def main(argv):
del argv # unused arg
tf.enable_v2_behavior()
tf.io.gfile.makedirs(FLAGS.output_dir)
logging.info('Saving checkpoints at %s', FLAGS.output_dir)
tf.random.set_seed(FLAGS.seed)
if FLAGS.use_gpu:
logging.info('Use GPU')
strategy = tf.distribute.MirroredStrategy()
else:
logging.info('Use TPU at %s',
FLAGS.tpu if FLAGS.tpu is not None else 'local')
resolver = tf.distribute.cluster_resolver.TPUClusterResolver(
tpu=FLAGS.tpu)
tf.config.experimental_connect_to_cluster(resolver)
tf.tpu.experimental.initialize_tpu_system(resolver)
strategy = tf.distribute.experimental.TPUStrategy(resolver)
ind_dataset_builder = ub.datasets.ClincIntentDetectionDataset(
batch_size=FLAGS.per_core_batch_size,
eval_batch_size=FLAGS.per_core_batch_size,
data_mode='ind')
ood_dataset_builder = ub.datasets.ClincIntentDetectionDataset(
batch_size=FLAGS.per_core_batch_size,
eval_batch_size=FLAGS.per_core_batch_size,
data_mode='ood')
dataset_builders = {
'clean': ind_dataset_builder,
'out_of_scope_requests': ood_dataset_builder
}
train_dataset = ind_dataset_builder.build(
split=ub.datasets.base.Split.TRAIN)
ds_info = ind_dataset_builder.info
feature_size = ds_info['feature_size']
# num_classes is number of valid intents plus out-of-scope intent
num_classes = ds_info['num_classes'] + 1
# vocab_size is total number of valid tokens plus the out-of-vocabulary token.
vocab_size = ind_dataset_builder.tokenizer.num_words + 1
batch_size = FLAGS.per_core_batch_size * FLAGS.num_cores
steps_per_epoch = ds_info['num_train_examples'] // batch_size
test_datasets = {}
steps_per_eval = {}
for dataset_name, dataset_builder in dataset_builders.items():
test_datasets[dataset_name] = dataset_builder.build(
split=ub.datasets.base.Split.TEST)
steps_per_eval[dataset_name] = (
dataset_builder.info['num_test_examples'] // batch_size)
if FLAGS.use_bfloat16:
policy = tf.keras.mixed_precision.experimental.Policy('mixed_bfloat16')
tf.keras.mixed_precision.experimental.set_policy(policy)
summary_writer = tf.summary.create_file_writer(
os.path.join(FLAGS.output_dir, 'summaries'))
premade_embedding_array = None
if FLAGS.word_embedding_dir:
with tf.io.gfile.GFile(FLAGS.word_embedding_dir,
'rb') as embedding_file:
premade_embedding_array = np.load(embedding_file)
with strategy.scope():
logging.info('Building %s model', FLAGS.model_family)
if FLAGS.model_family.lower() == 'textcnn':
model = cnn_model.textcnn(
filter_sizes=FLAGS.filter_sizes,
num_filters=FLAGS.num_filters,
num_classes=num_classes,
feature_size=feature_size,
vocab_size=vocab_size,
embed_size=FLAGS.embedding_size,
dropout_rate=FLAGS.dropout_rate,
l2=FLAGS.l2,
premade_embedding_arr=premade_embedding_array)
optimizer = tf.keras.optimizers.Adam(FLAGS.base_learning_rate)
elif FLAGS.model_family.lower() == 'bert':
bert_config_dir, bert_ckpt_dir = resolve_bert_ckpt_and_config_dir(
FLAGS.bert_dir, FLAGS.bert_config_dir, FLAGS.bert_ckpt_dir)
bert_config = bert_model.create_config(bert_config_dir)
model, bert_encoder = bert_model.create_model(
num_classes=num_classes,
feature_size=feature_size,
bert_config=bert_config)
optimizer = bert_model.create_optimizer(
FLAGS.base_learning_rate,
steps_per_epoch=steps_per_epoch,
epochs=FLAGS.train_epochs,
warmup_proportion=FLAGS.warmup_proportion)
else:
raise ValueError(
'model_family ({}) can only be TextCNN or BERT.'.format(
FLAGS.model_family))
logging.info('Model input shape: %s', model.input_shape)
logging.info('Model output shape: %s', model.output_shape)
logging.info('Model number of weights: %s', model.count_params())
metrics = {
'train/negative_log_likelihood':
tf.keras.metrics.Mean(),
'train/accuracy':
tf.keras.metrics.SparseCategoricalAccuracy(),
'train/loss':
tf.keras.metrics.Mean(),
'train/ece':
ed.metrics.ExpectedCalibrationError(num_bins=FLAGS.num_bins),
'test/negative_log_likelihood':
tf.keras.metrics.Mean(),
'test/accuracy':
tf.keras.metrics.SparseCategoricalAccuracy(),
'test/ece':
ed.metrics.ExpectedCalibrationError(num_bins=FLAGS.num_bins),
}
for dataset_name, test_dataset in test_datasets.items():
if dataset_name != 'clean':
metrics.update({
'test/nll_{}'.format(dataset_name):
tf.keras.metrics.Mean(),
'test/accuracy_{}'.format(dataset_name):
tf.keras.metrics.SparseCategoricalAccuracy(),
'test/ece_{}'.format(dataset_name):
ed.metrics.ExpectedCalibrationError(
num_bins=FLAGS.num_bins)
})
checkpoint = tf.train.Checkpoint(model=model, optimizer=optimizer)
latest_checkpoint = tf.train.latest_checkpoint(FLAGS.output_dir)
initial_epoch = 0
if latest_checkpoint:
# checkpoint.restore must be within a strategy.scope() so that optimizer
# slot variables are mirrored.
checkpoint.restore(latest_checkpoint)
logging.info('Loaded checkpoint %s', latest_checkpoint)
initial_epoch = optimizer.iterations.numpy() // steps_per_epoch
elif FLAGS.model_family.lower() == 'bert':
# load BERT from initial checkpoint
bert_checkpoint = tf.train.Checkpoint(model=bert_encoder)
bert_checkpoint.restore(
bert_ckpt_dir).assert_existing_objects_matched()
logging.info('Loaded BERT checkpoint %s', bert_ckpt_dir)
@tf.function
def train_step(iterator):
"""Training StepFn."""
def step_fn(inputs):
"""Per-Replica StepFn."""
features, labels = create_feature_and_label(
inputs, feature_size, model_family=FLAGS.model_family)
with tf.GradientTape() as tape:
# Set learning phase to enable dropout etc during training.
logits = model(features, training=True)
if FLAGS.use_bfloat16:
logits = tf.cast(logits, tf.float32)
negative_log_likelihood = tf.reduce_mean(
tf.keras.losses.sparse_categorical_crossentropy(
labels, logits, from_logits=True))
l2_loss = sum(model.losses)
loss = negative_log_likelihood + l2_loss
# Scale the loss given the TPUStrategy will reduce sum all gradients.
scaled_loss = loss / strategy.num_replicas_in_sync
grads = tape.gradient(scaled_loss, model.trainable_variables)
optimizer.apply_gradients(zip(grads, model.trainable_variables))
probs = tf.nn.softmax(logits)
metrics['train/ece'].update_state(labels, probs)
metrics['train/loss'].update_state(loss)
metrics['train/negative_log_likelihood'].update_state(
negative_log_likelihood)
metrics['train/accuracy'].update_state(labels, logits)
strategy.run(step_fn, args=(next(iterator), ))
@tf.function
def test_step(iterator, dataset_name):
"""Evaluation StepFn."""
def step_fn(inputs):
"""Per-Replica StepFn."""
features, labels = create_feature_and_label(
inputs, feature_size, model_family=FLAGS.model_family)
# Set learning phase to disable dropout etc during eval.
logits = model(features, training=False)
if FLAGS.use_bfloat16:
logits = tf.cast(logits, tf.float32)
probs = tf.nn.softmax(logits)
negative_log_likelihood = tf.reduce_mean(
tf.keras.losses.sparse_categorical_crossentropy(labels, probs))
if dataset_name == 'clean':
metrics['test/negative_log_likelihood'].update_state(
negative_log_likelihood)
metrics['test/accuracy'].update_state(labels, probs)
metrics['test/ece'].update_state(labels, probs)
else:
metrics['test/nll_{}'.format(dataset_name)].update_state(
negative_log_likelihood)
metrics['test/accuracy_{}'.format(dataset_name)].update_state(
labels, probs)
metrics['test/ece_{}'.format(dataset_name)].update_state(
labels, probs)
strategy.run(step_fn, args=(next(iterator), ))
train_iterator = iter(train_dataset)
start_time = time.time()
for epoch in range(initial_epoch, FLAGS.train_epochs):
logging.info('Starting to run epoch: %s', epoch)
for step in range(steps_per_epoch):
train_step(train_iterator)
current_step = epoch * steps_per_epoch + (step + 1)
max_steps = steps_per_epoch * FLAGS.train_epochs
time_elapsed = time.time() - start_time
steps_per_sec = float(current_step) / time_elapsed
eta_seconds = (max_steps - current_step) / steps_per_sec
message = ('{:.1%} completion: epoch {:d}/{:d}. {:.1f} steps/s. '
'ETA: {:.0f} min. Time elapsed: {:.0f} min'.format(
current_step / max_steps, epoch + 1,
FLAGS.train_epochs, steps_per_sec, eta_seconds / 60,
time_elapsed / 60))
if step % 20 == 0:
logging.info(message)
if epoch % FLAGS.evaluation_interval == 0:
for dataset_name, test_dataset in test_datasets.items():
test_iterator = iter(test_dataset)
logging.info('Testing on dataset %s', dataset_name)
for step in range(steps_per_eval[dataset_name]):
if step % 20 == 0:
logging.info(
'Starting to run eval step %s of epoch: %s', step,
epoch)
test_step(test_iterator, dataset_name)
logging.info('Done with testing on %s', dataset_name)
logging.info('Train Loss: %.4f, Accuracy: %.2f%%',
metrics['train/loss'].result(),
metrics['train/accuracy'].result() * 100)
logging.info('Test NLL: %.4f, Accuracy: %.2f%%',
metrics['test/negative_log_likelihood'].result(),
metrics['test/accuracy'].result() * 100)
total_results = {
name: metric.result()
for name, metric in metrics.items()
}
with summary_writer.as_default():
for name, result in total_results.items():
tf.summary.scalar(name, result, step=epoch + 1)
for metric in metrics.values():
metric.reset_states()
if (FLAGS.checkpoint_interval > 0
and (epoch + 1) % FLAGS.checkpoint_interval == 0):
checkpoint_name = checkpoint.save(
os.path.join(FLAGS.output_dir, 'checkpoint'))
logging.info('Saved checkpoint to %s', checkpoint_name)
def main(argv):
del argv # unused arg
if not FLAGS.use_gpu:
raise ValueError('Only GPU is currently supported.')
if FLAGS.num_cores > 1:
raise ValueError('Only a single accelerator is currently supported.')
tf.enable_v2_behavior()
tf.random.set_seed(FLAGS.seed)
tf.io.gfile.makedirs(FLAGS.output_dir)
ind_dataset_builder = ub.datasets.ClincIntentDetectionDataset(
batch_size=FLAGS.per_core_batch_size,
eval_batch_size=FLAGS.per_core_batch_size,
dataset_dir=FLAGS.dataset_dir,
data_mode='ind')
ood_dataset_builder = ub.datasets.ClincIntentDetectionDataset(
batch_size=FLAGS.per_core_batch_size,
eval_batch_size=FLAGS.per_core_batch_size,
dataset_dir=FLAGS.dataset_dir,
data_mode='ood')
dataset_builders = {
'clean': ind_dataset_builder,
'ood': ood_dataset_builder
}
ds_info = ind_dataset_builder.info
feature_size = ds_info['feature_size']
# num_classes is number of valid intents plus out-of-scope intent
num_classes = ds_info['num_classes'] + 1
# vocab_size is total number of valid tokens plus the out-of-vocabulary token.
vocab_size = ind_dataset_builder.tokenizer.num_words + 1
batch_size = FLAGS.per_core_batch_size * FLAGS.num_cores
test_datasets = {}
steps_per_eval = {}
for dataset_name, dataset_builder in dataset_builders.items():
test_datasets[dataset_name] = dataset_builder.build(
split=ub.datasets.base.Split.TEST)
steps_per_eval[dataset_name] = (
dataset_builder.info['num_test_examples'] // batch_size)
if FLAGS.model_family.lower() == 'textcnn':
model = cnn_model.textcnn(filter_sizes=FLAGS.filter_sizes,
num_filters=FLAGS.num_filters,
num_classes=num_classes,
feature_size=feature_size,
vocab_size=vocab_size,
embed_size=FLAGS.embedding_size,
dropout_rate=FLAGS.dropout_rate,
l2=FLAGS.l2)
elif FLAGS.model_family.lower() == 'bert':
bert_config_dir, _ = deterministic.resolve_bert_ckpt_and_config_dir(
FLAGS.bert_dir, FLAGS.bert_config_dir, FLAGS.bert_ckpt_dir)
bert_config = bert_model.create_config(bert_config_dir)
model, _ = bert_model.create_model(num_classes=num_classes,
feature_size=feature_size,
bert_config=bert_config)
else:
raise ValueError(
'model_family ({}) can only be TextCNN or BERT.'.format(
FLAGS.model_family))
logging.info('Model input shape: %s', model.input_shape)
logging.info('Model output shape: %s', model.output_shape)
logging.info('Model number of weights: %s', model.count_params())
# Search for checkpoints from their index file; then remove the index suffix.
ensemble_filenames = tf.io.gfile.glob(
os.path.join(FLAGS.checkpoint_dir, '**/*.index'))
ensemble_filenames = [filename[:-6] for filename in ensemble_filenames]
ensemble_size = len(ensemble_filenames)
logging.info('Ensemble size: %s', ensemble_size)
logging.info('Ensemble number of weights: %s',
ensemble_size * model.count_params())
logging.info('Ensemble filenames: %s', str(ensemble_filenames))
checkpoint = tf.train.Checkpoint(model=model)
# Write model predictions to files.
num_datasets = len(test_datasets)
for m, ensemble_filename in enumerate(ensemble_filenames):
checkpoint.restore(ensemble_filename)
for n, (name, test_dataset) in enumerate(test_datasets.items()):
filename = '{dataset}_{member}.npy'.format(dataset=name, member=m)
filename = os.path.join(FLAGS.output_dir, filename)
if not tf.io.gfile.exists(filename):
logits = []
test_iterator = iter(test_dataset)
for _ in range(steps_per_eval[name]):
inputs = next(test_iterator)
features, _ = deterministic.create_feature_and_label(
inputs, feature_size, model_family=FLAGS.model_family)
logits.append(model(features, training=False))
logits = tf.concat(logits, axis=0)
with tf.io.gfile.GFile(filename, 'w') as f:
np.save(f, logits.numpy())
percent = (m * num_datasets +
(n + 1)) / (ensemble_size * num_datasets)
message = (
'{:.1%} completion for prediction: ensemble member {:d}/{:d}. '
'Dataset {:d}/{:d}'.format(percent, m + 1, ensemble_size,
n + 1, num_datasets))
logging.info(message)
metrics = {
'test/negative_log_likelihood': tf.keras.metrics.Mean(),
'test/gibbs_cross_entropy': tf.keras.metrics.Mean(),
'test/accuracy': tf.keras.metrics.SparseCategoricalAccuracy(),
'test/ece':
ed.metrics.ExpectedCalibrationError(num_bins=FLAGS.num_bins),
}
for dataset_name, test_dataset in test_datasets.items():
if dataset_name != 'clean':
metrics.update({
'test/nll_{}'.format(dataset_name):
tf.keras.metrics.Mean(),
'test/accuracy_{}'.format(dataset_name):
tf.keras.metrics.SparseCategoricalAccuracy(),
'test/ece_{}'.format(dataset_name):
ed.metrics.ExpectedCalibrationError(num_bins=FLAGS.num_bins)
})
# Evaluate model predictions.
for n, (name, test_dataset) in enumerate(test_datasets.items()):
logits_dataset = []
for m in range(ensemble_size):
filename = '{dataset}_{member}.npy'.format(dataset=name, member=m)
filename = os.path.join(FLAGS.output_dir, filename)
with tf.io.gfile.GFile(filename, 'rb') as f:
logits_dataset.append(np.load(f))
logits_dataset = tf.convert_to_tensor(logits_dataset)
test_iterator = iter(test_dataset)
for step in range(steps_per_eval[name]):
inputs = next(test_iterator)
_, labels = deterministic.create_feature_and_label(
inputs, feature_size, model_family=FLAGS.model_family)
logits = logits_dataset[:, (step * batch_size):((step + 1) *
batch_size)]
labels = tf.cast(labels, tf.int32)
negative_log_likelihood = tf.reduce_mean(
ensemble_negative_log_likelihood(labels, logits))
per_probs = tf.nn.softmax(logits)
probs = tf.reduce_mean(per_probs, axis=0)
if name == 'clean':
gibbs_ce = tf.reduce_mean(gibbs_cross_entropy(labels, logits))
metrics['test/negative_log_likelihood'].update_state(
negative_log_likelihood)
metrics['test/gibbs_cross_entropy'].update_state(gibbs_ce)
metrics['test/accuracy'].update_state(labels, probs)
metrics['test/ece'].update_state(labels, probs)
else:
metrics['test/nll_{}'.format(name)].update_state(
negative_log_likelihood)
metrics['test/accuracy_{}'.format(name)].update_state(
labels, probs)
metrics['test/ece_{}'.format(name)].update_state(labels, probs)
message = (
'{:.1%} completion for evaluation: dataset {:d}/{:d}'.format(
(n + 1) / num_datasets, n + 1, num_datasets))
logging.info(message)
total_results = {name: metric.result() for name, metric in metrics.items()}
logging.info('Metrics: %s', total_results)