tf_agents.replay_buffers.py_hashed_replay_buffer.PyHashedReplayBuffer

A Python-based replay buffer with optimized underlying storage.

Inherits From: PyUniformReplayBuffer, ReplayBuffer

This replay buffer deduplicates data in the stored trajectories along the last axis of the observation, which is useful, e.g., if you are performing something like frame stacking. For example, if each observation is 4 stacked 84x84 grayscale images forming a shape [84, 84, 4], then the replay buffer will separate out each of the images and depuplicate across each trajectory in case an image is repeated.

data_spec An ArraySpec or a list/tuple/nest of ArraySpecs describing a single item that can be stored in this buffer.
capacity The maximum number of items that can be stored in the buffer.

capacity Returns the capacity of the replay buffer.
data_spec Returns the spec for items in the replay buffer.
size

stateful_dataset Returns whether the dataset of the replay buffer has stateful ops.

Methods

add_batch

View source

Adds a batch of items to the replay buffer.

Args
items An item or list/tuple/nest of items to be added to the replay buffer. items must match the data_spec of this class, with a batch_size dimension added to the beginning of each tensor/array.

Returns
Adds items to the replay buffer.

as_dataset

View source

Creates and returns a dataset that returns entries from the buffer.

A single entry from the dataset is the result of the following pipeline:

  • Sample sequences from the underlying data store
  • (optionally) Process them with sequence_preprocess_fn,
  • (optionally) Split them into subsequences of length num_steps
  • (optionally) Batch them into batches of size sample_batch_size.

In practice, this pipeline is executed in parallel as much as possible if num_parallel_calls != 1.

Some additional notes:

If num_steps is None, different replay buffers will behave differently. For example, TFUniformReplayBuffer will return single time steps without a time dimension. In contrast, e.g., EpisodicReplayBuffer will return full sequences (since each sequence may be an episode of unknown length, the outermost shape dimension will be None).

If sample_batch_size is None, no batching is performed; and there is no outer batch dimension in the returned Dataset entries. This setting is useful with variable episode lengths using e.g. EpisodicReplayBuffer, because it allows the user to get full episodes back, and use tf.data to build padded or truncated batches themselves.

If single_deterministic_pass == True, the replay buffer will make every attempt to ensure every time step is visited once and exactly once in a deterministic manner (though true determinism depends on the underlying data store). Additional work may be done to ensure minibatches do not have multiple rows from the same episode. In some cases, this may mean arguments like num_parallel_calls are ignored.

Args
sample_batch_size (Optional.) An optional batch_size to specify the number of items to return. If None (default), a single item is returned which matches the data_spec of this class (without a batch dimension). Otherwise, a batch of sample_batch_size items is returned, where each tensor in items will have its first dimension equal to sample_batch_size and the rest of the dimensions match the corresponding data_spec.
num_steps (Optional.) Optional way to specify that sub-episodes are desired. If None (default), a batch of single items is returned. Otherwise, a batch of sub-episodes is returned, where a sub-episode is a sequence of consecutive items in the replay_buffer. The returned tensors will have first dimension equal to sample_batch_size (if sample_batch_size is not None), subsequent dimension equal to num_steps, and remaining dimensions which match the data_spec of this class.
num_parallel_calls (Optional.) A tf.int32 scalar tf.Tensor, representing the number elements to process in parallel. If not specified, elements will be processed sequentially.
sequence_preprocess_fn (Optional) fn for preprocessing the collected data before it is split into subsequences of length num_steps. Defined in TFAgent.preprocess_sequence. Defaults to pass through.
single_deterministic_pass Python boolean. If True, the dataset will return a single deterministic pass through its underlying data. NOTE: If the buffer is modified while a Dataset iterator is iterating over this data, the iterator may miss any new data or otherwise have subtly invalid data.

Returns
A dataset of type tf.data.Dataset, elements of which are 2-tuples of:

  • An item or sequence of items or batch thereof
  • Auxiliary info for the items (i.e. ids, probs).

Raises
NotImplementedError If a non-default argument value is not supported.
ValueError If the data spec contains lists that must be converted to tuples.

clear

View source

Resets the contents of replay buffer.

Returns
Clears the replay buffer contents.

gather_all

View source

Returns all the items in buffer. (deprecated)

Returns
Returns all the items currently in the buffer. Returns a tensor of shape [B, T, ...] where B = batch size, T = timesteps, and the remaining shape is the shape spec of the items in the buffer.

get_next

View source

Returns an item or batch of items from the buffer. (deprecated)

Args
sample_batch_size (Optional.) An optional batch_size to specify the number of items to return. If None (default), a single item is returned which matches the data_spec of this class (without a batch dimension). Otherwise, a batch of sample_batch_size items is returned, where each tensor in items will have its first dimension equal to sample_batch_size and the rest of the dimensions match the corresponding data_spec. See examples below.
num_steps (Optional.) Optional way to specify that sub-episodes are desired. If None (default), in non-episodic replay buffers, a batch of single items is returned. In episodic buffers, full episodes are returned (note that sample_batch_size must be None in that case). Otherwise, a batch of sub-episodes is returned, where a sub-episode is a sequence of consecutive items in the replay_buffer. The returned tensors will have first dimension equal to sample_batch_size (if sample_batch_size is not None), subsequent dimension equal to num_steps, if time_stacked=True and remaining dimensions which match the data_spec of this class. See examples below.
time_stacked (Optional.) Boolean, when true and num_steps > 1 it returns the items stacked on the time dimension. See examples below for details. Examples of tensor shapes returned: (B = batch size, T = timestep, D = data spec) get_next(sample_batch_size=None, num_steps=None, time_stacked=True) return shape (non-episodic): [D] return shape (episodic): T, D get_next(sample_batch_size=B, num_steps=None, time_stacked=True) return shape (non-episodic): [B, D] return shape (episodic): Not supported get_next(sample_batch_size=B, num_steps=T, time_stacked=True) return shape: [B, T, D] get_next(sample_batch_size=None, num_steps=T, time_stacked=False) return shape: ([D], [D], ..) T tensors in the tuple get_next(sample_batch_size=B, num_steps=T, time_stacked=False) return shape: ([B, D], [B, D], ..) T tensors in the tuple

Returns
A 2-tuple containing:

  • An item or sequence of (optionally batched and stacked) items.
  • Auxiliary info for the items (i.e. ids, probs).

num_frames

View source

Returns the number of frames in the replay buffer.