager.entry_point(env_id)

    @staticmethod
    def is_bullet(env_id: str) -> bool:
        return "pybullet_envs" in ExperimentManager.entry_point(env_id)

    @staticmethod
    def is_robotics_env(env_id: str) -> bool:
        return "gym.envs.robotics" in ExperimentManager.entry_point(
            env_id
        ) or "panda_gym.envs" in ExperimentManager.entry_point(env_id)

    @staticmethod
    def is_panda_gym(env_id: str) -> bool:
        return "panda_gym.envs" in ExperimentManager.entry_point(env_id)

    def _maybe_normalize(self, env: VecEnv, eval_env: bool) -> VecEnv:
        """
        Wrap the env into a VecNormalize wrapper if needed
        and load saved statistics when present.

        :param env:
        :param eval_env:
        :return:
        """
        # Pretrained model, load normalization
        path_ = os.path.join(os.path.dirname(self.trained_agent), self.env_name)
        path_ = os.path.join(path_, "vecnormalize.pkl")

        if os.path.exists(path_):
            print("Loading saved VecNormalize stats")
            env = VecNormalize.load(path_, env)
            # Deactivate training and reward normalization
            if eval_env:
                env.training = False
                env.norm_reward = False

        elif self.normalize:
            # Copy to avoid changing default values by reference
            local_normalize_kwargs = self.normalize_kwargs.copy()
            # In eval env: turn off reward normalization and normalization stats updates.
            if eval_env:
                local_normalize_kwargs["norm_reward"] = False
                local_normalize_kwargs["training"] = False

            if self.verbose > 0:
                if len(local_normalize_kwargs) > 0:
                    print(f"Normalization activated: {local_normalize_kwargs}")
                else:
                    print("Normalizing input and reward")
            env = VecNormalize(env, **local_normalize_kwargs)
        return env

    def create_envs(self, n_envs: int, eval_env: bool = False, no_log: bool = False) -> VecEnv:
        """
        Create the environment and wrap it if necessary.

        :param n_envs:
        :param eval_env: Whether is it an environment used for evaluation or not
        :param no_log: Do not log training when doing hyperparameter optim
            (issue with writing the same file)
        :return: the vectorized environment, with appropriate wrappers
        """
        # Do not log eval env (issue with writing the same file)
        log_dir = None if eval_env or no_log else self.save_path

        # Special case for GoalEnvs: log success rate too
        if (
            "Neck" in self.env_name.gym_id
            or self.is_robotics_env(self.env_name.gym_id)
            or "parking-v0" in self.env_name.gym_id
            and len(self.monitor_kwargs) == 0  # do not overwrite custom kwargs
        ):
            self.monitor_kwargs = dict(info_keywords=("is_success",))

        spec = gym.spec(self.env_name.gym_id)

        # Define make_env here, so it works with subprocesses
        # when the registry was modified with `--gym-packages`
        # See https://github.com/HumanCompatibleAI/imitation/pull/160
        def make_env(**kwargs) -> gym.Env:
            return spec.make(**kwargs)

        env_kwargs = self.eval_env_kwargs if eval_env else self.env_kwargs

        # On most env, SubprocVecEnv does not help and is quite memory hungry,
        # therefore, we use DummyVecEnv by default
        env = make_vec_env(
            make_env,
            n_envs=n_envs,
            seed=self.seed,
            env_kwargs=env_kwargs,
            monitor_dir=log_dir,
            wrapper_class=self.env_wrapper,
            vec_env_cls=self.vec_env_class,  # type: ignore[arg-type]
            vec_env_kwargs=self.vec_env_kwargs,
            monitor_kwargs=self.monitor_kwargs,
        )

        if self.vec_env_wrapper is not None:
            env = self.vec_env_wrapper(env)

        # Wrap the env into a VecNormalize wrapper if needed
        # and load saved statistics when present
        env = self._maybe_normalize(env, eval_env)

        # Optional Frame-stacking
        if self.frame_stack is not None:
            n_stack = self.frame_stack
            env = VecFrameStack(env, n_stack)
            if self.verbose > 0:
                print(f"Stacking {n_stack} frames")

        if not is_vecenv_wrapped(env, VecTransposeImage):
            wrap_with_vectranspose = False
            if isinstance(env.observation_space, spaces.Dict):
                # If even one of the keys is an image-space in need of transpose, apply transpose
                # If the image spaces are not consistent (for instance, one is channel first,
                # the other channel last); VecTransposeImage will throw an error
                for space in env.observation_space.spaces.values():
                    wrap_with_vectranspose = wrap_with_vectranspose or (
                        is_image_space(space) and not is_image_space_channels_first(space)  # type: ignore[arg-type]
                    )
            else:
                wrap_with_vectranspose = is_image_space(env.observation_space) and not is_image_space_channels_first(
                    env.observation_space  # type: ignore[arg-type]
                )

            if wrap_with_vectranspose:
                if self.verbose >= 1:
                    print("Wrapping the env in a VecTransposeImage.")
                env = VecTransposeImage(env)

        return env

    def _load_pretrained_agent(self, hyperparams: Dict[str, Any], env: VecEnv) -> BaseAlgorithm:
        # Continue training
        print("Loading pretrained agent")
        # Policy should not be changed
        del hyperparams["policy"]

        if "policy_kwargs" in hyperparams.keys():
            del hyperparams["policy_kwargs"]

        model = ALGOS[self.algo].load(
            self.trained_agent,
            env=env,
            seed=self.seed,
            tensorboard_log=self.tensorboard_log,
            verbose=self.verbose,
            device=self.device,
            **hyperparams,
        )

        replay_buffer_path = os.path.join(os.path.dirname(self.trained_agent), "replay_buffer.pkl")

        if os.path.exists(replay_buffer_path):
            print("Loading replay buffer")
            # `truncate_last_traj` will be taken into account only if we use HER replay buffer
            assert hasattr(
                model, "load_replay_buffer"
            ), "The current model doesn't have a `load_replay_buffer` to load the replay buffer"
            model.load_replay_buffer(replay_buffer_path, truncate_last_traj=self.truncate_last_trajectory)
        return model

    def _create_sampler(self, sampler_method: str) -> BaseSampler:
        # n_warmup_steps: Disable pruner until the trial reaches the given number of steps.
        if sampler_method == "random":
            sampler: BaseSampler = RandomSampler(seed=self.seed)
        elif sampler_method == "tpe":
            sampler = TPESampler(n_startup_trials=self.n_startup_trials, seed=self.seed, multivariate=True)
        elif sampler_method == "skopt":
            from optuna.integration.skopt import SkoptSampler

            # cf https://scikit-optimize.github.io/#skopt.Optimizer
            # GP: gaussian process
            # Gradient boosted regression: GBRT
            sampler = SkoptSampler(skopt_kwargs={"base_estimator": "GP", "acq_func": "gp_hedge"})
        else:
            raise ValueError(f"Unknown sampler: {sampler_method}")
        return sampler

    def _create_pruner(self, pruner_method: str) -> BasePruner:
        if pruner_method == "halving":
            pruner: BasePruner = SuccessiveHalvingPruner(min_resource=1, reduction_factor=4, min_early_stopping_rate=0)
        elif pruner_method == "median":
            pruner = MedianPruner(n_startup_trials=self.n_startup_trials, n_warmup_steps=self.n_evaluations // 3)
        elif pruner_method == "none":
            # Do not prune
            pruner = NopPruner()
        else:
            raise ValueError(f"Unknown pruner: {pruner_method}")
        return pruner

    def objective(self, trial: optuna.Trial) -> float:
        kwargs = self._hyperparams.copy()

        n_envs = 1 if self.algo == "ars" else self.n_envs

        additional_args = {
            "using_her_replay_buffer": kwargs.get("replay_buffer_class") == HerReplayBuffer,
            "her_kwargs": kwargs.get("replay_buffer_kwargs", {}),
        }
        # Pass n_actions to initialize DDPG/TD3 noise sampler
        # Sample candidate hyperparameters
        sampled_hyperparams = HYPERPARAMS_SAMPLER[self.algo](trial, self.n_actions, n_envs, additional_args)
        kwargs.update(sampled_hyperparams)

        env = self.create_envs(n_envs, no_log=True)

        # By default, do not activate verbose output to keep
        # stdout clean with only the trial's results
        trial_verbosity = 0
        # Activate verbose mode for the trial in debug mode
        # See PR #214
        if self.verbose >= 2:
            trial_verbosity = self.verbose

        model = ALGOS[self.algo](
            env=env,
            tensorboard_log=None,
            # We do not seed the trial
            seed=None,
            verbose=trial_verbosity,
            device=self.device,
            **kwargs,
        )

        eval_env = self.create_envs(n_envs=self.n_eval_envs, eval_env=True)

        optuna_eval_freq = int(self.n_timesteps / self.n_evaluations)
        # Account for parallel envs
        optuna_eval_freq = max(optuna_eval_freq // self.n_envs, 1)
        # Use non-deterministic eval for Atari
        path = None
        if self.optimization_log_path is not None:
            path = os.path.join(self.optimization_log_path, f"trial_{trial.number!s}")
        callbacks = get_callback_list({"callback": self.specified_callbacks})
        eval_callback = TrialEvalCallback(
            eval_env,
            trial,
            best_model_save_path=path,
            log_path=path,
            n_eval_episodes=self.n_eval_episodes,
            eval_freq=optuna_eval_freq,
            deterministic=self.deterministic_eval,
        )
        callbacks.append(eval_callback)

        learn_kwargs = {}
        # Special case for ARS
        if self.algo == "ars" and self.n_envs > 1:
            learn_kwargs["async_eval"] = AsyncEval(
                [lambda: self.create_envs(n_envs=1, no_log=True) for _ in range(self.n_envs)], model.policy
            )

        try:
            model.learn(self.n_timesteps, callback=callbacks, **learn_kwargs)  # type: ignore[arg-type]
            # Free memory
            assert model.env is not None
            model.env.close()
            eval_env.close()
        except (AssertionError, ValueError) as e:
            # Sometimes, random hyperparams can generate NaN
            # Free memory
            assert model.env is not None
            model.env.close()
            eval_env.close()
            # Prune hyperparams that generate NaNs
            print(e)
            print("============")
            print("Sampled hyperparams:")
            pprint(sampled_hyperparams)
            raise optuna.exceptions.TrialPruned() from e
        is_pruned = eval_callback.is_pruned
        reward = eval_callback.last_mean_reward

        del model.env, eval_env
        del model

        if is_pruned:
            raise optuna.exceptions.TrialPruned()

        return reward

    def hyperparameters_optimization(self) -> None:
        if self.verbose > 0:
            print("Optimizing hyperparameters")

        if self.storage is not None and self.study_name is None:
            warnings.warn(
                f"You passed a remote storage: {self.storage} but no `--study-name`."
                "The study name will be generated by Optuna, make sure to re-use the same study name "
                "when you want to do distributed hyperparameter optimization."
            )

        if self.tensorboard_log is not None:
            warnings.warn("Tensorboard log is deactivated when running hyperparameter optimization")
            self.tensorboard_log = None

        # TODO: eval each hyperparams several times to account for noisy evaluation
        sampler = self._create_sampler(self.sampler)
        pruner = self._create_pruner(self.pruner)

        if self.verbose > 0:
            print(f"Sampler: {self.sampler} - Pruner: {self.pruner}")

        study = optuna.create_study(
            sampler=sampler,
            pruner=pruner,
            storage=self.storage,
            study_name=self.study_name,
            load_if_exists=True,
            direction="maximize",
        )

        try:
            if self.max_total_trials is not None:
                # Note: we count already running trials here otherwise we get
                #  (max_total_trials + number of workers) trials in total.
                counted_states = [
                    TrialState.COMPLETE,
                    TrialState.RUNNING,
                    TrialState.PRUNED,
                ]
                completed_trials = len(study.get_trials(states=counted_states))
                if completed_trials < self.max_total_trials:
                    study.optimize(
                        self.objective,
                        n_jobs=self.n_jobs,
                        callbacks=[
                            MaxTrialsCallback(
                                self.max_total_trials,
                                states=counted_states,
                            )
                        ],
                    )
            else:
                study.optimize(self.objective, n_jobs=self.n_jobs, n_trials=self.n_trials)
        except KeyboardInterrupt:
            pass

        print("Number of finished trials: ", len(study.trials))

        print("Best trial:")
        trial = study.best_trial

        print("Value: ", trial.value)

        print("Params: ")
        for key, value in trial.params.items():
            print(f"    {key}: {value}")

        report_name = (
            f"report_{self.env_name}_{self.n_trials}-trials-{self.n_timesteps}"
            f"-{self.sampler}-{self.pruner}_{int(time.time())}"
        )

        log_path = os.path.join(self.log_folder, self.algo, report_name)

        if self.verbose:
            print(f"Writing report to {log_path}")

        # Write report
        os.makedirs(os.path.dirname(log_path), exist_ok=True)
        study.trials_dataframe().to_csv(f"{log_path}.csv")

        # Save python object to inspect/re-use it later
        with open(f"{log_path}.pkl", "wb+") as f:
            pkl.dump(study, f)

        # Skip plots
        if self.no_optim_plots:
            return

        # Plot optimization result
        try:
            fig1 = plot_optimization_history(study)
            fig2 = plot_param_importances(study)

            fig1.show()
            fig2.show()
        except (ValueError, ImportError, RuntimeError):
            pass