基于pytorch使用BI-LSTM模型做中文文本分类

基于pytorch使用BI-LSTM模型做中文文本分类

主要涉及使用pandas从csv文件读取文本数据，将dataframe格式数据转为torch iteration 输出模型训练和预测。

import pandas as pd 
import numpy as np
import re
import torch.utils.data as data_utils
import torch
from torchtext.legacy import data 
from torchtext.legacy import datasets,vocab
from tqdm import tqdm
from torch import nn

1.读取数据与预处理

数据格式为csv，按逗号分隔，主要分为complain_content、label两列，其中complain_content为分词后的中文文本。

df = pd.read_csv("/home/jovyan/work/NLP/kesu_cate/data/kesu_clean_data.csv",encoding='utf-8')
df["complain_content"] = df["complain_content"].apply(lambda x : ' '.join(x.replace('[','').replace(']','').replace("'",'').replace(' ','').split(',')))  #
df["label"] = df["label"].astype(float)

#划分训练集、测试集、验证集
from sklearn.model_selection import train_test_split
train_data,test_data= train_test_split(df,test_size = 0.2,random_state = 1024)
train_data,valid_data= train_test_split(train_data,test_size = 0.2,random_state = 1024)

2.使用torchtext处理dataframe数据，并生成词典，并创建iteration

tokenize=lambda x:x.split(' ')
TEXT=data.Field(tokenize=tokenize,sequential=True)
LABEL=data.Field(sequential=False, use_vocab=False)

#数据处理类
class DatasetProcess(data.Dataset):
    def __init__(self,df,text_tield,label_field):
        fields=[('complain_content',text_tield),("label",label_field)]
        examples=[]
        for text , label in tqdm(zip(df['complain_content'], df['label'])):
            examples.append(data.Example.fromlist([text, label], fields)) 
        super(DatasetProcess,self).__init__(examples,fields)

train_data = DatasetProcess(train_data,TEXT,LABEL)
valid_data = DatasetProcess(valid_data,TEXT,LABEL)
test_data = DatasetProcess(test_data,TEXT,LABEL)

#加载腾讯NLP预训练结果 并建立词典
vectors = vocab.Vectors(name='/home/jovyan/work/NLP/pytoch/.vector_cache/tencent-ailab-embedding-zh-d100-v0.2.0-s/tencent-ailab-embedding-zh-d100-v0.2.0-s.txt')
TEXT.build_vocab(train_data,max_size=25000,vectors=vectors,unk_init=torch.Tensor.normal_)
LABEL.build_vocab(train_data)


train_iter,valid_iter,test_iter=data.BucketIterator.splits(
    (train_data,valid_data,test_data),
    batch_size=256,
    device='cpu',
    sort_key=lambda x: len(x.complain_content),
    sort_within_batch=False,
    repeat=False
)

3.构建BI-LSTM模型

#初始化参数
vocab_size = len(TEXT.vocab)
embedding_size = 100
output_size = 1
num_layers = 2 
dropout = 0.3 
pad_idx = TEXT.vocab.stoi[TEXT.pad_token]
hidden_size = 100
device = 'cpu'

#定义模型
class RNNModel(nn.Module):
    def __init__(self,vocab_size,embedding_size,output_size,hidden_size,num_layers,dropout,pad_idx):
        super(RNNModel,self).__init__()
        self.embed = nn.Embedding(vocab_size,embedding_size,padding_idx=pad_idx)
        self.lstm = nn.LSTM(embedding_size,hidden_size,bidirectional=True,num_layers=num_layers)
        self.linear = nn.Linear(hidden_size * num_layers,output_size)
        self.dropout = nn.Dropout(dropout)
        
    def forward(self,text):
        embeded = self.embed(text)
        embeded = self.dropout(embeded)
        output,(hidden,cell) = self.lstm(embeded)
        hidden = self.dropout(hidden.squeeze())
        hidden = torch.cat([hidden[-1],hidden[-2]],dim=1)
        return self.linear(hidden)

#定义评价函数和训练验证函数
def binary_accuracy(preds,y):
    rounded_preds = torch.round(torch.sigmoid(preds))
    correct = (rounded_preds == y).float()
    acc = correct.sum()/len(correct)
    return acc

def train(model,iterator,optimizer,crit):
    epoch_loss,epoch_acc = 0.,0.
    model.train()
    total_len = 0. 
    for batch in iterator:
        preds = model(batch.complain_content).squeeze()
        loss = crit(preds,batch.label.float())
        acc = binary_accuracy(preds,batch.label)
        
        optimizer.zero_grad()
        loss.backward()
        optimizer.step()
        
        epoch_loss += loss.item() * len(batch.label)
        epoch_acc += acc.item() * len(batch.label)
        total_len += len(batch.label)
        print(epoch_loss,epoch_acc,len(batch.label))
    return epoch_loss / total_len,epoch_acc/total_len

def evaluate(model,iterator,crit):
    epoch_loss,epoch_acc = 0.,0.
    model.eval()
    total_len = 0. 
    for batch in iterator:
        preds = model(batch.complain_content).squeeze()
        loss = crit(preds,batch.label.float())
        acc = binary_accuracy(preds,batch.label)
        
        epoch_loss += loss.item() * len(batch.label)
        epoch_acc += acc.item() * len(batch.label)
        total_len += len(batch.label) 
    model.train()
    return epoch_loss / total_len,epoch_acc/total_len


model = RNNModel(vocab_size,embedding_size,output_size,hidden_size,num_layers,dropout,pad_idx)

#初始化模型参数
pretrained_embedding = TEXT.vocab.vectors
model.embed.weight.data.copy_(pretrained_embedding)
unk_idx = TEXT.vocab.stoi[TEXT.unk_token]
model.embed.weight.data[pad_idx] = torch.zeros(embedding_size)
model.embed.weight.data[unk_idx] = torch.zeros(embedding_size)


#训练模型
n_epoch = 10 
best_valid_acc = 0. 
for epoch in range(n_epoch):
    train_loss,train_acc = train(model,train_iter,optimizer,crit)
    valid_loss,valid_acc = evaluate(model,valid_iter,crit)
    
    if valid_acc > best_valid_acc:
        best_valid_acc = valid_acc
        torch.save(model.state_dict(),'lstm_model_cn.pth')
    print("Epoch",epoch,"Train loss",train_loss,"Train Acc",train_acc)
    print("Epoch",epoch,"valid_loss",valid_loss,"valid Acc",valid_acc)



#测试模型
def removepad(arr):
    res = []
    for x in arr:
        if x != '<pad>':
            res.append(x)
    return res

def predict(test_iter):
    with torch.no_grad():
        for batch in test_iter:
            pred = torch.sigmoid(model(batch.complain_content))
            for i in range(len(batch.label)):
                index = batch.complain_content.t()[i]
                tmp = [TEXT.vocab.itos[x] for x in index]
                seq = ' '.join(removepad(tmp))
                
                print(pred[i].item(),seq)

参考链接

1.torchtext数据预处理：https://www.jianshu.com/p/8e891bdca78a
2.腾讯预训练词向量链接：https://ai.tencent.com/ailab/nlp/zh/download.html
3.加载教程:https://blog.csdn.net/nlpuser/article/details/83627709