def _store_transition( self, replay_buffer: ReplayBuffer, buffer_action: np.ndarray, new_obs: np.ndarray, reward: np.ndarray, done: np.ndarray, infos: List[Dict[str, Any]], ) -> None: """ Store transition in the replay buffer. We store the normalized action and the unnormalized observation. It also handles terminal observations (because VecEnv resets automatically). :param replay_buffer: Replay buffer object where to store the transition. :param buffer_action: normalized action :param new_obs: next observation in the current episode or first observation of the episode (when done is True) :param reward: reward for the current transition :param done: Termination signal :param infos: List of additional information about the transition. It may contain the terminal observations and information about timeout. """ # Store only the unnormalized version if self._vec_normalize_env is not None: new_obs_ = self._vec_normalize_env.get_original_obs() reward_ = self._vec_normalize_env.get_original_reward() else: # Avoid changing the original ones self._last_original_obs, new_obs_, reward_ = self._last_obs, new_obs, reward # As the VecEnv resets automatically, new_obs is already the # first observation of the next episode if done and infos[0].get("terminal_observation") is not None: next_obs = infos[0]["terminal_observation"] # VecNormalize normalizes the terminal observation if self._vec_normalize_env is not None: next_obs = self._vec_normalize_env.unnormalize_obs(next_obs) else: next_obs = new_obs_ replay_buffer.add( self._last_original_obs, next_obs, buffer_action, reward_, done, infos, ) self._last_obs = new_obs # Save the unnormalized observation if self._vec_normalize_env is not None: self._last_original_obs = new_obs_
def _store_transition_not_done_if_truncated( self, replay_buffer: ReplayBuffer, buffer_action: np.ndarray, new_obs: np.ndarray, reward: np.ndarray, done: np.ndarray, infos: List[Dict[str, Any]], ) -> None: """ Store transition in the replay buffer. We store the normalized action and the unnormalized observation. It also handles terminal observations (because VecEnv resets automatically). :param replay_buffer: Replay buffer object where to store the transition. :param buffer_action: normalized action :param new_obs: next observation in the current episode or first observation of the episode (when done is True) :param reward: reward for the current transition :param done: Termination signal :param infos: List of additional information about the transition. It contains the terminal observations. """ # Store only the unnormalized version if self._vec_normalize_env is not None: new_obs_ = self._vec_normalize_env.get_original_obs() reward_ = self._vec_normalize_env.get_original_reward() else: # Avoid changing the original ones self._last_original_obs, new_obs_, reward_ = self._last_obs, new_obs, reward # As the VecEnv resets automatically, new_obs is already the # first observation of the next episode if done and infos[0].get("terminal_observation") is not None: next_obs = infos[0]["terminal_observation"] # VecNormalize normalizes the terminal observation if self._vec_normalize_env is not None: next_obs = self._vec_normalize_env.unnormalize_obs(next_obs) else: next_obs = new_obs_ # NOTE: The monkey patch is inside the following block of code done_ = np.array([False]) if infos[0].get("TimeLimit.truncated", False) else done replay_buffer.add(self._last_original_obs, next_obs, buffer_action, reward_, done_) # replay_buffer.add(self._last_original_obs, next_obs, buffer_action, reward_, done) self._last_obs = new_obs # Save the unnormalized observation if self._vec_normalize_env is not None: self._last_original_obs = new_obs_
def learn(self, initial_models): mesa_algo = TD3( "MlpPolicy", self.env, verbose=1, learning_starts=1 ) # Note: Unecessarily initializes parameters (could speed up a bit by fixing)' mesa_algo.set_parameters(to_torch(initial_models), exact_match=False) LOG_DIR = "/home/jet/catkin_ws/src/marsha/marsha_ai/training/logs/" MODEL_DIR = "/home/jet/catkin_ws/src/marsha/marsha_ai/training/models/" callback_list = [] callback_list.append(TensorboardCallback()) callback_list.append( StopTrainingOnMaxEpisodes(max_episodes=5, verbose=1)) """callback_list.append(EvalCallback(self.env, best_model_save_path=MODEL_DIR, log_path=LOG_DIR, deterministic=True, eval_freq=5, n_eval_episodes=1))""" mesa_algo.learn(total_timesteps=1000, callback=callback_list ) #rospy.get_param("/hyperparameters/total_timesteps") print("finished training! Testing mesa network...") test_buffer = ReplayBuffer(100, TaskEnv.observation_space, TaskEnv.action_space, device="cuda") test_env = Monitor(self.env) done = False ob = test_env.reset() while not done: action, state = mesa_algo.predict(ob) next_ob, reward, done, info = test_env.step(action) test_buffer.add(ob, next_ob, action, reward, done, [info]) ob = next_ob meta_buffer = {"test": test_buffer, "train": mesa_algo.replay_buffer} optimized_mesa_parameters = mesa_algo.get_parameters() tf_mesa_models = from_torch(optimized_mesa_parameters) return meta_buffer, tf_mesa_models
# TRY NOT TO MODIFY: record rewards for plotting purposes for info in infos: if "episode" in info.keys(): print( f"global_step={global_step}, episode_reward={info['episode']['r']}" ) writer.add_scalar("charts/episodic_return", info["episode"]["r"], global_step) break # TRY NOT TO MODIFY: save data to reply buffer; handle `terminal_observation` real_next_obs = next_obs.copy() for idx, d in enumerate(dones): if d: real_next_obs[idx] = infos[idx]["terminal_observation"] rb.add(obs, real_next_obs, actions, rewards, dones, infos) # TRY NOT TO MODIFY: CRUCIAL step easy to overlook obs = next_obs # ALGO LOGIC: training. if global_step > args.learning_starts: data = rb.sample(args.batch_size) with torch.no_grad(): clipped_noise = (torch.randn_like(torch.Tensor(actions[0])) * args.policy_noise).clamp( -args.noise_clip, args.noise_clip) next_state_actions = ( target_actor.forward(data.next_observations) + clipped_noise.to(device)).clamp(
def main(args): if args.env_type == 'atari': # env = make_atari_env( # env_name=args.env_name, # action_repeat=args.action_repeat, # frame_stack=args.frame_stack # ) # eval_env = make_atari_env( # env_name=args.env_name, # action_repeat=args.action_repeat, # frame_stack=args.frame_stack # ) env = make_atari_env(env_id=args.env_name) eval_env = make_atari_env(env_id=args.env_name) elif args.env_type == 'dmc_locomotion': env = make_locomotion_env( env_name=args.env_name, seed=args.seed, episode_length=args.episode_length, from_pixels=args.pixel_obs, action_repeat=args.action_repeat, obs_height=args.obs_height, obs_width=args.obs_width, camera_id=args.env_camera_id, mode=args.mode ) eval_env = make_locomotion_env( env_name=args.env_name, seed=args.seed, episode_length=args.episode_length, from_pixels=args.pixel_obs, action_repeat=args.action_repeat, obs_height=args.obs_height, obs_width=args.obs_width, camera_id=args.env_camera_id, mode=args.mode ) # Initialize environment if args.seed is None: args.seed = np.random.randint(int(1e9)) utils.set_seed_everywhere(args.seed, env=env, eval_env=eval_env) utils.make_dir(args.work_dir) model_dir = utils.make_dir(os.path.join(args.work_dir, 'model')) video_dir = utils.make_dir(os.path.join(args.work_dir, 'video')) video = VideoRecorder(video_dir if args.save_video else None, args.env_type, height=448, width=448, camera_id=args.video_camera_id) # Prepare agent assert torch.cuda.is_available(), 'must have cuda enabled' device = torch.device(args.device) if args.env_type == 'atari': # replay_buffer = buffers.FrameStackReplayBuffer( # obs_space=env.observation_space, # action_space=env.action_space, # capacity=args.replay_buffer_capacity, # frame_stack=args.frame_stack, # device=device, # optimize_memory_usage=True, # ) from stable_baselines3.common.buffers import ReplayBuffer replay_buffer = ReplayBuffer( args.replay_buffer_capacity, env.observation_space, env.action_space, device, optimize_memory_usage=True, ) # replay_buffer = buffers.ReplayBuffer( # obs_space=env.observation_space, # action_space=env.action_space, # capacity=args.replay_buffer_capacity, # device=device, # optimize_memory_usage=True, # ) elif args.env_type == 'dmc_locomotion': replay_buffer = buffers.ReplayBuffer( obs_space=env.observation_space, action_space=env.action_space, capacity=args.replay_buffer_capacity, device=device, optimize_memory_usage=True, ) agent = make_agent( obs_space=env.observation_space, action_space=env.action_space, device=device, args=args ) logger = Logger(args.work_dir, log_frequency=args.log_freq, action_repeat=args.action_repeat, save_tb=args.save_tb) episode, episode_reward, episode_step, done = 0, 0, 0, True obs = env.reset() start_time = time.time() for step in range(args.train_steps + 1): # (chongyi zheng): we can also evaluate and save model when current episode is not finished # Evaluate agent periodically if step % args.eval_freq == 0: print('Evaluating:', args.work_dir) logger.log('eval/episode', episode, step) evaluate(eval_env, agent, video, args.num_eval_episodes, logger, step) # Save agent periodically if step % args.save_freq == 0 and step > 0: if args.save_model: agent.save(model_dir, step) if done: if step > 0: logger.log('train/duration', time.time() - start_time, step) start_time = time.time() logger.dump(step, ty='train', save=(step > args.init_steps)) logger.log('train/episode_reward', episode_reward, step) obs = env.reset() done = False episode_reward = 0 episode_step = 0 episode += 1 logger.log('train/episode', episode, step) # Sample action for data collection if step < args.init_steps: action = env.action_space.sample() else: # with utils.eval_mode(agent): action = agent.act(obs, True) # action, _ = agent.predict(obs, deterministic=False) if 'dqn' in args.algo: agent.schedule_exploration_rate(step, logger) # agent._update_current_progress_remaining(step, args.train_steps) # agent._on_step() # Run training update if step >= args.init_steps and step % args.train_freq == 0: # TODO (chongyi zheng): Do we need multiple updates after initial data collection? # num_updates = args.init_steps if step == args.init_steps else 1 # for _ in range(num_updates): # agent.update(replay_buffer, logger, step) for _ in range(args.num_train_iters): agent.update(replay_buffer, logger, step) # agent.train(batch_size=args.batch_size, gradient_steps=args.num_train_iters) # Take step next_obs, reward, done, info = env.step(action) # agent.replay_buffer.add(obs, next_obs, action, reward, done) # replay_buffer.add(obs, action, reward, next_obs, done) replay_buffer.add(np.expand_dims(np.asarray(obs), axis=0), np.expand_dims(np.asarray(next_obs), axis=0), np.expand_dims(action, axis=0), np.expand_dims(np.sign(reward), axis=0), np.expand_dims(done, axis=0)) episode_reward += reward obs = next_obs episode_step += 1
class ReservoirOffPolicyAlgorithm(BaseAlgorithm): """ The base for Off-Policy algorithms (ex: SAC/TD3) :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 buffer_size: size of the replay buffer :param learning_starts: how many steps of the model to collect transitions for before learning starts :param batch_size: Minibatch size for each gradient update :param tau: the soft update coefficient ("Polyak update", between 0 and 1) :param gamma: the discount factor :param train_freq: Update the model every ``train_freq`` steps. Set to `-1` to disable. :param gradient_steps: How many gradient steps to do after each rollout (see ``train_freq`` and ``n_episodes_rollout``) Set to ``-1`` means to do as many gradient steps as steps done in the environment during the rollout. :param n_episodes_rollout: Update the model every ``n_episodes_rollout`` episodes. Note that this cannot be used at the same time as ``train_freq``. Set to `-1` to disable. :param action_noise: the action noise type (None by default), this can help for hard exploration problem. Cf common.noise for the different action noise type. :param optimize_memory_usage: Enable a memory efficient variant of the replay buffer at a cost of more complexity. See https://github.com/DLR-RM/stable-baselines3/issues/37#issuecomment-637501195 :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 State Dependent Exploration (SDE) 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 use_sde_at_warmup: Whether to use gSDE instead of uniform sampling during the warm up phase (before learning starts) :param sde_support: Whether the model support gSDE or not :param remove_time_limit_termination: Remove terminations (dones) that are due to time limit. See https://github.com/hill-a/stable-baselines/issues/863 """ def __init__( self, policy: Type[BasePolicy], env: Union[GymEnv, str], policy_base: Type[BasePolicy], learning_rate: Union[float, Callable], buffer_size: int = int(1e6), learning_starts: int = 100, batch_size: int = 256, tau: float = 0.005, gamma: float = 0.99, train_freq: int = 1, gradient_steps: int = 1, n_episodes_rollout: int = -1, action_noise: Optional[ActionNoise] = None, optimize_memory_usage: bool = False, policy_kwargs: 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, use_sde_at_warmup: bool = False, sde_support: bool = True, remove_time_limit_termination: bool = False, ): super(ReservoirOffPolicyAlgorithm, self).__init__( policy=policy, env=env, policy_base=policy_base, learning_rate=learning_rate, policy_kwargs=policy_kwargs, tensorboard_log=tensorboard_log, verbose=verbose, device=device, support_multi_env=support_multi_env, create_eval_env=create_eval_env, monitor_wrapper=monitor_wrapper, seed=seed, use_sde=use_sde, sde_sample_freq=sde_sample_freq, ) self.buffer_size = buffer_size self.batch_size = batch_size self.learning_starts = learning_starts self.tau = tau self.gamma = gamma self.train_freq = train_freq self.gradient_steps = gradient_steps self.n_episodes_rollout = n_episodes_rollout self.action_noise = action_noise self.optimize_memory_usage = optimize_memory_usage # Remove terminations (dones) that are due to time limit # see https://github.com/hill-a/stable-baselines/issues/863 self.remove_time_limit_termination = remove_time_limit_termination if train_freq > 0 and n_episodes_rollout > 0: warnings.warn( "You passed a positive value for `train_freq` and `n_episodes_rollout`." "Please make sure this is intended. " "The agent will collect data by stepping in the environment " "until both conditions are true: " "`number of steps in the env` >= `train_freq` and " "`number of episodes` > `n_episodes_rollout`" ) self.actor = None # type: Optional[th.nn.Module] self.replay_buffer = None # type: Optional[ReservoirBuffer] self.current_experience_buffer = None # type: Optional[ReplayBuffer] # Update policy keyword arguments if sde_support: self.policy_kwargs["use_sde"] = self.use_sde # For gSDE only self.use_sde_at_warmup = use_sde_at_warmup self.update_env(env, support_multi_env=support_multi_env, create_eval_env=create_eval_env, monitor_wrapper=monitor_wrapper, is_reservoir_replay=True) def _setup_model(self) -> None: self._setup_lr_schedule() self.set_random_seed(self.seed) self.replay_buffer = ReservoirBuffer( self.buffer_size, self.observation_space, self.action_space, self.device, optimize_memory_usage=self.optimize_memory_usage, ) self.current_experience_buffer = ReplayBuffer( 1, self.observation_space, self.action_space, self.device, optimize_memory_usage=False, ) self.policy = self.policy_class( self.observation_space, self.action_space, self.lr_schedule, **self.policy_kwargs # pytype:disable=not-instantiable ) self.policy = self.policy.to(self.device) def save_replay_buffer(self, path: Union[str, pathlib.Path, io.BufferedIOBase]) -> None: """ Save the replay buffer as a pickle file. :param path: Path to the file where the replay buffer should be saved. if path is a str or pathlib.Path, the path is automatically created if necessary. """ assert self.replay_buffer is not None, "The replay buffer is not defined" save_to_pkl(path, self.replay_buffer, self.verbose) def load_replay_buffer(self, path: Union[str, pathlib.Path, io.BufferedIOBase]) -> None: """ Load a replay buffer from a pickle file. :param path: Path to the pickled replay buffer. """ self.replay_buffer = load_from_pkl(path, self.verbose) assert isinstance(self.replay_buffer, ReplayBuffer), "The replay buffer must inherit from ReplayBuffer class" def _setup_learn( self, total_timesteps: int, eval_env: Optional[GymEnv], callback: Union[None, Callable, List[BaseCallback], BaseCallback] = 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]: """ cf `BaseAlgorithm`. """ # Prevent continuity issue by truncating trajectory # when using memory efficient replay buffer # see https://github.com/DLR-RM/stable-baselines3/issues/46 truncate_last_traj = ( self.optimize_memory_usage and reset_num_timesteps and self.replay_buffer is not None and (self.replay_buffer.full or self.replay_buffer.pos > 0) ) if truncate_last_traj: warnings.warn( "The last trajectory in the replay buffer will be truncated, " "see https://github.com/DLR-RM/stable-baselines3/issues/46." "You should use `reset_num_timesteps=False` or `optimize_memory_usage=False`" "to avoid that issue." ) # Go to the previous index pos = (self.replay_buffer.pos - 1) % self.replay_buffer.buffer_size self.replay_buffer.dones[pos] = True return super()._setup_learn( total_timesteps, eval_env, callback, eval_freq, n_eval_episodes, log_path, reset_num_timesteps, tb_log_name ) def learn( self, total_timesteps: int, callback: MaybeCallback = None, log_interval: int = 4, eval_env: Optional[GymEnv] = None, eval_freq: int = -1, n_eval_episodes: int = 5, tb_log_name: str = "run", eval_log_path: Optional[str] = None, reset_num_timesteps: bool = True, ) -> "ReservoirOffPolicyAlgorithm": total_timesteps, callback = self._setup_learn( total_timesteps, eval_env, callback, eval_freq, n_eval_episodes, eval_log_path, reset_num_timesteps, tb_log_name ) callback.on_training_start(locals(), globals()) while self.num_timesteps < total_timesteps: rollout = self.collect_rollouts( self.env, n_episodes=self.n_episodes_rollout, n_steps=self.train_freq, action_noise=self.action_noise, callback=callback, learning_starts=self.learning_starts, replay_buffer=self.replay_buffer, log_interval=log_interval, ) if rollout.continue_training is False: break if self.num_timesteps > 0 and self.num_timesteps > self.learning_starts: # If no `gradient_steps` is specified, # do as many gradients steps as steps performed during the rollout gradient_steps = self.gradient_steps if self.gradient_steps > 0 else rollout.episode_timesteps self.train(batch_size=self.batch_size, gradient_steps=gradient_steps) callback.on_training_end() return self def train(self, gradient_steps: int, batch_size: int) -> None: """ Sample the replay buffer and do the updates (gradient descent and update target networks) """ raise NotImplementedError() def _sample_action( self, learning_starts: int, action_noise: Optional[ActionNoise] = None ) -> Tuple[np.ndarray, np.ndarray]: """ Sample an action according to the exploration policy. This is either done by sampling the probability distribution of the policy, or sampling a random action (from a uniform distribution over the action space) or by adding noise to the deterministic output. :param action_noise: Action noise that will be used for exploration Required for deterministic policy (e.g. TD3). This can also be used in addition to the stochastic policy for SAC. :param learning_starts: Number of steps before learning for the warm-up phase. :return: action to take in the environment and scaled action that will be stored in the replay buffer. The two differs when the action space is not normalized (bounds are not [-1, 1]). """ # Select action randomly or according to policy if self.num_timesteps < learning_starts and not (self.use_sde and self.use_sde_at_warmup): # Warmup phase unscaled_action = np.array([self.action_space.sample()]) else: # Note: when using continuous actions, # we assume that the policy uses tanh to scale the action # We use non-deterministic action in the case of SAC, for TD3, it does not matter unscaled_action, _ = self.predict(self._last_obs, deterministic=False) # Rescale the action from [low, high] to [-1, 1] if isinstance(self.action_space, gym.spaces.Box): scaled_action = self.policy.scale_action(unscaled_action) # Add noise to the action (improve exploration) if action_noise is not None: scaled_action = np.clip(scaled_action + action_noise(), -1, 1) # We store the scaled action in the buffer buffer_action = scaled_action action = self.policy.unscale_action(scaled_action) else: # Discrete case, no need to normalize or clip buffer_action = unscaled_action action = buffer_action return action, buffer_action def _dump_logs(self) -> None: """ Write log. """ try: fps = int(self.num_timesteps / (time.time() - self.start_time)) except ZeroDivisionError: warnings.warn("fps dump had zero division somehow, storing 0 instead.") fps = 0 logger.record("time/episodes", self._episode_num, exclude="tensorboard") if len(self.ep_info_buffer) > 0 and len(self.ep_info_buffer[0]) > 0: logger.record("rollout/ep_rew_mean", safe_mean([ep_info["r"] for ep_info in self.ep_info_buffer])) logger.record("rollout/ep_len_mean", safe_mean([ep_info["l"] for ep_info in self.ep_info_buffer])) logger.record("time/fps", fps) logger.record("time/time_elapsed", int(time.time() - self.start_time), exclude="tensorboard") logger.record("time/total timesteps", self.num_timesteps, exclude="tensorboard") if self.use_sde: logger.record("train/std", (self.actor.get_std()).mean().item()) if len(self.ep_success_buffer) > 0: logger.record("rollout/success rate", safe_mean(self.ep_success_buffer)) # Pass the number of timesteps for tensorboard logger.dump(step=self.num_timesteps) def _on_step(self) -> None: """ Method called after each step in the environment. It is meant to trigger DQN target network update but can be used for other purposes """ pass def collect_rollouts( self, env: VecEnv, callback: BaseCallback, n_episodes: int = 1, n_steps: int = -1, action_noise: Optional[ActionNoise] = None, learning_starts: int = 0, replay_buffer: Optional[ReservoirBuffer] = None, log_interval: Optional[int] = None, ) -> RolloutReturn: """ Collect experiences and store them into a ReplayBuffer. :param env: The training environment :param callback: Callback that will be called at each step (and at the beginning and end of the rollout) :param n_episodes: Number of episodes to use to collect rollout data You can also specify a ``n_steps`` instead :param n_steps: Number of steps to use to collect rollout data You can also specify a ``n_episodes`` instead. :param action_noise: Action noise that will be used for exploration Required for deterministic policy (e.g. TD3). This can also be used in addition to the stochastic policy for SAC. :param learning_starts: Number of steps before learning for the warm-up phase. :param replay_buffer: :param log_interval: Log data every ``log_interval`` episodes :return: """ episode_rewards, total_timesteps = [], [] total_steps, total_episodes = 0, 0 assert isinstance(env, VecEnv), "You must pass a VecEnv" assert env.num_envs == 1, "OffPolicyAlgorithm only support single environment" if self.use_sde: self.actor.reset_noise() callback.on_rollout_start() continue_training = True while total_steps < n_steps or total_episodes < n_episodes: done = False episode_reward, episode_timesteps = 0.0, 0 while not done: if self.use_sde and self.sde_sample_freq > 0 and total_steps % self.sde_sample_freq == 0: # Sample a new noise matrix self.actor.reset_noise() # Select action randomly or according to policy action, buffer_action = self._sample_action(learning_starts, action_noise) # Rescale and perform action new_obs, reward, done, infos = env.step(action) self.num_timesteps += 1 episode_timesteps += 1 total_steps += 1 # Give access to local variables callback.update_locals(locals()) # Only stop training if return value is False, not when it is None. if callback.on_step() is False: return RolloutReturn(0.0, total_steps, total_episodes, continue_training=False) episode_reward += reward # Retrieve reward and episode length if using Monitor wrapper self._update_info_buffer(infos, done) # Store data in replay buffer if replay_buffer is not None: # Store only the unnormalized version if self._vec_normalize_env is not None: new_obs_ = self._vec_normalize_env.get_original_obs() reward_ = self._vec_normalize_env.get_original_reward() else: # Avoid changing the original ones self._last_original_obs, new_obs_, reward_ = self._last_obs, new_obs, reward replay_buffer.add(self._last_original_obs, new_obs_, buffer_action, reward_, done, self.num_timesteps - 1) self.current_experience_buffer.add(self._last_original_obs, new_obs_, buffer_action, reward_, done) self._last_obs = new_obs # Save the unnormalized observation if self._vec_normalize_env is not None: self._last_original_obs = new_obs_ self._update_current_progress_remaining(self.num_timesteps, self._total_timesteps) # For DQN, check if the target network should be updated # and update the exploration schedule # For SAC/TD3, the update is done as the same time as the gradient update # see https://github.com/hill-a/stable-baselines/issues/900 self._on_step() if 0 < n_steps <= total_steps: break if done: total_episodes += 1 self._episode_num += 1 episode_rewards.append(episode_reward) total_timesteps.append(episode_timesteps) if action_noise is not None: action_noise.reset() # Log training infos if log_interval is not None and self._episode_num % log_interval == 0: self._dump_logs() mean_reward = np.mean(episode_rewards) if total_episodes > 0 else 0.0 callback.on_rollout_end() return RolloutReturn(mean_reward, total_steps, total_episodes, continue_training) def update_env(self, env, support_multi_env: bool = False, eval_env: Optional[GymEnv] = None, monitor_wrapper: bool = True, is_reservoir_replay: bool = True, **kwargs): """ Replace current env with new env. :param env: Gym environment (activated, not a string). :param support_multi_env: Whether the algorithm supports training with multiple environments (as in A2C) :param eval_env: Environment to use for evaluation. :param monitor_wrapper: When creating an environment, whether to wrap it or not in a Monitor wrapper. :param is_reservoir_replay: Whether experience replay is normal or reservoir :param kwargs: Does nothing, just so more arguments can pass without method failing :return: """ if self.replay_buffer is not None: if is_reservoir_replay: self.replay_buffer.is_reservoir = True else: self.replay_buffer.is_reservoir = False if env is not None: if eval_env is not None: self.eval_env = eval_env if monitor_wrapper: self.eval_env = Monitor(self.eval_env, filename=None) if monitor_wrapper: env = Monitor(env, filename=None) env = self._wrap_env(env, self.verbose) self.observation_space = env.observation_space self.action_space = env.action_space self.n_envs = env.num_envs self.env = env 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." ) def add_memories_from_another_replay_mem(self, another_replay_mem: ReplayBuffer): for i in range(another_replay_mem.buffer_size): self.replay_buffer.add( obs=another_replay_mem.observations[i], next_obs=another_replay_mem.next_observations[i], action=another_replay_mem.actions[i], reward=another_replay_mem.rewards[i], done=another_replay_mem.dones[i], experience_index=0, # doesn't matter )