def test_self_contained_example_noclip(self, clip_norm=-1):
client_data = _create_client_data()
model = MnistModel()
optimizer_fn = lambda: tf.keras.optimizers.SGD(learning_rate=0.1)
losses = []
for _ in range(2):
optimizer = optimizer_fn()
_initialize_optimizer_vars(model, optimizer)
server_message = dp_fedavg.BroadcastMessage(
model_weights=model.weights, dp_clip_norm=clip_norm)
outputs = dp_fedavg.client_update(model, client_data(),
server_message, optimizer)
losses.append(outputs.model_output.numpy())
self.assertAllEqual(int(outputs.client_weight.numpy()), 2)
self.assertLess(losses[1], losses[0])
def test_self_contained_example(self, clip_norm, simulation_flag):
client_data = _create_client_data()
model = MnistModel()
optimizer_fn = lambda: tf.keras.optimizers.SGD(learning_rate=0.1)
losses = []
for _ in range(2):
optimizer = optimizer_fn()
_initialize_optimizer_vars(model, optimizer)
server_message = dp_fedavg.BroadcastMessage(
model_weights=model.weights, dp_clip_norm=clip_norm)
outputs = dp_fedavg.client_update(model, client_data(), server_message,
optimizer, simulation_flag)
losses.append(outputs.model_output.numpy())
weights_delta_norm = tf.linalg.global_norm(
tf.nest.flatten(outputs.weights_delta))
self.assertLessEqual(weights_delta_norm, clip_norm)
self.assertAllEqual(int(outputs.client_weight.numpy()), 2)
self.assertLess(losses[1], losses[0])