def test_run(self):
num_classes = 10
data, model = core.test_util.create_toy_example(
num_clients=10,
num_clusters=4,
num_classes=num_classes,
num_examples=5,
seed=0)
rng_seq = core.PRNGSequence(0)
algorithm = mime_lite.MimeLite(
federated_data=data,
model=model,
base_optimizer=core.get_optimizer(
core.OptimizerName.MOMENTUM, learning_rate=0.2, momentum=0.9),
hparams=mime_lite.MimeLiteHParams(
train_data_hparams=core.ClientDataHParams(batch_size=100),
combined_data_hparams=core.ClientDataHParams(batch_size=100),
server_learning_rate=1.0),
rng_seq=rng_seq,
)
dataset = core.create_tf_dataset_for_clients(data).batch(50)
state = algorithm.init_state()
init_metrics = core.evaluate_single_client(
dataset=dataset, model=model, params=state.params)
for _ in range(10):
state = algorithm.run_round(state, data.client_ids)
metrics = core.evaluate_single_client(
dataset=dataset, model=model, params=state.params)
self.assertLess(metrics['loss'], init_metrics['loss'])
self.assertGreater(metrics['accuracy'], init_metrics['accuracy'])
def test_run(self):
data, model = core.test_util.create_toy_example(num_clients=10,
num_clusters=4,
num_classes=10,
num_examples=5,
seed=0)
dataset = core.create_tf_dataset_for_clients(data).batch(50)
algorithm = fed_avg.FedAvg(
federated_data=data,
model=model,
client_optimizer=core.get_optimizer(core.OptimizerName.SGD,
learning_rate=0.1),
server_optimizer=core.get_optimizer(core.OptimizerName.MOMENTUM,
learning_rate=2.0,
momentum=0.9),
hparams=fed_avg.FedAvgHParams(
train_data_hparams=core.ClientDataHParams(batch_size=100,
num_epochs=1)),
rng_seq=core.PRNGSequence(0))
state = algorithm.init_state()
init_metrics = core.evaluate_single_client(dataset=dataset,
model=model,
params=state.params)
for _ in range(10):
state = algorithm.run_round(state, data.client_ids)
metrics = core.evaluate_single_client(dataset=dataset,
model=model,
params=state.params)
self.assertLess(metrics['loss'], init_metrics['loss'])
self.assertGreater(metrics['accuracy'], init_metrics['accuracy'])
def __call__(self, state: Any, round_num: int) -> core.MetricResults:
random_state = get_pseudo_random_state(round_num,
self._sample_client_random_seed)
client_ids = list(
random_state.choice(self._federated_data.client_ids,
size=self._num_clients_per_round,
replace=False))
combined_dataset = core.create_tf_dataset_for_clients(
self._federated_data, client_ids=client_ids)
return core.evaluate_single_client(combined_dataset, self._model,
state.params)
def test_run(self):
num_classes = 10
num_clusters = 3
federated_data, model = core.test_util.create_toy_example(
num_clients=10,
num_clusters=num_clusters,
num_classes=num_classes,
num_examples=5,
seed=0)
rng_seq = core.PRNGSequence(0)
algorithm = hyp_cluster.HypCluster(
federated_data=federated_data,
model=model,
client_optimizer=core.get_optimizer(core.OptimizerName.SGD,
learning_rate=0.1),
server_optimizer=core.get_optimizer(core.OptimizerName.SGD,
learning_rate=1.0),
hparams=hyp_cluster.HypClusterHParams(
train_data_hparams=core.ClientDataHParams(batch_size=5),
num_clusters=num_clusters),
rng_seq=rng_seq,
)
state = algorithm.init_state()
for _ in range(10):
state = algorithm.run_round(state, federated_data.client_ids)
with self.subTest('num_clusters'):
self.assertLen(state.cluster_params, num_clusters)
with self.subTest('maximization'):
data_hparams = core.ClientDataHParams(batch_size=5)
cluster_client_ids = hyp_cluster.maximization(
federated_data, federated_data.client_ids, model,
state.cluster_params, data_hparams)
for cluster_id, client_ids in enumerate(cluster_client_ids):
for client_id in client_ids:
dataset = federated_data.create_tf_dataset_for_client(
client_id)
dataset = core.preprocess_tf_dataset(dataset, data_hparams)
cluster_loss = []
for params in state.cluster_params:
cluster_loss.append(
core.evaluate_single_client(
dataset, model, params)['loss'])
# Cluster should be the best params for a client because clients are
# clustered based on empirical loss.
self.assertEqual(cluster_id,
cluster_loss.index(min(cluster_loss)))
def __call__(self, state: Any, round_num: int) -> core.MetricResults:
del round_num
return core.evaluate_single_client(self._dataset, self._model,
state.params)
def __call__(self, state: Any, round_num: int) -> core.MetricResults:
client_ids = self._sample_clients(round_num)
combined_dataset = core.create_tf_dataset_for_clients(
self._federated_data, client_ids=client_ids)
return core.evaluate_single_client(combined_dataset, self._model,
state.params)