'can be a convolutional model (cnn) or a two hidden-layer '
'densely connected network (2nn).')
flags.DEFINE_integer('client_batch_size', 20,
'Batch size used on the client.')
flags.DEFINE_integer('clients_per_round', 10,
'How many clients to sample per round.')
flags.DEFINE_integer(
'client_epochs_per_round', 1,
'Number of client (inner optimizer) epochs per federated round.')
flags.DEFINE_boolean(
'uniform_weighting', False,
'Whether to weigh clients uniformly. If false, clients '
'are weighted by the number of samples.')
# Optimizer configuration (this defines one or more flags per optimizer).
utils_impl.define_optimizer_flags('server')
utils_impl.define_optimizer_flags('client')
# Differential privacy flags
flags.DEFINE_float('clip', 0.05, 'Initial clip.')
flags.DEFINE_float('noise_multiplier', None,
'Noise multiplier. If None, no DP is used.')
flags.DEFINE_float('adaptive_clip_learning_rate', 0,
'Adaptive clip learning rate.')
flags.DEFINE_float('target_unclipped_quantile', 0.5,
'Target unclipped quantile.')
flags.DEFINE_float(
'clipped_count_budget_allocation', 0.1,
'Fraction of privacy budget to allocate for clipped counts.')
flags.DEFINE_boolean(
'per_vector_clipping', False, 'Use per-vector clipping'
'in filenames.')
flags.DEFINE_integer('random_seed', 0, 'Random seed for the experiment.')
# Training hyperparameters
flags.DEFINE_integer('total_rounds', 200,
'Number of total training rounds.')
flags.DEFINE_integer('rounds_per_eval', 1, 'How often to evaluate')
flags.DEFINE_integer('train_clients_per_round', 2,
'How many clients to sample per round.')
flags.DEFINE_integer('client_epochs_per_round', 1,
'Number of epochs in the client to take per round.')
flags.DEFINE_integer('batch_size', 20, 'Batch size used on the client.')
flags.DEFINE_integer('num_pseudo_clients', 1, 'Number of pseudo-clients.')
# Optimizer configuration (this defines one or more flags per optimizer).
utils_impl.define_optimizer_flags('server',
defaults=dict(learning_rate=1.0))
utils_impl.define_optimizer_flags('client',
defaults=dict(learning_rate=0.2))
# Differential privacy hyperparameters
flags.DEFINE_float('clip', 0.05, 'Initial clip.')
flags.DEFINE_float('noise_multiplier', 1.0, 'Noise multiplier.')
flags.DEFINE_float('adaptive_clip_learning_rate', 0,
'Adaptive clip learning rate.')
flags.DEFINE_float('target_unclipped_quantile', 0.5,
'Target unclipped quantile.')
flags.DEFINE_float(
'clipped_count_budget_allocation', 0.1,
'Fraction of privacy budget to allocate for clipped counts.')
flags.DEFINE_boolean('use_per_vector', False, 'Use per-vector clipping.')
'sentences to use per user.')
flags.DEFINE_integer('batch_size', 8, 'Batch size used.')
flags.DEFINE_integer('sequence_length', 20, 'Max sequence length to use.')
flags.DEFINE_integer('epochs', 3, 'Number of epochs to train for.')
flags.DEFINE_integer('shuffle_buffer_size', 1000,
'Buffer size for data shuffling.')
flags.DEFINE_integer('num_validation_examples', 10000,
'Number of examples to take for validation set.')
flags.DEFINE_integer('num_test_examples', 10000,
'Number of examples to take for test set.')
flags.DEFINE_integer('tensorboard_update_frequency', 100 * 1000,
'Number of steps between tensorboard logging calls.')
flags.DEFINE_string('root_output_dir', '/tmp/centralized_stackoverflow/',
'Root directory for writing experiment output.')
flags.DEFINE_integer('steps_per_epoch', None, 'Steps per epoch')
utils_impl.define_optimizer_flags('centralized')
# Modeling flags
flags.DEFINE_integer(
'vocab_size', 10000,
'Size of the vocab to use; results in most `vocab_size` number of most '
'common words used as vocabulary.')
flags.DEFINE_integer('embedding_size', 96,
'Dimension of word embedding to use.')
flags.DEFINE_integer('latent_size', 512,
'Dimension of latent size to use in recurrent cell')
flags.DEFINE_integer('num_layers', 1,
'Number of stacked recurrent layers to use.')
flags.DEFINE_boolean(
'lstm', True,
'Boolean indicating LSTM recurrent cell. If False, GRU is used.')
def test_define_optimizer_unused_default(self):
with self.assertRaisesRegex(ValueError, 'not consumed'):
# Use a different prefix to avoid declaring duplicate flags:
utils_impl.define_optimizer_flags('client', defaults=dict(lr=1.25))
def test_get_optimizer_from_flags(self):
utils_impl.define_optimizer_flags('server',
defaults=dict(learning_rate=1.25))
self.assertEqual(FLAGS.server_learning_rate, 1.25)
optimizer = utils_impl.get_optimizer_from_flags('server')
self.assertEqual(optimizer.get_config()['learning_rate'], 1.25)
def setUpModule():
# Create flags here to ensure duplicate flags are not created.
utils_impl.define_optimizer_flags(TEST_SERVER_FLAG_PREFIX)
utils_impl.define_optimizer_flags(TEST_CLIENT_FLAG_PREFIX)
