optimization

optimization.centralized module

federated.optimization.centralized.centralized_pipeline(name: str, output: str, epochs: int, batch_size: int, optimizer: str, model: str, learning_rate: float) → float

Function runs centralized training pipeline.

Args:

name (str): Name of the experiment.

output (str): Where to save config files. Defaults to history.

epochs (int): Number of epochs. Defaults to 15.

batch_size (int): Batch size. Defaults to 32.

optimizer (str): Which optimizer to use. Defaults to sgd.

model (str): Which model to use.

learning_rate (float): Learning rate to be used. Defaults to 1.0.

Returns:

float: Training time after centralized learning.

optimization.federated module

federated.optimization.federated.federated_pipeline(name: str, aggregation_method: str, client_weighting: str, keras_model_fn: str, server_optimizer_fn: str, server_optimizer_lr: float, client_optimizer_fn: str, client_optimizer_lr: float, data_selector: str, output: str, client_epochs: int, batch_size: int, number_of_clients: int, number_of_clients_per_round: int, number_of_rounds: int, iterations: int, v: float, compression: bool, dp: bool, noise_multiplier: float, clipping_value: float, seed: Optional[int] = None) → float

Function runs federated training pipeline on the dataset.

Args:

name (str): Experiment name.

aggregation_method (str): Aggregation method. Defaults to “fedavg”.

client_weighting (str): Client weighting. Either Uniform or Data dependent. Defaults to NUM_EXAMPLES.

keras_model_fn (str): Keras Model.

server_optimizer_fn (str): Server Optimizer. Defaults to sgd.

server_optimizer_lr (float): Learning rate for server optimizer. Defaults to 1.0.

client_optimizer_fn (str): Client optimizer. Defaults to sgd.

client_optimizer_lr (float): Learning rate for client optimizer. Defaults to 0.02.

data_selector (str): Data distribution. Defaults to non-iid.

output (str): Where to log files. Defaults to history.

client_epochs (int): Number of client epochs. Defaults to 10.

batch_size (int): Batch size. Defaults to 32.

number_of_clients (int): Number of clients. Defaults to 10.

number_of_clients_per_round (int): Number of clients per round. Defaults to 5.

number_of_rounds (int): Number of global rounds. Defaults to 15.

iterations (int): Number of RFA iterations. Defaults to 3.

v (float): L2 threshold. Defaults to 1e-6.

compression (bool): If the data should be compressed. Defaults to False.

dp (bool): If the differential privacy should be applied. Defaults to False.

noise_multiplier (float): The noise multipler that shoud be used with DP.

clipping_norm (float): The clipping norm used with DP.

seed (int, optional): Random seed. Defaults to None.

Returns:

float: Returns training time after federated learning.

federated.optimization.federated.get_optimizer(optimizer: str, learning_rate: float) → Callable[], tensorflow.python.keras.optimizer_v2.optimizer_v2.OptimizerV2]

Function for getting correct optimizer.

Args:

optimizer (str): Optimizer to be used.

learning_rate (float): Learning rate.

Returns:

Callable[[], tf.keras.optimizers.Optimizer]: Returns function for optimizer.

federated.optimization.federated.iterative_process_fn(tff_model: tensorflow_federated.python.learning.model.Model, server_optimizer_fn: Callable[], tensorflow.python.keras.optimizer_v2.optimizer_v2.OptimizerV2], aggregation_method: str = 'fedavg', client_optimizer_fn: Optional[Callable[], tensorflow.python.keras.optimizer_v2.optimizer_v2.OptimizerV2]] = None, iterations: Optional[int] = None, client_weighting: Optional[tensorflow_federated.python.learning.federated_averaging.ClientWeighting] = None, v: Optional[float] = None, compression: bool = False, model_update_aggregation_factory: Optional[Callable[], tensorflow_federated.python.aggregators.factory.UnweightedAggregationFactory]] = None) → tensorflow_federated.python.core.templates.iterative_process.IterativeProcess

Function builds an iterative process that performs federated aggregation. The function offers federated averaging, federated stochastic gradient descent and robust federated aggregation.

Args:

tff_model (tff.learning.Model): Federated model object.

server_optimizer_fn (Callable[[], tf.keras.optimizers.Optimizer]): Server optimizer function.

aggregation_method (str, optional): Aggregation method. Defaults to “fedavg”.

client_optimizer_fn (Callable[[], tf.keras.optimizers.Optimizer], optional): Client optimizer function. Defaults to None.

iterations (int, optional): [description]. Defaults to None.

client_weighting (tff.learning.ClientWeighting, optional): Client weighting. Defaults to None.

v (float, optional): L2 threshold. Defaults to None.

compression (bool, optional): If the model should be compressed. Defaults to False.

model_update_aggregation_factory (Callable[ [], tff.aggregators.UnweightedAggregationFactory ], optional): If the model should be trained with DP. Defaults to None.

Returns:

tff.templates.IterativeProcess: An Iterative Process.