import collections import logging import numpy as np import platform import random from typing import List, Dict # Import ray before psutil will make sure we use psutil's bundled version import ray # noqa F401 import psutil # noqa E402 from ray.rllib.execution.segment_tree import SumSegmentTree, MinSegmentTree from ray.rllib.policy.rnn_sequencing import \ timeslice_along_seq_lens_with_overlap from ray.rllib.policy.sample_batch import SampleBatch, MultiAgentBatch, \ DEFAULT_POLICY_ID from ray.rllib.utils.annotations import DeveloperAPI from ray.util.iter import ParallelIteratorWorker from ray.util.debug import log_once from ray.rllib.utils.timer import TimerStat from ray.rllib.utils.window_stat import WindowStat from ray.rllib.utils.typing import SampleBatchType # Constant that represents all policies in lockstep replay mode. _ALL_POLICIES = "__all__" logger = logging.getLogger(__name__) def warn_replay_buffer_size(*, item: SampleBatchType, num_items: int) -> None: """Warn if the configured replay buffer size is too large.""" if log_once("replay_buffer_size"): item_size = item.size_bytes() psutil_mem = psutil.virtual_memory() total_gb = psutil_mem.total / 1e9 mem_size = num_items * item_size / 1e9 msg = ("Estimated max memory usage for replay buffer is {} GB " "({} batches of size {}, {} bytes each), " "available system memory is {} GB".format( mem_size, num_items, item.count, item_size, total_gb)) if mem_size > total_gb: raise ValueError(msg) elif mem_size > 0.2 * total_gb: logger.warning(msg) else: logger.info(msg) @DeveloperAPI class ReplayBuffer: @DeveloperAPI def __init__(self, size: int): """Create Prioritized Replay buffer. Args: size (int): Max number of timesteps to store in the FIFO buffer. """ self._storage = [] self._maxsize = size self._next_idx = 0 self._hit_count = np.zeros(size) self._eviction_started = False self._num_timesteps_added = 0 self._num_timesteps_added_wrap = 0 self._num_timesteps_sampled = 0 self._evicted_hit_stats = WindowStat("evicted_hit", 1000) self._est_size_bytes = 0 def __len__(self) -> int: return len(self._storage) @DeveloperAPI def add(self, item: SampleBatchType, weight: float) -> None: warn_replay_buffer_size( item=item, num_items=self._maxsize / item.count) assert item.count > 0, item self._num_timesteps_added += item.count self._num_timesteps_added_wrap += item.count if self._next_idx >= len(self._storage): self._storage.append(item) self._est_size_bytes += item.size_bytes() else: self._storage[self._next_idx] = item # Wrap around storage as a circular buffer once we hit maxsize. if self._num_timesteps_added_wrap >= self._maxsize: self._eviction_started = True self._num_timesteps_added_wrap = 0 self._next_idx = 0 else: self._next_idx += 1 if self._eviction_started: self._evicted_hit_stats.push(self._hit_count[self._next_idx]) self._hit_count[self._next_idx] = 0 def _encode_sample(self, idxes: List[int]) -> SampleBatchType: out = SampleBatch.concat_samples([self._storage[i] for i in idxes]) out.decompress_if_needed() return out @DeveloperAPI def sample(self, num_items: int) -> SampleBatchType: """Sample a batch of experiences. Args: num_items (int): Number of items to sample from this buffer. Returns: SampleBatchType: concatenated batch of items. """ idxes = [ random.randint(0, len(self._storage) - 1) for _ in range(num_items) ] self._num_sampled += num_items return self._encode_sample(idxes) @DeveloperAPI def stats(self, debug=False) -> dict: data = { "added_count": self._num_timesteps_added, "sampled_count": self._num_timesteps_sampled, "est_size_bytes": self._est_size_bytes, "num_entries": len(self._storage), } if debug: data.update(self._evicted_hit_stats.stats()) return data @DeveloperAPI class PrioritizedReplayBuffer(ReplayBuffer): @DeveloperAPI def __init__(self, size: int, alpha: float): """Create Prioritized Replay buffer. Args: size (int): Max number of items to store in the FIFO buffer. alpha (float): how much prioritization is used (0 - no prioritization, 1 - full prioritization). See also: ReplayBuffer.__init__() """ super(PrioritizedReplayBuffer, self).__init__(size) assert alpha > 0 self._alpha = alpha it_capacity = 1 while it_capacity < size: it_capacity *= 2 self._it_sum = SumSegmentTree(it_capacity) self._it_min = MinSegmentTree(it_capacity) self._max_priority = 1.0 self._prio_change_stats = WindowStat("reprio", 1000) @DeveloperAPI def add(self, item: SampleBatchType, weight: float) -> None: idx = self._next_idx super(PrioritizedReplayBuffer, self).add(item, weight) if weight is None: weight = self._max_priority self._it_sum[idx] = weight**self._alpha self._it_min[idx] = weight**self._alpha def _sample_proportional(self, num_items: int) -> List[int]: res = [] for _ in range(num_items): # TODO(szymon): should we ensure no repeats? mass = random.random() * self._it_sum.sum(0, len(self._storage)) idx = self._it_sum.find_prefixsum_idx(mass) res.append(idx) return res @DeveloperAPI def sample(self, num_items: int, beta: float) -> SampleBatchType: """Sample a batch of experiences and return priority weights, indices. Args: num_items (int): Number of items to sample from this buffer. beta (float): To what degree to use importance weights (0 - no corrections, 1 - full correction). Returns: SampleBatchType: Concatenated batch of items including "weights" and "batch_indexes" fields denoting IS of each sampled transition and original idxes in buffer of sampled experiences. """ assert beta >= 0.0 idxes = self._sample_proportional(num_items) weights = [] batch_indexes = [] p_min = self._it_min.min() / self._it_sum.sum() max_weight = (p_min * len(self._storage))**(-beta) for idx in idxes: p_sample = self._it_sum[idx] / self._it_sum.sum() weight = (p_sample * len(self._storage))**(-beta) count = self._storage[idx].count # If zero-padded, count will not be the actual batch size of the # data. if isinstance(self._storage[idx], SampleBatch) and \ self._storage[idx].zero_padded: actual_size = self._storage[idx].max_seq_len else: actual_size = count weights.extend([weight / max_weight] * actual_size) batch_indexes.extend([idx] * actual_size) self._num_timesteps_sampled += count batch = self._encode_sample(idxes) # Note: prioritization is not supported in lockstep replay mode. if isinstance(batch, SampleBatch): batch["weights"] = np.array(weights) batch["batch_indexes"] = np.array(batch_indexes) return batch @DeveloperAPI def update_priorities(self, idxes: List[int], priorities: List[float]) -> None: """Update priorities of sampled transitions. sets priority of transition at index idxes[i] in buffer to priorities[i]. Parameters ---------- idxes: [int] List of idxes of sampled transitions priorities: [float] List of updated priorities corresponding to transitions at the sampled idxes denoted by variable `idxes`. """ assert len(idxes) == len(priorities) for idx, priority in zip(idxes, priorities): assert priority > 0 assert 0 <= idx < len(self._storage) delta = priority**self._alpha - self._it_sum[idx] self._prio_change_stats.push(delta) self._it_sum[idx] = priority**self._alpha self._it_min[idx] = priority**self._alpha self._max_priority = max(self._max_priority, priority) @DeveloperAPI def stats(self, debug: bool = False) -> Dict: parent = ReplayBuffer.stats(self, debug) if debug: parent.update(self._prio_change_stats.stats()) return parent # Visible for testing. _local_replay_buffer = None class LocalReplayBuffer(ParallelIteratorWorker): """A replay buffer shard storing data for all policies (in multiagent setup). Ray actors are single-threaded, so for scalability, multiple replay actors may be created to increase parallelism.""" def __init__(self, num_shards: int = 1, learning_starts: int = 1000, buffer_size: int = 10000, replay_batch_size: int = 1, prioritized_replay_alpha: float = 0.6, prioritized_replay_beta: float = 0.4, prioritized_replay_eps: float = 1e-6, replay_mode: str = "independent", replay_sequence_length: int = 1, replay_burn_in: int = 0, replay_zero_init_states: bool = True): """Initializes a LocalReplayBuffer instance. Args: num_shards (int): The number of buffer shards that exist in total (including this one). learning_starts (int): Number of timesteps after which a call to `replay()` will yield samples (before that, `replay()` will return None). buffer_size (int): The size of the buffer. Note that when `replay_sequence_length` > 1, this is the number of sequences (not single timesteps) stored. replay_batch_size (int): The batch size to be sampled (in timesteps). Note that if `replay_sequence_length` > 1, `self.replay_batch_size` will be set to the number of sequences sampled (B). prioritized_replay_alpha (float): Alpha parameter for a prioritized replay buffer. prioritized_replay_beta (float): Beta parameter for a prioritized replay buffer. prioritized_replay_eps (float): Epsilon parameter for a prioritized replay buffer. replay_mode (str): One of "independent" or "lockstep". Determined, whether in the multiagent case, sampling is done across all agents/policies equally. replay_sequence_length (int): The sequence length (T) of a single sample. If > 1, we will sample B x T from this buffer. replay_burn_in (int): The burn-in length in case `replay_sequence_length` > 0. This is the number of timesteps each sequence overlaps with the previous one to generate a better internal state (=state after the burn-in), instead of starting from 0.0 each RNN rollout. replay_zero_init_states (bool): Whether the initial states in the buffer (if replay_sequence_length > 0) are alwayas 0.0 or should be updated with the previous train_batch state outputs. """ self.replay_starts = learning_starts // num_shards self.buffer_size = buffer_size // num_shards self.replay_batch_size = replay_batch_size self.prioritized_replay_beta = prioritized_replay_beta self.prioritized_replay_eps = prioritized_replay_eps self.replay_mode = replay_mode self.replay_sequence_length = replay_sequence_length self.replay_burn_in = replay_burn_in self.replay_zero_init_states = replay_zero_init_states if replay_sequence_length > 1: self.replay_batch_size = int( max(1, replay_batch_size // replay_sequence_length)) logger.info( "Since replay_sequence_length={} and replay_batch_size={}, " "we will replay {} sequences at a time.".format( replay_sequence_length, replay_batch_size, self.replay_batch_size)) if replay_mode not in ["lockstep", "independent"]: raise ValueError("Unsupported replay mode: {}".format(replay_mode)) def gen_replay(): while True: yield self.replay() ParallelIteratorWorker.__init__(self, gen_replay, False) def new_buffer(): return PrioritizedReplayBuffer( self.buffer_size, alpha=prioritized_replay_alpha) self.replay_buffers = collections.defaultdict(new_buffer) # Metrics self.add_batch_timer = TimerStat() self.replay_timer = TimerStat() self.update_priorities_timer = TimerStat() self.num_added = 0 # Make externally accessible for testing. global _local_replay_buffer _local_replay_buffer = self # If set, return this instead of the usual data for testing. self._fake_batch = None @staticmethod def get_instance_for_testing(): global _local_replay_buffer return _local_replay_buffer def get_host(self) -> str: return platform.node() def add_batch(self, batch: SampleBatchType) -> None: # Make a copy so the replay buffer doesn't pin plasma memory. batch = batch.copy() # Handle everything as if multiagent if isinstance(batch, SampleBatch): batch = MultiAgentBatch({DEFAULT_POLICY_ID: batch}, batch.count) with self.add_batch_timer: # Lockstep mode: Store under _ALL_POLICIES key (we will always # only sample from all policies at the same time). if self.replay_mode == "lockstep": # Note that prioritization is not supported in this mode. for s in batch.timeslices(self.replay_sequence_length): self.replay_buffers[_ALL_POLICIES].add(s, weight=None) else: for policy_id, sample_batch in batch.policy_batches.items(): if self.replay_sequence_length == 1: timeslices = sample_batch.timeslices(1) else: timeslices = timeslice_along_seq_lens_with_overlap( sample_batch=sample_batch, zero_pad_max_seq_len=self.replay_sequence_length, pre_overlap=self.replay_burn_in, zero_init_states=self.replay_zero_init_states, ) for time_slice in timeslices: # If SampleBatch has prio-replay weights, average # over these to use as a weight for the entire # sequence. if "weights" in time_slice: weight = np.mean(time_slice["weights"]) else: weight = None self.replay_buffers[policy_id].add( time_slice, weight=weight) self.num_added += batch.count def replay(self) -> SampleBatchType: if self._fake_batch: fake_batch = SampleBatch(self._fake_batch) return MultiAgentBatch({ DEFAULT_POLICY_ID: fake_batch }, fake_batch.count) if self.num_added < self.replay_starts: return None with self.replay_timer: # Lockstep mode: Sample from all policies at the same time an # equal amount of steps. if self.replay_mode == "lockstep": return self.replay_buffers[_ALL_POLICIES].sample( self.replay_batch_size, beta=self.prioritized_replay_beta) else: samples = {} for policy_id, replay_buffer in self.replay_buffers.items(): samples[policy_id] = replay_buffer.sample( self.replay_batch_size, beta=self.prioritized_replay_beta) return MultiAgentBatch(samples, self.replay_batch_size) def update_priorities(self, prio_dict: Dict) -> None: with self.update_priorities_timer: for policy_id, (batch_indexes, td_errors) in prio_dict.items(): new_priorities = ( np.abs(td_errors) + self.prioritized_replay_eps) self.replay_buffers[policy_id].update_priorities( batch_indexes, new_priorities) def stats(self, debug: bool = False) -> Dict: stat = { "add_batch_time_ms": round(1000 * self.add_batch_timer.mean, 3), "replay_time_ms": round(1000 * self.replay_timer.mean, 3), "update_priorities_time_ms": round( 1000 * self.update_priorities_timer.mean, 3), } for policy_id, replay_buffer in self.replay_buffers.items(): stat.update({ "policy_{}".format(policy_id): replay_buffer.stats(debug=debug) }) return stat ReplayActor = ray.remote(num_cpus=0)(LocalReplayBuffer)