def test_bert_trainer(self):
"""Validate that the Keras object can be created."""
# Build a transformer network to use within the BERT trainer.
vocab_size = 100
sequence_length = 512
test_network = networks.TransformerEncoder(
vocab_size=vocab_size,
num_layers=2,
sequence_length=sequence_length)
# Create a BERT trainer with the created network.
num_classes = 3
bert_trainer_model = bert_token_classifier.BertTokenClassifier(
test_network, num_classes=num_classes)
# Create a set of 2-dimensional inputs (the first dimension is implicit).
word_ids = tf.keras.Input(shape=(sequence_length, ), dtype=tf.int32)
mask = tf.keras.Input(shape=(sequence_length, ), dtype=tf.int32)
type_ids = tf.keras.Input(shape=(sequence_length, ), dtype=tf.int32)
# Invoke the trainer model on the inputs. This causes the layer to be built.
sequence_outs = bert_trainer_model([word_ids, mask, type_ids])
# Validate that the outputs are of the expected shape.
expected_classification_shape = [None, sequence_length, num_classes]
self.assertAllEqual(expected_classification_shape,
sequence_outs.shape.as_list())
def test_serialize_deserialize(self):
"""Validate that the BERT trainer can be serialized and deserialized."""
# Build a transformer network to use within the BERT trainer. (Here, we use
# a short sequence_length for convenience.)
test_network = networks.TransformerEncoder(vocab_size=100,
num_layers=2,
sequence_length=5)
# Create a BERT trainer with the created network. (Note that all the args
# are different, so we can catch any serialization mismatches.)
bert_trainer_model = bert_token_classifier.BertTokenClassifier(
test_network,
num_classes=4,
initializer='zeros',
output='predictions')
# Create another BERT trainer via serialization and deserialization.
config = bert_trainer_model.get_config()
new_bert_trainer_model = (
bert_token_classifier.BertTokenClassifier.from_config(config))
# Validate that the config can be forced to JSON.
_ = new_bert_trainer_model.to_json()
# If the serialization was successful, the new config should match the old.
self.assertAllEqual(bert_trainer_model.get_config(),
new_bert_trainer_model.get_config())
def test_bert_trainer_tensor_call(self):
"""Validate that the Keras object can be invoked."""
# Build a transformer network to use within the BERT trainer. (Here, we use
# a short sequence_length for convenience.)
test_network = networks.BertEncoder(
vocab_size=100, num_layers=2, max_sequence_length=2)
# Create a BERT trainer with the created network.
bert_trainer_model = bert_token_classifier.BertTokenClassifier(
test_network, num_classes=2)
# Create a set of 2-dimensional data tensors to feed into the model.
word_ids = tf.constant([[1, 1], [2, 2]], dtype=tf.int32)
mask = tf.constant([[1, 1], [1, 0]], dtype=tf.int32)
type_ids = tf.constant([[1, 1], [2, 2]], dtype=tf.int32)
# Invoke the trainer model on the tensors. In Eager mode, this does the
# actual calculation. (We can't validate the outputs, since the network is
# too complex: this simply ensures we're not hitting runtime errors.)
_ = bert_trainer_model([word_ids, mask, type_ids])
