ResNet代码复现＋超详细注释（PyTorch）

关于ResNet的原理和具体细节，可参见上篇解读：经典神经网络论文超详细解读（五）——ResNet（残差网络）学习笔记（翻译＋精读＋代码复现）

接下来我们就来复现一下代码。

源代码比较复杂，感兴趣的同学可以上官网学习：

https://github.com/pytorch/vision/tree/master/torchvision

本篇是简化版本

一、BasicBlock模块

BasicBlock结构图如图所示：

BasicBlock是基础版本，主要用来构建ResNet18和ResNet34网络，里面只包含两个卷积层，使用了两个 3*3 的卷积，通道数都是64，卷积后接着 BN 和 ReLU。

右边的曲线就是Shortcut Connections，将输入x加到输出。

代码：

'''-------------一、BasicBlock模块-----------------------------'''
# 用于ResNet18和ResNet34基本残差结构块
class BasicBlock(nn.Module):
    def __init__(self, inchannel, outchannel, stride=1):
        super(BasicBlock, self).__init__()
        self.left = nn.Sequential(
            nn.Conv2d(inchannel, outchannel, kernel_size=3, stride=stride, padding=1, bias=False),
            nn.BatchNorm2d(outchannel),
            nn.ReLU(inplace=True), #inplace=True表示进行原地操作，一般默认为False，表示新建一个变量存储操作
            nn.Conv2d(outchannel, outchannel, kernel_size=3, stride=1, padding=1, bias=False),
            nn.BatchNorm2d(outchannel)
        )
        self.shortcut = nn.Sequential()
        #论文中模型架构的虚线部分，需要下采样
        if stride != 1 or inchannel != outchannel:
            self.shortcut = nn.Sequential(
                nn.Conv2d(inchannel, outchannel, kernel_size=1, stride=stride, bias=False),
                nn.BatchNorm2d(outchannel)
            )

    def forward(self, x):
        out = self.left(x) #这是由于残差块需要保留原始输入
        out += self.shortcut(x)#这是ResNet的核心，在输出上叠加了输入x
        out = F.relu(out)
        return out

二、Bottleneck 模块

Bottleneck结构图如图所示：

Bottleneck主要用在ResNet50及以上的网络结构，与BasicBlock不同的是这里有 3 个卷积，分别为 11，33，11大小的卷积核，分别用于压缩维度、卷积处理、恢复维度。
这里的通道数是变化的，11卷积层的作用就是用于改变特征图的通数，使得可以和恒等映射x相叠加，另外这里的1*1卷积层改变维度的很重要的一点是可以降低网络参数量，这也是为什么更深层的网络采用BottleNeck而不是BasicBlock的原因。

注意：这里outchannel / 4是因为Bottleneck层输出通道都是输入的4倍

代码：

'''-------------二、Bottleneck模块-----------------------------'''
# 用于ResNet50及以上的残差结构块
class Bottleneck(nn.Module):
    def __init__(self, inchannel, outchannel, stride=1):
        super(Bottleneck, self).__init__()
        self.left = nn.Sequential(
            nn.Conv2d(inchannel, int(outchannel / 4), kernel_size=1, stride=stride, padding=0, bias=False),
            nn.BatchNorm2d(int(outchannel / 4)),
            nn.ReLU(inplace=True),
            nn.Conv2d(int(outchannel / 4), int(outchannel / 4), kernel_size=3, stride=1, padding=1, bias=False),
            nn.BatchNorm2d(int(outchannel / 4)),
            nn.ReLU(inplace=True),
            nn.Conv2d(int(outchannel / 4), outchannel, kernel_size=1, stride=1, padding=0, bias=False),
            nn.BatchNorm2d(outchannel),
        )
        self.shortcut = nn.Sequential()
        if stride != 1 or inchannel != outchannel:
            self.shortcut = nn.Sequential(
                nn.Conv2d(inchannel, outchannel, kernel_size=1, stride=stride, bias=False),
                nn.BatchNorm2d(outchannel)
            )

    def forward(self, x):
        out = self.left(x)
        y = self.shortcut(x)
        out += self.shortcut(x)
        out = F.relu(out)
        return out

三、ResNet主体

介绍了上述BasicBlock基础块和BotteNeck结构后，我们就可以搭建ResNet结构了。

5种不同层数的ResNet结构图如图所示：

代码：

ResNet18

'''----------ResNet18----------'''
class ResNet_18(nn.Module):
    def __init__(self, ResidualBlock, num_classes=10):
        super(ResNet_18, self).__init__()
        self.inchannel = 64
        self.conv1 = nn.Sequential(
            nn.Conv2d(3, 64, kernel_size=3, stride=1, padding=1, bias=False),
            nn.BatchNorm2d(64),
            nn.ReLU(),
        )
        self.layer1 = self.make_layer(ResidualBlock, 64, 2, stride=1)
        self.layer2 = self.make_layer(ResidualBlock, 128, 2, stride=2)
        self.layer3 = self.make_layer(ResidualBlock, 256, 2, stride=2)
        self.layer4 = self.make_layer(ResidualBlock, 512, 2, stride=2)
        self.fc = nn.Linear(512, num_classes)

    def make_layer(self, block, channels, num_blocks, stride):
        strides = [stride] + [1] * (num_blocks - 1)  # strides=[1,1]
        layers = []
        for stride in strides:
            layers.append(block(self.inchannel, channels, stride))
            self.inchannel = channels
        return nn.Sequential(*layers)

    def forward(self, x):  # 3*32*32
        out = self.conv1(x)  # 64*32*32
        out = self.layer1(out)  # 64*32*32
        out = self.layer2(out)  # 128*16*16
        out = self.layer3(out)  # 256*8*8
        out = self.layer4(out)  # 512*4*4
        out = F.avg_pool2d(out, 4)  # 512*1*1
        out = out.view(out.size(0), -1)  # 512
        out = self.fc(out)
        return out

ResNet34

'''----------ResNet34----------'''
class ResNet_34(nn.Module):
    def __init__(self, ResidualBlock, num_classes=10):
        super(ResNet_34, self).__init__()
        self.inchannel = 64
        self.conv1 = nn.Sequential(
            nn.Conv2d(3, 64, kernel_size=3, stride=1, padding=1, bias=False),
            nn.BatchNorm2d(64),
            nn.ReLU(),
        )
        self.layer1 = self.make_layer(ResidualBlock, 64, 3, stride=1)
        self.layer2 = self.make_layer(ResidualBlock, 128, 4, stride=2)
        self.layer3 = self.make_layer(ResidualBlock, 256, 6, stride=2)
        self.layer4 = self.make_layer(ResidualBlock, 512, 3, stride=2)
        self.fc = nn.Linear(512, num_classes)

    def make_layer(self, block, channels, num_blocks, stride):
        strides = [stride] + [1] * (num_blocks - 1)  # strides=[1,1]
        layers = []
        for stride in strides:
            layers.append(block(self.inchannel, channels, stride))
            self.inchannel = channels
        return nn.Sequential(*layers)

    def forward(self, x):  # 3*32*32
        out = self.conv1(x)  # 64*32*32
        out = self.layer1(out)  # 64*32*32
        out = self.layer2(out)  # 128*16*16
        out = self.layer3(out)  # 256*8*8
        out = self.layer4(out)  # 512*4*4
        out = F.avg_pool2d(out, 4)  # 512*1*1
        out = out.view(out.size(0), -1)  # 512
        out = self.fc(out)
        return out

ResNet50

'''---------ResNet50--------'''
class ResNet_50(nn.Module):
    def __init__(self, ResidualBlock, num_classes=10):
        super(ResNet_50, self).__init__()
        self.inchannel = 64
        self.conv1 = nn.Sequential(
            nn.Conv2d(3, 64, kernel_size=3, stride=1, padding=1, bias=False),
            nn.BatchNorm2d(64),
            nn.ReLU(),
        )
        self.layer1 = self.make_layer(ResidualBlock, 256, 3, stride=1)
        self.layer2 = self.make_layer(ResidualBlock, 512, 4, stride=2)
        self.layer3 = self.make_layer(ResidualBlock, 1024, 6, stride=2)
        self.layer4 = self.make_layer(ResidualBlock, 2048, 3, stride=2)
        self.fc = nn.Linear(512 * 4, num_classes)
        # **************************

    def make_layer(self, block, channels, num_blocks, stride):
        strides = [stride] + [1] * (num_blocks - 1)  # strides=[1,1]
        layers = []
        for stride in strides:
            layers.append(block(self.inchannel, channels, stride))
            self.inchannel = channels
        return nn.Sequential(*layers)

    def forward(self, x):  # 3*32*32
        out = self.conv1(x)  # 64*32*32
        out = self.layer1(out)  # 64*32*32
        out = self.layer2(out)  # 128*16*16
        out = self.layer3(out)  # 256*8*8
        out = self.layer4(out)  # 512*4*4
        out = F.avg_pool2d(out, 4)  # 512*1*1
        # print(out.size())
        out = out.view(out.size(0), -1)  # 512
        out = self.fc(out)
        return out

四、完整代码

import torch
import torch.nn as nn
import torch.nn.functional as F

'''-------------一、BasicBlock模块-----------------------------'''
# 用于ResNet18和ResNet34基本残差结构块
class BasicBlock(nn.Module):
    def __init__(self, inchannel, outchannel, stride=1):
        super(BasicBlock, self).__init__()
        self.left = nn.Sequential(
            nn.Conv2d(inchannel, outchannel, kernel_size=3, stride=stride, padding=1, bias=False),
            nn.BatchNorm2d(outchannel),
            nn.ReLU(inplace=True), #inplace=True表示进行原地操作，一般默认为False，表示新建一个变量存储操作
            nn.Conv2d(outchannel, outchannel, kernel_size=3, stride=1, padding=1, bias=False),
            nn.BatchNorm2d(outchannel)
        )
        self.shortcut = nn.Sequential()
        #论文中模型架构的虚线部分，需要下采样
        if stride != 1 or inchannel != outchannel:
            self.shortcut = nn.Sequential(
                nn.Conv2d(inchannel, outchannel, kernel_size=1, stride=stride, bias=False),
                nn.BatchNorm2d(outchannel)
            )

    def forward(self, x):
        out = self.left(x) #这是由于残差块需要保留原始输入
        out += self.shortcut(x)#这是ResNet的核心，在输出上叠加了输入x
        out = F.relu(out)
        return out

'''-------------二、Bottleneck模块-----------------------------'''
# 用于ResNet50及以上的残差结构块
class Bottleneck(nn.Module):
    def __init__(self, inchannel, outchannel, stride=1):
        super(Bottleneck, self).__init__()
        self.left = nn.Sequential(
            nn.Conv2d(inchannel, int(outchannel / 4), kernel_size=1, stride=stride, padding=0, bias=False),
            nn.BatchNorm2d(int(outchannel / 4)),
            nn.ReLU(inplace=True),
            nn.Conv2d(int(outchannel / 4), int(outchannel / 4), kernel_size=3, stride=1, padding=1, bias=False),
            nn.BatchNorm2d(int(outchannel / 4)),
            nn.ReLU(inplace=True),
            nn.Conv2d(int(outchannel / 4), outchannel, kernel_size=1, stride=1, padding=0, bias=False),
            nn.BatchNorm2d(outchannel),
        )
        self.shortcut = nn.Sequential()
        if stride != 1 or inchannel != outchannel:
            self.shortcut = nn.Sequential(
                nn.Conv2d(inchannel, outchannel, kernel_size=1, stride=stride, bias=False),
                nn.BatchNorm2d(outchannel)
            )

    def forward(self, x):
        out = self.left(x)
        y = self.shortcut(x)
        out += self.shortcut(x)
        out = F.relu(out)
        return out

'''-------------ResNet18---------------'''
class ResNet_18(nn.Module):
    def __init__(self, ResidualBlock, num_classes=10):
        super(ResNet_18, self).__init__()
        self.inchannel = 64
        self.conv1 = nn.Sequential(
            nn.Conv2d(3, 64, kernel_size=3, stride=1, padding=1, bias=False),
            nn.BatchNorm2d(64),
            nn.ReLU(),
        )
        self.layer1 = self.make_layer(ResidualBlock, 64, 2, stride=1)
        self.layer2 = self.make_layer(ResidualBlock, 128, 2, stride=2)
        self.layer3 = self.make_layer(ResidualBlock, 256, 2, stride=2)
        self.layer4 = self.make_layer(ResidualBlock, 512, 2, stride=2)
        self.fc = nn.Linear(512, num_classes)

    def make_layer(self, block, channels, num_blocks, stride):
        strides = [stride] + [1] * (num_blocks - 1)  # strides=[1,1]
        layers = []
        for stride in strides:
            layers.append(block(self.inchannel, channels, stride))
            self.inchannel = channels
        return nn.Sequential(*layers)

    def forward(self, x):  # 3*32*32
        out = self.conv1(x)  # 64*32*32
        out = self.layer1(out)  # 64*32*32
        out = self.layer2(out)  # 128*16*16
        out = self.layer3(out)  # 256*8*8
        out = self.layer4(out)  # 512*4*4
        out = F.avg_pool2d(out, 4)  # 512*1*1
        out = out.view(out.size(0), -1)  # 512
        out = self.fc(out)
        return out

'''-------------ResNet34---------------'''
class ResNet_34(nn.Module):
    def __init__(self, ResidualBlock, num_classes=10):
        super(ResNet_34, self).__init__()
        self.inchannel = 64
        self.conv1 = nn.Sequential(
            nn.Conv2d(3, 64, kernel_size=3, stride=1, padding=1, bias=False),
            nn.BatchNorm2d(64),
            nn.ReLU(),
        )
        self.layer1 = self.make_layer(ResidualBlock, 64, 3, stride=1)
        self.layer2 = self.make_layer(ResidualBlock, 128, 4, stride=2)
        self.layer3 = self.make_layer(ResidualBlock, 256, 6, stride=2)
        self.layer4 = self.make_layer(ResidualBlock, 512, 3, stride=2)
        self.fc = nn.Linear(512, num_classes)

    def make_layer(self, block, channels, num_blocks, stride):
        strides = [stride] + [1] * (num_blocks - 1)  # strides=[1,1]
        layers = []
        for stride in strides:
            layers.append(block(self.inchannel, channels, stride))
            self.inchannel = channels
        return nn.Sequential(*layers)

    def forward(self, x):  # 3*32*32
        out = self.conv1(x)  # 64*32*32
        out = self.layer1(out)  # 64*32*32
        out = self.layer2(out)  # 128*16*16
        out = self.layer3(out)  # 256*8*8
        out = self.layer4(out)  # 512*4*4
        out = F.avg_pool2d(out, 4)  # 512*1*1
        out = out.view(out.size(0), -1)  # 512
        out = self.fc(out)
        return out

'''-------------ResNet50---------------'''
class ResNet_50(nn.Module):
    def __init__(self, ResidualBlock, num_classes=10):
        super(ResNet_50, self).__init__()
        self.inchannel = 64
        self.conv1 = nn.Sequential(
            nn.Conv2d(3, 64, kernel_size=3, stride=1, padding=1, bias=False),
            nn.BatchNorm2d(64),
            nn.ReLU(),
        )
        self.layer1 = self.make_layer(ResidualBlock, 256, 3, stride=1)
        self.layer2 = self.make_layer(ResidualBlock, 512, 4, stride=2)
        self.layer3 = self.make_layer(ResidualBlock, 1024, 6, stride=2)
        self.layer4 = self.make_layer(ResidualBlock, 2048, 3, stride=2)
        self.fc = nn.Linear(512 * 4, num_classes)
        # **************************

    def make_layer(self, block, channels, num_blocks, stride):
        strides = [stride] + [1] * (num_blocks - 1)  # strides=[1,1]
        layers = []
        for stride in strides:
            layers.append(block(self.inchannel, channels, stride))
            self.inchannel = channels
        return nn.Sequential(*layers)

    def forward(self, x):  # 3*32*32
        out = self.conv1(x)  # 64*32*32
        out = self.layer1(out)  # 64*32*32
        out = self.layer2(out)  # 128*16*16
        out = self.layer3(out)  # 256*8*8
        out = self.layer4(out)  # 512*4*4
        out = F.avg_pool2d(out, 4)  # 512*1*1
        # print(out.size())
        out = out.view(out.size(0), -1)  # 512
        out = self.fc(out)
        return out

def ResNet18():
    return ResNet_18(BasicBlock)

def ResNet34():
    return ResNet_34(BasicBlock)

def ResNet50():
    return ResNet_50(Bottleneck)

本篇到这里就结束啦，有什么问题，欢迎大家留言讨论~

标签： pytorch 深度学习人工智能

本文转载自: https://blog.csdn.net/weixin_43334693/article/details/128409032
版权归原作者 路人贾'ω' 所有，如有侵权，请联系我们删除。

ResNet代码复现＋超详细注释（PyTorch）

一、BasicBlock模块

代码：

二、Bottleneck 模块

代码：

三、ResNet主体

代码：

ResNet18

ResNet34

ResNet50

四、完整代码

发表评论

“ResNet代码复现＋超详细注释（PyTorch）”的评论:

关于作者

overfit同步小助手

相关阅读

文章导航