PyTorch 搭建 ResNet 模型对 CIFAR10 数据集分类总结

发表于 2022-09-18 更新于 2023-11-24 分类于技术，学习阅读次数：本文字数： 11k 阅读时长 ≈ 10 分钟

本次试验将以 [Deep Residual Learning for Image Recognition] 作为主要参考来源，使用 PyTorch 搭建 ResNet 模型实现对 CIFAR10 数据集的训练和测试

ResNet 模型搭建

结构分析

通用框架

根据 Deep Residual Learning for Image Recognition 论文中的信息，可以得到常规 ResNet 模型的通用框架如下

layer_name	out_size	(18/34 layers) out_channel	(50/101/152 layers) out_channel	kernel_size	stride	padding
Input	224*224	3	3	None	None	None
Conv1	112*112	64	64	7	2	3
Maxpool	56*56	64	64	3	2	1
Conv2_x	56*56	64	64*4=256	-	-	-
Conv3_x	28*28	128	128*4=512	-	-	-
Conv4_x	14*14	256	256*4=1024	-	-	-
Conv5_x	7*7	512	512*4=2048	-	-	-
Avgpool	1*1	512	2048	None	None	None
Flatten	2048	1	1	None	None	None
FC	1000	1	1	None	None	None

其中 Conv2_x 、 Conv3_x 、Conv4_x 、 Conv5_x 层可由 BasicBlock 和 Bottleneck 两种基本模型组合而成

layer_name	ResNet18	ResNet34	ResNet50	ResNet101	ResNet152
Conv2_x	BasicBlock*2	BasicBlock*3	Bottleneck*3	Bottleneck*3	Bottleneck*3
Conv3_x	BasicBlock*2	BasicBlock*4	Bottleneck*4	Bottleneck*4	Bottleneck*8
Conv4_x	BasicBlock*2	BasicBlock*6	Bottleneck*6	Bottleneck*23	Bottleneck*36
Conv5_x	BasicBlock*2	BasicBlock*3	Bottleneck*3	Bottleneck*3	Bottleneck*3

基础结构

ResNet 18/34 使用的 BasicBlock 结构如下

layer_name	in_size	out_size	out_channel	kernel_size	stride	padding
Conv1	x*x	(x/stride)*(x/stride)	out_channel	3	stride	1
Conv2	x'*x'	x'*x'	out_channel	3	1	1
identity

ResNet 50/101/152 使用的 Bottleneck 结构如下

layer_name	in_size	out_size	out_channel	kernel_size	stride	padding
Conv1	x*x	(x/stride)*(x/stride)	out_channel	1	1	0
Conv2	x'*x'	x'*x'	out_channel	3	stride	1
Conv3	x'*x'	x'*x'	out_channel	1	1	0
identity

stride 和 identity

当基础结构是 Conv3_x 、Conv4_x 、 Conv5_x 的第一层时， stride=2 且 identity 为下采样后的输入

nn.Sequential(
    nn.Conv2d(self.in_channel, out_channel * block.expansion, kernel_size=1, stride=stride),
    nn.BatchNorm2d(out_channel * block.expansion),
)

当基础结构是 Conv3_x 、Conv4_x 、 Conv5_x 的其他层或在 Conv2_x 层时， stride=1 且 identity 为输入本身

1	nn.Sequential()

网络实现

BasicBlock

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

    def __init__(
        self,
        in_channel: int,
        out_channel: int,
        stride: int = 1,
        downsample: Optional[nn.Module] = None,
    ) -> None:
        super().__init__()
        # transforming (batch_size * x * x * input_channel) to (batch_size * x * x * output_channel)
        #                                                   or (batch_size * x/2 * x/2 * output_channel)
        # floor(((x - 3 + 2 * 1) / stride) + 1) => floor(x) stride = 1
        #                                       => floor(x/2) stride = 2
        self.conv1 = nn.Conv2d(in_channel, out_channel, kernel_size=3, stride=stride, padding=1, bias=False)
        self.bn1 = nn.BatchNorm2d(out_channel)
        # transforming (batch_size * x' * x' * output_channel) to (batch_size * x' * x' * output_channel)
        # floor(((x' - 3 + 2 * 1) / 1) + 1) => floor(x')
        self.conv2 = nn.Conv2d(out_channel, out_channel, kernel_size=3, stride=1, padding=1, bias=False)
        self.bn2 = nn.BatchNorm2d(out_channel)
        self.relu = nn.ReLU(inplace=True)
        self.downsample = downsample
        self.stride = stride

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

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

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

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

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

        return out

Bottleneck

class Bottleneck(nn.Module):
    expansion: int = 4

    def __init__(
        self,
        in_channel: int,
        out_channel: int,
        stride: int = 1,
        downsample: Optional[nn.Module] = None,
    ) -> None:
        super().__init__()

        # transforming (batch_size * x * x * output_channel) to (batch_size * x * x * output_channel)
        # floor(((x - 3 + 2 * 1) / 1) + 1) => floor(x)
        self.conv1 = nn.Conv2d(in_channel, out_channel, kernel_size=1, bias=False)
        self.bn1 = nn.BatchNorm2d(out_channel)
        # transforming (batch_size * x * x * output_channel) to (batch_size * x * x * output_channel)
        #                                                    or (batch_size * x/2 * x/2 * output_channel)
        # floor(((x - 3 + 2 * 1) / stride) + 1) => floor(x) stride = 1
        #                                       => floor(x/2) stride = 2
        self.conv2 = nn.Conv2d(out_channel, out_channel, kernel_size=3, stride=stride, padding=1, bias=False)
        self.bn2 = nn.BatchNorm2d(out_channel)
        # transforming (batch_size * x' * x' * output_channel) to (batch_size * x' * x' * (output_channel* expansion))
        # floor(((x' - 3 + 2 * 1) / 1) + 1) => floor(x')
        self.conv3 = nn.Conv2d(out_channel, out_channel * self.expansion, kernel_size=1, bias=False)
        self.bn3 = nn.BatchNorm2d(out_channel * self.expansion)
        self.relu = nn.ReLU(inplace=True)
        self.downsample = downsample
        self.stride = stride

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

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

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

        return out

通用框架

class ResNet(nn.Module):
    def __init__(self, block: Type[Union[BasicBlock, Bottleneck]], num_block: List[int], num_classes: int = 1000) -> None:
        super().__init__()
        self.in_channel = 64
        # transforming (batch_size * 224 * 224 * input_channel) to (batch_size * 112 * 112 * 64)
        # floor(((224 - 7 + 2 * 3) / 2) + 1) => floor(112.5) => floor(112)
        self.conv1 = nn.Sequential(
            nn.Conv2d(3, self.in_channel, kernel_size=7, stride=2, padding=3, bias=False),
            nn.BatchNorm2d(self.in_channel),
            nn.ReLU(inplace=True),
        )
        # transforming (batch_size * 112 * 112 * 64) to (batch_size * 56 * 56 * 64)
        # floor(((112 - 3 + 2 * 1) / 2) + 1) => floor(56.5) => floor(56)
        self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
        # transforming (batch_size * 56 * 56 * 64) to (batch_size * 56 * 56 * (64 * block.expansion))
        self.conv2_x = self._make_layer(block, 64, num_block[0], stride=1)
        # transforming (batch_size * 56 * 56 * (64 * block.expansion)) to (batch_size * 28 * 28 * (128 * block.expansion))
        self.conv3_x = self._make_layer(block, 128, num_block[1], stride=2)
        # transforming (batch_size * 28 * 28 * (128 * block.expansion)) to (batch_size * 14 * 14 * (256 * block.expansion))
        self.conv4_x = self._make_layer(block, 256, num_block[2], stride=2)
        # transforming (batch_size * 14 * 14 * (256 * block.expansion)) to (batch_size * 7 * 7 * (512 * block.expansion))
        self.conv5_x = self._make_layer(block, 512, num_block[3], stride=2)
        # transforming (batch_size * 7 * 7 * (512 * block.expansion)) to (batch_size * 1 * 1 * (512 * block.expansion))
        self.avgpool = nn.AdaptiveAvgPool2d((1, 1))
        # transforming (batch_size * 2048) to (batch_size * num_classes)
        self.fc = nn.Linear(512 * block.expansion, num_classes)

    def _make_layer(
        self, block: Type[Union[BasicBlock, Bottleneck]], out_channel: int, num_blocks: int, stride: int = 1
    ) -> nn.Sequential:
        downsample = None
        if stride != 1:
            downsample = nn.Sequential(
                nn.Conv2d(self.in_channel, out_channel * block.expansion, kernel_size=1, stride=stride),
                nn.BatchNorm2d(out_channel * block.expansion),
            )
        layers = []
        layers.append(block(self.in_channel, out_channel, stride, downsample))
        self.in_channel = out_channel * block.expansion
        for _ in range(1, num_blocks):
            layers.append(block(self.in_channel, out_channel))
        return nn.Sequential(*layers)

    def forward(self, x: Tensor) -> Tensor:
        x = self.conv1(x)
        x = self.maxpool(x)
        x = self.conv2_x(x)
        x = self.conv3_x(x)
        x = self.conv4_x(x)
        x = self.conv5_x(x)
        x = self.avgpool(x)
        x = torch.flatten(x, 1)
        x = self.fc(x)

        return x

构造网络

def ResNet18() -> ResNet:
    return ResNet(BasicBlock, [2, 2, 2, 2])


def ResNet34() -> ResNet:
    return ResNet(BasicBlock, [3, 4, 6, 3])


def ResNet50() -> ResNet:
    return ResNet(Bottleneck, [3, 4, 6, 3])


def ResNet101() -> ResNet:
    return ResNet(Bottleneck, [3, 4, 23, 3])


def ResNet152() -> ResNet:
    return ResNet(Bottleneck, [3, 8, 36, 3])

参考网页

PyTorch - SOURCE CODE FOR TORCHVISION.MODELS.RESNET

明素 - ResNet详解

回顾

`nn.Conv2d()` 函数 `bias` 参数的设置

当 nn.Conv2d() 后接 nn.BatchNorm2d() 时，可以把 bias 参数设置为 False

因为在 BN 层中，输入是否存在偏置不影响输出结果

不添加偏置还可以减少显卡内存的占用

参考网页

7s记忆的鱼 - 【pytorch】Conv2d()里面的参数bias什么时候加，什么时候不加？

`nn.AdaptiveAvgPool2d` 函数

参考网页

`*参数` 的作用

*参数 可以解压参数

a = (0,1,2,3,4,5,6,7,8,9)
b = [0,1,2,3,4,5,6,7,8,9]
print(*a)
print(*b)

将 List 和 Tuple 中的元素逐一解压出来

1 2	0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9

参考网页

TEDxPY - Python *args 用法笔记

`pip install` 默认安装在 `base` 环境

使用 pip install 时改用如下指令即可安装到当前虚拟环境中

1	python -m pip install **

参考网页

timertimer - 在conda虚拟环境中用pip安装包总是在base环境中的解决办法

CIFAR10 特化模型

layer_name	out_size	out_channel	kernel_size	stride	padding
Input	32*32	3	None	None	None
Conv1	32*32	16	1	1	31
Conv2_x	32*32	16	64*4=256	-	-	-
Conv3_x	16*16	32	128*4=512	-	-	-
Conv4_x	8*8	64	256*4=1024	-	-	-
Avgpool	1*1	64	None	None	None
Flatten	64	1	None	None	None
FC	10	1	None	None	None

其中 Conv2_x 、 Conv3_x 、Conv4_x 层由 Block 组成

layer_name	ResNet_CIFAR10
Conv2_x	Block*n
Conv3_x	Block*n
Conv4_x	Block*n

Block 结构如下

layer_name	in_size	out_size	out_channel	kernel_size	stride	padding
Conv1	x*x	(x/stride)*(x/stride)	out_channel	3	stride	1
Conv2	x'*x'	x'*x'	out_channel	3	1	1
identity

当基础结构是 Conv3_x 、Conv4_x 、 Conv5_x 的第一层时， stride=2 且 identity 为下采样后的输入

nn.Sequential(
    nn.Conv2d(self.in_channel, out_channel * block.expansion, kernel_size=1, stride=stride),
    nn.BatchNorm2d(out_channel * block.expansion),
)

当基础结构是 Conv3_x 、Conv4_x 、 Conv5_x 的其他层或在 Conv2_x 层时， stride=1 且 identity 为输入本身

1	nn.Sequential()

模型实现

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

    def __init__(
        self,
        in_channel: int,
        out_channel: int,
        stride: int = 1,
        downsample: Optional[nn.Module] = None,
    ) -> None:
        super().__init__()
        # transforming (batch_size * x * x * input_channel) to (batch_size * x * x * output_channel)
        #                                                   or (batch_size * x/2 * x/2 * output_channel)
        # floor(((x - 3 + 2 * 1) / stride) + 1) => floor(x) stride = 1
        #                                       => floor(x/2) stride = 2
        self.conv1 = nn.Conv2d(in_channel, out_channel, kernel_size=3, stride=stride, padding=1, bias=False)
        self.bn1 = nn.BatchNorm2d(out_channel)
        # transforming (batch_size * x' * x' * output_channel) to (batch_size * x' * x' * output_channel)
        # floor(((x' - 3 + 2 * 1) / 1) + 1) => floor(x')
        self.conv2 = nn.Conv2d(out_channel, out_channel, kernel_size=3, stride=1, padding=1, bias=False)
        self.bn2 = nn.BatchNorm2d(out_channel)
        self.relu = nn.ReLU(inplace=True)
        self.downsample = downsample
        self.stride = stride

    def forward(self, x: Tensor) -> Tensor:
        if self.downsample is not None:
            identity = self.downsample(x)
        else:
            identity = x

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

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

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

        return out

class ResNet(nn.Module):
    def __init__(self, block: Block, num_block: List[int], num_classes: int = 10) -> None:
        super().__init__()
        self.in_channel = 16
        # transforming (batch_size * 32 * 32 * input_channel) to (batch_size * 32 * 32 * 16)
        # floor(((32 - 3 + 2 * 1) / 1) + 1) => floor(112.5) => floor(112)
        self.conv1 = nn.Sequential(
            nn.Conv2d(3, self.in_channel, kernel_size=3, padding=1, bias=False),
            nn.BatchNorm2d(self.in_channel),
            nn.ReLU(inplace=True),
        )
        # transforming (batch_size * 32 * 32 * 16) to (batch_size * 32 * 32 * 16)
        self.conv2_x = self._make_layer(block, 16, num_block, stride=1)
        # transforming (batch_size * 32 * 32 * 16) to (batch_size * 16 * 16 * 32)
        self.conv3_x = self._make_layer(block, 32, num_block, stride=2)
        # transforming (batch_size * 16 * 16 * 16) to (batch_size * 8 * 8 * 64)
        self.conv4_x = self._make_layer(block, 64, num_block, stride=2)
        # transforming (batch_size * 8 * 8 * 64) to (batch_size * 1 * 1 * 64)
        self.avgpool = nn.AdaptiveAvgPool2d((1, 1))
        # transforming (batch_size * 64) to (batch_size * num_classes)
        self.fc = nn.Linear(64 * block.expansion, num_classes)

        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)

    def _make_layer(self, block: Block, out_channel: int, num_blocks: int, stride: int = 1) -> nn.Sequential:
        downsample = None
        if stride != 1:
            downsample = nn.Sequential(
                nn.Conv2d(self.in_channel, out_channel * block.expansion, kernel_size=1, stride=stride),
                nn.BatchNorm2d(out_channel * block.expansion),
            )
        layers = []
        layers.append(block(self.in_channel, out_channel, stride, downsample))
        self.in_channel = out_channel * block.expansion
        for _ in range(1, num_blocks):
            layers.append(block(self.in_channel, out_channel))
        return nn.Sequential(*layers)

    def forward(self, x: Tensor) -> Tensor:
        x = self.conv1(x)
        x = self.conv2_x(x)
        x = self.conv3_x(x)
        x = self.conv4_x(x)
        x = self.avgpool(x)
        x = torch.flatten(x, 1)
        x = self.fc(x)

        return x

def ResNet20() -> ResNet:
    return ResNet(Block, 3)


def ResNet32() -> ResNet:
    return ResNet(Block, 5)


def ResNet44() -> ResNet:
    return ResNet(Block, 7)


def ResNet56() -> ResNet:
    return ResNet(Block, 9)


def ResNet110() -> ResNet:
    return ResNet(Block, 18)


def ResNet1202() -> ResNet:
    return ResNet(Block, 200)

模型训练

模型训练内容与 AlexNet_CIFAR10 项目相似，相同之处不再赘述

封装自定义 Python 库

此次实验中将 AlexNet_CIFAR10 项目中计算数据集均值和方差封装在 utils 文件夹下

需要在 utils 文件夹下生成空的 __init__.py 文件，声明 utils 文件夹为封装好的 Python 库

ResNet 模型搭建

结构分析

通用框架

基础结构

stride 和 identity

网络实现

BasicBlock

Bottleneck

通用框架

构造网络

参考网页

回顾

nn.Conv2d() 函数 bias 参数的设置

参考网页

nn.AdaptiveAvgPool2d 函数

参考网页

*参数 的作用

参考网页

pip install 默认安装在 base 环境

参考网页

CIFAR10 特化模型

模型实现

模型训练

封装自定义 Python 库

`nn.Conv2d()` 函数 `bias` 参数的设置

`nn.AdaptiveAvgPool2d` 函数

`*参数` 的作用

`pip install` 默认安装在 `base` 环境