1. 为什么需要残差网络在2015年之前深度学习领域普遍认为网络越深性能越好。从LeNet到AlexNet再到VGG和GoogLeNet网络深度确实在不断加深。但研究者很快发现一个奇怪现象当网络深度超过某个阈值后继续增加层数反而会导致训练误差和测试误差同时增大。这不是过拟合问题因为过拟合应该表现为训练误差降低而测试误差升高。这种现象被称为退化问题Degradation Problem。想象一下你正在教一个学生做数学题给他100道练习题后成绩提高了但给他1000道题后成绩反而下降——这显然不是因为他学过头了而是训练方法出了问题。残差网络ResNet的提出者何恺明团队发现深层网络难以训练的关键在于随着网络加深梯度信号在反向传播时会逐渐减弱。即使使用了BatchNorm等技巧深层网络的优化仍然困难。这就好比你在玩传话游戏一句话经过20个人传递后最后一个人听到的内容可能已经面目全非。2. 残差块ResNet的核心创新2.1 残差学习的基本思想传统神经网络直接学习目标映射H(x)而ResNet改为学习残差映射F(x) H(x) - x。这个看似简单的改动带来了深远影响恒等映射的保底作用当F(x)0时H(x)x网络至少能保持浅层网络的性能梯度高速公路通过跨层连接shortcut梯度可以直接回传到浅层特征复用深层网络可以专注于学习新增特征而非重复学习浅层特征用一个生活类比假设你要从北京到上海传统网络像让你一步步走到上海而ResNet则是先让你坐高铁到南京学习残差再补充走到上海这段距离。2.2 两种残差块结构ResNet使用了两种基本模块适用于不同深度的网络BasicBlock用于ResNet-18/34class BasicBlock(nn.Module): def __init__(self, in_channels, out_channels, stride1): super().__init__() self.conv1 nn.Conv2d(in_channels, out_channels, kernel_size3, stridestride, padding1, biasFalse) self.bn1 nn.BatchNorm2d(out_channels) self.conv2 nn.Conv2d(out_channels, out_channels, kernel_size3, stride1, padding1, biasFalse) self.bn2 nn.BatchNorm2d(out_channels) # 当输入输出维度不一致时需要1x1卷积调整维度 self.shortcut nn.Sequential() if stride ! 1 or in_channels ! out_channels: self.shortcut nn.Sequential( nn.Conv2d(in_channels, out_channels, kernel_size1, stridestride, biasFalse), nn.BatchNorm2d(out_channels) ) def forward(self, x): out F.relu(self.bn1(self.conv1(x))) out self.bn2(self.conv2(out)) out self.shortcut(x) # 关键残差连接 return F.relu(out)Bottleneck用于ResNet-50/101/152class Bottleneck(nn.Module): expansion 4 # 输出通道的倍乘系数 def __init__(self, in_channels, out_channels, stride1): super().__init__() mid_channels out_channels // self.expansion self.conv1 nn.Conv2d(in_channels, mid_channels, kernel_size1, stride1, biasFalse) self.bn1 nn.BatchNorm2d(mid_channels) self.conv2 nn.Conv2d(mid_channels, mid_channels, kernel_size3, stridestride, padding1, biasFalse) self.bn2 nn.BatchNorm2d(mid_channels) self.conv3 nn.Conv2d(mid_channels, out_channels, kernel_size1, stride1, biasFalse) self.bn3 nn.BatchNorm2d(out_channels) self.shortcut nn.Sequential() if stride ! 1 or in_channels ! out_channels: self.shortcut nn.Sequential( nn.Conv2d(in_channels, out_channels, kernel_size1, stridestride, biasFalse), nn.BatchNorm2d(out_channels) ) def forward(self, x): out F.relu(self.bn1(self.conv1(x))) out F.relu(self.bn2(self.conv2(out))) out self.bn3(self.conv3(out)) out self.shortcut(x) return F.relu(out)Bottleneck结构通过1x1卷积先降维再升维大幅减少了参数量。例如对于输入256维的特征BasicBlock需要3x3x256x256 3x3x256x256 ≈ 1.18M参数Bottleneck只需1x1x256x64 3x3x64x64 1x1x64x256 ≈ 70k参数3. 完整ResNet架构实现3.1 网络组装逻辑ResNet通过_make_layer函数将残差块组装成各个阶段class ResNet(nn.Module): def __init__(self, block, layers, num_classes1000): super().__init__() self.in_channels 64 # 初始卷积层 self.conv1 nn.Conv2d(3, 64, kernel_size7, stride2, padding3, biasFalse) self.bn1 nn.BatchNorm2d(64) self.maxpool nn.MaxPool2d(kernel_size3, stride2, padding1) # 四个残差阶段 self.layer1 self._make_layer(block, 64, layers[0], stride1) self.layer2 self._make_layer(block, 128, layers[1], stride2) self.layer3 self._make_layer(block, 256, layers[2], stride2) self.layer4 self._make_layer(block, 512, layers[3], stride2) # 分类头 self.avgpool nn.AdaptiveAvgPool2d((1, 1)) self.fc nn.Linear(512 * block.expansion, num_classes) def _make_layer(self, block, out_channels, blocks, stride1): layers [] # 第一个块可能需要下采样 layers.append(block(self.in_channels, out_channels, stride)) self.in_channels out_channels * block.expansion # 后续块保持维度不变 for _ in range(1, blocks): layers.append(block(self.in_channels, out_channels)) return nn.Sequential(*layers) def forward(self, x): x F.relu(self.bn1(self.conv1(x))) x self.maxpool(x) x self.layer1(x) x self.layer2(x) x self.layer3(x) x self.layer4(x) x self.avgpool(x) x torch.flatten(x, 1) x self.fc(x) return x3.2 不同深度ResNet配置def resnet18(num_classes1000): return ResNet(BasicBlock, [2, 2, 2, 2], num_classes) def resnet34(num_classes1000): return ResNet(BasicBlock, [3, 4, 6, 3], num_classes) def resnet50(num_classes1000): return ResNet(Bottleneck, [3, 4, 6, 3], num_classes) def resnet101(num_classes1000): return ResNet(Bottleneck, [3, 4, 23, 3], num_classes) def resnet152(num_classes1000): return ResNet(Bottleneck, [3, 8, 36, 3], num_classes)4. 实战在Fashion-MNIST上训练ResNet4.1 数据准备与增强虽然Fashion-MNIST原始尺寸是28x28但我们可以将其上采样到224x224以适应ResNettransform transforms.Compose([ transforms.Resize(224), transforms.ToTensor(), transforms.Normalize((0.5,), (0.5,)) ]) train_set datasets.FashionMNIST(root./data, trainTrue, downloadTrue, transformtransform) test_set datasets.FashionMNIST(root./data, trainFalse, downloadTrue, transformtransform) train_loader DataLoader(train_set, batch_size64, shuffleTrue) test_loader DataLoader(test_set, batch_size64, shuffleFalse)4.2 模型训练技巧学习率预热初始几轮使用较小学习率余弦退火平滑降低学习率混合精度训练节省显存并加速model resnet18(num_classes10) model.conv1 nn.Conv2d(1, 64, kernel_size7, stride2, padding3, biasFalse) # 修改输入通道 device torch.device(cuda if torch.cuda.is_available() else cpu) model model.to(device) criterion nn.CrossEntropyLoss() optimizer torch.optim.SGD(model.parameters(), lr0.1, momentum0.9, weight_decay1e-4) scheduler torch.optim.lr_scheduler.CosineAnnealingLR(optimizer, T_max10) scaler torch.cuda.amp.GradScaler() # 混合精度训练 for epoch in range(10): model.train() for images, labels in train_loader: images, labels images.to(device), labels.to(device) optimizer.zero_grad() with torch.cuda.amp.autocast(): outputs model(images) loss criterion(outputs, labels) scaler.scale(loss).backward() scaler.step(optimizer) scaler.update() scheduler.step() # 测试集评估 model.eval() correct 0 with torch.no_grad(): for images, labels in test_loader: images, labels images.to(device), labels.to(device) outputs model(images) _, predicted torch.max(outputs.data, 1) correct (predicted labels).sum().item() print(fEpoch {epoch1}, Accuracy: {100 * correct / len(test_set):.2f}%)5. 进阶技巧与常见问题5.1 残差连接的设计变体Pre-activation将BN和ReLU放在卷积之前ResNet V2Group Conv在残差块中使用分组卷积ResNeXtAttention机制引入SE模块SENet5.2 训练深层ResNet的注意事项使用Kaiming初始化nn.init.kaiming_normal_(m.weight, modefan_out, nonlinearityrelu)对于超过100层的网络建议增加batch size至少256使用warmup学习率策略添加额外的正则化如Label Smoothing5.3 模型部署优化剪枝移除不重要的通道量化将FP32转为INT8TensorRT加速转换模型为引擎文件# 示例模型量化 quantized_model torch.quantization.quantize_dynamic( model, {nn.Linear, nn.Conv2d}, dtypetorch.qint8 )在实际项目中ResNet-18量化后模型大小可减少4倍推理速度提升2-3倍。我曾在一个工业质检项目中通过量化将ResNet-50的推理时间从15ms降到6ms满足了产线实时检测的需求。