学习参考杨老师的博客，请支持原文

一、Keras_bert 基础知识

1.1、kert_bert库安装

Kert_bert库是基于Keras库对Bert进行封装的Python库，可以实现直接对官方预训练权重的调用训练

但是该库在安装使用方面对于版本的兼容性要求比较高，python版本、tensorfolw版本、Keras版本、Kert_bert版本的相互之间需要相互的兼容。

本人由于习惯于使用Python3.6版本，基本都是在3.6版本上，安装tensoflow与keras库。tensorflow与keras的对应关系：

tensorflow2.0以上会出现和keras版本问题,在安装Kert_bert库出现很多次的Keras和keras_bert模型不兼容问题,多次搜索调试后选择版本如下：

tensorflow： 2.2.0
keras: 2.4.3
keras_bert: 0.89.0
Python 3.6.13

该版本，目前使用未出现问题。

1.2 Tokenizer文本拆分

keras_bert内嵌的Tokenizer可以实现对于字单位级别的拆分，并且生成相应的id，通过字典存放相应的token和id映射。

Tookenizer需要传入相应的字符所对应的id映射字典token_dict，其中[CLS]代表句子的起始，[SEP]代表句子的结束，[UNK]代表该该词未登录token_dict

from keras_bert import Tokenizer

token_dict = {
    '[CLS]': 0,
    '[SEP]': 1,
    '今': 2,
    '天': 3,
    '气': 4,
    '很':5,
    '好':6,
    '[UNK]':7 ,
}

#生成分词器-tokenizer对象
tokenizer = Tokenizer(token_dict)

#用生成的分词器对象对句子拆分
print('拆分：',tokenizer.tokenize('今天天气很好'))  

#用生成的分词器对句子的字符进行ID映射编码
indices, segments = tokenizer.encode('今天天气很好')
print('索引：',indices)    #字对应索引

# -----------------------多句子训练------------------------------------
print('-----------传入两句话，max_len=10------------------------------')
tokenizer.tokenize(first='今天天气很好', second='今天天蓝很不错')

indices, segments = tokenizer.encode(first='今天天气很好', second='今天天蓝很不错', max_len=10)
print(indices)
#索引对应位置上字属于第一句话还是第二句话 
print(segments)
print('-------------传入两句话，max_len=20-------------------------')
indices, segments = tokenizer.encode(first='今天天气很好', second='今天天蓝很不错', max_len=20)
print(indices)
#索引对应位置上字属于第一句话还是第二句话 
print(segments)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36

拆分： [‘[CLS]’, ‘今’, ‘天’, ‘天’, ‘气’, ‘很’, ‘好’, ‘[SEP]’]
索引： [0, 2, 3, 3, 4, 5, 6, 1]
-----------传入两句话，max_len=10------------------------------
[0, 2, 3, 3, 4, 1, 2, 3, 3, 1]
[0, 0, 0, 0, 0, 0, 1, 1, 1, 1]
-------------传入两句话，max_len=20-------------------------
[0, 2, 3, 3, 4, 5, 6, 1, 2, 3, 3, 7, 5, 7, 7, 1, 0, 0, 0, 0]
[0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0]

可以看出当出现token_dict未登录的词就转成了nuk，编码为7
当句子长度超过maxlen就都会被裁切掉
当句子长度不足maxlen，就会填充0

1.3 训练和使用

可见在使用之前需要先获取token_dict，才能进行下一步,用keras_bert中的，get_base_dict可以获取初始的tokendict，再对每个句子中的词进行编码

import keras
from keras_bert import get_base_dict, get_model, compile_model, gen_batch_inputs

#输入样本：toy玩具
sentence_pairs = [
    [['all', 'work', 'and', 'no', 'play'], ['makes', 'jack', 'a', 'dull', 'boy']],
    [['from', 'the', 'day', 'forth'], ['my', 'arm', 'changed']],
    [['and', 'a', 'voice', 'echoed'], ['power', 'give', 'me', 'more', 'power']],
]


#构建token词典
token_dict = get_base_dict()  # A dict that contains some special tokens
print('初始的tokendict:',token_dict)

for pairs in sentence_pairs:
    print('pairs:',pairs)
    print('pairs[0] + pairs[1]:',pairs[0] + pairs[1])
    for token in pairs[0] + pairs[1]:

        print('token:',token)
        if token not in token_dict:
            token_dict[token] = len(token_dict)
token_list = list(token_dict.keys())  # Used for selecting a random word
print('-----------------token_dict------------------')
print(token_dict)
print('-----------------token_list------------------')
print('token_list:',token_list)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28

初始的tokendict: {‘’: 0, ‘[UNK]’: 1, ‘[CLS]’: 2, ‘[SEP]’: 3, ‘[MASK]’: 4}
pairs: [[‘all’, ‘work’, ‘and’, ‘no’, ‘play’], [‘makes’, ‘jack’, ‘a’, ‘dull’, ‘boy’]]
pairs[0] + pairs[1]: [‘all’, ‘work’, ‘and’, ‘no’, ‘play’, ‘makes’, ‘jack’, ‘a’, ‘dull’, ‘boy’]
token: all
token: work
token: and
token: no
token: play
token: makes
token: jack
token: a
token: dull
token: boy
pairs: [[‘from’, ‘the’, ‘day’, ‘forth’], [‘my’, ‘arm’, ‘changed’]]
pairs[0] + pairs[1]: [‘from’, ‘the’, ‘day’, ‘forth’, ‘my’, ‘arm’, ‘changed’]
token: from
token: the
token: day
token: forth
token: my
token: arm
token: changed
pairs: [[‘and’, ‘a’, ‘voice’, ‘echoed’], [‘power’, ‘give’, ‘me’, ‘more’, ‘power’]]
pairs[0] + pairs[1]: [‘and’, ‘a’, ‘voice’, ‘echoed’, ‘power’, ‘give’, ‘me’, ‘more’, ‘power’]
token: and
token: a
token: voice
token: echoed
token: power
token: give
token: me
token: more
token: power
-----------------token_dict------------------
{‘’: 0, ‘[UNK]’: 1, ‘[CLS]’: 2, ‘[SEP]’: 3, ‘[MASK]’: 4, ‘all’: 5, ‘work’: 6, ‘and’: 7, ‘no’: 8, ‘play’: 9, ‘makes’: 10, ‘jack’: 11, ‘a’: 12, ‘dull’: 13, ‘boy’: 14, ‘from’: 15, ‘the’: 16, ‘day’: 17, ‘forth’: 18, ‘my’: 19, ‘arm’: 20, ‘changed’: 21, ‘voice’: 22, ‘echoed’: 23, ‘power’: 24, ‘give’: 25, ‘me’: 26, ‘more’: 27}
-----------------token_list------------------
token_list: [‘’, ‘[UNK]’, ‘[CLS]’, ‘[SEP]’, ‘[MASK]’, ‘all’, ‘work’, ‘and’, ‘no’, ‘play’, ‘makes’, ‘jack’, ‘a’, ‘dull’, ‘boy’, ‘from’, ‘the’, ‘day’, ‘forth’, ‘my’, ‘arm’, ‘changed’, ‘voice’, ‘echoed’, ‘power’, ‘give’, ‘me’, ‘more’]

构建模型

#构建训练模型
model = get_model(
    token_num=len(token_dict),
    head_num=5,
    transformer_num=12,
    embed_dim=25,
    feed_forward_dim=100,
    seq_len=20,
    pos_num=20,
    dropout_rate=0.05,
)
compile_model(model)        #模型编译
model.summary()             #模型简介
1
2
3
4
5
6
7
8
9
10
11
12
13

模型训练

def _generator():
    while True:
        yield gen_batch_inputs(
            sentence_pairs,
            token_dict,     #传入token_dict
            token_list,     #传入token_list
            seq_len=20,
            mask_rate=0.3,
            swap_sentence_rate=1.0,
        )

#模型训练        
model.fit_generator(
    generator=_generator(), #传入构建的函数
    steps_per_epoch=1000,
    epochs=100,
    validation_data=_generator(),
    validation_steps=100,
    callbacks=[
        keras.callbacks.EarlyStopping(monitor='val_loss', patience=5)
    ],
)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22

使用模型

#使用训练好的模型
inputs, output_layer = get_model(
    token_num=len(token_dict),
    head_num=5,
    transformer_num=12,
    embed_dim=25,
    feed_forward_dim=100,
    seq_len=20,
    pos_num=20,
    dropout_rate=0.05,
    training=False,      # The input layers and output layer will be returned if `training` is `False`
    trainable=False,     # Whether the model is trainable. The default value is the same with `training`
    output_layer_num=4,  # The number of layers whose outputs will be concatenated as a single output.
                         # Only available when `training` is `False`.
)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15

1.4 AdamWarmup

AdamWarmup优化器( optimizer )要用于预热( warmup )和衰减( decay)。学习率经过warmpup_ steps步后将达到Ir ,经过decay_ steps步后将衰减到min_ Ir。 中, calc_ train_ steps是 辅助函数,用于计算这两个步骤。

import numpy as np
from keras_bert import AdamWarmup, calc_train_steps

#生成随机数
train_x = np.random.standard_normal((1024, 100))
print(train_x)

#分批训练
total_steps, warmup_steps = calc_train_steps(
    num_example=train_x.shape[0],
    batch_size=32,
    epochs=10,
    warmup_proportion=0.1,
)

optimizer = AdamWarmup(total_steps, warmup_steps, lr=1e-3, min_lr=1e-5)
print(optimizer)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17

1.5 获取预先训练的检测点

用get_pretrained函数获取

from keras_bert import get_pretrained, PretrainedList, get_checkpoint_paths


#下载解压数据
model_path = get_pretrained(PretrainedList.multi_cased_base) #模型路径
paths = get_checkpoint_paths(model_path)                     #获取预训练地址

print('paths.config:',paths.config)

print('paths.checkpoint:',paths.checkpoint)

print('paths.vocab:', paths.vocab)
1
2
3
4
5
6
7
8
9
10
11
12

paths.config: C:\Users\N.keras\datasets\multi_cased_L-12_H-768_A-12\bert_config.json
paths.checkpoint: C:\Users\N.keras\datasets\multi_cased_L-12_H-768_A-12\bert_model.ckpt
paths.vocab: C:\Users\N.keras\datasets\multi_cased_L-12_H-768_A-12\vocab.txt
​

1.6抽取特征

如果需要tokens或句子的特征,则可以使用辅助函数extract_ embeddings。 下面的代码能提取所有的tokens特征。

需要传入模型的上文获取的模型地址，和待抽取的句子文本

from keras_bert import extract_embeddings

# model_path = get_pretrained(PretrainedList.multi_cased_base)
texts = ['all work and no play', 'makes jack a dull boy~']

embeddings = extract_embeddings(model_path, texts)
print('------------------句子特征------------------')
print(embeddings)
1
2
3
4
5
6
7
8

------------------句子特征------------------
[array([[ 0.07531555, -0.15103097, 0.16370851, …, 0.77671117,
0.02733378, -0.0297538 ],
[-0.12954001, -0.35776633, 0.02477884, …, 1.4055997 ,
0.16759634, -0.29797873],
[-0.26747486, -0.26116127, 0.11113451, …, 1.3588182 ,
0.10443275, -0.415785 ],
…,
[-0.34335068, -0.252737 , -0.68407285, …, 1.3278028 ,
-0.15623422, -0.4789365 ],
[-0.24592108, -0.12426493, -0.07056627, …, 1.3984789 ,
-0.03810839, -0.19883244],
[-0.02836535, -0.2510882 , 0.3347783 , …, 0.7883287 ,
0.05263783, -0.11937716]], dtype=float32), array([[ 0.30098903, -0.22291777, 0.6170633 , …, 0.5640485 ,
-0.03552696, 0.15263349],
[ 0.10523404, -0.2641968 , 0.6566459 , …, 1.0994014 ,
-0.13300316, 0.0962064 ],
[-0.2954503 , -0.35092717, 0.45172513, …, -0.24845225,
-0.4451233 , -0.0801053 ],
…,
[ 0.43035302, -0.31948596, 0.4378098 , …, 0.17429134,
-0.41476333, -0.09128644],
[ 0.47151148, -0.29935226, 0.6846881 , …, 0.34857082,
-0.11684854, -0.11591156],
[ 0.26193255, -0.16078277, 0.83464456, …, 0.53757495,
-0.04961903, 0.0441534 ]], dtype=float32)]

二、中文文本Bert预训练模型

2.1下载中文语料

首先,我们下载官方预训练模型chinese_L-12_H-768_A-12。

Google提供了多种预训练好的Bert模型,有针对不同语言的和不同模型大小。对于中文模型,我们使用Bert Base Chinese。

需要先下载好预训练的模型，解压到相应的位置，再读取相应文件夹的位置。

• Bert源代码: htps://github com/google-research/bert
• Bert预训练模型: https://storage.googleapis.com/bert_models/2018_11_03/chinese_L-12_H-768_A-12.zip

import os
from keras_bert import load_vocabulary
from keras_bert import Tokenizer
from keras_bert import load_trained_model_from_checkpoint
 
#设置预训练模型的路径
pretrained_path = r'G:\chinese_L-12_H-768_A-12'
config_path = os.path.join(pretrained_path, 'bert_config.json')
checkpoint_path = os.path.join(pretrained_path, 'bert_model.ckpt')
vocab_path = os.path.join(pretrained_path, 'vocab.txt')
 
#构建中文字典ID映射
token_dict = load_vocabulary(vocab_path)
print(token_dict)
print('中文ID字典长度:',len(token_dict))
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15

可以看到一共21128个词，同时也构建出来了相应的中文ID映射，就可以接着往下

#Tokenization
tokenizer = Tokenizer(token_dict)
print(tokenizer)

#加载预训练模型
model = load_trained_model_from_checkpoint(config_path, checkpoint_path)
print(model)
1
2
3
4
5
6
7

2.2 从预加载模型中提取中文字特征

对自定义语料进行tokenizer处理,并使用预训练模型提取输入文本的特征。

仅提取出前5的位的向量值来查看

#-------------------------------第二步 特征提取--------------------------------- 
text = '语言模型'

#先用分词器对象，将目标文本拆分，标准化
tokens = tokenizer.tokenize(text)
print('标准化：',tokens)


#获取对应字在字典中的ID映射
indices, segments = tokenizer.encode(first=text, max_len=512)
print('----------------ID映射----------------')
print(indices[:10])
print(segments[:10])
 
    
#根据ID提取模型中的字特征
print('---------------提取字特征-------------')
predicts = model.predict([np.array([indices]), np.array([segments])])[0] #获取
for i, token in enumerate(tokens):                                       #输出
    print(token, predicts[i].tolist()[:5])

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21

标准化： [‘[CLS]’, ‘语’, ‘言’, ‘模’, ‘型’, ‘[SEP]’]
----------------ID映射----------------
[101, 6427, 6241, 3563, 1798, 102, 0, 0, 0, 0]
[0, 0, 0, 0, 0, 0, 0, 0, 0, 0]
---------------提取字特征-------------
[CLS] [-0.6339437961578369, 0.2029203623533249, 0.08105022460222244, -0.0326896607875824, 0.5675354599952698]
语 [-0.7589297890663147, 0.09625156223773956, 1.072314739227295, 0.006224252283573151, 0.6886608004570007]
言 [0.5497941970825195, -0.7931232452392578, 0.44259175658226013, -0.7123060822486877, 1.2053987979888916]
模 [-0.29216861724853516, 0.606366753578186, 0.4984245002269745, -0.42493146657943726, 0.426719069480896]
型 [-0.7458041906356812, 0.4949134290218353, 0.7189165353775024, -0.8728538155555725, 0.8354955911636353]
[SEP] [-0.8752526640892029, -0.21610988676548004, 1.339908480644226, -0.10673223435878754, 0.3961635231971741]

2.3 多句子特征提取

和前面一样，可以实现多个句子的特征提取

#----------------------------第三步 多句子特征提取------------------------------
text1 = '语言模型'
text2 = "你好"
tokens1 = tokenizer.tokenize(text1)
print(tokens1)
tokens2 = tokenizer.tokenize(text2)
print(tokens2)
 
indices_new, segments_new = tokenizer.encode(first=text1, second=text2 ,max_len=512)
print(indices_new[:10])
#[101, 6427, 6241, 3563, 1798, 102, 0, 0, 0, 0]
print(segments_new[:10])
#[0, 0, 0, 0, 0, 0, 0, 0, 0, 0]
 
#提取特征
predicts_new = model.predict([np.array([indices_new]), np.array([segments_new])])[0]
for i, token in enumerate(tokens1):
    print(token, predicts_new[i].tolist()[:5])
for i, token in enumerate(tokens2):
    print(token, predicts_new[i].tolist()[:5])
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20

[‘[CLS]’, ‘语’, ‘言’, ‘模’, ‘型’, ‘[SEP]’]
[‘[CLS]’, ‘你’, ‘好’, ‘[SEP]’]
[101, 6427, 6241, 3563, 1798, 102, 872, 1962, 102, 0]
[0, 0, 0, 0, 0, 0, 1, 1, 1, 0]
[CLS] [-0.3404940962791443, 0.5169003009796143, 0.8958081603050232, -0.5850763916969299, 0.1620779037475586]
语 [-0.6919717788696289, 0.37331458926200867, 1.3196662664413452, -0.0865214616060257, 0.5522887110710144]
言 [0.6706017851829529, -0.5946153402328491, 0.4751562178134918, -0.7590199112892151, 0.9860224723815918]
模 [-0.4227488040924072, 0.7286509871482849, 0.5555989742279053, -0.43479853868484497, 0.39219915866851807]
型 [-0.5974094271659851, 0.5976635217666626, 0.7734537124633789, -1.0439568758010864, 0.8142789006233215]
[SEP] [-1.1663365364074707, 0.541653037071228, 1.396380066871643, 0.014762230217456818, -0.20481276512145996]
[CLS] [-0.3404940962791443, 0.5169003009796143, 0.8958081603050232, -0.5850763916969299, 0.1620779037475586]
你 [-0.6919717788696289, 0.37331458926200867, 1.3196662664413452, -0.0865214616060257, 0.5522887110710144]
好 [0.6706017851829529, -0.5946153402328491, 0.4751562178134918, -0.7590199112892151, 0.9860224723815918]
[SEP] [-0.4227488040924072, 0.7286509871482849, 0.5555989742279053, -0.43479853868484497, 0.39219915866851807]

2.4 根据预训练的模型向量实现字词预测

可以根据相应字向量，实现对于未知词的预测。

将一句话中的某个字mask掉，再预测。

#------------------------------第四步 字词预测填充------------------------------
#加载语言模型
model = load_trained_model_from_checkpoint(config_path, checkpoint_path, training=True)

token_dict_rev = {v: k for k, v in token_dict.items()}

token_ids, segment_ids = tokenizer.encode(
    u'数学是利用符号语言研究数量、结构、变化以及空间等概念的一门学科', 
    max_len=512)

#mask掉“数学”
print(token_ids[1],token_ids[2])
token_ids[1] = token_ids[2] = tokenizer._token_dict['[MASK]']
masks = np.array([[0, 1, 1] + [0] * (512 - 3)])
 
#模型预测被mask掉的部分
probas = model.predict([np.array([token_ids]), np.array([segment_ids]), masks])[0]
pred_indice = probas[0][1:3].argmax(axis=1).tolist()
print('Fill with: ', list(map(lambda x: token_dict_rev[x], pred_indice)))
#Fill with:  ['数', '学']

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21

3144 2110
Fill with: [‘数’, ‘学’]