def test_run_simple_model(self):
vocab_size = 6
mask_model = shakespeare_models.create_recurrent_model(
vocab_size, sequence_length=5)
mask_model.compile(optimizer='sgd',
loss='sparse_categorical_crossentropy',
metrics=[keras_metrics.MaskedCategoricalAccuracy()])
no_mask_model = shakespeare_models.create_recurrent_model(
vocab_size, sequence_length=5, mask_zero=False)
no_mask_model.compile(
optimizer='sgd',
loss='sparse_categorical_crossentropy',
metrics=[keras_metrics.MaskedCategoricalAccuracy()])
constant_test_weights = tf.nest.map_structure(tf.ones_like,
mask_model.weights)
mask_model.set_weights(constant_test_weights)
no_mask_model.set_weights(constant_test_weights)
# `tf.data.Dataset.from_tensor_slices` aggresively coalesces the input into
# a single tensor, but we want a tuple of two tensors per example, so we
# apply a transformation to split.
def split_to_tuple(t):
return (t[0, :], t[1, :])
data = tf.data.Dataset.from_tensor_slices([
([0, 1, 2, 3, 4], [1, 2, 3, 4, 0]),
([2, 3, 4, 0, 1], [3, 4, 0, 1, 2]),
]).map(split_to_tuple).batch(2)
mask_metrics = mask_model.evaluate(data)
no_mask_metrics = no_mask_model.evaluate(data)
self.assertNotAllClose(mask_metrics, no_mask_metrics, atol=1e-3)
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,
sequence_length: Optional[int] = 80,
max_batches: Optional[int] = None):
"""Trains a two-layer RNN on Shakespeare next-character-prediction.
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.
sequence_length: The sequence length used for Shakespeare preprocessing.
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, eval_dataset = shakespeare_dataset.get_centralized_datasets(
train_batch_size=batch_size,
max_train_batches=max_batches,
max_test_batches=max_batches,
sequence_length=sequence_length)
pad_token, _, _, _ = shakespeare_dataset.get_special_tokens()
model = shakespeare_models.create_recurrent_model(
vocab_size=VOCAB_SIZE, sequence_length=sequence_length)
model.compile(
optimizer=optimizer,
loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True),
metrics=[
keras_metrics.MaskedCategoricalAccuracy(masked_tokens=[pad_token])
])
centralized_training_loop.run(keras_model=model,
train_dataset=train_dataset,
validation_dataset=eval_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)
def main(argv):
if len(argv) > 1:
raise app.UsageError('Too many command-line arguments.')
train_client_data, test_client_data = (
tff.simulation.datasets.shakespeare.load_data())
def preprocess(ds):
return shakespeare_dataset.convert_snippets_to_character_sequence_examples(
dataset=ds,
batch_size=FLAGS.batch_size,
epochs=1,
shuffle_buffer_size=0,
sequence_length=FLAGS.shakespeare_sequence_length)
train_dataset = train_client_data.create_tf_dataset_from_all_clients()
if FLAGS.shuffle_train_data:
train_dataset = train_dataset.shuffle(buffer_size=10000)
train_dataset = preprocess(train_dataset)
eval_dataset = preprocess(
test_client_data.create_tf_dataset_from_all_clients())
optimizer = optimizer_utils.create_optimizer_fn_from_flags('centralized')()
pad_token, _, _, _ = shakespeare_dataset.get_special_tokens()
model = shakespeare_models.create_recurrent_model(
vocab_size=VOCAB_SIZE,
sequence_length=FLAGS.shakespeare_sequence_length)
model.compile(
optimizer=optimizer,
loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True),
metrics=[
keras_metrics.MaskedCategoricalAccuracy(masked_tokens=[pad_token])
])
hparams_dict = collections.OrderedDict([(name, FLAGS[name].value)
for name in hparam_flags])
centralized_training_loop.run(keras_model=model,
train_dataset=train_dataset,
validation_dataset=eval_dataset,
experiment_name=FLAGS.experiment_name,
root_output_dir=FLAGS.root_output_dir,
num_epochs=FLAGS.num_epochs,
hparams_dict=hparams_dict,
decay_epochs=FLAGS.decay_epochs,
lr_decay=FLAGS.lr_decay)
def main(argv):
if len(argv) > 1:
raise app.UsageError('Too many command-line arguments.')
experiment_output_dir = FLAGS.root_output_dir
tensorboard_dir = os.path.join(experiment_output_dir, 'logdir',
FLAGS.experiment_name)
results_dir = os.path.join(experiment_output_dir, 'results',
FLAGS.experiment_name)
for path in [experiment_output_dir, tensorboard_dir, results_dir]:
try:
tf.io.gfile.makedirs(path)
except tf.errors.OpError:
pass # Directory already exists.
hparam_dict = collections.OrderedDict([(name, FLAGS[name].value)
for name in hparam_flags])
hparam_dict['results_file'] = results_dir
hparams_file = os.path.join(results_dir, 'hparams.csv')
logging.info('Saving hyper parameters to: [%s]', hparams_file)
utils_impl.atomic_write_to_csv(pd.Series(hparam_dict), hparams_file)
train_client_data, test_client_data = (
tff.simulation.datasets.shakespeare.load_data())
def preprocess(ds):
return shakespeare_dataset.convert_snippets_to_character_sequence_examples(
ds, FLAGS.batch_size, epochs=1).cache()
train_dataset = train_client_data.create_tf_dataset_from_all_clients()
if FLAGS.shuffle_train_data:
train_dataset = train_dataset.shuffle(buffer_size=10000)
train_dataset = preprocess(train_dataset)
eval_dataset = preprocess(
test_client_data.create_tf_dataset_from_all_clients())
optimizer = optimizer_utils.create_optimizer_fn_from_flags('centralized')()
pad_token, _, _, _ = shakespeare_dataset.get_special_tokens()
# Vocabulary with one OOV ID and zero for the mask.
vocab_size = len(shakespeare_dataset.CHAR_VOCAB) + 2
model = shakespeare_models.create_recurrent_model(
vocab_size=vocab_size, batch_size=FLAGS.batch_size)
model.compile(
optimizer=optimizer,
loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True),
metrics=[
keras_metrics.MaskedCategoricalAccuracy(masked_tokens=[pad_token])
])
logging.info('Training model:')
logging.info(model.summary())
csv_logger_callback = keras_callbacks.AtomicCSVLogger(results_dir)
tensorboard_callback = tf.keras.callbacks.TensorBoard(
log_dir=tensorboard_dir)
# Reduce the learning rate every 20 epochs.
def decay_lr(epoch, lr):
if (epoch + 1) % 20 == 0:
return lr * 0.1
else:
return lr
lr_callback = tf.keras.callbacks.LearningRateScheduler(decay_lr, verbose=1)
history = model.fit(
train_dataset,
validation_data=eval_dataset,
epochs=FLAGS.num_epochs,
callbacks=[lr_callback, tensorboard_callback, csv_logger_callback])
logging.info('Final metrics:')
for name in ['loss', 'accuracy']:
metric = history.history['val_{}'.format(name)][-1]
logging.info('\t%s: %.4f', name, metric)
def model_builder():
"""Constructs a `tf.keras.Model` to train."""
return shakespeare_models.create_recurrent_model(
vocab_size=VOCAB_SIZE, sequence_length=FLAGS.sequence_length)