利用pytorch复现resnet对cifar-10进行分类

一、resnet创新点

传统的卷积网络在网络很深的时候，会出现梯度消失或者梯度爆炸的现象而resnet就能很好的解决这个问题。

resnet最为创新的一点是残差结构，它使用了一种连接方式叫做“shortcut connection”，顾名思义，shortcut就是“抄近道”的意思。示意图如下。

它对每层的输入做一个reference（X）, 学习形成残差函数， 而不是学习一些没有reference（X）的函数。这种残差函数更容易优化，能使网络层数大大加深。

resnet还有一种结构，用在resnet50/101/102上。

极大程度上的减小了参数量，它将两个3x3的卷积层替换为1x1 + 3x3 + 1x1。第一个1x1的卷积把256维channel降到64维，然后在最后通过1x1卷积恢复，整体上用的参数数目：1x1x256x64 + 3x3x64x64 + 1x1x64x256 = 69632，而不使用bottleneck的话就是两个3x3x256的卷积，参数数目: 3x3x256x256x2 = 1179648，差了16.94倍。

二、代码部分

（一）数据预处理

引入包库

import torch
import torch.nn as nn
import torch.optim as optim
import torchvision
import torchvision.transforms as transforms
import os
import argparse
from torch.autograd import Variable

参数设置

parser = argparse.ArgumentParser(description='cifar10')
parser.add_argument('--lr', default=1e-2, help='learning rate')
parser.add_argument('--epoch', default=20, help='time for ergodic')
parser.add_argument('--pre_epoch', default=0, help='begin epoch')
# 输出结果保存路径
parser.add_argument('--outf', default='./model/', help='folder to output images and model checkpoints')  
# 恢复训练时的模型路径
parser.add_argument('--pre_model', default=True, help='use pre-model')  
args = parser.parse_args()
# 使用gpu
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

图像处理

transform_train = transforms.Compose([
    # transforms.RandomCrop(32, padding=4),  # 先四周填充0，再把图像随机裁剪成32*32
    #     transforms.Resize(256),
    #     transforms.CenterCrop(224),#从中心开始裁剪
    transforms.RandomHorizontalFlip(),  # 图像一半的概率翻转，一半的概率不翻转
    transforms.ToTensor(),
    # transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)),
    transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225))
])

transform_test = transforms.Compose([
    transforms.ToTensor(),
        # transforms.Resize(256),
        # transforms.CenterCrop(224),#从中心开始裁剪
    # transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)),
    transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225))
])

加载数据

trainset = torchvision.datasets.CIFAR10(root='./data/', train=True, download=False, transform=transform_train)
trainloader = torch.utils.data.DataLoader(trainset, batch_size=128, shuffle=True, num_workers=0)
testset = torchvision.datasets.CIFAR10(root='./data/', train=False, download=False, transform=transform_test)
testloader = torch.utils.data.DataLoader(testset, batch_size=128, shuffle=False, num_workers=0)
# Cifar-10的标签
classes = ('plane', 'car', 'bird', 'cat', 'deer', 'dog', 'frog', 'horse', 'ship', 'truck')

定义模型

import torch
import torch.nn as nn
import torch.nn.functional as F
#定义残差块
class ResidualBlock(nn.Module):
    def __init__(self, inchannel, outchannel, stride):
        super(ResidualBlock, self).__init__()
        self.left = nn.Sequential(
            nn.Conv2d(inchannel, outchannel, kernel_size=3, stride=stride, padding=1, bias=False),
            nn.BatchNorm2d(outchannel),
            nn.ReLU(inplace=True),
            nn.Conv2d(outchannel, outchannel, kernel_size=3, stride=1, padding=1, bias=False),
            nn.BatchNorm2d(outchannel)
        )
        self.shortcut = nn.Sequential()
        if stride != 1 or inchannel != outchannel:
            self.shortcut = nn.Sequential(
                nn.Conv2d(inchannel, outchannel, kernel_size=1, stride=stride, bias=False),
                nn.BatchNorm2d(outchannel)
            )
 
    def forward(self, x):
        out = self.left(x)
        out += self.shortcut(x)
        out = F.relu(out)
        return out
#resnet主体
class ResNet(nn.Module):
    def __init__(self, ResidualBlock, num_classes=10):
        super(ResNet, self).__init__()
        self.inchannel = 64
        self.conv1 = nn.Sequential(
            nn.Conv2d(3, 64, kernel_size=3, stride=1, padding=1, bias=False),
            nn.BatchNorm2d(64),
            nn.ReLU(),
        )
        self.layer1 = self.make_layer(ResidualBlock, 64,  2, stride=1)
        self.layer2 = self.make_layer(ResidualBlock, 128, 2, stride=2)
        self.layer3 = self.make_layer(ResidualBlock, 256, 2, stride=2)
        self.layer4 = self.make_layer(ResidualBlock, 512, 2, stride=2)
        self.fc = nn.Linear(512, num_classes)
 
    def make_layer(self, block, channels, num_blocks, stride):
        strides = [stride] + [1] * (num_blocks - 1)   #strides=[1,1]
        print(strides)
        layers = []
        for stride in strides:
            layers.append(block(self.inchannel, channels, stride))
            self.inchannel = channels
        return nn.Sequential(*layers)
    def forward(self, x):
        out = self.conv1(x)
        out = self.layer1(out)
        out = self.layer2(out)
        out = self.layer3(out)
        out = self.layer4(out)
        out = F.avg_pool2d(out,4)
        out = out.view(out.size(0), -1)
        out = self.fc(out)
        return out
#定义resnet18
def ResNet18():
 
    return ResNet(ResidualBlock)
 
if __name__=='__main__':
 
    model = ResNet18()
    print(model)
    input = torch.randn(1, 3, 32, 32)
    out = model(input)
    print(out.shape)

定义损失函数和优化器

criterion = nn.CrossEntropyLoss()
optimizer = optim.SGD(net.parameters(), lr=args.lr, momentum=0.9,
                      weight_decay=5e-4) 

训练部分

best_test_acc = 0
pre_epoch = args.pre_epoch

if __name__ == "__main__":
    if not os.path.exists('./model'):
        os.mkdir('./model')
    #writer = SummaryWriter(log_dir='./log')
    print("Start Training, VGG-16...")
    with open("acc.txt", "w") as acc_f:
        with open("log.txt", "w") as log_f:
            for epoch in range(pre_epoch, args.epoch):
                print('\nEpoch: %d' % (epoch + 1))
                # 开始训练
                net.train()
                print(net)
                # 总损失
                sum_loss = 0.0
                # 准确率
                accuracy = 0.0
                total = 0.0

                for i, data in enumerate(trainloader):
                    # 准备数据
                    length = len(trainloader)  # 数据大小
                    inputs, labels = data  # 取出数据
                    #inputs, labels = inputs.to(device), labels.to(device)
                    optimizer.zero_grad()  # 梯度初始化为零（因为一个batch的loss关于weight的导数是所有sample的loss关于weight的导数的累加和）
                    inputs, labels = Variable(inputs), Variable(labels)
                    # forward + backward + optimize
                    outputs = net(inputs)  # 前向传播求出预测值
                    loss = criterion(outputs, labels)  # 求loss
                    loss.backward()  # 反向传播求梯度
                    optimizer.step()  # 更新参数

                    # 每一个batch输出对应的损失loss和准确率accuracy
                    sum_loss += loss.item()
                    _, predicted = torch.max(outputs.data, 1)  # 返回每一行中最大值的那个元素，且返回其索引
                    total += labels.size(0)
                    accuracy += predicted.eq(labels.data).cpu().sum()  # 预测值和真实值进行比较，将数据放到cpu上并且求和

                    print('[epoch:%d, iter:%d] Loss: %.03f | Acc: %.3f%% '
                          % (epoch + 1, (i + 1 + epoch * length), sum_loss / (i + 1), 100. * accuracy / total))

                    # 写入日志
                    log_f.write('[epoch:%d, iter:%d] |Loss: %.03f | Acc: %.3f%% '
                                % (epoch + 1, (i + 1 + epoch * length), sum_loss / (i + 1), 100. * accuracy / total))
                    log_f.write('\n')
                    log_f.flush()

                # 每一个训练epoch完成测试准确率
                print("Waiting for test...")
                # 在上下文环境中切断梯度计算，在此模式下，每一步的计算结果中requires_grad都是False，即使input设置为requires_grad=True
                with torch.no_grad():
                    accuracy = 0
                    total = 0
                    for data in testloader:
                        # 开始测试
                        net.eval()

                        images, labels = data
                        #images, labels = images.to(device), labels.to(device)

                        outputs = net(images)

                        _, predicted = torch.max(outputs.data, 1)  # 返回每一行中最大值的那个元素，且返回其索引(得分高的那一类)
                        total += labels.size(0)
                        accuracy += (predicted == labels).sum()

                    # 输出测试准确率
                    print('测试准确率为: %.3f%%' % (100 * accuracy / total))
                    acc = 100. * accuracy / total

                    # 写入tensorboard
                    #writer.add_scalar('accuracy/test', acc, epoch)

                    # 将测试结果写入文件
                    print('Saving model...')
                    torch.save(net.state_dict(), '%s/net_%3d.pth' % (args.outf, epoch + 1))
                    acc_f.write("epoch = %03d, accuracy = %.3f%%" % (epoch + 1, acc))
                    acc_f.write('\n')
                    acc_f.flush()

                    # 记录最佳的测试准确率
                    if acc > best_test_acc:
                        print('Saving Best Model...')
                        # 存储状态
                        state = {
                            'state_dict': net.state_dict(),
                            'acc': acc,
                            'epoch': epoch + 1,

                        }
                        # 没有就创建checkpoint文件夹
                        if not os.path.isdir('checkpoint'):
                            os.mkdir('checkpoint')
                        # best_acc_f = open("best_acc.txt","w")
                        # best_acc_f.write("epoch = %03d, accuracy = %.3f%%" % (epoch + 1, acc))
                        # best_acc_f.close()
                        torch.save(state, './checkpoint/ckpt.t7')
                        best_test_acc = acc

            # 训练结束
            print("Training Finished, Total Epoch = %d" % epoch)

三、训练结果

最终resnet18在测试集上准确率为0.76，resnet50在测试集上准确率为0.77