def test_build_to_ids_fn_truncates(self):
vocab = ['A', 'B', 'C']
max_seq_len = 1
bos = stackoverflow_dataset.get_special_tokens(len(vocab)).bos
to_ids_fn = stackoverflow_dataset.build_to_ids_fn(vocab, max_seq_len)
data = {'tokens': 'A B C'}
processed = to_ids_fn(data)
self.assertAllEqual(self.evaluate(processed), [bos, 1])
def test_build_to_ids_fn_embeds_all_vocab(self):
vocab = ['A', 'B', 'C']
max_seq_len = 5
special_tokens = stackoverflow_dataset.get_special_tokens(len(vocab))
bos = special_tokens.bos
eos = special_tokens.eos
to_ids_fn = stackoverflow_dataset.build_to_ids_fn(vocab, max_seq_len)
data = {'tokens': 'A B C'}
processed = to_ids_fn(data)
self.assertAllEqual(self.evaluate(processed), [bos, 1, 2, 3, eos])
def test_oov_token_correct(self):
vocab = ['A', 'B', 'C']
max_seq_len = 5
num_oov_buckets = 2
to_ids_fn = stackoverflow_dataset.build_to_ids_fn(
vocab, max_seq_len, num_oov_buckets=num_oov_buckets)
oov_tokens = stackoverflow_dataset.get_special_tokens(
len(vocab), num_oov_buckets=num_oov_buckets).oov
data = {'tokens': 'A B D'}
processed = to_ids_fn(data)
self.assertLen(oov_tokens, num_oov_buckets)
self.assertIn(self.evaluate(processed)[3], oov_tokens)
def test_pad_token_correct(self):
vocab = ['A', 'B', 'C']
max_seq_len = 5
to_ids_fn = stackoverflow_dataset.build_to_ids_fn(vocab, max_seq_len)
special_tokens = stackoverflow_dataset.get_special_tokens(len(vocab))
pad, bos, eos = special_tokens.pad, special_tokens.bos, special_tokens.eos
data = {'tokens': 'A B C'}
processed = to_ids_fn(data)
batched_ds = tf.data.Dataset.from_tensor_slices([processed]).padded_batch(
1, padded_shapes=[6])
sample_elem = next(iter(batched_ds))
self.assertAllEqual(self.evaluate(sample_elem), [[bos, 1, 2, 3, eos, pad]])
def main(argv):
if len(argv) > 1:
raise app.UsageError('Expected no command-line arguments, '
'got: {}'.format(argv))
tff.backends.native.set_local_execution_context(max_fanout=10)
model_builder = functools.partial(
stackoverflow_models.create_recurrent_model,
vocab_size=FLAGS.vocab_size,
embedding_size=FLAGS.embedding_size,
latent_size=FLAGS.latent_size,
num_layers=FLAGS.num_layers,
shared_embedding=FLAGS.shared_embedding)
loss_builder = functools.partial(
tf.keras.losses.SparseCategoricalCrossentropy, from_logits=True)
special_tokens = stackoverflow_dataset.get_special_tokens(FLAGS.vocab_size)
pad_token = special_tokens.pad
oov_tokens = special_tokens.oov
eos_token = special_tokens.eos
def metrics_builder():
return [
keras_metrics.MaskedCategoricalAccuracy(name='accuracy_with_oov',
masked_tokens=[pad_token]),
keras_metrics.MaskedCategoricalAccuracy(name='accuracy_no_oov',
masked_tokens=[pad_token] +
oov_tokens),
# Notice BOS never appears in ground truth.
keras_metrics.MaskedCategoricalAccuracy(
name='accuracy_no_oov_or_eos',
masked_tokens=[pad_token, eos_token] + oov_tokens),
keras_metrics.NumBatchesCounter(),
keras_metrics.NumTokensCounter(masked_tokens=[pad_token]),
]
datasets = stackoverflow_dataset.construct_word_level_datasets(
FLAGS.vocab_size, FLAGS.client_batch_size,
FLAGS.client_epochs_per_round, FLAGS.sequence_length,
FLAGS.max_elements_per_user, FLAGS.num_validation_examples)
train_dataset, validation_dataset, test_dataset = datasets
if FLAGS.uniform_weighting:
def client_weight_fn(local_outputs):
del local_outputs
return 1.0
else:
def client_weight_fn(local_outputs):
return tf.cast(tf.squeeze(local_outputs['num_tokens']), tf.float32)
def model_fn():
return tff.learning.from_keras_model(
model_builder(),
loss_builder(),
input_spec=validation_dataset.element_spec,
metrics=metrics_builder())
if FLAGS.noise_multiplier is not None:
if not FLAGS.uniform_weighting:
raise ValueError(
'Differential privacy is only implemented for uniform weighting.'
)
dp_query = tff.utils.build_dp_query(
clip=FLAGS.clip,
noise_multiplier=FLAGS.noise_multiplier,
expected_total_weight=FLAGS.clients_per_round,
adaptive_clip_learning_rate=FLAGS.adaptive_clip_learning_rate,
target_unclipped_quantile=FLAGS.target_unclipped_quantile,
clipped_count_budget_allocation=FLAGS.
clipped_count_budget_allocation,
expected_clients_per_round=FLAGS.clients_per_round)
weights_type = tff.learning.framework.weights_type_from_model(model_fn)
aggregation_process = tff.utils.build_dp_aggregate_process(
weights_type.trainable, dp_query)
else:
aggregation_process = None
server_optimizer_fn = optimizer_utils.create_optimizer_fn_from_flags(
'server')
client_optimizer_fn = optimizer_utils.create_optimizer_fn_from_flags(
'client')
iterative_process = tff.learning.build_federated_averaging_process(
model_fn=model_fn,
server_optimizer_fn=server_optimizer_fn,
client_weight_fn=client_weight_fn,
client_optimizer_fn=client_optimizer_fn,
aggregation_process=aggregation_process)
client_datasets_fn = training_utils.build_client_datasets_fn(
train_dataset, FLAGS.clients_per_round)
evaluate_fn = training_utils.build_centralized_evaluate_fn(
model_builder=model_builder,
eval_dataset=validation_dataset,
loss_builder=loss_builder,
metrics_builder=metrics_builder)
test_fn = training_utils.build_centralized_evaluate_fn(
model_builder=model_builder,
# Use both val and test for symmetry with other experiments, which
# evaluate on the entire test set.
eval_dataset=validation_dataset.concatenate(test_dataset),
loss_builder=loss_builder,
metrics_builder=metrics_builder)
logging.info('Training model:')
logging.info(model_builder().summary())
hparam_dict = utils_impl.lookup_flag_values(utils_impl.get_hparam_flags())
training_loop_dict = utils_impl.lookup_flag_values(training_loop_flags)
training_loop.run(iterative_process=iterative_process,
client_datasets_fn=client_datasets_fn,
validation_fn=evaluate_fn,
test_fn=test_fn,
hparam_dict=hparam_dict,
**training_loop_dict)
def run_federated(
iterative_process_builder: Callable[...,
tff.templates.IterativeProcess],
client_epochs_per_round: int,
client_batch_size: int,
clients_per_round: int,
max_batches_per_client: Optional[int] = -1,
client_datasets_random_seed: Optional[int] = None,
vocab_size: Optional[int] = 10000,
num_oov_buckets: Optional[int] = 1,
sequence_length: Optional[int] = 20,
max_elements_per_user: Optional[int] = 1000,
num_validation_examples: Optional[int] = 10000,
embedding_size: Optional[int] = 96,
latent_size: Optional[int] = 670,
num_layers: Optional[int] = 1,
shared_embedding: Optional[bool] = False,
total_rounds: Optional[int] = 1500,
experiment_name: Optional[str] = 'federated_so_nwp',
root_output_dir: Optional[str] = '/tmp/fed_opt',
max_eval_batches: Optional[int] = None,
**kwargs):
"""Runs an iterative process on the Stack Overflow next word prediction task.
This method will load and pre-process dataset and construct a model used for
the task. It then uses `iterative_process_builder` to create an iterative
process that it applies to the task, using
`federated_research.utils.training_loop`.
We assume that the iterative process has the following functional type
signatures:
* `initialize`: `( -> S@SERVER)` where `S` represents the server state.
* `next`: `<S@SERVER, {B*}@CLIENTS> -> <S@SERVER, T@SERVER>` where `S`
represents the server state, `{B*}` represents the client datasets,
and `T` represents a python `Mapping` object.
Moreover, the server state must have an attribute `model` of type
`tff.learning.ModelWeights`.
Args:
iterative_process_builder: A function that accepts a no-arg `model_fn`, a
`client_weight_fn` and returns a `tff.templates.IterativeProcess`. The
`model_fn` must return a `tff.learning.Model`.
client_epochs_per_round: An integer representing the number of epochs of
training performed per client in each training round.
client_batch_size: An integer representing the batch size used on clients.
clients_per_round: An integer representing the number of clients
participating in each round.
max_batches_per_client: An optional int specifying the number of batches
taken by each client at each round. If `-1`, the entire client dataset is
used.
client_datasets_random_seed: An optional int used to seed which clients are
sampled at each round. If `None`, no seed is used.
vocab_size: Integer dictating the number of most frequent words to use in
the vocabulary.
num_oov_buckets: The number of out-of-vocabulary buckets to use.
sequence_length: The maximum number of words to take for each sequence.
max_elements_per_user: The maximum number of elements processed for each
client's dataset.
num_validation_examples: The number of test examples to use for validation.
embedding_size: The dimension of the word embedding layer.
latent_size: The dimension of the latent units in the recurrent layers.
num_layers: The number of stacked recurrent layers to use.
shared_embedding: Boolean indicating whether to tie input and output
embeddings.
total_rounds: The number of federated training rounds.
experiment_name: The name of the experiment being run. This will be appended
to the `root_output_dir` for purposes of writing outputs.
root_output_dir: The name of the root output directory for writing
experiment outputs.
max_eval_batches: If set to a positive integer, evaluation datasets are
capped to at most that many batches. If set to None or a nonpositive
integer, the full evaluation datasets are used.
**kwargs: Additional arguments configuring the training loop. For details
on supported arguments, see
`federated_research/utils/training_utils.py`.
"""
model_builder = functools.partial(
stackoverflow_models.create_recurrent_model,
vocab_size=vocab_size,
num_oov_buckets=num_oov_buckets,
embedding_size=embedding_size,
latent_size=latent_size,
num_layers=num_layers,
shared_embedding=shared_embedding)
loss_builder = functools.partial(
tf.keras.losses.SparseCategoricalCrossentropy, from_logits=True)
special_tokens = stackoverflow_dataset.get_special_tokens(
vocab_size, num_oov_buckets)
pad_token = special_tokens.pad
oov_tokens = special_tokens.oov
eos_token = special_tokens.eos
def metrics_builder():
return [
keras_metrics.MaskedCategoricalAccuracy(name='accuracy_with_oov',
masked_tokens=[pad_token]),
keras_metrics.MaskedCategoricalAccuracy(name='accuracy_no_oov',
masked_tokens=[pad_token] +
oov_tokens),
# Notice BOS never appears in ground truth.
keras_metrics.MaskedCategoricalAccuracy(
name='accuracy_no_oov_or_eos',
masked_tokens=[pad_token, eos_token] + oov_tokens),
keras_metrics.NumBatchesCounter(),
keras_metrics.NumTokensCounter(masked_tokens=[pad_token])
]
train_clientdata, _, _ = tff.simulation.datasets.stackoverflow.load_data()
# TODO(b/161914546): consider moving evaluation to use
# `tff.learning.build_federated_evaluation` to get metrics over client
# distributions, as well as the example weight means from this centralized
# evaluation.
_, validation_dataset, test_dataset = stackoverflow_dataset.get_centralized_datasets(
vocab_size=vocab_size,
max_seq_len=sequence_length,
train_batch_size=client_batch_size,
max_validation_batches=max_eval_batches,
max_test_batches=max_eval_batches,
num_validation_examples=num_validation_examples,
num_oov_buckets=num_oov_buckets)
train_dataset_preprocess_comp = stackoverflow_dataset.create_train_dataset_preprocess_fn(
vocab=stackoverflow_dataset.create_vocab(vocab_size),
num_oov_buckets=num_oov_buckets,
client_batch_size=client_batch_size,
client_epochs_per_round=client_epochs_per_round,
max_seq_len=sequence_length,
max_training_elements_per_user=max_elements_per_user,
max_batches_per_user=max_batches_per_client)
input_spec = train_dataset_preprocess_comp.type_signature.result.element
def tff_model_fn() -> tff.learning.Model:
return tff.learning.from_keras_model(keras_model=model_builder(),
input_spec=input_spec,
loss=loss_builder(),
metrics=metrics_builder())
def client_weight_fn(local_outputs):
# Num_tokens is a tensor with type int64[1], to use as a weight need
# a float32 scalar.
return tf.cast(tf.squeeze(local_outputs['num_tokens']), tf.float32)
training_process = iterative_process_builder(
tff_model_fn, client_weight_fn=client_weight_fn)
training_process = tff.simulation.compose_dataset_computation_with_iterative_process(
train_dataset_preprocess_comp, training_process)
client_datasets_fn = training_utils.build_client_datasets_fn(
dataset=train_clientdata,
clients_per_round=clients_per_round,
random_seed=client_datasets_random_seed)
evaluate_fn = training_utils.build_centralized_evaluate_fn(
model_builder=model_builder,
eval_dataset=validation_dataset,
loss_builder=loss_builder,
metrics_builder=metrics_builder)
test_fn = training_utils.build_centralized_evaluate_fn(
model_builder=model_builder,
# Use both val and test for symmetry with other experiments, which
# evaluate on the entire test set.
eval_dataset=validation_dataset.concatenate(test_dataset),
loss_builder=loss_builder,
metrics_builder=metrics_builder)
logging.info('Training model:')
logging.info(model_builder().summary())
training_loop.run(iterative_process=training_process,
client_datasets_fn=client_datasets_fn,
validation_fn=evaluate_fn,
test_fn=test_fn,
total_rounds=total_rounds,
experiment_name=experiment_name,
root_output_dir=root_output_dir,
**kwargs)
def run_centralized(optimizer: tf.keras.optimizers.Optimizer,
experiment_name: str,
root_output_dir: str,
num_epochs: int,
batch_size: int,
decay_epochs: Optional[int] = None,
lr_decay: Optional[float] = None,
hparams_dict: Optional[Mapping[str, Any]] = None,
vocab_size: Optional[int] = 10000,
num_oov_buckets: Optional[int] = 1,
sequence_length: Optional[int] = 20,
num_validation_examples: Optional[int] = 10000,
embedding_size: Optional[int] = 96,
latent_size: Optional[int] = 670,
num_layers: Optional[int] = 1,
shared_embedding: Optional[bool] = False,
max_batches: Optional[int] = None):
"""Trains an RNN on the Stack Overflow next word prediction task.
Args:
optimizer: A `tf.keras.optimizers.Optimizer` used to perform training.
experiment_name: The name of the experiment. Part of the output directory.
root_output_dir: The top-level output directory for experiment runs. The
`experiment_name` argument will be appended, and the directory will
contain tensorboard logs, metrics written as CSVs, and a CSV of
hyperparameter choices (if `hparams_dict` is used).
num_epochs: The number of training epochs.
batch_size: The batch size, used for train, validation, and test.
decay_epochs: The number of epochs of training before decaying the learning
rate. If None, no decay occurs.
lr_decay: The amount to decay the learning rate by after `decay_epochs`
training epochs have occurred.
hparams_dict: A mapping with string keys representing the hyperparameters
and their values. If not None, this is written to CSV.
vocab_size: Integer dictating the number of most frequent words to use in
the vocabulary.
num_oov_buckets: The number of out-of-vocabulary buckets to use.
sequence_length: The maximum number of words to take for each sequence.
num_validation_examples: The number of test examples to use for validation.
embedding_size: The dimension of the word embedding layer.
latent_size: The dimension of the latent units in the recurrent layers.
num_layers: The number of stacked recurrent layers to use.
shared_embedding: Boolean indicating whether to tie input and output
embeddings.
max_batches: If set to a positive integer, datasets are capped to at most
that many batches. If set to None or a nonpositive integer, the full
datasets are used.
"""
train_dataset, validation_dataset, test_dataset = stackoverflow_dataset.get_centralized_datasets(
vocab_size=vocab_size,
max_seq_len=sequence_length,
train_batch_size=batch_size,
max_train_batches=max_batches,
max_validation_batches=max_batches,
max_test_batches=max_batches,
num_validation_examples=num_validation_examples,
num_oov_buckets=num_oov_buckets,
)
model = stackoverflow_models.create_recurrent_model(
vocab_size=vocab_size,
num_oov_buckets=num_oov_buckets,
name='stackoverflow-lstm',
embedding_size=embedding_size,
latent_size=latent_size,
num_layers=num_layers,
shared_embedding=shared_embedding)
special_tokens = stackoverflow_dataset.get_special_tokens(
vocab_size=vocab_size, num_oov_buckets=num_oov_buckets)
pad_token = special_tokens.pad
oov_tokens = special_tokens.oov
eos_token = special_tokens.eos
model.compile(
loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True),
optimizer=optimizer,
metrics=[
keras_metrics.MaskedCategoricalAccuracy(
name='accuracy_with_oov', masked_tokens=[pad_token]),
keras_metrics.MaskedCategoricalAccuracy(
name='accuracy_no_oov', masked_tokens=[pad_token] + oov_tokens),
keras_metrics.MaskedCategoricalAccuracy(
name='accuracy_no_oov_or_eos',
masked_tokens=[pad_token, eos_token] + oov_tokens),
])
centralized_training_loop.run(
keras_model=model,
train_dataset=train_dataset,
validation_dataset=validation_dataset,
test_dataset=test_dataset,
experiment_name=experiment_name,
root_output_dir=root_output_dir,
num_epochs=num_epochs,
hparams_dict=hparams_dict,
decay_epochs=decay_epochs,
lr_decay=lr_decay)