1. ResNet论文详解

1.1. Introduction

一般网络越深，特征就越丰富，模型效果也就越好。在深度重要的驱动下，出现了2个问题：

1. 梯度消失和梯度爆炸：

   • 梯度消失：误差梯度<1，当网络层数增多时，最终求的梯度会以指数形式衰减
   • 梯度爆炸：误差梯度>1，当网络层数增多时，最终求的梯度会以指数形式增加
   • 解决方式：

     1. Xavier 初始化、Kaiming 初始化等

     2. Batch Normalization

2. 退化问题：在适当深度的模型中添加更多的层会导致更高的训练误差，如下图：

在本文中，我们通过残差结构来解决退化问题。

原本是通过堆叠非线性层来适合$H(x)$，现在是让这些非线性层来适合$F(x)$，原始映射被表示为：$H(x):=F(x)+x$

• $H(x):$ 原本需要学习的映射
• $F(x):$ 现在需要学习的映射
• $x:$ 单位映射，可以跳过一层或多层的连接(shortcut connection)

实验表示：

1. 极深残差网络很容易优化
2. 很容易获得网络深度带来的准确性的提高

1.2. Deep Residual Learning

1.2.1 Residual Learning

退化问题表明：很难通过多个非线性层来逼近单位映射，因为如果可以的话，那么更深的模型的训练误差应该不大于更浅层的对应模型。而对应残差网络，如果单位映射为最优的话，求解器可以简单地将多个非线性层地权值置为0从而逼近单位映射。

1.2.2 Identity Mapping by Shortcuts

在本文中，残差块定义如下：

• $x$和$F$的维度一致： y = F ( x , { W i } ) + x y=F(x,\{W_i\})+x y=F(x,{Wi​})+x
• $x$和$F$的维度不一致： y = F ( x , { W i } ) + W s x y=F(x,\{W_i\})+W_sx y=F(x,{Wi​})+Ws​x（$W_s$用来匹配维度）

残差函数$F$是灵活的，主要表现层数的个数和层的类别上

1. 在本文的实验中涉及的$F$，它有两层或三层，也可以有更多层，但是如果只有一层，则就类似于线性层：$y=W_{1}x+x$，我们没有观察到任何优势。
2. 尽管上面公式表示法是全连接层，但它同样适用于卷积层

1.3. Network Architectures

在一般网络结构的基础上，插入shortcut connection，将网络变为对应的残差网络。

当输入和输出的维数相同时：对应于上图的实线shortcut connection，处理方式：

• 直接使用单位映射

当输入和输出的维数不相同时：对应于上图的虚线shortcut connection，处理方式：

• shortcut connection仍然使用单位映射，增加维数用0填充，此方法不引入额外的参数
• 使用1x1卷积来匹配维度，文中称为projection shortcut

当跨越两种尺寸的特征图时，执行步幅为2的。有上表黄色部分可知。

1.4. Experiments

1.4.1 Residual Networks VS. Plain Networks

Plain Networks

上图表示：34层网络比18层网络具有更高的验证误差，尽管18层网络的解空间是34层网络的子空间，但在整个训练过程中，34层网络的训练误差都比较大，这说明了退化问题

我们认为这种优化困难不可能是梯度消失引起的，因为这些网络都使用了BN进行训练，保证前向传播的信号具有非零方差。我们还验证了反向传播的梯度在BN上表现出健康的范数。所以向前和向后的信号都不会消失。

Residual Networks

从图4和表2可知：

1. 34层的残差网络比18层残差网络有着相当低的训练误差，并可推广到验证数据，这说明退化问题得到很好的解决，这说明了残差网络在极深网络的有效性
2. 通过比较18层网络和18层残差网络，发现残差网络收敛的更快

1.4.2 Identity VS. Projection Shortcuts

在表3中，我们比较了3个选项（projection：在文中指利用1x1卷积来进行改变维度）

1. A：用0填充以增加维度
2. B：projection shortcuts用于增加维度，其他shortcuts为单位映射
3. C：所有shortcuts均为projection

由表3可知：

• B略优于A：A中的0填充的维度没有进行残差学习
• C略优于B：多个projection引入了额外的参数
• A、B和C的细小差异表明：projection对于解决退化问题不是必须的，因此在本文的其余部分，我们不使用C以减少内存/时间复杂度和模型的规模，使用B

1.4.3 Deeper Bottleneck Architectures

1x1卷积用于升维或降维，可以减少网络的参数

• 上图左：256x3x3x256+256x3x3x256=1179648
• 上图右：256x1x1x64+64x3x3x64+64x1x1x256=69632

故在更深的残差网络中，使用上图右的结构，可以大大减少网络的参数，减小模型的规模

2. 基于Pytorch代码复现

2.1 模型搭建

import torch.nn as nn
import torch
from torchsummary import summary
import torchvision.models as modelsclass BasicBlock(nn.Module):expansion = 1def __init__(self, in_channel, out_channel, stride=1, downsample=None):super(BasicBlock, self).__init__()self.conv1 = nn.Conv2d(in_channels=in_channel, out_channels=out_channel, kernel_size=3, stride=stride, padding=1, bias=False)self.bn1 = nn.BatchNorm2d(out_channel)self.relu = nn.ReLU(inplace=True)self.conv2 = nn.Conv2d(in_channels=out_channel, out_channels=out_channel, kernel_size=3, stride=1, padding=1, bias=False)self.bn2 = nn.BatchNorm2d(out_channel)self.downsample = downsampledef forward(self, x):identity = xif self.downsample is not None:identity = self.downsample(x)out = self.conv1(x)out = self.bn1(out)out = self.relu(out)out = self.conv2(out)out = self.bn2(out)out += identityout = self.relu(out)return outclass Bottleneck(nn.Module):expansion = 4def __init__(self, in_channel, out_channel, stride=1, downsample=None):super(Bottleneck, self).__init__()self.conv1 = nn.Conv2d(in_channels=in_channel, out_channels=out_channel, kernel_size=1, stride=1, bias=False)self.bn1 = nn.BatchNorm2d(out_channel)self.conv2 = nn.Conv2d(in_channels=out_channel, out_channels=out_channel, kernel_size=3, stride=stride, padding=1, bias=False)self.bn2 = nn.Conv2d(out_channel)self.conv3 = nn.Conv2d(in_channels=out_channel, out_channels=out_channel*self.expansion, kernel_size=1, stride=1, bias=False)self.bn3 = nn.BatchNorm2d(out_channel*self.expansion)self.relu = nn.ReLU(inplace=True)self.downsample = downsampledef forward(self, x):identity = xif self.downsample is not None:identity = self.downsample(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)out += identityout = self.relu(out)return outclass ResNet(nn.Module):def __init__(self, block, block_num, num_classes=1000, include_top=True):super(ResNet, self).__init__()self.include_top = include_topself.in_channel = 64self.conv1 = nn.Conv2d(3, self.in_channel, kernel_size=7, stride=2, padding=3, bias=False)self.bn1 = nn.BatchNorm2d(self.in_channel)self.relu = nn.ReLU(inplace=True)self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)self.layer1 = self._make_layer(block, 64, block_num[0])self.layer2 = self._make_layer(block, 128, block_num[1], stride=2)self.layer3 = self._make_layer(block, 256, block_num[2], stride=2)self.layer4 = self._make_layer(block, 512, block_num[3], stride=2)if self.include_top: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')def _make_layer(self, block, channel, block_num, stride=1):downsample = Noneif stride != 1 or self.in_channel != channel * block.expansion:downsample = nn.Sequential(nn.Conv2d(self.in_channel, channel * block.expansion, kernel_size=1, stride=stride, bias=False),nn.BatchNorm2d(channel*block.expansion))layers = []layers.append(block(self.in_channel, channel, downsample=downsample, stride=stride))self.in_channel = channel * block.expansionfor _ in range(1, block_num):layers.append(block(self.in_channel, channel))return nn.Sequential(*layers)def forward(self, x):x = self.conv1(x)x = self.bn1(x)x = self.relu(x)x = self.maxpool(x)x = self.layer1(x)x = self.layer2(x)x = self.layer3(x)x = self.layer4(x)if self.include_top:x = self.avgpool(x)x = torch.flatten(x, 1)x = self.fc(x)return xdef resnet34(num_classes=1000, include_top=True):return ResNet(BasicBlock, [3, 4, 6, 3], num_classes=num_classes, include_top=include_top)def resnet101(num_classes=1000, include_top=True):return ResNet(Bottleneck, [3, 4, 23, 3], num_classes=num_classes, include_top=include_top)def read_resnet34():device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')model = models.resnet34(pretrained=True)model.to(device)print(model)summary(model, input_size=(3, 224, 224))def get_resnet34(flag, num_classes):if flag:net = models.resnet34(pretrained=True)num_input = net.fc.in_featuresnet.fc = nn.Linear(num_input, num_classes)else:net = resnet34(num_classes)return net

2.2 训练结果如下

1. 训练数据集与验证集大小以及训练参数

Using 3306 images for training, 364 images for validation
Using cuda GeForce RTX 2060 device for training
lr: 0.0001
batch_size: 16

2. 使用自己定义的网络训练结果

[epoch 1/10] train_loss: 1.309 val_acc: 0.555
[epoch 2/10] train_loss: 1.146 val_acc: 0.604
[epoch 3/10] train_loss: 1.029 val_acc: 0.643
[epoch 4/10] train_loss: 0.935 val_acc: 0.695
[epoch 5/10] train_loss: 0.919 val_acc: 0.615
[epoch 6/10] train_loss: 0.860 val_acc: 0.723
[epoch 7/10] train_loss: 0.841 val_acc: 0.690
[epoch 8/10] train_loss: 0.819 val_acc: 0.725
[epoch 9/10] train_loss: 0.800 val_acc: 0.745
[epoch 10/10] train_loss: 0.783 val_acc: 0.725
Best acc: 0.745
Finished Training
Train 耗时为：281.2s

4. 使用预训练模型参数训练结果

[epoch 1/10] train_loss: 0.492 val_acc: 0.896
[epoch 2/10] train_loss: 0.327 val_acc: 0.896
[epoch 3/10] train_loss: 0.285 val_acc: 0.909
[epoch 4/10] train_loss: 0.273 val_acc: 0.904
[epoch 5/10] train_loss: 0.205 val_acc: 0.901
[epoch 6/10] train_loss: 0.245 val_acc: 0.898
[epoch 7/10] train_loss: 0.200 val_acc: 0.923
[epoch 8/10] train_loss: 0.196 val_acc: 0.923
[epoch 9/10] train_loss: 0.179 val_acc: 0.929
[epoch 10/10] train_loss: 0.173 val_acc: 0.926
Best acc: 0.929
Finished Training
Train 耗时为：281.3s

上一篇：GoogLeNet
下一篇：DenseNet
完整代码