tf_agents.replay_buffers.reverb_utils.ReverbTrajectorySequenceObserver

Reverb trajectory sequence observer.

Inherits From: ReverbAddTrajectoryObserver

tf_agents.replay_buffers.reverb_utils.ReverbTrajectorySequenceObserver(
    py_client: tf_agents.typing.types.ReverbClient,
    table_name: Union[Text, Sequence[Text]],
    sequence_length: int,
    stride_length: int = 1,
    priority: Union[float, int] = 1,
    pad_end_of_episodes: bool = False,
    tile_end_of_episodes: bool = False
)

This is equivalent to ReverbAddTrajectoryObserver but sequences are not cut when a boundary trajectory is seen. This allows for sequences to be sampled with boundaries anywhere in the sequence rather than just at the end.

Consider using this observer when you want to create training experience that can encompass any subsequence of the observed trajectories.

Note: Counting of steps in drivers does not include boundary steps. To guarantee only 1 item is pushed to the replay when collecting n steps with a sequence_length of n make sure to set the stride_length

Args
py_client	Python client for the reverb replay server.
table_name	The table name(s) where samples will be written to.
sequence_length	The sequence_length used to write to the given table.
stride_length	The integer stride for the sliding window for overlapping sequences. The default value of 1 creates an item for every window. Using L = sequence_length this means items are created for times {0, 1, .., L-1}, {1, 2, .., L}, .... In contrast, stride_length = L will create an item only for disjoint windows {0, 1, ..., L-1}, {L, ..., 2 * L - 1}, ....
priority	Initial priority for new samples in the RB.
pad_end_of_episodes	At the end of an episode, the cache is dropped by default. When pad_end_of_episodes = True, the cache gets padded with boundary steps (last->first) with 0 values everywhere and padded items of sequence_length are written to Reverb.
tile_end_of_episodes	If pad_end_of_episodes is True then, the last padded item starts with a boundary step from the episode. When this option is True the following items will be generated: F, M, L, P M, L, P, P L, P, P, P If False, only a single one will be generated: F, M, L, P For training recurrent models on environments where required information is only available at the start of the episode it is useful to set tile_end_of_episodes=False and the sequence_length to be the length of the longest episode.

Raises
ValueError	If tile_end_of_episodes is set without pad_end_of_episodes.

Methods

close

View source

close() -> None

Closes the writer of the observer.

Note: Using the observer after it is closed (and not reopened) is not supported.

flush

View source

flush()

Ensures that items are pushed to the service.

Note: The items are not always immediately pushed. This method is often needed when rate_limiter_timeout_ms is set for the replay buffer. By calling this method before the learner.run(), we ensure that there is enough data to be consumed.

open

View source

open() -> None

Open the writer of the observer.

reset

View source

reset(
    write_cached_steps: bool = True
) -> None

Resets the state of the observer.

Note: Reset should be called only after all collection has finished in a standard workflow. No need to manually call reset between episodes.

Args
write_cached_steps	boolean flag indicating whether we want to write the cached trajectory. When this argument is True, the function attempts to write the cached data before resetting (optionally with padding). Otherwise, the cached data gets dropped.

__call__

View source

__call__(
    trajectory: tf_agents.trajectories.Trajectory
) -> None

Writes the trajectory into the underlying replay buffer.

Allows trajectory to be a flattened trajectory. No batch dimension allowed.

Args
trajectory	The trajectory to be written which could be (possibly nested) trajectory object or a flattened version of a trajectory. It assumes there is no batch dimension.