def instantiate_from_cfg(config: BertPretrainerConfig,
encoder_network: Optional[tf.keras.Model] = None):
"""Instantiates a BertPretrainer from the config."""
encoder_cfg = config.encoder
if encoder_network is None:
encoder_network = networks.TransformerEncoder(
vocab_size=encoder_cfg.vocab_size,
hidden_size=encoder_cfg.hidden_size,
num_layers=encoder_cfg.num_layers,
num_attention_heads=encoder_cfg.num_attention_heads,
intermediate_size=encoder_cfg.intermediate_size,
activation=tf_utils.get_activation(encoder_cfg.hidden_activation),
dropout_rate=encoder_cfg.dropout_rate,
attention_dropout_rate=encoder_cfg.attention_dropout_rate,
max_sequence_length=encoder_cfg.max_position_embeddings,
type_vocab_size=encoder_cfg.type_vocab_size,
initializer=tf.keras.initializers.TruncatedNormal(
stddev=encoder_cfg.initializer_range))
if config.cls_heads:
classification_heads = [
layers.ClassificationHead(**cfg.as_dict())
for cfg in config.cls_heads
]
else:
classification_heads = []
return bert_pretrainer.BertPretrainerV2(
config.num_masked_tokens,
mlm_initializer=tf.keras.initializers.TruncatedNormal(
stddev=encoder_cfg.initializer_range),
encoder_network=encoder_network,
classification_heads=classification_heads)
def test_bert_pretrainerv2(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.
bert_trainer_model = bert_pretrainer.BertPretrainerV2(
encoder_network=test_network)
num_token_predictions = 20
# 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)
lm_mask = tf.keras.Input(shape=(num_token_predictions, ),
dtype=tf.int32)
# Invoke the trainer model on the inputs. This causes the layer to be built.
outputs = bert_trainer_model([word_ids, mask, type_ids, lm_mask])
# Validate that the outputs are of the expected shape.
expected_lm_shape = [None, num_token_predictions, vocab_size]
self.assertAllEqual(expected_lm_shape,
outputs['lm_output'].shape.as_list())
def test_multiple_cls_outputs(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
hidden_size = 48
num_layers = 2
test_network = networks.BertEncoder(
vocab_size=vocab_size,
num_layers=num_layers,
hidden_size=hidden_size,
max_sequence_length=sequence_length,
dict_outputs=True)
bert_trainer_model = bert_pretrainer.BertPretrainerV2(
encoder_network=test_network,
classification_heads=[layers.MultiClsHeads(
inner_dim=5, cls_list=[('foo', 2), ('bar', 3)])])
num_token_predictions = 20
# Create a set of 2-dimensional inputs (the first dimension is implicit).
inputs = dict(
input_word_ids=tf.keras.Input(shape=(sequence_length,), dtype=tf.int32),
input_mask=tf.keras.Input(shape=(sequence_length,), dtype=tf.int32),
input_type_ids=tf.keras.Input(shape=(sequence_length,), dtype=tf.int32),
masked_lm_positions=tf.keras.Input(
shape=(num_token_predictions,), dtype=tf.int32))
# Invoke the trainer model on the inputs. This causes the layer to be built.
outputs = bert_trainer_model(inputs)
self.assertEqual(outputs['foo'].shape.as_list(), [None, 2])
self.assertEqual(outputs['bar'].shape.as_list(), [None, 3])
def test_bert_pretrainerv2(self, dict_outputs, return_all_encoder_outputs,
use_customized_masked_lm):
"""Validate that the Keras object can be created."""
# Build a transformer network to use within the BERT trainer.
vocab_size = 100
sequence_length = 512
hidden_size = 48
num_layers = 2
test_network = networks.BertEncoder(
vocab_size=vocab_size,
num_layers=num_layers,
hidden_size=hidden_size,
max_sequence_length=sequence_length,
return_all_encoder_outputs=return_all_encoder_outputs,
dict_outputs=dict_outputs)
# Create a BERT trainer with the created network.
if use_customized_masked_lm:
customized_masked_lm = layers.MaskedLM(
embedding_table=test_network.get_embedding_table())
else:
customized_masked_lm = None
bert_trainer_model = bert_pretrainer.BertPretrainerV2(
encoder_network=test_network, customized_masked_lm=customized_masked_lm)
num_token_predictions = 20
# Create a set of 2-dimensional inputs (the first dimension is implicit).
inputs = dict(
input_word_ids=tf.keras.Input(shape=(sequence_length,), dtype=tf.int32),
input_mask=tf.keras.Input(shape=(sequence_length,), dtype=tf.int32),
input_type_ids=tf.keras.Input(shape=(sequence_length,), dtype=tf.int32),
masked_lm_positions=tf.keras.Input(
shape=(num_token_predictions,), dtype=tf.int32))
# Invoke the trainer model on the inputs. This causes the layer to be built.
outputs = bert_trainer_model(inputs)
has_encoder_outputs = dict_outputs or return_all_encoder_outputs
if has_encoder_outputs:
self.assertSameElements(
outputs.keys(),
['sequence_output', 'pooled_output', 'mlm_logits', 'encoder_outputs'])
self.assertLen(outputs['encoder_outputs'], num_layers)
else:
self.assertSameElements(
outputs.keys(), ['sequence_output', 'pooled_output', 'mlm_logits'])
# Validate that the outputs are of the expected shape.
expected_lm_shape = [None, num_token_predictions, vocab_size]
self.assertAllEqual(expected_lm_shape,
outputs['mlm_logits'].shape.as_list())
expected_sequence_output_shape = [None, sequence_length, hidden_size]
self.assertAllEqual(expected_sequence_output_shape,
outputs['sequence_output'].shape.as_list())
expected_pooled_output_shape = [None, hidden_size]
self.assertAllEqual(expected_pooled_output_shape,
outputs['pooled_output'].shape.as_list())
def instantiate_pretrainer_from_cfg(
config: BertPretrainerConfig,
encoder_network: Optional[tf.keras.Model] = None
) -> bert_pretrainer.BertPretrainerV2:
"""Instantiates a BertPretrainer from the config."""
encoder_cfg = config.encoder
if encoder_network is None:
encoder_network = encoders.instantiate_encoder_from_cfg(encoder_cfg)
return bert_pretrainer.BertPretrainerV2(
mlm_activation=tf_utils.get_activation(encoder_cfg.hidden_activation),
mlm_initializer=tf.keras.initializers.TruncatedNormal(
stddev=encoder_cfg.initializer_range),
encoder_network=encoder_network,
classification_heads=instantiate_classification_heads_from_cfgs(
config.cls_heads))
def test_v2_serialize_deserialize(self):
"""Validate that the BERT trainer can be serialized and deserialized."""
# Build a transformer network to use within the BERT trainer.
test_network = networks.BertEncoderV2(vocab_size=100, num_layers=2)
# 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_pretrainer.BertPretrainerV2(
encoder_network=test_network)
# Create another BERT trainer via serialization and deserialization.
config = bert_trainer_model.get_config()
new_bert_trainer_model = bert_pretrainer.BertPretrainerV2.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())
