# Copyright 2019 The Texar Authors. 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.
"""
BERT encoder.
"""
from typing import Optional, Union
import torch
from torch import nn
from texar.torch.core import layers
from texar.torch.modules.embedders.embedders import WordEmbedder
from texar.torch.modules.embedders.position_embedders import PositionEmbedder
from texar.torch.modules.encoders.encoder_base import EncoderBase
from texar.torch.modules.encoders.transformer_encoder import TransformerEncoder
from texar.torch.modules.pretrained.bert import PretrainedBERTMixin
__all__ = [
"BERTEncoder",
]
[docs]class BERTEncoder(EncoderBase, PretrainedBERTMixin):
r"""Raw BERT Transformer for encoding sequences. Please see
:class:`~texar.torch.modules.PretrainedBERTMixin` for a brief description
of BERT.
This module basically stacks
:class:`~texar.torch.modules.WordEmbedder`,
:class:`~texar.torch.modules.PositionEmbedder`,
:class:`~texar.torch.modules.TransformerEncoder` and a dense
pooler.
Args:
pretrained_model_name (optional): a `str`, the name
of pre-trained model (e.g., ``bert-base-uncased``). Please refer to
:class:`~texar.torch.modules.PretrainedBERTMixin` for
all supported models.
If `None`, the model name in :attr:`hparams` is used.
cache_dir (optional): the path to a folder in which the
pre-trained models will be cached. If `None` (default),
a default directory (``texar_data`` folder under user's home
directory) will be used.
hparams (dict or HParams, optional): Hyperparameters. Missing
hyperparameter will be set to default values. See
:meth:`default_hparams` for the hyperparameter structure
and default values.
"""
def __init__(self,
pretrained_model_name: Optional[str] = None,
cache_dir: Optional[str] = None,
hparams=None):
super().__init__(hparams=hparams)
self.load_pretrained_config(pretrained_model_name, cache_dir)
# Word embedding
self.word_embedder = WordEmbedder(
vocab_size=self._hparams.vocab_size,
hparams=self._hparams.embed)
# Segment embedding for each type of tokens
self.segment_embedder = None
if self._hparams.get('type_vocab_size', 0) > 0:
if self.pretrained_model_name is not None and \
self.pretrained_model_name.startswith('spanbert'):
# Do not construct segment_embedder for SpanBERT
pass
else:
self.segment_embedder = WordEmbedder(
vocab_size=self._hparams.type_vocab_size,
hparams=self._hparams.segment_embed)
# Position embedding
self.position_embedder = PositionEmbedder(
position_size=self._hparams.position_size,
hparams=self._hparams.position_embed)
# The BERT encoder (a TransformerEncoder)
self.encoder = TransformerEncoder(hparams=self._hparams.encoder)
self.pooler = nn.Sequential(
nn.Linear(self._hparams.hidden_size, self._hparams.hidden_size),
nn.Tanh())
self.init_pretrained_weights()
[docs] def reset_parameters(self):
initialize = layers.get_initializer(self._hparams.initializer)
if initialize is not None:
# Do not re-initialize LayerNorm modules.
for name, param in self.named_parameters():
if name.split('.')[-1] == 'weight' and 'layer_norm' not in name:
initialize(param)
[docs] @staticmethod
def default_hparams():
r"""Returns a dictionary of hyperparameters with default values.
* The encoder arch is determined by the constructor argument
:attr:`pretrained_model_name` if it's specified. In this case,
`hparams` are ignored.
* Otherwise, the encoder arch is determined by
`hparams['pretrained_model_name']` if it's specified. All other
configurations in `hparams` are ignored.
* If the above two are `None`, the encoder arch is defined by the
configurations in `hparams` and weights are randomly initialized.
.. code-block:: python
{
"pretrained_model_name": "bert-base-uncased",
"embed": {
"dim": 768,
"name": "word_embeddings"
},
"vocab_size": 30522,
"segment_embed": {
"dim": 768,
"name": "token_type_embeddings"
},
"type_vocab_size": 2,
"position_embed": {
"dim": 768,
"name": "position_embeddings"
},
"position_size": 512,
"encoder": {
"dim": 768,
"embedding_dropout": 0.1,
"multihead_attention": {
"dropout_rate": 0.1,
"name": "self",
"num_heads": 12,
"num_units": 768,
"output_dim": 768,
"use_bias": True
},
"name": "encoder",
"num_blocks": 12,
"eps": 1e-12,
"poswise_feedforward": {
"layers": [
{
"kwargs": {
"in_features": 768,
"out_features": 3072,
"bias": True
},
"type": "Linear"
},
{"type": "BertGELU"},
{
"kwargs": {
"in_features": 3072,
"out_features": 768,
"bias": True
},
"type": "Linear"
}
]
},
"residual_dropout": 0.1,
"use_bert_config": True
},
"hidden_size": 768,
"initializer": None,
"name": "bert_encoder",
}
Here:
The default parameters are values for uncased BERT-Base model.
`"pretrained_model_name"`: str or None
The name of the pre-trained BERT model. If None, the model
will be randomly initialized.
`"embed"`: dict
Hyperparameters for word embedding layer.
`"vocab_size"`: int
The vocabulary size of `inputs` in BERT model.
`"segment_embed"`: dict
Hyperparameters for segment embedding layer.
`"type_vocab_size"`: int
The vocabulary size of the `segment_ids` passed into `BertModel`.
`"position_embed"`: dict
Hyperparameters for position embedding layer.
`"position_size"`: int
The maximum sequence length that this model might ever be used with.
`"encoder"`: dict
Hyperparameters for the TransformerEncoder.
See :func:`~texar.torch.modules.TransformerEncoder.default_hparams`
for details.
`"hidden_size"`: int
Size of the pooler dense layer.
`"eps"`: float
Epsilon values for layer norm layers.
`"initializer"`: dict, optional
Hyperparameters of the default initializer that initializes
variables created in this module.
See :func:`~texar.torch.core.get_initializer` for details.
`"name"`: str
Name of the module.
"""
return {
'pretrained_model_name': 'bert-base-uncased',
'embed': {
'dim': 768,
'name': 'word_embeddings'
},
'vocab_size': 30522,
'segment_embed': {
'dim': 768,
'name': 'token_type_embeddings'
},
'type_vocab_size': 2,
'position_embed': {
'dim': 768,
'name': 'position_embeddings'
},
'position_size': 512,
'encoder': {
'dim': 768,
'embedding_dropout': 0.1,
'multihead_attention': {
'dropout_rate': 0.1,
'name': 'self',
'num_heads': 12,
'num_units': 768,
'output_dim': 768,
'use_bias': True
},
'name': 'encoder',
'num_blocks': 12,
'eps': 1e-12,
'poswise_feedforward': {
'layers': [
{
'kwargs': {
'in_features': 768,
'out_features': 3072,
'bias': True
},
'type': 'Linear'
},
{"type": "BertGELU"},
{
'kwargs': {
'in_features': 3072,
'out_features': 768,
'bias': True
},
'type': 'Linear'
}
]
},
'residual_dropout': 0.1,
'use_bert_config': True
},
'hidden_size': 768,
'initializer': None,
'name': 'bert_encoder',
'@no_typecheck': ['pretrained_model_name']
}
[docs] def forward(self, # type: ignore
inputs: Union[torch.Tensor, torch.LongTensor],
sequence_length: Optional[torch.LongTensor] = None,
segment_ids: Optional[torch.LongTensor] = None):
r"""Encodes the inputs. Note that the SpanBERT model does not use
segmentation embedding. As a result, SpanBERT does not require
`segment_ids` as an input when you use pre-trained SpanBERT checkpoint
files.
Args:
inputs: Either a **2D Tensor** of shape `[batch_size, max_time]`,
containing the ids of tokens in input sequences, or
a **3D Tensor** of shape `[batch_size, max_time, vocab_size]`,
containing soft token ids (i.e., weights or probabilities)
used to mix the embedding vectors.
segment_ids (optional): A 2D Tensor of shape
`[batch_size, max_time]`, containing the segment ids
of tokens in input sequences. If `None` (default), a
tensor with all elements set to zero is used.
sequence_length (optional): A 1D Tensor of shape `[batch_size]`.
Input tokens beyond respective sequence lengths are masked
out automatically.
Returns:
A pair :attr:`(outputs, pooled_output)`
- :attr:`outputs`: A Tensor of shape
`[batch_size, max_time, dim]` containing the encoded vectors.
- :attr:`pooled_output`: A Tensor of size
`[batch_size, hidden_size]` which is the output of a pooler
pre-trained on top of the hidden state associated to the first
character of the input (`CLS`), see BERT's paper.
"""
if inputs.dim() == 2:
word_embeds = self.word_embedder(ids=inputs)
elif inputs.dim() == 3:
word_embeds = self.word_embedder(soft_ids=inputs)
else:
raise ValueError("'inputs' should be a 2D or 3D tensor.")
batch_size = inputs.size(0)
pos_length = inputs.new_full((batch_size,), inputs.size(1),
dtype=torch.int64)
pos_embeds = self.position_embedder(sequence_length=pos_length)
if self.segment_embedder is not None:
if segment_ids is None:
segment_ids = torch.zeros((inputs.size(0), inputs.size(1)),
dtype=torch.long,
device=inputs.device)
segment_embeds = self.segment_embedder(segment_ids)
inputs_embeds = word_embeds + segment_embeds + pos_embeds
else:
inputs_embeds = word_embeds + pos_embeds
if sequence_length is None:
sequence_length = inputs.new_full((batch_size,), inputs.size(1),
dtype=torch.int64)
output = self.encoder(inputs_embeds, sequence_length)
# taking the hidden state corresponding to the first token.
first_token_tensor = output[:, 0, :]
pooled_output = self.pooler(first_token_tensor)
return output, pooled_output
@property
def output_size(self):
r"""The feature size of :meth:`forward` output
:attr:`pooled_output`.
"""
return self._hparams.hidden_size