# coding=utf-8
# Copyright 2018 The OpenAI Team Authors and HuggingFace Inc. team.
# Copyright (c) 2018, NVIDIA CORPORATION. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
""" TF 2.0 ALBERT model. """
from dataclasses import dataclass
from typing import Optional, Tuple
import tensorflow as tf
from .activations_tf import get_tf_activation
from .configuration_albert import AlbertConfig
from .file_utils import (
MULTIPLE_CHOICE_DUMMY_INPUTS,
ModelOutput,
add_code_sample_docstrings,
add_start_docstrings,
add_start_docstrings_to_callable,
replace_return_docstrings,
)
from .modeling_tf_bert import TFBertSelfAttention
from .modeling_tf_outputs import (
TFBaseModelOutput,
TFBaseModelOutputWithPooling,
TFMaskedLMOutput,
TFMultipleChoiceModelOutput,
TFQuestionAnsweringModelOutput,
TFSequenceClassifierOutput,
TFTokenClassifierOutput,
)
from .modeling_tf_utils import (
TFMaskedLanguageModelingLoss,
TFMultipleChoiceLoss,
TFPreTrainedModel,
TFQuestionAnsweringLoss,
TFSequenceClassificationLoss,
TFTokenClassificationLoss,
get_initializer,
keras_serializable,
shape_list,
)
from .tokenization_utils import BatchEncoding
from .utils import logging
logger = logging.get_logger(__name__)
_CONFIG_FOR_DOC = "AlbertConfig"
_TOKENIZER_FOR_DOC = "AlbertTokenizer"
TF_ALBERT_PRETRAINED_MODEL_ARCHIVE_LIST = [
"albert-base-v1",
"albert-large-v1",
"albert-xlarge-v1",
"albert-xxlarge-v1",
"albert-base-v2",
"albert-large-v2",
"albert-xlarge-v2",
"albert-xxlarge-v2",
# See all ALBERT models at https://huggingface.co/models?filter=albert
]
class TFAlbertEmbeddings(tf.keras.layers.Layer):
"""Construct the embeddings from word, position and token_type embeddings."""
def __init__(self, config, **kwargs):
super().__init__(**kwargs)
self.config = config
self.vocab_size = config.vocab_size
self.position_embeddings = tf.keras.layers.Embedding(
config.max_position_embeddings,
config.embedding_size,
embeddings_initializer=get_initializer(self.config.initializer_range),
name="position_embeddings",
)
self.token_type_embeddings = tf.keras.layers.Embedding(
config.type_vocab_size,
config.embedding_size,
embeddings_initializer=get_initializer(self.config.initializer_range),
name="token_type_embeddings",
)
# self.LayerNorm is not snake-cased to stick with TensorFlow model variable name and be able to load
# any TensorFlow checkpoint file
self.LayerNorm = tf.keras.layers.LayerNormalization(epsilon=config.layer_norm_eps, name="LayerNorm")
self.dropout = tf.keras.layers.Dropout(config.hidden_dropout_prob)
def build(self, input_shape):
"""Build shared word embedding layer """
with tf.name_scope("word_embeddings"):
# Create and initialize weights. The random normal initializer was chosen
# arbitrarily, and works well.
self.word_embeddings = self.add_weight(
"weight",
shape=[self.config.vocab_size, self.config.embedding_size],
initializer=get_initializer(self.config.initializer_range),
)
super().build(input_shape)
def call(
self,
input_ids=None,
position_ids=None,
token_type_ids=None,
inputs_embeds=None,
mode="embedding",
training=False,
):
"""Get token embeddings of inputs.
Args:
inputs: list of three int64 tensors with shape [batch_size, length]: (input_ids, position_ids, token_type_ids)
mode: string, a valid value is one of "embedding" and "linear".
Returns:
outputs: (1) If mode == "embedding", output embedding tensor, float32 with
shape [batch_size, length, embedding_size]; (2) mode == "linear", output
linear tensor, float32 with shape [batch_size, length, vocab_size].
Raises:
ValueError: if mode is not valid.
Shared weights logic adapted from
https://github.com/tensorflow/models/blob/a009f4fb9d2fc4949e32192a944688925ef78659/official/transformer/v2/embedding_layer.py#L24
"""
if mode == "embedding":
return self._embedding(input_ids, position_ids, token_type_ids, inputs_embeds, training=training)
elif mode == "linear":
return self._linear(input_ids)
else:
raise ValueError("mode {} is not valid.".format(mode))
def _embedding(self, input_ids, position_ids, token_type_ids, inputs_embeds, training=False):
"""Applies embedding based on inputs tensor."""
assert not (input_ids is None and inputs_embeds is None)
if input_ids is not None:
input_shape = shape_list(input_ids)
else:
input_shape = shape_list(inputs_embeds)[:-1]
seq_length = input_shape[1]
if position_ids is None:
position_ids = tf.range(seq_length, dtype=tf.int32)[tf.newaxis, :]
if token_type_ids is None:
token_type_ids = tf.fill(input_shape, 0)
if inputs_embeds is None:
inputs_embeds = tf.gather(self.word_embeddings, input_ids)
position_embeddings = self.position_embeddings(position_ids)
token_type_embeddings = self.token_type_embeddings(token_type_ids)
embeddings = inputs_embeds + position_embeddings + token_type_embeddings
embeddings = self.LayerNorm(embeddings)
embeddings = self.dropout(embeddings, training=training)
return embeddings
def _linear(self, inputs):
"""Computes logits by running inputs through a linear layer.
Args:
inputs: A float32 tensor with shape [batch_size, length, embedding_size]
Returns:
float32 tensor with shape [batch_size, length, vocab_size].
"""
batch_size = shape_list(inputs)[0]
length = shape_list(inputs)[1]
x = tf.reshape(inputs, [-1, self.config.embedding_size])
logits = tf.matmul(x, self.word_embeddings, transpose_b=True)
return tf.reshape(logits, [batch_size, length, self.config.vocab_size])
class TFAlbertSelfAttention(tf.keras.layers.Layer):
def __init__(self, config, **kwargs):
super().__init__(**kwargs)
if config.hidden_size % config.num_attention_heads != 0:
raise ValueError(
"The hidden size (%d) is not a multiple of the number of attention "
"heads (%d)" % (config.hidden_size, config.num_attention_heads)
)
self.num_attention_heads = config.num_attention_heads
assert (
config.hidden_size % config.num_attention_heads == 0
), f"Hidden size {config.hidden_size} not dividable by number of heads {config.num_attention_heads}"
self.attention_head_size = int(config.hidden_size / config.num_attention_heads)
self.all_head_size = self.num_attention_heads * self.attention_head_size
self.output_attentions = config.output_attentions
self.query = tf.keras.layers.Dense(
self.all_head_size, kernel_initializer=get_initializer(config.initializer_range), name="query"
)
self.key = tf.keras.layers.Dense(
self.all_head_size, kernel_initializer=get_initializer(config.initializer_range), name="key"
)
self.value = tf.keras.layers.Dense(
self.all_head_size, kernel_initializer=get_initializer(config.initializer_range), name="value"
)
self.dropout = tf.keras.layers.Dropout(config.attention_probs_dropout_prob)
def transpose_for_scores(self, x, batch_size):
x = tf.reshape(x, (batch_size, -1, self.num_attention_heads, self.attention_head_size))
return tf.transpose(x, perm=[0, 2, 1, 3])
def call(self, hidden_states, attention_mask, head_mask, output_attentions, training=False):
batch_size = shape_list(hidden_states)[0]
mixed_query_layer = self.query(hidden_states)
mixed_key_layer = self.key(hidden_states)
mixed_value_layer = self.value(hidden_states)
query_layer = self.transpose_for_scores(mixed_query_layer, batch_size)
key_layer = self.transpose_for_scores(mixed_key_layer, batch_size)
value_layer = self.transpose_for_scores(mixed_value_layer, batch_size)
# Take the dot product between "query" and "key" to get the raw attention scores.
# (batch size, num_heads, seq_len_q, seq_len_k)
attention_scores = tf.matmul(query_layer, key_layer, transpose_b=True)
# scale attention_scores
dk = tf.cast(shape_list(key_layer)[-1], tf.float32)
attention_scores = attention_scores / tf.math.sqrt(dk)
if attention_mask is not None:
# Apply the attention mask is (precomputed for all layers in TFAlbertModel call() function)
attention_scores = attention_scores + attention_mask
# Normalize the attention scores to probabilities.
attention_probs = tf.nn.softmax(attention_scores, axis=-1)
# 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.dropout(attention_probs, training=training)
# Mask heads if we want to
if head_mask is not None:
attention_probs = attention_probs * head_mask
context_layer = tf.matmul(attention_probs, value_layer)
context_layer = tf.transpose(context_layer, perm=[0, 2, 1, 3])
context_layer = tf.reshape(
context_layer, (batch_size, -1, self.all_head_size)
) # (batch_size, seq_len_q, all_head_size)
outputs = (context_layer, attention_probs) if output_attentions else (context_layer,)
return outputs
class TFAlbertSelfOutput(tf.keras.layers.Layer):
def __init__(self, config, **kwargs):
super().__init__(**kwargs)
self.dense = tf.keras.layers.Dense(
config.hidden_size, kernel_initializer=get_initializer(config.initializer_range), name="dense"
)
self.LayerNorm = tf.keras.layers.LayerNormalization(epsilon=config.layer_norm_eps, name="LayerNorm")
self.dropout = tf.keras.layers.Dropout(config.hidden_dropout_prob)
def call(self, hidden_states, input_tensor, training=False):
hidden_states = self.dense(hidden_states)
hidden_states = self.dropout(hidden_states, training=training)
hidden_states = self.LayerNorm(hidden_states + input_tensor)
return hidden_states
class TFAlbertAttention(TFBertSelfAttention):
""" Contains the complete attention sublayer, including both dropouts and layer norm. """
def __init__(self, config, **kwargs):
super().__init__(config, **kwargs)
self.hidden_size = config.hidden_size
self.output_attentions = config.output_attentions
self.dense = tf.keras.layers.Dense(
config.hidden_size, kernel_initializer=get_initializer(config.initializer_range), name="dense"
)
self.LayerNorm = tf.keras.layers.LayerNormalization(epsilon=config.layer_norm_eps, name="LayerNorm")
self.pruned_heads = set()
# Two different dropout probabilities; see https://github.com/google-research/albert/blob/master/modeling.py#L971-L993
self.attention_dropout = tf.keras.layers.Dropout(config.attention_probs_dropout_prob)
self.output_dropout = tf.keras.layers.Dropout(config.hidden_dropout_prob)
def prune_heads(self, heads):
raise NotImplementedError
def call(self, input_tensor, attention_mask, head_mask, output_attentions, training=False):
batch_size = shape_list(input_tensor)[0]
mixed_query_layer = self.query(input_tensor)
mixed_key_layer = self.key(input_tensor)
mixed_value_layer = self.value(input_tensor)
query_layer = self.transpose_for_scores(mixed_query_layer, batch_size)
key_layer = self.transpose_for_scores(mixed_key_layer, batch_size)
value_layer = self.transpose_for_scores(mixed_value_layer, batch_size)
# Take the dot product between "query" and "key" to get the raw attention scores.
# (batch size, num_heads, seq_len_q, seq_len_k)
attention_scores = tf.matmul(query_layer, key_layer, transpose_b=True)
# scale attention_scores
dk = tf.cast(shape_list(key_layer)[-1], tf.float32)
attention_scores = attention_scores / tf.math.sqrt(dk)
if attention_mask is not None:
# Apply the attention mask is (precomputed for all layers in TFBertModel call() function)
attention_scores = attention_scores + attention_mask
# Normalize the attention scores to probabilities.
attention_probs = tf.nn.softmax(attention_scores, axis=-1)
# 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_dropout(attention_probs, training=training)
# Mask heads if we want to
if head_mask is not None:
attention_probs = attention_probs * head_mask
context_layer = tf.matmul(attention_probs, value_layer)
context_layer = tf.transpose(context_layer, perm=[0, 2, 1, 3])
context_layer = tf.reshape(
context_layer, (batch_size, -1, self.all_head_size)
) # (batch_size, seq_len_q, all_head_size)
self_outputs = (context_layer, attention_probs) if output_attentions else (context_layer,)
hidden_states = self_outputs[0]
hidden_states = self.dense(hidden_states)
hidden_states = self.output_dropout(hidden_states, training=training)
attention_output = self.LayerNorm(hidden_states + input_tensor)
# add attentions if we output them
outputs = (attention_output,) + self_outputs[1:]
return outputs
class TFAlbertLayer(tf.keras.layers.Layer):
def __init__(self, config, **kwargs):
super().__init__(**kwargs)
self.attention = TFAlbertAttention(config, name="attention")
self.ffn = tf.keras.layers.Dense(
config.intermediate_size, kernel_initializer=get_initializer(config.initializer_range), name="ffn"
)
if isinstance(config.hidden_act, str):
self.activation = get_tf_activation(config.hidden_act)
else:
self.activation = config.hidden_act
self.ffn_output = tf.keras.layers.Dense(
config.hidden_size, kernel_initializer=get_initializer(config.initializer_range), name="ffn_output"
)
self.full_layer_layer_norm = tf.keras.layers.LayerNormalization(
epsilon=config.layer_norm_eps, name="full_layer_layer_norm"
)
self.dropout = tf.keras.layers.Dropout(config.hidden_dropout_prob)
def call(self, hidden_states, attention_mask, head_mask, output_attentions, training=False):
attention_outputs = self.attention(
hidden_states, attention_mask, head_mask, output_attentions, training=training
)
ffn_output = self.ffn(attention_outputs[0])
ffn_output = self.activation(ffn_output)
ffn_output = self.ffn_output(ffn_output)
ffn_output = self.dropout(ffn_output, training=training)
hidden_states = self.full_layer_layer_norm(ffn_output + attention_outputs[0])
# add attentions if we output them
outputs = (hidden_states,) + attention_outputs[1:]
return outputs
class TFAlbertLayerGroup(tf.keras.layers.Layer):
def __init__(self, config, **kwargs):
super().__init__(**kwargs)
self.output_attentions = config.output_attentions
self.output_hidden_states = config.output_hidden_states
self.albert_layers = [
TFAlbertLayer(config, name="albert_layers_._{}".format(i)) for i in range(config.inner_group_num)
]
def call(self, hidden_states, attention_mask, head_mask, output_attentions, output_hidden_states, training=False):
layer_hidden_states = ()
layer_attentions = ()
for layer_index, albert_layer in enumerate(self.albert_layers):
layer_output = albert_layer(
hidden_states, attention_mask, head_mask[layer_index], output_attentions, training=training
)
hidden_states = layer_output[0]
if output_attentions:
layer_attentions = layer_attentions + (layer_output[1],)
if output_hidden_states:
layer_hidden_states = layer_hidden_states + (hidden_states,)
outputs = (hidden_states,)
if output_hidden_states:
outputs = outputs + (layer_hidden_states,)
if output_attentions:
outputs = outputs + (layer_attentions,)
# last-layer hidden state, (layer hidden states), (layer attentions)
return outputs
class TFAlbertTransformer(tf.keras.layers.Layer):
def __init__(self, config, **kwargs):
super().__init__(**kwargs)
self.config = config
self.embedding_hidden_mapping_in = tf.keras.layers.Dense(
config.hidden_size,
kernel_initializer=get_initializer(config.initializer_range),
name="embedding_hidden_mapping_in",
)
self.albert_layer_groups = [
TFAlbertLayerGroup(config, name="albert_layer_groups_._{}".format(i))
for i in range(config.num_hidden_groups)
]
def call(
self,
hidden_states,
attention_mask,
head_mask,
output_attentions,
output_hidden_states,
return_dict,
training=False,
):
hidden_states = self.embedding_hidden_mapping_in(hidden_states)
all_attentions = () if output_attentions else None
all_hidden_states = (hidden_states,) if output_hidden_states else None
for i in range(self.config.num_hidden_layers):
# Number of layers in a hidden group
layers_per_group = int(self.config.num_hidden_layers / self.config.num_hidden_groups)
# Index of the hidden group
group_idx = int(i / (self.config.num_hidden_layers / self.config.num_hidden_groups))
layer_group_output = self.albert_layer_groups[group_idx](
hidden_states,
attention_mask,
head_mask[group_idx * layers_per_group : (group_idx + 1) * layers_per_group],
output_attentions,
output_hidden_states,
training=training,
)
hidden_states = layer_group_output[0]
if output_attentions:
all_attentions = all_attentions + layer_group_output[-1]
if output_hidden_states:
all_hidden_states = all_hidden_states + (hidden_states,)
if not return_dict:
return tuple(v for v in [hidden_states, all_hidden_states, all_attentions] if v is not None)
return TFBaseModelOutput(
last_hidden_state=hidden_states, hidden_states=all_hidden_states, attentions=all_attentions
)
class TFAlbertPreTrainedModel(TFPreTrainedModel):
"""An abstract class to handle weights initialization and
a simple interface for downloading and loading pretrained models.
"""
config_class = AlbertConfig
base_model_prefix = "albert"
class TFAlbertMLMHead(tf.keras.layers.Layer):
def __init__(self, config, input_embeddings, **kwargs):
super().__init__(**kwargs)
self.vocab_size = config.vocab_size
self.dense = tf.keras.layers.Dense(
config.embedding_size, kernel_initializer=get_initializer(config.initializer_range), name="dense"
)
if isinstance(config.hidden_act, str):
self.activation = get_tf_activation(config.hidden_act)
else:
self.activation = config.hidden_act
self.LayerNorm = tf.keras.layers.LayerNormalization(epsilon=config.layer_norm_eps, name="LayerNorm")
# The output weights are the same as the input embeddings, but there is
# an output-only bias for each token.
self.decoder = input_embeddings
def build(self, input_shape):
self.bias = self.add_weight(shape=(self.vocab_size,), initializer="zeros", trainable=True, name="bias")
self.decoder_bias = self.add_weight(
shape=(self.vocab_size,), initializer="zeros", trainable=True, name="decoder/bias"
)
super().build(input_shape)
def call(self, hidden_states):
hidden_states = self.dense(hidden_states)
hidden_states = self.activation(hidden_states)
hidden_states = self.LayerNorm(hidden_states)
hidden_states = self.decoder(hidden_states, mode="linear") + self.decoder_bias
return hidden_states
@keras_serializable
class TFAlbertMainLayer(tf.keras.layers.Layer):
config_class = AlbertConfig
def __init__(self, config, **kwargs):
super().__init__(**kwargs)
self.num_hidden_layers = config.num_hidden_layers
self.output_attentions = config.output_attentions
self.output_hidden_states = config.output_hidden_states
self.return_dict = config.use_return_dict
self.embeddings = TFAlbertEmbeddings(config, name="embeddings")
self.encoder = TFAlbertTransformer(config, name="encoder")
self.pooler = tf.keras.layers.Dense(
config.hidden_size,
kernel_initializer=get_initializer(config.initializer_range),
activation="tanh",
name="pooler",
)
def get_input_embeddings(self):
return self.embeddings
def set_input_embeddings(self, value):
self.embeddings.word_embeddings = value
self.embeddings.vocab_size = value.shape[0]
def _resize_token_embeddings(self, new_num_tokens):
raise NotImplementedError
def _prune_heads(self, heads_to_prune):
"""Prunes heads of the model.
heads_to_prune: dict of {layer_num: list of heads to prune in this layer}
See base class PreTrainedModel
"""
raise NotImplementedError
def call(
self,
inputs,
attention_mask=None,
token_type_ids=None,
position_ids=None,
head_mask=None,
inputs_embeds=None,
output_attentions=None,
output_hidden_states=None,
return_dict=None,
training=False,
):
if isinstance(inputs, (tuple, list)):
input_ids = inputs[0]
attention_mask = inputs[1] if len(inputs) > 1 else attention_mask
token_type_ids = inputs[2] if len(inputs) > 2 else token_type_ids
position_ids = inputs[3] if len(inputs) > 3 else position_ids
head_mask = inputs[4] if len(inputs) > 4 else head_mask
inputs_embeds = inputs[5] if len(inputs) > 5 else inputs_embeds
output_attentions = inputs[6] if len(inputs) > 6 else output_attentions
output_hidden_states = inputs[7] if len(inputs) > 7 else output_hidden_states
return_dict = inputs[8] if len(inputs) > 8 else return_dict
assert len(inputs) <= 9, "Too many inputs."
elif isinstance(inputs, (dict, BatchEncoding)):
input_ids = inputs.get("input_ids")
attention_mask = inputs.get("attention_mask", attention_mask)
token_type_ids = inputs.get("token_type_ids", token_type_ids)
position_ids = inputs.get("position_ids", position_ids)
head_mask = inputs.get("head_mask", head_mask)
inputs_embeds = inputs.get("inputs_embeds", inputs_embeds)
output_attentions = inputs.get("output_attentions", output_attentions)
output_hidden_states = inputs.get("output_hidden_states", output_hidden_states)
return_dict = inputs.get("return_dict", return_dict)
assert len(inputs) <= 9, "Too many inputs."
else:
input_ids = inputs
output_attentions = output_attentions if output_attentions is not None else self.output_attentions
output_hidden_states = output_hidden_states if output_hidden_states is not None else self.output_hidden_states
return_dict = return_dict if return_dict is not None else self.return_dict
if input_ids is not None and inputs_embeds is not None:
raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
elif input_ids is not None:
input_shape = shape_list(input_ids)
elif inputs_embeds is not None:
input_shape = shape_list(inputs_embeds)[:-1]
else:
raise ValueError("You have to specify either input_ids or inputs_embeds")
if attention_mask is None:
attention_mask = tf.fill(input_shape, 1)
if token_type_ids is None:
token_type_ids = tf.fill(input_shape, 0)
# We create a 3D attention mask from a 2D tensor mask.
# Sizes are [batch_size, 1, 1, to_seq_length]
# So we can broadcast to [batch_size, num_heads, from_seq_length, to_seq_length]
# this attention mask is more simple than the triangular masking of causal attention
# used in OpenAI GPT, we just need to prepare the broadcast dimension here.
extended_attention_mask = attention_mask[:, tf.newaxis, tf.newaxis, :]
# 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.
# Since we are adding it to the raw scores before the softmax, this is
# effectively the same as removing these entirely.
extended_attention_mask = tf.cast(extended_attention_mask, tf.float32)
extended_attention_mask = (1.0 - extended_attention_mask) * -10000.0
# Prepare head mask if needed
# 1.0 in head_mask indicate we keep the head
# attention_probs has shape bsz x n_heads x N x N
# input head_mask has shape [num_heads] or [num_hidden_layers x num_heads]
# and head_mask is converted to shape [num_hidden_layers x batch x num_heads x seq_length x seq_length]
if head_mask is not None:
raise NotImplementedError
else:
head_mask = [None] * self.num_hidden_layers
# head_mask = tf.constant([0] * self.num_hidden_layers)
embedding_output = self.embeddings(input_ids, position_ids, token_type_ids, inputs_embeds, training=training)
encoder_outputs = self.encoder(
embedding_output,
extended_attention_mask,
head_mask,
output_attentions,
output_hidden_states,
return_dict,
training=training,
)
sequence_output = encoder_outputs[0]
pooled_output = self.pooler(sequence_output[:, 0])
if not return_dict:
return (
sequence_output,
pooled_output,
) + encoder_outputs[1:]
return TFBaseModelOutputWithPooling(
last_hidden_state=sequence_output,
pooler_output=pooled_output,
hidden_states=encoder_outputs.hidden_states,
attentions=encoder_outputs.attentions,
)
[docs]@dataclass
class TFAlbertForPreTrainingOutput(ModelOutput):
"""
Output type of :class:`~transformers.TFAlbertForPreTrainingModel`.
Args:
prediction_logits (:obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length, config.vocab_size)`):
Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
sop_logits (:obj:`tf.Tensor` of shape :obj:`(batch_size, 2)`):
Prediction scores of the next sequence prediction (classification) head (scores of True/False
continuation before SoftMax).
hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
Tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
of shape :obj:`(batch_size, sequence_length, hidden_size)`.
Hidden-states of the model at the output of each layer plus the initial embedding outputs.
attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
Tuple of :obj:`tf.Tensor` (one for each layer) of shape
:obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
heads.
"""
prediction_logits: tf.Tensor = None
sop_logits: tf.Tensor = None
hidden_states: Optional[Tuple[tf.Tensor]] = None
attentions: Optional[Tuple[tf.Tensor]] = None
ALBERT_START_DOCSTRING = r"""
This model is a `tf.keras.Model <https://www.tensorflow.org/api_docs/python/tf/keras/Model>`__ sub-class.
Use it as a regular TF 2.0 Keras Model and
refer to the TF 2.0 documentation for all matter related to general usage and behavior.
.. _`ALBERT: A Lite BERT for Self-supervised Learning of Language Representations`:
https://arxiv.org/abs/1909.11942
.. _`tf.keras.Model`:
https://www.tensorflow.org/versions/r2.0/api_docs/python/tf/keras/Model
.. note::
TF 2.0 models accepts two formats as inputs:
- having all inputs as keyword arguments (like PyTorch models), or
- having all inputs as a list, tuple or dict in the first positional arguments.
This second option is useful when using :obj:`tf.keras.Model.fit()` method which currently requires having
all the tensors in the first argument of the model call function: :obj:`model(inputs)`.
If you choose this second option, there are three possibilities you can use to gather all the input Tensors
in the first positional argument :
- a single Tensor with input_ids only and nothing else: :obj:`model(inputs_ids)`
- a list of varying length with one or several input Tensors IN THE ORDER given in the docstring:
:obj:`model([input_ids, attention_mask])` or :obj:`model([input_ids, attention_mask, token_type_ids])`
- a dictionary with one or several input Tensors associated to the input names given in the docstring:
:obj:`model({'input_ids': input_ids, 'token_type_ids': token_type_ids})`
Args:
config (:class:`~transformers.AlbertConfig`): Model configuration class with all the parameters of the model.
Initializing with a config file does not load the weights associated with the model, only the configuration.
Check out the :meth:`~transformers.PreTrainedModel.from_pretrained` method to load the model weights.
"""
ALBERT_INPUTS_DOCSTRING = r"""
Args:
input_ids (:obj:`Numpy array` or :obj:`tf.Tensor` of shape :obj:`{0}`):
Indices of input sequence tokens in the vocabulary.
Indices can be obtained using :class:`transformers.AlbertTokenizer`.
See :func:`transformers.PreTrainedTokenizer.encode` and
:func:`transformers.PreTrainedTokenizer.__call__` for details.
`What are input IDs? <../glossary.html#input-ids>`__
attention_mask (:obj:`Numpy array` or :obj:`tf.Tensor` of shape :obj:`{0}`, `optional`):
Mask to avoid performing attention on padding token indices.
Mask values selected in ``[0, 1]``:
``1`` for tokens that are NOT MASKED, ``0`` for MASKED tokens.
`What are attention masks? <../glossary.html#attention-mask>`__
token_type_ids (:obj:`Numpy array` or :obj:`tf.Tensor` of shape :obj:`{0}`, `optional`):
Segment token indices to indicate first and second portions of the inputs.
Indices are selected in ``[0, 1]``: ``0`` corresponds to a `sentence A` token, ``1``
corresponds to a `sentence B` token
`What are token type IDs? <../glossary.html#token-type-ids>`_
position_ids (:obj:`Numpy array` or :obj:`tf.Tensor` of shape :obj:`{0}`, `optional`):
Indices of positions of each input sequence tokens in the position embeddings.
Selected in the range ``[0, config.max_position_embeddings - 1]``.
`What are position IDs? <../glossary.html#position-ids>`_
head_mask (:obj:`Numpy array` or :obj:`tf.Tensor` of shape :obj:`(num_heads,)` or :obj:`(num_layers, num_heads)`, `optional`):
Mask to nullify selected heads of the self-attention modules.
Mask values selected in ``[0, 1]``:
``1`` indicates the head is **not masked**, ``0`` indicates the head is **masked**.
inputs_embeds (:obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`, `optional`):
Optionally, instead of passing :obj:`input_ids` you can choose to directly pass an embedded representation.
This is useful if you want more control over how to convert `input_ids` indices into associated vectors
than the model's internal embedding lookup matrix.
training (:obj:`boolean`, `optional`, defaults to :obj:`False`):
Whether to activate dropout modules (if set to :obj:`True`) during training or to de-activate them
(if set to :obj:`False`) for evaluation.
output_attentions (:obj:`bool`, `optional`):
If set to ``True``, the attentions tensors of all attention layers are returned. See ``attentions`` under returned tensors for more detail.
output_hidden_states (:obj:`bool`, `optional`):
If set to ``True``, the hidden states of all layers are returned. See ``hidden_states`` under returned tensors for more detail.
return_dict (:obj:`bool`, `optional`):
If set to ``True``, the model will return a :class:`~transformers.file_utils.ModelOutput` instead of a
plain tuple.
"""
[docs]@add_start_docstrings(
"The bare Albert Model transformer outputing raw hidden-states without any specific head on top.",
ALBERT_START_DOCSTRING,
)
class TFAlbertModel(TFAlbertPreTrainedModel):
def __init__(self, config, *inputs, **kwargs):
super().__init__(config, *inputs, **kwargs)
self.albert = TFAlbertMainLayer(config, name="albert")
[docs] @add_start_docstrings_to_callable(ALBERT_INPUTS_DOCSTRING.format("(batch_size, sequence_length)"))
@add_code_sample_docstrings(
tokenizer_class=_TOKENIZER_FOR_DOC,
checkpoint="albert-base-v2",
output_type=TFBaseModelOutputWithPooling,
config_class=_CONFIG_FOR_DOC,
)
def call(self, inputs, **kwargs):
outputs = self.albert(inputs, **kwargs)
return outputs
[docs]@add_start_docstrings(
"""Albert Model with two heads on top for pre-training:
a `masked language modeling` head and a `sentence order prediction` (classification) head. """,
ALBERT_START_DOCSTRING,
)
class TFAlbertForPreTraining(TFAlbertPreTrainedModel):
def __init__(self, config, *inputs, **kwargs):
super().__init__(config, *inputs, **kwargs)
self.num_labels = config.num_labels
self.albert = TFAlbertMainLayer(config, name="albert")
self.predictions = TFAlbertMLMHead(config, self.albert.embeddings, name="predictions")
self.sop_classifier = TFAlbertSOPHead(config, name="sop_classifier")
[docs] def get_output_embeddings(self):
return self.albert.embeddings
[docs] @add_start_docstrings_to_callable(ALBERT_INPUTS_DOCSTRING.format("(batch_size, sequence_length)"))
@replace_return_docstrings(output_type=TFAlbertForPreTrainingOutput, config_class=_CONFIG_FOR_DOC)
def call(self, inputs, **kwargs):
r"""
Return:
Examples::
>>> import tensorflow as tf
>>> from transformers import AlbertTokenizer, TFAlbertForPreTraining
>>> tokenizer = AlbertTokenizer.from_pretrained('albert-base-v2')
>>> model = TFAlbertForPreTraining.from_pretrained('albert-base-v2')
>>> input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True))[None, :] # Batch size 1
>>> outputs = model(input_ids)
>>> prediction_scores, sop_scores = outputs[:2]
"""
return_dict = kwargs.get("return_dict")
return_dict = return_dict if return_dict is not None else self.albert.return_dict
outputs = self.albert(inputs, **kwargs)
sequence_output, pooled_output = outputs[:2]
prediction_scores = self.predictions(sequence_output)
sop_scores = self.sop_classifier(pooled_output, training=kwargs.get("training", False))
if not return_dict:
return (prediction_scores, sop_scores) + outputs[2:]
return TFAlbertForPreTrainingOutput(
prediction_logits=prediction_scores,
sop_logits=sop_scores,
hidden_states=outputs.hidden_states,
attentions=outputs.attentions,
)
class TFAlbertSOPHead(tf.keras.layers.Layer):
def __init__(self, config, **kwargs):
super().__init__(**kwargs)
self.dropout = tf.keras.layers.Dropout(config.classifier_dropout_prob)
self.classifier = tf.keras.layers.Dense(
config.num_labels,
kernel_initializer=get_initializer(config.initializer_range),
name="classifier",
)
def call(self, pooled_output, training: bool):
dropout_pooled_output = self.dropout(pooled_output, training=training)
logits = self.classifier(dropout_pooled_output)
return logits
[docs]@add_start_docstrings(
"""Albert Model transformer with a sequence classification/regression head on top (a linear layer on top of
the pooled output) e.g. for GLUE tasks. """,
ALBERT_START_DOCSTRING,
)
class TFAlbertForSequenceClassification(TFAlbertPreTrainedModel, TFSequenceClassificationLoss):
def __init__(self, config, *inputs, **kwargs):
super().__init__(config, *inputs, **kwargs)
self.num_labels = config.num_labels
self.albert = TFAlbertMainLayer(config, name="albert")
self.dropout = tf.keras.layers.Dropout(config.classifier_dropout_prob)
self.classifier = tf.keras.layers.Dense(
config.num_labels, kernel_initializer=get_initializer(config.initializer_range), name="classifier"
)
[docs] @add_start_docstrings_to_callable(ALBERT_INPUTS_DOCSTRING)
@add_code_sample_docstrings(
tokenizer_class=_TOKENIZER_FOR_DOC,
checkpoint="albert-base-v2",
output_type=TFSequenceClassifierOutput,
config_class=_CONFIG_FOR_DOC,
)
def call(
self,
inputs=None,
attention_mask=None,
token_type_ids=None,
position_ids=None,
head_mask=None,
inputs_embeds=None,
output_attentions=None,
output_hidden_states=None,
return_dict=None,
labels=None,
training=False,
):
r"""
labels (:obj:`tf.Tensor` of shape :obj:`(batch_size,)`, `optional`):
Labels for computing the sequence classification/regression loss.
Indices should be in ``[0, ..., config.num_labels - 1]``.
If ``config.num_labels == 1`` a regression loss is computed (Mean-Square loss),
If ``config.num_labels > 1`` a classification loss is computed (Cross-Entropy).
"""
return_dict = return_dict if return_dict is not None else self.albert.return_dict
if isinstance(inputs, (tuple, list)):
labels = inputs[9] if len(inputs) > 9 else labels
if len(inputs) > 9:
inputs = inputs[:9]
elif isinstance(inputs, (dict, BatchEncoding)):
labels = inputs.pop("labels", labels)
outputs = self.albert(
inputs,
attention_mask=attention_mask,
token_type_ids=token_type_ids,
position_ids=position_ids,
head_mask=head_mask,
inputs_embeds=inputs_embeds,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
training=training,
)
pooled_output = outputs[1]
pooled_output = self.dropout(pooled_output, training=training)
logits = self.classifier(pooled_output)
loss = None if labels is None else self.compute_loss(labels, logits)
if not return_dict:
output = (logits,) + outputs[2:]
return ((loss,) + output) if loss is not None else output
return TFSequenceClassifierOutput(
loss=loss,
logits=logits,
hidden_states=outputs.hidden_states,
attentions=outputs.attentions,
)
[docs]@add_start_docstrings(
"""Albert Model with a token classification head on top (a linear layer on top of
the hidden-states output) e.g. for Named-Entity-Recognition (NER) tasks. """,
ALBERT_START_DOCSTRING,
)
class TFAlbertForTokenClassification(TFAlbertPreTrainedModel, TFTokenClassificationLoss):
def __init__(self, config, *inputs, **kwargs):
super().__init__(config, *inputs, **kwargs)
self.num_labels = config.num_labels
self.albert = TFAlbertMainLayer(config, name="albert")
self.dropout = tf.keras.layers.Dropout(config.hidden_dropout_prob)
self.classifier = tf.keras.layers.Dense(
config.num_labels, kernel_initializer=get_initializer(config.initializer_range), name="classifier"
)
[docs] @add_start_docstrings_to_callable(ALBERT_INPUTS_DOCSTRING)
@add_code_sample_docstrings(
tokenizer_class=_TOKENIZER_FOR_DOC,
checkpoint="albert-base-v2",
output_type=TFTokenClassifierOutput,
config_class=_CONFIG_FOR_DOC,
)
def call(
self,
inputs=None,
attention_mask=None,
token_type_ids=None,
position_ids=None,
head_mask=None,
inputs_embeds=None,
output_attentions=None,
output_hidden_states=None,
return_dict=None,
labels=None,
training=False,
):
r"""
labels (:obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):
Labels for computing the token classification loss.
Indices should be in ``[0, ..., config.num_labels - 1]``.
"""
return_dict = return_dict if return_dict is not None else self.albert.return_dict
if isinstance(inputs, (tuple, list)):
labels = inputs[9] if len(inputs) > 9 else labels
if len(inputs) > 9:
inputs = inputs[:9]
elif isinstance(inputs, (dict, BatchEncoding)):
labels = inputs.pop("labels", labels)
outputs = self.albert(
inputs,
attention_mask=attention_mask,
token_type_ids=token_type_ids,
position_ids=position_ids,
head_mask=head_mask,
inputs_embeds=inputs_embeds,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
training=training,
)
sequence_output = outputs[0]
sequence_output = self.dropout(sequence_output, training=training)
logits = self.classifier(sequence_output)
loss = None if labels is None else self.compute_loss(labels, logits)
if not return_dict:
output = (logits,) + outputs[2:]
return ((loss,) + output) if loss is not None else output
return TFTokenClassifierOutput(
loss=loss,
logits=logits,
hidden_states=outputs.hidden_states,
attentions=outputs.attentions,
)
[docs]@add_start_docstrings(
"""Albert Model with a span classification head on top for extractive question-answering tasks like SQuAD (a linear layers on top of the hidden-states output to compute `span start logits` and `span end logits`). """,
ALBERT_START_DOCSTRING,
)
class TFAlbertForQuestionAnswering(TFAlbertPreTrainedModel, TFQuestionAnsweringLoss):
def __init__(self, config, *inputs, **kwargs):
super().__init__(config, *inputs, **kwargs)
self.num_labels = config.num_labels
self.albert = TFAlbertMainLayer(config, name="albert")
self.qa_outputs = tf.keras.layers.Dense(
config.num_labels, kernel_initializer=get_initializer(config.initializer_range), name="qa_outputs"
)
[docs] @add_start_docstrings_to_callable(ALBERT_INPUTS_DOCSTRING)
@add_code_sample_docstrings(
tokenizer_class=_TOKENIZER_FOR_DOC,
checkpoint="albert-base-v2",
output_type=TFQuestionAnsweringModelOutput,
config_class=_CONFIG_FOR_DOC,
)
def call(
self,
inputs=None,
attention_mask=None,
token_type_ids=None,
position_ids=None,
head_mask=None,
inputs_embeds=None,
output_attentions=None,
output_hidden_states=None,
return_dict=None,
start_positions=None,
end_positions=None,
training=False,
):
r"""
start_positions (:obj:`tf.Tensor` of shape :obj:`(batch_size,)`, `optional`):
Labels for position (index) of the start of the labelled span for computing the token classification loss.
Positions are clamped to the length of the sequence (`sequence_length`).
Position outside of the sequence are not taken into account for computing the loss.
end_positions (:obj:`tf.Tensor` of shape :obj:`(batch_size,)`, `optional`):
Labels for position (index) of the end of the labelled span for computing the token classification loss.
Positions are clamped to the length of the sequence (`sequence_length`).
Position outside of the sequence are not taken into account for computing the loss.
"""
return_dict = return_dict if return_dict is not None else self.albert.return_dict
if isinstance(inputs, (tuple, list)):
start_positions = inputs[9] if len(inputs) > 9 else start_positions
end_positions = inputs[10] if len(inputs) > 10 else end_positions
if len(inputs) > 9:
inputs = inputs[:9]
elif isinstance(inputs, (dict, BatchEncoding)):
start_positions = inputs.pop("start_positions", start_positions)
end_positions = inputs.pop("end_positions", start_positions)
outputs = self.albert(
inputs,
attention_mask=attention_mask,
token_type_ids=token_type_ids,
position_ids=position_ids,
head_mask=head_mask,
inputs_embeds=inputs_embeds,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
training=training,
)
sequence_output = outputs[0]
logits = self.qa_outputs(sequence_output)
start_logits, end_logits = tf.split(logits, 2, axis=-1)
start_logits = tf.squeeze(start_logits, axis=-1)
end_logits = tf.squeeze(end_logits, axis=-1)
loss = None
if start_positions is not None and end_positions is not None:
labels = {"start_position": start_positions}
labels["end_position"] = end_positions
loss = self.compute_loss(labels, (start_logits, end_logits))
if not return_dict:
output = (start_logits, end_logits) + outputs[2:]
return ((loss,) + output) if loss is not None else output
return TFQuestionAnsweringModelOutput(
loss=loss,
start_logits=start_logits,
end_logits=end_logits,
hidden_states=outputs.hidden_states,
attentions=outputs.attentions,
)