系列文章目录

1. pytorch MNIST数据集无法正常加载的解决办法（ HTTP Error 503: Service Unavailable）
2. Python 手写数字识别的实现(pytorch框架) 超详细版本
3. pytorch 手写数字识别 新网络设计和学习率探索

前言

这是中国科学院大学深度学习的课程作业

在pytorch 手写数字识别 新网络设计和学习率探索中我探索了一个新结构的CNN网络用于手写数字识别，可参看。

分步骤解释

1. 首先导入需要的包

import torch import torchvision import torchvision.transforms as transforms 

2. 设置超参数，每个参数解释见注释

n_epochs = 5 # 模型训练5轮 log_interval = 30 #控制打印频率的，设n = 30*batch_size，即n张图后打印一次进度 DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu") # 根据设备是否支持GPU来选择硬件 size = 32 # 对输入图片进行处理，拉伸为32*32的图片，这是为了复刻手写数字识别的神经网络，其输入为32*32的灰度图像 learn_rate = 0.03 # 学习率 momentum = 0.1 # 动量 

3. 加载数据集（见blog：pytorch集成的数据集无法访问，采用了指定url方法）

!wget www.di.ens.fr/~lelarge/MNIST.tar.gz !tar -zxvf MNIST.tar.gz from torchvision.datasets import MNIST transform = transforms.Compose( [ transforms.Resize(size), transforms.ToTensor(), transforms.Normalize((0.5), (0.5))]) # 正则化处理，相当于z-score trainset = MNIST(root = './', train=True, download=True, transform=transform) trainloader = torch.utils.data.DataLoader(trainset, batch_size=64, shuffle=True, num_workers=2) testset = MNIST(root = './', train=False, download=True, transform=transform) testloader = torch.utils.data.DataLoader(testset, batch_size=1000, shuffle=True, num_workers=2) # classes = ('1', '2', '3', '4', '5', '6', '7', '8', '9', '0') 

这一步需要说明的是，对加载来的图像进行了增强处理（transforms.Compose()），主要是为了防止过拟合、
详细解释推荐博客：https://medium.com/@CinnamonAITaiwan/cnn%E5%85%A5%E9%96%80-%E5%9C%96%E5%83%8F%E5%A2%9E%E5%BC%B7-fa654d36dafc
后面将通过进一步实验探索图像增强对结果的影响

4. 打印测试集的标签和tensor大小

examples = enumerate(testloader) batch_idx, (example_data, example_targets) = next(examples) print(example_targets) print(example_data.shape) 

tensor([3, 0, 9, 7, 6, 1, 4, 9, 3, 1, 9, 6, 1, 2, 4, 2, 6, 0, 1, 4, 7, 3, 7, 7, 7, 1, 6, 6, 9, 0, 8, 6, 9, 2, 8, 3, 7, 0, 3, 5, 1, 1, 5, 6, 1, 6, 8, 6, 5, 2, 5, 1, 4, 8, 8, 1, 4, 2, 1, 8, 6, 2, 4, 9, 3, 0, 7, 5, 2, 2, 4, 8, 7, 4, 9, 2, 2, 7, 7, 8, 1, 7, 4, 8, 7, 8, 9, 1, 5, 5, 4, 8, 0, 5, 4, 9, 5, 1, 1, 3, 9, 4, 9, 6, 6, 7, 8, 0, 8, 1, 6, 6, 5, 6, 3, 6, 2, 2, 1, 3, 9, 2, 1, 3, 3, 0, 2, 6, 3, 6, 2, 9, 2, 9, 4, 9, 8, 7, 8, 7, 6, 0, 6, 4, 6, 0, 3, 9, 5, 3, 6, 9, 4, 0, 3, 2, 7, 0, 8, 7, 2, 3, 4, 7, 8, 1, 5, 5, 4, 9, 1, 5, 1, 2, 9, 8, 0, 4, 7, 9, 8, 8, 3, 5, 1, 1, 0, 5, 8, 3, 1, 2, 6, 8, 3, 3, 4, 3, 0, 4, 4, 7, 5, 4, 9, 1, 3, 9, 9, 4, 9, 0, 7, 1, 0, 4, 0, 3, 7, 3, 4, 8, 4, 6, 5, 1, 4, 1, 8, 8, 6, 5, 1, 0, 9, 9, 7, 5, 6, 9, 1, 8, 7, 3, 3, 9, 8, 8, 3, 2, 7, 8, 1, 4, 2, 9, 8, 9, 4, 9, 2, 4, 4, 3, 5, 1, 5, 9, 1, 1, 7, 9, 5, 5, 3, 1, 5, 6, 9, 2, 8, 4, 3, 7, 1, 7, 2, 3, 1, 6, 6, 8, 8, 0, 1, 3, 9, 0, 3, 0, 9, 7, 7, 9, 2, 0, 8, 3, 9, 5, 9, 6, 5, 6, 4, 4, 6, 0, 3, 3, 1, 2, 0, 5, 6, 6, 1, 0, 1, 4, 8, 8, 3, 3, 5, 0, 5, 4, 7, 9, 1, 1, 7, 4, 2, 8, 6, 9, 7, 6, 7, 2, 3, 9, 3, 7, 9, 2, 8, 3, 0, 8, 7, 3, 2, 6, 8, 8, 0, 7, 2, 9, 5, 7, 1, 7, 9, 1, 9, 4, 5, 3, 7, 8, 3, 9, 7, 0, 8, 8, 7, 8, 8, 3, 8, 3, 6, 8, 2, 9, 6, 1, 5, 4, 7, 0, 8, 8, 9, 5, 9, 4, 1, 0, 4, 4, 6, 2, 9, 7, 1, 0, 3, 3, 0, 2, 5, 7, 3, 7, 7, 7, 4, 0, 8, 7, 3, 1, 2, 4, 8, 0, 1, 3, 9, 7, 7, 3, 5, 1, 3, 2, 3, 2, 2, 2, 6, 6, 8, 7, 8, 6, 0, 4, 4, 1, 3, 7, 4, 6, 8, 2, 3, 9, 4, 6, 9, 4, 2, 6, 1, 9, 1, 9, 0, 3, 6, 3, 6, 9, 6, 3, 7, 1, 4, 8, 4, 4, 6, 6, 2, 4, 1, 0, 2, 0, 4, 1, 3, 2, 0, 5, 2, 5, 1, 6, 9, 4, 5, 9, 0, 0, 3, 1, 9, 6, 2, 8, 3, 9, 3, 1, 6, 4, 1, 4, 4, 8, 6, 7, 1, 8, 5, 4, 6, 4, 8, 6, 8, 5, 9, 4, 2, 4, 2, 9, 3, 2, 4, 5, 9, 9, 7, 4, 3, 7, 4, 9, 1, 9, 6, 4, 3, 8, 3, 5, 9, 9, 3, 2, 0, 7, 6, 3, 3, 1, 8, 8, 8, 0, 4, 3, 9, 7, 2, 0, 7, 9, 0, 3, 7, 2, 6, 1, 6, 3, 9, 0, 6, 4, 9, 5, 9, 8, 7, 1, 9, 7, 8, 2, 0, 7, 1, 3, 9, 4, 4, 2, 1, 4, 9, 9, 7, 0, 1, 6, 9, 8, 4, 2, 6, 0, 4, 5, 1, 2, 9, 5, 1, 9, 4, 0, 7, 5, 1, 2, 5, 2, 9, 9, 4, 7, 6, 8, 9, 4, 7, 8, 9, 6, 9, 6, 9, 8, 1, 9, 3, 2, 7, 7, 5, 8, 8, 3, 2, 5, 8, 9, 1, 0, 1, 9, 1, 3, 6, 1, 4, 1, 4, 5, 3, 0, 5, 4, 8, 3, 5, 9, 6, 0, 6, 5, 8, 3, 8, 1, 1, 9, 2, 7, 9, 7, 3, 4, 2, 8, 2, 2, 6, 5, 9, 7, 9, 7, 0, 2, 0, 2, 6, 4, 9, 8, 4, 7, 4, 9, 6, 1, 2, 7, 6, 3, 0, 2, 1, 7, 4, 5, 8, 6, 6, 1, 5, 6, 2, 4, 2, 8, 3, 6, 7, 5, 7, 7, 9, 4, 0, 3, 1, 6, 6, 2, 3, 1, 6, 9, 2, 9, 0, 9, 3, 3, 5, 9, 9, 1, 3, 0, 7, 6, 8, 3, 8, 2, 5, 7, 8, 6, 2, 5, 1, 8, 2, 4, 3, 5, 0, 8, 8, 1, 1, 6, 9, 4, 9, 4, 5, 2, 1, 7, 2, 8, 9, 7, 8, 8, 4, 9, 0, 5, 8, 6, 4, 2, 4, 2, 6, 0, 9, 9, 1, 1, 8, 5, 0, 5, 5, 9, 9, 2, 6, 3, 6, 5, 7, 4, 2, 0, 0, 4, 3, 0, 7, 9, 5, 8, 2, 8, 2, 5, 1, 7, 7, 0, 1, 9, 9, 7, 4, 4, 5, 6, 4, 4, 3, 5, 9, 2, 6, 0, 0, 7, 6, 7, 3, 2, 3, 9, 8, 2, 1, 8, 3, 2, 6, 6, 9, 2, 4, 1, 8, 7, 0, 7, 2, 2, 6, 0, 3, 5, 6, 2, 8, 4, 4, 0, 1, 6, 4, 3, 3, 1, 6, 2, 9, 9, 0, 4, 9, 8, 0, 0, 4, 1, 9, 5, 5, 5, 8, 2, 7, 7, 8, 0, 4, 9, 8, 5, 7, 1, 4, 1, 1, 9, 0, 6, 1, 8, 4, 3, 1, 6, 4, 1, 0, 1, 6, 0, 5, 5, 6, 9, 1, 3, 5, 1, 9, 0, 4, 8]) torch.Size([1000, 1, 32, 32]) 

5. 展示6张图片看一看

import matplotlib.pyplot as plt fig = plt.figure() for i in range(6): plt.subplot(2,3,i+1) plt.tight_layout() plt.imshow(example_data[i][0], cmap='gray', interpolation='none') plt.title("Ground Truth: {}".format(example_targets[i])) plt.xticks([]) plt.yticks([]) plt.show() 

6. 设计LeNet-5网络
   复现LeNet-5网络，结构图如下：
   
   一些解释
   输入层
   – 3232的图片，也就是相当于1024个神经
   C1层(卷积层)
   – 选择6个55的卷积核，得到6个大小为32-5+1=28的特征图，也就是
   神经的个数为62828=4704
   S2层(下采样层)
   – 每个下抽样节点的4个输入节点求和后取平均(平均池化)，均值乘上一个权重参数加上一个偏置参数作为激活函数的输入，激活函数的输出即是下一层节点的值。池化核大小选择22，得到6个 1414大小特征图
   C3层(卷积层)
   – 用55的卷积核对S2层输出的特征图进行卷积后，得到6张1010新图片，然后将这6张图片相加在一起，然后加一个偏置项b，然后用激活函数进行映射，就可以得到1张1010的特征图。我们希望得到16张1010的 特征图，因此我们就需要参数个数 为 16*(6*(55))=166*(55)个参数
   S4层(下采样层)
   – 对C3的16张1010特征图进行最大池化，池化核大小为22，得到16
   张大小为55的特征图。神经个数已经减少为:1655=400
   C5层(卷积层)
   – 用5*5的卷积核进行卷积，然后我们希望得到120个特征图，特征图 大小为5-5+1=1。神经个数为120
   F6层(全连接层)
   – 有84个节点，该层的训练参数和连接数都是(120+1)x84=10164
   Output层
   – 共有10个节点，分别代表数字0到9，如果节点i的输出值为0，则网
   识别的结果是数字i。

根据上图，首先设置各层的参数。详细见注释

import torch.nn as nn import torch.nn.functional as F import torch.optim as optim class Net(nn.Module): def __init__(self): super(Net, self).__init__() self.conv1 = nn.Conv2d(1, 6, kernel_size=5, stride=1, padding=0, bias=True) # C1层使用单通道，6深度的卷积核，卷积核大小为5 self.max_pool_1 = nn.MaxPool2d(2) # maxpooling从28降到14个像素点，故采用大小为2最大池化 self.conv2 = nn.Conv2d(6, 16, kernel_size=5, stride=1, padding=0, bias=True) # 第二次卷积 self.max_pool_2 = nn.MaxPool2d(2) # 第二次maxpooling self.conv3 = nn.Conv2d(16, 120, kernel_size=5, stride=1, padding=0, bias=F) #第三次卷积，得到120张大小为1的“图像” self.fc1 = nn.Linear(120,84) # 全连接层 self.fc2 = nn.Linear(84,10) # 全连接层 def forward(self, x): # print("0:"+ str(x.size())) # 输出tensor大小 x = F.relu(self.conv1(x)) # 用激活函数处理卷积结果，激活函数结果再在下一步做maxpooling x = self.max_pool_1(x) # print("1:"+ str(x.size())) x = F.relu(self.conv2(x)) # 同上 x = self.max_pool_2(x) # print("2:"+ str(x.size())) x = F.relu(self.conv3(x)) # print("3:"+ str(x.size())) # x = F.dropout(x, training=self.training) x = x.view(-1, 120) # 把120张大小为1的图像当成一个长度为120的一维tensor x = F.relu(self.fc1(x)) x = F.relu(self.fc2(x)) return F.log_softmax(x) 

7. 实例化神经网络

network = Net().to(DEVICE) optimizer = optim.SGD(network.parameters(), lr = learn_rate, momentum=momentum) # 学习率，动量 

8. 设置需要输出参数的存储变量

train_losses = [] train_counter = [] test_losses = [] test_counter = [i*len(trainloader.dataset) for i in range(n_epochs + 1)] #test_losses为横坐标，test_losses为纵坐标 test_acc = [] 

8. 定义训练函数

def train(epoch, device): network.train() # 调用上一步实例化对象network中的方法（该方法包内已经写好） for batch_idx, (data, target) in enumerate(trainloader): # 按batch_size为集合对象进行逐个处理 data, target = data.to(device), target.to(device) # data是图片，target是标签，device是为了有GPU情况下使用GPU加速 optimizer.zero_grad() # 开始进行BP之前将梯度设置为零，因为PyTorch会在随后的BP中累积梯度 output = network(data) loss = F.nll_loss(output, target) # 函数全称是negative log likelihood loss，下面博客有详细解释 # https://blog.csdn.net/weixin_/article/details/ loss.backward() # 根据误差进行BP optimizer.step() if batch_idx % log_interval == 0: # 控制输出频率 print('Train Epoch: {} [{}/{} ({:.0f}%)]\tLoss: {:.6f}'.format( epoch, batch_idx * len(data), len(trainloader.dataset), 100. * batch_idx / len(trainloader), loss.item())) train_losses.append(loss.item()) # 记录并储存train loss train_counter.append( (batch_idx*64) + ((epoch-1)*len(trainloader.dataset))) 

9. 定义测试函数

def test(device): network.eval() test_loss = 0 correct = 0 with torch.no_grad(): for data, target in testloader: data, target = data.to(device), target.to(device) output = network(data) test_loss += F.nll_loss(output, target, size_average=False).item() pred = output.data.max(1, keepdim=True)[1] # 找到概率最大的下标，为预测类别 correct += pred.eq(target.data.view_as(pred)).sum() # x下面都是记录数据用于绘图，不再解释 test_loss /= len(testloader.dataset) test_losses.append(test_loss) print('\nTest set: Avg. loss: {:.4f}, Accuracy: {}/{} ({:.0f}%)\n'.format( test_loss, correct, len(testloader.dataset), 100. * correct / len(testloader.dataset))) test_acc.append(correct / len(testloader.dataset)) 

10. 训练模型，绘制图像
    在这儿我只训练了5轮（为了节约时间),可以增大训练轮数，获得更精确的结果。

test(DEVICE) for epoch in range(1, n_epochs + 1): train(epoch, DEVICE) test(DEVICE) import matplotlib.pyplot as plt fig = plt.figure() plt.plot(train_counter, train_losses, color='blue') # train_losses变化曲线 plt.plot(test_counter, test_acc, color='yellow') # test集accuracy变化曲线 plt.scatter(test_counter, test_losses, color='red') # test集loss散点图 plt.legend(['Train Loss', 'Test Loss'], loc='upper right') plt.xlabel('number of training examples seen') plt.ylabel('negative log likelihood loss and test accuarcy') plt.show() 

结果

11. 测试模型
    使用6张图来测试一下，是否能正常读取

examples = enumerate(testloader) batch_idx, (example_data, example_targets) = next(examples) with torch.no_grad(): example_data, example_targets = example_data.to(DEVICE), example_targets.to(DEVICE) output = network(example_data) fig = plt.figure() for i in range(6): plt.subplot(2,3,i+1) plt.tight_layout() plt.imshow(example_data[i][0].cpu().clone().numpy(), cmap='gray', interpolation='none') plt.title("Prediction: {}".format( output.data.max(1, keepdim=True)[1][i].item())) plt.xticks([]) plt.yticks([]) plt.show() 

colab文件链接

https://colab.research.google.com/drive/14yVmcQTSLrWm5Fwm7zvw8zK_AODYoxrQ?usp=sharing