用残差网络实现MNIST数据集手写数字识别

1.残差网络

本文为用带残差块的CNN网络实现MNIST数据集手写数字的识别。

关于残差网络，知乎上有篇文章讲的不错，供参考：详解残差网络

残差网络比起LeNet等简单的神经网络，不同之初在于，多了一个连接线。

左边为基础的CNN结构，右边为带残差的网络结构

残差块是目前网络模型中，一个跟经典、很基础的结构，像DenseNet就是基于残差块来提出的，一个新的网络模型。

2.MNIST数据集

参考笔者的上篇博客：CNN实现MNIST数据集手写数字识别

3.模型结构

Residual Block：残差块

其结构为：

对x做两次卷积后与 x相加，再做激活

4.代码实现（pytorch）

import torch
from torchvision import transforms
from torchvision import datasets
from torch.utils.data import DataLoader
import torch.optim as optim
import torch.nn.functional as F
import matplotlib.pyplot as plt



batch_size = 64

transform = transforms.Compose([
    transforms.ToTensor(),
    transforms.Normalize((0.1307),(0.3081)) #两个参数，平均值和标准差

])

train_dataset = datasets.MNIST(
    root="../dataset/mnist/",
    train= True,
    download= True,
    transform= transform
)

train_loader = DataLoader(train_dataset,
                          shuffle = True,
                          batch_size = batch_size)

test_dataset = datasets.MNIST(
    root="../dataset/mnist/",
    train=False,
    download=True,
    transform=transform
)

test_loder = DataLoader(test_dataset,
                        shuffle = True,
                        batch_size = batch_size)


class ResidualBlock(torch.nn.Module):
    def __init__(self, channels):
        super(ResidualBlock, self).__init__()
        self.channels = channels
        self.conv1 = torch.nn.Conv2d(channels, channels, kernel_size=3, padding=1)
        self.conv2 = torch.nn.Conv2d(channels, channels, kernel_size=3, padding=1)

    def forward(self, x):
        y = F.relu(self.conv1(x))
        y = self.conv2(y)
        return F.relu(x + y)

'''
CLASS torch.nn.Conv2d(in_channels, out_channels, kernel_size, stride=1, padding=0, 
dilation=1, groups=1, bias=True, padding_mode='zeros', device=None, dtype=None)
'''

'''
CLASS torch.nn.MaxPool2d(kernel_size, stride=None, padding=0, 
dilation=1, return_indices=False, ceil_mode=False)
'''


class Net(torch.nn.Module):
    def __init__(self):
        super(Net, self).__init__()
        self.conv1 = torch.nn.Conv2d(in_channels=1, out_channels=16, kernel_size=5)
        self.conv2 = torch.nn.Conv2d(in_channels=16, out_channels=32, kernel_size=5)  # 88 = 24x3 + 16

        self.rblock1 = ResidualBlock(16)
        self.rblock2 = ResidualBlock(32)

        self.maxpooling = torch.nn.MaxPool2d(2)

        # 建议读者在实现时，可以做增加几个全连接层，参考笔者博客：
        #https://blog.csdn.net/t18438605018/article/details/122137737?spm=1001.2014.3001.5501
        self.linear1 = torch.nn.Linear(512, 10)

    def forward(self, x):
        in_size = x.size(0)
        x = self.maxpooling(F.relu(self.conv1(x)))
        x = self.rblock1(x)
        x = self.maxpooling(F.relu(self.conv2(x)))
        x = self.rblock2(x)

        x = x.view(in_size, -1) # Flatten操作
        x = self.linear1(x)
        return x





model = Net()
#有GPU就使用GPU，没有就是用CPU
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model.to(device)

criterion = torch.nn.CrossEntropyLoss()
optimizer = optim.SGD(model.parameters(), lr=0.01, momentum= 0.5)


def train(epoch):
    total = 0
    running_loss = 0.0
    train_loss = 0.0 #记录每次epoch的损失
    accuracy = 0 #记录每次epoch的accuracy
    for batch_id, data in enumerate(train_loader,0):
        inputs, target = data
        inputs, target = inputs.to(device), target.to(device)
        optimizer.zero_grad()
        # forword + backward + update
        outputs = model(inputs)
        loss = criterion(outputs, target)

        _, predicted = torch.max(outputs.data, dim=1)
        accuracy += (predicted == target).sum().item()
        total += target.size(0)

        loss.backward()
        optimizer.step()

        running_loss += loss.item()
        train_loss = running_loss
        #每迭代300次，求一下这三百次迭代的平均
        if batch_id % 300 == 299:
            print('[%d, %5d] loss: %.3f' %(epoch+1, batch_id+1, running_loss / 300))
            running_loss = 0.0
    print('第 %d epoch的 Accuracy on train set: %d %%, Loss on train set: %f' % (epoch + 1, 100 * accuracy / total, train_loss))

    #返回acc和loss
    return 1.0 * accuracy / total, train_loss


def validation(epoch):
    correct = 0
    total = 0
    val_loss = 0.0
    with torch.no_grad():
        for data in test_loder:
            images, target = data
            images, target = images.to(device), target.to(device)
            outputs = model(images)
            loss = criterion(outputs, target)
            val_loss += loss.item()
            _, predicted = torch.max(outputs.data, dim=1)
            total += target.size(0)
            correct += (predicted == target).sum().item()
    print('第 %d epoch的 Accuracy on validation set: %d %%, Loss on validation set: %f' %(epoch+1,100*correct / total, val_loss))

    #返回acc和loss
    return 1.0 * correct / total, val_loss



def draw_in_one(list,epoch):
    # x_axix，train_pn_dis这些都是长度相同的list()
    # 开始画图
    x_axix = [x for x in range(1, epoch+1)] #把ranage转化为list
    train_acc = list[0]
    train_loss = list[1]
    val_acc = list[2]
    val_loss = list[3]
    #sub_axix = filter(lambda x: x % 200 == 0, x_axix)
    plt.title('Result Analysis')
    plt.plot(x_axix, train_acc, color='green', label='training accuracy')
    plt.plot(x_axix, train_loss, color='red', label='training loss')
    plt.plot(x_axix, val_acc, color='skyblue', label='val accuracy')
    plt.plot(x_axix, val_loss, color='blue', label='val loss')
    plt.legend()  # 显示图例
    plt.xlabel('epoch times')
    plt.ylabel('rate')
    plt.show()
    # python 一个折线图绘制多个曲线
if __name__ == '__main__':

    train_loss = []
    train_acc = []

    val_loss = []
    val_acc = []
    epoches = 10
    list = []
    for epoch in range(epoches):
        acc1, loss1 = train(epoch)

        train_loss.append(loss1)
        train_acc.append(acc1)

        acc2, loss2 = validation(epoch)

        val_loss.append(loss2)
        val_acc.append(acc2)
    # 四幅图合并绘制
    list.append(train_acc)
    list.append(train_loss)
    list.append(val_acc)
    list.append(val_loss)
    draw_in_one(list, epoches)

本文代码与CNN实现MNIST数据集手写数字识别代码不同之处，仅在于网络模型换了。其它均未更改。

在验证集上，识别的准确率达到99%。

四条曲线：

控制台输出信息：

E:\anaconda3\envs\pytorch\python.exe D:/PycharmProjects/pytorchProject/ReNet实现手写数字识别.py
[1,   300] loss: 0.593
[1,   600] loss: 0.153
[1,   900] loss: 0.118
第 1 epoch的 Accuracy on train set: 91 %, Loss on train set: 3.547480
第 1 epoch的 Accuracy on validation set: 97 %, Loss on validation set: 12.158820
[2,   300] loss: 0.087
[2,   600] loss: 0.083
[2,   900] loss: 0.081
第 2 epoch的 Accuracy on train set: 97 %, Loss on train set: 2.241397
第 2 epoch的 Accuracy on validation set: 98 %, Loss on validation set: 9.401881
[3,   300] loss: 0.059
[3,   600] loss: 0.060
[3,   900] loss: 0.062
第 3 epoch的 Accuracy on train set: 98 %, Loss on train set: 2.196551
第 3 epoch的 Accuracy on validation set: 98 %, Loss on validation set: 7.038621
[4,   300] loss: 0.051
[4,   600] loss: 0.050
[4,   900] loss: 0.043
第 4 epoch的 Accuracy on train set: 98 %, Loss on train set: 1.987330
第 4 epoch的 Accuracy on validation set: 98 %, Loss on validation set: 6.167475
[5,   300] loss: 0.046
[5,   600] loss: 0.039
[5,   900] loss: 0.038
第 5 epoch的 Accuracy on train set: 98 %, Loss on train set: 1.205675
第 5 epoch的 Accuracy on validation set: 98 %, Loss on validation set: 5.971746
[6,   300] loss: 0.035
[6,   600] loss: 0.039
[6,   900] loss: 0.035
第 6 epoch的 Accuracy on train set: 98 %, Loss on train set: 1.088960
第 6 epoch的 Accuracy on validation set: 98 %, Loss on validation set: 5.528260
[7,   300] loss: 0.029
[7,   600] loss: 0.034
[7,   900] loss: 0.030
第 7 epoch的 Accuracy on train set: 99 %, Loss on train set: 1.450512
第 7 epoch的 Accuracy on validation set: 98 %, Loss on validation set: 5.239810
[8,   300] loss: 0.027
[8,   600] loss: 0.026
[8,   900] loss: 0.026
第 8 epoch的 Accuracy on train set: 99 %, Loss on train set: 1.436349
第 8 epoch的 Accuracy on validation set: 98 %, Loss on validation set: 5.200812
[9,   300] loss: 0.025
[9,   600] loss: 0.025
[9,   900] loss: 0.023
第 9 epoch的 Accuracy on train set: 99 %, Loss on train set: 1.118738
第 9 epoch的 Accuracy on validation set: 98 %, Loss on validation set: 5.235084
[10,   300] loss: 0.025
[10,   600] loss: 0.022
[10,   900] loss: 0.022
第 10 epoch的 Accuracy on train set: 99 %, Loss on train set: 0.706974
第 10 epoch的 Accuracy on validation set: 99 %, Loss on validation set: 4.611128

Process finished with exit code 0

如果本文帮助了你，请给作者一个关注吧！