Source code for stable_baselines3.common.base_class

"""Abstract base classes for RL algorithms."""

import io
import pathlib
import time
from abc import ABC, abstractmethod
from collections import deque
from typing import Any, Dict, Iterable, List, Optional, Tuple, Type, Union

import gym
import numpy as np
import torch as th

from stable_baselines3.common import utils
from stable_baselines3.common.callbacks import BaseCallback, CallbackList, ConvertCallback, EvalCallback
from stable_baselines3.common.env_util import is_wrapped
from stable_baselines3.common.logger import Logger
from stable_baselines3.common.monitor import Monitor
from stable_baselines3.common.noise import ActionNoise
from stable_baselines3.common.policies import BasePolicy, get_policy_from_name
from stable_baselines3.common.preprocessing import check_for_nested_spaces, is_image_space, is_image_space_channels_first
from stable_baselines3.common.save_util import load_from_zip_file, recursive_getattr, recursive_setattr, save_to_zip_file
from stable_baselines3.common.type_aliases import GymEnv, MaybeCallback, Schedule
from stable_baselines3.common.utils import (
    check_for_correct_spaces,
    get_device,
    get_schedule_fn,
    set_random_seed,
    update_learning_rate,
)
from stable_baselines3.common.vec_env import (
    DummyVecEnv,
    VecEnv,
    VecNormalize,
    VecTransposeImage,
    is_vecenv_wrapped,
    unwrap_vec_normalize,
)


def maybe_make_env(env: Union[GymEnv, str, None], verbose: int) -> Optional[GymEnv]:
    """If env is a string, make the environment; otherwise, return env.

    :param env: The environment to learn from.
    :param verbose: logging verbosity
    :return A Gym (vector) environment.
    """
    if isinstance(env, str):
        if verbose >= 1:
            print(f"Creating environment from the given name '{env}'")
        env = gym.make(env)
    return env


[docs]class BaseAlgorithm(ABC): """ The base of RL algorithms :param policy: Policy object :param env: The environment to learn from (if registered in Gym, can be str. Can be None for loading trained models) :param policy_base: The base policy used by this method :param learning_rate: learning rate for the optimizer, it can be a function of the current progress remaining (from 1 to 0) :param policy_kwargs: Additional arguments to be passed to the policy on creation :param tensorboard_log: the log location for tensorboard (if None, no logging) :param verbose: The verbosity level: 0 none, 1 training information, 2 debug :param device: Device on which the code should run. By default, it will try to use a Cuda compatible device and fallback to cpu if it is not possible. :param support_multi_env: Whether the algorithm supports training with multiple environments (as in A2C) :param create_eval_env: Whether to create a second environment that will be used for evaluating the agent periodically. (Only available when passing string for the environment) :param monitor_wrapper: When creating an environment, whether to wrap it or not in a Monitor wrapper. :param seed: Seed for the pseudo random generators :param use_sde: Whether to use generalized State Dependent Exploration (gSDE) instead of action noise exploration (default: False) :param sde_sample_freq: Sample a new noise matrix every n steps when using gSDE Default: -1 (only sample at the beginning of the rollout) :param supported_action_spaces: The action spaces supported by the algorithm. """ def __init__( self, policy: Type[BasePolicy], env: Union[GymEnv, str, None], policy_base: Type[BasePolicy], learning_rate: Union[float, Schedule], policy_kwargs: Optional[Dict[str, Any]] = None, tensorboard_log: Optional[str] = None, verbose: int = 0, device: Union[th.device, str] = "auto", support_multi_env: bool = False, create_eval_env: bool = False, monitor_wrapper: bool = True, seed: Optional[int] = None, use_sde: bool = False, sde_sample_freq: int = -1, supported_action_spaces: Optional[Tuple[gym.spaces.Space, ...]] = None, ): if isinstance(policy, str) and policy_base is not None: self.policy_class = get_policy_from_name(policy_base, policy) else: self.policy_class = policy self.device = get_device(device) if verbose > 0: print(f"Using {self.device} device") self.env = None # type: Optional[GymEnv] # get VecNormalize object if needed self._vec_normalize_env = unwrap_vec_normalize(env) self.verbose = verbose self.policy_kwargs = {} if policy_kwargs is None else policy_kwargs self.observation_space = None # type: Optional[gym.spaces.Space] self.action_space = None # type: Optional[gym.spaces.Space] self.n_envs = None self.num_timesteps = 0 # Used for updating schedules self._total_timesteps = 0 self.eval_env = None self.seed = seed self.action_noise = None # type: Optional[ActionNoise] self.start_time = None self.policy = None self.learning_rate = learning_rate self.tensorboard_log = tensorboard_log self.lr_schedule = None # type: Optional[Schedule] self._last_obs = None # type: Optional[Union[np.ndarray, Dict[str, np.ndarray]]] self._last_episode_starts = None # type: Optional[np.ndarray] # When using VecNormalize: self._last_original_obs = None # type: Optional[Union[np.ndarray, Dict[str, np.ndarray]]] self._episode_num = 0 # Used for gSDE only self.use_sde = use_sde self.sde_sample_freq = sde_sample_freq # Track the training progress remaining (from 1 to 0) # this is used to update the learning rate self._current_progress_remaining = 1 # Buffers for logging self.ep_info_buffer = None # type: Optional[deque] self.ep_success_buffer = None # type: Optional[deque] # For logging (and TD3 delayed updates) self._n_updates = 0 # type: int # The logger object self._logger = None # type: Logger # Whether the user passed a custom logger or not self._custom_logger = False # Create and wrap the env if needed if env is not None: if isinstance(env, str): if create_eval_env: self.eval_env = maybe_make_env(env, self.verbose) env = maybe_make_env(env, self.verbose) env = self._wrap_env(env, self.verbose, monitor_wrapper) self.observation_space = env.observation_space self.action_space = env.action_space self.n_envs = env.num_envs self.env = env if supported_action_spaces is not None: assert isinstance(self.action_space, supported_action_spaces), ( f"The algorithm only supports {supported_action_spaces} as action spaces " f"but {self.action_space} was provided" ) if not support_multi_env and self.n_envs > 1: raise ValueError( "Error: the model does not support multiple envs; it requires " "a single vectorized environment." ) if self.use_sde and not isinstance(self.action_space, gym.spaces.Box): raise ValueError("generalized State-Dependent Exploration (gSDE) can only be used with continuous actions.") @staticmethod def _wrap_env(env: GymEnv, verbose: int = 0, monitor_wrapper: bool = True) -> VecEnv: """ " Wrap environment with the appropriate wrappers if needed. For instance, to have a vectorized environment or to re-order the image channels. :param env: :param verbose: :param monitor_wrapper: Whether to wrap the env in a ``Monitor`` when possible. :return: The wrapped environment. """ if not isinstance(env, VecEnv): if not is_wrapped(env, Monitor) and monitor_wrapper: if verbose >= 1: print("Wrapping the env with a `Monitor` wrapper") env = Monitor(env) if verbose >= 1: print("Wrapping the env in a DummyVecEnv.") env = DummyVecEnv([lambda: env]) # Make sure that dict-spaces are not nested (not supported) check_for_nested_spaces(env.observation_space) if isinstance(env.observation_space, gym.spaces.Dict): for space in env.observation_space.spaces.values(): if isinstance(space, gym.spaces.Dict): raise ValueError("Nested observation spaces are not supported (Dict spaces inside Dict space).") if not is_vecenv_wrapped(env, VecTransposeImage): wrap_with_vectranspose = False if isinstance(env.observation_space, gym.spaces.Dict): # If even one of the keys is a 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) ) else: wrap_with_vectranspose = is_image_space(env.observation_space) and not is_image_space_channels_first( env.observation_space ) if wrap_with_vectranspose: if verbose >= 1: print("Wrapping the env in a VecTransposeImage.") env = VecTransposeImage(env) return env @abstractmethod def _setup_model(self) -> None: """Create networks, buffer and optimizers."""
[docs] def set_logger(self, logger: Logger) -> None: """ Setter for for logger object. .. warning:: When passing a custom logger object, this will overwrite ``tensorboard_log`` and ``verbose`` settings passed to the constructor. """ self._logger = logger # User defined logger self._custom_logger = True
@property def logger(self) -> Logger: """Getter for the logger object.""" return self._logger def _get_eval_env(self, eval_env: Optional[GymEnv]) -> Optional[GymEnv]: """ Return the environment that will be used for evaluation. :param eval_env:) :return: """ if eval_env is None: eval_env = self.eval_env if eval_env is not None: eval_env = self._wrap_env(eval_env, self.verbose) assert eval_env.num_envs == 1 return eval_env def _setup_lr_schedule(self) -> None: """Transform to callable if needed.""" self.lr_schedule = get_schedule_fn(self.learning_rate) def _update_current_progress_remaining(self, num_timesteps: int, total_timesteps: int) -> None: """ Compute current progress remaining (starts from 1 and ends to 0) :param num_timesteps: current number of timesteps :param total_timesteps: """ self._current_progress_remaining = 1.0 - float(num_timesteps) / float(total_timesteps) def _update_learning_rate(self, optimizers: Union[List[th.optim.Optimizer], th.optim.Optimizer]) -> None: """ Update the optimizers learning rate using the current learning rate schedule and the current progress remaining (from 1 to 0). :param optimizers: An optimizer or a list of optimizers. """ # Log the current learning rate self.logger.record("train/learning_rate", self.lr_schedule(self._current_progress_remaining)) if not isinstance(optimizers, list): optimizers = [optimizers] for optimizer in optimizers: update_learning_rate(optimizer, self.lr_schedule(self._current_progress_remaining)) def _excluded_save_params(self) -> List[str]: """ Returns the names of the parameters that should be excluded from being saved by pickling. E.g. replay buffers are skipped by default as they take up a lot of space. PyTorch variables should be excluded with this so they can be stored with ``th.save``. :return: List of parameters that should be excluded from being saved with pickle. """ return [ "policy", "device", "env", "eval_env", "replay_buffer", "rollout_buffer", "_vec_normalize_env", "_episode_storage", "_logger", "_custom_logger", ] def _get_torch_save_params(self) -> Tuple[List[str], List[str]]: """ Get the name of the torch variables that will be saved with PyTorch ``th.save``, ``th.load`` and ``state_dicts`` instead of the default pickling strategy. This is to handle device placement correctly. Names can point to specific variables under classes, e.g. "policy.optimizer" would point to ``optimizer`` object of ``self.policy`` if this object. :return: List of Torch variables whose state dicts to save (e.g. th.nn.Modules), and list of other Torch variables to store with ``th.save``. """ state_dicts = ["policy"] return state_dicts, [] def _init_callback( self, callback: MaybeCallback, eval_env: Optional[VecEnv] = None, eval_freq: int = 10000, n_eval_episodes: int = 5, log_path: Optional[str] = None, ) -> BaseCallback: """ :param callback: Callback(s) called at every step with state of the algorithm. :param eval_freq: How many steps between evaluations; if None, do not evaluate. :param n_eval_episodes: How many episodes to play per evaluation :param n_eval_episodes: Number of episodes to rollout during evaluation. :param log_path: Path to a folder where the evaluations will be saved :return: A hybrid callback calling `callback` and performing evaluation. """ # Convert a list of callbacks into a callback if isinstance(callback, list): callback = CallbackList(callback) # Convert functional callback to object if not isinstance(callback, BaseCallback): callback = ConvertCallback(callback) # Create eval callback in charge of the evaluation if eval_env is not None: eval_callback = EvalCallback( eval_env, best_model_save_path=log_path, log_path=log_path, eval_freq=eval_freq, n_eval_episodes=n_eval_episodes, ) callback = CallbackList([callback, eval_callback]) callback.init_callback(self) return callback def _setup_learn( self, total_timesteps: int, eval_env: Optional[GymEnv], callback: MaybeCallback = None, eval_freq: int = 10000, n_eval_episodes: int = 5, log_path: Optional[str] = None, reset_num_timesteps: bool = True, tb_log_name: str = "run", ) -> Tuple[int, BaseCallback]: """ Initialize different variables needed for training. :param total_timesteps: The total number of samples (env steps) to train on :param eval_env: Environment to use for evaluation. :param callback: Callback(s) called at every step with state of the algorithm. :param eval_freq: How many steps between evaluations :param n_eval_episodes: How many episodes to play per evaluation :param log_path: Path to a folder where the evaluations will be saved :param reset_num_timesteps: Whether to reset or not the ``num_timesteps`` attribute :param tb_log_name: the name of the run for tensorboard log :return: """ self.start_time = time.time() if self.ep_info_buffer is None or reset_num_timesteps: # Initialize buffers if they don't exist, or reinitialize if resetting counters self.ep_info_buffer = deque(maxlen=100) self.ep_success_buffer = deque(maxlen=100) if self.action_noise is not None: self.action_noise.reset() if reset_num_timesteps: self.num_timesteps = 0 self._episode_num = 0 else: # Make sure training timesteps are ahead of the internal counter total_timesteps += self.num_timesteps self._total_timesteps = total_timesteps # Avoid resetting the environment when calling ``.learn()`` consecutive times if reset_num_timesteps or self._last_obs is None: self._last_obs = self.env.reset() # pytype: disable=annotation-type-mismatch self._last_episode_starts = np.ones((self.env.num_envs,), dtype=bool) # Retrieve unnormalized observation for saving into the buffer if self._vec_normalize_env is not None: self._last_original_obs = self._vec_normalize_env.get_original_obs() if eval_env is not None and self.seed is not None: eval_env.seed(self.seed) eval_env = self._get_eval_env(eval_env) # Configure logger's outputs if no logger was passed if not self._custom_logger: self._logger = utils.configure_logger(self.verbose, self.tensorboard_log, tb_log_name, reset_num_timesteps) # Create eval callback if needed callback = self._init_callback(callback, eval_env, eval_freq, n_eval_episodes, log_path) return total_timesteps, callback def _update_info_buffer(self, infos: List[Dict[str, Any]], dones: Optional[np.ndarray] = None) -> None: """ Retrieve reward, episode length, episode success and update the buffer if using Monitor wrapper or a GoalEnv. :param infos: List of additional information about the transition. :param dones: Termination signals """ if dones is None: dones = np.array([False] * len(infos)) for idx, info in enumerate(infos): maybe_ep_info = info.get("episode") maybe_is_success = info.get("is_success") if maybe_ep_info is not None: self.ep_info_buffer.extend([maybe_ep_info]) if maybe_is_success is not None and dones[idx]: self.ep_success_buffer.append(maybe_is_success)
[docs] def get_env(self) -> Optional[VecEnv]: """ Returns the current environment (can be None if not defined). :return: The current environment """ return self.env
[docs] def get_vec_normalize_env(self) -> Optional[VecNormalize]: """ Return the ``VecNormalize`` wrapper of the training env if it exists. :return: The ``VecNormalize`` env. """ return self._vec_normalize_env
[docs] def set_env(self, env: GymEnv) -> None: """ Checks the validity of the environment, and if it is coherent, set it as the current environment. Furthermore wrap any non vectorized env into a vectorized checked parameters: - observation_space - action_space :param env: The environment for learning a policy """ # if it is not a VecEnv, make it a VecEnv # and do other transformations (dict obs, image transpose) if needed env = self._wrap_env(env, self.verbose) # Check that the observation spaces match check_for_correct_spaces(env, self.observation_space, self.action_space) self.n_envs = env.num_envs self.env = env
[docs] @abstractmethod def learn( self, total_timesteps: int, callback: MaybeCallback = None, log_interval: int = 100, tb_log_name: str = "run", eval_env: Optional[GymEnv] = None, eval_freq: int = -1, n_eval_episodes: int = 5, eval_log_path: Optional[str] = None, reset_num_timesteps: bool = True, ) -> "BaseAlgorithm": """ Return a trained model. :param total_timesteps: The total number of samples (env steps) to train on :param callback: callback(s) called at every step with state of the algorithm. :param log_interval: The number of timesteps before logging. :param tb_log_name: the name of the run for TensorBoard logging :param eval_env: Environment that will be used to evaluate the agent :param eval_freq: Evaluate the agent every ``eval_freq`` timesteps (this may vary a little) :param n_eval_episodes: Number of episode to evaluate the agent :param eval_log_path: Path to a folder where the evaluations will be saved :param reset_num_timesteps: whether or not to reset the current timestep number (used in logging) :return: the trained model """
[docs] def predict( self, observation: np.ndarray, state: Optional[np.ndarray] = None, mask: Optional[np.ndarray] = None, deterministic: bool = False, ) -> Tuple[np.ndarray, Optional[np.ndarray]]: """ Get the model's action(s) from an observation :param observation: the input observation :param state: The last states (can be None, used in recurrent policies) :param mask: The last masks (can be None, used in recurrent policies) :param deterministic: Whether or not to return deterministic actions. :return: the model's action and the next state (used in recurrent policies) """ return self.policy.predict(observation, state, mask, deterministic)
[docs] def set_random_seed(self, seed: Optional[int] = None) -> None: """ Set the seed of the pseudo-random generators (python, numpy, pytorch, gym, action_space) :param seed: """ if seed is None: return set_random_seed(seed, using_cuda=self.device.type == th.device("cuda").type) self.action_space.seed(seed) if self.env is not None: self.env.seed(seed) if self.eval_env is not None: self.eval_env.seed(seed)
[docs] def set_parameters( self, load_path_or_dict: Union[str, Dict[str, Dict]], exact_match: bool = True, device: Union[th.device, str] = "auto", ) -> None: """ Load parameters from a given zip-file or a nested dictionary containing parameters for different modules (see ``get_parameters``). :param load_path_or_iter: Location of the saved data (path or file-like, see ``save``), or a nested dictionary containing nn.Module parameters used by the policy. The dictionary maps object names to a state-dictionary returned by ``torch.nn.Module.state_dict()``. :param exact_match: If True, the given parameters should include parameters for each module and each of their parameters, otherwise raises an Exception. If set to False, this can be used to update only specific parameters. :param device: Device on which the code should run. """ params = None if isinstance(load_path_or_dict, dict): params = load_path_or_dict else: _, params, _ = load_from_zip_file(load_path_or_dict, device=device) # Keep track which objects were updated. # `_get_torch_save_params` returns [params, other_pytorch_variables]. # We are only interested in former here. objects_needing_update = set(self._get_torch_save_params()[0]) updated_objects = set() for name in params: attr = None try: attr = recursive_getattr(self, name) except Exception: # What errors recursive_getattr could throw? KeyError, but # possible something else too (e.g. if key is an int?). # Catch anything for now. raise ValueError(f"Key {name} is an invalid object name.") if isinstance(attr, th.optim.Optimizer): # Optimizers do not support "strict" keyword... # Seems like they will just replace the whole # optimizer state with the given one. # On top of this, optimizer state-dict # seems to change (e.g. first ``optim.step()``), # which makes comparing state dictionary keys # invalid (there is also a nesting of dictionaries # with lists with dictionaries with ...), adding to the # mess. # # TL;DR: We might not be able to reliably say # if given state-dict is missing keys. # # Solution: Just load the state-dict as is, and trust # the user has provided a sensible state dictionary. attr.load_state_dict(params[name]) else: # Assume attr is th.nn.Module attr.load_state_dict(params[name], strict=exact_match) updated_objects.add(name) if exact_match and updated_objects != objects_needing_update: raise ValueError( "Names of parameters do not match agents' parameters: " f"expected {objects_needing_update}, got {updated_objects}" )
[docs] @classmethod def load( cls, path: Union[str, pathlib.Path, io.BufferedIOBase], env: Optional[GymEnv] = None, device: Union[th.device, str] = "auto", custom_objects: Optional[Dict[str, Any]] = None, **kwargs, ) -> "BaseAlgorithm": """ Load the model from a zip-file :param path: path to the file (or a file-like) where to load the agent from :param env: the new environment to run the loaded model on (can be None if you only need prediction from a trained model) has priority over any saved environment :param device: Device on which the code should run. :param custom_objects: Dictionary of objects to replace upon loading. If a variable is present in this dictionary as a key, it will not be deserialized and the corresponding item will be used instead. Similar to custom_objects in ``keras.models.load_model``. Useful when you have an object in file that can not be deserialized. :param kwargs: extra arguments to change the model when loading """ data, params, pytorch_variables = load_from_zip_file(path, device=device, custom_objects=custom_objects) # Remove stored device information and replace with ours if "policy_kwargs" in data: if "device" in data["policy_kwargs"]: del data["policy_kwargs"]["device"] if "policy_kwargs" in kwargs and kwargs["policy_kwargs"] != data["policy_kwargs"]: raise ValueError( f"The specified policy kwargs do not equal the stored policy kwargs." f"Stored kwargs: {data['policy_kwargs']}, specified kwargs: {kwargs['policy_kwargs']}" ) if "observation_space" not in data or "action_space" not in data: raise KeyError("The observation_space and action_space were not given, can't verify new environments") if env is not None: # Wrap first if needed env = cls._wrap_env(env, data["verbose"]) # Check if given env is valid check_for_correct_spaces(env, data["observation_space"], data["action_space"]) else: # Use stored env, if one exists. If not, continue as is (can be used for predict) if "env" in data: env = data["env"] # noinspection PyArgumentList model = cls( # pytype: disable=not-instantiable,wrong-keyword-args policy=data["policy_class"], env=env, device=device, _init_setup_model=False, # pytype: disable=not-instantiable,wrong-keyword-args ) # load parameters model.__dict__.update(data) model.__dict__.update(kwargs) model._setup_model() # put state_dicts back in place model.set_parameters(params, exact_match=True, device=device) # put other pytorch variables back in place if pytorch_variables is not None: for name in pytorch_variables: # Skip if PyTorch variable was not defined (to ensure backward compatibility). # This happens when using SAC/TQC. # SAC has an entropy coefficient which can be fixed or optimized. # If it is optimized, an additional PyTorch variable `log_ent_coef` is defined, # otherwise it is initialized to `None`. if pytorch_variables[name] is None: continue # Set the data attribute directly to avoid issue when using optimizers # See https://github.com/DLR-RM/stable-baselines3/issues/391 recursive_setattr(model, name + ".data", pytorch_variables[name].data) # Sample gSDE exploration matrix, so it uses the right device # see issue #44 if model.use_sde: model.policy.reset_noise() # pytype: disable=attribute-error return model
[docs] def get_parameters(self) -> Dict[str, Dict]: """ Return the parameters of the agent. This includes parameters from different networks, e.g. critics (value functions) and policies (pi functions). :return: Mapping of from names of the objects to PyTorch state-dicts. """ state_dicts_names, _ = self._get_torch_save_params() params = {} for name in state_dicts_names: attr = recursive_getattr(self, name) # Retrieve state dict params[name] = attr.state_dict() return params
[docs] def save( self, path: Union[str, pathlib.Path, io.BufferedIOBase], exclude: Optional[Iterable[str]] = None, include: Optional[Iterable[str]] = None, ) -> None: """ Save all the attributes of the object and the model parameters in a zip-file. :param path: path to the file where the rl agent should be saved :param exclude: name of parameters that should be excluded in addition to the default ones :param include: name of parameters that might be excluded but should be included anyway """ # Copy parameter list so we don't mutate the original dict data = self.__dict__.copy() # Exclude is union of specified parameters (if any) and standard exclusions if exclude is None: exclude = [] exclude = set(exclude).union(self._excluded_save_params()) # Do not exclude params if they are specifically included if include is not None: exclude = exclude.difference(include) state_dicts_names, torch_variable_names = self._get_torch_save_params() all_pytorch_variables = state_dicts_names + torch_variable_names for torch_var in all_pytorch_variables: # We need to get only the name of the top most module as we'll remove that var_name = torch_var.split(".")[0] # Any params that are in the save vars must not be saved by data exclude.add(var_name) # Remove parameter entries of parameters which are to be excluded for param_name in exclude: data.pop(param_name, None) # Build dict of torch variables pytorch_variables = None if torch_variable_names is not None: pytorch_variables = {} for name in torch_variable_names: attr = recursive_getattr(self, name) pytorch_variables[name] = attr # Build dict of state_dicts params_to_save = self.get_parameters() save_to_zip_file(path, data=data, params=params_to_save, pytorch_variables=pytorch_variables)