tff.program.TensorBoardReleaseManager

A tff.program.ReleaseManager that releases values to TensorBoard.

Inherits From: ReleaseManager

tff.program.TensorBoardReleaseManager(
    summary_dir: Union[str, os.PathLike[str]]
)

A tff.program.TensorBoardReleaseManager is a utility for releasing values from a federated program to TensorBoard and is used to release values from platform storage to customer storage in a federated program.

Values are released as summary data using tf.summary. When the value is released, if the value is a value reference or a structure containing value references, each value reference is materialized. The value is then flattened and released as summary data. The structure of the value is used as the name of the summary data. Scalar values are released using tf.summary.scalar and non-scalar values are released using tf.summary.histogram.

Warning: The summary data can only contain booleans, integers, unsigned integers, and floats, releasing any other values will be silently ignored.

See https://www.tensorflow.org/api_docs/python/tf/summary for more information about summary data and how to visualize summary data using TensorBoard.

Args
summary_dir	A path on the file system to save release values. If this path does not exist it will be created.

Raises
ValueError	If summary_dir is an empty string.

Methods

release

View source

release(
    value: tff.program.ReleasableStructure,
    key
) -> value_reference.MaterializedValue

Releases value from a federated program.

Args
value	A tff.program.ReleasableStructure to release.
key	A integer used to reference the released value; key represents a step in a federated program.