import logging import time from abc import ABC, abstractmethod from collections import defaultdict, deque from collections.abc import Callable from dataclasses import asdict import numpy as np import tqdm from tianshou.data import ( AsyncCollector, CollectStats, EpochStats, InfoStats, ReplayBuffer, SequenceSummaryStats, ) from tianshou.data.collector import BaseCollector, CollectStatsBase from tianshou.policy import BasePolicy from tianshou.policy.base import TrainingStats from tianshou.trainer.utils import gather_info, test_episode from tianshou.utils import ( BaseLogger, DummyTqdm, LazyLogger, MovAvg, tqdm_config, ) from tianshou.utils.logging import set_numerical_fields_to_precision from tianshou.utils.torch_utils import policy_within_training_step log = logging.getLogger(__name__) class BaseTrainer(ABC): """An iterator base class for trainers. Returns an iterator that yields a 3-tuple (epoch, stats, info) of train results on every epoch. :param policy: an instance of the :class:`~tianshou.policy.BasePolicy` class. :param batch_size: the batch size of sample data, which is going to feed in the policy network. If None, will use the whole buffer in each gradient step. :param train_collector: the collector used for training. :param test_collector: the collector used for testing. If it's None, then no testing will be performed. :param buffer: the replay buffer used for off-policy algorithms or for pre-training. If a policy overrides the ``process_buffer`` method, the replay buffer will be pre-processed before training. :param max_epoch: the maximum number of epochs for training. The training process might be finished before reaching ``max_epoch`` if ``stop_fn`` is set. :param step_per_epoch: the number of transitions collected per epoch. :param repeat_per_collect: the number of repeat time for policy learning, for example, set it to 2 means the policy needs to learn each given batch data twice. Only used in on-policy algorithms :param episode_per_test: the number of episodes for one policy evaluation. :param update_per_step: only used in off-policy algorithms. How many gradient steps to perform per step in the environment (i.e., per sample added to the buffer). :param step_per_collect: the number of transitions the collector would collect before the network update, i.e., trainer will collect "step_per_collect" transitions and do some policy network update repeatedly in each epoch. :param episode_per_collect: the number of episodes the collector would collect before the network update, i.e., trainer will collect "episode_per_collect" episodes and do some policy network update repeatedly in each epoch. :param train_fn: a hook called at the beginning of training in each epoch. It can be used to perform custom additional operations, with the signature ``f(num_epoch: int, step_idx: int) -> None``. :param test_fn: a hook called at the beginning of testing in each epoch. It can be used to perform custom additional operations, with the signature ``f(num_epoch: int, step_idx: int) -> None``. :param save_best_fn: a hook called when the undiscounted average mean reward in evaluation phase gets better, with the signature ``f(policy: BasePolicy) -> None``. :param save_checkpoint_fn: a function to save training process and return the saved checkpoint path, with the signature ``f(epoch: int, env_step: int, gradient_step: int) -> str``; you can save whatever you want. :param resume_from_log: resume env_step/gradient_step and other metadata from existing tensorboard log. :param stop_fn: a function with signature ``f(mean_rewards: float) -> bool``, receives the average undiscounted returns of the testing result, returns a boolean which indicates whether reaching the goal. :param reward_metric: a function with signature ``f(rewards: np.ndarray with shape (num_episode, agent_num)) -> np.ndarray with shape (num_episode,)``, used in multi-agent RL. We need to return a single scalar for each episode's result to monitor training in the multi-agent RL setting. This function specifies what is the desired metric, e.g., the reward of agent 1 or the average reward over all agents. :param logger: A logger that logs statistics during training/testing/updating. To not log anything, keep the default logger. :param verbose: whether to print status information to stdout. If set to False, status information will still be logged (provided that logging is enabled via the `logging` module). :param show_progress: whether to display a progress bar when training. :param test_in_train: whether to test in the training phase. """ __doc__: str @staticmethod def gen_doc(learning_type: str) -> str: """Document string for subclass trainer.""" step_means = f'The "step" in {learning_type} trainer means ' if learning_type != "offline": step_means += "an environment step (a.k.a. transition)." else: # offline step_means += "a gradient step." trainer_name = learning_type.capitalize() + "Trainer" return f"""An iterator class for {learning_type} trainer procedure. Returns an iterator that yields a 3-tuple (epoch, stats, info) of train results on every epoch. {step_means} Example usage: :: trainer = {trainer_name}(...) for epoch, epoch_stat, info in trainer: print("Epoch:", epoch) print(epoch_stat) print(info) do_something_with_policy() query_something_about_policy() make_a_plot_with(epoch_stat) display(info) - epoch int: the epoch number - epoch_stat dict: a large collection of metrics of the current epoch - info dict: result returned from :func:`~tianshou.trainer.gather_info` You can even iterate on several trainers at the same time: :: trainer1 = {trainer_name}(...) trainer2 = {trainer_name}(...) for result1, result2, ... in zip(trainer1, trainer2, ...): compare_results(result1, result2, ...) """ def __init__( self, policy: BasePolicy, max_epoch: int, batch_size: int | None, train_collector: BaseCollector | None = None, test_collector: BaseCollector | None = None, buffer: ReplayBuffer | None = None, step_per_epoch: int | None = None, repeat_per_collect: int | None = None, episode_per_test: int | None = None, update_per_step: float = 1.0, step_per_collect: int | None = None, episode_per_collect: int | None = None, train_fn: Callable[[int, int], None] | None = None, test_fn: Callable[[int, int | None], None] | None = None, stop_fn: Callable[[float], bool] | None = None, save_best_fn: Callable[[BasePolicy], None] | None = None, save_checkpoint_fn: Callable[[int, int, int], str] | None = None, resume_from_log: bool = False, reward_metric: Callable[[np.ndarray], np.ndarray] | None = None, logger: BaseLogger = LazyLogger(), verbose: bool = True, show_progress: bool = True, test_in_train: bool = True, ): self.policy = policy if buffer is not None: buffer = policy.process_buffer(buffer) self.buffer = buffer self.train_collector = train_collector self.test_collector = test_collector self.logger = logger self.start_time = time.time() self.stat: defaultdict[str, MovAvg] = defaultdict(MovAvg) self.best_reward = 0.0 self.best_reward_std = 0.0 self.start_epoch = 0 # This is only used for logging but creeps into the implementations # of the trainers. I believe it would be better to remove self._gradient_step = 0 self.env_step = 0 self.policy_update_time = 0.0 self.max_epoch = max_epoch self.step_per_epoch = step_per_epoch # either on of these two self.step_per_collect = step_per_collect self.episode_per_collect = episode_per_collect self.update_per_step = update_per_step self.repeat_per_collect = repeat_per_collect self.episode_per_test = episode_per_test self.batch_size = batch_size self.train_fn = train_fn self.test_fn = test_fn self.stop_fn = stop_fn self.save_best_fn = save_best_fn self.save_checkpoint_fn = save_checkpoint_fn self.reward_metric = reward_metric self.verbose = verbose self.show_progress = show_progress self.test_in_train = test_in_train self.resume_from_log = resume_from_log self.is_run = False self.last_rew, self.last_len = 0.0, 0.0 self.epoch = self.start_epoch self.best_epoch = self.start_epoch self.stop_fn_flag = False self.iter_num = 0 def _reset_collectors(self, reset_buffer: bool = False) -> None: if self.train_collector is not None: self.train_collector.reset(reset_buffer=reset_buffer) if self.test_collector is not None: self.test_collector.reset(reset_buffer=reset_buffer) def reset(self, reset_collectors: bool = True, reset_buffer: bool = False) -> None: """Initialize or reset the instance to yield a new iterator from zero.""" self.is_run = False self.env_step = 0 if self.resume_from_log: ( self.start_epoch, self.env_step, self._gradient_step, ) = self.logger.restore_data() self.last_rew, self.last_len = 0.0, 0.0 self.start_time = time.time() if reset_collectors: self._reset_collectors(reset_buffer=reset_buffer) if self.train_collector is not None and ( self.train_collector.policy != self.policy or self.test_collector is None ): self.test_in_train = False if self.test_collector is not None: assert self.episode_per_test is not None assert not isinstance(self.test_collector, AsyncCollector) # Issue 700 test_result = test_episode( self.test_collector, self.test_fn, self.start_epoch, self.episode_per_test, self.logger, self.env_step, self.reward_metric, ) assert test_result.returns_stat is not None # for mypy self.best_epoch = self.start_epoch self.best_reward, self.best_reward_std = ( test_result.returns_stat.mean, test_result.returns_stat.std, ) if self.save_best_fn: self.save_best_fn(self.policy) self.epoch = self.start_epoch self.stop_fn_flag = False self.iter_num = 0 def __iter__(self): # type: ignore self.reset(reset_collectors=True, reset_buffer=False) return self def __next__(self) -> EpochStats: """Perform one epoch (both train and eval).""" self.epoch += 1 self.iter_num += 1 if self.iter_num > 1: # iterator exhaustion check if self.epoch > self.max_epoch: raise StopIteration # exit flag 1, when stop_fn succeeds in train_step or test_step if self.stop_fn_flag: raise StopIteration progress = tqdm.tqdm if self.show_progress else DummyTqdm # perform n step_per_epoch with progress(total=self.step_per_epoch, desc=f"Epoch #{self.epoch}", **tqdm_config) as t: train_stat: CollectStatsBase while t.n < t.total and not self.stop_fn_flag: train_stat, update_stat, self.stop_fn_flag = self.training_step() if isinstance(train_stat, CollectStats): pbar_data_dict = { "env_step": str(self.env_step), "rew": f"{self.last_rew:.2f}", "len": str(int(self.last_len)), "n/ep": str(train_stat.n_collected_episodes), "n/st": str(train_stat.n_collected_steps), } t.update(train_stat.n_collected_steps) else: pbar_data_dict = {} t.update() pbar_data_dict = set_numerical_fields_to_precision(pbar_data_dict) pbar_data_dict["gradient_step"] = str(self._gradient_step) t.set_postfix(**pbar_data_dict) if self.stop_fn_flag: break if t.n <= t.total and not self.stop_fn_flag: t.update() # for offline RL if self.train_collector is None: assert self.buffer is not None batch_size = self.batch_size or len(self.buffer) self.env_step = self._gradient_step * batch_size test_stat = None if not self.stop_fn_flag: self.logger.save_data( self.epoch, self.env_step, self._gradient_step, self.save_checkpoint_fn, ) # test if self.test_collector is not None: test_stat, self.stop_fn_flag = self.test_step() info_stat = gather_info( start_time=self.start_time, policy_update_time=self.policy_update_time, gradient_step=self._gradient_step, best_reward=self.best_reward, best_reward_std=self.best_reward_std, train_collector=self.train_collector, test_collector=self.test_collector, ) self.logger.log_info_data(asdict(info_stat), self.epoch) # in case trainer is used with run(), epoch_stat will not be returned return EpochStats( epoch=self.epoch, train_collect_stat=train_stat, test_collect_stat=test_stat, training_stat=update_stat, info_stat=info_stat, ) def test_step(self) -> tuple[CollectStats, bool]: """Perform one testing step.""" assert self.episode_per_test is not None assert self.test_collector is not None stop_fn_flag = False test_stat = test_episode( self.test_collector, self.test_fn, self.epoch, self.episode_per_test, self.logger, self.env_step, self.reward_metric, ) assert test_stat.returns_stat is not None # for mypy rew, rew_std = test_stat.returns_stat.mean, test_stat.returns_stat.std if self.best_epoch < 0 or self.best_reward < rew: self.best_epoch = self.epoch self.best_reward = float(rew) self.best_reward_std = rew_std if self.save_best_fn: self.save_best_fn(self.policy) log_msg = ( f"Epoch #{self.epoch}: test_reward: {rew:.6f} ± {rew_std:.6f}," f" best_reward: {self.best_reward:.6f} ± " f"{self.best_reward_std:.6f} in #{self.best_epoch}" ) log.info(log_msg) if self.verbose: print(log_msg, flush=True) if self.stop_fn and self.stop_fn(self.best_reward): stop_fn_flag = True return test_stat, stop_fn_flag def training_step(self) -> tuple[CollectStatsBase, TrainingStats | None, bool]: """Perform one training iteration. A training iteration includes collecting data (for online RL), determining whether to stop training, and performing a policy update if the training iteration should continue. :return: the iteration's collect stats, training stats, and a flag indicating whether to stop training. If training is to be stopped, no gradient steps will be performed and the training stats will be `None`. """ with policy_within_training_step(self.policy): should_stop_training = False collect_stats: CollectStatsBase | CollectStats if self.train_collector is not None: collect_stats = self._collect_training_data() should_stop_training = self._update_best_reward_and_return_should_stop_training( collect_stats, ) else: assert self.buffer is not None, "Either train_collector or buffer must be provided." collect_stats = CollectStatsBase( n_collected_episodes=len(self.buffer), ) if not should_stop_training: training_stats = self.policy_update_fn(collect_stats) else: training_stats = None return collect_stats, training_stats, should_stop_training def _collect_training_data(self) -> CollectStats: """Performs training data collection. :return: the data collection stats """ assert self.episode_per_test is not None assert self.train_collector is not None if self.train_fn: self.train_fn(self.epoch, self.env_step) collect_stats = self.train_collector.collect( n_step=self.step_per_collect, n_episode=self.episode_per_collect, ) self.env_step += collect_stats.n_collected_steps if collect_stats.n_collected_episodes > 0: assert collect_stats.returns_stat is not None # for mypy assert collect_stats.lens_stat is not None # for mypy self.last_rew = collect_stats.returns_stat.mean self.last_len = collect_stats.lens_stat.mean if self.reward_metric: # TODO: move inside collector rew = self.reward_metric(collect_stats.returns) collect_stats.returns = rew collect_stats.returns_stat = SequenceSummaryStats.from_sequence(rew) self.logger.log_train_data(asdict(collect_stats), self.env_step) return collect_stats # TODO (maybe): separate out side effect, simplify name? def _update_best_reward_and_return_should_stop_training( self, collect_stats: CollectStats, ) -> bool: """If `test_in_train` and `stop_fn` are set, will compute the `stop_fn` on the mean return of the training data. Then, if the `stop_fn` is True there, will collect test data also compute the stop_fn of the mean return on it. Finally, if the latter is also True, will return True. **NOTE:** has a side effect of updating the best reward and corresponding std. :param collect_stats: the data collection stats :return: flag indicating whether to stop training """ should_stop_training = False # Because we need to evaluate the policy, we need to temporarily leave the "is_training_step" semantics with policy_within_training_step(self.policy, enabled=False): if ( collect_stats.n_collected_episodes > 0 and self.test_in_train and self.stop_fn and self.stop_fn(collect_stats.returns_stat.mean) # type: ignore ): assert self.test_collector is not None assert self.episode_per_test is not None and self.episode_per_test > 0 test_result = test_episode( self.test_collector, self.test_fn, self.epoch, self.episode_per_test, self.logger, self.env_step, ) assert test_result.returns_stat is not None # for mypy if self.stop_fn(test_result.returns_stat.mean): should_stop_training = True self.best_reward = test_result.returns_stat.mean self.best_reward_std = test_result.returns_stat.std return should_stop_training # TODO: move moving average computation and logging into its own logger # TODO: maybe think about a command line logger instead of always printing data dict def _update_moving_avg_stats_and_log_update_data(self, update_stat: TrainingStats) -> None: """Log losses, update moving average stats, and also modify the smoothed_loss in update_stat.""" cur_losses_dict = update_stat.get_loss_stats_dict() update_stat.smoothed_loss = self._update_moving_avg_stats_and_get_averaged_data( cur_losses_dict, ) self.logger.log_update_data(asdict(update_stat), self._gradient_step) # TODO: seems convoluted, there should be a better way of dealing with the moving average stats def _update_moving_avg_stats_and_get_averaged_data( self, data: dict[str, float], ) -> dict[str, float]: """Add entries to the moving average object in the trainer and retrieve the averaged results. :param data: any entries to be tracked in the moving average object. :return: A dictionary containing the averaged values of the tracked entries. """ smoothed_data = {} for key, loss_item in data.items(): self.stat[key].add(loss_item) smoothed_data[key] = self.stat[key].get() return smoothed_data @abstractmethod def policy_update_fn( self, collect_stats: CollectStatsBase, ) -> TrainingStats: """Policy update function for different trainer implementation. :param collect_stats: provides info about the most recent collection. In the offline case, this will contain stats of the whole dataset """ def run(self, reset_prior_to_run: bool = True) -> InfoStats: """Consume iterator. See itertools - recipes. Use functions that consume iterators at C speed (feed the entire iterator into a zero-length deque). """ if reset_prior_to_run: self.reset() try: self.is_run = True deque(self, maxlen=0) # feed the entire iterator into a zero-length deque info = gather_info( start_time=self.start_time, policy_update_time=self.policy_update_time, gradient_step=self._gradient_step, best_reward=self.best_reward, best_reward_std=self.best_reward_std, train_collector=self.train_collector, test_collector=self.test_collector, ) finally: self.is_run = False return info def _sample_and_update(self, buffer: ReplayBuffer) -> TrainingStats: """Sample a mini-batch, perform one gradient step, and update the _gradient_step counter.""" self._gradient_step += 1 # Note: since sample_size=batch_size, this will perform # exactly one gradient step. This is why we don't need to calculate the # number of gradient steps, like in the on-policy case. update_stat = self.policy.update(sample_size=self.batch_size, buffer=buffer) self._update_moving_avg_stats_and_log_update_data(update_stat) return update_stat class OfflineTrainer(BaseTrainer): """Offline trainer, samples mini-batches from buffer and passes them to update. Uses a buffer directly and usually does not have a collector. """ # for mypy assert isinstance(BaseTrainer.__doc__, str) __doc__ += BaseTrainer.gen_doc("offline") + "\n".join(BaseTrainer.__doc__.split("\n")[1:]) def policy_update_fn( self, collect_stats: CollectStatsBase | None = None, ) -> TrainingStats: """Perform one off-line policy update.""" assert self.buffer update_stat = self._sample_and_update(self.buffer) # logging self.policy_update_time += update_stat.train_time return update_stat class OffpolicyTrainer(BaseTrainer): """Offpolicy trainer, samples mini-batches from buffer and passes them to update. Note that with this trainer, it is expected that the policy's `learn` method does not perform additional mini-batching but just updates params from the received mini-batch. """ # for mypy assert isinstance(BaseTrainer.__doc__, str) __doc__ += BaseTrainer.gen_doc("offpolicy") + "\n".join(BaseTrainer.__doc__.split("\n")[1:]) def policy_update_fn( self, # TODO: this is the only implementation where collect_stats is actually needed. Maybe change interface? collect_stats: CollectStatsBase, ) -> TrainingStats: """Perform `update_per_step * n_collected_steps` gradient steps by sampling mini-batches from the buffer. :param collect_stats: the :class:`~TrainingStats` instance returned by the last gradient step. Some values in it will be replaced by their moving averages. """ assert self.train_collector is not None n_collected_steps = collect_stats.n_collected_steps n_gradient_steps = round(self.update_per_step * n_collected_steps) if n_gradient_steps == 0: raise ValueError( f"n_gradient_steps is 0, n_collected_steps={n_collected_steps}, " f"update_per_step={self.update_per_step}", ) for _ in range(n_gradient_steps): update_stat = self._sample_and_update(self.train_collector.buffer) # logging self.policy_update_time += update_stat.train_time # TODO: only the last update_stat is returned, should be improved return update_stat class OnpolicyTrainer(BaseTrainer): """On-policy trainer, passes the entire buffer to .update and resets it after. Note that it is expected that the learn method of a policy will perform batching when using this trainer. """ # for mypy assert isinstance(BaseTrainer.__doc__, str) __doc__ = BaseTrainer.gen_doc("onpolicy") + "\n".join(BaseTrainer.__doc__.split("\n")[1:]) def policy_update_fn( self, result: CollectStatsBase | None = None, ) -> TrainingStats: """Perform one on-policy update by passing the entire buffer to the policy's update method.""" assert self.train_collector is not None training_stat = self.policy.update( sample_size=0, buffer=self.train_collector.buffer, # Note: sample_size is None, so the whole buffer is used for the update. # The kwargs are in the end passed to the .learn method, which uses # batch_size to iterate through the buffer in mini-batches # Off-policy algos typically don't use the batch_size kwarg at all batch_size=self.batch_size, repeat=self.repeat_per_collect, ) # just for logging, no functional role self.policy_update_time += training_stat.train_time # TODO: remove the gradient step counting in trainers? Doesn't seem like # it's important and it adds complexity self._gradient_step += 1 if self.batch_size is None: self._gradient_step += 1 elif self.batch_size > 0: self._gradient_step += int((len(self.train_collector.buffer) - 0.1) // self.batch_size) # Note: this is the main difference to the off-policy trainer! # The second difference is that batches of data are sampled without replacement # during training, whereas in off-policy or offline training, the batches are # sampled with replacement (and potentially custom prioritization). self.train_collector.reset_buffer(keep_statistics=True) # The step is the number of mini-batches used for the update, so essentially self._update_moving_avg_stats_and_log_update_data(training_stat) return training_stat