一.Resnet论文分析

1.研究动机

此前，人们认为通过堆叠网络结构增加网络深度可以提升网络的学习能力。然而一直增加网络深度到一定程度之后，模型的能力反而缩减了。

论文作者提出以下问题：Is learning better networks as easy as stacking more layers?

我们知道神经网络的深度太大可能会造成以下问题：
1.梯度消失和梯度爆炸
2.过拟合

针对梯度消失和梯度爆炸问题，我们使用normalized initialization和batch normalization方法可以解决。

通过上图我们也发现，56层的网络的training error和test error都比较大，所以这也不是过拟合问题。

现在假设一个浅层网络已经达到一个较好的效果，如果增加网络深度，但使新增的网络什么也不干，即identity mapping，理论上来说，较深的网络在性能上至少应该和
之前的浅层网络一致。然而现实却是，增加网络深度以后，性能反而下降了。作者认为是网络深度加深以后，导致了整个神经网络变得更加难以优化。

2.解决方案

由此，作者提出residual learning。把网络中的几层看成是一个整体，输入是X，输出为H(X)。原本的映射是从X到H(X)。作者提出，令F(X)=H(X)-X,令网络拟合F(X)，
再使用shortcut connection(短路连接),从而达到和映射H(X)一样的效果。

作者假设通过residual learning 可以使得深层网络的优化变得简单。最后实验的结果也表明确实如此。

3.网络结构

resnet具体结构：

(1).Block之前的操作

输入为3*224*224，进行一次kernel_size=7, stride=2, padding=3的卷积操作，以及一次kernel_size=3, stride=2, padding=1的池化操作，输出为64*56*56。

(2).两种不同的块

左边叫building block,右边叫bottleneck。building block用于组成resnet18/34, bottleneck用于组成resnet50/101/152

以resnet18为例：
第一层：

这种操作执行两次。

第二层：

需要注意stride在第一块中的第一个卷积操作中，stride=2，从而达到减小尺寸的目的。

后面的layers和第二层大致相同，不再赘述。

组成resnet50/101/152的bottleneck也与building block类似，不再赘述。

自己在实现网络时，需要注意的是每一步的padding和stride，从而保持尺寸一致。

最后附上pytorch实现代码：

#import package
import os
os.environ["CUDA_VISIBLE_DEVICES"] = "0,1,2,3,4,5,6"
import torch
import torchvision
import torch.nn as nn
from torchvision import transforms
import torch.nn.functional as F

#device configuration
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")

class building_block(nn.Module):
    def __init__(self, in_channels, out_channels, stride=1, downsample=None):
        super(building_block, self).__init__()
        self.conv1 = nn.Conv2d(in_channels, out_channels, 3, stride, padding=1, bias=False)
        self.bn1 = nn.BatchNorm2d(out_channels)
        self.relu = nn.ReLU(inplace=True)
        self.conv2 = nn.Conv2d(out_channels, out_channels, 3, stride, padding=1, bias=False)
        self.bn1 = nn.BatchNorm2d(out_channels)
        self.downsample = downsample
    
    def forward(self, input):
        residual = input
        x = self.conv1(input)
        x = self.bn1(x)
        x = self.relu(x)
        x = self.conv2(x)
        x = self.bn2(x)
        if self.downsample:
            residual = self.downsample(input)
        x = x + residual
        output = self.relu(x)
        return output

class bottle_neck(nn.Module):
    def __init__(self, in_channels, out_channels, stride=1, downsampling=False, expansion=4):
        super(bottle_neck, self).__init__()
        self.expansion = expansion
        self.downsampling = downsampling
        self.bottleneck = nn.Sequential(
            nn.Conv2d(in_channels=in_channels, out_channels=out_channels, kernel_size=1, stride=1, bias=False),
            nn.BatchNorm2d(out_channels),
            nn.ReLU(inplace=True),
            nn.Conv2d(in_channels=out_channels, out_channels=out_channels, kernel_size=3, stride=stride, padding=1, bias=False),
            nn.BatchNorm2d(out_channels),
            nn.ReLU(inplace=True),
            nn.Conv2d(in_channels=out_channels, out_channels=out_channels*self.expansion, kernel_size=1, stride=1, bias=False),
            nn.BatchNorm2d(out_channels*self.expansion),
        )
        if self.downsampling:
            self.downsample = nn.Sequential(
                nn.Conv2d(in_channels=in_channels, out_channels=out_channels*self.expansion, kernel_size=1, stride=stride, bias=False),
                nn.BatchNorm2d(out_channels*self.expansion)
            )
        self.relu = nn.ReLU(inplace=True)
        
    def forward(self, x):
        residual = x
        out = self.bottleneck(x)
        
        if self.downsampling:
            residual = self.downsample(x)
        out += residual
        out = self.relu(out)
        return out

def Conv1(in_channels, out_channels, stride = 2):
    return nn.Sequential(
        nn.Conv2d(in_channels=in_channels, out_channels=out_channels, kernel_size=7, stride=stride, padding=3, bias=False),
        nn.BatchNorm2d(out_channels),
        nn.ReLU(inplace=True),
        nn.MaxPool2d(kernel_size=3, stride=2, padding=1),
    )
    
    
class ResNet(nn.Module):
    def __init__(self, blocks, num_classes, expansion = 4):
        super(ResNet, self).__init__()
        self.expansion = expansion
        self.conv1 = Conv1(in_channels=3, out_channels=64)
        self.layer1 = self.make_layer(in_channels=64, out_channels=64, block=blocks[0], stride=1)
        self.layer2 = self.make_layer(in_channels=256, out_channels=128, block=blocks[1], stride=2)
        self.layer3 = self.make_layer(in_channels=512, out_channels=256, block=blocks[2], stride=2)
        self.layer4 = self.make_layer(in_channels=1024, out_channels=512, block=blocks[3], stride=2)
        self.avgpool = nn.AvgPool2d(7, stride=1)
        self.fc = nn.Linear(2048,num_classes)

    def make_layer(self, in_channels, out_channels, block, stride):
        layers = []
        layers.append(bottle_neck(in_channels, out_channels, stride, downsampling=True))
        for i in range(1, block):
            layers.append(bottle_neck(out_channels*self.expansion, out_channels))
        
        return nn.Sequential(*layers)
    
    def forward(self, x):
        x = self.conv1(x)
        x = self.layer1(x)
        x = self.layer2(x)
        x = self.layer3(x)
        x = self.layer4(x)
        x = self.avgpool(x)
        x = x.view(x.size(0), -1)
        x = self.fc(x)
        return x


def ResNet50():
    return ResNet([3, 4, 6, 3], num_classes = 7)

def ResNet101():
    return ResNet([3, 4, 23, 3], num_classes = 7)

def ResNet152():
    return ResNet([3, 8, 36, 3], num_classes = 7)
    



#model = torchvision.models.resnet50()
model = ResNet101()
print(model)

input = torch.randn(1, 3, 224, 224)
out = model(input)
print(out.shape)

Resnet网络结构及pytorch实现代码

一.Resnet论文分析

1.研究动机

2.解决方案

3.网络结构

推荐阅读