def evaluate(self, ins: fl.EvaluateIns) -> fl.EvaluateRes:
weights = fl.parameters_to_weights(ins[0])
# Use provided weights to update the local model
self.model.set_weights(weights)
# Evaluate the updated model on the local dataset
loss, _ = self.model.evaluate(self.x_test,
self.y_test,
batch_size=len(self.x_test),
verbose=2)
# Return the number of evaluation examples and the evaluation result (loss)
return len(self.x_test), float(loss)
def test_fit(self):
"""This test is currently quite simple and should be improved"""
# Prepare
client = GrpcClientProxy(cid="1", bridge=self.bridge_mock)
parameters = flower.weights_to_parameters([np.ones((2, 2))])
ins: flower.FitIns = (parameters, {})
# Execute
parameters_prime, num_examples, _ = client.fit(ins=ins)
# Assert
assert parameters_prime.tensor_type == "np"
assert flower.parameters_to_weights(parameters_prime) == []
assert num_examples == 10
def fit(self, ins: flwr.FitIns) -> flwr.FitRes:
weights: flwr.Weights = flwr.parameters_to_weights(ins[0])
config = ins[1]
log(
DEBUG,
"fit on %s (examples: %s), config %s",
self.cid,
self.num_examples_train,
config,
)
# Training configuration
# epoch_global = int(config["epoch_global"])
epochs = int(config["epochs"])
batch_size = int(config["batch_size"])
# lr_initial = float(config["lr_initial"])
# lr_decay = float(config["lr_decay"])
timeout = int(config["timeout"])
partial_updates = bool(int(config["partial_updates"]))
# Use provided weights to update the local model
self.model.set_weights(weights)
# Train the local model using the local dataset
completed, fit_duration, num_examples = custom_fit(
model=self.model,
dataset=self.ds_train,
num_epochs=epochs,
batch_size=batch_size,
callbacks=[],
delay_factor=self.delay_factor,
timeout=timeout,
)
log(DEBUG, "client %s had fit_duration %s", self.cid, fit_duration)
# Compute the maximum number of examples which could have been processed
num_examples_ceil = self.num_examples_train * epochs
# Return empty update if local update could not be completed in time
if not completed and not partial_updates:
parameters = flwr.weights_to_parameters([])
return parameters, num_examples, num_examples_ceil
# Return the refined weights and the number of examples used for training
parameters = flwr.weights_to_parameters(self.model.get_weights())
return parameters, num_examples, num_examples_ceil
def evaluate(self, ins: flwr.EvaluateIns) -> flwr.EvaluateRes:
weights = flwr.parameters_to_weights(ins[0])
config = ins[1]
log(
DEBUG,
"evaluate on %s (examples: %s), config %s",
self.cid,
self.num_examples_test,
config,
)
# Use provided weights to update the local model
self.model.set_weights(weights)
# Evaluate the updated model on the local dataset
loss, _ = keras_evaluate(self.model,
self.ds_test,
batch_size=self.num_examples_test)
# Return the number of evaluation examples and the evaluation result (loss)
return self.num_examples_test, loss
def fit(self, ins: fl.FitIns) -> fl.FitRes:
weights: fl.Weights = fl.parameters_to_weights(ins[0])
config = ins[1]
# Get training
epochs = int(config["epochs"])
batch_size = int(config["batch_size"])
# Use provided weights to update the local model
self.model.set_weights(weights)
# Train the local model using the local dataset
self.model.fit(self.x_train,
self.y_train,
epochs=epochs,
batch_size=batch_size,
verbose=2)
# Return the refined weights and the number of examples used for training
weights_prime = fl.weights_to_parameters(self.model.get_weights())
num_examples = len(self.x_train)
return weights_prime, num_examples, num_examples