使用bert进行中文文本分类
文章目录
- 一、Bert预训练模型的使用
-
- 安装transformers
- BertTokenizer
- BertModel
- 二、使用Bert进行中文文本分类
- 参考
一、Bert预训练模型的使用
安装transformers
pip install transformers
BertTokenizer
Bert的分词器，附带Bert的字典，因为词向量化需要找到字典中对应的序号，才能找到它对应的词向量
from transformers import BertTokenizer
import torch
tokenizers = BertTokenizer.from_pretrained('bert-base-uncased') # 加载base模型的对应的切词器
print(tokenizers)
token = tokenizers.tokenize("I love music")
print(token)
indexes = tokenizers.convert_tokens_to_ids(token)
print(indexes)
id2token = tokenizers.convert_ids_to_tokens(indexes)
print(id2token)
encoder = tokenizers.encode('I love music')
print(encoder)
encoder_tensor = torch.tensor(encoder)
print(f"encoder_tensor: {encoder_tensor}, the size of encoder_tensor is: {encoder_tensor.size()}")
cls = tokenizers._convert_token_to_id('[CLS]')
sep = tokenizers._convert_token_to_id("[SEP]")
print(cls, sep)
'''
PreTrainedTokenizer(name_or_path='bert-base-uncased', vocab_size=30522, model_max_len=512, is_fast=False, padding_side='right', truncation_side='right', special_tokens={'unk_token': '[UNK]', 'sep_token': '[SEP]', 'pad_token': '[PAD]', 'cls_token': '[CLS]', 'mask_token': '[MASK]'})
['i', 'love', 'music']
[1045, 2293, 2189]
['i', 'love', 'music']
[101, 1045, 2293, 2189, 102]
encoder_tensor: tensor([ 101, 1045, 2293, 2189, 102]), the size of encoder_tensor is: torch.Size([5])
101 102
'''
BertModel
输入BertModel中的句子的开头和结尾需要自己添加上[CLS]和[SEP]，才能够正确使用Bert，或者直接调用BertTokenizer.encode则会自动添加
转换为id即可输入Bert获得对应的句子编码和词向量
import torch
from transformers import BertTokenizer, BertModel
tokenizer = BertTokenizer.from_pretrained("bert-base-uncased")
model = BertModel.from_pretrained('bert-base-uncased')
music = torch.tensor(tokenizer.encode('I love music')).unsqueeze(0) # batch_size=1
bert = torch.tensor(tokenizer.encode("I am using the bert model")).unsqueeze(0) # batch_size=1
output_music = model(music)
music_word_embedding = output_music[0] # 词的向量表表示
music_sentence_embedding = output_music[1] # 句子的向量表示
print(f"word embedding: {music_word_embedding.shape}")
print(f"sentence embedding: {music_sentence_embedding.shape}")
'''
word embedding: torch.Size([1, 5, 768])
sentence embedding: torch.Size([1, 768])
'''
[CLS]和[SEP]也会被转换为向量输出
import torch
from transformers import BertTokenizer, BertModel
tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
model = BertModel.from_pretrained('bert-base-uncased')
morning = torch.tensor(tokenizer.encode('good morning')).unsqueeze(0)
night = torch.tensor(tokenizer.encode('good night')).unsqueeze(0)
morning_embedding = model(morning)
night_embedding = model(night)
print(f"morning good: {morning_embedding[0][0][1]}")
print(f"night good: {night_embedding[0][0][1]}")
'''
morning good: tensor([-5.8899e-02, 4.4347e-02, 8.4845e-01, -1.1691e+00, -4.5506e-01,
6.7695e-02, -6.9360e-02, 1.6030e+00, -1.2251e-01, -1.6857e+00,
1.2733e-01, -4.0276e-01, 1.2708e-01, 1.2586e-01, -1.8597e-01,
night good: tensor([ 1.0142e+00, 1.2357e-01, 1.2800e+00, -1.2907e+00, -4.6667e-01,
-1.9426e-01, 5.0342e-01, 1.7515e+00, 6.8127e-02, -1.3032e+00,
-9.8331e-02, 2.1038e-01, 6.1536e-02, 1.5393e-02, -3.9809e-01,
'''
看的出来，同一个词，在不同的句子中词向量是不同的，说明Bert能够很好解决一词多义的问题
二、使用Bert进行中文文本分类
BertForSequenceClassification 是在 BertModel 的基础上，添加了一个线性层 + 激活函数，用于分类。我们会使用model = BertForSequenceClassification.from_pretrained("bert-base-uncased", config=config)来加载模型，那么线性层 + 激活函数的权重就会随机初始化。我们的目的，就是通过微调，学习到线性层 + 激活函数的权重
# -*- coding: utf-8 -*-
# @Time : 2022/12/4 10:53
# @Author : 楚楚
# @File : 04文本分类.py
# @Software: PyCharm
import pandas as pd
import torch.nn as nn
from torch.utils.data import Dataset, DataLoader
import torch
import pandas as od
from torch.utils.data import TensorDataset, DataLoader, RandomSampler, SequentialSampler
from transformers import BertTokenizer, BertForSequenceClassification, AdamW, BertConfig
from datetime import datetime
# hyper parameters
HIDDEN_DROPOUT_PROB = 0.3
NUM_LABELS = 6
LR = 1e-5
WEIGHT_DECAY = 1e-2
EPOCH = 100
BATCH_SIZE = 16
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
# file path
data_path = "./dataset/"
vocab_file = data_path + "vocab.txt"
train = data_path + "train.xlsx"
test = data_path + "test.xlsx"
print("loading BERT tokenizer...")
tokenizer = BertTokenizer(vocab_file=vocab_file)
'''
BertTokenizer(vocab_file=vocab_file) 等价于 BertTokenizer.from_pretrained("hfl/chinese-bert-wwm" )
'''
config = BertConfig.from_pretrained("hfl/chinese-bert-wwm", num_labels=NUM_LABELS,
hidden_dropout_prob=HIDDEN_DROPOUT_PROB)
model = BertForSequenceClassification.from_pretrained("hfl/chinese-bert-wwm", config=config)
model.to(device)
'''
颜色对应的序号
0：蓝帽、1：白帽、2：红帽、3：黄帽、4：黑帽、5：绿帽
'''
class SixHatDataset(Dataset):
def __init__(self, path_to_file):
super(SixHatDataset, self).__init__()
self.label = ['蓝帽', '白帽', '红帽', '黄帽', '黑帽', '绿帽']
self.label2id = {'蓝帽': 0, '白帽': 1, '红帽': 2, '黄帽': 3, '黑帽': 4, '绿帽': 5}
self.dataset = pd.read_excel(path_to_file, keep_default_na=False) # 读取train.xlsx/test.xlsx中的数据
self.label_text = [] # 存放标签以及标签对应的文本（列表中存放的数据类型为字典类型 label: text)
for label in self.label:
self.hat_dataset = self.dataset.loc[:, label]
self.read_each_hat()
# 读取excel表中每一个帽子对应的文本数据
def read_each_hat(self):
label = self.hat_dataset.name # 获取对应的标签
label = self.label2id.get(label)
for text in self.hat_dataset:
if text != '':
self.label_text.append({label: text})
def __len__(self):
return len(self.label_text)
def __getitem__(self, idx):
label_text = self.label_text[idx]
label = list(label_text.keys())[0]
text = label_text.get(label)
label = torch.tensor(label, dtype=torch.long)
return text, label
train_dataset = SixHatDataset(train)
train_dataloader = DataLoader(dataset=train_dataset, batch_size=BATCH_SIZE, shuffle=True, num_workers=0)
test_dataset = SixHatDataset(test)
test_dataloader = DataLoader(dataset=test_dataset, batch_size=BATCH_SIZE, shuffle=True, num_workers=0)
# 定义优化器和损失函数
optimizer = AdamW(model.parameters(), lr=LR)
criterion = nn.CrossEntropyLoss().to(device)
# 定义训练的函数
def train(epoch, dataloader):
model.train()
epoch_acc = 0
for idx, data in enumerate(dataloader):
text, label = data
label = label.to(device)
tokenize_text = tokenizer(text, max_length=100, add_special_tokens=True, truncation=True, padding=True,
return_tensors='pt')
tokenize_text = tokenize_text.to(device)
optimizer.zero_grad()
'''
SequenceClassifierOutput(loss=tensor(2.0327, grad_fn=), logits=tensor([[-0.1659, 0.7432, 0.9424, -0.3815, 0.1794, 0.1559],
[-1.1916, 0.2135, 1.0156, -0.5150, 0.7795, -0.0261],
[-0.3528, 0.1796, 1.1230, -0.8721, 0.4448, 0.8882],
[-0.9661, 0.0696, 1.0002, -0.3308, 0.8832, 0.1922],
[-0.4554, 0.4486, 0.9846, -0.3371, 0.8539, -0.3214],
[-1.3695, -0.2882, 0.5169, 0.4508, 1.1330, 0.2997],
[-0.5405, 0.0763, 1.2337, -0.2260, 0.6922, -0.2044],
[-0.9427, -0.0595, 1.7682, -1.0026, 0.4901, -0.1369],
[-1.1734, 0.1412, 2.0086, -0.5898, 0.8525, 0.0528],
[-0.7478, -0.3635, 1.2168, -0.5125, 1.2169, 0.3979],
[-1.2102, 0.2823, 0.9883, -0.5061, 0.5131, -0.0209],
[-1.0257, -0.0059, 1.0093, -0.8454, 1.1518, 0.1737],
[-0.6118, 0.2500, 1.3389, -0.7910, 0.0835, 0.4923],
[-0.5885, -0.0195, 1.0697, 0.0891, 0.9630, -0.0917],
[-0.7473, 0.1327, 1.0242, -0.4896, 0.2457, 0.2772],
[-0.6821, -0.1901, 0.6271, -0.2386, 0.1395, -0.0949]],
grad_fn=), hidden_states=None, attentions=None)
'''
output = model(**tokenize_text, labels=label)
# y_pred_prob = logits : [batch_size, num_labels]
y_pred_prob = output[1]
y_pred_label = y_pred_prob.argmax(dim=1)
# loss = loss
loss = output[0]
acc = ((y_pred_label == label.view(-1)).sum()).item()
loss.backward()
optimizer.step()
if idx % 10 == 0:
print(f'train epoch:{epoch}, loss: {loss}')
now = datetime.now()
now = now.strftime("%Y-%m-%d %H:%M:%S")
content = f"{now}\tloss: {loss}\n"
with open('information.txt', 'a+', encoding='utf-8') as file:
file.write(content)
epoch_acc += acc
accuracy = epoch_acc / len(train_dataset)
print(f"训练集上的准确率：{accuracy:.4f}%")
now = datetime.now()
now = now.strftime("%Y-%m-%d %H:%M:%S")
content = f"{now}\t训练集上的准确率：{accuracy:.4f}%\n"
with open('information.txt', 'a+', encoding='utf-8') as file:
file.write(content)
def validate(epoch, dataloader):
model.eval()
total_loss = 0
epoch_acc = 0
with torch.no_grad():
for _, data in enumerate(dataloader):
text, label = data
'''
result:
{'input_ids': tensor([[ 101, 1266, 776, 2356, 3308, 7345, 1277, 1266, 1724, 4384, 704, 6662,
102, 0, 0]]), 'token_type_ids': tensor([[0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]]), 'attention_mask': tensor([[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0]])}
token_type_ids: segment encoding
attention_mask: 避免填充的数据参与Self-Attention的计算
tokenizer():
return_tensors='pt': 表示返回的是pytorch的tensor
'''
tokenize_text = tokenizer(text, max_length=100, padding=True, truncation=True, add_special_tokens=True,
return_tensors='pt')
tokenize_text = tokenize_text.to(device)
label = label.to(device)
output = model(**tokenize_text, labels=label)
y_pred_label = output[1].argmax(dim=1)
loss = output[0]
total_loss += loss
acc = (y_pred_label == label.view(-1)).sum().item()
epoch_acc += acc
print(f"测试集上的loss：{total_loss}")
accuracy = epoch_acc / len(test_dataset)
print(f"测试集准确率：{accuracy}%")
print("模型保存成功")
torch.save(model.state_dict(), f'./model/classification_{epoch}.pth')
now = datetime.now()
now = now.strftime("%Y-%m-%d %H:%M:%S")
content = f"{now}\t测试集上的loss：{total_loss}，测试集上的准确率：{accuracy}\n"
with open('information.txt', 'a+', encoding='utf-8') as file:
file.write(content)
for i in range(EPOCH):
content = f"{'-' * 20}epoch{i + 1}{'-' * 20}\n"
content = content + f"{'*' * 10}训练开始{'*' * 10}\n"
with open('information.txt', 'a+', encoding='utf-8') as file:
file.write(content)
train(i, train_dataloader)
content = f"{'*' * 10}测试开始{'*' * 10}\n"
with open('information.txt', 'a+', encoding='utf-8') as file:
file.write(content)
validate(i, test_dataloader)
词汇表 vocab.txt 来自于哈工大的中文预训练语言模型 BERT-wwm, Chinese，地址： 中文BERT-wwm
参考
1、transformer包中的bert预训练模型的调用详解
2、How does max_length, padding and truncation arguments work in HuggingFace’ BertTokenizerFast.from_pretrained(‘bert-base-uncased’)?
3、明明pip install transformers了，但调用模型的时候还会报错
4、中文BERT-wwm
5、NLP（二十八）：BertForSequenceClassification进行文本分类，基于transformers