xmlns="http://www.w3.org/2000/svg"

style="display:

none;">

[深度学习网络从入门到入土]

残差网络ResNet

个人导航

知乎：https://www.zhihu.com/people/byzh_rc

CSDN：https://blog.csdn.net/qq_54636039

注：本文仅对所述内容做了框架性引导，具体细节可查询其余相关资料or源码

参考文章：各方资料

参考资料

Deep

Residual

Recognition.

背景

在

2014–2015

进入“越深越好”的阶段：

• AlexNet：8

  层

问题来了：当网络超过

层后，训练误差反而上升

这不是过拟合，而是优化困难（degradation

problem）

resnet横空出世:让网络学习“残差”，而不是直接学习映射

传统网络：

=

H(x)=F(x)

style="margin-right:

0.0813em;">H(x)=

style="margin-right:

0.1389em;">F(x)

/>ResNet：

0.0813em;">H(x)=

style="margin-right:

0.1389em;">F(x)+x

架构(公式)

/>

1.BasicBlock（ResNet18/34）

Conv
ReLU
ReLU

y

=

0.0359em;">y=ReLU(

style="margin-right:

0.1389em;">F(x)+x)

style="margin-right:

0.1389em;">F(x)=

style="margin-right:

0.1389em;">W

style="height:

0.05em;">2​

style="height:

0.15em;">

style="margin-right:

0.0359em;">σ(

style="margin-right:

0.1389em;">W

style="height:

0.05em;">1​

style="height:

0.15em;">x)

2.Bottleneck（ResNet50/101/152）

当网络变得非常深时，使用瓶颈结构：

1×1(降维)
3×3(提取特征)
1×1(升维)

F

=

0.1389em;">F(x)=

style="margin-right:

0.1389em;">W

style="height:

0.05em;">3​

style="height:

0.15em;">

style="margin-right:

0.0359em;">σ(

style="margin-right:

0.1389em;">W

style="height:

0.05em;">2​

style="height:

0.15em;">

style="margin-right:

0.0359em;">σ(

style="margin-right:

0.1389em;">W

style="height:

0.05em;">1​

style="height:

0.15em;">x))

3.Shortcut

类型

情况1：尺寸相同

=

0.0359em;">y=

style="margin-right:

0.1389em;">F(x)+x

/>情况2：尺寸不同（下采样）

=

0.0359em;">y=

style="margin-right:

0.1389em;">F(x)+

style="margin-right:

0.1389em;">W

style="height:

0.05em;">s​

style="height:

0.15em;">x

style="color:

purple;">W

style="height:

purple;">s​

style="height:

0.15em;">

style="color:

purple;">Conv

/>

创新点

1.残差连接（Skip

Connection）

允许梯度直接传播

2.

层首次成功训练

3.

结构简洁但极强

成为后续几乎所有视觉网络的基础(DenseNet,

U-Net)

为什么

1

style="height:

0.0556em;">∂x

style="top:

0.0359em;">y​

style="height:

0.686em;">=

style="height:

0.0556em;">∂x

style="top:

0.1389em;">F(x)​

style="height:

0.686em;">+1

/>即梯度中始终存在

项：

• 梯度不会消失
• 网络可以直接传递恒等映射

代码实现

/>

importtorchimporttorch.nnasnnimporttorch.nn.functionalasFfrombyzh.ai.Butilsimportb_get_paramsclassBasicBlock(nn.Module):"""
ResNet18/34
"""expansion=1def__init__(self,in_ch,out_ch,stride=1):super().__init__()self.conv1=nn.Conv2d(in_ch,out_ch,3,stride,1,bias=False)self.bn1=nn.BatchNorm2d(out_ch)self.conv2=nn.Conv2d(out_ch,out_ch,3,1,1,bias=False)self.bn2=nn.BatchNorm2d(out_ch)self.shortcut=nn.Sequential()ifstride!=1orin_ch!=out_ch:self.shortcut=nn.Sequential(nn.Conv2d(in_ch,out_ch,1,stride,bias=False),nn.BatchNorm2d(out_ch))defforward(self,x):out=torch.relu(self.bn1(self.conv1(x)))out=self.bn2(self.conv2(out))out+=self.shortcut(x)out=torch.relu(out)returnoutclassBottleneck(nn.Module):"""
ResNet50/101/152
"""expansion=4#
=
4def__init__(self,in_ch,out_ch,stride=1):super().__init__()#
1x1
降维self.conv1=nn.Conv2d(in_ch,out_ch,kernel_size=1,bias=False)self.bn1=nn.BatchNorm2d(out_ch)#
3x3
下采样）self.conv2=nn.Conv2d(out_ch,out_ch,kernel_size=3,stride=stride,padding=1,bias=False)self.bn2=nn.BatchNorm2d(out_ch)#
1x1
升维self.conv3=nn.Conv2d(out_ch,out_ch*self.expansion,kernel_size=1,bias=False)self.bn3=nn.BatchNorm2d(out_ch*self.expansion)self.shortcut=nn.Sequential()ifstride!=1orin_ch!=out_ch*self.expansion:self.shortcut=nn.Sequential(nn.Conv2d(in_ch,out_ch*self.expansion,kernel_size=1,stride=stride,bias=False),nn.BatchNorm2d(out_ch*self.expansion))defforward(self,x):out=torch.relu(self.bn1(self.conv1(x)))out=torch.relu(self.bn2(self.conv2(out)))out=self.bn3(self.conv3(out))out+=self.shortcut(x)out=torch.relu(out)returnoutclassResNet(nn.Module):"""
input
"""def__init__(self,block,layers,num_classes=1000):super().__init__()self.in_ch=64self.conv1=nn.Conv2d(3,64,7,2,3,bias=False)self.bn1=nn.BatchNorm2d(64)self.maxpool=nn.MaxPool2d(3,2,1)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)def_make_layer(self,block,out_ch,blocks,stride=1):layers=[]layers.append(block(self.in_ch,out_ch,stride))self.in_ch=out_ch*block.expansionfor_inrange(1,blocks):layers.append(block(self.in_ch,out_ch))returnnn.Sequential(*layers)defforward(self,x):x=torch.relu(self.bn1(self.conv1(x)))x=self.maxpool(x)x=self.layer1(x)x=self.layer2(x)x=self.layer3(x)x=self.layer4(x)x=self.avgpool(x)x=torch.flatten(x,1)x=self.fc(x)returnxclassB_ResNet18_Paper(ResNet):"""
input
"""def__init__(self,num_classes=1000):block=BasicBlock
layers=[2,2,2,2]super().__init__(block=block,layers=layers,num_classes=num_classes)classB_ResNet34_Paper(ResNet):"""
input
"""def__init__(self,num_classes=1000):block=BasicBlock
layers=[3,4,6,3]super().__init__(block=block,layers=layers,num_classes=num_classes)classB_ResNet50_Paper(ResNet):"""
input
"""def__init__(self,num_classes=1000):block=Bottleneck
layers=[3,4,6,3]super().__init__(block=block,layers=layers,num_classes=num_classes)classB_ResNet101_Paper(ResNet):"""
input
"""def__init__(self,num_classes=1000):block=Bottleneck
layers=[3,4,23,3]super().__init__(block=block,layers=layers,num_classes=num_classes)classB_ResNet152_Paper(ResNet):"""
input
"""def__init__(self,num_classes=1000):block=Bottleneck
layers=[3,8,36,3]super().__init__(block=block,layers=layers,num_classes=num_classes)if__name__=='__main__':#
ResNet18net=B_ResNet18_Paper(num_classes=1000)a=torch.randn(50,3,224,224)result=net(a)print(result.shape)print(f"参数量:{b_get_params(net)}")#
ResNet34net=B_ResNet34_Paper(num_classes=1000)a=torch.randn(50,3,224,224)result=net(a)print(result.shape)print(f"参数量:{b_get_params(net)}")#
ResNet50net=B_ResNet50_Paper(num_classes=1000)a=torch.randn(50,3,224,224)result=net(a)print(result.shape)print(f"参数量:{b_get_params(net)}")#
ResNet101net=B_ResNet101_Paper(num_classes=1000)a=torch.randn(50,3,224,224)result=net(a)print(result.shape)print(f"参数量:{b_get_params(net)}")#
ResNet152net=B_ResNet152_Paper(num_classes=1000)a=torch.randn(50,3,224,224)result=net(a)print(result.shape)print(f"参数量:{b_get_params(net)}")#
60_192_808

项目实例

库环境:

numpy==1.26.4
...

ResNet18训练MNIST数据集:

#
copy
runimporttorchimporttorch.nn.functionalasFfromuploadToPypi_ai.byzh.ai.Bdataimportb_stratified_indicesfrombyzh.ai.BtrainerimportB_Classification_Trainerfrombyzh.ai.BdataimportB_Download_MNIST,b_get_dataloader_from_tensor#
from
uploadToPypi_ai.byzh.ai.Bmodel.study_cnn
import
B_ResNet18_Paperfrombyzh.ai.Bmodel.study_cnnimportB_ResNet18_Paperfrombyzh.ai.Butilsimportb_get_device#####
hyper
#####epochs=10lr=1e-3batch_size=32device=b_get_device(use_idle_gpu=True)#####
data
#####downloader=B_Download_MNIST(save_dir='D:/study_cnn/datasets/MNIST')data_dict=downloader.get_data()X_train=data_dict['X_train_standard']y_train=data_dict['y_train']X_test=data_dict['X_test_standard']y_test=data_dict['y_test']num_classes=data_dict['num_classes']num_samples=data_dict['num_samples']indices=b_stratified_indices(y_train,num_samples//5)X_train=X_train[indices]X_train=F.interpolate(X_train,size=(224,224),mode='bilinear')X_train=X_train.repeat(1,3,1,1)y_train=y_train[indices]indices=b_stratified_indices(y_test,num_samples//5)X_test=X_test[indices]X_test=F.interpolate(X_test,size=(224,224),mode='bilinear')X_test=X_test.repeat(1,3,1,1)y_test=y_test[indices]train_dataloader,val_dataloader=b_get_dataloader_from_tensor(X_train,y_train,X_test,y_test,batch_size=batch_size)#####
model
#####model=B_ResNet18_Paper(num_classes=num_classes)#####
else
#####optimizer=torch.optim.Adam(model.parameters(),lr=lr)criterion=torch.nn.CrossEntropyLoss()#####
trainer
#####trainer=B_Classification_Trainer(model=model,optimizer=optimizer,criterion=criterion,train_loader=train_dataloader,val_loader=val_dataloader,device=device)trainer.set_writer1('./runs/resnet18/log.txt')#####
run
#####trainer.train_eval_s(epochs=epochs)#####
calculate
#####trainer.draw_loss_acc('./runs/resnet18/loss_acc.png',y_lim=False)trainer.save_best_checkpoint('./runs/resnet18/best_checkpoint.pth')trainer.calculate_model()