def get_masked_lm_output(bert_config, input_tensor, output_weights, positions):
"""Get loss and log probs for the masked LM."""
input_tensor = gather_indexes(input_tensor, positions)
with tf.variable_scope("cls/predictions"):
# We apply one more non-linear transformation before the output layer.
# This matrix is not used after pre-training.
with tf.variable_scope("transform"):
input_tensor = tf.layers.dense(
input_tensor,
units=bert_config.hidden_size,
activation=modeling.get_activation(bert_config.hidden_act),
kernel_initializer=modeling.create_initializer(
bert_config.initializer_range))
input_tensor = modeling.layer_norm(input_tensor)
# The output weights are the same as the input embeddings, but there is
# an output-only bias for each token.
output_bias = tf.get_variable("output_bias",
shape=[bert_config.vocab_size],
initializer=tf.zeros_initializer())
logits = tf.matmul(input_tensor, output_weights, transpose_b=True)
logits = tf.nn.bias_add(logits, output_bias)
return logits
def __init__(self, config, input_embedding, input_mask=None):
input_shape = modeling.get_shape_list(input_embedding, expected_rank=3)
batch_size = input_shape[0]
seq_length = input_shape[1]
if input_mask is None:
input_mask = tf.ones(shape=[batch_size, seq_length],
dtype=tf.int32)
# Keep variable names the same as BERT
with tf.variable_scope("bert"):
with tf.variable_scope("encoder"):
attention_mask = modeling.create_attention_mask_from_input_mask(
input_embedding, input_mask)
all_encoder_layers = modeling.transformer_model(
input_tensor=input_embedding,
attention_mask=attention_mask,
hidden_size=config.hidden_size,
num_hidden_layers=config.num_hidden_layers,
num_attention_heads=config.num_attention_heads,
intermediate_size=config.intermediate_size,
intermediate_act_fn=modeling.get_activation(
config.hidden_act),
hidden_dropout_prob=config.hidden_dropout_prob,
attention_probs_dropout_prob=config.
attention_probs_dropout_prob,
initializer_range=config.initializer_range,
do_return_all_layers=True)
self.sequence_output = all_encoder_layers[-1]
def __init__(
self,
):
self.X = tf.placeholder(tf.int32, [None, None])
model = modeling.BertModel(
config=bert_config,
is_training=False,
input_ids=self.X,
use_one_hot_embeddings=False)
output_layer = model.get_sequence_output()
embedding = model.get_embedding_table()
with tf.variable_scope('cls/predictions'):
with tf.variable_scope('transform'):
input_tensor = tf.layers.dense(
output_layer,
units = bert_config.hidden_size,
activation = modeling.get_activation(bert_config.hidden_act),
kernel_initializer = modeling.create_initializer(
bert_config.initializer_range
),
)
input_tensor = modeling.layer_norm(input_tensor)
output_bias = tf.get_variable(
'output_bias',
shape = [bert_config.vocab_size],
initializer = tf.zeros_initializer(),
)
logits = tf.matmul(input_tensor, embedding, transpose_b = True)
self.logits = tf.nn.bias_add(logits, output_bias)
def get_masked_lm_output(bert_config, input_tensor, output_weights, positions, top_k_indices, truncation_factor):
"""Get loss and log probs for the masked LM."""
input_tensor = gather_indexes(input_tensor, positions)
with tf.variable_scope("cls/predictions"):
# We apply one more non-linear transformation before the output layer.
# This matrix is not used after pre-training.
with tf.variable_scope("transform"):
input_tensor = tf.layers.dense(
input_tensor,
units=bert_config.hidden_size,
activation=modeling.get_activation(bert_config.hidden_act),
kernel_initializer=modeling.create_initializer(
bert_config.initializer_range))
input_tensor = modeling.layer_norm(input_tensor)
# The output weights are the same as the input embeddings, but there is
# an output-only bias for each token.
output_bias = tf.get_variable(
"output_bias",
shape=[bert_config.vocab_size],
initializer=tf.zeros_initializer())
logits = tf.matmul(input_tensor, output_weights, transpose_b=True)
logits = tf.nn.bias_add(logits, output_bias)
log_probs_student = tf.nn.log_softmax(logits, axis=-1)
probs_student = tf.nn.softmax(logits, axis=-1)
prob_shape = tf.shape(log_probs_student)
new_shape = [prob_shape[0], truncation_factor] #[batch_size*seq_len,truncation_factor]
top_k_indices = tf.reshape(top_k_indices, new_shape)
top_k_log_probs_student = tf.batch_gather(log_probs_student, top_k_indices)
top_k_probs_student = tf.batch_gather(probs_student, top_k_indices)
return top_k_log_probs_student, top_k_probs_student
def get_masked_lm_output(self):
self.input_tensor = self.gather_indexes()
with tf.variable_scope("cls/predictions"):
# We apply one more non-linear transformation before the output layer.
# This matrix is not used after pre-training.
with tf.variable_scope("transform"):
self.input_tensor = tf.layers.dense(
self.input_tensor,
units=self.bert_config.hidden_size,
activation=modeling.get_activation(
self.bert_config.hidden_act),
kernel_initializer=modeling.create_initializer(
self.bert_config.initializer_range))
self.input_tensor = modeling.layer_norm(self.input_tensor)
# The output weights are the same as the input embeddings, but there is
# an output-only bias for each token.
output_bias = tf.get_variable("output_bias",
shape=[self.bert_config.vocab_size],
initializer=tf.zeros_initializer())
logits = tf.matmul(self.input_tensor,
self.output_weights,
transpose_b=True)
logits = tf.nn.bias_add(logits, output_bias)
log_probs = tf.nn.log_softmax(logits, axis=-1)
flat_masked_lm_ids = tf.reshape(self.masked_lm_ids, [-1])
one_hot_labels = tf.one_hot(flat_masked_lm_ids,
depth=self.bert_config.vocab_size,
dtype=tf.float32)
per_example_loss = -tf.reduce_sum(log_probs * one_hot_labels,
axis=[-1])
# TODO: dynamic gather from per_example_loss???
loss = tf.reshape(per_example_loss,
[-1, tf.shape(self.masked_lm_positions)[1]])
return loss
def compute_image_transformer(
self,
input_ids,
input_image,
input_image_mask,
input_positions,
reuse=None,
):
"""Build the image transformer."""
with tf.variable_scope(self.scope_prefix + "transformer", reuse=reuse):
with tf.variable_scope("bridge"):
image_emb = tf.layers.dense(
inputs=input_image,
units=self.config.hidden_size,
activation=tf.nn.relu,
kernel_initializer=modeling.create_initializer(
self.config.initializer_range),
reuse=reuse)
with tf.variable_scope("embeddings"):
input_emb = tf.gather(self.embedding_table, input_ids)
image_emb = tf.concat([input_emb, image_emb], axis=1)
batch_size = tensor_utils.shape(image_emb, 0)
sequence_length = tensor_utils.shape(image_emb, 1)
position_emb = tf.gather(self.image_region_table,
input_positions)
position_emb = tf.pad(position_emb, [[0, 0], [1, 0], [0, 0]])
input_order = tf.range(tensor_utils.shape(image_emb, 1))
input_order = tf.tile(tf.expand_dims(input_order, 0),
[tensor_utils.shape(image_emb, 0), 1])
order_emb = tf.gather(self.image_order_table, input_order)
input_segment_id = tf.fill([batch_size, sequence_length],
self.IMG)
segment_emb = tf.gather(self.segment_table, input_segment_id)
input_emb = image_emb + position_emb + order_emb + segment_emb
input_emb = modeling.layer_norm_and_dropout(
input_emb, self.config.hidden_dropout_prob)
with tf.variable_scope("image/encoder"):
sequence_output, output_cache = compute_transformer(
input_tensor=input_emb,
attention_mask=tf.expand_dims(input_image_mask, 1),
hidden_size=self.config.hidden_size,
num_hidden_layers=self.config.num_hidden_layers,
num_attention_heads=self.config.num_attention_heads,
intermediate_size=self.config.intermediate_size,
intermediate_act_fn=modeling.get_activation(
self.config.hidden_act),
hidden_dropout_prob=self.config.hidden_dropout_prob,
attention_probs_dropout_prob=(
self.config.attention_probs_dropout_prob),
initializer_range=self.config.initializer_range,
input_cache=None)
return sequence_output, output_cache
def build_bert_model(self, input_ids, input_mask, token_type_ids):
with tf.variable_scope('bert'):
with tf.variable_scope("embeddings"):
# Perform embedding lookup on the word ids.
(embedding_output, _) = modeling.embedding_lookup(
input_ids=input_ids,
vocab_size=self.bert_config.vocab_size,
embedding_size=self.bert_config.hidden_size,
initializer_range=self.bert_config.initializer_range,
word_embedding_name="word_embeddings",
use_one_hot_embeddings=False)
# Add positional embeddings and token type embeddings, then layer
# normalize and perform dropout.
embedding_output = modeling.embedding_postprocessor(
input_tensor=embedding_output,
use_token_type=True,
token_type_ids=token_type_ids,
token_type_vocab_size=self.bert_config.type_vocab_size,
token_type_embedding_name="token_type_embeddings",
use_position_embeddings=True,
position_embedding_name="position_embeddings",
initializer_range=self.bert_config.initializer_range,
max_position_embeddings=self.bert_config.
max_position_embeddings,
dropout_prob=self.bert_config.hidden_dropout_prob)
with tf.variable_scope("encoder"):
# This converts a 2D mask of shape [batch_size, seq_length] to a 3D
# mask of shape [batch_size, seq_length, seq_length] which is used
# for the attention scores.
attention_mask = modeling.create_attention_mask_from_input_mask(
input_ids, input_mask)
# Run the stacked transformer, only fetching the final lyaer
# `final_layer` shape = [batch_size, seq_length, hidden_size].
self.all_encoder_layers = modeling.transformer_model(
input_tensor=embedding_output,
attention_mask=attention_mask,
hidden_size=self.bert_config.hidden_size,
num_hidden_layers=self.bert_config.num_hidden_layers,
num_attention_heads=self.bert_config.num_attention_heads,
intermediate_size=self.bert_config.intermediate_size,
intermediate_act_fn=modeling.get_activation(
self.bert_config.hidden_act),
hidden_dropout_prob=self.bert_config.hidden_dropout_prob,
attention_probs_dropout_prob=\
self.bert_config.attention_probs_dropout_prob,
initializer_range=self.bert_config.initializer_range,
do_return_all_layers=True
)
self.sequence_output = self.all_encoder_layers[-1]
def get_masked_lm_output(bert_config, input_tensor, output_weights, positions,
label_weights, truncated_masked_lm_probs_teacher,
top_k_indices, truncation_factor):
"""Get loss and log probs for the masked LM."""
input_tensor = gather_indexes(input_tensor, positions)
with tf.variable_scope("cls/predictions"):
# We apply one more non-linear transformation before the output layer.
# This matrix is not used after pre-training.
with tf.variable_scope("transform"):
input_tensor = tf.layers.dense(
input_tensor,
units=bert_config.hidden_size,
activation=modeling.get_activation(bert_config.hidden_act),
kernel_initializer=modeling.create_initializer(
bert_config.initializer_range))
input_tensor = modeling.layer_norm(input_tensor)
# The output weights are the same as the input embeddings, but there is
# an output-only bias for each token.
output_bias = tf.get_variable("output_bias",
shape=[bert_config.vocab_size],
initializer=tf.zeros_initializer())
logits = tf.matmul(input_tensor, output_weights, transpose_b=True)
logits = tf.nn.bias_add(logits, output_bias)
log_probs_student = tf.nn.log_softmax(logits, axis=-1)
label_weights = tf.reshape(label_weights, [-1])
prob_shape = tf.shape(log_probs_student)
new_shape = [prob_shape[0], truncation_factor
] #[batch_size*seq_len,truncation_factor]
top_k_indices = tf.reshape(top_k_indices, new_shape)
top_k_log_probs_student = tf.batch_gather(log_probs_student,
top_k_indices)
truncated_masked_lm_probs_teacher = tf.reshape(
truncated_masked_lm_probs_teacher, new_shape)
# The `positions` tensor might be zero-padded (if the sequence is too
# short to have the maximum number of predictions). The `label_weights`
# tensor has a value of 1.0 for every real prediction and 0.0 for the
# padding predictions.
per_example_loss = -tf.reduce_sum(
truncated_masked_lm_probs_teacher * top_k_log_probs_student,
axis=[-1])
numerator = tf.reduce_sum(label_weights * per_example_loss)
denominator = tf.reduce_sum(label_weights) + 1e-5
loss = numerator / denominator
return (loss, per_example_loss, log_probs_student)
def get_masked_lm_output(bert_config, input_tensor, output_weights, positions,
label_ids, label_weights):
"""Get loss and log probs for the masked LM."""
input_tensor = gather_indexes(input_tensor, positions)
with tf.variable_scope("cls/predictions"):
# We apply one more non-linear transformation before the output layer.
# This matrix is not used after pre-training.
with tf.variable_scope("transform"):
input_tensor = tf.layers.dense(
input_tensor,
units=bert_config.hidden_size,
activation=modeling.get_activation(bert_config.hidden_act),
kernel_initializer=modeling.create_initializer(
bert_config.initializer_range))
input_tensor = modeling.layer_norm(input_tensor)
# The output weights are the same as the input embeddings, but there is
# an output-only bias for each token.
output_bias = tf.get_variable("output_bias",
shape=[bert_config.vocab_size],
initializer=tf.zeros_initializer())
logits = tf.matmul(input_tensor, output_weights, transpose_b=True)
logits = tf.nn.bias_add(logits, output_bias)
log_probs = tf.nn.log_softmax(logits, axis=-1)
label_ids = tf.reshape(label_ids, [-1])
label_weights_flat = tf.reshape(label_weights, [-1])
one_hot_labels = tf.one_hot(label_ids,
depth=bert_config.vocab_size,
dtype=tf.float32)
# The `positions` tensor might be zero-padded (if the sequence is too
# short to have the maximum number of predictions). The `label_weights`
# tensor has a value of 1.0 for every real prediction and 0.0 for the
# padding predictions.
per_example_loss = -tf.reduce_sum(log_probs * one_hot_labels,
axis=[-1])
numerator = tf.reduce_sum(label_weights_flat * per_example_loss)
denominator = tf.reduce_sum(label_weights_flat) + 1e-5
loss = numerator / denominator
batch_size = tf.cast(tf.shape(label_weights)[0], tf.float32)
print('==============')
print(label_weights.shape)
print('==============')
loss = batch_size * loss
return (loss, per_example_loss, log_probs)
def build(self, unused_input_shapes):
"""Implements build() for the layer."""
self.lm_dense = tf.keras.layers.Dense(
self.config.hidden_size,
activation=modeling.get_activation(self.config.hidden_act),
kernel_initializer=self.initializer)
self.lm_bias = self.add_weight(
shape=[self.config.vocab_size],
name='lm_bias',
initializer=tf.keras.initializers.Zeros())
self.lm_layer_norm = tf.keras.layers.LayerNormalization(
axis=-1, epsilon=1e-12)
self.next_sentence_dense = tf.keras.layers.Dense(
self.num_next_sentence_label, kernel_initializer=self.initializer)
super(BertPretrainLayer, self).build(unused_input_shapes)
def bert_module_fn(is_training):
"""Spec function for a token embedding module."""
input_ids = tf.placeholder(shape=[None, None], dtype=tf.int32, name="input_ids")
bert_config = modeling.BertConfig.from_json_file(config_path)
model = modeling.BertModel(config=bert_config, is_training=is_training,
input_ids=input_ids)
output_layer = model.get_sequence_output()
embedding = model.get_embedding_table()
with tf.variable_scope('cls/predictions'):
with tf.variable_scope('transform'):
input_tensor = tf.layers.dense(
output_layer,
units=bert_config.hidden_size,
activation=modeling.get_activation(bert_config.hidden_act),
kernel_initializer=modeling.create_initializer(
bert_config.initializer_range
),
)
input_tensor = modeling.layer_norm(input_tensor)
output_bias = tf.get_variable(
'output_bias',
shape=[bert_config.vocab_size],
initializer=tf.zeros_initializer(),
)
logits = tf.matmul(input_tensor, embedding, transpose_b=True)
logits = tf.nn.bias_add(logits, output_bias)
config_file = tf.constant(value=config_path, dtype=tf.string, name="config_file")
vocab_file = tf.constant(value=vocab_path, dtype=tf.string, name="vocab_file")
lower_case = tf.constant(do_lower_case)
tf.add_to_collection(tf.GraphKeys.ASSET_FILEPATHS, config_file)
tf.add_to_collection(tf.GraphKeys.ASSET_FILEPATHS, vocab_file)
input_map = {"input_ids": input_ids}
output_map = {"logits": logits}
output_info_map = {"vocab_file": vocab_file,
"do_lower_case": lower_case}
hub.add_signature(name="tokens", inputs=input_map, outputs=output_map)
hub.add_signature(name="tokenization_info", inputs={}, outputs=output_info_map)
def forward(x, segment, masks, y, reuse=False, config=bert_config):
with tf.variable_scope('bert', reuse=reuse):
model = modeling.BertModel(
config=config,
is_training=training,
input_ids=x,
input_mask=masks,
token_type_ids=segment,
use_one_hot_embeddings=False,
)
memory = model.get_sequence_output()
with tf.variable_scope('bert', reuse=True):
Y_seq_len = tf.count_nonzero(y, 1, dtype=tf.int32)
y_masks = tf.sequence_mask(Y_seq_len,
tf.reduce_max(Y_seq_len),
dtype=tf.float32)
model = modeling_decoder.BertModel(
config=config,
is_training=training,
input_ids=y,
input_mask=y_masks,
memory=memory,
memory_mask=masks,
use_one_hot_embeddings=False,
)
output_layer = model.get_sequence_output()
embedding = model.get_embedding_table()
with tf.variable_scope('cls/predictions', reuse=reuse):
with tf.variable_scope('transform'):
input_tensor = tf.layers.dense(
output_layer,
units=config.hidden_size,
activation=modeling.get_activation(
bert_config.hidden_act),
kernel_initializer=modeling.create_initializer(
bert_config.initializer_range),
)
input_tensor = modeling.layer_norm(input_tensor)
output_bias = tf.get_variable(
'output_bias',
shape=[bert_config.vocab_size],
initializer=tf.zeros_initializer(),
)
logits = tf.matmul(input_tensor, embedding, transpose_b=True)
return logits
def compute_logits(self, target_emb, reuse=None):
"""Compute logits for word prediction."""
with tf.variable_scope(self.scope_prefix + "cls/predictions",
reuse=reuse):
with tf.variable_scope("transform"):
target_emb = tf.layers.dense(
target_emb,
units=self.config.hidden_size,
activation=modeling.get_activation(self.config.hidden_act),
kernel_initializer=self.initializer)
target_emb = modeling.layer_norm(target_emb)
output_bias = tf.get_variable("output_bias",
shape=[self.config.vocab_size],
initializer=tf.zeros_initializer())
logits = tf.matmul(target_emb, self.embedding_table, transpose_b=True)
logits = tf.nn.bias_add(logits, output_bias)
return logits
def get_masked_lm_output(bert_config, input_tensor, output_weights, positions,
label_ids):
"""Get loss and log probs for the masked LM."""
print("input tensor before gather_indexes:", input_tensor)
input_tensor = gather_indexes(input_tensor, positions)
print("input tensor before gather_indexes:", input_tensor)
with tf.variable_scope("cls/predictions"):
# We apply one more non-linear transformation before the output layer.
# This matrix is not used after pre-training.
with tf.variable_scope("transform"):
input_tensor = tf.layers.dense(
input_tensor,
units=bert_config.hidden_size,
activation=modeling.get_activation(bert_config.hidden_act),
kernel_initializer=modeling.create_initializer(
bert_config.initializer_range))
input_tensor = modeling.layer_norm(input_tensor)
# The output weights are the same as the input embeddings, but there is
# an output-only bias for each token.
output_bias = tf.get_variable("output_bias",
shape=[bert_config.vocab_size],
initializer=tf.zeros_initializer())
logits = tf.matmul(input_tensor, output_weights, transpose_b=True)
logits = tf.nn.bias_add(logits, output_bias)
log_probs = tf.nn.log_softmax(logits, axis=-1)
print(label_ids)
label_ids = tf.reshape(label_ids, [-1])
print(label_ids)
one_hot_labels = tf.one_hot(label_ids,
depth=bert_config.vocab_size,
dtype=tf.float32)
print(one_hot_labels)
print(log_probs)
per_example_loss = -tf.reduce_sum(log_probs * one_hot_labels,
axis=[-1])
print(per_example_loss)
loss = tf.reshape(per_example_loss, [-1, tf.shape(positions)[1]])
print('positions: ', positions)
print('loss', loss)
# TODO: dynamic gather from per_example_loss???
return loss
def __init__(self, config, input_hidden, embedding_table):
# Keep variable names the same as BERT
with tf.variable_scope("cls"):
with tf.variable_scope("predictions"):
with tf.variable_scope("transform"):
self.transformed_output = tf.layers.dense(
input_hidden,
config.hidden_size,
activation=modeling.get_activation(config.hidden_act),
kernel_initializer=modeling.create_initializer(
config.initializer_range))
self.transformed_output = modeling.layer_norm(
self.transformed_output)
output_bias = tf.Variable(tf.zeros([config.vocab_size]),
name="output_bias")
self.final_output = tf.add(
tf.matmul(self.transformed_output,
tf.transpose(embedding_table)), output_bias)
self.probs = tf.nn.softmax(self.final_output,
name='token_probs')
def compute_transformer(
self,
input_ids,
input_segment_id,
input_positions,
attention_mask,
input_cache=None,
reuse=None,
conditional=False,
):
"""Build the full text transformer."""
with tf.variable_scope(self.scope_prefix + "transformer", reuse=reuse):
with tf.variable_scope("embeddings"):
token_emb = tf.gather(self.embedding_table, input_ids)
segment_embed = tf.gather(self.segment_table, input_segment_id)
if conditional:
position_table = self.condition_position_table
else:
position_table = self.position_table
position_emb = tf.gather(position_table, input_positions)
input_emb = token_emb + segment_embed + position_emb
input_emb = modeling.layer_norm_and_dropout(
input_emb, self.config.hidden_dropout_prob)
with tf.variable_scope("encoder"):
sequence_output, output_cache = compute_transformer(
input_tensor=input_emb,
attention_mask=attention_mask,
hidden_size=self.config.hidden_size,
num_hidden_layers=self.config.num_hidden_layers,
num_attention_heads=self.config.num_attention_heads,
intermediate_size=self.config.intermediate_size,
intermediate_act_fn=modeling.get_activation(
self.config.hidden_act),
hidden_dropout_prob=self.config.hidden_dropout_prob,
attention_probs_dropout_prob=(
self.config.attention_probs_dropout_prob),
initializer_range=self.config.initializer_range,
input_cache=input_cache)
return sequence_output, output_cache
def get_masked_lm_output(bert_config, input_tensor, output_weights, positions,
truncation_factor):
"""Get loss and log probs for the masked LM."""
input_tensor = gather_indexes(input_tensor, positions)
with tf.variable_scope("cls/predictions"):
# We apply one more non-linear transformation before the output layer.
# This matrix is not used after pre-training.
with tf.variable_scope("transform"):
input_tensor = tf.layers.dense(
input_tensor,
units=bert_config.hidden_size,
activation=modeling.get_activation(bert_config.hidden_act),
kernel_initializer=modeling.create_initializer(
bert_config.initializer_range))
input_tensor = modeling.layer_norm(input_tensor)
# The output weights are the same as the input embeddings, but there is
# an output-only bias for each token.
output_bias = tf.get_variable("output_bias",
shape=[bert_config.vocab_size],
initializer=tf.zeros_initializer())
logits = tf.matmul(input_tensor, output_weights, transpose_b=True)
logits = tf.nn.bias_add(logits, output_bias)
masked_lm_probs = tf.nn.softmax(logits, axis=-1)
trunc_masked_lm_probs, top_indices = tf.math.top_k(masked_lm_probs,
k=truncation_factor,
sorted=False)
max_predictions_per_seq = positions.get_shape().as_list()[1]
truncation_factor_ = top_indices.get_shape().as_list()[1]
trunc_masked_lm_probs = tf.reshape(
trunc_masked_lm_probs,
[-1, max_predictions_per_seq, truncation_factor_])
top_indices = tf.reshape(
top_indices, [-1, max_predictions_per_seq, truncation_factor_])
return trunc_masked_lm_probs, top_indices
def __init__(
self,
):
BERT_CONFIG = "PATH_TO/multi_cased_L-12_H-768_A-12/bert_config.json"
bert_config = modeling.BertConfig.from_json_file(BERT_CONFIG)
self.X = tf.placeholder(tf.int32, [None, None])
model = modeling.BertModel(
config=bert_config,
is_training=False,
input_ids=self.X,
use_one_hot_embeddings=False,
)
output_layer = model.get_sequence_output()
embedding = model.get_embedding_table()
output_layer = tf.reshape(output_layer, [-1, bert_config.hidden_size])
with tf.variable_scope("cls/predictions"):
with tf.variable_scope("transform"):
input_tensor = tf.layers.dense(
output_layer,
units=bert_config.hidden_size,
activation=modeling.get_activation(bert_config.hidden_act),
kernel_initializer=modeling.create_initializer(
bert_config.initializer_range
),
)
input_tensor = modeling.layer_norm(input_tensor)
output_bias = tf.get_variable(
"output_bias",
shape=[bert_config.vocab_size],
initializer=tf.zeros_initializer(),
)
logits = tf.matmul(input_tensor, embedding, transpose_b=True)
print("---")
self.logits = tf.nn.bias_add(logits, output_bias)
def __init__(
self,
bert_config,
input_ids,
input_mask,
token_type_ids,
Y,
is_training=True,
):
self.X = input_ids
self.segment_ids = token_type_ids
self.input_masks = input_mask
self.Y = Y
self.X_seq_len = tf.count_nonzero(self.X, 1, dtype=tf.int32)
self.Y_seq_len = tf.count_nonzero(self.Y, 1, dtype=tf.int32)
batch_size = tf.shape(self.X)[0]
model = modeling.BertModel(
config=bert_config,
is_training=is_training,
input_ids=self.X,
input_mask=self.input_masks,
token_type_ids=self.segment_ids,
use_one_hot_embeddings=False,
)
print(bert_config.__dict__)
BASE_PARAMS = defaultdict(
lambda: None,
default_batch_size=2048,
default_batch_size_tpu=32768,
max_length=bert_config.max_position_embeddings,
initializer_gain=1.0,
vocab_size=bert_config.vocab_size,
hidden_size=bert_config.hidden_size,
num_hidden_layers=bert_config.num_hidden_layers,
num_heads=bert_config.num_attention_heads,
filter_size=bert_config.intermediate_size,
layer_postprocess_dropout=0.1,
attention_dropout=0.1,
relu_dropout=0.1,
label_smoothing=0.1,
learning_rate=1.0,
learning_rate_decay_rate=1.0,
learning_rate_warmup_steps=16000,
optimizer_adam_beta1=0.9,
optimizer_adam_beta2=0.997,
optimizer_adam_epsilon=1e-09,
extra_decode_length=50,
beam_size=4,
alpha=0.6,
use_tpu=False,
static_batch=False,
allow_ffn_pad=True,
)
self.decoder_stack = DecoderStack(BASE_PARAMS, is_training)
attention_bias = model_utils.get_padding_bias(self.X)
output_layer = model.get_sequence_output()
pooled_output = model.get_pooled_output()
embedding = model.get_embedding_table()
with tf.name_scope('decode'):
mask = tf.to_float(tf.not_equal(self.Y, 0))
decoder_inputs = tf.gather(embedding, self.Y)
decoder_inputs *= tf.expand_dims(mask, -1)
with tf.name_scope('shift_targets'):
decoder_inputs = tf.pad(decoder_inputs,
[[0, 0], [1, 0], [0, 0]])[:, :-1, :]
with tf.name_scope('add_pos_encoding'):
length = tf.shape(decoder_inputs)[1]
decoder_inputs += model_utils.get_position_encoding(
length, BASE_PARAMS['hidden_size'])
if is_training:
decoder_inputs = tf.nn.dropout(
decoder_inputs,
1 - BASE_PARAMS['layer_postprocess_dropout'])
decoder_self_attention_bias = model_utils.get_decoder_self_attention_bias(
length)
outputs = self.decoder_stack(
decoder_inputs,
output_layer,
decoder_self_attention_bias,
attention_bias,
)
with tf.variable_scope('cls/predictions'):
with tf.variable_scope('transform'):
input_tensor = tf.layers.dense(
outputs,
units=bert_config.hidden_size,
activation=modeling.get_activation(bert_config.hidden_act),
kernel_initializer=modeling.create_initializer(
bert_config.initializer_range),
)
input_tensor = modeling.layer_norm(input_tensor)
output_bias = tf.get_variable(
'output_bias',
shape=[bert_config.vocab_size],
initializer=tf.zeros_initializer(),
)
self.training_logits = tf.matmul(input_tensor,
embedding,
transpose_b=True)
print(self.training_logits)
def __init__(self, bert_config, tokenizer, cls, sep):
_graph = tf.Graph()
with _graph.as_default():
self.X = tf.placeholder(tf.int32, [None, None])
self.top_p = tf.placeholder(tf.float32, None)
self.top_k = tf.placeholder(tf.int32, None)
self.k = tf.placeholder(tf.int32, None)
self.temperature = tf.placeholder(tf.float32, None)
self.indices = tf.placeholder(tf.int32, [None, None])
self._tokenizer = tokenizer
self._cls = cls
self._sep = sep
self.model = modeling.BertModel(
config = bert_config,
is_training = False,
input_ids = self.X,
use_one_hot_embeddings = False,
)
self.logits = self.model.get_pooled_output()
output_layer = self.model.get_sequence_output()
embedding = self.model.get_embedding_table()
with tf.variable_scope('cls/predictions'):
with tf.variable_scope('transform'):
input_tensor = tf.layers.dense(
output_layer,
units = bert_config.hidden_size,
activation = modeling.get_activation(
bert_config.hidden_act
),
kernel_initializer = modeling.create_initializer(
bert_config.initializer_range
),
)
input_tensor = modeling.layer_norm(input_tensor)
output_bias = tf.get_variable(
'output_bias',
shape = [bert_config.vocab_size],
initializer = tf.zeros_initializer(),
)
logits = tf.matmul(input_tensor, embedding, transpose_b = True)
self._logits = tf.nn.bias_add(logits, output_bias)
self._log_softmax = tf.nn.log_softmax(self._logits)
logits = tf.gather_nd(self._logits, self.indices)
logits = logits / self.temperature
def necleus():
return top_p_logits(logits, self.top_p)
def select_k():
return top_k_logits(logits, self.top_k)
logits = tf.cond(self.top_p > 0, necleus, select_k)
self.samples = tf.multinomial(
logits, num_samples = self.k, output_dtype = tf.int32
)
self._sess = tf.InteractiveSession()
self._sess.run(tf.global_variables_initializer())
var_lists = tf.get_collection(
tf.GraphKeys.TRAINABLE_VARIABLES, scope = 'bert'
)
cls = tf.get_collection(
tf.GraphKeys.TRAINABLE_VARIABLES, scope = 'cls'
)
self._saver = tf.train.Saver(var_list = var_lists + cls)
attns = _extract_attention_weights(
bert_config.num_hidden_layers, tf.get_default_graph()
)
self.attns = attns
def __init__(self,
config,
is_training,
input_ids,
image_embeddings,
input_mask=None,
token_type_ids=None,
use_one_hot_embeddings=False,
scope=None):
"""Constructor for a visually grounded BertModel.
Args:
config: `BertConfig` instance.
is_training: bool. true for training model, false for eval model. Controls
whether dropout will be applied.
input_ids: int32 Tensor of shape [batch_size, seq_length].
image_embeddings: float32 Tensor of shape [batch_size, seq_length, depth].
input_mask: (optional) int32 Tensor of shape [batch_size, seq_length].
token_type_ids: (optional) int32 Tensor of shape [batch_size, seq_length].
use_one_hot_embeddings: (optional) bool. Whether to use one-hot word
embeddings or tf.embedding_lookup() for the word embeddings.
scope: (optional) variable scope. Defaults to "bert".
Raises:
ValueError: The config is invalid or one of the input tensor shapes
is invalid.
"""
config = copy.deepcopy(config)
if not is_training:
config.hidden_dropout_prob = 0.0
config.attention_probs_dropout_prob = 0.0
text_input_shape = modeling.get_shape_list(input_ids, expected_rank=2)
batch_size = text_input_shape[0]
text_seq_length = text_input_shape[1]
if input_mask is None:
input_mask = tf.ones(shape=[batch_size, text_seq_length],
dtype=tf.int32)
if token_type_ids is None:
token_type_ids = tf.zeros(shape=[batch_size, text_seq_length],
dtype=tf.int32)
with tf.variable_scope(scope, default_name="bert"):
with tf.variable_scope("embeddings"):
# Perform embedding lookup on the word ids.
(self.embedding_output,
self.embedding_table) = modeling.embedding_lookup(
input_ids=input_ids,
vocab_size=config.vocab_size,
embedding_size=config.hidden_size,
initializer_range=config.initializer_range,
word_embedding_name="word_embeddings",
use_one_hot_embeddings=use_one_hot_embeddings)
# Add positional embeddings and token type embeddings, then layer
# normalize and perform dropout.
self.embedding_output = modeling.embedding_postprocessor(
input_tensor=self.embedding_output,
use_token_type=True,
token_type_ids=token_type_ids,
token_type_vocab_size=config.type_vocab_size,
token_type_embedding_name="token_type_embeddings",
use_position_embeddings=True,
position_embedding_name="position_embeddings",
initializer_range=config.initializer_range,
max_position_embeddings=config.max_position_embeddings,
dropout_prob=config.hidden_dropout_prob)
# Add image embeddings the rest of the input embeddings.
self.embedding_output += tf.layers.dense(
image_embeddings,
config.hidden_size,
activation=tf.tanh,
kernel_initializer=modeling.create_initializer(
config.initializer_range))
with tf.variable_scope("encoder"):
# This converts a 2D mask of shape [batch_size, seq_length] to a 3D
# mask of shape [batch_size, seq_length, seq_length] which is used
# for the attention scores.
attention_mask = modeling.create_attention_mask_from_input_mask(
self.embedding_output, input_mask)
# Run the stacked transformer.
# `sequence_output` shape = [batch_size, seq_length, hidden_size].
self.all_encoder_layers = modeling.transformer_model(
input_tensor=self.embedding_output,
attention_mask=attention_mask,
hidden_size=config.hidden_size,
num_hidden_layers=config.num_hidden_layers,
num_attention_heads=config.num_attention_heads,
intermediate_size=config.intermediate_size,
intermediate_act_fn=modeling.get_activation(
config.hidden_act),
hidden_dropout_prob=config.hidden_dropout_prob,
attention_probs_dropout_prob=config.
attention_probs_dropout_prob,
initializer_range=config.initializer_range,
do_return_all_layers=True)
self.sequence_output = self.all_encoder_layers[-1]
# The "pooler" converts the encoded sequence tensor of shape
# [batch_size, seq_length, hidden_size] to a tensor of shape
# [batch_size, hidden_size]. This is necessary for segment-level
# (or segment-pair-level) classification tasks where we need a fixed
# dimensional representation of the segment.
with tf.variable_scope("pooler"):
# We "pool" the model by simply taking the hidden state corresponding
# to the first token. We assume that this has been pre-trained
first_token_tensor = tf.squeeze(self.sequence_output[:,
0:1, :],
axis=1)
self.pooled_output = tf.layers.dense(
first_token_tensor,
config.hidden_size,
activation=tf.tanh,
kernel_initializer=modeling.create_initializer(
config.initializer_range))
def __init__(self, config, is_training, input_tensor, input_mask,
token_type_ids):
"""Constructor for BertFlexEmbeddingModel.
Args:
config: `BertConfig` instance.
is_training: bool. true for training model, false for eval model. Controls
whether dropout will be applied.
input_tensor: float32 Tensor of shape [batch_size, seq_length,
hidden_size].
input_mask: (optional) int32 Tensor of shape [batch_size, seq_length].
token_type_ids: (optional) int32 Tensor of shape [batch_size, seq_length].
Raises:
ValueError: The config is invalid or one of the input tensor shapes
is invalid.
"""
config = copy.deepcopy(config)
if not is_training:
config.hidden_dropout_prob = 0.0
config.attention_probs_dropout_prob = 0.0
with tf.variable_scope("bert", reuse=tf.compat.v1.AUTO_REUSE):
with tf.variable_scope("embeddings"):
# Add positional embeddings and token type embeddings, then layer
# normalize and perform dropout.
self.embedding_output = modeling.embedding_postprocessor(
input_tensor=input_tensor,
use_token_type=True,
token_type_ids=token_type_ids,
token_type_vocab_size=config.type_vocab_size,
token_type_embedding_name="token_type_embeddings",
use_position_embeddings=True,
position_embedding_name="position_embeddings",
initializer_range=config.initializer_range,
max_position_embeddings=config.max_position_embeddings,
dropout_prob=config.hidden_dropout_prob)
with tf.variable_scope("encoder"):
# This converts a 2D mask of shape [batch_size, seq_length] to a 3D
# mask of shape [batch_size, seq_length, seq_length] which is used
# for the attention scores.
attention_mask = modeling.create_attention_mask_from_input_mask(
input_tensor, input_mask)
# Run the stacked transformer.
# `sequence_output` shape = [batch_size, seq_length, hidden_size].
self.all_encoder_layers = modeling.transformer_model(
input_tensor=self.embedding_output,
attention_mask=attention_mask,
hidden_size=config.hidden_size,
num_hidden_layers=config.num_hidden_layers,
num_attention_heads=config.num_attention_heads,
intermediate_size=config.intermediate_size,
intermediate_act_fn=modeling.get_activation(
config.hidden_act),
hidden_dropout_prob=config.hidden_dropout_prob,
attention_probs_dropout_prob=config.
attention_probs_dropout_prob,
initializer_range=config.initializer_range,
do_return_all_layers=True)
self.sequence_output = self.all_encoder_layers[-1]
# The "pooler" converts the encoded sequence tensor of shape
# [batch_size, seq_length, hidden_size] to a tensor of shape
# [batch_size, hidden_size]. This is necessary for segment-level
# (or segment-pair-level) classification tasks where we need a fixed
# dimensional representation of the segment.
with tf.variable_scope("pooler"):
# We "pool" the model by simply taking the hidden state corresponding
# to the first token. We assume that this has been pre-trained
first_token_tensor = tf.squeeze(self.sequence_output[:,
0:1, :],
axis=1)
self.pooled_output = tf.layers.dense(
first_token_tensor,
config.hidden_size,
activation=tf.tanh,
kernel_initializer=modeling.create_initializer(
config.initializer_range))
def __init__(
self,
input_ids,
input_mask,
token_type_ids,
Y,
learning_rate=2e-5,
is_training=True,
):
self.X = input_ids
self.segment_ids = token_type_ids
self.input_masks = input_mask
self.Y = Y
self.X_seq_len = tf.count_nonzero(self.X, 1, dtype=tf.int32)
self.Y_seq_len = tf.count_nonzero(self.Y, 1, dtype=tf.int32)
batch_size = tf.shape(self.X)[0]
model = modeling.BertModel(
config=bert_config,
is_training=is_training,
input_ids=self.X,
input_mask=self.input_masks,
token_type_ids=self.segment_ids,
use_one_hot_embeddings=False,
)
self.decoder_stack = DecoderStack(BASE_PARAMS, is_training)
attention_bias = model_utils.get_padding_bias(self.X)
output_layer = model.get_sequence_output()
pooled_output = model.get_pooled_output()
embedding = model.get_embedding_table()
with tf.name_scope('decode'):
mask = tf.to_float(tf.not_equal(self.Y, 0))
decoder_inputs = tf.gather(embedding, self.Y)
decoder_inputs *= tf.expand_dims(mask, -1)
with tf.name_scope('shift_targets'):
decoder_inputs = tf.pad(decoder_inputs,
[[0, 0], [1, 0], [0, 0]])[:, :-1, :]
with tf.name_scope('add_pos_encoding'):
length = tf.shape(decoder_inputs)[1]
decoder_inputs += model_utils.get_position_encoding(
length, BASE_PARAMS['hidden_size'])
if is_training:
decoder_inputs = tf.nn.dropout(
decoder_inputs,
1 - BASE_PARAMS['layer_postprocess_dropout'])
decoder_self_attention_bias = model_utils.get_decoder_self_attention_bias(
length)
outputs = self.decoder_stack(
decoder_inputs,
output_layer,
decoder_self_attention_bias,
attention_bias,
)
with tf.variable_scope('cls/predictions'):
with tf.variable_scope('transform'):
input_tensor = tf.layers.dense(
outputs,
units=bert_config.hidden_size,
activation=modeling.get_activation(bert_config.hidden_act),
kernel_initializer=modeling.create_initializer(
bert_config.initializer_range),
)
input_tensor = modeling.layer_norm(input_tensor)
output_bias = tf.get_variable(
'output_bias',
shape=[bert_config.vocab_size],
initializer=tf.zeros_initializer(),
)
self.training_logits = tf.matmul(input_tensor,
embedding,
transpose_b=True)
print(self.training_logits)
def create_mask_model(bert_config, is_training, input_ids, input_mask,
segment_ids, mask_positions, use_one_hot_embeddings):
"""Creates a classification model."""
#print("create mask model ----------------------------------------------")
model = modeling.BertModel(config=bert_config,
is_training=is_training,
input_ids=input_ids,
input_mask=input_mask,
token_type_ids=segment_ids,
use_one_hot_embeddings=use_one_hot_embeddings)
# Get the logits for the start and end predictions.
final_hidden = model.get_sequence_output()
final_hidden_shape = modeling.get_shape_list(final_hidden, expected_rank=3)
batch_size = final_hidden_shape[0]
seq_length = final_hidden_shape[1]
hidden_size = final_hidden_shape[2]
output_weights = tf.get_variable(
"cls/nq/output_weights", [2, hidden_size + 12],
initializer=tf.truncated_normal_initializer(stddev=0.02))
output_bias = tf.get_variable("cls/nq/output_bias", [2],
initializer=tf.zeros_initializer())
final_hidden_matrix = tf.reshape(final_hidden,
[batch_size * seq_length, hidden_size])
mask_positions_matrix = tf.cast(tf.reshape(mask_positions,
[batch_size * seq_length, 1]),
dtype=tf.float32)
padding = tf.zeros([batch_size * seq_length, 11], dtype=tf.float32)
mask_positions_matrix = tf.concat([mask_positions_matrix, padding],
axis=-1)
final_hidden_matrix = tf.concat(
[final_hidden_matrix, mask_positions_matrix], axis=-1)
final_hidden_matrix = tf.reshape(
final_hidden_matrix, [batch_size, seq_length, hidden_size + 12])
attention_mask = modeling.create_attention_mask_from_input_mask(
input_ids, input_mask)
config = bert_config
all_encoder_layers = modeling.transformer_model(
input_tensor=final_hidden_matrix,
attention_mask=attention_mask,
hidden_size=config.hidden_size + 12, # input hidden size
num_hidden_layers=1, #config.num_hidden_layers,
num_attention_heads=config.num_attention_heads,
intermediate_size=config.intermediate_size,
intermediate_act_fn=modeling.get_activation(config.hidden_act),
hidden_dropout_prob=config.hidden_dropout_prob,
attention_probs_dropout_prob=config.attention_probs_dropout_prob,
initializer_range=config.initializer_range,
do_return_all_layers=True)
#print(all_encoder_layers.shape)
transformer_output_matrix = all_encoder_layers[-1]
transformer_output_matrix = tf.reshape(
transformer_output_matrix, [batch_size * seq_length, hidden_size + 12])
logits = tf.matmul(transformer_output_matrix,
output_weights,
transpose_b=True)
logits = tf.nn.bias_add(logits, output_bias)
logits = tf.reshape(logits, [batch_size, seq_length, 2])
logits = tf.transpose(logits, [2, 0, 1])
unstacked_logits = tf.unstack(logits, axis=0)
(start_logits, end_logits) = (unstacked_logits[0], unstacked_logits[1])
# Get the logits for the answer type prediction.
answer_type_output_layer = model.get_pooled_output()
answer_type_hidden_size = answer_type_output_layer.shape[-1].value
num_answer_types = 5 # YES, NO, UNKNOWN, SHORT, LONG
answer_type_output_weights = tf.get_variable(
"answer_type_output_weights",
[num_answer_types, answer_type_hidden_size],
initializer=tf.truncated_normal_initializer(stddev=0.02))
answer_type_output_bias = tf.get_variable(
"answer_type_output_bias", [num_answer_types],
initializer=tf.zeros_initializer())
answer_type_logits = tf.matmul(answer_type_output_layer,
answer_type_output_weights,
transpose_b=True)
answer_type_logits = tf.nn.bias_add(answer_type_logits,
answer_type_output_bias)
return (start_logits, end_logits, answer_type_logits)
def main(args):
bert_config = modeling.BertConfig.from_json_file(args.config)
bert_config.hidden_dropout_prob = 0.0
bert_config.attention_probs_dropout_prob = 0.0
batch_size = args.batch_size
avg_seq_len = args.avg_seq_length
max_seq_len = args.max_seq_length
tf_dtype = tf.float16 if args.precision == 'fp16' else tf.float32
# fake input array length
input_len = np.random.randint(low=2 * avg_seq_len - max_seq_len,
high=max_seq_len + 1,
size=(batch_size),
dtype=np.int32)
valid_word_num = sum(input_len)
# fake input id and mask
input_ids = np.random.randint(low=0,
high=bert_config.vocab_size,
size=(batch_size, max_seq_len),
dtype=np.int32)
input_mask = np.zeros((batch_size, max_seq_len), dtype=np.int32)
for b_idx, s_len in enumerate(input_len):
input_mask[b_idx][:s_len] = 1
input_ids_tensor = tf.convert_to_tensor(input_ids, dtype=tf.int32)
input_mask_tensor = tf.convert_to_tensor(input_mask, dtype=tf.int32)
# fake embedding output
embed_output = np.random.randn(batch_size, max_seq_len,
bert_config.hidden_size)
input_tensor = tf.convert_to_tensor(embed_output, dtype=tf_dtype)
# keep attention_mask for compatible reason
att_mask = np.tile(input_mask, max_seq_len)
att_mask = att_mask.reshape(batch_size, max_seq_len, max_seq_len)
attention_mask = tf.convert_to_tensor(att_mask, dtype=tf_dtype)
# input info
valid_word_num = sum(input_len)
print("Valid word num : {}/{}, avg sequence length : {:.6} ".format(
valid_word_num, batch_size * max_seq_len, valid_word_num / batch_size))
# bert with standard transformer
std_bert = modeling.transformer_model(
input_tensor=input_tensor,
attention_mask=attention_mask,
hidden_size=bert_config.hidden_size,
num_hidden_layers=bert_config.num_hidden_layers,
num_attention_heads=bert_config.num_attention_heads,
intermediate_size=bert_config.intermediate_size,
intermediate_act_fn=modeling.get_activation(bert_config.hidden_act),
hidden_dropout_prob=bert_config.hidden_dropout_prob,
attention_probs_dropout_prob=bert_config.attention_probs_dropout_prob,
initializer_range=bert_config.initializer_range,
do_return_all_layers=False)
config = tf.ConfigProto()
config.graph_options.optimizer_options.global_jit_level = tf.OptimizerOptions.ON_1
with tf.Session(config=config) as sess:
# init weights
sess.run(tf.global_variables_initializer())
# get transformer weights
all_vars = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES)
transformer_vars = [v for v in all_vars if v.name.startswith('layer')]
weights_value = sess.run(transformer_vars)
# bert with effective transformer
et_bert = effective_transformer.get_sequence_output(
max_batch_size=batch_size,
max_seq_length=max_seq_len,
config=bert_config,
attention_mask=attention_mask,
input_mask=input_mask_tensor,
from_tensor=input_tensor,
weights_value=weights_value,
)
# diff
val1 = sess.run(std_bert).reshape(-1, 768)
val2 = sess.run(et_bert).reshape(-1, 768)
diff = []
for b_idx, s_len in enumerate(input_len):
for w_idx in range(s_len):
idx = b_idx * args.max_seq_length + w_idx
diff.append(np.fabs(val1[idx] - val2[idx]).max())
print("max diff : {:.6}, avg diff : {:.6}.".format(
max(diff),
sum(diff) / len(diff)))
def time_inference(output_tensor):
iter_num = 128
# warm up
for i in range(10):
sess.run(output_tensor)
beg = datetime.now()
for i in range(iter_num):
sess.run(output_tensor)
end = datetime.now()
return (end - beg).total_seconds() * 1000 / iter_num # ms
print("xla cost : {:.6} ms".format(time_inference(std_bert)))
print("et cost : {:.6} ms".format(time_inference(et_bert)))
def __call__(self, features, hidden_feature, mode, problem_name):
"""Get loss and log probs for the masked LM.
DO NOT CHANGE THE VARAIBLE SCOPE.
"""
seq_hidden_feature = hidden_feature['seq']
positions = features['masked_lm_positions']
input_tensor = gather_indexes(seq_hidden_feature, positions)
output_weights = hidden_feature['embed_table']
label_ids = features['masked_lm_ids']
label_weights = features['masked_lm_weights']
with tf.variable_scope("cls/predictions"):
# We apply one more non-linear transformation before the output layer.
# This matrix is not used after pre-training.
with tf.variable_scope("transform"):
input_tensor = tf.layers.dense(
input_tensor,
units=self.params.mask_lm_hidden_size,
activation=modeling.get_activation(
self.params.mask_lm_hidden_act),
kernel_initializer=modeling.create_initializer(
self.params.mask_lm_initializer_range))
input_tensor = modeling.layer_norm(input_tensor)
# The output weights are the same as the input embeddings, but there is
# an output-only bias for each token.
output_bias = tf.get_variable("output_bias",
shape=[self.params.vocab_size],
initializer=tf.zeros_initializer())
logits = tf.matmul(input_tensor, output_weights, transpose_b=True)
logits = tf.nn.bias_add(logits, output_bias)
self.logits = logits
log_probs = tf.nn.log_softmax(logits, axis=-1)
if mode == tf.estimator.ModeKeys.PREDICT:
self.prob = log_probs
return self.prob
else:
label_ids = tf.reshape(label_ids, [-1])
label_weights = tf.reshape(label_weights, [-1])
one_hot_labels = tf.one_hot(label_ids,
depth=self.params.vocab_size,
dtype=tf.float32)
# The `positions` tensor might be zero-padded (if the sequence is too
# short to have the maximum number of predictions). The `label_weights`
# tensor has a value of 1.0 for every real prediction and 0.0 for the
# padding predictions.
per_example_loss = - \
tf.reduce_sum(log_probs * one_hot_labels, axis=[-1])
numerator = tf.reduce_sum(label_weights * per_example_loss)
denominator = tf.reduce_sum(label_weights) + 1e-5
loss = numerator / denominator
if mode == tf.estimator.ModeKeys.TRAIN:
self.loss = loss
return self.loss
else:
def metric_fn(masked_lm_example_loss, masked_lm_log_probs,
masked_lm_ids, masked_lm_weights):
"""Computes the loss and accuracy of the model."""
masked_lm_log_probs = tf.reshape(
masked_lm_log_probs,
[-1, masked_lm_log_probs.shape[-1]])
masked_lm_predictions = tf.argmax(masked_lm_log_probs,
axis=-1,
output_type=tf.int32)
masked_lm_example_loss = tf.reshape(
masked_lm_example_loss, [-1])
masked_lm_ids = tf.reshape(masked_lm_ids, [-1])
masked_lm_weights = tf.reshape(masked_lm_weights, [-1])
masked_lm_accuracy = tf.metrics.accuracy(
labels=masked_lm_ids,
predictions=masked_lm_predictions,
weights=masked_lm_weights)
masked_lm_mean_loss = tf.metrics.mean(
values=masked_lm_example_loss,
weights=masked_lm_weights)
return {
"masked_lm_accuracy": masked_lm_accuracy,
"masked_lm_loss": masked_lm_mean_loss,
}
eval_metrics = (metric_fn(per_example_loss, log_probs,
label_ids, label_weights), loss)
self.eval_metrics = eval_metrics
return self.eval_metrics
def aggregate_embedding(embeddings,
segment_idx,
aggregator,
config=None,
aux=None,
name=None):
# segment_idx denotes different needles, rather than rows.
if aggregator == 'segment_sqrt_n':
denom = to_col(
tf.sqrt(
tf.to_float(
tf.segment_sum(tf.ones_like(segment_idx), segment_idx))))
output_layer = tf.div_no_nan(tf.segment_sum(embeddings, segment_idx),
denom,
name=name)
elif aggregator in ['segment_sum', 'segment_mean']:
output_layer = getattr(tf, aggregator)(embeddings,
segment_idx,
name=name)
else:
del embeddings
assert aggregator.startswith('transformer')
flags = {}
if '^' in aggregator:
flags = [
kv.split('@')
for kv in filter(None,
aggregator.split('^')[1].split(','))
]
flags = {k: eval(v) for k, v in flags}
assert config is not None and aux is not None
needle_pos = aux['needle_pos']
embedding_output = aux['sequence_output']
batch_idx2 = aux['batch_idx2'] # different rows.
is_training = aux['is_training']
attention_mask = get_dense_mask(needle_pos, batch_idx2,
tf.shape(embedding_output)[:2])
with tf.variable_scope('final_transformer'):
all_encoder_layers = modeling.transformer_model(
input_tensor=embedding_output,
attention_mask=attention_mask,
hidden_size=config.
hidden_size, # this must agree with input width.
num_hidden_layers=flags.get('num_hidden_layers', 1),
num_attention_heads=flags.get('num_attention_heads',
config.num_attention_heads),
intermediate_size=flags.get('intermediate_size',
config.intermediate_size),
intermediate_act_fn=modeling.get_activation(config.hidden_act),
hidden_dropout_prob=flags.get('hidden_dropout_prob',
config.hidden_dropout_prob) *
int(is_training),
attention_probs_dropout_prob=int(is_training) *
flags.get('attention_probs_dropout_prob',
config.attention_probs_dropout_prob),
initializer_range=config.initializer_range,
do_return_all_layers=True)
first_token_tensor = all_encoder_layers[-1][:, 0, :]
output_layer = tf.layers.dense(
first_token_tensor,
config.hidden_size,
activation=tf.tanh,
kernel_initializer=modeling.create_initializer(
config.initializer_range))
return output_layer
def __init__(self,
config,
use_one_hot_embeddings=True,
num_labels=2,
max_seq_length=128):
"""Constructor for BertModel.
Args:
config: `BertConfig` instance.
is_training: bool. true for training model, false for eval model. Controls
whether dropout will be applied.
input_ids: int32 Tensor of shape [batch_size, seq_length].
input_mask: (optional) int32 Tensor of shape [batch_size, seq_length].
token_type_ids: (optional) int32 Tensor of shape [batch_size, seq_length].
use_one_hot_embeddings: (optional) bool. Whether to use one-hot word
embeddings or tf.embedding_lookup() for the word embeddings. On the TPU,
it is much faster if this is True, on the CPU or GPU, it is faster if
this is False.
scope: (optional) variable scope. Defaults to "bert".
Raises:
ValueError: The config is invalid or one of the input tensor shapes
is invalid.
"""
self.input_ids = tf.placeholder(dtype=tf.int32,
shape=(None, max_seq_length))
self.input_mask = tf.placeholder(dtype=tf.int8,
shape=(None, max_seq_length))
config = copy.deepcopy(config)
input_shape = modeling.get_shape_list(self.input_ids, expected_rank=2)
batch_size = input_shape[0]
seq_length = input_shape[1]
token_type_ids = tf.zeros(shape=[batch_size, seq_length],
dtype=tf.int32)
with tf.variable_scope("bert", reuse=tf.AUTO_REUSE):
with tf.variable_scope("embeddings", reuse=tf.AUTO_REUSE):
# Perform embedding lookup on the word ids.
(self.embedding_output,
self.embedding_table) = modeling.embedding_lookup(
input_ids=self.input_ids,
vocab_size=config.vocab_size,
embedding_size=config.hidden_size,
initializer_range=config.initializer_range,
word_embedding_name="word_embeddings",
use_one_hot_embeddings=use_one_hot_embeddings)
# Add positional embeddings and token type embeddings, then layer
# normalize and perform dropout.
self.embedding_output = modeling.embedding_postprocessor(
input_tensor=self.embedding_output,
use_token_type=True,
token_type_ids=token_type_ids,
token_type_vocab_size=config.type_vocab_size,
token_type_embedding_name="token_type_embeddings",
use_position_embeddings=True,
position_embedding_name="position_embeddings",
initializer_range=config.initializer_range,
max_position_embeddings=config.max_position_embeddings,
dropout_prob=config.hidden_dropout_prob)
with tf.variable_scope("encoder", reuse=tf.AUTO_REUSE):
# This converts a 2D mask of shape [batch_size, seq_length] to a 3D
# mask of shape [batch_size, seq_length, seq_length] which is used
# for the attention scores.
attention_mask = modeling.create_attention_mask_from_input_mask(
self.input_ids, self.input_mask)
# Run the stacked transformer.
# `sequence_output` shape = [batch_size, seq_length, hidden_size].
self.all_encoder_layers = modeling.transformer_model(
input_tensor=self.embedding_output,
attention_mask=attention_mask,
hidden_size=config.hidden_size,
num_hidden_layers=config.num_hidden_layers,
num_attention_heads=config.num_attention_heads,
intermediate_size=config.intermediate_size,
intermediate_act_fn=modeling.get_activation(
config.hidden_act),
hidden_dropout_prob=config.hidden_dropout_prob,
attention_probs_dropout_prob=config.
attention_probs_dropout_prob,
initializer_range=config.initializer_range,
do_return_all_layers=True)
self.sequence_output = self.all_encoder_layers[-1]
# The "pooler" converts the encoded sequence tensor of shape
# [batch_size, seq_length, hidden_size] to a tensor of shape
# [batch_size, hidden_size]. This is necessary for segment-level
# (or segment-pair-level) classification tasks where we need a fixed
# dimensional representation of the segment.
with tf.variable_scope("pooler", reuse=tf.AUTO_REUSE):
# We "pool" the model by simply taking the hidden state corresponding
# to the first token. We assume that this has been pre-trained
first_token_tensor = tf.squeeze(self.sequence_output[:,
0:1, :],
axis=1)
self.pooled_output = tf.layers.dense(
first_token_tensor,
config.hidden_size,
activation=tf.tanh,
kernel_initializer=modeling.create_initializer(
config.initializer_range))
# define output_weights and output_bias
hidden_size = self.pooled_output.shape[-1].value
with tf.variable_scope("", reuse=tf.AUTO_REUSE):
self.output_weights = tf.get_variable(
"output_weights", [num_labels, hidden_size],
initializer=tf.truncated_normal_initializer(stddev=0.02))
self.output_bias = tf.get_variable(
"output_bias", [num_labels],
initializer=tf.zeros_initializer())
def create_bilstm_classification_model(bert_config,
is_training,
response_input_ids,
response_input_mask,
response_segment_ids,
response_text_len,
response_labels,
random_forward_input_ids,
random_forward_input_mask,
random_forward_segment_ids,
random_forward_text_len,
random_backward_input_ids,
random_backward_input_mask,
random_backward_segment_ids,
random_backward_text_len,
random_labels,
swap_forward_input_ids,
swap_forward_input_mask,
swap_forward_segment_ids,
swap_forward_text_len,
swap_backward_input_ids,
swap_backward_input_mask,
swap_backward_segment_ids,
swap_backward_text_len,
swap_labels,
nli_forward_input_ids,
nli_forward_input_mask,
nli_forward_segment_ids,
nli_forward_text_len,
nli_backward_input_ids,
nli_backward_input_mask,
nli_backward_segment_ids,
nli_backward_text_len,
nli_labels,
num_nli_labels,
use_one_hot_embeddings,
l2_reg_lambda=0.1,
dropout_rate=1.0,
lstm_size=None,
num_layers=1):
config = copy.deepcopy(bert_config)
if not is_training:
config.hidden_dropout_prob = 0.0
config.attention_probs_dropout_prob = 0.0
with tf.variable_scope("bert", reuse=tf.AUTO_REUSE):
with tf.variable_scope("embeddings", reuse=tf.AUTO_REUSE):
(response_embedding_output,
response_embedding_table) = modeling.embedding_lookup(
input_ids=response_input_ids,
vocab_size=config.vocab_size,
embedding_size=config.hidden_size,
initializer_range=config.initializer_range,
word_embedding_name="word_embeddings",
use_one_hot_embeddings=use_one_hot_embeddings)
response_embedding_output = modeling.embedding_postprocessor(
input_tensor=response_embedding_output,
use_token_type=not config.roberta,
token_type_ids=response_segment_ids,
token_type_vocab_size=config.type_vocab_size,
token_type_embedding_name="token_type_embeddings",
use_position_embeddings=True,
position_embedding_name="position_embeddings",
initializer_range=config.initializer_range,
max_position_embeddings=config.max_position_embeddings,
dropout_prob=config.hidden_dropout_prob,
roberta=config.roberta)
# random detection
# Perform embedding lookup on the word ids.
(random_foward_embedding_output,
random_forward_embedding_table) = modeling.embedding_lookup(
input_ids=random_forward_input_ids,
vocab_size=config.vocab_size,
embedding_size=config.hidden_size,
initializer_range=config.initializer_range,
word_embedding_name="word_embeddings",
use_one_hot_embeddings=use_one_hot_embeddings)
# Perform embedding lookup on the word ids.
(random_backward_embedding_output,
random_backward_embedding_table) = modeling.embedding_lookup(
input_ids=random_backward_input_ids,
vocab_size=config.vocab_size,
embedding_size=config.hidden_size,
initializer_range=config.initializer_range,
word_embedding_name="word_embeddings",
use_one_hot_embeddings=use_one_hot_embeddings)
# Add positional embeddings and token type embeddings, then layer
# normalize and perform dropout.
random_foward_embedding_output = modeling.embedding_postprocessor(
input_tensor=random_foward_embedding_output,
use_token_type=not config.roberta,
token_type_ids=random_forward_segment_ids,
token_type_vocab_size=config.type_vocab_size,
token_type_embedding_name="token_type_embeddings",
use_position_embeddings=True,
position_embedding_name="position_embeddings",
initializer_range=config.initializer_range,
max_position_embeddings=config.max_position_embeddings,
dropout_prob=config.hidden_dropout_prob,
roberta=config.roberta)
random_backward_embedding_output = modeling.embedding_postprocessor(
input_tensor=random_backward_embedding_output,
use_token_type=not config.roberta,
token_type_ids=random_backward_segment_ids,
token_type_vocab_size=config.type_vocab_size,
token_type_embedding_name="token_type_embeddings",
use_position_embeddings=True,
position_embedding_name="position_embeddings",
initializer_range=config.initializer_range,
max_position_embeddings=config.max_position_embeddings,
dropout_prob=config.hidden_dropout_prob,
roberta=config.roberta)
# swap detection
(swap_foward_embedding_output,
swap_forward_embedding_table) = modeling.embedding_lookup(
input_ids=swap_forward_input_ids,
vocab_size=config.vocab_size,
embedding_size=config.hidden_size,
initializer_range=config.initializer_range,
word_embedding_name="word_embeddings",
use_one_hot_embeddings=use_one_hot_embeddings)
(swap_backward_embedding_output,
swap_backward_embedding_table) = modeling.embedding_lookup(
input_ids=swap_backward_input_ids,
vocab_size=config.vocab_size,
embedding_size=config.hidden_size,
initializer_range=config.initializer_range,
word_embedding_name="word_embeddings",
use_one_hot_embeddings=use_one_hot_embeddings)
swap_foward_embedding_output = modeling.embedding_postprocessor(
input_tensor=swap_foward_embedding_output,
use_token_type=not config.roberta,
token_type_ids=swap_forward_segment_ids,
token_type_vocab_size=config.type_vocab_size,
token_type_embedding_name="token_type_embeddings",
use_position_embeddings=True,
position_embedding_name="position_embeddings",
initializer_range=config.initializer_range,
max_position_embeddings=config.max_position_embeddings,
dropout_prob=config.hidden_dropout_prob,
roberta=config.roberta)
swap_backward_embedding_output = modeling.embedding_postprocessor(
input_tensor=swap_backward_embedding_output,
use_token_type=not config.roberta,
token_type_ids=swap_backward_segment_ids,
token_type_vocab_size=config.type_vocab_size,
token_type_embedding_name="token_type_embeddings",
use_position_embeddings=True,
position_embedding_name="position_embeddings",
initializer_range=config.initializer_range,
max_position_embeddings=config.max_position_embeddings,
dropout_prob=config.hidden_dropout_prob,
roberta=config.roberta)
# # generic detection
# (generic_foward_embedding_output, generic_forward_embedding_table) = modeling.embedding_lookup(
# input_ids=generic_forward_input_ids,
# vocab_size=config.vocab_size,
# embedding_size=config.hidden_size,
# initializer_range=config.initializer_range,
# word_embedding_name="word_embeddings",
# use_one_hot_embeddings=use_one_hot_embeddings)
# (generic_backward_embedding_output, generic_backward_embedding_table) = modeling.embedding_lookup(
# input_ids=generic_backward_input_ids,
# vocab_size=config.vocab_size,
# embedding_size=config.hidden_size,
# initializer_range=config.initializer_range,
# word_embedding_name="word_embeddings",
# use_one_hot_embeddings=use_one_hot_embeddings)
# generic_foward_embedding_output = modeling.embedding_postprocessor(
# input_tensor=generic_foward_embedding_output,
# use_token_type=not config.roberta,
# token_type_ids=generic_forward_segment_ids,
# token_type_vocab_size=config.type_vocab_size,
# token_type_embedding_name="token_type_embeddings",
# use_position_embeddings=True,
# position_embedding_name="position_embeddings",
# initializer_range=config.initializer_range,
# max_position_embeddings=config.max_position_embeddings,
# dropout_prob=config.hidden_dropout_prob,
# roberta=config.roberta)
# generic_backward_embedding_output = modeling.embedding_postprocessor(
# input_tensor=generic_backward_embedding_output,
# use_token_type=not config.roberta,
# token_type_ids=generic_backward_segment_ids,
# token_type_vocab_size=config.type_vocab_size,
# token_type_embedding_name="token_type_embeddings",
# use_position_embeddings=True,
# position_embedding_name="position_embeddings",
# initializer_range=config.initializer_range,
# max_position_embeddings=config.max_position_embeddings,
# dropout_prob=config.hidden_dropout_prob,
# roberta=config.roberta)
# nli detection
(nli_foward_embedding_output,
nli_forward_embedding_table) = modeling.embedding_lookup(
input_ids=nli_forward_input_ids,
vocab_size=config.vocab_size,
embedding_size=config.hidden_size,
initializer_range=config.initializer_range,
word_embedding_name="word_embeddings",
use_one_hot_embeddings=use_one_hot_embeddings)
(nli_backward_embedding_output,
nli_backward_embedding_table) = modeling.embedding_lookup(
input_ids=nli_backward_input_ids,
vocab_size=config.vocab_size,
embedding_size=config.hidden_size,
initializer_range=config.initializer_range,
word_embedding_name="word_embeddings",
use_one_hot_embeddings=use_one_hot_embeddings)
nli_foward_embedding_output = modeling.embedding_postprocessor(
input_tensor=nli_foward_embedding_output,
use_token_type=not config.roberta,
token_type_ids=nli_forward_segment_ids,
token_type_vocab_size=config.type_vocab_size,
token_type_embedding_name="token_type_embeddings",
use_position_embeddings=True,
position_embedding_name="position_embeddings",
initializer_range=config.initializer_range,
max_position_embeddings=config.max_position_embeddings,
dropout_prob=config.hidden_dropout_prob,
roberta=config.roberta)
nli_backward_embedding_output = modeling.embedding_postprocessor(
input_tensor=nli_backward_embedding_output,
use_token_type=not config.roberta,
token_type_ids=nli_backward_segment_ids,
token_type_vocab_size=config.type_vocab_size,
token_type_embedding_name="token_type_embeddings",
use_position_embeddings=True,
position_embedding_name="position_embeddings",
initializer_range=config.initializer_range,
max_position_embeddings=config.max_position_embeddings,
dropout_prob=config.hidden_dropout_prob,
roberta=config.roberta)
with tf.variable_scope("encoder", reuse=tf.AUTO_REUSE):
response_attention_mask = modeling.create_attention_mask_from_input_mask(
response_input_ids, response_input_mask)
# [batch_size, from_seq_length, to_seq_length]
# mask future tokens
diag_vals = tf.ones_like(response_attention_mask[0, :, :])
tril = tf.linalg.LinearOperatorLowerTriangular(
diag_vals).to_dense()
future_masks = tf.tile(tf.expand_dims(
tril, 0), [tf.shape(response_attention_mask)[0], 1, 1])
response_attention_mask = tf.math.multiply(response_attention_mask,
future_masks)
# Run the stacked transformer.
# `sequence_output` shape = [batch_size, seq_length, hidden_size].
response_all_encoder_layers = modeling.transformer_model(
input_tensor=response_embedding_output,
attention_mask=response_attention_mask,
hidden_size=config.hidden_size,
num_hidden_layers=config.num_hidden_layers,
num_attention_heads=config.num_attention_heads,
intermediate_size=config.intermediate_size,
intermediate_act_fn=modeling.get_activation(config.hidden_act),
hidden_dropout_prob=config.hidden_dropout_prob,
attention_probs_dropout_prob=config.
attention_probs_dropout_prob,
initializer_range=config.initializer_range,
do_return_all_layers=True)
# random detection
# This converts a 2D mask of shape [batch_size, seq_length] to a 3D
# mask of shape [batch_size, seq_length, seq_length] which is used
# for the attention scores.
random_forward_attention_mask = modeling.create_attention_mask_from_input_mask(
random_forward_input_ids, random_forward_input_mask)
random_backward_attention_mask = modeling.create_attention_mask_from_input_mask(
random_backward_input_ids, random_backward_input_mask)
# Run the stacked transformer.
# `sequence_output` shape = [batch_size, seq_length, hidden_size].
random_forward_all_encoder_layers = modeling.transformer_model(
input_tensor=random_foward_embedding_output,
attention_mask=random_forward_attention_mask,
hidden_size=config.hidden_size,
num_hidden_layers=config.num_hidden_layers,
num_attention_heads=config.num_attention_heads,
intermediate_size=config.intermediate_size,
intermediate_act_fn=modeling.get_activation(config.hidden_act),
hidden_dropout_prob=config.hidden_dropout_prob,
attention_probs_dropout_prob=config.
attention_probs_dropout_prob,
initializer_range=config.initializer_range,
do_return_all_layers=True)
random_backward_all_encoder_layers = modeling.transformer_model(
input_tensor=random_backward_embedding_output,
attention_mask=random_backward_attention_mask,
hidden_size=config.hidden_size,
num_hidden_layers=config.num_hidden_layers,
num_attention_heads=config.num_attention_heads,
intermediate_size=config.intermediate_size,
intermediate_act_fn=modeling.get_activation(config.hidden_act),
hidden_dropout_prob=config.hidden_dropout_prob,
attention_probs_dropout_prob=config.
attention_probs_dropout_prob,
initializer_range=config.initializer_range,
do_return_all_layers=True)
# swap detection
swap_forward_attention_mask = modeling.create_attention_mask_from_input_mask(
swap_forward_input_ids, swap_forward_input_mask)
swap_backward_attention_mask = modeling.create_attention_mask_from_input_mask(
swap_backward_input_ids, swap_backward_input_mask)
swap_forward_all_encoder_layers = modeling.transformer_model(
input_tensor=swap_foward_embedding_output,
attention_mask=swap_forward_attention_mask,
hidden_size=config.hidden_size,
num_hidden_layers=config.num_hidden_layers,
num_attention_heads=config.num_attention_heads,
intermediate_size=config.intermediate_size,
intermediate_act_fn=modeling.get_activation(config.hidden_act),
hidden_dropout_prob=config.hidden_dropout_prob,
attention_probs_dropout_prob=config.
attention_probs_dropout_prob,
initializer_range=config.initializer_range,
do_return_all_layers=True)
swap_backward_all_encoder_layers = modeling.transformer_model(
input_tensor=swap_backward_embedding_output,
attention_mask=swap_backward_attention_mask,
hidden_size=config.hidden_size,
num_hidden_layers=config.num_hidden_layers,
num_attention_heads=config.num_attention_heads,
intermediate_size=config.intermediate_size,
intermediate_act_fn=modeling.get_activation(config.hidden_act),
hidden_dropout_prob=config.hidden_dropout_prob,
attention_probs_dropout_prob=config.
attention_probs_dropout_prob,
initializer_range=config.initializer_range,
do_return_all_layers=True)
# # generic detection
# generic_forward_attention_mask = modeling.create_attention_mask_from_input_mask(generic_forward_input_ids,
# generic_forward_input_mask)
# generic_backward_attention_mask = modeling.create_attention_mask_from_input_mask(generic_backward_input_ids,
# generic_backward_input_mask)
# generic_forward_all_encoder_layers = modeling.transformer_model(
# input_tensor=generic_foward_embedding_output,
# attention_mask=generic_forward_attention_mask,
# hidden_size=config.hidden_size,
# num_hidden_layers=config.num_hidden_layers,
# num_attention_heads=config.num_attention_heads,
# intermediate_size=config.intermediate_size,
# intermediate_act_fn=modeling.get_activation(config.hidden_act),
# hidden_dropout_prob=config.hidden_dropout_prob,
# attention_probs_dropout_prob=config.attention_probs_dropout_prob,
# initializer_range=config.initializer_range,
# do_return_all_layers=True)
# generic_backward_all_encoder_layers = modeling.transformer_model(
# input_tensor=generic_backward_embedding_output,
# attention_mask=generic_backward_attention_mask,
# hidden_size=config.hidden_size,
# num_hidden_layers=config.num_hidden_layers,
# num_attention_heads=config.num_attention_heads,
# intermediate_size=config.intermediate_size,
# intermediate_act_fn=modeling.get_activation(config.hidden_act),
# hidden_dropout_prob=config.hidden_dropout_prob,
# attention_probs_dropout_prob=config.attention_probs_dropout_prob,
# initializer_range=config.initializer_range,
# do_return_all_layers=True)
# nli detection
nli_forward_attention_mask = modeling.create_attention_mask_from_input_mask(
nli_forward_input_ids, nli_forward_input_mask)
nli_backward_attention_mask = modeling.create_attention_mask_from_input_mask(
nli_backward_input_ids, nli_backward_input_mask)
nli_forward_all_encoder_layers = modeling.transformer_model(
input_tensor=nli_foward_embedding_output,
attention_mask=nli_forward_attention_mask,
hidden_size=config.hidden_size,
num_hidden_layers=config.num_hidden_layers,
num_attention_heads=config.num_attention_heads,
intermediate_size=config.intermediate_size,
intermediate_act_fn=modeling.get_activation(config.hidden_act),
hidden_dropout_prob=config.hidden_dropout_prob,
attention_probs_dropout_prob=config.
attention_probs_dropout_prob,
initializer_range=config.initializer_range,
do_return_all_layers=True)
nli_backward_all_encoder_layers = modeling.transformer_model(
input_tensor=nli_backward_embedding_output,
attention_mask=nli_backward_attention_mask,
hidden_size=config.hidden_size,
num_hidden_layers=config.num_hidden_layers,
num_attention_heads=config.num_attention_heads,
intermediate_size=config.intermediate_size,
intermediate_act_fn=modeling.get_activation(config.hidden_act),
hidden_dropout_prob=config.hidden_dropout_prob,
attention_probs_dropout_prob=config.
attention_probs_dropout_prob,
initializer_range=config.initializer_range,
do_return_all_layers=True)
random_forward_embedding = random_forward_all_encoder_layers[-2]
random_backward_embedding = random_backward_all_encoder_layers[-2]
swap_forward_embedding = swap_forward_all_encoder_layers[-2]
swap_backward_embedding = swap_backward_all_encoder_layers[-2]
# generic_forward_embedding = generic_forward_all_encoder_layers[-2]
# generic_backward_embedding = generic_backward_all_encoder_layers[-2]
nli_forward_embedding = nli_forward_all_encoder_layers[-2]
nli_backward_embedding = nli_backward_all_encoder_layers[-2]
response_embedding = response_all_encoder_layers[-2]
response_embedding_shape = modeling.get_shape_list(response_embedding,
expected_rank=3)
with tf.variable_scope("lm_head", reuse=tf.AUTO_REUSE):
response_logits = tf.layers.dense(response_embedding,
config.hidden_size,
activation=None)
response_logits = modeling.gelu(response_logits)
response_logits = modeling.layer_norm(response_logits)
response_outputs = tf.layers.dense(
response_logits,
config.vocab_size,
activation=None,
use_bias=True,
bias_initializer=tf.zeros_initializer())
response_one_hot = tf.one_hot(response_labels,
depth=config.vocab_size,
dtype=tf.float32)
lm_cost = tf.nn.softmax_cross_entropy_with_logits(
labels=response_one_hot, logits=response_outputs)
sequence_mask = tf.sequence_mask(response_text_len,
maxlen=response_embedding_shape[1],
dtype=tf.float32)
masked_lm_cost = tf.math.multiply(lm_cost, sequence_mask)
final_lm_loss = tf.reduce_mean(
tf.math.divide(tf.reduce_sum(masked_lm_cost, axis=1),
tf.cast(response_text_len, dtype=tf.float32)))
perplexity = tf.exp(
tf.math.divide(tf.reduce_sum(masked_lm_cost, axis=1),
tf.cast(response_text_len, dtype=tf.float32)))
random_forward_embedding_shape = modeling.get_shape_list(
random_forward_embedding, expected_rank=3)
random_backward_embedding_shape = modeling.get_shape_list(
random_backward_embedding, expected_rank=3)
assert random_forward_embedding_shape[
2] == random_backward_embedding_shape[2]
random_forward_embedding = tf.transpose(random_forward_embedding,
[1, 0, 2])
random_backward_embedding = tf.transpose(random_backward_embedding,
[1, 0, 2])
random_forward_input_mask = tf.cast(
tf.transpose(random_forward_input_mask, [1, 0]), tf.float32)
random_backward_input_mask = tf.cast(
tf.transpose(random_backward_input_mask, [1, 0]), tf.float32)
swap_forward_embedding_shape = modeling.get_shape_list(
swap_forward_embedding, expected_rank=3)
swap_backward_embedding_shape = modeling.get_shape_list(
swap_backward_embedding, expected_rank=3)
assert swap_forward_embedding_shape[2] == swap_backward_embedding_shape[2]
swap_forward_embedding = tf.transpose(swap_forward_embedding, [1, 0, 2])
swap_backward_embedding = tf.transpose(swap_backward_embedding, [1, 0, 2])
swap_forward_input_mask = tf.cast(
tf.transpose(swap_forward_input_mask, [1, 0]), tf.float32)
swap_backward_input_mask = tf.cast(
tf.transpose(swap_backward_input_mask, [1, 0]), tf.float32)
# generic_forward_embedding_shape = modeling.get_shape_list(generic_forward_embedding, expected_rank=3)
# generic_backward_embedding_shape = modeling.get_shape_list(generic_backward_embedding, expected_rank=3)
# assert generic_forward_embedding_shape[2] == generic_backward_embedding_shape[2]
# generic_forward_embedding = tf.transpose(generic_forward_embedding, [1, 0, 2])
# generic_backward_embedding = tf.transpose(generic_backward_embedding, [1, 0, 2])
# generic_forward_input_mask = tf.cast(tf.transpose(generic_forward_input_mask, [1, 0]), tf.float32)
# generic_backward_input_mask = tf.cast(tf.transpose(generic_backward_input_mask, [1, 0]), tf.float32)
nli_forward_embedding_shape = modeling.get_shape_list(
nli_forward_embedding, expected_rank=3)
nli_backward_embedding_shape = modeling.get_shape_list(
nli_backward_embedding, expected_rank=3)
assert nli_forward_embedding_shape[2] == nli_backward_embedding_shape[2]
nli_forward_embedding = tf.transpose(nli_forward_embedding, [1, 0, 2])
nli_backward_embedding = tf.transpose(nli_backward_embedding, [1, 0, 2])
nli_forward_input_mask = tf.cast(
tf.transpose(nli_forward_input_mask, [1, 0]), tf.float32)
nli_backward_input_mask = tf.cast(
tf.transpose(nli_backward_input_mask, [1, 0]), tf.float32)
model = HadeModel(
x_random_forward=random_forward_embedding,
x_random_mask_forward=random_forward_input_mask,
x_random_length_forward=random_forward_text_len,
x_random_backward=random_backward_embedding,
x_random_mask_backward=random_backward_input_mask,
x_random_length_backward=random_backward_text_len,
y_random=random_labels,
x_swap_forward=swap_forward_embedding,
x_swap_mask_forward=swap_forward_input_mask,
x_swap_length_forward=swap_forward_text_len,
x_swap_backward=swap_backward_embedding,
x_swap_mask_backward=swap_backward_input_mask,
x_swap_length_backward=swap_backward_text_len,
y_swap=swap_labels,
# x_generic_forward=generic_forward_embedding,
# x_generic_mask_forward=generic_forward_input_mask,
# x_generic_length_forward=generic_forward_text_len,
# x_generic_backward=generic_backward_embedding,
# x_generic_mask_backward=generic_backward_input_mask,
# x_generic_length_backward=generic_backward_text_len, y_generic=generic_labels,
x_nli_forward=nli_forward_embedding,
x_nli_mask_forward=nli_forward_input_mask,
x_nli_length_forward=nli_forward_text_len,
x_nli_backward=nli_backward_embedding,
x_nli_mask_backward=nli_backward_input_mask,
x_nli_length_backward=nli_backward_text_len,
y_nli=nli_labels,
embedding_dim=random_forward_embedding_shape[2],
num_nli_labels=num_nli_labels,
hidden_size=lstm_size,
l2_reg_lambda=l2_reg_lambda,
num_layers=num_layers,
dropout_rate=dropout_rate,
is_training=is_training)
random_prob, swap_prob, nli_prob, total_cost = model.create_model()
return random_prob, swap_prob, nli_prob, total_cost, final_lm_loss, perplexity
