bert微调做二分类任务代码

下面是使用PyTorch实现BERT微调进行二分类任务的代码示例：

import torch
from torch.utils.data import TensorDataset, DataLoader, RandomSampler, SequentialSampler
from transformers import BertTokenizer, BertForSequenceClassification, AdamW, get_linear_schedule_with_warmup

# 定义一些基本参数
MAX_LEN = 128
BATCH_SIZE = 32
EPOCHS = 4
LEARNING_RATE = 2e-5

# 加载数据
train_texts = [...] # 训练集文本
train_labels = [...] # 训练集标签
test_texts = [...] # 测试集文本
test_labels = [...] # 测试集标签

# 加载BERT tokenizer
tokenizer = BertTokenizer.from_pretrained('bert-base-uncased', do_lower_case=True)

# 对训练集和测试集进行编码
train_encodings = tokenizer(train_texts, truncation=True, padding=True, max_length=MAX_LEN)
test_encodings = tokenizer(test_texts, truncation=True, padding=True, max_length=MAX_LEN)

# 将编码转换为PyTorch tensors
train_labels = torch.tensor(train_labels)
test_labels = torch.tensor(test_labels)
train_dataset = TensorDataset(torch.tensor(train_encodings['input_ids']), torch.tensor(train_encodings['attention_mask']), train_labels)
test_dataset = TensorDataset(torch.tensor(test_encodings['input_ids']), torch.tensor(test_encodings['attention_mask']), test_labels)

# 创建数据加载器
train_sampler = RandomSampler(train_dataset)
train_dataloader = DataLoader(train_dataset, sampler=train_sampler, batch_size=BATCH_SIZE)
test_sampler = SequentialSampler(test_dataset)
test_dataloader = DataLoader(test_dataset, sampler=test_sampler, batch_size=BATCH_SIZE)

# 加载BERT模型
model = BertForSequenceClassification.from_pretrained('bert-base-uncased', num_labels=2)

# 定义优化器和学习率调度器
optimizer = AdamW(model.parameters(), lr=LEARNING_RATE, eps=1e-8)
total_steps = len(train_dataloader) * EPOCHS
scheduler = get_linear_schedule_with_warmup(optimizer, num_warmup_steps=0, num_training_steps=total_steps)

# 训练BERT模型
for epoch in range(EPOCHS):
    for step, batch in enumerate(train_dataloader):
        # 将batch转换为GPU tensor
        batch = tuple(t.to('cuda') for t in batch)
        inputs = {'input_ids': batch[0], 'attention_mask': batch[1], 'labels': batch[2]}
        outputs = model(**inputs)
        loss = outputs[0]
        loss.backward()
        torch.nn.utils.clip_grad_norm_(model.parameters(), 1.0)
        optimizer.step()
        scheduler.step()
        optimizer.zero_grad()
    
    # 在测试集上评估模型
    model.eval()
    test_loss, test_accuracy = 0, 0
    nb_test_steps, nb_test_examples = 0, 0
    for batch in test_dataloader:
        batch = tuple(t.to('cuda') for t in batch)
        inputs = {'input_ids': batch[0], 'attention_mask': batch[1], 'labels': batch[2]}
        with torch.no_grad():
            outputs = model(**inputs)
        logits = outputs[1]
        logits = logits.detach().cpu().numpy()
        label_ids = inputs['labels'].cpu().numpy()
        tmp_test_accuracy = accuracy(logits, label_ids)
        test_accuracy += tmp_test_accuracy
        nb_test_examples += inputs['input_ids'].size(0)
        nb_test_steps += 1
    test_accuracy = test_accuracy / nb_test_examples
    print('Epoch: {}, Test Accuracy: {}'.format(epoch, test_accuracy))

其中，`train_texts`和`train_labels`是训练集的文本和标签，`test_texts`和`test_labels`是测试集的文本和标签。`tokenizer`用来将文本转换为BERT的输入格式，`BertForSequenceClassification`是带有分类层的BERT模型，`AdamW`是用于优化模型的优化器，`get_linear_schedule_with_warmup`是用于调整学习率的调度器。在训练过程中，我们使用了梯度裁剪来避免梯度爆炸，同时在每个epoch结束后在测试集上评估模型的性能。