tff.simulation.run_training_process

Runs a federated training_process.

tff.simulation.run_training_process(
    training_process: tff.templates.IterativeProcess,
    training_selection_fn: Callable[[int], Any],
    total_rounds: int,
    evaluation_fn: Optional[Callable[[Any, Any], MetricsType]] = None,
    evaluation_selection_fn: Optional[Callable[[int], Any]] = None,
    rounds_per_evaluation: int = 1,
    program_state_manager: Optional[tff.program.ProgramStateManager] = None,
    rounds_per_saving_program_state: int = 1,
    metrics_managers: Optional[Iterable[release_manager_lib.ReleaseManager]] = None
)

The following tff.Computation types signaures are required:

• training_process.initialize: ( -> state).
• training_process.next: <state, client_data> -> <state, metrics>
• evaulation_fn: <state, client_data> -> metrics

This function performs up to total_rounds updates to the state of the given training_process. At each training round, this update occurs by invoking training_process.next with state and the output of training_selection_fn. Depending on rounds_per_evaluation and rounds_per_saving_program_state, each training round may be followed by an invocation of the evaluation_fn and by saving the program state.

Note: Round 0 represents saving an initial program model state and computing initial evaluation metrics and round 1 through total_rounds + 1 represent the training rounds.

In addition to the training metrics and evaluation metrics, this function adds the following performance metrics (key and descriptions):

• tff.simulation.ROUND_NUMBER_KEY: The round number.
• tff.simulation.TRAINING_TIME_KEY: The amount of time (in seconds) it takes to run one round of training.
• tff.simulation.EVALUATION_TIME_KEY: The amount of time (in seconds) it takes to run one round of evaluation.

Args
training_process	A tff.templates.IterativeProcess to run for training.
training_selection_fn	A Callable accepting an integer round number, and returning a list of client data to use for trainig in that round.
total_rounds	The number of training rounds to run.
evaluation_fn	An optional callable accepting the state of training_process and the output of evaluation_selection_fn, and returning a mutable mapping with string-valued keys.
evaluation_selection_fn	A optional Callable accepting an integer round number, and returning a list of client data to use for evaluation in that round.
rounds_per_evaluation	The number of training rounds to run between each invocation of evaluation_fn.
program_state_manager	An optional tff.program.ProgramStateManager to use to save program state for fault tolerance.
rounds_per_saving_program_state	The number of training rounds to run between saving program state.
metrics_managers	An optional list of tff.program.ReleaseManagerss to use to save metrics.

Returns
The state of the training process after training.