def test_update_state_with_no_special_character(self):
metric = keras_metrics.MaskedCategoricalAccuracy()
metric.update_state(
y_true=[[1, 2, 3, 4], [0, 0, 0, 0]],
y_pred=[
# A batch with 100% accruacy.
[
[0.1, 0.9, 0.1, 0.1, 0.1],
[0.1, 0.1, 0.9, 0.1, 0.1],
[0.1, 0.1, 0.1, 0.9, 0.1],
[0.1, 0.1, 0.1, 0.1, 0.9],
],
# A batch with 50% accruacy.
[
[0.1, 0.9, 0.1, 0.1, 0.1],
[0.9, 0.1, 0.1, 0.1, 0.1],
[0.1, 0.1, 0.1, 0.9, 0.1],
[0.9, 0.1, 0.1, 0.1, 0.0],
],
])
self.assertEqual(self.evaluate(metric.result()), 6 / 8.0)
metric.update_state(
y_true=[[0, 4, 1, 2]],
y_pred=[
# A batch with 25% accruacy.
[
[0.9, 0.1, 0.1, 0.1, 0.1],
[0.1, 0.1, 0.1, 0.1, 0.9],
[0.1, 0.1, 0.1, 0.9, 0.1],
[0.1, 0.1, 0.1, 0.1, 0.9],
],
])
self.assertAllClose(self.evaluate(metric.result()), 8 / 12.0)
def test_update_state_with_special_character(self):
metric = keras_metrics.MaskedCategoricalAccuracy(masked_tokens=[4])
metric.update_state(
y_true=[[1, 2, 3, 4], [0, 0, 0, 0]],
y_pred=[
# A batch with 100% accruacy.
[
[0.1, 0.9, 0.1, 0.1, 0.1],
[0.1, 0.1, 0.9, 0.1, 0.1],
[0.1, 0.1, 0.1, 0.9, 0.1],
[0.1, 0.1, 0.1, 0.1, 0.9],
],
# A batch with 50% accruacy.
[
[0.1, 0.9, 0.1, 0.1, 0.1],
[0.9, 0.1, 0.1, 0.1, 0.1],
[0.1, 0.1, 0.1, 0.9, 0.1],
[0.9, 0.1, 0.1, 0.1, 0.0],
],
])
self.assertAllClose(self.evaluate(metric.result()), 5 / 7.0)
metric.update_state(
y_true=[[0, 4, 1, 2]],
y_pred=[
# A batch with 33% accruacy.
[
[0.9, 0.1, 0.1, 0.1, 0.1],
[0.1, 0.1, 0.1, 0.1, 0.9],
[0.1, 0.1, 0.1, 0.9, 0.1],
[0.1, 0.1, 0.1, 0.1, 0.9],
],
])
self.assertAllClose(self.evaluate(metric.result()), 6 / 10.0)
def test_constructor_no_masked_token(self):
metric_name = 'my_test_metric'
metric = keras_metrics.MaskedCategoricalAccuracy(name=metric_name)
self.assertIsInstance(metric, tf.keras.metrics.Metric)
self.assertEqual(metric.name, metric_name)
self.assertAllEqual(metric.get_config()['masked_tokens'], [])
self.assertEqual(self.evaluate(metric.result()), 0.0)
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 metrics_builder():
"""Returns a `list` of `tf.keras.metric.Metric` objects."""
pad_token, _, _, _ = shakespeare_dataset.get_special_tokens()
return [
keras_metrics.NumBatchesCounter(),
keras_metrics.NumExamplesCounter(),
keras_metrics.NumTokensCounter(masked_tokens=[pad_token]),
keras_metrics.MaskedCategoricalAccuracy(masked_tokens=[pad_token]),
]
def test_weighted_update_state_with_scalar_weight(self):
metric = keras_metrics.MaskedCategoricalAccuracy()
metric.update_state(
y_true=[[1, 2, 3, 4]],
y_pred=[
# A batch with 50% accuracy.
[
[0.9, 0.1, 0.1, 0.1, 0.1],
[0.1, 0.9, 0.1, 0.1, 0.1],
[0.1, 0.1, 0.1, 0.9, 0.1],
[0.1, 0.1, 0.1, 0.1, 0.9],
],
],
sample_weight=1.0)
self.assertAllClose(self.evaluate(metric.result()), .5)
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 test_update_state_with_all_tokens_masked(self):
metric = keras_metrics.MaskedCategoricalAccuracy(masked_tokens=[1, 2, 3, 4])
metric.update_state(
# All batches should be masked.
y_true=[[1, 2, 3, 4], [4, 3, 2, 1]],
y_pred=[
[
[0.1, 0.9, 0.1, 0.1, 0.1],
[0.1, 0.1, 0.9, 0.1, 0.1],
[0.1, 0.1, 0.1, 0.9, 0.1],
[0.1, 0.1, 0.1, 0.1, 0.9],
],
[
[0.1, 0.9, 0.1, 0.1, 0.1],
[0.9, 0.1, 0.1, 0.1, 0.1],
[0.1, 0.1, 0.1, 0.9, 0.1],
[0.9, 0.1, 0.1, 0.1, 0.0],
],
])
self.assertAllClose(self.evaluate(metric.result()), 0.0)
def test_update_state_with_multiple_tokens_masked(self):
metric = keras_metrics.MaskedCategoricalAccuracy(masked_tokens=[1, 2, 3, 4])
metric.update_state(
y_true=[[1, 2, 3, 4], [0, 0, 0, 0]],
y_pred=[
[
# This batch should be masked.
[0.1, 0.9, 0.1, 0.1, 0.1],
[0.1, 0.1, 0.9, 0.1, 0.1],
[0.1, 0.1, 0.1, 0.9, 0.1],
[0.1, 0.1, 0.1, 0.1, 0.9],
],
[
# Batch with 50% accuracy
[0.1, 0.9, 0.1, 0.1, 0.1],
[0.9, 0.1, 0.1, 0.1, 0.1],
[0.1, 0.1, 0.1, 0.9, 0.1],
[0.9, 0.1, 0.1, 0.1, 0.0],
],
])
self.assertAllClose(self.evaluate(metric.result()), 0.5)
def test_weighted_update_state_special_character_rank_2_sample_weight(self):
metric = keras_metrics.MaskedCategoricalAccuracy(masked_tokens=[4])
metric.update_state(
y_true=[[1, 2, 3, 4], [0, 0, 0, 0]],
y_pred=[
# A batch with 100% accuracy.
[
[0.1, 0.9, 0.1, 0.1, 0.1],
[0.1, 0.1, 0.9, 0.1, 0.1],
[0.1, 0.1, 0.1, 0.9, 0.1],
[0.1, 0.1, 0.1, 0.1, 0.9],
],
# A batch with 50% accuracy.
[
[0.1, 0.9, 0.1, 0.1, 0.1],
[0.9, 0.1, 0.1, 0.1, 0.1],
[0.1, 0.1, 0.1, 0.9, 0.1],
[0.9, 0.1, 0.1, 0.1, 0.0],
],
],
# A weight for each `y_true` scalar.
sample_weight=[[1.0, 2.0, 1.0, 2.0], [1.0, 2.0, 1.0, 2.0]])
self.assertAllClose(self.evaluate(metric.result()), (6 + 2) / 10.0)
def test_weighted_update_state_no_special_character(self):
metric = keras_metrics.MaskedCategoricalAccuracy()
metric.update_state(
y_true=[[1, 2, 3, 4], [0, 0, 0, 0]],
y_pred=[
# A batch with 100% accuracy.
[
[0.1, 0.9, 0.1, 0.1, 0.1],
[0.1, 0.1, 0.9, 0.1, 0.1],
[0.1, 0.1, 0.1, 0.9, 0.1],
[0.1, 0.1, 0.1, 0.1, 0.9],
],
# A batch with 50% accuracy.
[
[0.1, 0.9, 0.1, 0.1, 0.1],
[0.9, 0.1, 0.1, 0.1, 0.1],
[0.1, 0.1, 0.1, 0.9, 0.1],
[0.9, 0.1, 0.1, 0.1, 0.0],
],
],
# A weight for each `y_true` scalar.
sample_weight=[1.0, 2.0, 1.0, 2.0, 1.0, 2.0, 1.0, 2.0])
self.assertAllClose(self.evaluate(metric.result()), (6 + 4) / 12.0)
metric.update_state(
y_true=[[0, 4, 1, 2]],
y_pred=[
# A batch with 25% accruacy.
[
[0.9, 0.1, 0.1, 0.1, 0.1],
[0.1, 0.1, 0.1, 0.1, 0.9],
[0.1, 0.1, 0.1, 0.9, 0.1],
[0.1, 0.1, 0.1, 0.1, 0.9],
],
],
sample_weight=[1.0, 1.0, 2.0, 2.0])
self.assertAllClose(self.evaluate(metric.result()), (6 + 4 + 2) / 18.0)
def test_weighted_update_state_with_masked_token(self):
metric = keras_metrics.MaskedCategoricalAccuracy(masked_tokens=[4])
metric.update_state(
y_true=[[1, 2, 3, 4], [0, 0, 0, 0]],
y_pred=[
# A batch with 100% accuracy.
[
[0.1, 0.9, 0.1, 0.1, 0.1],
[0.1, 0.1, 0.9, 0.1, 0.1],
[0.1, 0.1, 0.1, 0.9, 0.1],
[0.1, 0.1, 0.1, 0.1, 0.9],
],
# A batch with 50% accuracy.
[
[0.1, 0.9, 0.1, 0.1, 0.1],
[0.9, 0.1, 0.1, 0.1, 0.1],
[0.1, 0.1, 0.1, 0.9, 0.1],
[0.9, 0.1, 0.1, 0.1, 0.0],
],
],
# A weight for each `y_true` scalar.
sample_weight=[[1.0, 2.0, 1.0, 2.0], [1.0, 2.0, 1.0, 2.0]])
self.assertAllClose(self.evaluate(metric.result()), (4 + 4) / 10.0)
metric.update_state(
y_true=[[0, 4, 1, 2]],
y_pred=[
# A batch with 25% accruacy.
[
[0.9, 0.1, 0.1, 0.1, 0.1],
[0.1, 0.1, 0.1, 0.1, 0.9],
[0.1, 0.1, 0.1, 0.9, 0.1],
[0.1, 0.1, 0.1, 0.1, 0.9],
],
],
sample_weight=[1.0, 1.0, 2.0, 2.0])
self.assertAllClose(self.evaluate(metric.result()), (4 + 4 + 1) / 15.0)
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)
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 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, _, _, _ = dataset.get_special_tokens()
# Vocabulary with one OOV ID and zero for the mask.
vocab_size = len(dataset.CHAR_VOCAB) + 2
model = 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 run_experiment():
"""Runs the training experiment."""
_, validation_dataset, test_dataset = dataset.construct_word_level_datasets(
FLAGS.vocab_size, FLAGS.batch_size, 1, FLAGS.sequence_length, -1,
FLAGS.num_validation_examples)
train_dataset = dataset.get_centralized_train_dataset(
FLAGS.vocab_size, FLAGS.batch_size, FLAGS.sequence_length,
FLAGS.shuffle_buffer_size)
model = models.create_recurrent_model(
vocab_size=FLAGS.vocab_size,
name='stackoverflow-lstm',
embedding_size=FLAGS.embedding_size,
latent_size=FLAGS.latent_size,
num_layers=FLAGS.num_layers,
shared_embedding=FLAGS.shared_embedding)
logging.info('Training model: %s', model.summary())
optimizer = optimizer_utils.create_optimizer_fn_from_flags('centralized')()
pad_token, oov_token, _, eos_token = dataset.get_special_tokens(
FLAGS.vocab_size)
model.compile(
loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True),
optimizer=optimizer,
metrics=[
# Plus 4 for pad, oov, bos, eos
keras_metrics.MaskedCategoricalAccuracy(
name='accuracy_with_oov', masked_tokens=[pad_token]),
keras_metrics.MaskedCategoricalAccuracy(
name='accuracy_no_oov', masked_tokens=[pad_token, oov_token]),
keras_metrics.MaskedCategoricalAccuracy(
name='accuracy_no_oov_or_eos',
masked_tokens=[pad_token, oov_token, eos_token]),
])
train_results_path = os.path.join(FLAGS.root_output_dir, 'train_results',
FLAGS.experiment_name)
test_results_path = os.path.join(FLAGS.root_output_dir, 'test_results',
FLAGS.experiment_name)
train_csv_logger = keras_callbacks.AtomicCSVLogger(train_results_path)
test_csv_logger = keras_callbacks.AtomicCSVLogger(test_results_path)
log_dir = os.path.join(FLAGS.root_output_dir, 'logdir', FLAGS.experiment_name)
try:
tf.io.gfile.makedirs(log_dir)
tf.io.gfile.makedirs(train_results_path)
tf.io.gfile.makedirs(test_results_path)
except tf.errors.OpError:
pass # log_dir already exists.
train_tensorboard_callback = tf.keras.callbacks.TensorBoard(
log_dir=log_dir,
write_graph=True,
update_freq=FLAGS.tensorboard_update_frequency)
test_tensorboard_callback = tf.keras.callbacks.TensorBoard(log_dir=log_dir)
# Write the hyperparameters to a CSV:
hparam_dict = collections.OrderedDict([
(name, FLAGS[name].value) for name in hparam_flags
])
hparams_file = os.path.join(FLAGS.root_output_dir, FLAGS.experiment_name,
'hparams.csv')
utils_impl.atomic_write_to_csv(pd.Series(hparam_dict), hparams_file)
model.fit(
train_dataset,
epochs=FLAGS.epochs,
verbose=0,
validation_data=validation_dataset,
callbacks=[train_csv_logger, train_tensorboard_callback])
score = model.evaluate(
test_dataset,
verbose=0,
callbacks=[test_csv_logger, test_tensorboard_callback])
logging.info('Final test loss: %.4f', score[0])
logging.info('Final test accuracy: %.4f', score[1])
