@浙大疏錦行?python day50.

• 在預訓練模型（resnet18）中添加cbam注意力機制，需要修改模型的架構，同時應該考慮插入的cbam注意力機制模塊的位置；

import torch
import torch.nn as nn
from torchvision import models# 自定義ResNet18模型，插入CBAM模塊
class ResNet18_CBAM(nn.Module):def __init__(self, num_classes=10, pretrained=True, cbam_ratio=16, cbam_kernel=7):super().__init__()# 加載預訓練ResNet18self.backbone = models.resnet18(pretrained=pretrained) # 修改首層卷積以適應32x32輸入（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層（因輸入尺寸小）# 在每個殘差塊組后添加CBAM模塊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)# 修改分類頭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]# 第一層殘差塊 + CBAMx = self.backbone.layer1(x)  # [B, 64, 32, 32]x = self.cbam_layer1(x)# 第二層殘差塊 + CBAMx = self.backbone.layer2(x)  # [B, 128, 16, 16]x = self.cbam_layer2(x)# 第三層殘差塊 + CBAMx = self.backbone.layer3(x)  # [B, 256, 8, 8]x = self.cbam_layer3(x)# 第四層殘差塊 + CBAMx = self.backbone.layer4(x)  # [B, 512, 4, 4]x = self.cbam_layer4(x)# 全局平均池化 + 分類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# 初始化模型并移至設備
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)

• 修改模型結構后，需要考慮模型訓練的策略，一般來說可以先凍結原有的部分進行訓練以期待新增的部分可以獲得一個不錯的表現；之后解凍原有部分中的高層layer并賦予一個較低的學習率來保證不會出現不應該的錯誤；最后解凍所有參數，也是賦予較低的學習率，來學習最終的端到端任務。

def set_trainable_layers(model, trainable_parts):print(f"\n---> 解凍以下部分并設為可訓練: {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# 初始化歷史記錄列表，與你的要求一致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()# --- 動態調整學習率和凍結層 ---if epoch == 1:print("\n" + "="*50 + "\n🚀 **階段 1：訓練注意力模塊和分類頭**\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?? **階段 2：解凍高層卷積層 (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🛰? **階段 3：解凍所有層，進行全局微調**\n" + "="*50)for param in model.parameters(): param.requires_grad = Trueoptimizer = optim.Adam(model.parameters(), lr=1e-5)# --- 訓練循環 ---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()# 記錄每個iteration的損失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個batch打印一次if (batch_idx + 1) % 100 == 0:print(f'Epoch: {epoch}/{epochs} | Batch: {batch_idx+1}/{len(train_loader)} 'f'| 單Batch損失: {iter_loss:.4f} | 累計平均損失: {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)# --- 測試循環 ---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)# 打印每個epoch的最終結果print(f'Epoch {epoch}/{epochs} 完成 | 耗時: {time.time() - epoch_start_time:.2f}s | 訓練準確率: {epoch_train_acc:.2f}% | 測試準確率: {epoch_test_acc:.2f}%')# 訓練結束后調用繪圖函數print("\n訓練完成! 開始繪制結果圖表...")plot_iter_losses(all_iter_losses, iter_indices)plot_epoch_metrics(train_acc_history, test_acc_history, train_loss_history, test_loss_history)# 返回最終的測試準確率return epoch_test_accmodel = ResNet18_CBAM().to(device)
criterion = nn.CrossEntropyLoss()
epochs = 50print("開始使用帶分階段微調策略的ResNet18+CBAM模型進行訓練...")
final_accuracy = train_staged_finetuning(model, criterion, train_loader, test_loader, device, epochs)
print(f"訓練完成！最終測試準確率: {final_accuracy:.2f}%")torch.save(model.state_dict(), 'resnet18_cbam_finetuned.pth')
print("模型已保存為: resnet18_cbam_finetuned.pth")