def call(self, inputs, mode="bert", **kwargs):
"""Implements call() for the layer.
Args:
inputs: packed input tensors.
mode: string, `bert` or `encoder`.
Returns:
Output tensor of the last layer for BERT training (mode=`bert`) which
is a float Tensor of shape [batch_size, seq_length, hidden_size] or
a list of output tensors for encoder usage (mode=`encoder`).
"""
unpacked_inputs = tf_utils.unpack_inputs(inputs)
input_word_ids = unpacked_inputs[0]
input_mask = unpacked_inputs[1]
input_type_ids = unpacked_inputs[2]
word_embeddings = self.embedding_lookup(input_word_ids)
embedding_tensor = self.embedding_postprocessor(
word_embeddings=word_embeddings, token_type_ids=input_type_ids)
if self.float_type == tf.float16:
embedding_tensor = tf.cast(embedding_tensor, tf.float16)
attention_mask = None
if input_mask is not None:
attention_mask = create_attention_mask_from_input_mask(
input_word_ids, input_mask)
if mode == "encoder":
return self.encoder(embedding_tensor,
attention_mask,
return_all_layers=True)
sequence_output = self.encoder(embedding_tensor, attention_mask)
first_token_tensor = tf.squeeze(sequence_output[:, 0:1, :], axis=1)
pooled_output = self.pooler_transform(first_token_tensor)
return (pooled_output, sequence_output)
def call(self, inputs, **kwargs):
"""Implements call() for the layer."""
(input_tensor, attention_mask) = tf_utils.unpack_inputs(inputs)
attention_output = self.attention_layer(from_tensor=input_tensor,
to_tensor=input_tensor,
attention_mask=attention_mask,
**kwargs)
attention_output = self.attention_output_dense(attention_output)
attention_output = self.attention_dropout(attention_output,
training=kwargs.get(
'training', False))
# Use float32 in keras layer norm and the gelu activation in the
# intermediate dense layer for numeric stability
attention_output = self.attention_layer_norm(input_tensor +
attention_output)
if self.float_type == tf.float16:
attention_output = tf.cast(attention_output, tf.float16)
intermediate_output = self.intermediate_dense(attention_output)
if self.float_type == tf.float16:
intermediate_output = tf.cast(intermediate_output, tf.float16)
layer_output = self.output_dense(intermediate_output)
layer_output = self.output_dropout(layer_output,
training=kwargs.get(
'training', False))
# Use float32 in keras layer norm for numeric stability
layer_output = self.output_layer_norm(layer_output + attention_output)
if self.float_type == tf.float16:
layer_output = tf.cast(layer_output, tf.float16)
return layer_output
def call(self, inputs, return_all_layers=False, **kwargs):
"""Implements call() for the layer.
Args:
inputs: packed inputs.
return_all_layers: bool, whether to return outputs of all layers inside
encoders.
Returns:
Output tensor of the last layer or a list of output tensors.
"""
unpacked_inputs = tf_utils.unpack_inputs(inputs)
input_tensor = unpacked_inputs[0]
attention_mask = unpacked_inputs[1]
output_tensor = input_tensor
all_layer_outputs = []
for i in range(self.num_hidden_layers):
output_tensor = self.shared_layer(output_tensor, attention_mask,
**kwargs)
all_layer_outputs.append(output_tensor)
if return_all_layers:
return all_layer_outputs
return all_layer_outputs[-1]
def call(self, inputs):
"""Implements call() for the layer."""
unpacked_inputs = tf_utils.unpack_inputs(inputs)
lm_output = unpacked_inputs[0]
sentence_output = unpacked_inputs[1]
lm_label_ids = unpacked_inputs[2]
lm_label_ids = tf.keras.backend.reshape(lm_label_ids, [-1])
lm_label_ids_one_hot = tf.keras.backend.one_hot(lm_label_ids,
self.config.vocab_size)
lm_label_weights = tf.keras.backend.cast(unpacked_inputs[3], tf.float32)
lm_label_weights = tf.keras.backend.reshape(lm_label_weights, [-1])
lm_per_example_loss = -tf.keras.backend.sum(
lm_output * lm_label_ids_one_hot, axis=[-1])
numerator = tf.keras.backend.sum(lm_label_weights * lm_per_example_loss)
denominator = tf.keras.backend.sum(lm_label_weights) + 1e-5
mask_label_loss = numerator / denominator
sentence_labels = unpacked_inputs[4]
sentence_labels = tf.keras.backend.reshape(sentence_labels, [-1])
sentence_label_one_hot = tf.keras.backend.one_hot(sentence_labels, 2)
per_example_loss_sentence = -tf.keras.backend.sum(
sentence_label_one_hot * sentence_output, axis=-1)
sentence_loss = tf.keras.backend.mean(per_example_loss_sentence)
loss = mask_label_loss + sentence_loss
# TODO(hongkuny): Avoids the hack and switches add_loss.
final_loss = tf.fill(
tf.keras.backend.shape(per_example_loss_sentence), loss)
self._add_metrics(lm_output, lm_label_ids, lm_label_weights,
lm_per_example_loss, sentence_output, sentence_labels,
per_example_loss_sentence)
return final_loss
def call(self, inputs, **kwargs):
"""Implements call() for the layer."""
(from_tensor, to_tensor,
attention_mask) = tf_utils.unpack_inputs(inputs)
# Scalar dimensions referenced here:
# B = batch size (number of sequences)
# F = `from_tensor` sequence length
# T = `to_tensor` sequence length
# N = `num_attention_heads`
# H = `size_per_head`
# `query_tensor` = [B, F, N ,H]
query_tensor = self.query_dense(from_tensor)
# `key_tensor` = [B, T, N, H]
key_tensor = self.key_dense(to_tensor)
# `value_tensor` = [B, T, N, H]
value_tensor = self.value_dense(to_tensor)
# Take the dot product between "query" and "key" to get the raw
# attention scores.
attention_scores = tf.einsum("BTNH,BFNH->BNFT", key_tensor,
query_tensor)
attention_scores = tf.multiply(
attention_scores, 1.0 / math.sqrt(float(self.size_per_head)))
if attention_mask is not None:
# `attention_mask` = [B, 1, F, T]
attention_mask = tf.expand_dims(attention_mask, axis=[1])
# Since attention_mask is 1.0 for positions we want to attend and 0.0 for
# masked positions, this operation will create a tensor which is 0.0 for
# positions we want to attend and -10000.0 for masked positions.
adder = (1.0 - tf.cast(attention_mask,
attention_scores.dtype)) * -10000.0
# Since we are adding it to the raw scores before the softmax, this is
# effectively the same as removing these entirely.
attention_scores += adder
# Normalize the attention scores to probabilities.
# `attention_probs` = [B, N, F, T]
attention_probs = tf.nn.softmax(attention_scores)
# This is actually dropping out entire tokens to attend to, which might
# seem a bit unusual, but is taken from the original Transformer paper.
attention_probs = self.attention_probs_dropout(attention_probs,
training=kwargs.get(
'training', False))
# `context_layer` = [B, F, N, H]
context_tensor = tf.einsum("BNFT,BTNH->BFNH", attention_probs,
value_tensor)
return context_tensor
def call(self, inputs):
"""Implements call() for the layer."""
unpacked_inputs = tf_utils.unpack_inputs(inputs)
pooled_output = unpacked_inputs[0]
sequence_output = unpacked_inputs[1]
masked_lm_positions = unpacked_inputs[2]
mask_lm_input_tensor = gather_indexes(sequence_output, masked_lm_positions)
lm_output = self.lm_dense(mask_lm_input_tensor)
lm_output = self.lm_layer_norm(lm_output)
lm_output = tf.matmul(lm_output, self.embedding_table, transpose_b=True)
lm_output = tf.nn.bias_add(lm_output, self.output_bias)
lm_output = tf.nn.log_softmax(lm_output, axis=-1)
return lm_output
def call(self, inputs):
"""Implements call() for the layer."""
unpacked_inputs = tf_utils.unpack_inputs(inputs)
lm_output = unpacked_inputs[0]
lm_label_ids = unpacked_inputs[1]
lm_label_ids = tf.keras.backend.reshape(lm_label_ids, [-1])
lm_label_ids_one_hot = tf.keras.backend.one_hot(lm_label_ids,
self.config.vocab_size)
lm_label_weights = tf.keras.backend.cast(unpacked_inputs[2], tf.float32)
lm_label_weights = tf.keras.backend.reshape(lm_label_weights, [-1])
lm_per_example_loss = -tf.keras.backend.sum(
lm_output * lm_label_ids_one_hot, axis=[-1])
numerator = tf.keras.backend.sum(lm_label_weights * lm_per_example_loss)
denominator = tf.keras.backend.sum(lm_label_weights) + 1e-5
mask_label_loss = numerator / denominator
final_loss = mask_label_loss
self._add_metrics(lm_output, lm_label_ids, lm_label_weights,
lm_per_example_loss)
return final_loss
def call(self, inputs, **kwargs):
"""Implements call() for the layer."""
unpacked_inputs = tf_utils.unpack_inputs(inputs)
word_embeddings = unpacked_inputs[0]
token_type_ids = unpacked_inputs[1]
input_shape = tf_utils.get_shape_list(word_embeddings, expected_rank=3)
batch_size = input_shape[0]
seq_length = input_shape[1]
width = input_shape[2]
output = word_embeddings
if self.use_type_embeddings:
flat_token_type_ids = tf.reshape(token_type_ids, [-1])
one_hot_ids = tf.one_hot(flat_token_type_ids,
depth=self.token_type_vocab_size,
dtype=self.dtype)
token_type_embeddings = tf.matmul(one_hot_ids,
self.type_embeddings)
token_type_embeddings = tf.reshape(token_type_embeddings,
[batch_size, seq_length, width])
output += token_type_embeddings
if self.use_position_embeddings:
position_embeddings = tf.expand_dims(tf.slice(
self.position_embeddings, [0, 0], [seq_length, width]),
axis=0)
output += position_embeddings
output = self.output_layer_norm(output)
output = self.output_dropout(output,
training=kwargs.get('training', False))
projected_output = self.projection(output)
return projected_output
def call(self, inputs, **kwargs):
"""Implements call() for the layer."""
unpacked_inputs = tf_utils.unpack_inputs(inputs)
sequence_output = unpacked_inputs[0]
p_mask = unpacked_inputs[1]
cls_index = unpacked_inputs[2]
start_positions = unpacked_inputs[3]
_, seq_len, _ = sequence_output.shape.as_list()
sequence_output = tf.transpose(sequence_output, [1, 0, 2])
start_logits = self.start_logits_proj_layer(sequence_output)
start_logits = tf.transpose(tf.squeeze(start_logits, -1), [1, 0])
start_logits_masked = start_logits * (1 - p_mask) - 1e30 * p_mask
start_log_probs = tf.nn.log_softmax(start_logits_masked, -1)
if kwargs.get("training", False):
# during training, compute the end logits based on the
# ground truth of the start position
start_positions = tf.reshape(start_positions, [-1])
start_index = tf.one_hot(start_positions,
depth=seq_len,
axis=-1,
dtype=tf.float32)
start_features = tf.einsum('lbh,bl->bh', sequence_output,
start_index)
start_features = tf.tile(start_features[None], [seq_len, 1, 1])
end_logits = self.end_logits_proj_layer0(
tf.concat([sequence_output, start_features], axis=-1))
end_logits = self.end_logits_layer_norm(end_logits)
end_logits = self.end_logits_proj_layer1(end_logits)
end_logits = tf.transpose(tf.squeeze(end_logits, -1), [1, 0])
end_logits_masked = end_logits * (1 - p_mask) - 1e30 * p_mask
end_log_probs = tf.nn.log_softmax(end_logits_masked, -1)
else:
start_top_log_probs, start_top_index = tf.nn.top_k(
start_log_probs, k=self.start_n_top)
start_index = tf.one_hot(start_top_index,
depth=seq_len,
axis=-1,
dtype=tf.float32)
start_features = tf.einsum('lbh,bkl->bkh', sequence_output,
start_index)
end_input = tf.tile(sequence_output[:, :, None],
[1, 1, self.start_n_top, 1])
start_features = tf.tile(start_features[None], [seq_len, 1, 1, 1])
end_input = tf.concat([end_input, start_features], axis=-1)
end_logits = self.end_logits_proj_layer0(end_input)
end_logits = tf.reshape(end_logits,
[seq_len, -1, self.hidden_size])
end_logits = self.end_logits_layer_norm(end_logits)
end_logits = tf.reshape(
end_logits, [seq_len, -1, self.start_n_top, self.hidden_size])
end_logits = self.end_logits_proj_layer1(end_logits)
end_logits = tf.reshape(end_logits,
[seq_len, -1, self.start_n_top])
end_logits = tf.transpose(end_logits, [1, 2, 0])
end_logits_masked = end_logits * (
1 - p_mask[:, None]) - 1e30 * p_mask[:, None]
end_log_probs = tf.nn.log_softmax(end_logits_masked, -1)
end_top_log_probs, end_top_index = tf.nn.top_k(end_log_probs,
k=self.end_n_top)
end_top_log_probs = tf.reshape(
end_top_log_probs, [-1, self.start_n_top * self.end_n_top])
end_top_index = tf.reshape(end_top_index,
[-1, self.start_n_top * self.end_n_top])
# an additional layer to predict answerability
# get the representation of CLS
cls_index = tf.one_hot(cls_index, seq_len, axis=-1, dtype=tf.float32)
cls_feature = tf.einsum('lbh,bl->bh', sequence_output, cls_index)
# get the representation of START
start_p = tf.nn.softmax(start_logits_masked,
axis=-1,
name='softmax_start')
start_feature = tf.einsum('lbh,bl->bh', sequence_output, start_p)
ans_feature = tf.concat([start_feature, cls_feature], -1)
ans_feature = self.answer_class_proj_layer0(ans_feature)
ans_feature = self.ans_feature_dropout(ans_feature,
training=kwargs.get(
'training', False))
cls_logits = self.answer_class_proj_layer1(ans_feature)
cls_logits = tf.squeeze(cls_logits, -1)
if kwargs.get("training", False):
return (start_log_probs, end_log_probs, cls_logits)
else:
return (start_top_log_probs, start_top_index, end_top_log_probs,
end_top_index, cls_logits)
