python训练营day50
知识点回顾：
- resnet结构解析
- CBAM放置位置的思考
- 针对预训练模型的训练策略
- 差异化学习率
- 三阶段微调
ps：今日的代码训练时长较长，3080ti大概需要40min的训练时长
作业：
- 好好理解下resnet18的模型结构
- 尝试对vgg16+cbam进行微调策略
import torch
import torch.nn as nn
from torchvision import models# 自定义ResNet18模型，插入CBAM模块
class ResNet18_CBAM(nn.Module):def __init__(self, num_classes=10, pretrained=True, cbam_ratio=16, cbam_kernel=7):super().__init__()# 加载预训练ResNet18self.backbone = models.resnet18(pretrained=pretrained) # 修改首层卷积以适应32x32输入（CIFAR10）self.backbone.conv1 = nn.Conv2d(in_channels=3, out_channels=64, kernel_size=3, stride=1, padding=1, bias=False)self.backbone.maxpool = nn.Identity() # 移除原始MaxPool层（因输入尺寸小）# 在每个残差块组后添加CBAM模块self.cbam_layer1 = CBAM(in_channels=64, ratio=cbam_ratio, kernel_size=cbam_kernel)self.cbam_layer2 = CBAM(in_channels=128, ratio=cbam_ratio, kernel_size=cbam_kernel)self.cbam_layer3 = CBAM(in_channels=256, ratio=cbam_ratio, kernel_size=cbam_kernel)self.cbam_layer4 = CBAM(in_channels=512, ratio=cbam_ratio, kernel_size=cbam_kernel)# 修改分类头self.backbone.fc = nn.Linear(in_features=512, out_features=num_classes)def forward(self, x):# 主干特征提取x = self.backbone.conv1(x)x = self.backbone.bn1(x)x = self.backbone.relu(x) # [B, 64, 32, 32]# 第一层残差块 + CBAMx = self.backbone.layer1(x) # [B, 64, 32, 32]x = self.cbam_layer1(x)# 第二层残差块 + CBAMx = self.backbone.layer2(x) # [B, 128, 16, 16]x = self.cbam_layer2(x)# 第三层残差块 + CBAMx = self.backbone.layer3(x) # [B, 256, 8, 8]x = self.cbam_layer3(x)# 第四层残差块 + CBAMx = self.backbone.layer4(x) # [B, 512, 4, 4]x = self.cbam_layer4(x)# 全局平均池化 + 分类x = self.backbone.avgpool(x) # [B, 512, 1, 1]x = torch.flatten(x, 1) # [B, 512]x = self.backbone.fc(x) # [B, 10]return x# 初始化模型并移至设备
model = ResNet18_CBAM().to(device)
criterion = nn.CrossEntropyLoss()
optimizer = optim.Adam(model.parameters(), lr=0.001)
scheduler = optim.lr_scheduler.ReduceLROnPlateau(optimizer, mode='min', patience=3, factor=0.5)import time# ======================================================================
# 4. 结合了分阶段策略和详细打印的训练函数
# ======================================================================
def set_trainable_layers(model, trainable_parts):print(f"\n---> 解冻以下部分并设为可训练: {trainable_parts}")for name, param in model.named_parameters():param.requires_grad = Falsefor part in trainable_parts:if part in name:param.requires_grad = Truebreakdef train_staged_finetuning(model, criterion, train_loader, test_loader, device, epochs):optimizer = None# 初始化历史记录列表，与你的要求一致all_iter_losses, iter_indices = [], []train_acc_history, test_acc_history = [], []train_loss_history, test_loss_history = [], []for epoch in range(1, epochs + 1):epoch_start_time = time.time()# --- 动态调整学习率和冻结层 ---if epoch == 1:print("\n" + "="*50 + "\n🚀 **阶段 1：训练注意力模块和分类头**\n" + "="*50)set_trainable_layers(model, ["cbam", "backbone.fc"])optimizer = optim.Adam(filter(lambda p: p.requires_grad, model.parameters()), lr=1e-3)elif epoch == 6:print("\n" + "="*50 + "\n✈️ **阶段 2：解冻高层卷积层 (layer3, layer4)**\n" + "="*50)set_trainable_layers(model, ["cbam", "backbone.fc", "backbone.layer3", "backbone.layer4"])optimizer = optim.Adam(filter(lambda p: p.requires_grad, model.parameters()), lr=1e-4)elif epoch == 21:print("\n" + "="*50 + "\n🛰️ **阶段 3：解冻所有层，进行全局微调**\n" + "="*50)for param in model.parameters(): param.requires_grad = Trueoptimizer = optim.Adam(model.parameters(), lr=1e-5)# --- 训练循环 ---model.train()running_loss, correct, total = 0.0, 0, 0for batch_idx, (data, target) in enumerate(train_loader):data, target = data.to(device), target.to(device)optimizer.zero_grad()output = model(data)loss = criterion(output, target)loss.backward()optimizer.step()# 记录每个iteration的损失iter_loss = loss.item()all_iter_losses.append(iter_loss)iter_indices.append((epoch - 1) * len(train_loader) + batch_idx + 1)running_loss += iter_loss_, predicted = output.max(1)total += target.size(0)correct += predicted.eq(target).sum().item()# 按你的要求，每100个batch打印一次if (batch_idx + 1) % 100 == 0:print(f'Epoch: {epoch}/{epochs} | Batch: {batch_idx+1}/{len(train_loader)} 'f'| 单Batch损失: {iter_loss:.4f} | 累计平均损失: {running_loss/(batch_idx+1):.4f}')epoch_train_loss = running_loss / len(train_loader)epoch_train_acc = 100. * correct / totaltrain_loss_history.append(epoch_train_loss)train_acc_history.append(epoch_train_acc)# --- 测试循环 ---model.eval()test_loss, correct_test, total_test = 0, 0, 0with torch.no_grad():for data, target in test_loader:data, target = data.to(device), target.to(device)output = model(data)test_loss += criterion(output, target).item()_, predicted = output.max(1)total_test += target.size(0)correct_test += predicted.eq(target).sum().item()epoch_test_loss = test_loss / len(test_loader)epoch_test_acc = 100. * correct_test / total_testtest_loss_history.append(epoch_test_loss)test_acc_history.append(epoch_test_acc)# 打印每个epoch的最终结果print(f'Epoch {epoch}/{epochs} 完成 | 耗时: {time.time() - epoch_start_time:.2f}s | 训练准确率: {epoch_train_acc:.2f}% | 测试准确率: {epoch_test_acc:.2f}%')# 训练结束后调用绘图函数print("\n训练完成! 开始绘制结果图表...")plot_iter_losses(all_iter_losses, iter_indices)plot_epoch_metrics(train_acc_history, test_acc_history, train_loss_history, test_loss_history)# 返回最终的测试准确率return epoch_test_acc# ======================================================================
# 5. 绘图函数定义
# ======================================================================
def plot_iter_losses(losses, indices):plt.figure(figsize=(10, 4))plt.plot(indices, losses, 'b-', alpha=0.7, label='Iteration Loss')plt.xlabel('Iteration（Batch序号）')plt.ylabel('损失值')plt.title('每个 Iteration 的训练损失')plt.legend()plt.grid(True)plt.tight_layout()plt.show()def plot_epoch_metrics(train_acc, test_acc, train_loss, test_loss):epochs = range(1, len(train_acc) + 1)plt.figure(figsize=(12, 4))plt.subplot(1, 2, 1)plt.plot(epochs, train_acc, 'b-', label='训练准确率')plt.plot(epochs, test_acc, 'r-', label='测试准确率')plt.xlabel('Epoch')plt.ylabel('准确率 (%)')plt.title('训练和测试准确率')plt.legend(); plt.grid(True)plt.subplot(1, 2, 2)plt.plot(epochs, train_loss, 'b-', label='训练损失')plt.plot(epochs, test_loss, 'r-', label='测试损失')plt.xlabel('Epoch')plt.ylabel('损失值')plt.title('训练和测试损失')plt.legend(); plt.grid(True)plt.tight_layout()plt.show()# ======================================================================
# 6. 执行训练
# ======================================================================
model = ResNet18_CBAM().to(device)
criterion = nn.CrossEntropyLoss()
epochs = 50print("开始使用带分阶段微调策略的ResNet18+CBAM模型进行训练...")
final_accuracy = train_staged_finetuning(model, criterion, train_loader, test_loader, device, epochs)
print(f"训练完成！最终测试准确率: {final_accuracy:.2f}%")# torch.save(model.state_dict(), 'resnet18_cbam_finetuned.pth')
# print("模型已保存为: resnet18_cbam_finetuned.pth")@浙大疏锦行