def test_fed_avg_with_adaptive_client_and_server(self, optimizer):
federated_data = [[_batch_fn()]]
client_optimizer_fn = functools.partial(optimizer, epsilon=0.01)
server_optimizer_fn = functools.partial(optimizer, epsilon=0.01)
iterative_process = fed_avg_client_opt.build_iterative_process(
_uncompiled_model_builder,
client_optimizer_fn=client_optimizer_fn,
server_optimizer_fn=server_optimizer_fn)
_, train_outputs = self._run_rounds(iterative_process, federated_data,
5)
self.assertLess(train_outputs[4]['loss'], train_outputs[0]['loss'])
def test_fed_avg_decreases_loss(self):
federated_data = [[_batch_fn()]]
client_optimizer_fn = tf.keras.optimizers.SGD
server_optimizer_fn = tf.keras.optimizers.SGD
iterative_process = fed_avg_client_opt.build_iterative_process(
_uncompiled_model_builder,
client_optimizer_fn=client_optimizer_fn,
server_optimizer_fn=server_optimizer_fn)
_, train_outputs = self._run_rounds(iterative_process, federated_data,
5)
self.assertLess(train_outputs[4]['loss'], train_outputs[0]['loss'])
def test_state_types(self):
federated_data = [[_batch_fn()]]
client_optimizer_fn = functools.partial(tf.keras.optimizers.Adam,
epsilon=0.01)
server_optimizer_fn = tf.keras.optimizers.SGD
iterative_process = fed_avg_client_opt.build_iterative_process(
_uncompiled_model_builder,
client_optimizer_fn=client_optimizer_fn,
server_optimizer_fn=server_optimizer_fn)
state, _ = self._run_rounds(iterative_process, federated_data, 1)
self.assertIsInstance(state, fed_avg_client_opt.ServerState)
self.assertIsInstance(state.model, tff.learning.ModelWeights)
def test_client_state_aggregate_mean(self, n, m):
dataset1 = [_batch_fn() for _ in range(n)]
dataset2 = [_batch_fn() for _ in range(m)]
federated_data = [dataset1, dataset2]
client_optimizer_fn = functools.partial(tf.keras.optimizers.Adam,
epsilon=0.01)
server_optimizer_fn = tf.keras.optimizers.SGD
iterative_process = fed_avg_client_opt.build_iterative_process(
_uncompiled_model_builder,
client_optimizer_fn=client_optimizer_fn,
server_optimizer_fn=server_optimizer_fn,
client_opt_weight_fn=lambda x: 1.0)
state = iterative_process.initialize()
state, _ = iterative_process.next(state, federated_data)
client_opt_iteration = state.client_optimizer_state.iterations
self.assertEqual(client_opt_iteration, (n + m) // 2)
def test_get_model_weights(self):
federated_data = [[_batch_fn()]]
iterative_process = fed_avg_client_opt.build_iterative_process(
_uncompiled_model_builder,
client_optimizer_fn=tf.keras.optimizers.SGD,
server_optimizer_fn=tf.keras.optimizers.SGD)
state = iterative_process.initialize()
self.assertIsInstance(iterative_process.get_model_weights(state),
tff.learning.ModelWeights)
self.assertAllClose(
state.model.trainable,
iterative_process.get_model_weights(state).trainable)
for _ in range(3):
state, _ = iterative_process.next(state, federated_data)
self.assertIsInstance(iterative_process.get_model_weights(state),
tff.learning.ModelWeights)
self.assertAllClose(
state.model.trainable,
iterative_process.get_model_weights(state).trainable)