ResNet-18网络模型+原理解析+Pytorch实现+手写模型

ResNet-18 通过引入残差连接（Skip Connection） ，有效解决了深层卷积神经网络（CNN）中常见的梯度消失 与网络退化问题。

论文 https://arxiv.org/pdf/1512.03385v1

对于传统的深度学习网络，我们普遍认为网络深度越深（参数越多）非线性的表达能力越强，该网络所能学习到的东西就越多。 凭借这一基本规则，经典的CNN网络从LetNet-5（5层）和AlexNet（8层）发展到VGGNet(16-19)，再到后来GoogleNet（22层）。 但是后来发现传统的CNN网络结构随着层数加深到一定程度之后，越深的网络反而效果更差，过深的网络竟然使分类的准确率下降了,如图所示 56层网络的错误率 高于 20层的错误率

网络退化

网络退化（Network Degradation）是指随着神经网络深度的增加，模型的性能并未如预期那样持续提升，反而可能出现饱和乃至下降的现象。这种现象并非由过拟合引起，因为在更深的网络中，即使在训练数据上的表现也开始变差，导致训练误差和测试误差之间的差距可能增大。

---

残差结构

传统深层网络直接学习映射 (H(x))，层数加深时梯度消失、性能下降。 ResNet 改为学习残差 ： H(x)=F(x)+x

• x：输入特征（跳跃连接，直接传到输出）

• F(x)：残差函数（2层3×3卷积+BN+ReLU）

• 优势：

  • 梯度可沿跳跃连接直达浅层，缓解梯度消失

  • 网络易拟合恒等映射（F(x)\to0)），深层训练更稳

关键特性

1. 轻量高效 ：约11M参数 、1.8G FLOPs，推理快，适合移动端/嵌入式。

2. 模块化设计：4个Stage，每阶段2个Block，易扩展（如ResNet-34/50）。

3. 强表征能力：通道从64→512，空间逐级下采样，兼顾细粒度与全局特征。

4. 训练稳定：残差连接让深层梯度有效回传，收敛快、泛化性好。

典型应用

• 图像分类：ImageNet、CIFAR-10/100、自定义数据集

• 迁移学习：特征提取、微调小样本任务

• 检测/分割骨干：Faster R-CNN、YOLO、Mask R-CNN、U-Net

• 轻量化部署：手机、边缘设备、实时系统

ResNet-18 结构总览

（输入 3×224×224） 共18个可训练层 （17卷积+1全连接），分5阶段：

阶段	操作	输出尺寸	通道数	层数
输入	RGB图像	224×224	3	-
Stage1	7×7卷积(64, stride=2)+BN+ReLU+MaxPool(3×3, stride=2)	56×56	64	1
Stage2	2×BasicBlock（3×3卷积×2）	56×56	64	4
Stage3	2×BasicBlock（第一个stride=2下采样）	28×28	128	4
Stage4	2×BasicBlock（第一个stride=2下采样）	14×14	256	4
Stage5	2×BasicBlock（第一个stride=2下采样）	7×7	512	4
分类头	全局平均池化+全连接(1000)	1×1	512→1000	1

BasicBlock（基础残差块）

• 左图：Conv(3×3)→BN→ReLU→Conv(3×3)→BN

• 右图跳跃连接：维度匹配时用恒等映射 ；不匹配时用1×1卷积 调整通道/步长

代码示例

​

ResNet18
@Author:  CSDN@王_teacher  bilibili@王_teacher
@Date: 2024/5/26
@Description: RestNet18 纯手写
"""
import torch
import torch.nn as nn
from torchvision.models import resnet18
​
# 纯平铺版 ResNet18，没有任何子模块、没有任何函数调用
class ResNet18(nn.Module):
    def __init__(self, num_classes=1000):
        super().__init__()
​
        # ==================== Stem 层 ====================
        self.conv1 = nn.Conv2d(3, 64, kernel_size=7, stride=2, padding=3, bias=False)
        self.bn1 = nn.BatchNorm2d(64)
        self.relu = nn.ReLU(inplace=True)
        self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
​
        # ==================== Layer 1 (64, 无下采样) ====================
        # Block 1
        self.layer1_conv1 = nn.Conv2d(64, 64, 3, 1, 1, bias=False)
        self.layer1_bn1 = nn.BatchNorm2d(64)
        self.layer1_conv2 = nn.Conv2d(64, 64, 3, 1, 1, bias=False)
        self.layer1_bn2 = nn.BatchNorm2d(64)
​
        # Block 2
        self.layer1_conv3 = nn.Conv2d(64, 64, 3, 1, 1, bias=False)
        self.layer1_bn3 = nn.BatchNorm2d(64)
        self.layer1_conv4 = nn.Conv2d(64, 64, 3, 1, 1, bias=False)
        self.layer1_bn4 = nn.BatchNorm2d(64)
​
        # ==================== Layer 2 (128, 下采样) ====================
        # Downsample
        self.downsample2 = nn.Sequential(
            nn.Conv2d(64, 128, 1, 2, bias=False),
            nn.BatchNorm2d(128)
        )
        # Block 1
        self.layer2_conv1 = nn.Conv2d(64, 128, 3, 2, 1, bias=False)
        self.layer2_bn1 = nn.BatchNorm2d(128)
        self.layer2_conv2 = nn.Conv2d(128, 128, 3, 1, 1, bias=False)
        self.layer2_bn2 = nn.BatchNorm2d(128)
​
        # Block 2
        self.layer2_conv3 = nn.Conv2d(128, 128, 3, 1, 1, bias=False)
        self.layer2_bn3 = nn.BatchNorm2d(128)
        self.layer2_conv4 = nn.Conv2d(128, 128, 3, 1, 1, bias=False)
        self.layer2_bn4 = nn.BatchNorm2d(128)
​
        # ==================== Layer3 (256, 下采样) ====================
        self.downsample3 = nn.Sequential(
            nn.Conv2d(128, 256, 1, 2, bias=False),
            nn.BatchNorm2d(256)
        )
        self.layer3_conv1 = nn.Conv2d(128, 256, 3, 2, 1, bias=False)
        self.layer3_bn1 = nn.BatchNorm2d(256)
        self.layer3_conv2 = nn.Conv2d(256, 256, 3, 1, 1, bias=False)
        self.layer3_bn2 = nn.BatchNorm2d(256)
​
        self.layer3_conv3 = nn.Conv2d(256, 256, 3, 1, 1, bias=False)
        self.layer3_bn3 = nn.BatchNorm2d(256)
        self.layer3_conv4 = nn.Conv2d(256, 256, 3, 1, 1, bias=False)
        self.layer3_bn4 = nn.BatchNorm2d(256)
​
        # ==================== Layer4 (512, 下采样) ====================
        self.downsample4 = nn.Sequential(
            nn.Conv2d(256, 512, 1, 2, bias=False),
            nn.BatchNorm2d(512)
        )
        self.layer4_conv1 = nn.Conv2d(256, 512, 3, 2, 1, bias=False)
        self.layer4_bn1 = nn.BatchNorm2d(512)
        self.layer4_conv2 = nn.Conv2d(512, 512, 3, 1, 1, bias=False)
        self.layer4_bn2 = nn.BatchNorm2d(512)
​
        self.layer4_conv3 = nn.Conv2d(512, 512, 3, 1, 1, bias=False)
        self.layer4_bn3 = nn.BatchNorm2d(512)
        self.layer4_conv4 = nn.Conv2d(512, 512, 3, 1, 1, bias=False)
        self.layer4_bn4 = nn.BatchNorm2d(512)
​
        # ==================== 头部 ====================
        self.avgpool = nn.AdaptiveAvgPool2d((1, 1))
        self.fc = nn.Linear(512, num_classes)
​
    def forward(self, x):
        # Stem
        x = self.conv1(x)
        x = self.bn1(x)
        x = self.relu(x)
        x = self.maxpool(x)
​
        # ==================== Layer 1 ====================
        identity = x
        x = self.layer1_conv1(x)
        x = self.layer1_bn1(x)
        x = self.relu(x)
        x = self.layer1_conv2(x)
        x = self.layer1_bn2(x)
        x += identity
        x = self.relu(x)
​
        identity = x
        x = self.layer1_conv3(x)
        x = self.layer1_bn3(x)
        x = self.relu(x)
        x = self.layer1_conv4(x)
        x = self.layer1_bn4(x)
        x += identity
        x = self.relu(x)
​
        # ==================== Layer 2 ====================
        identity = self.downsample2(x)
        x = self.layer2_conv1(x)
        x = self.layer2_bn1(x)
        x = self.relu(x)
        x = self.layer2_conv2(x)
        x = self.layer2_bn2(x)
        x += identity
        x = self.relu(x)
​
        identity = x
        x = self.layer2_conv3(x)
        x = self.layer2_bn3(x)
        x = self.relu(x)
        x = self.layer2_conv4(x)
        x = self.layer2_bn4(x)
        x += identity
        x = self.relu(x)
​
        # ==================== Layer3 ====================
        identity = self.downsample3(x)
        x = self.layer3_conv1(x)
        x = self.layer3_bn1(x)
        x = self.relu(x)
        x = self.layer3_conv2(x)
        x = self.layer3_bn2(x)
        x += identity
        x = self.relu(x)
​
        identity = x
        x = self.layer3_conv3(x)
        x = self.layer3_bn3(x)
        x = self.relu(x)
        x = self.layer3_conv4(x)
        x = self.layer3_bn4(x)
        x += identity
        x = self.relu(x)
​
        # ==================== Layer4 ====================
        identity = self.downsample4(x)
        x = self.layer4_conv1(x)
        x = self.layer4_bn1(x)
        x = self.relu(x)
        x = self.layer4_conv2(x)
        x = self.layer4_bn2(x)
        x += identity
        x = self.relu(x)
​
        identity = x
        x = self.layer4_conv3(x)
        x = self.layer4_bn3(x)
        x = self.relu(x)
        x = self.layer4_conv4(x)
        x = self.layer4_bn4(x)
        x += identity
        x = self.relu(x)
​
        # 分类头
        x = self.avgpool(x)
        x = torch.flatten(x, 1)
        x = self.fc(x)
        return x
​
if __name__ == '__main__':
    torch.manual_seed(0)
​
    # 模型
    model = ResNet18()
​
​
    x = torch.randn(1,3,224,224)
    model.eval()
    py = model.forward(x)
    print(py.shape)