python打卡day37 - 技术栈

@疏锦行

知识点回顾：

过拟合的判断：测试集和训练集同步打印指标
模型的保存和加载

a. 仅保存权重

b. 保存权重和模型

c. 保存全部信息checkpoint，还包含训练状态

早停策略

**作业：**对信贷数据集训练后保存权重，加载权重后继续训练50轮，并采取早停策略

复制代码

# 保存模型权重
torch.save(model.state_dict(), 'credit_model_weights.pth')

# 加载模型权重
model.load_state_dict(torch.load('credit_model_weights.pth'))

# 设置继续训练的轮数
additional_epochs = 50

for epoch in range(additional_epochs):
    # 前向传播
    outputs = model(X_train_tensor)
    loss = criterion(outputs, y_train_tensor)

    # 反向传播和优化
    optimizer.zero_grad()
    loss.backward()
    optimizer.step()

    if (epoch + 1) % 10 == 0:
        print(f'Epoch [{epoch+1}/{additional_epochs}], Loss: {loss.item():.4f}')

# 保存继续训练后的模型权重
torch.save(model.state_dict(), 'credit_model_weights_continued.pth')
# 早停策略参数
patience = 10  # 容忍验证集损失不下降的最大轮数
best_val_loss = float('inf')
counter = 0

for epoch in range(num_epochs):
    # 训练代码
    model.train()
    outputs = model(X_train_tensor)
    train_loss = criterion(outputs, y_train_tensor)
    optimizer.zero_grad()
    train_loss.backward()
    optimizer.step()

    # 验证代码
    model.eval()
    with torch.no_grad():
        val_outputs = model(X_val_tensor)
        val_loss = criterion(val_outputs, y_val_tensor)

    print(f'Epoch [{epoch+1}/{num_epochs}], Train Loss: {train_loss.item():.4f}, Val Loss: {val_loss.item():.4f}')

    # 早停策略逻辑
    if val_loss < best_val_loss:
        best_val_loss = val_loss
        counter = 0
        # 保存最佳模型权重
        torch.save(model.state_dict(), 'best_credit_model_weights.pth')
    else:
        counter += 1
        if counter >= patience:
            print('Early stopping!')
            break