def test_serialize_deserialize(self):
"""Validate that the ELECTRA 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_generator_network = networks.TransformerEncoder(
vocab_size=100, num_layers=4, sequence_length=3)
test_discriminator_network = networks.TransformerEncoder(
vocab_size=100, num_layers=4, sequence_length=3)
# Create a ELECTRA trainer with the created network. (Note that all the args
# are different, so we can catch any serialization mismatches.)
electra_trainer_model = electra_pretrainer.ElectraPretrainer(
generator_network=test_generator_network,
discriminator_network=test_discriminator_network,
vocab_size=100,
num_classes=2,
sequence_length=3,
num_token_predictions=2)
# Create another BERT trainer via serialization and deserialization.
config = electra_trainer_model.get_config()
new_electra_trainer_model = electra_pretrainer.ElectraPretrainer.from_config(
config)
# Validate that the config can be forced to JSON.
_ = new_electra_trainer_model.to_json()
# If the serialization was successful, the new config should match the old.
self.assertAllEqual(electra_trainer_model.get_config(),
new_electra_trainer_model.get_config())
def test_electra_pretrainer(self):
"""Validate that the Keras object can be created."""
# Build a transformer network to use within the ELECTRA trainer.
vocab_size = 100
sequence_length = 512
test_generator_network = networks.TransformerEncoder(
vocab_size=vocab_size,
num_layers=2,
max_sequence_length=sequence_length)
test_discriminator_network = networks.TransformerEncoder(
vocab_size=vocab_size,
num_layers=2,
max_sequence_length=sequence_length)
# Create a ELECTRA trainer with the created network.
num_classes = 3
num_token_predictions = 2
eletrca_trainer_model = electra_pretrainer.ElectraPretrainer(
generator_network=test_generator_network,
discriminator_network=test_discriminator_network,
vocab_size=vocab_size,
num_classes=num_classes,
num_token_predictions=num_token_predictions,
disallow_correct=True)
# 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_positions = tf.keras.Input(shape=(num_token_predictions, ),
dtype=tf.int32)
lm_ids = tf.keras.Input(shape=(num_token_predictions, ),
dtype=tf.int32)
inputs = {
'input_word_ids': word_ids,
'input_mask': mask,
'input_type_ids': type_ids,
'masked_lm_positions': lm_positions,
'masked_lm_ids': lm_ids
}
# Invoke the trainer model on the inputs. This causes the layer to be built.
outputs = eletrca_trainer_model(inputs)
lm_outs = outputs['lm_outputs']
cls_outs = outputs['sentence_outputs']
disc_logits = outputs['disc_logits']
disc_label = outputs['disc_label']
# Validate that the outputs are of the expected shape.
expected_lm_shape = [None, num_token_predictions, vocab_size]
expected_classification_shape = [None, num_classes]
expected_disc_logits_shape = [None, sequence_length]
expected_disc_label_shape = [None, sequence_length]
self.assertAllEqual(expected_lm_shape, lm_outs.shape.as_list())
self.assertAllEqual(expected_classification_shape,
cls_outs.shape.as_list())
self.assertAllEqual(expected_disc_logits_shape,
disc_logits.shape.as_list())
self.assertAllEqual(expected_disc_label_shape,
disc_label.shape.as_list())
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 create_lm_model(self,
vocab_size,
sequence_length,
hidden_size,
num_predictions,
output="predictions"):
# First, create a transformer stack that we can use to get the LM's
# vocabulary weight.
xformer_stack = networks.TransformerEncoder(
vocab_size=vocab_size,
num_layers=1,
sequence_length=sequence_length,
hidden_size=hidden_size,
num_attention_heads=4,
)
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)
_ = xformer_stack([word_ids, mask, type_ids])
# Create a maskedLM from the transformer stack.
test_layer = layers.MaskedLM(
embedding_table=xformer_stack.get_embedding_table(), output=output)
# Create a model from the masked LM layer.
lm_input_tensor = tf.keras.Input(shape=(sequence_length, hidden_size))
masked_lm_positions = tf.keras.Input(shape=(num_predictions, ),
dtype=tf.int32)
output = test_layer(lm_input_tensor,
masked_positions=masked_lm_positions)
return tf.keras.Model([lm_input_tensor, masked_lm_positions], output)
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.
bert_trainer_model = bert_span_labeler.BertSpanLabeler(test_network)
# 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.
cls_outs = bert_trainer_model([word_ids, mask, type_ids])
# Validate that there are 2 outputs are of the expected shape.
self.assertEqual(2, len(cls_outs))
expected_shape = [None, sequence_length]
for out in cls_outs:
self.assertAllEqual(expected_shape, out.shape.as_list())
def _create_bert_model(cfg):
"""Creates a BERT keras core model from BERT configuration.
Args:
cfg: A `BertConfig` to create the core model.
Returns:
A TransformerEncoder netowork.
"""
bert_encoder = networks.TransformerEncoder(
vocab_size=cfg.vocab_size,
hidden_size=cfg.hidden_size,
num_layers=cfg.num_hidden_layers,
num_attention_heads=cfg.num_attention_heads,
intermediate_size=cfg.intermediate_size,
activation=activations.gelu,
dropout_rate=cfg.hidden_dropout_prob,
attention_dropout_rate=cfg.attention_probs_dropout_prob,
<<<<<<< HEAD
sequence_length=cfg.max_position_embeddings,
=======
max_sequence_length=cfg.max_position_embeddings,
>>>>>>> a811a3b7e640722318ad868c99feddf3f3063e36
type_vocab_size=cfg.type_vocab_size,
initializer=tf.keras.initializers.TruncatedNormal(
stddev=cfg.initializer_range),
embedding_width=cfg.embedding_size)
return bert_encoder
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 get_transformer_encoder(bert_config,
sequence_length):
"""Gets a 'TransformerEncoder' object.
Args:
bert_config: A 'modeling.BertConfig' or 'modeling.AlbertConfig' object.
sequence_length: Maximum sequence length of the training data.
Returns:
A networks.TransformerEncoder object.
"""
kwargs = dict(
vocab_size=bert_config.vocab_size,
hidden_size=bert_config.hidden_size,
num_layers=bert_config.num_hidden_layers,
num_attention_heads=bert_config.num_attention_heads,
intermediate_size=bert_config.intermediate_size,
activation=tf_utils.get_activation(bert_config.hidden_act),
dropout_rate=bert_config.hidden_dropout_prob,
attention_dropout_rate=bert_config.attention_probs_dropout_prob,
sequence_length=sequence_length,
max_sequence_length=bert_config.max_position_embeddings,
type_vocab_size=bert_config.type_vocab_size,
initializer=tf.keras.initializers.TruncatedNormal(
stddev=bert_config.initializer_range))
if isinstance(bert_config, albert_configs.AlbertConfig):
kwargs['embedding_width'] = bert_config.embedding_size
return networks.AlbertTransformerEncoder(**kwargs)
else:
assert isinstance(bert_config, configs.BertConfig)
return networks.TransformerEncoder(**kwargs)
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_dual_encoder_tensor_call(self, hidden_size, output):
"""Validate that the Keras object can be invoked."""
# Build a transformer network to use within the dual encoder model. (Here,
# we use # a short sequence_length for convenience.)
sequence_length = 2
test_network = networks.TransformerEncoder(
vocab_size=100, num_layers=2, sequence_length=sequence_length)
# Create a dual encoder model with the created network.
dual_encoder_model = dual_encoder.DualEncoder(
test_network, max_seq_length=sequence_length, output=output)
# 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 model 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.)
if output == 'logits':
_ = dual_encoder_model(
[word_ids, mask, type_ids, word_ids, mask, type_ids])
elif output == 'predictions':
_ = dual_encoder_model([word_ids, mask, type_ids])
def _get_transformer_encoder(bert_config,
sequence_length,
float_dtype=tf.float32):
"""Gets a 'TransformerEncoder' object.
Args:
bert_config: A 'modeling.BertConfig' object.
sequence_length: Maximum sequence length of the training data.
float_dtype: tf.dtype, tf.float32 or tf.float16.
Returns:
A networks.TransformerEncoder object.
"""
return networks.TransformerEncoder(
vocab_size=bert_config.vocab_size,
hidden_size=bert_config.hidden_size,
num_layers=bert_config.num_hidden_layers,
num_attention_heads=bert_config.num_attention_heads,
intermediate_size=bert_config.intermediate_size,
activation=tf_utils.get_activation(bert_config.hidden_act),
dropout_rate=bert_config.hidden_dropout_prob,
attention_dropout_rate=bert_config.attention_probs_dropout_prob,
sequence_length=sequence_length,
max_sequence_length=bert_config.max_position_embeddings,
type_vocab_size=bert_config.type_vocab_size,
initializer=tf.keras.initializers.TruncatedNormal(
stddev=bert_config.initializer_range),
float_dtype=float_dtype.name)
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 get_transformer_encoder(bert_config,
sequence_length,
transformer_encoder_cls=None):
"""Gets a 'TransformerEncoder' object.
Args:
bert_config: A 'modeling.BertConfig' or 'modeling.AlbertConfig' object.
sequence_length: Maximum sequence length of the training data.
transformer_encoder_cls: A EncoderScaffold class. If it is None, uses the
default BERT encoder implementation.
Returns:
A networks.TransformerEncoder object.
"""
if transformer_encoder_cls is not None:
# TODO(hongkuny): evaluate if it is better to put cfg definition in gin.
embedding_cfg = dict(
vocab_size=bert_config.vocab_size,
type_vocab_size=bert_config.type_vocab_size,
hidden_size=bert_config.hidden_size,
seq_length=sequence_length,
max_seq_length=bert_config.max_position_embeddings,
initializer=tf.keras.initializers.TruncatedNormal(
stddev=bert_config.initializer_range),
dropout_rate=bert_config.hidden_dropout_prob,
)
hidden_cfg = dict(
num_attention_heads=bert_config.num_attention_heads,
intermediate_size=bert_config.intermediate_size,
intermediate_activation=tf_utils.get_activation(bert_config.hidden_act),
dropout_rate=bert_config.hidden_dropout_prob,
attention_dropout_rate=bert_config.attention_probs_dropout_prob,
)
kwargs = dict(embedding_cfg=embedding_cfg, hidden_cfg=hidden_cfg,
num_hidden_instances=bert_config.num_hidden_layers,)
# Relies on gin configuration to define the Transformer encoder arguments.
return transformer_encoder_cls(**kwargs)
kwargs = dict(
vocab_size=bert_config.vocab_size,
hidden_size=bert_config.hidden_size,
num_layers=bert_config.num_hidden_layers,
num_attention_heads=bert_config.num_attention_heads,
intermediate_size=bert_config.intermediate_size,
activation=tf_utils.get_activation(bert_config.hidden_act),
dropout_rate=bert_config.hidden_dropout_prob,
attention_dropout_rate=bert_config.attention_probs_dropout_prob,
sequence_length=sequence_length,
max_sequence_length=bert_config.max_position_embeddings,
type_vocab_size=bert_config.type_vocab_size,
initializer=tf.keras.initializers.TruncatedNormal(
stddev=bert_config.initializer_range))
if isinstance(bert_config, albert_configs.AlbertConfig):
kwargs['embedding_width'] = bert_config.embedding_size
return networks.AlbertTransformerEncoder(**kwargs)
else:
assert isinstance(bert_config, configs.BertConfig)
return networks.TransformerEncoder(**kwargs)
def get_nhnet_layers(params: configs.NHNetConfig):
"""Creates a Mult-doc encoder/decoder.
Args:
params: ParamsDict.
Returns:
two keras Layers, bert_model_layer and decoder_layer
"""
input_ids = tf.keras.layers.Input(
shape=(None,), name="input_ids", dtype=tf.int32)
input_mask = tf.keras.layers.Input(
shape=(None,), name="input_mask", dtype=tf.int32)
segment_ids = tf.keras.layers.Input(
shape=(None,), name="segment_ids", dtype=tf.int32)
bert_config = utils.get_bert_config_from_params(params)
bert_model_layer = networks.TransformerEncoder(
vocab_size=bert_config.vocab_size,
hidden_size=bert_config.hidden_size,
num_layers=bert_config.num_hidden_layers,
num_attention_heads=bert_config.num_attention_heads,
intermediate_size=bert_config.intermediate_size,
activation=tf_utils.get_activation(bert_config.hidden_act),
dropout_rate=bert_config.hidden_dropout_prob,
attention_dropout_rate=bert_config.attention_probs_dropout_prob,
sequence_length=None,
max_sequence_length=bert_config.max_position_embeddings,
type_vocab_size=bert_config.type_vocab_size,
initializer=tf.keras.initializers.TruncatedNormal(
stddev=bert_config.initializer_range),
return_all_encoder_outputs=True,
name="bert_encoder")
bert_model_layer([input_ids, input_mask, segment_ids])
input_ids = tf.keras.layers.Input(
shape=(None, None), name="input_ids", dtype=tf.int32)
all_encoder_outputs = tf.keras.layers.Input((None, None, params.hidden_size),
dtype=tf.float32)
target_ids = tf.keras.layers.Input(
shape=(None,), name="target_ids", dtype=tf.int32)
doc_attention_probs = tf.keras.layers.Input(
(params.num_decoder_attn_heads, None, None), dtype=tf.float32)
# pylint: disable=protected-access
decoder_layer = decoder.Decoder(params, bert_model_layer._embedding_layer)
# pylint: enable=protected-access
cross_attention_bias = decoder.AttentionBias(bias_type="multi_cross")(
input_ids)
self_attention_bias = decoder.AttentionBias(bias_type="decoder_self")(
target_ids)
decoder_inputs = dict(
attention_bias=cross_attention_bias,
self_attention_bias=self_attention_bias,
target_ids=target_ids,
all_encoder_outputs=all_encoder_outputs,
doc_attention_probs=doc_attention_probs)
_ = decoder_layer(decoder_inputs)
return bert_model_layer, decoder_layer
def get_bert2bert_layers(params: configs.BERT2BERTConfig):
"""Creates a Bert2Bert stem model and returns Bert encoder/decoder.
We use funtional-style to create stem model because we need to make all layers
built to restore variables in a customized way. The layers are called with
placeholder inputs to make them fully built.
Args:
params: ParamsDict.
Returns:
two keras Layers, bert_model_layer and decoder_layer
"""
input_ids = tf.keras.layers.Input(
shape=(None,), name="input_ids", dtype=tf.int32)
input_mask = tf.keras.layers.Input(
shape=(None,), name="input_mask", dtype=tf.int32)
segment_ids = tf.keras.layers.Input(
shape=(None,), name="segment_ids", dtype=tf.int32)
target_ids = tf.keras.layers.Input(
shape=(None,), name="target_ids", dtype=tf.int32)
bert_config = utils.get_bert_config_from_params(params)
bert_model_layer = networks.TransformerEncoder(
vocab_size=bert_config.vocab_size,
hidden_size=bert_config.hidden_size,
num_layers=bert_config.num_hidden_layers,
num_attention_heads=bert_config.num_attention_heads,
intermediate_size=bert_config.intermediate_size,
activation=tf_utils.get_activation(bert_config.hidden_act),
dropout_rate=bert_config.hidden_dropout_prob,
attention_dropout_rate=bert_config.attention_probs_dropout_prob,
<<<<<<< HEAD
sequence_length=None,
=======
>>>>>>> a811a3b7e640722318ad868c99feddf3f3063e36
max_sequence_length=bert_config.max_position_embeddings,
type_vocab_size=bert_config.type_vocab_size,
initializer=tf.keras.initializers.TruncatedNormal(
stddev=bert_config.initializer_range),
return_all_encoder_outputs=True,
name="bert_encoder")
all_encoder_outputs, _ = bert_model_layer(
[input_ids, input_mask, segment_ids])
# pylint: disable=protected-access
decoder_layer = decoder.Decoder(params, bert_model_layer._embedding_layer)
# pylint: enable=protected-access
cross_attention_bias = decoder.AttentionBias(bias_type="single_cross")(
input_ids)
self_attention_bias = decoder.AttentionBias(bias_type="decoder_self")(
target_ids)
decoder_inputs = dict(
attention_bias=cross_attention_bias,
self_attention_bias=self_attention_bias,
target_ids=target_ids,
all_encoder_outputs=all_encoder_outputs)
_ = decoder_layer(decoder_inputs)
return bert_model_layer, decoder_layer
def __init__(self,
context_feature_columns,
example_feature_columns,
bert_config_file,
bert_max_seq_length,
bert_output_dropout,
name="tfrbert",
**kwargs):
"""Initializes an instance of TFRBertRankingNetwork.
Args:
context_feature_columns: A dict containing all the context feature columns
used by the network. Keys are feature names, and values are instances of
classes derived from `_FeatureColumn`.
example_feature_columns: A dict containing all the example feature columns
used by the network. Keys are feature names, and values are instances of
classes derived from `_FeatureColumn`.
bert_config_file: (string) path to Bert configuration file.
bert_max_seq_length: (int) maximum input sequence length (#words) after
WordPiece tokenization. Sequences longer than this will be truncated,
and shorter than this will be padded.
bert_output_dropout: When not `None`, the probability will be used as the
dropout probability for BERT output.
name: name of Keras network.
**kwargs: keyword arguments.
"""
super(TFRBertRankingNetwork,
self).__init__(context_feature_columns=context_feature_columns,
example_feature_columns=example_feature_columns,
name=name,
**kwargs)
self._bert_config_file = bert_config_file
self._bert_max_seq_length = bert_max_seq_length
self._bert_output_dropout = bert_output_dropout
bert_config = configs.BertConfig.from_json_file(self._bert_config_file)
self._bert_encoder = tfmodel_networks.TransformerEncoder(
vocab_size=bert_config.vocab_size,
hidden_size=bert_config.hidden_size,
num_layers=bert_config.num_hidden_layers,
num_attention_heads=bert_config.num_attention_heads,
intermediate_size=bert_config.intermediate_size,
activation=activations.gelu,
dropout_rate=bert_config.hidden_dropout_prob,
attention_dropout_rate=bert_config.attention_probs_dropout_prob,
sequence_length=self._bert_max_seq_length,
max_sequence_length=bert_config.max_position_embeddings,
type_vocab_size=bert_config.type_vocab_size,
initializer=tf.keras.initializers.TruncatedNormal(
stddev=bert_config.initializer_range))
self._dropout_layer = tf.keras.layers.Dropout(
rate=self._bert_output_dropout)
self._score_layer = tf.keras.layers.Dense(units=1, name="score")
def test_electra_pretrainer(self):
"""Validate that the Keras object can be created."""
# Build a transformer network to use within the ELECTRA trainer.
vocab_size = 100
sequence_length = 512
test_generator_network = networks.TransformerEncoder(
<<<<<<< HEAD
vocab_size=vocab_size, num_layers=2, sequence_length=sequence_length)
test_discriminator_network = networks.TransformerEncoder(
vocab_size=vocab_size, num_layers=2, sequence_length=sequence_length)
def test_electra_trainer_tensor_call(self):
"""Validate that the Keras object can be invoked."""
# Build a transformer network to use within the ELECTRA trainer. (Here, we
# use a short sequence_length for convenience.)
test_generator_network = networks.TransformerEncoder(vocab_size=100,
num_layers=4,
sequence_length=3)
test_discriminator_network = networks.TransformerEncoder(
vocab_size=100, num_layers=4, sequence_length=3)
# Create a ELECTRA trainer with the created network.
eletrca_trainer_model = electra_pretrainer.ElectraPretrainer(
generator_network=test_generator_network,
discriminator_network=test_discriminator_network,
vocab_size=100,
num_classes=2,
sequence_length=3,
last_hidden_dim=768,
num_token_predictions=2)
# Create a set of 2-dimensional data tensors to feed into the model.
word_ids = tf.constant([[1, 1, 1], [2, 2, 2]], dtype=tf.int32)
mask = tf.constant([[1, 1, 1], [1, 0, 0]], dtype=tf.int32)
type_ids = tf.constant([[1, 1, 1], [2, 2, 2]], dtype=tf.int32)
lm_positions = tf.constant([[0, 1], [0, 2]], dtype=tf.int32)
lm_ids = tf.constant([[10, 20], [20, 30]], dtype=tf.int32)
inputs = {
'input_word_ids': word_ids,
'input_mask': mask,
'input_type_ids': type_ids,
'masked_lm_positions': lm_positions,
'masked_lm_ids': lm_ids
}
# 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.)
_, _, _, _ = eletrca_trainer_model(inputs)
def instantiate_encoder_from_cfg(
config: TransformerEncoderConfig) -> networks.TransformerEncoder:
"""Instantiate a Transformer encoder network from TransformerEncoderConfig."""
encoder_network = networks.TransformerEncoder(
vocab_size=config.vocab_size,
hidden_size=config.hidden_size,
num_layers=config.num_layers,
num_attention_heads=config.num_attention_heads,
intermediate_size=config.intermediate_size,
activation=tf_utils.get_activation(config.hidden_activation),
dropout_rate=config.dropout_rate,
attention_dropout_rate=config.attention_dropout_rate,
max_sequence_length=config.max_position_embeddings,
type_vocab_size=config.type_vocab_size,
initializer=tf.keras.initializers.TruncatedNormal(
stddev=config.initializer_range))
return encoder_network
def test_bert_trainer_named_compilation(self):
"""Validate compilation using explicit output names."""
# Build a transformer network to use within the BERT trainer.
vocab_size = 100
test_network = networks.TransformerEncoder(vocab_size=vocab_size,
num_layers=2)
# Create a BERT trainer with the created network.
bert_trainer_model = bert_span_labeler.BertSpanLabeler(test_network)
# Attempt to compile the model using a string-keyed dict of output names to
# loss functions. This will validate that the outputs are named as we
# expect.
bert_trainer_model.compile(optimizer='sgd',
loss={
'start_positions': 'mse',
'end_positions': 'mse'
})
def test_dual_encoder(self, hidden_size, output):
"""Validate that the Keras object can be created."""
# Build a transformer network to use within the dual encoder model.
vocab_size = 100
sequence_length = 512
test_network = networks.TransformerEncoder(
vocab_size=vocab_size,
num_layers=2,
hidden_size=hidden_size,
sequence_length=sequence_length)
# Create a dual encoder model with the created network.
dual_encoder_model = dual_encoder.DualEncoder(
test_network, max_seq_length=sequence_length, output=output)
# Create a set of 2-dimensional inputs (the first dimension is implicit).
left_word_ids = tf.keras.Input(shape=(sequence_length, ),
dtype=tf.int32)
left_mask = tf.keras.Input(shape=(sequence_length, ), dtype=tf.int32)
left_type_ids = tf.keras.Input(shape=(sequence_length, ),
dtype=tf.int32)
right_word_ids = tf.keras.Input(shape=(sequence_length, ),
dtype=tf.int32)
right_mask = tf.keras.Input(shape=(sequence_length, ), dtype=tf.int32)
right_type_ids = tf.keras.Input(shape=(sequence_length, ),
dtype=tf.int32)
if output == 'logits':
outputs = dual_encoder_model([
left_word_ids, left_mask, left_type_ids, right_word_ids,
right_mask, right_type_ids
])
left_encoded, _ = outputs
elif output == 'predictions':
left_encoded = dual_encoder_model(
[left_word_ids, left_mask, left_type_ids])
# Validate that the outputs are of the expected shape.
expected_encoding_shape = [None, 768]
self.assertAllEqual(expected_encoding_shape,
left_encoded.shape.as_list())
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.TransformerEncoder(vocab_size=100,
num_layers=2)
# Create a BERT trainer with the created network.
bert_trainer_model = bert_span_labeler.BertSpanLabeler(test_network)
# 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])
def test_serialize_deserialize(self):
"""Validate that the dual encoder model can be serialized / deserialized."""
# Build a transformer network to use within the dual encoder model. (Here,
# we use a short sequence_length for convenience.)
sequence_length = 32
test_network = networks.TransformerEncoder(
vocab_size=100, num_layers=2, sequence_length=sequence_length)
# Create a dual encoder model with the created network. (Note that all the
# args are different, so we can catch any serialization mismatches.)
dual_encoder_model = dual_encoder.DualEncoder(
test_network, max_seq_length=sequence_length, output='predictions')
# Create another dual encoder model via serialization and deserialization.
config = dual_encoder_model.get_config()
new_dual_encoder = dual_encoder.DualEncoder.from_config(config)
# Validate that the config can be forced to JSON.
_ = new_dual_encoder.to_json()
# If the serialization was successful, the new config should match the old.
self.assertAllEqual(dual_encoder_model.get_config(),
new_dual_encoder.get_config())
def _create_bert_model(cfg):
"""Creates a BERT keras core model from BERT configuration.
Args:
cfg: A `BertConfig` to create the core model.
Returns:
A keras model.
"""
bert_encoder = networks.TransformerEncoder(
vocab_size=cfg.vocab_size,
hidden_size=cfg.hidden_size,
num_layers=cfg.num_hidden_layers,
num_attention_heads=cfg.num_attention_heads,
intermediate_size=cfg.intermediate_size,
activation=activations.gelu,
dropout_rate=cfg.hidden_dropout_prob,
attention_dropout_rate=cfg.attention_probs_dropout_prob,
sequence_length=cfg.max_position_embeddings,
type_vocab_size=cfg.type_vocab_size,
initializer=tf.keras.initializers.TruncatedNormal(
stddev=cfg.initializer_range))
return bert_encoder
attention_dropout_rate=bert_config.attention_probs_dropout_prob,
<<<<<<< HEAD
sequence_length=sequence_length,
=======
>>>>>>> a811a3b7e640722318ad868c99feddf3f3063e36
max_sequence_length=bert_config.max_position_embeddings,
type_vocab_size=bert_config.type_vocab_size,
embedding_width=bert_config.embedding_size,
initializer=tf.keras.initializers.TruncatedNormal(
stddev=bert_config.initializer_range))
if isinstance(bert_config, albert_configs.AlbertConfig):
return networks.AlbertTransformerEncoder(**kwargs)
else:
assert isinstance(bert_config, configs.BertConfig)
kwargs['output_range'] = output_range
return networks.TransformerEncoder(**kwargs)
def pretrain_model(bert_config,
seq_length,
max_predictions_per_seq,
initializer=None,
use_next_sentence_label=True,
return_core_pretrainer_model=False):
"""Returns model to be used for pre-training.
Args:
bert_config: Configuration that defines the core BERT model.
seq_length: Maximum sequence length of the training data.
max_predictions_per_seq: Maximum number of tokens in sequence to mask out
and use for pretraining.
def classifier_model(bert_config,
float_type,
num_labels,
max_seq_length,
final_layer_initializer=None,
hub_module_url=None):
"""BERT classifier model in functional API style.
Construct a Keras model for predicting `num_labels` outputs from an input with
maximum sequence length `max_seq_length`.
Args:
bert_config: BertConfig, the config defines the core BERT model.
float_type: dtype, tf.float32 or tf.bfloat16.
num_labels: integer, the number of classes.
max_seq_length: integer, the maximum input sequence length.
final_layer_initializer: Initializer for final dense layer. Defaulted
TruncatedNormal initializer.
hub_module_url: TF-Hub path/url to Bert module.
Returns:
Combined prediction model (words, mask, type) -> (one-hot labels)
BERT sub-model (words, mask, type) -> (bert_outputs)
"""
if final_layer_initializer is not None:
initializer = final_layer_initializer
else:
initializer = tf.keras.initializers.TruncatedNormal(
stddev=bert_config.initializer_range)
if not hub_module_url:
bert_encoder = networks.TransformerEncoder(
vocab_size=bert_config.vocab_size,
hidden_size=bert_config.hidden_size,
num_layers=bert_config.num_hidden_layers,
num_attention_heads=bert_config.num_attention_heads,
intermediate_size=bert_config.intermediate_size,
activation=tf_utils.get_activation('gelu'),
dropout_rate=bert_config.hidden_dropout_prob,
attention_dropout_rate=bert_config.attention_probs_dropout_prob,
sequence_length=max_seq_length,
max_sequence_length=bert_config.max_position_embeddings,
type_vocab_size=bert_config.type_vocab_size,
initializer=tf.keras.initializers.TruncatedNormal(
stddev=bert_config.initializer_range))
return bert_classifier.BertClassifier(
bert_encoder,
num_classes=num_labels,
dropout_rate=bert_config.hidden_dropout_prob,
initializer=initializer), bert_encoder
input_word_ids = tf.keras.layers.Input(shape=(max_seq_length, ),
dtype=tf.int32,
name='input_word_ids')
input_mask = tf.keras.layers.Input(shape=(max_seq_length, ),
dtype=tf.int32,
name='input_mask')
input_type_ids = tf.keras.layers.Input(shape=(max_seq_length, ),
dtype=tf.int32,
name='input_type_ids')
bert_model = hub.KerasLayer(hub_module_url, trainable=True)
pooled_output, _ = bert_model([input_word_ids, input_mask, input_type_ids])
output = tf.keras.layers.Dropout(
rate=bert_config.hidden_dropout_prob)(pooled_output)
output = tf.keras.layers.Dense(num_labels,
kernel_initializer=initializer,
name='output',
dtype=float_type)(output)
return tf.keras.Model(inputs={
'input_word_ids': input_word_ids,
'input_mask': input_mask,
'input_type_ids': input_type_ids
},
outputs=output), bert_model
