keras-bert是Keras框架加载BERT模型的Python第三方模块，在之前的文章中，笔者介绍了如何使用keras-bret来实现不同的NLP任务，比如：
- NLP（三十四）使用keras-bert实现序列标注任务
- NLP（三十五）使用keras-bert实现文本多分类任务
- NLP（三十六）使用keras-bert实现文本多标签分类任务
- NLP（三十七）使用keras-bert实现英语序列标注任务
- NLP（三十九）使用keras-bert实现完形填空及简单的文本纠错功能
- NLP（四十二）人物关系分类的再次尝试
本文将介绍两种使用keras-bert加载BERT模型的方法。使用的Python环境如下：
python==3.7.0
tensorflow==1.14.0
Keras==2.2.4
keras-bert==0.83.0
加载的模型为Google官方发布的BERT中文预训练模型。创建的模型为BERT+Bi-LSTM+CRF，其中对BERT进行微调。
方法1
方法1的完整代码如下：
# -*- coding: utf-8 -*-
from keras.layers import *
from keras.models import Model
from keras.utils import plot_model
from keras_bert import load_trained_model_from_checkpoint
from keras_contrib.layers import CRF
# 创建BERT-BiLSTM-CRF模型
model_path = "./chinese_L-12_H-768_A-12/"
bert = load_trained_model_from_checkpoint(
model_path + "bert_config.json",
model_path + "bert_model.ckpt",
seq_len=128
)
# make bert layer trainable
for layer in bert.layers:
layer.trainable = True
x1 = Input(shape=(None,))
x2 = Input(shape=(None,))
bert_out = bert([x1, x2])
lstm_out = Bidirectional(LSTM(64,
return_sequences=True,
dropout=0.2,
recurrent_dropout=0.2))(bert_out)
crf_out = CRF(8, sparse_target=True)(lstm_out)
model = Model([x1, x2], crf_out)
model.summary()
plot_model(model, to_file="model.png")
输出的模型结构如下：
__________________________________________________________________________________________________
Layer (type) Output Shape Param # Connected to
==================================================================================================
input_1 (InputLayer) (None, None) 0
__________________________________________________________________________________________________
input_2 (InputLayer) (None, None) 0
__________________________________________________________________________________________________
model_2 (Model) multiple 101382144 input_1[0][0]
input_2[0][0]
__________________________________________________________________________________________________
bidirectional_1 (Bidirectional) (None, None, 128) 426496 model_2[1][0]
__________________________________________________________________________________________________
crf_1 (CRF) (None, None, 8) 1112 bidirectional_1[0][0]
==================================================================================================
Total params: 101,809,752
Trainable params: 101,809,752
Non-trainable params: 0
__________________________________________________________________________________________________
可以看到，该方法加载BERT，会把BERT模型整体当做一个输出形状为multiple的层，我们无法得知BERT模型的具体层信息，好处是我们的模型结构会显得比较简单（略去了BERT的细节）。
方法2
方法2加载BERT模型的Python代码如下：
# -*- coding: utf-8 -*-
from keras.layers import *
from keras.models import Model
from keras.utils import plot_model
from keras_bert import load_trained_model_from_checkpoint
from keras_contrib.layers import CRF
# 创建BERT-BiLSTM-CRF模型
model_path = "./chinese_L-12_H-768_A-12/"
bert = load_trained_model_from_checkpoint(
model_path + "bert_config.json",
model_path + "bert_model.ckpt",
seq_len=128
)
# make bert layer trainable
for layer in bert.layers:
layer.trainable = True
lstm_out = Bidirectional(LSTM(64,
return_sequences=True,
dropout=0.2,
recurrent_dropout=0.2))(bert.output)
crf_out = CRF(8, sparse_target=True)(lstm_out)
model = Model(bert.input, crf_out)
model.summary()
plot_model(model, to_file="model.png")
输出的模型结构如下：
__________________________________________________________________________________________________
Layer (type) Output Shape Param # Connected to
==================================================================================================
Input-Token (InputLayer) (None, 128) 0
__________________________________________________________________________________________________
Input-Segment (InputLayer) (None, 128) 0
__________________________________________________________________________________________________
Embedding-Token (TokenEmbedding [(None, 128, 768), ( 16226304 Input-Token[0][0]
__________________________________________________________________________________________________
Embedding-Segment (Embedding) (None, 128, 768) 1536 Input-Segment[0][0]
__________________________________________________________________________________________________
Embedding-Token-Segment (Add) (None, 128, 768) 0 Embedding-Token[0][0]
Embedding-Segment[0][0]
__________________________________________________________________________________________________
Embedding-Position (PositionEmb (None, 128, 768) 98304 Embedding-Token-Segment[0][0]
__________________________________________________________________________________________________
Embedding-Dropout (Dropout) (None, 128, 768) 0 Embedding-Position[0][0]
__________________________________________________________________________________________________
Embedding-Norm (LayerNormalizat (None, 128, 768) 1536 Embedding-Dropout[0][0]
__________________________________________________________________________________________________
Encoder-1-MultiHeadSelfAttentio (None, 128, 768) 2362368 Embedding-Norm[0][0]
__________________________________________________________________________________________________
Encoder-1-MultiHeadSelfAttentio (None, 128, 768) 0 Encoder-1-MultiHeadSelfAttention[
__________________________________________________________________________________________________
Encoder-1-MultiHeadSelfAttentio (None, 128, 768) 0 Embedding-Norm[0][0]
Encoder-1-MultiHeadSelfAttention-
__________________________________________________________________________________________________
Encoder-1-MultiHeadSelfAttentio (None, 128, 768) 1536 Encoder-1-MultiHeadSelfAttention-
__________________________________________________________________________________________________
Encoder-1-FeedForward (FeedForw (None, 128, 768) 4722432 Encoder-1-MultiHeadSelfAttention-
__________________________________________________________________________________________________
Encoder-1-FeedForward-Dropout ( (None, 128, 768) 0 Encoder-1-FeedForward[0][0]
__________________________________________________________________________________________________
Encoder-1-FeedForward-Add (Add) (None, 128, 768) 0 Encoder-1-MultiHeadSelfAttention-
Encoder-1-FeedForward-Dropout[0][
__________________________________________________________________________________________________
Encoder-1-FeedForward-Norm (Lay (None, 128, 768) 1536 Encoder-1-FeedForward-Add[0][0]
__________________________________________________________________________________________________
Encoder-2-MultiHeadSelfAttentio (None, 128, 768) 2362368 Encoder-1-FeedForward-Norm[0][0]
__________________________________________________________________________________________________
Encoder-2-MultiHeadSelfAttentio (None, 128, 768) 0 Encoder-2-MultiHeadSelfAttention[
__________________________________________________________________________________________________
Encoder-2-MultiHeadSelfAttentio (None, 128, 768) 0 Encoder-1-FeedForward-Norm[0][0]
Encoder-2-MultiHeadSelfAttention-
__________________________________________________________________________________________________
Encoder-2-MultiHeadSelfAttentio (None, 128, 768) 1536 Encoder-2-MultiHeadSelfAttention-
__________________________________________________________________________________________________
Encoder-2-FeedForward (FeedForw (None, 128, 768) 4722432 Encoder-2-MultiHeadSelfAttention-
__________________________________________________________________________________________________
Encoder-2-FeedForward-Dropout ( (None, 128, 768) 0 Encoder-2-FeedForward[0][0]
__________________________________________________________________________________________________
Encoder-2-FeedForward-Add (Add) (None, 128, 768) 0 Encoder-2-MultiHeadSelfAttention-
Encoder-2-FeedForward-Dropout[0][
__________________________________________________________________________________________________
Encoder-2-FeedForward-Norm (Lay (None, 128, 768) 1536 Encoder-2-FeedForward-Add[0][0]
__________________________________________________________________________________________________
Encoder-3-MultiHeadSelfAttentio (None, 128, 768) 2362368 Encoder-2-FeedForward-Norm[0][0]
__________________________________________________________________________________________________
Encoder-3-MultiHeadSelfAttentio (None, 128, 768) 0 Encoder-3-MultiHeadSelfAttention[
__________________________________________________________________________________________________
Encoder-3-MultiHeadSelfAttentio (None, 128, 768) 0 Encoder-2-FeedForward-Norm[0][0]
Encoder-3-MultiHeadSelfAttention-
__________________________________________________________________________________________________
Encoder-3-MultiHeadSelfAttentio (None, 128, 768) 1536 Encoder-3-MultiHeadSelfAttention-
__________________________________________________________________________________________________
Encoder-3-FeedForward (FeedForw (None, 128, 768) 4722432 Encoder-3-MultiHeadSelfAttention-
__________________________________________________________________________________________________
Encoder-3-FeedForward-Dropout ( (None, 128, 768) 0 Encoder-3-FeedForward[0][0]
__________________________________________________________________________________________________
Encoder-3-FeedForward-Add (Add) (None, 128, 768) 0 Encoder-3-MultiHeadSelfAttention-
Encoder-3-FeedForward-Dropout[0][
__________________________________________________________________________________________________
Encoder-3-FeedForward-Norm (Lay (None, 128, 768) 1536 Encoder-3-FeedForward-Add[0][0]
__________________________________________________________________________________________________
Encoder-4-MultiHeadSelfAttentio (None, 128, 768) 2362368 Encoder-3-FeedForward-Norm[0][0]
__________________________________________________________________________________________________
Encoder-4-MultiHeadSelfAttentio (None, 128, 768) 0 Encoder-4-MultiHeadSelfAttention[
__________________________________________________________________________________________________
Encoder-4-MultiHeadSelfAttentio (None, 128, 768) 0 Encoder-3-FeedForward-Norm[0][0]
Encoder-4-MultiHeadSelfAttention-
__________________________________________________________________________________________________
Encoder-4-MultiHeadSelfAttentio (None, 128, 768) 1536 Encoder-4-MultiHeadSelfAttention-
__________________________________________________________________________________________________
Encoder-4-FeedForward (FeedForw (None, 128, 768) 4722432 Encoder-4-MultiHeadSelfAttention-
__________________________________________________________________________________________________
Encoder-4-FeedForward-Dropout ( (None, 128, 768) 0 Encoder-4-FeedForward[0][0]
__________________________________________________________________________________________________
Encoder-4-FeedForward-Add (Add) (None, 128, 768) 0 Encoder-4-MultiHeadSelfAttention-
Encoder-4-FeedForward-Dropout[0][
__________________________________________________________________________________________________
Encoder-4-FeedForward-Norm (Lay (None, 128, 768) 1536 Encoder-4-FeedForward-Add[0][0]
__________________________________________________________________________________________________
Encoder-5-MultiHeadSelfAttentio (None, 128, 768) 2362368 Encoder-4-FeedForward-Norm[0][0]
__________________________________________________________________________________________________
Encoder-5-MultiHeadSelfAttentio (None, 128, 768) 0 Encoder-5-MultiHeadSelfAttention[
__________________________________________________________________________________________________
Encoder-5-MultiHeadSelfAttentio (None, 128, 768) 0 Encoder-4-FeedForward-Norm[0][0]
Encoder-5-MultiHeadSelfAttention-
__________________________________________________________________________________________________
Encoder-5-MultiHeadSelfAttentio (None, 128, 768) 1536 Encoder-5-MultiHeadSelfAttention-
__________________________________________________________________________________________________
Encoder-5-FeedForward (FeedForw (None, 128, 768) 4722432 Encoder-5-MultiHeadSelfAttention-
__________________________________________________________________________________________________
Encoder-5-FeedForward-Dropout ( (None, 128, 768) 0 Encoder-5-FeedForward[0][0]
__________________________________________________________________________________________________
Encoder-5-FeedForward-Add (Add) (None, 128, 768) 0 Encoder-5-MultiHeadSelfAttention-
Encoder-5-FeedForward-Dropout[0][
__________________________________________________________________________________________________
Encoder-5-FeedForward-Norm (Lay (None, 128, 768) 1536 Encoder-5-FeedForward-Add[0][0]
__________________________________________________________________________________________________
Encoder-6-MultiHeadSelfAttentio (None, 128, 768) 2362368 Encoder-5-FeedForward-Norm[0][0]
__________________________________________________________________________________________________
Encoder-6-MultiHeadSelfAttentio (None, 128, 768) 0 Encoder-6-MultiHeadSelfAttention[
__________________________________________________________________________________________________
Encoder-6-MultiHeadSelfAttentio (None, 128, 768) 0 Encoder-5-FeedForward-Norm[0][0]
Encoder-6-MultiHeadSelfAttention-
__________________________________________________________________________________________________
Encoder-6-MultiHeadSelfAttentio (None, 128, 768) 1536 Encoder-6-MultiHeadSelfAttention-
__________________________________________________________________________________________________
Encoder-6-FeedForward (FeedForw (None, 128, 768) 4722432 Encoder-6-MultiHeadSelfAttention-
__________________________________________________________________________________________________
Encoder-6-FeedForward-Dropout ( (None, 128, 768) 0 Encoder-6-FeedForward[0][0]
__________________________________________________________________________________________________
Encoder-6-FeedForward-Add (Add) (None, 128, 768) 0 Encoder-6-MultiHeadSelfAttention-
Encoder-6-FeedForward-Dropout[0][
__________________________________________________________________________________________________
Encoder-6-FeedForward-Norm (Lay (None, 128, 768) 1536 Encoder-6-FeedForward-Add[0][0]
__________________________________________________________________________________________________
Encoder-7-MultiHeadSelfAttentio (None, 128, 768) 2362368 Encoder-6-FeedForward-Norm[0][0]
__________________________________________________________________________________________________
Encoder-7-MultiHeadSelfAttentio (None, 128, 768) 0 Encoder-7-MultiHeadSelfAttention[
__________________________________________________________________________________________________
Encoder-7-MultiHeadSelfAttentio (None, 128, 768) 0 Encoder-6-FeedForward-Norm[0][0]
Encoder-7-MultiHeadSelfAttention-
__________________________________________________________________________________________________
Encoder-7-MultiHeadSelfAttentio (None, 128, 768) 1536 Encoder-7-MultiHeadSelfAttention-
__________________________________________________________________________________________________
Encoder-7-FeedForward (FeedForw (None, 128, 768) 4722432 Encoder-7-MultiHeadSelfAttention-
__________________________________________________________________________________________________
Encoder-7-FeedForward-Dropout ( (None, 128, 768) 0 Encoder-7-FeedForward[0][0]
__________________________________________________________________________________________________
Encoder-7-FeedForward-Add (Add) (None, 128, 768) 0 Encoder-7-MultiHeadSelfAttention-
Encoder-7-FeedForward-Dropout[0][
__________________________________________________________________________________________________
Encoder-7-FeedForward-Norm (Lay (None, 128, 768) 1536 Encoder-7-FeedForward-Add[0][0]
__________________________________________________________________________________________________
Encoder-8-MultiHeadSelfAttentio (None, 128, 768) 2362368 Encoder-7-FeedForward-Norm[0][0]
__________________________________________________________________________________________________
Encoder-8-MultiHeadSelfAttentio (None, 128, 768) 0 Encoder-8-MultiHeadSelfAttention[
__________________________________________________________________________________________________
Encoder-8-MultiHeadSelfAttentio (None, 128, 768) 0 Encoder-7-FeedForward-Norm[0][0]
Encoder-8-MultiHeadSelfAttention-
__________________________________________________________________________________________________
Encoder-8-MultiHeadSelfAttentio (None, 128, 768) 1536 Encoder-8-MultiHeadSelfAttention-
__________________________________________________________________________________________________
Encoder-8-FeedForward (FeedForw (None, 128, 768) 4722432 Encoder-8-MultiHeadSelfAttention-
__________________________________________________________________________________________________
Encoder-8-FeedForward-Dropout ( (None, 128, 768) 0 Encoder-8-FeedForward[0][0]
__________________________________________________________________________________________________
Encoder-8-FeedForward-Add (Add) (None, 128, 768) 0 Encoder-8-MultiHeadSelfAttention-
Encoder-8-FeedForward-Dropout[0][
__________________________________________________________________________________________________
Encoder-8-FeedForward-Norm (Lay (None, 128, 768) 1536 Encoder-8-FeedForward-Add[0][0]
__________________________________________________________________________________________________
Encoder-9-MultiHeadSelfAttentio (None, 128, 768) 2362368 Encoder-8-FeedForward-Norm[0][0]
__________________________________________________________________________________________________
Encoder-9-MultiHeadSelfAttentio (None, 128, 768) 0 Encoder-9-MultiHeadSelfAttention[
__________________________________________________________________________________________________
Encoder-9-MultiHeadSelfAttentio (None, 128, 768) 0 Encoder-8-FeedForward-Norm[0][0]
Encoder-9-MultiHeadSelfAttention-
__________________________________________________________________________________________________
Encoder-9-MultiHeadSelfAttentio (None, 128, 768) 1536 Encoder-9-MultiHeadSelfAttention-
__________________________________________________________________________________________________
Encoder-9-FeedForward (FeedForw (None, 128, 768) 4722432 Encoder-9-MultiHeadSelfAttention-
__________________________________________________________________________________________________
Encoder-9-FeedForward-Dropout ( (None, 128, 768) 0 Encoder-9-FeedForward[0][0]
__________________________________________________________________________________________________
Encoder-9-FeedForward-Add (Add) (None, 128, 768) 0 Encoder-9-MultiHeadSelfAttention-
Encoder-9-FeedForward-Dropout[0][
__________________________________________________________________________________________________
Encoder-9-FeedForward-Norm (Lay (None, 128, 768) 1536 Encoder-9-FeedForward-Add[0][0]
__________________________________________________________________________________________________
Encoder-10-MultiHeadSelfAttenti (None, 128, 768) 2362368 Encoder-9-FeedForward-Norm[0][0]
__________________________________________________________________________________________________
Encoder-10-MultiHeadSelfAttenti (None, 128, 768) 0 Encoder-10-MultiHeadSelfAttention
__________________________________________________________________________________________________
Encoder-10-MultiHeadSelfAttenti (None, 128, 768) 0 Encoder-9-FeedForward-Norm[0][0]
Encoder-10-MultiHeadSelfAttention
__________________________________________________________________________________________________
Encoder-10-MultiHeadSelfAttenti (None, 128, 768) 1536 Encoder-10-MultiHeadSelfAttention
__________________________________________________________________________________________________
Encoder-10-FeedForward (FeedFor (None, 128, 768) 4722432 Encoder-10-MultiHeadSelfAttention
__________________________________________________________________________________________________
Encoder-10-FeedForward-Dropout (None, 128, 768) 0 Encoder-10-FeedForward[0][0]
__________________________________________________________________________________________________
Encoder-10-FeedForward-Add (Add (None, 128, 768) 0 Encoder-10-MultiHeadSelfAttention
Encoder-10-FeedForward-Dropout[0]
__________________________________________________________________________________________________
Encoder-10-FeedForward-Norm (La (None, 128, 768) 1536 Encoder-10-FeedForward-Add[0][0]
__________________________________________________________________________________________________
Encoder-11-MultiHeadSelfAttenti (None, 128, 768) 2362368 Encoder-10-FeedForward-Norm[0][0]
__________________________________________________________________________________________________
Encoder-11-MultiHeadSelfAttenti (None, 128, 768) 0 Encoder-11-MultiHeadSelfAttention
__________________________________________________________________________________________________
Encoder-11-MultiHeadSelfAttenti (None, 128, 768) 0 Encoder-10-FeedForward-Norm[0][0]
Encoder-11-MultiHeadSelfAttention
__________________________________________________________________________________________________
Encoder-11-MultiHeadSelfAttenti (None, 128, 768) 1536 Encoder-11-MultiHeadSelfAttention
__________________________________________________________________________________________________
Encoder-11-FeedForward (FeedFor (None, 128, 768) 4722432 Encoder-11-MultiHeadSelfAttention
__________________________________________________________________________________________________
Encoder-11-FeedForward-Dropout (None, 128, 768) 0 Encoder-11-FeedForward[0][0]
__________________________________________________________________________________________________
Encoder-11-FeedForward-Add (Add (None, 128, 768) 0 Encoder-11-MultiHeadSelfAttention
Encoder-11-FeedForward-Dropout[0]
__________________________________________________________________________________________________
Encoder-11-FeedForward-Norm (La (None, 128, 768) 1536 Encoder-11-FeedForward-Add[0][0]
__________________________________________________________________________________________________
Encoder-12-MultiHeadSelfAttenti (None, 128, 768) 2362368 Encoder-11-FeedForward-Norm[0][0]
__________________________________________________________________________________________________
Encoder-12-MultiHeadSelfAttenti (None, 128, 768) 0 Encoder-12-MultiHeadSelfAttention
__________________________________________________________________________________________________
Encoder-12-MultiHeadSelfAttenti (None, 128, 768) 0 Encoder-11-FeedForward-Norm[0][0]
Encoder-12-MultiHeadSelfAttention
__________________________________________________________________________________________________
Encoder-12-MultiHeadSelfAttenti (None, 128, 768) 1536 Encoder-12-MultiHeadSelfAttention
__________________________________________________________________________________________________
Encoder-12-FeedForward (FeedFor (None, 128, 768) 4722432 Encoder-12-MultiHeadSelfAttention
__________________________________________________________________________________________________
Encoder-12-FeedForward-Dropout (None, 128, 768) 0 Encoder-12-FeedForward[0][0]
__________________________________________________________________________________________________
Encoder-12-FeedForward-Add (Add (None, 128, 768) 0 Encoder-12-MultiHeadSelfAttention
Encoder-12-FeedForward-Dropout[0]
__________________________________________________________________________________________________
Encoder-12-FeedForward-Norm (La (None, 128, 768) 1536 Encoder-12-FeedForward-Add[0][0]
__________________________________________________________________________________________________
bidirectional_1 (Bidirectional) (None, 128, 128) 426496 Encoder-12-FeedForward-Norm[0][0]
__________________________________________________________________________________________________
crf_1 (CRF) (None, 128, 8) 1112 bidirectional_1[0][0]
==================================================================================================
Total params: 101,809,752
Trainable params: 101,809,752
Non-trainable params: 0
__________________________________________________________________________________________________
可以看到，该方法加载BERT模型，可以完整地看到BERT具体层信息，而不是把BERT模型当成一个层来看，更像是BERT finetune的感觉。
总结
本文较为简单，介绍了两种使用keras-bert加载BERT模型的方法。之所以笔者在此介绍这些加载方法，是为了后续方便使用对抗训练FGM来增加模型效果，FGM对抗训练需要我们对Embedding层做扰动。
本文到此结束，感谢大家的阅读~
2021年3月31日于上海浦东~