PyTorch：3-基础实战

注：所有资料来源且归属于thorough-pytorch(https://datawhalechina.github.io/thorough-pytorch/)，下文仅为学习记录

3.1：ResNet基本介绍

退化现象（degradation）：增加网络层数的过程中，随着训练准确率逐渐饱和，继续增加层数，训练准确率出现下降的现象。且这种下降不是过拟合。

快捷连接（shortcut connection）：将输入直接连接到后面的层，一定程度缓解了梯度消失和梯度爆炸，消除深度过大导致神经网络训练困难的问题。

梯度消失和梯度爆炸的根源：DNN结构，和，反向传播算法

梯度爆炸：网络层之间的梯度（值大于 1.0）重复相乘导致的指数级增长

梯度消失：网络层之间的梯度（值小于 1.0）重复相乘导致的指数级变小

3.2：torchvision的源代码

卷积核封装

封装3x3和1x1卷积核

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,			# 输入通道数
        out_planes,			# 输出通道数
        kernel_size=3,		# 卷积核尺寸
        stride=stride,		# 步长
        padding=dilation,	# 填充
        groups=groups,		# 分组
        bias=False,			# 偏移量
        dilation=dilation,	# 空洞卷积的中间间隔
    )


def conv1x1(in_planes: int, out_planes: int, stride: int = 1) -> nn.Conv2d:
    """1x1 convolution"""
    return nn.Conv2d(in_planes, out_planes, kernel_size=1, stride=stride, bias=False)
	# 解释同上

基本模块设计

ResNet常见的大小有ResNet-18，ResNet-34，ResNet-50、ResNet-101和ResNet-152，其中网络后面的数字代表的是网络的层数。

两个基本模块：BasicBlock和BottleNeck

两个block类输入一个通道为in_planes维的度特征图，输出一个planes*block.expansion维的特征图，其中planes的数目大小等于in_planes。

支路上的downsample操作：对shortcut支路进行大小或维度上的调整。

shortcut connection

【1】同等维度的映射：输入输出直接相加
$$F(x)+x$$
【2】不同维度的映射：给x补充一个线性映射来匹配维度（通常是1x1卷积）

basic block

BasicBlock模块用来构建resnet18和resnet34

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().__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)
        self.bn1 = norm_layer(planes)
        self.relu = nn.ReLU(inplace=True)
        self.conv2 = conv3x3(planes, planes)
        self.bn2 = norm_layer(planes)
        self.downsample = downsample
        self.stride = stride

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

        out = self.conv1(x)  	# 对x做卷积 
        out = self.bn1(out)  	# 对x归一化 
        out = self.relu(out)  	# 对x用激活函数

        out = self.conv2(out)  	# 对x做卷积
        out = self.bn2(out)  	# 归一化

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

        out += identity  		# 进行downsample
        out = self.relu(out)

        return out

bottle neck

BottleNeck模块用来构建resnet50，resnet101和resnet152

class Bottleneck(nn.Module):
    expansion: int = 4  # 对输出通道进行倍增

    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().__init__()
        if norm_layer is None:
            norm_layer = nn.BatchNorm2d
        width = int(planes * (base_width / 64.0)) * groups
        # Both self.conv2 and self.downsample layers downsample the input when stride != 1
        self.conv1 = conv1x1(inplanes, width)
        self.bn1 = norm_layer(width)
        self.conv2 = conv3x3(width, width, stride, groups, dilation)
        self.bn2 = norm_layer(width)
        self.conv3 = conv1x1(width, planes * self.expansion)
        self.bn3 = norm_layer(planes * self.expansion)
        self.relu = nn.ReLU(inplace=True)
        self.downsample = downsample
        self.stride = stride

    # Bottleneckd forward函数和BasicBlock类似
    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

网络整体结构

class ResNet(nn.Module):
    def __init__(
        self,
        block: Type[Union[BasicBlock, Bottleneck]], # 选择基本模块
        layers: List[int],							# 每一层block的数目构成 -> [3,4,6,3]
        num_classes: int = 1000, 					# 分类数目
        zero_init_residual: bool = False, 			# 初始化
        
        #######其他卷积构成，与本文ResNet无关######
        groups: int = 1,
        width_per_group: int = 64,
        replace_stride_with_dilation: Optional[List[bool]] = None,
        #########################################
        
        norm_layer: Optional[Callable[..., nn.Module]] = None, # norm层
    ) -> None:
        super().__init__()
        _log_api_usage_once(self)
        if norm_layer is None:
            norm_layer = nn.BatchNorm2d
        self._norm_layer = norm_layer
		
        self.inplanes = 64 # 输入通道
        
        #######其他卷积构成，与本文ResNet无关######
        self.dilation = 1 # 空洞卷积
        if replace_stride_with_dilation is None:
            # 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 "
                f"or a 3-element tuple, got {replace_stride_with_dilation}"
            )
        self.groups = groups
        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)
        # 通过_make_layer带到层次化设计的效果
        self.layer1 = self._make_layer(block, 64, layers[0])  # 对应着conv2_x
        self.layer2 = self._make_layer(block, 128, layers[1], stride=2, dilate=replace_stride_with_dilation[0])  # 对应着conv3_x
        self.layer3 = self._make_layer(block, 256, layers[2], stride=2, dilate=replace_stride_with_dilation[1])  # 对应着conv4_x
        self.layer4 = self._make_layer(block, 512, layers[3], stride=2, dilate=replace_stride_with_dilation[2])  # 对应着conv5_x
        # 分类头
        self.avgpool = nn.AdaptiveAvgPool2d((1, 1))
        self.fc = nn.Linear(512 * block.expansion, num_classes)
		
        # 模型初始化
        for m in self.modules():
            if isinstance(m, nn.Conv2d):
                nn.init.kaiming_normal_(m.weight, mode="fan_out", nonlinearity="relu")
            elif isinstance(m, (nn.BatchNorm2d, nn.GroupNorm)):
                nn.init.constant_(m.weight, 1)
                nn.init.constant_(m.bias, 0)

         
        if zero_init_residual:
            for m in self.modules():
                if isinstance(m, Bottleneck) and m.bn3.weight is not None:
                    nn.init.constant_(m.bn3.weight, 0)  # type: ignore[arg-type]
                elif isinstance(m, BasicBlock) and m.bn2.weight is not None:
                    nn.init.constant_(m.bn2.weight, 0)  # type: ignore[arg-type]
	# 层次化设计
    def _make_layer(
        self,
        block: Type[Union[BasicBlock, Bottleneck]], # 基本构成模块选择
        planes: int,  								# 输入的通道
        blocks: int, 								# 模块数目
        stride: int = 1, 							# 步长
        dilate: bool = False, 						# 空洞卷积，与本文无关
    ) -> nn.Sequential:
        norm_layer = self._norm_layer
        downsample = None 							# 是否采用下采样
        
        ####################无关#####################
        previous_dilation = self.dilation 
        if dilate:
            self.dilation *= stride
            stride = 1
        #############################################
        
        if stride != 1 or self.inplanes != planes * block.expansion:
            downsample = nn.Sequential(
                conv1x1(self.inplanes, planes * block.expansion, stride),
                norm_layer(planes * block.expansion),
            )
		
        # 使用layers存储每个layer
        layers = []
        layers.append(
            block(
                self.inplanes, planes, stride, downsample, self.groups, self.base_width, previous_dilation, norm_layer
            )
        )
        self.inplanes = planes * block.expansion
        for _ in range(1, blocks):
            layers.append(
                block(
                    self.inplanes,
                    planes,
                    groups=self.groups,
                    base_width=self.base_width,
                    dilation=self.dilation,
                    norm_layer=norm_layer,
                )
            )
		# 将layers通过nn.Sequential转化为网络
        return nn.Sequential(*layers)

    def _forward_impl(self, x: Tensor) -> Tensor:
        # See note [TorchScript super()]
        x = self.conv1(x)  		# conv1   x shape [1 64 112 112]
        x = self.bn1(x)   		# 归一化处理   
        x = self.relu(x)  		# 激活函数
        x = self.maxpool(x)  	# conv2_x的3x3 maxpool, x shape [1 64 56 56]

        x = self.layer1(x) # layer 1
        x = self.layer2(x) # layer 2
        x = self.layer3(x) # layer 3
        x = self.layer4(x) # layer 4

        x = self.avgpool(x) 	# 自适应池化
        x = torch.flatten(x, 1) 
        x = self.fc(x) 			# 分类

        return x

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

模型步骤

【1】首先是一个7 x 7的卷积作用在输入的3维图片上，并输入一个64维的特征图（即self.inplanes的初始值），通过BatchNorm层，ReLU层，MaxPool层。

【2】然后经过_make_layer()函数构建的4层layer。

【3】最后经过一个AveragePooling层，再经过一个fc层得到分类输出。

【4】在网络搭建起来后，还对模型的参数(Conv2d、BatchNorm2d、last BN)进行了初始化。

一个_make_layer()构建一个layer层，每一个layer层是两种基本模块的堆叠。

输入参数中block代表该layer堆叠模块的类型，可选BasicBlock或者BottleNeck。blocks代表该layer中堆叠的block的数目；planes与该layer最终输出的维度数有关，注意最终输出的维度数为planes * block.expansion。

变体

【1】Wider ResNet。

【2】DarkNet53。只是使用到了残差连接从而复用特征。

【3】ResNeXt。提出了一种介于普通卷积核深度可分离卷积的这种策略：分组卷积。通过控制分组的数量（基数）来达到两种策略的平衡。分组卷积的思想源自Inception，ResNeXt的每个分支的拓扑结构是相同的。

3.3：模型保存

【1】确定保存路径

【2】调用save函数

save_path = "./FahionModel.pkl"
torch.save(model, save_path)