resnet.py源码简要分析

上篇博客主要是阅读论文《Deep Residual Learning for Image Recognition》时，按照自己简单理解做的笔记，本文主要根据自己的理解，先对论文中的部分代码从总体结构上进行简单分析，很多具体细节暂时还未完全明白。官方代码链接如下：https://github.com/pytorch/vision/blob/master/torchvision/models/resnet.py

1、model_urls

model_urls = {		#这里保存的地址里是不同网络对应的预训练权重
    'resnet18': 'https://download.pytorch.org/models/resnet18-f37072fd.pth',
    'resnet34': 'https://download.pytorch.org/models/resnet34-b627a593.pth',
    'resnet50': 'https://download.pytorch.org/models/resnet50-0676ba61.pth',
}

2、基本卷积结构

def conv3x3(in_planes: int, out_planes: int, stride: int = 1, groups: int = 1, dilation: int = 1) -> nn.Conv2d:
    """3x3 convolution with padding"""
    return nn.Conv2d(in_planes, out_planes, kernel_size=3, stride=stride,
                     padding=dilation, groups=groups, bias=False, dilation=dilation)	#使用nn网络中的2维卷积封装成resnet中常用到的3x3卷积结构

def conv1x1(in_planes: int, out_planes: int, stride: int = 1) -> nn.Conv2d:				#封装成1x1卷积结构
    """1x1 convolution"""
return nn.Conv2d(in_planes, out_planes, kernel_size=1, stride=stride, bias=False)

3、残差网络结构

残差结构由BasicBlock类和Bottleneck类构成，两个类都继承自nn.Module类，并且都重写了类中的init()方法和forward()方法，这两个类实现的都是残差网络部分，区别在于其中包含的卷积层数不同，因此也适用于不同类型的resnet网络。
BasicBlock主要用来构建resnet18和resnet34网络，里面只包含两个卷积层，使用的都是前面已经封装好的conv3x3的卷积结构，具体实现如下：

class BasicBlock(nn.Module):
    expansion: int = 1
    def __init__(
        self,
        inplanes: int,		#输入通道数	
        planes: int,		#输出通道数
        stride: int = 1,	#步长
        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:
            norm_layer = nn.BatchNorm2d
        if groups != 1 or base_width != 64:
            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)	#第一层使用3x3卷积
        self.bn1 = norm_layer(planes)					#归一化
        self.relu = nn.ReLU(inplace=True)				#激活函数
        self.conv2 = conv3x3(planes, planes)			#第二层使用的也是3x3卷积
        self.bn2 = norm_layer(planes)
        self.downsample = downsample
        self.stride = stride

    def forward(self, x: Tensor) -> Tensor:		#前向传播函数
        identity = x			#保存输入值

        out = self.conv1(x)		#第一层卷积
        out = self.bn1(out)		#BN层
        out = self.relu(out)	#relu激活函数

        out = self.conv2(out)	#第二层卷积
        out = self.bn2(out)	bn

        if self.downsample is not None:	#当尺寸不匹配时，需要使用下采样方法调整维度使其一致
            identity = self.downsample(x)

        out += identity					#这里与传来的输入值直接相加作为结果输出，即是残差网络的核心。
        out = self.relu(out)			#再使用一层激活函数

        return out

Bottleneck主要用在resnet50及以上的网络结构，它和BasicBlock不同，内部是一个三层残差网块，使用的也是前面封装好的conv1x1，conv3x3，conv1x1的卷积结构，这三层卷积分别用来压缩维度、卷积处理和恢复维度。

class Bottleneck(nn.Module):
    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
        # Both self.conv2 and self.downsample layers downsample the input when stride != 1
        self.conv1 = conv1x1(inplanes, width)	#1x1卷积网络
        self.bn1 = norm_layer(width)			#归一化
        self.conv2 = conv3x3(width, width, stride, groups, dilation)	#3x3网络
        self.bn2 = norm_layer(width)
        self.conv3 = conv1x1(width, planes * self.expansion)	#1x1网络
        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:	#前向传播函数
        identity = x			#保存输入值

        out = self.conv1(x)		#1x1卷积层
        out = self.bn1(out)
        out = self.relu(out)	#激活函数

        out = self.conv2(out)	#3x3卷积层
        out = self.bn2(out)
        out = self.relu(out)

        out = self.conv3(out)	#1x1卷积层
        out = self.bn3(out)

        if self.downsample is not None:
            identity = self.downsample(x)	#维度或通道数不匹配时，调整后再做后续运算

        out += identity			#加上输入值再作为输出结果
        out = self.relu(out)	#激活函数

        return out

4、resnet类

resnet类也继承pytorch中的基类nn.Module，重写了方法init和forward，在init方法中定义好网络的各个层次，在forward方法中定义前向传播过程。

class ResNet(nn.Module):
    def __init__(
        self,
        block: Type[Union[BasicBlock, Bottleneck]],	#表明使用的是哪种类型的残差块，是BasicBlock或者Bottleneck
        layers: List[int],			#每个残差块的个数，比如resnet34， layers=[3,4,6,3]，表示第一层有3个残差块，第二层有4个残差块，第三层有6个残差块，第四层有3个残差块。
    ) -> None:
        super(ResNet, self).__init__()
        ......
        self.conv1 = nn.Conv2d(3, self.inplanes, kernel_size=7, stride=2, padding=3,
                               bias=False)		#第一层卷积网络
        self.bn1 = norm_layer(self.inplanes)	#BN层
        self.relu = nn.ReLU(inplace=True)		#激活函数
        self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)		#最大池化
        self.layer1 = self._make_layer(block, 64, layers[0])	#第一个参数表示采用的block类型（basicblock或者bottleneck），按照通道维度数目（第二个参数）的不同分为四大类不同的层次结构（layer1-layer4），每一层内都有不同的block数目，具体数量由layers[0]-layers[3]（第三个参数）表示。
        self.layer2 = self._make_layer(block, 128, layers[1], stride=2,
                                       dilate=replace_stride_with_dilation[0])
        self.layer3 = self._make_layer(block, 256, layers[2], stride=2,
                                       dilate=replace_stride_with_dilation[1])
        self.layer4 = self._make_layer(block, 512, layers[3], stride=2,
                                       dilate=replace_stride_with_dilation[2])
        self.avgpool = nn.AdaptiveAvgPool2d((1, 1))			#平均池化下采样
        self.fc = nn.Linear(512 * block.expansion, num_classes)
		......
		......
        if stride != 1 or self.inplanes != planes * block.expansion:
            downsample = nn.Sequential(			#当维度不匹配时，调用它实现维度匹配
                conv1x1(self.inplanes, planes * block.expansion, stride),
                norm_layer(planes * block.expansion),
            )

        layers = []
        layers.append(block(self.inplanes, planes, stride, downsample, self.groups,
                            self.base_width, previous_dilation, norm_layer))	#每一类self.layer的第一层需要downsample，以保证运算的通道数相匹配，所以单独列出来
        self.inplanes = planes * block.expansion
        for _ in range(1, blocks):		#接下来将blocks中剩下的残差结构保存在layers列表中
            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:	#前向传播函数
        # See note [TorchScript super()]
        x = self.conv1(x)	#7x7的卷积核，步长为2
        x = self.bn1(x)		
        x = self.relu(x)	#激活函数
        x = self.maxpool(x)	#最大池化

        x = self.layer1(x)	#接下来是四个卷积块
        x = self.layer2(x)
        x = self.layer3(x)
        x = self.layer4(x)

        x = self.avgpool(x)	#平均池化
        x = torch.flatten(x, 1)
        x = self.fc(x)		#全连接层

        return x

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

5、调用resnet类

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:
        state_dict = load_state_dict_from_url(model_urls[arch],
                                              progress=progress)
        model.load_state_dict(state_dict)
    return model

_resnet('resnet18', BasicBlock, [2, 2, 2, 2], pretrained, progress,**kwargs)
_resnet('resnet34', BasicBlock, [3, 4, 6, 3], pretrained, progress,**kwargs)
_resnet('resnet50', Bottleneck, [3, 4, 6, 3], pretrained, progress,**kwargs)
_resnet('resnet101', Bottleneck, [3, 4, 23, 3], pretrained, progress,**kwargs)
_resnet('resnet152', Bottleneck, [3, 8, 36, 3], pretrained, progress,**kwargs)

调用resnet类时，第一个参数对应不同层数的网络结构，第二个参数对应残差网络的类型，可以看出，resnet18和resnet34网络使用的是BasicBlock，当网络层数加深时，就需要使用Bottleneck，方便网络模型的存储和计算。第三个参数里的四个元素分别表示不同构建块中的块数，其中resnet34和resnet50的数目完全一样，但二者使用残差网络的类型不同。

参考链接：
https://www.jianshu.com/p/085f4c8256f1
https://www.cnblogs.com/wzyuan/p/9880342.html
https://blog.csdn.net/xz1308579340/article/details/86527971