CIFAR-10 数据集实战 —— 构建 ResNet18 神经网络

如果不了解 ResNet 的同学可以先看我的这篇博客 ResNet 论文阅读

首先实现一个 Residual Block

import torch
from torch import nn
from torch.nn import functional as F

class ResBlk(nn.Module):
    def __init__(self, ch_in, ch_out, stride=1):
        super(ResBlk, self).__init__()
        self.conv1 = nn.Conv2d(ch_in, ch_out, kernel_size=3, stride=stride, padding=1)
        self.bn1 = nn.BatchNorm2d(ch_out)
        
        self.conv2 = nn.Conv2d(ch_out, ch_out, kernel_size=3, stride=1, padding=1)
        self.bn2 = nn.BatchNorm2d(ch_out)
        
        if ch_out == ch_in:
            self.extra = nn.Sequential()
        else:
            self.extra = nn.Sequential(
                
                # 1×1的卷积作用是修改输入x的channel
                # [b, ch_in, h, w] => [b, ch_out, h, w]
                nn.Conv2d(ch_in, ch_out, kernel_size=1, stride=stride),
                nn.BatchNorm2d(ch_out),
            )
        
    def forward(self, x):
        out = F.relu(self.bn1(self.conv1(x)))
        out = self.bn2(self.conv2(out))

        # short cut
        out = self.extra(x) + out
        out = F.relu(out)
        
        return out

Block 中进行了正则化处理，以使 train 过程更快更稳定。同时要考虑，如果两元素的 ch_in 和 ch_out 不匹配，进行加法时会报错，因此需要判断一下，如果不相等，就用 1×1 的卷积调整一下

测试一下

blk = ResBlk(64, 128, stride=2)
tmp = torch.randn(2, 64, 32, 32)
out = blk(tmp)
print(out.shape)

输出的 shape 大小是 torch.Size([2, 128, 16, 16])

这里解释一下，为什么有的层要专门设置 stride。对于一个 Residual block，channel 从 64 增大到 128，如果所有的 stride 都是 1，padding 也是 1，那么图片的 w 和 h 也不会变，但是 channel 增大了，此时就会导致整个网络的参数增多。而这才仅仅一个 Block，更不用说后面的 FC 以及更多 Block 了，所以 stride 不能全部设置为 1，不要让网络的参数一直增大

然后我们搭建完整的 ResNet-18

class ResNet18(nn.Module):
    def __init__(self):
        super(ResNet18, self).__init__()
        
        self.conv1 = nn.Sequential(
            nn.Conv2d(3, 64, kernel_size=3, stride=3, padding=0),
            nn.BatchNorm2d(64),
        )
        # followed 4 blocks
        
        # [b, 64, h, w] => [b, 128, h, w]
        self.blk1 = ResBlk(64, 128, stride=2)
        # [b, 128, h, w] => [b, 256, h, w]
        self.blk2 = ResBlk(128, 256, stride=2)
        # [b, 256, h, w] => [b, 512, h, w]
        self.blk3 = ResBlk(256, 512, stride=2)
        # [b, 512, h, w] => [b, 512, h, w]
        self.blk4 = ResBlk(512, 512, stride=2)
        
        self.outlayer = nn.Linear(512*1*1, 10)
    
    def forward(self, x):
        x = F.relu(self.conv1(x))
        
        # 经过四个blk以后 [b, 64, h, w] => [b, 512, h, w]
        x = self.blk1(x)
        x = self.blk2(x)
        x = self.blk3(x)
        x = self.blk4(x)
        
        x = self.outlayer(x)
        
        return x

测试一下

x = torch.randn(2, 3, 32, 32)
model = ResNet18()
out = model(x)
print("ResNet:", out.shape)

结果报错了，错误信息如下

size mismatch, m1: [2048 x 2], m2: [512 x 10] at /pytorch/aten/src/TH/generic/THTensorMath.cpp:961

问题在于我们最后定义线性层的输入维度，和上一层 Block 的输出维度不匹配，在 ResNet18 的最后一个 Block 运行结束后打印一下当前 x 的 shape，结果是 torch.Size([2, 512, 2, 2])

解决办法有很多，可以修改线性层的输入进行匹配，也可以在最后一层 Block 后面再进行一些操作，使其与 512 匹配

先给出修改后的代码，在做解释

class ResNet18(nn.Module):
    def __init__(self):
        super(ResNet18, self).__init__()
        
        self.conv1 = nn.Sequential(
            nn.Conv2d(3, 64, kernel_size=3, stride=3, padding=0),
            nn.BatchNorm2d(64),
        )
        # followed 4 blocks
        
        # [b, 64, h, w] => [b, 128, h, w]
        self.blk1 = ResBlk(64, 128, stride=2)
        # [b, 128, h, w] => [b, 256, h, w]
        self.blk2 = ResBlk(128, 256, stride=2)
        # [b, 256, h, w] => [b, 512, h, w]
        self.blk3 = ResBlk(256, 512, stride=2)
        # [b, 512, h, w] => [b, 512, h, w]
        self.blk4 = ResBlk(512, 512, stride=2)
        
        self.outlayer = nn.Linear(512*1*1, 10)
    
    def forward(self, x):
        x = F.relu(self.conv1(x))
        
        # 经过四个blk以后 [b, 64, h, w] => [b, 512, h, w]
        x = self.blk1(x)
        x = self.blk2(x)
        x = self.blk3(x)
        x = self.blk4(x)
        
        # print("after conv:", x.shape) # [b, 512, 2, 2]
        
        # [b, 512, h, w] => [b, 512, 1, 1]
        x = F.adaptive_avg_pool2d(x, [1, 1])
        
        x = x.view(x.size(0), -1) # [b, 512, 1, 1] => [b, 512*1*1]
        x = self.outlayer(x)
        
        return x

这里我采用的是第二种方法，在最后一个 Block 结束以后，接了一个自适应的 pooling 层，这个 pooling 的作用是将不论输入的宽高是多少，全部输出称宽高都是 1 的 tensor，其他维度保持不变。然后再做一个 reshape 操作，将 [batchsize, 512, 1, 1]reshape 成 [batchsize, 512*1*1] 大小的 tensor，这样就和接下来的线性层对上了，线性层的输入大小是 512，输出是 10。因此整个网络最终输出的 shape 就是 [batchsize, 10]

最后我们把之前训练 LeNet5 的代码拷贝过来，将里面的 model=LeNet5() 改为 model=ResNet18() 就行了。完整代码如下

import torch
from torch import nn, optim
import torch.nn.functional as F
from torch.utils.data import DataLoader
from torchvision import datasets, transforms


batch_size=32
cifar_train = datasets.CIFAR10(root='cifar', train=True, transform=transforms.Compose([
    transforms.Resize([32, 32]),
    transforms.ToTensor(),
]), download=True)

cifar_train = DataLoader(cifar_train, batch_size=batch_size, shuffle=True)

cifar_test = datasets.CIFAR10(root='cifar', train=False, transform=transforms.Compose([
    transforms.Resize([32, 32]),
    transforms.ToTensor(),
]), download=True)
    
cifar_test = DataLoader(cifar_test, batch_size=batch_size, shuffle=True)      

class ResBlk(nn.Module):
    def __init__(self, ch_in, ch_out, stride=1):
        super(ResBlk, self).__init__()
        self.conv1 = nn.Conv2d(ch_in, ch_out, kernel_size=3, stride=stride, padding=1)
        self.bn1 = nn.BatchNorm2d(ch_out)
        
        self.conv2 = nn.Conv2d(ch_out, ch_out, kernel_size=3, stride=1, padding=1)
        self.bn2 = nn.BatchNorm2d(ch_out)
        
        if ch_out == ch_in:
            self.extra = nn.Sequential()
        else:
            self.extra = nn.Sequential(
                
                # 1×1的卷积作用是修改输入x的channel
                # [b, ch_in, h, w] => [b, ch_out, h, w]
                nn.Conv2d(ch_in, ch_out, kernel_size=1, stride=stride),
                nn.BatchNorm2d(ch_out),
            )
        
    def forward(self, x):
        out = F.relu(self.bn1(self.conv1(x)))
        out = self.bn2(self.conv2(out))

        # short cut
        out = self.extra(x) + out
        out = F.relu(out)
        
        return out
        
class ResNet18(nn.Module):
    def __init__(self):
        super(ResNet18, self).__init__()
        
        self.conv1 = nn.Sequential(
            nn.Conv2d(3, 64, kernel_size=3, stride=3, padding=0),
            nn.BatchNorm2d(64),
        )
        # followed 4 blocks
        
        # [b, 64, h, w] => [b, 128, h, w]
        self.blk1 = ResBlk(64, 128, stride=2)
        # [b, 128, h, w] => [b, 256, h, w]
        self.blk2 = ResBlk(128, 256, stride=2)
        # [b, 256, h, w] => [b, 512, h, w]
        self.blk3 = ResBlk(256, 512, stride=2)
        # [b, 512, h, w] => [b, 512, h, w]
        self.blk4 = ResBlk(512, 512, stride=2)
        
        self.outlayer = nn.Linear(512*1*1, 10)
    
    def forward(self, x):
        x = F.relu(self.conv1(x))
        
        # 经过四个blk以后 [b, 64, h, w] => [b, 512, h, w]
        x = self.blk1(x)
        x = self.blk2(x)
        x = self.blk3(x)
        x = self.blk4(x)
        
        # print("after conv:", x.shape) # [b, 512, 2, 2]
        
        # [b, 512, h, w] => [b, 512, 1, 1]
        x = F.adaptive_avg_pool2d(x, [1, 1])
        
        x = x.view(x.size(0), -1) # [b, 512, 1, 1] => [b, 512*1*1]
        x = self.outlayer(x)
        
        return x

def main():

    ##########  train  ##########
    #device = torch.device('cuda')
    #model = ResNet18().to(device)
    criteon = nn.CrossEntropyLoss()
    model = ResNet18()
    optimizer = optim.Adam(model.parameters(), 1e-3)
    for epoch in range(1000):
        model.train()
        for batchidx, (x, label) in enumerate(cifar_train):
            #x, label = x.to(device), label.to(device)
            logits = model(x)
            # logits: [b, 10]
            # label:  [b]
            loss = criteon(logits, label)
            
            # backward
            optimizer.zero_grad()
            loss.backward()
            optimizer.step()
        
        print('train:', epoch, loss.item())
        
        ########## test  ##########
        model.eval()
        with torch.no_grad():
            total_correct = 0
            total_num = 0
            for x, label in cifar_test:
                # x, label = x.to(device), label.to(device)

                # [b]
                logits = model(x)
                # [b]
                pred = logits.argmax(dim=1)
                # [b] vs [b]
                total_correct += torch.eq(pred, label).float().sum().item()
                total_num += x.size(0)
            acc = total_correct / total_num
            print('test:', epoch, acc)

if __name__ == '__main__':
    main()

ResNet 和 LeNet 相比，准确率提升的很快，但是由于层数增加，不可避免的会导致运行时间增加，如果没有 GPU，运行一个 epoch 大概要 15 分钟。读者同样可以在此基础上修改网络结构，运用一些 tricks，比方说一开始就对图片做一个 Normalize 等