def __init__(self, capacity: int = 10000, storage_unit: str = "timesteps", alpha: float = 1.0, **kwargs): """Initializes a PrioritizedReplayBuffer instance. Args: capacity: Max number of timesteps to store in the FIFO buffer. After reaching this number, older samples will be dropped to make space for new ones. storage_unit: Either 'timesteps', 'sequences' or 'episodes'. Specifies how experiences are stored. alpha: How much prioritization is used (0.0=no prioritization, 1.0=full prioritization). **kwargs: Forward compatibility kwargs. """ ReplayBuffer.__init__(self, capacity, storage_unit, **kwargs) assert alpha > 0 self._alpha = alpha # Segment tree must have capacity that is a power of 2 it_capacity = 1 while it_capacity < self.capacity: 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)
def __init__(self, capacity: int = 10000, storage_unit: str = "timesteps"): """Initializes a ReservoirBuffer instance. Args: capacity: Max number of timesteps to store in the FIFO buffer. After reaching this number, older samples will be dropped to make space for new ones. storage_unit: Either 'sequences' or 'timesteps'. Specifies how experiences are stored. """ ReplayBuffer.__init__(self, capacity, storage_unit) self._num_add_calls = 0 self._num_evicted = 0
def __init__(self, capacity: int = 10000, storage_unit: str = "timesteps", num_shards: int = 1, learning_starts: int = 1000, replay_mode: str = "independent", replay_sequence_length: int = 1, replay_burn_in: int = 0, replay_zero_init_states: bool = True, underlying_buffer_config: dict = None, **kwargs): """Initializes a MultiAgentReplayBuffer instance. Args: num_shards: The number of buffer shards that exist in total (including this one). storage_unit: Either 'timesteps', 'sequences' or 'episodes'. Specifies how experiences are stored. If they are stored in episodes, replay_sequence_length is ignored. learning_starts: Number of timesteps after which a call to `sample()` will yield samples (before that, `sample()` will return None). capacity: The capacity of the buffer, measured in `storage_unit`. replay_mode: One of "independent" or "lockstep". Determines, whether batches are sampled independently or to an equal amount. replay_sequence_length: The sequence length (T) of a single sample. If > 1, we will sample B x T from this buffer. This only has an effect if storage_unit is 'timesteps'. replay_burn_in: 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. This only has an effect if storage_unit is `sequences`. replay_zero_init_states: 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. underlying_buffer_config: A config that contains all necessary constructor arguments and arguments for methods to call on the underlying buffers. **kwargs: Forward compatibility kwargs. """ shard_capacity = capacity // num_shards ReplayBuffer.__init__(self, capacity, storage_unit) # If the user provides an underlying buffer config, we use to # instantiate and interact with underlying buffers self.underlying_buffer_config = underlying_buffer_config if self.underlying_buffer_config is not None: self.underlying_buffer_call_args = self.underlying_buffer_config else: self.underlying_buffer_call_args = {} self.replay_starts = learning_starts // num_shards 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 and self._storage_unit is not StorageUnit.SEQUENCES): logger.warning( "MultiAgentReplayBuffer configured with " "`replay_sequence_length={}`, but `storage_unit={}`. " "replay_sequence_length will be ignored and set to 1.".format( replay_sequence_length, storage_unit)) self.replay_sequence_length = 1 if replay_sequence_length == 1 and self._storage_unit is StorageUnit.SEQUENCES: logger.warning( "MultiAgentReplayBuffer configured with " "`replay_sequence_length={}`, but `storage_unit={}`. " "This will result in sequences equal to timesteps.".format( replay_sequence_length, storage_unit)) if replay_mode in ["lockstep", ReplayMode.LOCKSTEP]: self.replay_mode = ReplayMode.LOCKSTEP if self._storage_unit in [ StorageUnit.EPISODES, StorageUnit.SEQUENCES ]: raise ValueError("MultiAgentReplayBuffer does not support " "lockstep mode with storage unit `episodes`" "or `sequences`.") elif replay_mode in ["independent", ReplayMode.INDEPENDENT]: self.replay_mode = ReplayMode.INDEPENDENT else: raise ValueError("Unsupported replay mode: {}".format(replay_mode)) if self.underlying_buffer_config: ctor_args = { **{ "capacity": shard_capacity, "storage_unit": StorageUnit.FRAGMENTS }, **self.underlying_buffer_config, } def new_buffer(): return from_config(self.underlying_buffer_config["type"], ctor_args) else: # Default case def new_buffer(): self.underlying_buffer_call_args = {} return ReplayBuffer( self.capacity, storage_unit=StorageUnit.FRAGMENTS, ) self.replay_buffers = collections.defaultdict(new_buffer) # Metrics. self.add_batch_timer = TimerStat() self.replay_timer = TimerStat() self._num_added = 0
def __init__(self, capacity: int = 10000, storage_unit: str = "timesteps", num_shards: int = 1, learning_starts: int = 1000, 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, **kwargs): """Initializes a MultiAgentReplayBuffer instance. Args: num_shards: The number of buffer shards that exist in total (including this one). learning_starts: Number of timesteps after which a call to `replay()` will yield samples (before that, `replay()` will return None). capacity: The capacity of the buffer. Note that when `replay_sequence_length` > 1, this is the number of sequences (not single timesteps) stored. replay_batch_size: 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: Alpha parameter for a prioritized replay buffer. Use 0.0 for no prioritization. prioritized_replay_beta: Beta parameter for a prioritized replay buffer. prioritized_replay_eps: Epsilon parameter for a prioritized replay buffer. replay_mode: One of "independent" or "lockstep". Determined, whether in the multiagent case, sampling is done across all agents/policies equally. replay_sequence_length: The sequence length (T) of a single sample. If > 1, we will sample B x T from this buffer. replay_burn_in: 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: 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. **kwargs: Forward compatibility kwargs. """ shard_capacity = capacity // num_shards ReplayBuffer.__init__(self, shard_capacity, storage_unit) self.replay_starts = learning_starts // 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 new_buffer(): if prioritized_replay_alpha == 0.0: return ReplayBuffer(self.capacity) else: return PrioritizedReplayBuffer(self.capacity, 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