选取经典的早期Pytorch官方实现代码进行分析

https://github.com/pytorch/vision/blob/9a481d0bec2700763a799ff148fe2e083b575441/torchvision/models/resnet.py

各种ResNet网络是由BasicBlock或者bottleneck构成的 
    
    
    
，它们是构成深度残差网络的基本模块

ResNet主体

ResNet的大部分各种结构是1层conv+4个block+1层fc

class ResNet(nn.Module): def __init__(self, block, layers, zero_init_residual=False): super(ResNet, self).__init__() self.inplanes = 64 self.conv1 = nn.Conv2d(3, 64, kernel_size=7, stride=2, padding=3, bias=False) self.bn1 = nn.BatchNorm2d(64) self.relu = nn.ReLU(inplace=True) self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1) self.layer1 = self._make_layer(block, 64, layers[0]) self.layer2 = self._make_layer(block, 128, layers[1], stride=2) self.layer3 = self._make_layer(block, 256, layers[2], stride=2) self.layer4 = self._make_layer(block, 512, layers[3], stride=2) self.avgpool = nn.AdaptiveAvgPool2d((1, 1)) self.fc = nn.Linear(512 * 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.init.constant_(m.weight, 1) nn.init.constant_(m.bias, 0) # Zero-initialize the last BN in each residual branch, # so that the residual branch starts with zeros, and each residual block behaves like an identity. # This improves the model by 0.2~0.3% according to https://arxiv.org/abs/1706.02677 if zero_init_residual: for m in self.modules(): if isinstance(m, Bottleneck): nn.init.constant_(m.bn3.weight, 0) elif isinstance(m, BasicBlock): nn.init.constant_(m.bn2.weight, 0) def _make_layer(self, block, planes, blocks, stride=1): downsample = None if stride != 1 or self.inplanes != planes * block.expansion: # normly happened when stride = 2 downsample = nn.Sequential( conv1x1(self.inplanes, planes * block.expansion, stride), nn.BatchNorm2d(planes * block.expansion), ) layers = [] layers.append(block(self.inplanes, planes, stride, downsample)) self.inplanes = planes * block.expansion for _ in range(1, blocks): # only the first block need downsample thus there is no downsample and stride = 2 layers.append(block(self.inplanes, planes)) return nn.Sequential(*layers) def forward(self, x): x = self.conv1(x) x = self.bn1(x) x = self.relu(x) x = self.maxpool(x) c2 = self.layer1(x) c3 = self.layer2(c2) c4 = self.layer3(c3) c5 = self.layer4(c4) x = self.avgpool(x) x = x.view(x.size(0), -1) x = self.fc(x) return c5

需要注意的是最后的avgpool是全局的平均池化

BasicBlock

class BasicBlock(nn.Module): expansion = 1 def __init__(self, inplanes, planes, stride=1, downsample=None): # here planes names channel number super(BasicBlock, self).__init__() self.conv1 = conv3x3(inplanes, planes, stride) self.bn1 = nn.BatchNorm2d(planes) self.relu = nn.ReLU(inplace=True) self.conv2 = conv3x3(planes, planes) self.bn2 = nn.BatchNorm2d(planes) self.downsample = downsample self.stride = stride def forward(self, x): identity = x out = self.conv1(x) out = self.bn1(out) out = self.relu(out) out = self.conv2(out) out = self.bn2(out) if self.downsample is not None: identity = self.downsample(x) out += identity out = self.relu(out) return out

图1. BasicBlock结构图1

ResNet18

对应的就是[2,2,2,2]

def resnet18(pretrained=False, **kwargs): """Constructs a ResNet-18 model. Args: pretrained (bool): If True, returns a model pre-trained on ImageNet """ model = ResNet(BasicBlock, [2, 2, 2, 2], **kwargs) if pretrained: print(Loading the pretrained model ...) # strict = False as we dont need fc layer params. model.load_state_dict(model_zoo.load_url(model_urls[resnet18]), strict=False) return model

ResNet34

def resnet34(pretrained=False, **kwargs): """Constructs a ResNet-34 model. Args: pretrained (bool): If True, returns a model pre-trained on ImageNet """ model = ResNet(BasicBlock, [3, 4, 6, 3], **kwargs) if pretrained: print(Loading the pretrained model ...) model.load_state_dict(model_zoo.load_url(model_urls[resnet34]), strict=False) return model

ResNet20

这个需要强调一下,正常的ResNet20应该是文章中提出
     


     


     


  ，针对cifar数据集设计的n=3时候, 1+6*3+1=20

class ResNet4Cifar(nn.Module): def __init__(self, block, num_block, num_classes=10): super().__init__() self.in_channels = 16 self.conv1 = nn.Sequential( nn.Conv2d(3, 16, kernel_size=3, padding=1, bias=False), nn.BatchNorm2d(16), nn.ReLU(inplace=True)) # we use a different inputsize than the original paper # so conv2_xs stride is 1 self.conv2_x = self._make_layer(block, 16, num_block[0], 1) self.conv3_x = self._make_layer(block, 32, num_block[1], 2) self.conv4_x = self._make_layer(block, 64, num_block[2], 2) self.avg_pool = nn.AdaptiveAvgPool2d((1, 1)) self.fc = nn.Linear(64 * block.expansion, num_classes) def _make_layer(self, block, out_channels, num_blocks, stride): strides = [stride] + [1] * (num_blocks - 1) layers = [] for stride in strides: layers.append(block(self.in_channels, out_channels, stride)) self.in_channels = out_channels * block.expansion return nn.Sequential(*layers) def forward(self, x): output = self.conv1(x) output = self.conv2_x(output) output = self.conv3_x(output) output = self.conv4_x(output) output = self.avg_pool(output) output = output.view(output.size(0), -1) output = self.fc(output) return output def resnet20(num_classes=10, **kargs): """ return a ResNet 20 object """ return ResNet4Cifar(BasicBlock, [3, 3, 3], num_classes=num_classes)

我们通过参数量的计算也为0.27M 




 




 




 
，和论文中的一致 

   

   

   

，对[1,3,32,32]的输入
     

     

     

 ，输出维度为[1,64,8,8]

图2 ResNet20参数量计算

但是也有一些文章只换了开头三层的3x3卷积层  




  




  




  ，通道数并没有采用16    
    
    
  、32  


     


     


     
、64 


  


  


  


，仍是4层的64


 




 




 




、128   

   

   

 、256   

     

     

     、512


     
     
     
，这样下来参数量是11.25M 
    
    
    
。针对的任务不同   




   




   




   ，但是如果不关注原始网络结构 



 



 



 


，这一点可以忽略
     


     


     


  。

Bottleneck Block

Bottleneck Block中使用了1×1卷积层 




 




 




 
。如输入通道数为256               ，1×1卷积层会将通道数先降为64
    




    




    




   ，经过3×3卷积层后

















，再将通道数升为256 

   

   

   

。1×1卷积层的优势是在更深的网络中                ，用较小的参数量处理通道数很大的输入
     

     

     

 。

这种结构用在ResNet50



  




  




  




、ResNet101中  




  




  




  。

图2. Bottleneck 结构图1

class Bottleneck(nn.Module): expansion = 4 def __init__(self, inplanes, planes, stride=1, downsample=None): super(Bottleneck, self).__init__() self.conv1 = conv1x1(inplanes, planes) self.bn1 = nn.BatchNorm2d(planes) self.conv2 = conv3x3(planes, planes, stride) self.bn2 = nn.BatchNorm2d(planes) self.conv3 = conv1x1(planes, planes * self.expansion) self.bn3 = nn.BatchNorm2d(planes * self.expansion) self.relu = nn.ReLU(inplace=True) self.downsample = downsample self.stride = stride def forward(self, x): 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

ResNet50

图3. ResNet50结构图2

和以上的网络结构一样
     



     



     



   ，把Bottleneck按层数堆起来就可以了

def resnet50(pretrained=False, **kwargs): """Constructs a ResNet-50 model. Args: pretrained (bool): If True, returns a model pre-trained on ImageNet """ model = ResNet(Bottleneck, [3, 4, 6, 3], **kwargs) if pretrained: print(Loading the pretrained model ...) model.load_state_dict(model_zoo.load_url(model_urls[resnet50]), strict=False) return model

ResNet到底解决了什么问题

推荐看知乎问题Resnet到底在解决一个什么问题呢？

贴一些我比较喜欢的回答：

A. 对于

L

L

L层的网络来说
















，没有残差表示的Plain Net梯度相关性的衰减在

1

2

L

\frac{1}{2^L}

2L1​  
    
    
    
，而ResNet的衰减却只有

1

L

\frac{1}{\sqrt{L}}

L​1​  


  


  


  


。即使BN过后梯度的模稳定在了正常范围内
     


     


     


  ，但梯度的相关性实际上是随着层数增加持续衰减的
     
     
     
。而经过证明 




 




 




 
，ResNet可以有效减少这种相关性的衰减   




   




   




   。

B. 对于“梯度弥散    
    
    
  ”观点来说 

   

   

   

，在输出引入一个输入x的恒等映射
     

     

     

 ，则梯度也会对应地引入一个常数1  




  




  




  ，这样的网络的确不容易出现梯度值异常 


  


  


  


，在某种意义上
     
     
     
，起到了稳定梯度的作用 



 



 



 


。

C. 跳连接相加可以实现不同分辨率特征的组合   




   




   




   ，因为浅层容易有高分辨率但是低级语义的特征 



 



 



 


，而深层的特征有高级语义               ，但分辨率就很低了               。引入跳接实际上让模型自身有了更加“灵活  


     


     


     
”的结构
    




    




    




   ，即在训练过程本身

















，模型可以选择在每一个部分是“更多进行卷积与非线性变换


 




 




 




”还是“更多倾向于什么都不做   

   

   

 ”                ，抑或是将两者结合
    




    




    




   。模型在训练便可以自适应本身的结构

















。3

D. 当使用了残差网络时,就是加入了skip connection 结构,这时候由一个building block 的任务由: F(x) := H(x)
     



     



     



   ，变成了F(x) := H(x)-x对比这两个待拟合的函数, 拟合残差图更容易优化,也就是说:F(x) := H(x)-x比F(x) := H(x)更容易优化4. 举了一个差分放大器的例子：F是求和前网络映射
















，H是从输入到求和后的网络映射                。比如把5映射到5.1 
    
    
    
，那么引入残差前是F’(5)=5.1
     


     


     


  ，引入残差后是H(5)=5.1, H(5)=F(5)+5, F(5)=0.1
     



     



     



   。这里的F’和F都表示网络参数映射 




 




 




 
，引入残差后的映射对输出的变化更敏感
















。比如s输出从5.1变到5.2 

   

   

   

，映射F’的输出增加了1/51=2%
     

     

     

 ，而对于残差结构输出从5.1到5.2  




  




  




  ，映射F是从0.1到0.2 


  


  


  


，增加了100% 
    
    
    
。明显后者输出变化对权重的调整作用更大
     
     
     
，所以效果更好
     


     


     


  。残差的思想都是去掉相同的主体部分   




   




   




   ，从而突出微小的变化 




 




 




 
。

说法众多 



 



 



 


，好用就完事儿了嗷~

【pytorch系列】ResNet中的BasicBlock与bottleneck ↩︎ ↩︎

ResNet50网络结构图及结构详解 ↩︎

https://www.zhihu.com/question/64494691/answer/786270699 ↩︎

https://www.zhihu.com/question/64494691/answer/271335912 ↩︎