---

前言

昨天按照该文章（自然语言处理（NLP）Bert与Lstm结合）跑bert+bilstm分类的时候，没成功跑起来，于是自己修改了一下，成功运行后，记录在这篇博客中。

一、运行环境

• python==3.7
• pandas==1.3.0
• numpy==1.20.3
• scikit-learn==0.24.2
• torch==1.9.0
• transformers==4.8.2

二、数据

1、Bert下载地址bert-base-chinese
注意：

• 只需要下载三个文件：
  pytorch_model.bin
  vocab.txt
  config.json
• 如果下载下来的pytorch_model.bin是一串乱七八糟的字母，将其文件名改为pytorch_model.bin即可

2、完整代码与数据集链接

• Pytorch Bert+BiLstm二分类

3、目录结构

三、模型结构

代码如下：

class bert_lstm(nn.Module):
    def __init__(self, bertpath, hidden_dim, output_size,n_layers,bidirectional=True, drop_prob=0.5):
        super(bert_lstm, self).__init__()
 
        self.output_size = output_size
        self.n_layers = n_layers
        self.hidden_dim = hidden_dim
        self.bidirectional = bidirectional
        
        #Bert ----------------重点，bert模型需要嵌入到自定义模型里面
        self.bert=BertModel.from_pretrained(bertpath)
        for param in self.bert.parameters():
            param.requires_grad = True
        
        # LSTM layers
        self.lstm = nn.LSTM(768, hidden_dim, n_layers, batch_first=True,bidirectional=bidirectional)
        
        # dropout layer
        self.dropout = nn.Dropout(drop_prob)
        
        # linear and sigmoid layers
        if bidirectional:
            self.fc = nn.Linear(hidden_dim*2, output_size)
        else:
            self.fc = nn.Linear(hidden_dim, output_size)
          
        #self.sig = nn.Sigmoid()
 
    def forward(self, x, hidden):
        batch_size = x.size(0)
        #生成bert字向量
        x=self.bert(x)[0]     #bert 字向量
        
        # lstm_out
        #x = x.float()
        lstm_out, (hidden_last,cn_last) = self.lstm(x, hidden)
        #print(lstm_out.shape)   #[32,100,768]
        #print(hidden_last.shape)   #[4, 32, 384]
        #print(cn_last.shape)    #[4, 32, 384]
        
        #修改 双向的需要单独处理
        if self.bidirectional:
            #正向最后一层，最后一个时刻
            hidden_last_L=hidden_last[-2]
            #print(hidden_last_L.shape)  #[32, 384]
            #反向最后一层，最后一个时刻
            hidden_last_R=hidden_last[-1]
            #print(hidden_last_R.shape)   #[32, 384]
            #进行拼接
            hidden_last_out=torch.cat([hidden_last_L,hidden_last_R],dim=-1)
            #print(hidden_last_out.shape,'hidden_last_out')   #[32, 768]
        else:
            hidden_last_out=hidden_last[-1]   #[32, 384]
            
            
        # dropout and fully-connected layer
        out = self.dropout(hidden_last_out)
        #print(out.shape)    #[32,768]
        out = self.fc(out)
        
        return out
    
    def init_hidden(self, batch_size):
        weight = next(self.parameters()).data
        
        number = 1
        if self.bidirectional:
            number = 2
        
        if (USE_CUDA):
            hidden = (weight.new(self.n_layers*number, batch_size, self.hidden_dim).zero_().float().cuda(),
                      weight.new(self.n_layers*number, batch_size, self.hidden_dim).zero_().float().cuda()
                     )
        else:
            hidden = (weight.new(self.n_layers*number, batch_size, self.hidden_dim).zero_().float(),
                      weight.new(self.n_layers*number, batch_size, self.hidden_dim).zero_().float()
                     )
        
        return hidden

1、其实就是将embedding层换成了bert，所以Lstm的input_size 为bert的输出size，所以Lstm的第一个参数是768
2、由于是使用BiLstm，所以需要将最后时刻的正向最后一层与反向最后一层拼接起来，相应的Linear层的输入维度应是拼接后的维度
3、由于是2分类，所以output_size为2

四、训练

代码如下：

def train_model(config, data_train):
    net = bert_lstm(config.bert_path, 
                    config.hidden_dim, 
                    config.output_size,
                    config.n_layers, 
                    config.bidirectional)
    criterion = nn.CrossEntropyLoss()
    optimizer = torch.optim.Adam(net.parameters(), lr=config.lr)
    if(config.use_cuda):
        net.cuda()
    net.train()
    for e in range(config.epochs):
        # initialize hidden state
        h = net.init_hidden(config.batch_size)
        counter = 0
        # batch loop
        for inputs, labels in data_train:
            counter += 1
            
            if(config.use_cuda):
                inputs, labels = inputs.cuda(), labels.cuda()
            h = tuple([each.data for each in h])
            net.zero_grad()
            output= net(inputs, h)
            loss = criterion(output.squeeze(), labels.long())
            loss.backward()
            optimizer.step()
    
            # loss stats
            if counter % config.print_every == 0:
                net.eval()
                with torch.no_grad():
                    val_h = net.init_hidden(config.batch_size)
                    val_losses = []
                    for inputs, labels in valid_loader:
                        val_h = tuple([each.data for each in val_h])

                        if(config.use_cuda):
                            inputs, labels = inputs.cuda(), labels.cuda()

                        output = net(inputs, val_h)
                        val_loss = criterion(output.squeeze(), labels.long())

                        val_losses.append(val_loss.item())
                net.train()
                print("Epoch: {}/{}, ".format(e+1, config.epochs),
                    "Step: {}, ".format(counter),
                    "Loss: {:.6f}, ".format(loss.item()),
                    "Val Loss: {:.6f}".format(np.mean(val_losses)))
    torch.save(net.state_dict(), config.save_path)                

Epoch: 1/10,  Step: 10,  Loss: 0.684742,  Val Loss: 0.688432
Epoch: 1/10,  Step: 20,  Loss: 0.664885,  Val Loss: 0.671069
Epoch: 1/10,  Step: 30,  Loss: 0.613591,  Val Loss: 0.622387
Epoch: 1/10,  Step: 40,  Loss: 0.571192,  Val Loss: 0.562263
Epoch: 2/10,  Step: 10,  Loss: 0.283182,  Val Loss: 0.421199
Epoch: 2/10,  Step: 20,  Loss: 0.385077,  Val Loss: 0.361812
Epoch: 2/10,  Step: 30,  Loss: 0.348373,  Val Loss: 0.318632
Epoch: 2/10,  Step: 40,  Loss: 0.597140,  Val Loss: 0.314847
Epoch: 3/10,  Step: 10,  Loss: 0.194882,  Val Loss: 0.273278
Epoch: 3/10,  Step: 20,  Loss: 0.123732,  Val Loss: 0.343172
Epoch: 3/10,  Step: 30,  Loss: 0.115506,  Val Loss: 0.313013
Epoch: 3/10,  Step: 40,  Loss: 0.170411,  Val Loss: 0.282829
Epoch: 4/10,  Step: 10,  Loss: 0.150081,  Val Loss: 0.263128
Epoch: 4/10,  Step: 20,  Loss: 0.353257,  Val Loss: 0.326907
Epoch: 4/10,  Step: 30,  Loss: 0.037445,  Val Loss: 0.342072
Epoch: 4/10,  Step: 40,  Loss: 0.096485,  Val Loss: 0.331090
Epoch: 5/10,  Step: 10,  Loss: 0.062844,  Val Loss: 0.321690
Epoch: 5/10,  Step: 20,  Loss: 0.031070,  Val Loss: 0.316013
Epoch: 5/10,  Step: 30,  Loss: 0.058623,  Val Loss: 0.318129
Epoch: 5/10,  Step: 40,  Loss: 0.046849,  Val Loss: 0.335890
Epoch: 6/10,  Step: 10,  Loss: 0.147990,  Val Loss: 0.446133
Epoch: 6/10,  Step: 20,  Loss: 0.324785,  Val Loss: 0.448294
Epoch: 6/10,  Step: 30,  Loss: 0.034576,  Val Loss: 0.390784
Epoch: 6/10,  Step: 40,  Loss: 0.018516,  Val Loss: 0.446345
Epoch: 7/10,  Step: 10,  Loss: 0.041855,  Val Loss: 0.340792
Epoch: 7/10,  Step: 20,  Loss: 0.140505,  Val Loss: 0.433969
Epoch: 7/10,  Step: 30,  Loss: 0.015936,  Val Loss: 0.537444
Epoch: 7/10,  Step: 40,  Loss: 0.061285,  Val Loss: 0.688956
Epoch: 8/10,  Step: 10,  Loss: 0.027597,  Val Loss: 0.383043
Epoch: 8/10,  Step: 20,  Loss: 0.020746,  Val Loss: 0.344770
Epoch: 8/10,  Step: 30,  Loss: 0.066156,  Val Loss: 0.418557
Epoch: 8/10,  Step: 40,  Loss: 0.011359,  Val Loss: 0.434401
Epoch: 9/10,  Step: 10,  Loss: 0.014871,  Val Loss: 0.450156
Epoch: 9/10,  Step: 20,  Loss: 0.011917,  Val Loss: 0.455503
Epoch: 9/10,  Step: 30,  Loss: 0.139435,  Val Loss: 0.496916
Epoch: 9/10,  Step: 40,  Loss: 0.015576,  Val Loss: 0.499172
Epoch: 10/10,  Step: 10,  Loss: 0.006573,  Val Loss: 0.585796
Epoch: 10/10,  Step: 20,  Loss: 0.144999,  Val Loss: 0.514712
Epoch: 10/10,  Step: 30,  Loss: 0.037525,  Val Loss: 0.445692
Epoch: 10/10,  Step: 40,  Loss: 0.010959,  Val Loss: 0.382745

我的GPU为Quadro RTX 6000，没几分钟就训练完成了，还是比较快的。

五、测试及预测

def test_model(config, data_test):
    net = bert_lstm(config.bert_path, 
                config.hidden_dim, 
                config.output_size,
                config.n_layers, 
                config.bidirectional)
    net.load_state_dict(torch.load(config.save_path))
    net.cuda()
    criterion = nn.CrossEntropyLoss()
    test_losses = [] # track loss
    num_correct = 0
    
    # init hidden state
    h = net.init_hidden(config.batch_size)
    
    net.eval()
    # iterate over test data
    for inputs, labels in data_test:
        h = tuple([each.data for each in h])
        if(USE_CUDA):
            inputs, labels = inputs.cuda(), labels.cuda()
        output = net(inputs, h)
        test_loss = criterion(output.squeeze(), labels.long())
        test_losses.append(test_loss.item())
        
        output=torch.nn.Softmax(dim=1)(output)
        pred=torch.max(output, 1)[1]
        
        # compare predictions to true label
        correct_tensor = pred.eq(labels.long().view_as(pred))
        correct = np.squeeze(correct_tensor.numpy()) if not USE_CUDA else np.squeeze(correct_tensor.cpu().numpy())
        num_correct += np.sum(correct)
    print("Test loss: {:.3f}".format(np.mean(test_losses)))
    # accuracy over all test data
    test_acc = num_correct/len(data_test.dataset)
    print("Test accuracy: {:.3f}".format(test_acc))

def predict(test_comment_list, config):
    net = bert_lstm(config.bert_path, 
                config.hidden_dim, 
                config.output_size,
                config.n_layers, 
                config.bidirectional)
    net.load_state_dict(torch.load(config.save_path))
    net.cuda()
    result_comments=pretreatment(test_comment_list)   #预处理去掉标点符号
    #转换为字id
    tokenizer = BertTokenizer.from_pretrained(config.bert_path)
    result_comments_id = tokenizer(result_comments,
                                    padding=True,
                                    truncation=True,
                                    max_length=120,
                                    return_tensors='pt')
    tokenizer_id = result_comments_id['input_ids']
    # print(tokenizer_id.shape)
    inputs = tokenizer_id
    batch_size = inputs.size(0)
    # batch_size = 32
    # initialize hidden state
    h = net.init_hidden(batch_size)

    if(USE_CUDA):
        inputs = inputs.cuda()

    net.eval()
    with torch.no_grad():
        # get the output from the model
        output= net(inputs, h)
        output=torch.nn.Softmax(dim=1)(output)
        pred=torch.max(output, 1)[1]
        # printing output value, before rounding
        print('预测概率为: {:.6f}'.format(torch.max(output, 1)[0].item()))
        if(pred.item()==1):
            print("预测结果为:正向")
        else:
            print("预测结果为:负向")

Test loss: 0.390
Test accuracy: 0.837

predict :

test_comments =  ['这个菜真不错']
预测概率为: 0.981275
预测结果为:正向

完整代码与数据集链接：Pytorch Bert+BiLstm二分类