def test_channel_first_env(tmp_path): # test_cnn uses environment with HxWxC setup that is transposed, but we # also want to work with CxHxW envs directly without transposing wrapper. SAVE_NAME = "cnn_model.zip" # Create environment with transposed images (CxHxW). # If underlying CNN processes the data in wrong format, # it will raise an error of negative dimension sizes while creating convolutions env = FakeImageEnv(screen_height=40, screen_width=40, n_channels=1, discrete=True, channel_first=True) model = A2C("CnnPolicy", env, n_steps=100).learn(250) assert not is_vecenv_wrapped(model.get_env(), VecTransposeImage) obs = env.reset() action, _ = model.predict(obs, deterministic=True) model.save(tmp_path / SAVE_NAME) del model model = A2C.load(tmp_path / SAVE_NAME) # Check that the prediction is the same assert np.allclose(action, model.predict(obs, deterministic=True)[0]) os.remove(str(tmp_path / SAVE_NAME))
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]) if (is_image_space(env.observation_space) and not is_vecenv_wrapped(env, VecTransposeImage) and not is_image_space_channels_first(env.observation_space)): if verbose >= 1: print("Wrapping the env in a VecTransposeImage.") env = VecTransposeImage(env) # check if wrapper for dict support is needed when using HER if isinstance(env.observation_space, gym.spaces.dict.Dict): env = ObsDictWrapper(env) return env
def test_cnn(tmp_path, model_class): SAVE_NAME = "cnn_model.zip" # Fake grayscale with frameskip # Atari after preprocessing: 84x84x1, here we are using lower resolution # to check that the network handle it automatically env = FakeImageEnv(screen_height=40, screen_width=40, n_channels=1, discrete=model_class not in {SAC, TD3}) if model_class in {A2C, PPO}: kwargs = dict(n_steps=64) else: # Avoid memory error when using replay buffer # Reduce the size of the features kwargs = dict(buffer_size=250, policy_kwargs=dict(features_extractor_kwargs=dict(features_dim=32))) model = model_class("CnnPolicy", env, **kwargs).learn(250) # FakeImageEnv is channel last by default and should be wrapped assert is_vecenv_wrapped(model.get_env(), VecTransposeImage) obs = env.reset() action, _ = model.predict(obs, deterministic=True) model.save(tmp_path / SAVE_NAME) del model model = model_class.load(tmp_path / SAVE_NAME) # Check that the prediction is the same assert np.allclose(action, model.predict(obs, deterministic=True)[0]) os.remove(str(tmp_path / SAVE_NAME))
def test_cnn(tmp_path, model_class): SAVE_NAME = "cnn_model.zip" # Fake grayscale with frameskip # Atari after preprocessing: 84x84x1, here we are using lower resolution # to check that the network handle it automatically env = FakeImageEnv( screen_height=40, screen_width=40, n_channels=1, discrete=model_class not in {TQC}, ) kwargs = {} if model_class in {TQC, QRDQN}: # Avoid memory error when using replay buffer # Reduce the size of the features and the number of quantiles kwargs = dict( buffer_size=250, policy_kwargs=dict( n_quantiles=25, features_extractor_kwargs=dict(features_dim=32), ), ) else: kwargs = dict( buffer_size=250, policy_kwargs=dict(features_extractor_kwargs=dict( features_dim=32)), seed=1, ) model = model_class("CnnPolicy", env, **kwargs).learn(250) obs = env.reset() # FakeImageEnv is channel last by default and should be wrapped assert is_vecenv_wrapped(model.get_env(), VecTransposeImage) # Test stochastic predict with channel last input if model_class in {QRDQN, DQNClipped, DQNReg}: model.exploration_rate = 0.9 for _ in range(10): model.predict(obs, deterministic=False) action, _ = model.predict(obs, deterministic=True) model.save(tmp_path / SAVE_NAME) del model model = model_class.load(tmp_path / SAVE_NAME) # Check that the prediction is the same assert np.allclose(action, model.predict(obs, deterministic=True)[0]) os.remove(str(tmp_path / SAVE_NAME))
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
def _wrap_env(env: GymEnv, verbose: int = 0) -> VecEnv: if not isinstance(env, VecEnv): if verbose >= 1: print("Wrapping the env in a DummyVecEnv.") env = DummyVecEnv([lambda: env]) if (is_image_space(env.observation_space) and not is_vecenv_wrapped(env, VecTransposeImage) and not is_image_space_channels_first(env.observation_space)): if verbose >= 1: print("Wrapping the env in a VecTransposeImage.") env = VecTransposeImage(env) # check if wrapper for dict support is needed when using HER if isinstance(env.observation_space, gym.spaces.dict.Dict): env = ObsDictWrapper(env) 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 monitor_kwargs = {} # Special case for GoalEnvs: log success rate too if "Neck" in self.env_id or self.is_robotics_env( self.env_id) or "parking-v0" in self.env_id: monitor_kwargs = dict(info_keywords=("is_success", )) # On most env, SubprocVecEnv does not help and is quite memory hungry # therefore we use DummyVecEnv by default env = make_vec_env( env_id=self.env_id, n_envs=n_envs, seed=self.seed, env_kwargs=self.env_kwargs, monitor_dir=log_dir, wrapper_class=self.env_wrapper, vec_env_cls=self.vec_env_class, vec_env_kwargs=self.vec_env_kwargs, monitor_kwargs=monitor_kwargs, ) # 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, 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 self.verbose >= 1: print("Wrapping the env in a VecTransposeImage.") env = VecTransposeImage(env) return env
def evaluate_policy( model: "base_class.BaseAlgorithm", env: Union[gym.Env, VecEnv], n_eval_episodes: int = 10, deterministic: bool = True, render: bool = False, callback: Optional[Callable[[Dict[str, Any], Dict[str, Any]], None]] = None, reward_threshold: Optional[float] = None, return_episode_rewards: bool = False, warn: bool = True, ) -> Union[Tuple[float, float], Tuple[List[float], List[int]]]: """ Runs policy for ``n_eval_episodes`` episodes and returns average reward. If a vector env is passed in, this divides the episodes to evaluate onto the different elements of the vector env. This static division of work is done to remove bias. See https://github.com/DLR-RM/stable-baselines3/issues/402 for more details and discussion. .. note:: If environment has not been wrapped with ``Monitor`` wrapper, reward and episode lengths are counted as it appears with ``env.step`` calls. If the environment contains wrappers that modify rewards or episode lengths (e.g. reward scaling, early episode reset), these will affect the evaluation results as well. You can avoid this by wrapping environment with ``Monitor`` wrapper before anything else. :param model: The RL agent you want to evaluate. :param env: The gym environment or ``VecEnv`` environment. :param n_eval_episodes: Number of episode to evaluate the agent :param deterministic: Whether to use deterministic or stochastic actions :param render: Whether to render the environment or not :param callback: callback function to do additional checks, called after each step. Gets locals() and globals() passed as parameters. :param reward_threshold: Minimum expected reward per episode, this will raise an error if the performance is not met :param return_episode_rewards: If True, a list of rewards and episode lengths per episode will be returned instead of the mean. :param warn: If True (default), warns user about lack of a Monitor wrapper in the evaluation environment. :return: Mean reward per episode, std of reward per episode. Returns ([float], [int]) when ``return_episode_rewards`` is True, first list containing per-episode rewards and second containing per-episode lengths (in number of steps). """ is_monitor_wrapped = False # Avoid circular import from stable_baselines3.common.monitor import Monitor if not isinstance(env, VecEnv): env = DummyVecEnv([lambda: env]) is_monitor_wrapped = is_vecenv_wrapped(env, VecMonitor) or env.env_is_wrapped(Monitor)[0] if not is_monitor_wrapped and warn: warnings.warn( "Evaluation environment is not wrapped with a ``Monitor`` wrapper. " "This may result in reporting modified episode lengths and rewards, if other wrappers happen to modify these. " "Consider wrapping environment first with ``Monitor`` wrapper.", UserWarning, ) n_envs = env.num_envs episode_rewards = [] episode_lengths = [] episode_counts = np.zeros(n_envs, dtype="int") # Divides episodes among different sub environments in the vector as evenly as possible episode_count_targets = np.array([(n_eval_episodes + i) // n_envs for i in range(n_envs)], dtype="int") current_rewards = np.zeros(n_envs) current_lengths = np.zeros(n_envs, dtype="int") observations = env.reset() states = None while (episode_counts < episode_count_targets).any(): actions, states = model.predict(observations, state=states, deterministic=deterministic) observations, rewards, dones, infos = env.step(actions) current_rewards += rewards current_lengths += 1 for i in range(n_envs): if episode_counts[i] < episode_count_targets[i]: # unpack values so that the callback can access the local variables reward = rewards[i] done = dones[i] info = infos[i] if callback is not None: callback(locals(), globals()) if dones[i]: if is_monitor_wrapped: # Atari wrapper can send a "done" signal when # the agent loses a life, but it does not correspond # to the true end of episode if "episode" in info.keys(): # Do not trust "done" with episode endings. # Monitor wrapper includes "episode" key in info if environment # has been wrapped with it. Use those rewards instead. episode_rewards.append(info["episode"]["r"]) episode_lengths.append(info["episode"]["l"]) # Only increment at the real end of an episode episode_counts[i] += 1 else: episode_rewards.append(current_rewards[i]) episode_lengths.append(current_lengths[i]) episode_counts[i] += 1 current_rewards[i] = 0 current_lengths[i] = 0 if states is not None: states[i] *= 0 if render: env.render() mean_reward = np.mean(episode_rewards) std_reward = np.std(episode_rewards) if reward_threshold is not None: assert mean_reward > reward_threshold, "Mean reward below threshold: " f"{mean_reward:.2f} < {reward_threshold:.2f}" if return_episode_rewards: return episode_rewards, episode_lengths return mean_reward, std_reward
def evaluate_policy( model: "base_class.BaseAlgorithm", env: Union[gym.Env, VecEnv], n_eval_episodes: int = 10, deterministic: bool = True, render: bool = False, callback: Optional[Callable[[Dict[str, Any], Dict[str, Any]], None]] = None, reward_threshold: Optional[float] = None, return_episode_rewards: bool = False, warn: bool = True, ) -> Union[Tuple[float, float], Tuple[List[float], List[int]]]: """ Runs policy for ``n_eval_episodes`` episodes and returns average reward. This is made to work only with one env. .. note:: If environment has not been wrapped with ``Monitor`` wrapper, reward and episode lengths are counted as it appears with ``env.step`` calls. If the environment contains wrappers that modify rewards or episode lengths (e.g. reward scaling, early episode reset), these will affect the evaluation results as well. You can avoid this by wrapping environment with ``Monitor`` wrapper before anything else. :param model: The RL agent you want to evaluate. :param env: The gym environment. In the case of a ``VecEnv`` this must contain only one environment. :param n_eval_episodes: Number of episode to evaluate the agent :param deterministic: Whether to use deterministic or stochastic actions :param render: Whether to render the environment or not :param callback: callback function to do additional checks, called after each step. Gets locals() and globals() passed as parameters. :param reward_threshold: Minimum expected reward per episode, this will raise an error if the performance is not met :param return_episode_rewards: If True, a list of rewards and episode lengths per episode will be returned instead of the mean. :param warn: If True (default), warns user about lack of a Monitor wrapper in the evaluation environment. :return: Mean reward per episode, std of reward per episode. Returns ([float], [int]) when ``return_episode_rewards`` is True, first list containing per-episode rewards and second containing per-episode lengths (in number of steps). """ is_monitor_wrapped = False # Avoid circular import from stable_baselines3.common.env_util import is_wrapped from stable_baselines3.common.monitor import Monitor assert deterministic == False # print(deterministic) if isinstance(env, VecEnv): assert env.num_envs == 1, "You must pass only one environment when using this function" is_monitor_wrapped = is_vecenv_wrapped(env, VecMonitor) or env.env_is_wrapped(Monitor)[0] else: is_monitor_wrapped = is_wrapped(env, Monitor) if not is_monitor_wrapped and warn: warnings.warn( "Evaluation environment is not wrapped with a ``Monitor`` wrapper. " "This may result in reporting modified episode lengths and rewards, if other wrappers happen to modify these. " "Consider wrapping environment first with ``Monitor`` wrapper.", UserWarning, ) episode_rewards, episode_lengths = [], [] not_reseted = True while len(episode_rewards) < n_eval_episodes: # Number of loops here might differ from true episodes # played, if underlying wrappers modify episode lengths. # Avoid double reset, as VecEnv are reset automatically. if not isinstance(env, VecEnv) or not_reseted: obs = env.reset() not_reseted = False done, state = False, None episode_reward = 0.0 episode_length = 0 while not done: action, state = model.policy.predict(obs, state=state, deterministic=deterministic) obs, reward, done, info = env.step(action) episode_reward += reward if callback is not None: callback(locals(), globals()) episode_length += 1 if render: env.render() if is_monitor_wrapped: # Do not trust "done" with episode endings. # Remove vecenv stacking (if any) if isinstance(env, VecEnv): info = info[0] if "episode" in info.keys(): # Monitor wrapper includes "episode" key in info if environment # has been wrapped with it. Use those rewards instead. episode_rewards.append(info["episode"]["r"]) episode_lengths.append(info["episode"]["l"]) else: episode_rewards.append(episode_reward) episode_lengths.append(episode_length) mean_reward = np.mean(episode_rewards) std_reward = np.std(episode_rewards) if reward_threshold is not None: assert mean_reward > reward_threshold, "Mean reward below threshold: " f"{mean_reward:.2f} < {reward_threshold:.2f}" if return_episode_rewards: return episode_rewards, episode_lengths return mean_reward, std_reward