#------------------------------用50行程式碼搭建ResNet-------------------------------------------
from torch import nn
import torch as t
from torch.nn import functional as F

class ResidualBlock(nn.Module):
    #實現子module: Residual    Block
    def __init__(self,inchannel,outchannel,stride=1,shortcut=None):
        super(ResidualBlock,self).__init__()
        self.left=nn.Sequential(
            nn.Conv2d(inchannel,outchannel,3,stride,1,bias=False),
            nn.BatchNorm2d(outchannel),
            nn.ReLU(inplace=True),
            nn.Conv2d(outchannel,outchannel,3,1,1,bias=False),
            nn.BatchNorm2d(outchannel)
        )
        
        self.right=shortcut
        
    def forward(self,x):
        out=self.left(x)
        residual=x if self.right is None else self.right(x)
        out+=residual
        return F.relu(out)
    
    
class ResNet(nn.Module):
    #實現主module:ResNet34
    #ResNet34包含多個layer,每個layer又包含多個residual block
    #用子module實現residual block , 用 _make_layer 函式實現layer
    def __init__(self,num_classes=1000):
        super(ResNet,self).__init__()
        self.pre=nn.Sequential(
            nn.Conv2d(3,64,7,2,3,bias=False),
            nn.BatchNorm2d(64),
            nn.ReLU(inplace=True),
            nn.MaxPool2d(3,2,1)
        )
        #重複的layer,分別有3,4,6,3個residual block
        self.layer1=self._make_layer(64,64,3)
        self.layer2=self._make_layer(64,128,4,stride=2)
        self.layer3=self._make_layer(128,256,6,stride=2)
        self.layer4=self._make_layer(256,512,3,stride=2)
        
        #分類用的全連線
        self.fc=nn.Linear(512,num_classes)
        
    def _make_layer(self,inchannel,outchannel,block_num,stride=1):
        #構建layer,包含多個residual block
        shortcut=nn.Sequential(
            nn.Conv2d(inchannel,outchannel,1,stride,bias=False),
            nn.BatchNorm2d(outchannel))

        layers=[ ]
        layers.append(ResidualBlock(inchannel,outchannel,stride,shortcut))
        
        for i in range(1,block_num):
            layers.append(ResidualBlock(outchannel,outchannel))
        return nn.Sequential(*layers)
    
    def forward(self,x):
        x=self.pre(x)
        
        x=self.layer1(x)
        x=self.layer2(x)
        x=self.layer3(x)
        x=self.layer4(x)
        
        x=F.avg_pool2d(x,7)
        x=x.view(x.size(0),-1)
        return self.fc(x)

model=ResNet()
input=t.autograd.Variable(t.randn(1,3,224,224))
o=model(input)
print(o)

大致框架算是理解了，但是細節部分比如卷積層的輸入輸出的大小之類的，還需要仔細研究。

Pytorch學習系列（一）至（四）均摘自《深度學習框架PyTorch入門與實踐》陳雲

---------------------  
作者：尋找如意  
來源：CSDN  
原文：https://blog.csdn.net/qq_34447388/article/details/79503643  
版權宣告：本文為博主原創文章，轉載請附上博文連結！