def test_preprocess_fn_return_dataset_element_spec_oov_buckets(self):
ds = tf.data.Dataset.from_tensor_slices(TEST_DATA)
preprocess_fn = stackoverflow_word_prediction.create_preprocess_fn(
client_batch_size=32,
client_epochs_per_round=1,
max_sequence_length=10,
max_elements_per_client=100,
vocab=['one', 'must'],
num_oov_buckets=10)
preprocessed_ds = preprocess_fn(ds)
self.assertEqual(preprocessed_ds.element_spec,
(tf.TensorSpec(shape=[None, 10], dtype=tf.int64),
tf.TensorSpec(shape=[None, 10], dtype=tf.int64)))
def _preprocess_stackoverflow(vocab_size, num_oov_buckets, sequence_length,
num_validation_examples, client_batch_size,
client_epochs_per_round, max_elements_per_user):
"""Prepare stackoverflow dataset."""
train_clientdata, _, test_clientdata = (
tff.simulation.datasets.stackoverflow.load_data())
dataset_vocab = stackoverflow_dataset.create_vocab(vocab_size)
base_test_dataset = test_clientdata.create_tf_dataset_from_all_clients()
preprocess_val_and_test = stackoverflow_dataset.create_preprocess_fn(
vocab=dataset_vocab,
num_oov_buckets=num_oov_buckets,
client_batch_size=128,
client_epochs_per_round=client_epochs_per_round,
max_sequence_length=sequence_length,
max_elements_per_client=-1,
max_shuffle_buffer_size=1)
test_set = preprocess_val_and_test(
base_test_dataset.skip(num_validation_examples))
validation_set = preprocess_val_and_test(
base_test_dataset.take(num_validation_examples))
train_dataset_preprocess_comp = stackoverflow_dataset.create_preprocess_fn(
vocab=dataset_vocab,
num_oov_buckets=num_oov_buckets,
client_batch_size=client_batch_size,
client_epochs_per_round=client_epochs_per_round,
max_sequence_length=sequence_length,
max_elements_per_client=max_elements_per_user,
max_shuffle_buffer_size=max_elements_per_user)
@tff.tf_computation(tf.string)
def train_dataset_computation(client_id):
client_train_data = train_clientdata.dataset_computation(client_id)
return train_dataset_preprocess_comp(client_train_data)
return train_dataset_computation, train_clientdata, validation_set, test_set
def test_preprocess_fn_returns_correct_sequence(self):
ds = tf.data.Dataset.from_tensor_slices(TEST_DATA)
preprocess_fn = stackoverflow_word_prediction.create_preprocess_fn(
client_batch_size=32,
client_epochs_per_round=1,
max_sequence_length=6,
max_elements_per_client=100,
vocab=['one', 'must'],
num_oov_buckets=1)
preprocessed_ds = preprocess_fn(ds)
element = next(iter(preprocessed_ds))
# BOS is len(vocab)+2, EOS is len(vocab)+3, pad is 0, OOV is len(vocab)+1
self.assertAllEqual(self.evaluate(element[0]),
np.array([[4, 1, 2, 3, 5, 0]]))
def test_preprocess_fn_returns_correct_sequence_oov_buckets(self):
ds = tf.data.Dataset.from_tensor_slices(TEST_DATA)
preprocess_fn = stackoverflow_word_prediction.create_preprocess_fn(
client_batch_size=32,
client_epochs_per_round=1,
max_sequence_length=6,
max_elements_per_client=100,
vocab=['one', 'must'],
num_oov_buckets=3)
preprocessed_ds = preprocess_fn(ds)
element = next(iter(preprocessed_ds))
# BOS is len(vocab)+3+1
self.assertEqual(self.evaluate(element[0])[0][0], 6)
self.assertEqual(self.evaluate(element[0])[0][1], 1)
self.assertEqual(self.evaluate(element[0])[0][2], 2)
# OOV is [len(vocab)+1, len(vocab)+2, len(vocab)+3]
self.assertIn(self.evaluate(element[0])[0][3], [3, 4, 5])
# EOS is len(vocab)+3+2
self.assertEqual(self.evaluate(element[0])[0][4], 7)
# pad is 0
self.assertEqual(self.evaluate(element[0])[0][5], 0)
def run_federated(
iterative_process_builder: Callable[...,
tff.templates.IterativeProcess],
client_epochs_per_round: int,
client_batch_size: int,
clients_per_round: int,
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',
**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.
The iterative process must also have a callable attribute `get_model_weights`
that takes as input the state of the iterative process, and returns a
`tff.learning.ModelWeights` object.
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.
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.
**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_word_prediction.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_word_prediction.get_centralized_datasets(
vocab_size=vocab_size,
max_sequence_length=sequence_length,
num_validation_examples=num_validation_examples,
num_oov_buckets=num_oov_buckets)
train_dataset_preprocess_comp = stackoverflow_word_prediction.create_preprocess_fn(
vocab=stackoverflow_word_prediction.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_sequence_length=sequence_length,
max_elements_per_client=max_elements_per_user)
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)
iterative_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, iterative_process)
training_process.get_model_weights = iterative_process.get_model_weights
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)
validation_fn = lambda model_weights, round_num: evaluate_fn(model_weights)
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=validation_fn,
test_fn=test_fn,
total_rounds=total_rounds,
experiment_name=experiment_name,
root_output_dir=root_output_dir,
**kwargs)
def configure_training(
task_spec: training_specs.TaskSpec,
vocab_size: int = 10000,
num_oov_buckets: int = 1,
sequence_length: int = 20,
max_elements_per_user: int = 1000,
num_validation_examples: int = 10000,
embedding_size: int = 96,
latent_size: int = 670,
num_layers: int = 1,
shared_embedding: bool = False) -> training_specs.RunnerSpec:
"""Configures training for Stack Overflow next-word prediction.
This method will load and pre-process datasets and construct a model used for
the task. It then uses `iterative_process_builder` to create an iterative
process compatible with `federated_research.utils.training_loop`.
Args:
task_spec: A `TaskSpec` class for creating federated training tasks.
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.
Returns:
A `RunnerSpec` containing attributes used for running the newly created
federated task.
"""
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_word_prediction.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_word_prediction.get_centralized_datasets(
vocab_size=vocab_size,
max_sequence_length=sequence_length,
num_validation_examples=num_validation_examples,
num_oov_buckets=num_oov_buckets)
train_dataset_preprocess_comp = stackoverflow_word_prediction.create_preprocess_fn(
vocab=stackoverflow_word_prediction.create_vocab(vocab_size),
num_oov_buckets=num_oov_buckets,
client_batch_size=task_spec.client_batch_size,
client_epochs_per_round=task_spec.client_epochs_per_round,
max_sequence_length=sequence_length,
max_elements_per_client=max_elements_per_user)
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())
iterative_process = task_spec.iterative_process_builder(tff_model_fn)
@tff.tf_computation(tf.string)
def train_dataset_computation(client_id):
client_train_data = train_clientdata.dataset_computation(client_id)
return train_dataset_preprocess_comp(client_train_data)
training_process = tff.simulation.compose_dataset_computation_with_iterative_process(
train_dataset_computation, iterative_process)
client_ids_fn = training_utils.build_sample_fn(
train_clientdata.client_ids,
size=task_spec.clients_per_round,
replace=False,
random_seed=task_spec.client_datasets_random_seed)
# We convert the output to a list (instead of an np.ndarray) so that it can
# be used as input to the iterative process.
client_sampling_fn = lambda x: list(client_ids_fn(x))
training_process.get_model_weights = iterative_process.get_model_weights
centralized_validation_fn = training_utils.build_centralized_evaluate_fn(
model_builder=model_builder,
eval_dataset=validation_dataset,
loss_builder=loss_builder,
metrics_builder=metrics_builder)
def validation_fn(server_state, round_num):
del round_num
return centralized_validation_fn(
iterative_process.get_model_weights(server_state))
centralized_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)
def test_fn(server_state):
return centralized_test_fn(
iterative_process.get_model_weights(server_state))
return training_specs.RunnerSpec(iterative_process=training_process,
client_datasets_fn=client_sampling_fn,
validation_fn=validation_fn,
test_fn=test_fn)
def run_federated(iterative_process_builder: Callable[
..., tff.templates.IterativeProcess],
client_epochs_per_round: int,
client_batch_size: int,
clients_per_round: int,
max_elements_per_user: int,
total_rounds: int = 3000,
vocab_size: int = 10000,
num_oov_buckets: int = 1,
sequence_length: int = 20,
num_validation_examples: int = 10000,
dim_embed: int = 96,
dim_model: int = 512,
dim_hidden: int = 2048,
num_heads: int = 8,
num_layers: int = 1,
max_position_encoding: int = 1000,
dropout: float = 0.1,
client_datasets_random_seed: Optional[int] = None,
experiment_name: str = 'federated_stackoverflow',
root_output_dir: str = '/tmp/fedopt_guide',
max_val_test_batches: Optional[int] = None,
**kwargs) -> None:
"""Configures training for Stack Overflow next-word prediction.
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/fedopt_guide/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.
The iterative process must also have a callable attribute `get_model_weights`
that takes as input the state of the iterative process, and returns a
`tff.learning.ModelWeights` object.
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_elements_per_user: The maximum number of elements processed for each
client's dataset. This has be to a positive value or -1 (which means that
all elements are taken for training).
total_rounds: The number of federated training rounds.
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.
dim_embed: An integer for the dimension of the token embeddings.
dim_model: An integer for the dimension of features of MultiHeadAttention
layers.
dim_hidden: An integer for the dimension of hidden layers of the FFN.
num_heads: An integer for the number of attention heads.
num_layers: An integer for the number of Transformer blocks.
max_position_encoding: Maximum number of positions for position embeddings.
dropout: Dropout rate.
client_datasets_random_seed: An optional int used to seed which clients are
sampled at each round. If `None`, no seed is used.
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_val_test_batches: If set to a positive integer, val and test datasets
are capped to at most that many batches. If set to None or a nonpositive
integer, the full datasets are used.
**kwargs: Additional arguments configuring the training loop. For details on
supported arguments, see
`federated_research/fedopt_guide/training_utils.py`.
Returns:
A `RunnerSpec` containing attributes used for running the newly created
federated task.
"""
train_clientdata, _, _ = tff.simulation.datasets.stackoverflow.load_data()
_, validation_dataset, test_dataset = stackoverflow_word_prediction.get_centralized_datasets(
vocab_size=vocab_size,
max_sequence_length=sequence_length,
num_validation_examples=num_validation_examples,
num_oov_buckets=num_oov_buckets)
if max_val_test_batches and max_val_test_batches >= 1:
validation_dataset = validation_dataset.take(max_val_test_batches)
test_dataset = test_dataset.take(max_val_test_batches)
model_builder = functools.partial(
transformer_models.create_transformer_lm,
vocab_size=vocab_size,
num_oov_buckets=num_oov_buckets,
dim_embed=dim_embed,
dim_model=dim_model,
dim_hidden=dim_hidden,
num_heads=num_heads,
num_layers=num_layers,
max_position_encoding=max_position_encoding,
dropout=dropout,
name='stackoverflow-transformer')
loss_builder = functools.partial(
tf.keras.losses.SparseCategoricalCrossentropy, from_logits=True)
special_tokens = stackoverflow_word_prediction.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_dataset_preprocess_comp = stackoverflow_word_prediction.create_preprocess_fn(
vocab=stackoverflow_word_prediction.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_sequence_length=sequence_length,
max_elements_per_client=max_elements_per_user)
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)
iterative_process = iterative_process_builder(
tff_model_fn, client_weight_fn=client_weight_fn)
if hasattr(train_clientdata, 'dataset_computation'):
@tff.tf_computation(tf.string)
def train_dataset_computation(client_id):
client_train_data = train_clientdata.dataset_computation(client_id)
return train_dataset_preprocess_comp(client_train_data)
training_process = tff.simulation.compose_dataset_computation_with_iterative_process(
train_dataset_computation, iterative_process)
client_ids_fn = tff.simulation.build_uniform_sampling_fn(
train_clientdata.client_ids,
size=clients_per_round,
replace=False,
random_seed=client_datasets_random_seed)
# We convert the output to a list (instead of an np.ndarray) so that it can
# be used as input to the iterative process.
client_sampling_fn = lambda x: list(client_ids_fn(x))
else:
training_process = tff.simulation.compose_dataset_computation_with_iterative_process(
train_dataset_preprocess_comp, iterative_process)
client_sampling_fn = tff.simulation.build_uniform_client_sampling_fn(
dataset=train_clientdata,
clients_per_round=clients_per_round,
random_seed=client_datasets_random_seed)
training_process.get_model_weights = iterative_process.get_model_weights
evaluate_fn = tff.learning.build_federated_evaluation(tff_model_fn)
def validation_fn(model_weights, round_num):
del round_num
return evaluate_fn(model_weights, [validation_dataset])
def test_fn(model_weights):
return evaluate_fn(model_weights,
[validation_dataset.concatenate(test_dataset)])
logging.info('Training model:')
logging.info(model_builder().summary())
training_loop.run(iterative_process=training_process,
train_client_datasets_fn=client_sampling_fn,
evaluation_fn=validation_fn,
test_fn=test_fn,
total_rounds=total_rounds,
experiment_name=experiment_name,
root_output_dir=root_output_dir,
**kwargs)
