def _get_gan(with_dp=False):
gan_loss_fns = gan_losses.get_gan_loss_fns('wasserstein')
server_gen_optimizer = tf.keras.optimizers.Adam()
client_disc_optimizer = tf.keras.optimizers.Adam()
train_generator_fn = gan_training_tf_fns.create_train_generator_fn(
gan_loss_fns, server_gen_optimizer)
train_discriminator_fn = gan_training_tf_fns.create_train_discriminator_fn(
gan_loss_fns, client_disc_optimizer)
if with_dp:
dp_average_query = tensorflow_privacy.QuantileAdaptiveClipAverageQuery(
initial_l2_norm_clip=BEFORE_DP_L2_NORM_CLIP,
noise_multiplier=0.3,
target_unclipped_quantile=3,
learning_rate=0.1,
clipped_count_stddev=0.0,
expected_num_records=10,
denominator=10.0)
else:
dp_average_query = None
return tff_gans.GanFnsAndTypes(
generator_model_fn=one_dim_gan.create_generator,
discriminator_model_fn=one_dim_gan.create_discriminator,
dummy_gen_input=next(iter(one_dim_gan.create_generator_inputs())),
dummy_real_data=next(iter(one_dim_gan.create_real_data())),
train_generator_fn=train_generator_fn,
train_discriminator_fn=train_discriminator_fn,
server_disc_update_optimizer_fn=lambda: tf.keras.optimizers.SGD(lr=1.0
),
train_discriminator_dp_average_query=dp_average_query)
def test_build_gan_training_process(self, with_dp):
gan = _get_gan(with_dp)
process = tff_gans.build_gan_training_process(gan)
server_state = process.initialize()
if with_dp:
# Check that initial DP averaging aggregator state is correct.
dp_averaging_state = server_state.aggregation_state[0]
self.assertAlmostEqual(
dp_averaging_state.numerator_state.sum_state.l2_norm_clip,
BEFORE_DP_L2_NORM_CLIP,
places=5)
self.assertAlmostEqual(
dp_averaging_state.numerator_state.sum_state.stddev,
BEFORE_DP_STD_DEV,
places=5)
client_dataset_sizes = [1, 3]
client_gen_inputs = [
one_dim_gan.create_generator_inputs().take(i)
for i in client_dataset_sizes
]
client_real_inputs = [
one_dim_gan.create_real_data().take(i)
for i in client_dataset_sizes
]
num_rounds = 2
for _ in range(num_rounds):
server_state = process.next(
server_state,
one_dim_gan.create_generator_inputs().take(1),
client_gen_inputs, client_real_inputs)
# Check that server counters have incremented.
counters = server_state.counters
self.assertDictEqual(
counters, {
'num_rounds':
num_rounds,
'num_generator_train_examples':
one_dim_gan.BATCH_SIZE * num_rounds,
'num_discriminator_train_examples':
num_rounds * one_dim_gan.BATCH_SIZE *
sum(client_dataset_sizes),
})
if with_dp:
# Check that DP averaging aggregator state has updated properly over the
# above rounds.
dp_averaging_state = server_state.aggregation_state[0]
self.assertAlmostEqual(
dp_averaging_state.numerator_state.sum_state.l2_norm_clip,
AFTER_2_RDS_DP_L2_NORM_CLIP,
places=5)
self.assertAlmostEqual(
dp_averaging_state.numerator_state.sum_state.stddev,
AFTER_2_RDS_DP_STD_DEV,
places=5)
def test_client_computation(self, with_dp):
gan = _get_gan(with_dp)
client_comp = tff_gans.build_client_computation(gan)
generator = gan.generator_model_fn()
discriminator = gan.discriminator_model_fn()
from_server = gan_training_tf_fns.FromServer(
generator_weights=generator.weights,
discriminator_weights=discriminator.weights)
client_output = client_comp(
one_dim_gan.create_generator_inputs().take(10),
one_dim_gan.create_real_data().take(10), from_server)
self.assertDictEqual(
client_output.counters,
{'num_discriminator_train_examples': 10 * one_dim_gan.BATCH_SIZE})
def test_client_and_server_computations(self):
train_generator_fn, train_discriminator_fn = (
_get_train_generator_and_discriminator_fns())
# N.B. The way we are using datasets and re-using the same
# generator and discriminator doesn't really "make sense" from an ML
# perspective, but it's sufficient for testing. For more proper usage of
# these functions, see training_loops.py.
generator = one_dim_gan.create_generator()
discriminator = one_dim_gan.create_discriminator()
gen_inputs = one_dim_gan.create_generator_inputs()
real_data = one_dim_gan.create_real_data()
server_state = gan_training_tf_fns.server_initial_state(
generator, discriminator)
# The aggregation state (e.g., used for handling DP averaging) is
# initialized to be empty. A user of the `server_initial_state` is expected
# to take the output `ServerState` object and populate this field, most
# likely via an instance of tff.templates.AggregationProcess.
self.assertEmpty(server_state.aggregation_state)
client_output = gan_training_tf_fns.client_computation(
gen_inputs.take(3), real_data.take(3),
gan_training_tf_fns.FromServer(
generator_weights=server_state.generator_weights,
discriminator_weights=server_state.discriminator_weights),
generator, discriminator, train_discriminator_fn)
server_disc_update_optimizer = tf.keras.optimizers.Adam()
for _ in range(2): # Train for 2 rounds
server_state = gan_training_tf_fns.server_computation(
server_state, gen_inputs.take(3), client_output, generator,
discriminator, server_disc_update_optimizer, train_generator_fn,
NEW_DP_AVERAGING_STATE)
counters = self.evaluate(server_state.counters)
self.assertDictEqual(
counters, {
'num_rounds': 2,
'num_discriminator_train_examples': 2 * 3 * one_dim_gan.BATCH_SIZE,
'num_generator_train_examples': 2 * 3 * one_dim_gan.BATCH_SIZE
})
# DP averaging aggregation state updates properly in server_computation().
self.assertEqual(server_state.aggregation_state, NEW_DP_AVERAGING_STATE)
def test_client_and_server_computations(self):
train_generator_fn, train_discriminator_fn = (
_get_train_generator_and_discriminator_fns())
# N.B. The way we are using datasets and re-using the same
# generator and discriminator doesn't really "make sense" from an ML
# perspective, but it's sufficient for testing. For more proper usage of
# these functions, see training_loops.py.
generator = one_dim_gan.create_generator()
discriminator = one_dim_gan.create_discriminator()
gen_inputs = one_dim_gan.create_generator_inputs()
real_data = one_dim_gan.create_real_data()
server_state = gan_training_tf_fns.server_initial_state(
generator, discriminator, INIT_DP_AVERAGING_STATE)
# DP averaging aggregation state is initialized properly in
# server_initial_state().
self.assertEqual(server_state.dp_averaging_state, INIT_DP_AVERAGING_STATE)
client_output = gan_training_tf_fns.client_computation(
gen_inputs.take(3), real_data.take(3),
gan_training_tf_fns.FromServer(
generator_weights=server_state.generator_weights,
discriminator_weights=server_state.discriminator_weights),
generator, discriminator, train_discriminator_fn)
server_disc_update_optimizer = tf.keras.optimizers.Adam()
for _ in range(2): # Train for 2 rounds
server_state = gan_training_tf_fns.server_computation(
server_state, gen_inputs.take(3), client_output, generator,
discriminator, server_disc_update_optimizer, train_generator_fn,
NEW_DP_AVERAGING_STATE)
counters = self.evaluate(server_state.counters)
self.assertDictEqual(
counters, {
'num_rounds': 2,
'num_discriminator_train_examples': 2 * 3 * one_dim_gan.BATCH_SIZE,
'num_generator_train_examples': 2 * 3 * one_dim_gan.BATCH_SIZE
})
# DP averaging aggregation state updates properly in server_computation().
self.assertEqual(server_state.dp_averaging_state, NEW_DP_AVERAGING_STATE)
def test_tff_training_loop(self, dp_average_query, checkpoint):
if checkpoint:
root_checkpoint_dir = os.path.join(self.get_temp_dir(),
'checkpoints')
else:
root_checkpoint_dir = None
train_generator_fn, train_discriminator_fn = (
_get_train_generator_and_discriminator_fns())
gan = tff_gans.GanFnsAndTypes(
generator_model_fn=one_dim_gan.create_generator,
discriminator_model_fn=one_dim_gan.create_discriminator,
dummy_gen_input=next(iter(one_dim_gan.create_generator_inputs())),
dummy_real_data=next(iter(one_dim_gan.create_real_data())),
train_generator_fn=train_generator_fn,
train_discriminator_fn=train_discriminator_fn,
server_disc_update_optimizer_fn=lambda: tf.keras.optimizers.SGD(
lr=1.0),
train_discriminator_dp_average_query=dp_average_query)
gen_inputs = one_dim_gan.create_generator_inputs()
real_data = one_dim_gan.create_real_data()
client_disc_train_steps = 2
server_gen_train_steps = 3
server_gen_inputs = iter(gen_inputs.window(server_gen_train_steps))
client_gen_inputs = iter(gen_inputs.window(client_disc_train_steps))
client_real_data = iter(real_data.window(client_disc_train_steps))
def server_gen_inputs_fn(_):
return next(server_gen_inputs)
num_clients = 2
def client_datasets_fn(_):
return [(next(client_gen_inputs), next(client_real_data))
for _ in range(num_clients)]
server_state, _ = training_loops.federated_training_loop(
gan,
server_gen_inputs_fn=server_gen_inputs_fn,
client_datasets_fn=client_datasets_fn,
total_rounds=2,
rounds_per_checkpoint=1,
root_checkpoint_dir=root_checkpoint_dir)
self.assertDictEqual(
server_state.counters, {
'num_rounds':
2,
'num_generator_train_examples':
2 * 3 * one_dim_gan.BATCH_SIZE,
'num_discriminator_train_examples':
(2 * 2 * one_dim_gan.BATCH_SIZE * num_clients)
})
if checkpoint:
# TODO(b/141112101): We shouldn't need to re-create the gan, should be
# able to reuse the instance from above. See comment inside tff_gans.py.
train_generator_fn, train_discriminator_fn = (
_get_train_generator_and_discriminator_fns())
gan = tff_gans.GanFnsAndTypes(
generator_model_fn=one_dim_gan.create_generator,
discriminator_model_fn=one_dim_gan.create_discriminator,
dummy_gen_input=next(
iter(one_dim_gan.create_generator_inputs())),
dummy_real_data=next(iter(one_dim_gan.create_real_data())),
train_generator_fn=train_generator_fn,
train_discriminator_fn=train_discriminator_fn,
server_disc_update_optimizer_fn=lambda: tf.keras.optimizers.
SGD(lr=1.0),
train_discriminator_dp_average_query=dp_average_query)
# Train one more round, which should resume from the checkpoint.
server_state, _ = training_loops.federated_training_loop(
gan,
server_gen_inputs_fn=server_gen_inputs_fn,
client_datasets_fn=client_datasets_fn,
total_rounds=3,
rounds_per_checkpoint=1,
root_checkpoint_dir=root_checkpoint_dir)
# Note: It would be better to return something from
# federated_training_loop indicating the number of rounds trained in this
# invocation, so we could verify the checkpoint was read.
self.assertDictEqual(
server_state.counters, {
'num_rounds':
3,
'num_generator_train_examples':
3 * 3 * one_dim_gan.BATCH_SIZE,
'num_discriminator_train_examples':
(3 * 2 * one_dim_gan.BATCH_SIZE * num_clients)
})