# coding=utf-8
# Copyright 2018 The Google AI Language Team Authors and The 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 RoBERTa model. """
import tensorflow as tf
from .activations_tf import get_tf_activation
from .configuration_roberta import RobertaConfig
from .file_utils import (
MULTIPLE_CHOICE_DUMMY_INPUTS,
add_code_sample_docstrings,
add_start_docstrings,
add_start_docstrings_to_callable,
)
from .modeling_tf_bert import TFBertEmbeddings, TFBertMainLayer
from .modeling_tf_outputs import (
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_base import BatchEncoding
from .utils import logging
logger = logging.get_logger(__name__)
_CONFIG_FOR_DOC = "RobertaConfig"
_TOKENIZER_FOR_DOC = "RobertaTokenizer"
TF_ROBERTA_PRETRAINED_MODEL_ARCHIVE_LIST = [
"roberta-base",
"roberta-large",
"roberta-large-mnli",
"distilroberta-base",
# See all RoBERTa models at https://huggingface.co/models?filter=roberta
]
class TFRobertaEmbeddings(TFBertEmbeddings):
"""
Same as BertEmbeddings with a tiny tweak for positional embeddings indexing.
"""
def __init__(self, config, **kwargs):
super().__init__(config, **kwargs)
self.padding_idx = 1
def create_position_ids_from_input_ids(self, x):
"""Replace non-padding symbols with their position numbers. Position numbers begin at
padding_idx+1. Padding symbols are ignored. This is modified from fairseq's
`utils.make_positions`.
:param tf.Tensor x:
:return tf.Tensor:
"""
mask = tf.cast(tf.math.not_equal(x, self.padding_idx), dtype=tf.int32)
incremental_indicies = tf.math.cumsum(mask, axis=1) * mask
return incremental_indicies + self.padding_idx
def create_position_ids_from_inputs_embeds(self, inputs_embeds):
"""We are provided embeddings directly. We cannot infer which are padded so just generate
sequential position ids.
:param tf.Tensor inputs_embeds:
:return tf.Tensor:
"""
seq_length = shape_list(inputs_embeds)[1]
position_ids = tf.range(self.padding_idx + 1, seq_length + self.padding_idx + 1, dtype=tf.int32)[tf.newaxis, :]
return position_ids
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 position_ids is None:
if input_ids is not None:
# Create the position ids from the input token ids. Any padded tokens remain padded.
position_ids = self.create_position_ids_from_input_ids(input_ids)
else:
position_ids = self.create_position_ids_from_inputs_embeds(inputs_embeds)
return super()._embedding(input_ids, position_ids, token_type_ids, inputs_embeds, training=training)
@keras_serializable
class TFRobertaMainLayer(TFBertMainLayer):
"""
Same as TFBertMainLayer but uses TFRobertaEmbeddings.
"""
def __init__(self, config, **kwargs):
super().__init__(config, **kwargs)
self.embeddings = TFRobertaEmbeddings(config, name="embeddings")
class TFRobertaPreTrainedModel(TFPreTrainedModel):
"""An abstract class to handle weights initialization and
a simple interface for downloading and loading pretrained models.
"""
config_class = RobertaConfig
base_model_prefix = "roberta"
ROBERTA_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.
.. 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})`
Parameters:
config (:class:`~transformers.RobertaConfig`): 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.
"""
ROBERTA_INPUTS_DOCSTRING = r"""
Args:
input_ids (:obj:`Numpy array` or :obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length)`):
Indices of input sequence tokens in the vocabulary.
Indices can be obtained using :class:`transformers.RobertaTokenizer`.
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:`(batch_size, sequence_length)`, `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:`(batch_size, sequence_length)`, `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:`(batch_size, sequence_length)`, `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]``:
:obj:`1` indicates the head is **not masked**, :obj:`0` indicates the head is **masked**.
inputs_embeds (:obj:`Numpy array` or :obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length, embedding_dim)`, `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 RoBERTa Model transformer outputing raw hidden-states without any specific head on top.",
ROBERTA_START_DOCSTRING,
)
class TFRobertaModel(TFRobertaPreTrainedModel):
def __init__(self, config, *inputs, **kwargs):
super().__init__(config, *inputs, **kwargs)
self.roberta = TFRobertaMainLayer(config, name="roberta")
[docs] @add_start_docstrings_to_callable(ROBERTA_INPUTS_DOCSTRING)
@add_code_sample_docstrings(
tokenizer_class=_TOKENIZER_FOR_DOC,
checkpoint="roberta-base",
output_type=TFBaseModelOutputWithPooling,
config_class=_CONFIG_FOR_DOC,
)
def call(self, inputs, **kwargs):
outputs = self.roberta(inputs, **kwargs)
return outputs
class TFRobertaLMHead(tf.keras.layers.Layer):
"""Roberta Head for masked language modeling."""
def __init__(self, config, input_embeddings, **kwargs):
super().__init__(**kwargs)
self.vocab_size = config.vocab_size
self.dense = tf.keras.layers.Dense(
config.hidden_size, kernel_initializer=get_initializer(config.initializer_range), name="dense"
)
self.layer_norm = tf.keras.layers.LayerNormalization(epsilon=config.layer_norm_eps, name="layer_norm")
self.act = get_tf_activation("gelu")
# 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")
super().build(input_shape)
def call(self, features):
x = self.dense(features)
x = self.act(x)
x = self.layer_norm(x)
# project back to size of vocabulary with bias
x = self.decoder(x, mode="linear") + self.bias
return x
class TFRobertaClassificationHead(tf.keras.layers.Layer):
"""Head for sentence-level classification tasks."""
def __init__(self, config, **kwargs):
super().__init__(**kwargs)
self.dense = tf.keras.layers.Dense(
config.hidden_size,
kernel_initializer=get_initializer(config.initializer_range),
activation="tanh",
name="dense",
)
self.dropout = tf.keras.layers.Dropout(config.hidden_dropout_prob)
self.out_proj = tf.keras.layers.Dense(
config.num_labels, kernel_initializer=get_initializer(config.initializer_range), name="out_proj"
)
def call(self, features, training=False):
x = features[:, 0, :] # take <s> token (equiv. to [CLS])
x = self.dropout(x, training=training)
x = self.dense(x)
x = self.dropout(x, training=training)
x = self.out_proj(x)
return x
[docs]@add_start_docstrings(
"""RoBERTa Model transformer with a sequence classification/regression head on top (a linear layer
on top of the pooled output) e.g. for GLUE tasks. """,
ROBERTA_START_DOCSTRING,
)
class TFRobertaForSequenceClassification(TFRobertaPreTrainedModel, TFSequenceClassificationLoss):
def __init__(self, config, *inputs, **kwargs):
super().__init__(config, *inputs, **kwargs)
self.num_labels = config.num_labels
self.roberta = TFRobertaMainLayer(config, name="roberta")
self.classifier = TFRobertaClassificationHead(config, name="classifier")
[docs] @add_start_docstrings_to_callable(ROBERTA_INPUTS_DOCSTRING)
@add_code_sample_docstrings(
tokenizer_class=_TOKENIZER_FOR_DOC,
checkpoint="roberta-base",
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 :obj:`[0, ..., config.num_labels - 1]`.
If :obj:`config.num_labels == 1` a regression loss is computed (Mean-Square loss),
If :obj:`config.num_labels > 1` a classification loss is computed (Cross-Entropy).
"""
return_dict = return_dict if return_dict is not None else self.roberta.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.roberta(
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.classifier(sequence_output, training=training)
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(
"""RoBERTa 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. """,
ROBERTA_START_DOCSTRING,
)
class TFRobertaForTokenClassification(TFRobertaPreTrainedModel, TFTokenClassificationLoss):
def __init__(self, config, *inputs, **kwargs):
super().__init__(config, *inputs, **kwargs)
self.num_labels = config.num_labels
self.roberta = TFRobertaMainLayer(config, name="roberta")
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(ROBERTA_INPUTS_DOCSTRING)
@add_code_sample_docstrings(
tokenizer_class=_TOKENIZER_FOR_DOC,
checkpoint="roberta-base",
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.roberta.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.roberta(
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(
"""RoBERTa 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`). """,
ROBERTA_START_DOCSTRING,
)
class TFRobertaForQuestionAnswering(TFRobertaPreTrainedModel, TFQuestionAnsweringLoss):
def __init__(self, config, *inputs, **kwargs):
super().__init__(config, *inputs, **kwargs)
self.num_labels = config.num_labels
self.roberta = TFRobertaMainLayer(config, name="roberta")
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(ROBERTA_INPUTS_DOCSTRING)
@add_code_sample_docstrings(
tokenizer_class=_TOKENIZER_FOR_DOC,
checkpoint="roberta-base",
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.roberta.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.roberta(
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,
)