开局一张图，首先抛出resnet18的网络架构（完整版放在文章最下方）

下面，再配合pytorch官方代码，解析一下resnet18。以resnet18为切入点，由浅入深，理解resnet架构

源码解析

demo

import torch
import torchvision.models as models
resnet18 = models.resnet18()

input = torch.randn(32,3,224,224)

output = resnet18(input)
print(resnet18)

ctrl+鼠标左键点击resnet18，进入resnet.py文件下
映入眼帘的是resnet18的构造函数

#构造函数  conv1为 1 层，conv2、conv3、conv4、conv5均为 4 层（每个 layer 有 2 个 BasicBlock，每个 BasicBlock 有 2 个卷积层），总共为 16 层，最后一层全连接层，总层数=1 + 4 × 4 + 1 = 18，依此类推。
def resnet18(pretrained: bool = False, progress: bool = True, **kwargs: Any) -> ResNet:
    r"""ResNet-18 model from
    `"Deep Residual Learning for Image Recognition" <https://arxiv.org/pdf/1512.03385.pdf>`_.

    Args:
        pretrained (bool): If True, returns a model pre-trained on ImageNet
        progress (bool): If True, displays a progress bar of the download to stderr
    """
    return _resnet('resnet18', BasicBlock, [2, 2, 2, 2], pretrained, progress,
                   **kwargs)

构造函数用到了_resnet()方法来创建网络

def _resnet(
    arch: str,
    block: Type[Union[BasicBlock, Bottleneck]],
    layers: List[int],
    pretrained: bool,
    progress: bool,
    **kwargs: Any
) -> ResNet:
    model = ResNet(block, layers, **kwargs)
    if pretrained: #加载预训练参数
        state_dict = load_state_dict_from_url(model_urls[arch],
                                              progress=progress)
        model.load_state_dict(state_dict)
    return model

在_resnet()方法中，又调用了ResNet()方法创建模型

class ResNet(nn.Module):

    def __init__(
        self,
        block: Type[Union[BasicBlock, Bottleneck]],
        layers: List[int],
        num_classes: int = 1000,
        zero_init_residual: bool = False,
        groups: int = 1,
        width_per_group: int = 64,
        replace_stride_with_dilation: Optional[List[bool]] = None,
        norm_layer: Optional[Callable[..., nn.Module]] = None
    ) -> None:
        super(ResNet, self).__init__()
        if norm_layer is None:#判断是否传入 norm_layer，没有传入，则使用 BatchNorm2d
            norm_layer = nn.BatchNorm2d
        self._norm_layer = norm_layer

        self.inplanes = 64
        self.dilation = 1
        if replace_stride_with_dilation is None: #判断是否传入空洞卷积参数 replace_stride_with_dilation，如果不指定，则赋值为 [False, False, False]，表示不使用空洞卷积。
            # each element in the tuple indicates if we should replace
            # the 2x2 stride with a dilated convolution instead
            replace_stride_with_dilation = [False, False, False]
        if len(replace_stride_with_dilation) != 3:
            raise ValueError("replace_stride_with_dilation should be None "
                             "or a 3-element tuple, got {}".format(replace_stride_with_dilation))
        self.groups = groups #读取分组卷积的参数
        self.base_width = width_per_group
        self.conv1 = nn.Conv2d(3, self.inplanes, kernel_size=7, stride=2, padding=3,
                               bias=False)
        self.bn1 = norm_layer(self.inplanes)
        self.relu = nn.ReLU(inplace=True)
        self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
        self.layer1 = self._make_layer(block, 64, layers[0]) #这个 layer 的参数没有指定 stride，默认 stride=1，因此这个 layer 不会改变图片大小
        self.layer2 = self._make_layer(block, 128, layers[1], stride=2,
                                       dilate=replace_stride_with_dilation[0]) #这个layer的stride=2，会降采样
        self.layer3 = self._make_layer(block, 256, layers[2], stride=2,
                                       dilate=replace_stride_with_dilation[1]) #这个layer的stride=2，会降采样
        self.layer4 = self._make_layer(block, 512, layers[3], stride=2,
                                       dilate=replace_stride_with_dilation[2]) #这个layer的stride=2，会降采样
        self.avgpool = nn.AdaptiveAvgPool2d((1, 1))
        self.fc = nn.Linear(512 * block.expansion, num_classes)

        #网络参数初始化
        for m in self.modules():
            if isinstance(m, nn.Conv2d): #卷积层采用 kaiming_normal_() 初始化方法  参考：https://arxiv.org/abs/1502.01852
                nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')
            elif isinstance(m, (nn.BatchNorm2d, nn.GroupNorm)): #bn 层和 GroupNorm 层初始化为 weight=1，bias=0
                nn.init.constant_(m.weight, 1)
                nn.init.constant_(m.bias, 0)

        # Zero-initialize the last BN in each residual branch,
        # so that the residual branch starts with zeros, and each residual block behaves like an identity.
        # This improves the model by 0.2~0.3% according to https://arxiv.org/abs/1706.02677
        #每个 BasicBlock 和 Bottleneck 的最后一层 bn 的 weight=0，可以提升准确率 0.2~0.3%。
        if zero_init_residual:
            for m in self.modules():
                if isinstance(m, Bottleneck):
                    nn.init.constant_(m.bn3.weight, 0)  # type: ignore[arg-type]
                elif isinstance(m, BasicBlock):
                    nn.init.constant_(m.bn2.weight, 0)  # type: ignore[arg-type]

    def _make_layer(self, block: Type[Union[BasicBlock, Bottleneck]], planes: int, blocks: int,
                    stride: int = 1, dilate: bool = False) -> nn.Sequential:
        norm_layer = self._norm_layer
        downsample = None
        previous_dilation = self.dilation
        if dilate:#判断空洞卷积，计算 previous_dilation 参数
            self.dilation *= stride
            stride = 1
        if stride != 1 or self.inplanes != planes * block.expansion: #判断 stride 是否为 1，输入通道和输出通道是否相等。如果这两个条件都不成立，那么表明需要建立一个 1 X 1 的卷积层，来改变通道数和改变图片大小。具体是建立 downsample 层，包括 conv1x1 -> norm_layer。
            downsample = nn.Sequential(
                conv1x1(self.inplanes, planes * block.expansion, stride),
                norm_layer(planes * block.expansion),
            )

        layers = []
        #建立第一个 block，把 downsample 传给 block 作为降采样的层，并且 stride 也使用传入的 stride（stride=2）
        layers.append(block(self.inplanes, planes, stride, downsample, self.groups,
                            self.base_width, previous_dilation, norm_layer))
        self.inplanes = planes * block.expansion #改变通道数self.inplanes = planes * block.expansion。备注：在 BasicBlock 里，expansion=1，因此这一步不会改变通道数；在 Bottleneck 里，expansion=4，因此这一步会改变通道数
        #图片经过第一个 block后，就会改变通道数和图片大小。接下来 for 循环添加剩下的 block。从第 2 个 block 起，输入和输出通道数是相等的，因此就不用传入 downsample 和 stride（那么 block 的 stride 默认使用 1）
        for _ in range(1, blocks): #继续添加这个 layer 里接下来的 BasicBlock
            layers.append(block(self.inplanes, planes, groups=self.groups,
                                base_width=self.base_width, dilation=self.dilation,
                                norm_layer=norm_layer))

        return nn.Sequential(*layers)

    def _forward_impl(self, x: Tensor) -> Tensor: #整个resnet18的前向传播结构，可以参考文章最后的完整结构图来理解
        # See note [TorchScript super()]
        x = self.conv1(x)
        print("131254214")
        print(x.size())
        x = self.bn1(x)
        x = self.relu(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 x

    def forward(self, x: Tensor) -> Tensor:
        return self._forward_impl(x)

可以看到，上面每个layer都是使用_make_layer()方法来创建层，下面来解析_make_layer()

    def _make_layer(self, block: Type[Union[BasicBlock, Bottleneck]], planes: int, blocks: int,
                    stride: int = 1, dilate: bool = False) -> nn.Sequential:
        norm_layer = self._norm_layer
        downsample = None
        previous_dilation = self.dilation
        if dilate:#判断空洞卷积，计算 previous_dilation 参数
            self.dilation *= stride
            stride = 1
        if stride != 1 or self.inplanes != planes * block.expansion: #判断 stride 是否为 1，输入通道和输出通道是否相等。如果这两个条件都不成立，那么表明需要建立一个 1 X 1 的卷积层，来改变通道数和改变图片大小。具体是建立 downsample 层，包括 conv1x1 -> norm_layer。
            downsample = nn.Sequential(
                conv1x1(self.inplanes, planes * block.expansion, stride),
                norm_layer(planes * block.expansion),
            )

        layers = []
        #建立第一个 block，把 downsample 传给 block 作为降采样的层，并且 stride 也使用传入的 stride（stride=2）
        layers.append(block(self.inplanes, planes, stride, downsample, self.groups,
                            self.base_width, previous_dilation, norm_layer))
        self.inplanes = planes * block.expansion #改变通道数self.inplanes = planes * block.expansion。备注：在 BasicBlock 里，expansion=1，因此这一步不会改变通道数；在 Bottleneck 里，expansion=4，因此这一步会改变通道数
        #图片经过第一个 block后，就会改变通道数和图片大小。接下来 for 循环添加剩下的 block。从第 2 个 block 起，输入和输出通道数是相等的，因此就不用传入 downsample 和 stride（那么 block 的 stride 默认使用 1）
        for _ in range(1, blocks): #继续添加这个 layer 里接下来的 BasicBlock
            layers.append(block(self.inplanes, planes, groups=self.groups,
                                base_width=self.base_width, dilation=self.dilation,
                                norm_layer=norm_layer))

        return nn.Sequential(*layers)

上面的过程如下图所示：

resnet18用的是BasicBlock

class BasicBlock(nn.Module):
    expansion: int = 1

    def __init__(
        self,
        inplanes: int,
        planes: int,
        stride: int = 1,
        downsample: Optional[nn.Module] = None,
        groups: int = 1,
        base_width: int = 64,
        dilation: int = 1,
        norm_layer: Optional[Callable[..., nn.Module]] = None
    ) -> None:
        super(BasicBlock, self).__init__()
        if norm_layer is None: #判断是否传入了 norm_layer 层，如果没有，则使用 BatchNorm2d。
            norm_layer = nn.BatchNorm2d
        if groups != 1 or base_width != 64: #校验参数：groups == 1，base_width == 64，dilation == 1。也就是说，在 BasicBlock 中，不使用空洞卷积和分组卷积。
            raise ValueError('BasicBlock only supports groups=1 and base_width=64')
        if dilation > 1:
            raise NotImplementedError("Dilation > 1 not supported in BasicBlock")
        # Both self.conv1 and self.downsample layers downsample the input when stride != 1
        self.conv1 = conv3x3(inplanes, planes, stride) #定义第 1 组 conv3x3 -> norm_layer -> relu，这里使用传入的 stride 和 inplanes。（如果是 layer2 ，layer3 ，layer4 里的第一个 BasicBlock，那么 stride=2，这里会降采样和改变通道数）。
        self.bn1 = norm_layer(planes)
        self.relu = nn.ReLU(inplace=True)
        self.conv2 = conv3x3(planes, planes) #定义第 2 组 conv3x3 -> norm_layer -> relu，这里不使用传入的 stride （默认为 1），输入通道数和输出通道数使用planes，也就是不需要降采样和改变通道数。
        self.bn2 = norm_layer(planes)
        self.downsample = downsample
        self.stride = stride

    def forward(self, x: Tensor) -> Tensor:
        identity = x

        out = self.conv1(x)
        out = self.bn1(out)
        out = self.relu(out)

        out = self.conv2(out)
        out = self.bn2(out)

        if self.downsample is not None:
            identity = self.downsample(x)

        out += identity
        out = self.relu(out)

        return out

备注：resnet18,34用的是BasicBlock，而resnet50,101等都用的是Bottleneck结构

Bottleneck结构源码如下：

class Bottleneck(nn.Module):
    # Bottleneck in torchvision places the stride for downsampling at 3x3 convolution(self.conv2)
    # while original implementation places the stride at the first 1x1 convolution(self.conv1)
    # according to "Deep residual learning for image recognition"https://arxiv.org/abs/1512.03385.
    # This variant is also known as ResNet V1.5 and improves accuracy according to
    # https://ngc.nvidia.com/catalog/model-scripts/nvidia:resnet_50_v1_5_for_pytorch.

    expansion: int = 4

    def __init__(
        self,
        inplanes: int,
        planes: int,
        stride: int = 1,
        downsample: Optional[nn.Module] = None,
        groups: int = 1,
        base_width: int = 64,
        dilation: int = 1,
        norm_layer: Optional[Callable[..., nn.Module]] = None
    ) -> None:
        super(Bottleneck, self).__init__()
        if norm_layer is None:
            norm_layer = nn.BatchNorm2d
        width = int(planes * (base_width / 64.)) * groups #计算 width，等于传入的 planes，用于中间的3 × 3卷积。
        # Both self.conv2 and self.downsample layers downsample the input when stride != 1
        self.conv1 = conv1x1(inplanes, width) #定义第 1 组 conv1x1 -> norm_layer，这里不使用传入的 stride，使用 width，作用是进行降维，减少通道数。
        self.bn1 = norm_layer(width)
        self.conv2 = conv3x3(width, width, stride, groups, dilation) #定义第 2 组 conv3x3 -> norm_layer，这里使用传入的 stride，输入通道数和输出通道数使用width。（如果是 layer2 ，layer3 ，layer4 里的第一个 Bottleneck，那么 stride=2，这里会降采样）。
        self.bn2 = norm_layer(width)
        self.conv3 = conv1x1(width, planes * self.expansion) #定义第 3 组 conv1x1 -> norm_layer，这里不使用传入的 stride，使用 planes * self.expansion，作用是进行升维，增加通道数。
        self.bn3 = norm_layer(planes * self.expansion)
        self.relu = nn.ReLU(inplace=True)
        self.downsample = downsample
        self.stride = stride

    def forward(self, x: Tensor) -> Tensor:
        identity = x

        out = self.conv1(x)
        out = self.bn1(out)
        out = self.relu(out)

        out = self.conv2(out)
        out = self.bn2(out)
        out = self.relu(out)

        out = self.conv3(out)
        out = self.bn3(out)

        if self.downsample is not None:
            identity = self.downsample(x)

        out += identity
        out = self.relu(out)

        return out

BasicBlock与Bottleneck的不同：

• BasicBlock 中有 1 个 3 × 3 3 \times 3 3×3卷积层，如果是 layer 的第一个 BasicBlock，那么第一个卷积层的 stride=2，作用是进行降采样。
• Bottleneck 中有 2 个 1 × 1 1 \times 1 1×1卷积层， 1 个 3 × 3 3 \times 3 3×3卷积层。先经过第 1 个 1 × 1 1 \times 1 1×1卷积层，进行降维，然后经过 3 × 3 3 \times 3 3×3卷积层（如果是 layer 的第一个 Bottleneck，那么 3 × 3 3 \times 3 3×3卷积层的 stride=2，作用是进行降采样），最后经过 1 × 1 1 \times 1 1×1卷积层，进行升维 。

最后抛出resnet18的完整架构图：

标签：layer,self,ResNet,残差,stride,block,out,norm,里程碑
来源： https://blog.csdn.net/wl1780852311/article/details/123113586