tff.framework.UnplacedExecutorFactory

View source on GitHub

ExecutorFactory to construct executors which cannot understand placement.

Inherits From: ExecutorFactory

tff.framework.UnplacedExecutorFactory(
    *,
    support_sequence_ops: bool = False,
    can_resolve_references: bool = True,
    server_device: Optional[tf.config.LogicalDevice] = None,
    client_devices: Optional[Sequence[tf.config.LogicalDevice]] = (),
    leaf_executor_fn=tff.framework.EagerTFExecutor
)

This factory constructs executors which represent "local execution": work that happens at the clients, at the server, or without placements. As such, this executor manages the placement of work on local executors.

Methods

clean_up_executor

View source

clean_up_executor(
    cardinalities: tff.framework.CardinalitiesType
)

Releases any resources associated to the given cardinalities.

Note that calling this method may invalidate the state of any executors which have previously been returned by the factory with the cardinalities argument ; create_executor should be called again if a new executor which is safe to use is desired.

Args
cardinalities	The cardinalities of the executor whose state we wish to clear.

create_executor

View source

create_executor(
    *,
    cardinalities: Optional[tff.framework.CardinalitiesType] = None,
    placement: Optional[tff.framework.PlacementLiteral] = None
) -> tff.framework.Executor

Abstract method to construct instance of executor_base.Executor.

create_executor must accept a dict mapping placements.PlacementLiterals to ints, and return an executor_base.Executor.

Args
cardinalities	a dict mapping instances of placements.PlacementLiteral to ints, specifying the population size at each placement.

Returns
Instance of executor_base.Executor.