MNIST数据集手写数字识别

1 数据集介绍

MNIST 包括6万张28x28的训练样本，1万张测试样本，很多教程都会对它”下手”几乎成为一个 “典范”，可以说它就是计算机视觉里面的Hello World。所以我们这里也会使用MNIST来进行实战。

前面在介绍卷积神经网络的时候说到过LeNet-5，LeNet-5之所以强大就是因为在当时的环境下将MNIST数据的识别率提高到了99%，这里我们也自己从头搭建一个卷积神经网络，也达到99%的准确率

2 手写数字识别

首先，我们定义一些超参数

BATCH_SIZE=512 #大概需要2G的显存
EPOCHS=20 # 总共训练批次
DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu") # 让torch判断是否使用GPU，建议使用GPU环境，因为会快很多

因为Pytorch里面包含了MNIST的数据集，所以我们这里直接使用即可。如果第一次执行会生成data文件夹，并且需要一些时间下载，如果以前下载过就不会再次下载了

由于官方已经实现了dataset，所以这里可以直接使用DataLoader来对数据进行读取

train_loader = torch.utils.data.DataLoader(datasets.MNIST('data', train=True, download=True, transform=transforms.Compose([transforms.ToTensor(),transforms.Normalize((0.1307,), (0.3081,))])),batch_size=BATCH_SIZE, shuffle=True)

Downloading http://yann.lecun.com/exdb/mnist/train-images-idx3-ubyte.gz
Downloading http://yann.lecun.com/exdb/mnist/train-labels-idx1-ubyte.gz
Downloading http://yann.lecun.com/exdb/mnist/t10k-images-idx3-ubyte.gz
Downloading http://yann.lecun.com/exdb/mnist/t10k-labels-idx1-ubyte.gz
Processing...
Done!

测试集

test_loader = torch.utils.data.DataLoader(datasets.MNIST('data', train=False, transform=transforms.Compose([transforms.ToTensor(),transforms.Normalize((0.1307,), (0.3081,))])),batch_size=BATCH_SIZE, shuffle=True)

下面我们定义一个网络，网络包含两个卷积层，conv1和conv2，然后紧接着两个线性层作为输出，最后输出10个维度，这10个维度我们作为0-9的标识来确定识别出的是那个数字

在这里建议大家将每一层的输入和输出维度都作为注释标注出来，这样后面阅读代码的会方便很多

class ConvNet(nn.Module):def __init__(self):super().__init__()# batch*1*28*28（每次会送入batch个样本，输入通道数1（黑白图像），图像分辨率是28x28）# 下面的卷积层Conv2d的第一个参数指输入通道数，第二个参数指输出通道数，第三个参数指卷积核的大小self.conv1 = nn.Conv2d(1, 10, 5) # 输入通道数1，输出通道数10，核的大小5self.conv2 = nn.Conv2d(10, 20, 3) # 输入通道数10，输出通道数20，核的大小3# 下面的全连接层Linear的第一个参数指输入通道数，第二个参数指输出通道数self.fc1 = nn.Linear(20*10*10, 500) # 输入通道数是2000，输出通道数是500self.fc2 = nn.Linear(500, 10) # 输入通道数是500，输出通道数是10，即10分类def forward(self,x):in_size = x.size(0) # 在本例中in_size=512，也就是BATCH_SIZE的值。输入的x可以看成是512*1*28*28的张量。out = self.conv1(x) # batch*1*28*28 -> batch*10*24*24（28x28的图像经过一次核为5x5的卷积，输出变为24x24）out = F.relu(out) # batch*10*24*24（激活函数ReLU不改变形状））out = F.max_pool2d(out, 2, 2) # batch*10*24*24 -> batch*10*12*12（2*2的池化层会减半）out = self.conv2(out) # batch*10*12*12 -> batch*20*10*10（再卷积一次，核的大小是3）out = F.relu(out) # batch*20*10*10out = out.view(in_size, -1) # batch*20*10*10 -> batch*2000（out的第二维是-1，说明是自动推算，本例中第二维是20*10*10）out = self.fc1(out) # batch*2000 -> batch*500out = F.relu(out) # batch*500out = self.fc2(out) # batch*500 -> batch*10out = F.log_softmax(out, dim=1) # 计算log(softmax(x))return out

我们实例化一个网络，实例化后使用.to方法将网络移动到GPU

优化器我们也直接选择简单暴力的Adam

model = ConvNet().to(DEVICE)
optimizer = optim.Adam(model.parameters())

下面定义一下训练的函数，我们将训练的所有操作都封装到这个函数中

def train(model, device, train_loader, optimizer, epoch):model.train()for batch_idx, (data, target) in enumerate(train_loader):data, target = data.to(device), target.to(device)optimizer.zero_grad()output = model(data)loss = F.nll_loss(output, target)loss.backward()optimizer.step()if(batch_idx+1)%30 == 0: print('Train Epoch: {} [{}/{} ({:.0f}%)]\tLoss: {:.6f}'.format(epoch, batch_idx * len(data), len(train_loader.dataset),100. * batch_idx / len(train_loader), loss.item()))

测试的操作也一样封装成一个函数

def test(model, device, test_loader):model.eval()test_loss = 0correct = 0with torch.no_grad():for data, target in test_loader:data, target = data.to(device), target.to(device)output = model(data)test_loss += F.nll_loss(output, target, reduction='sum').item() # 将一批的损失相加pred = output.max(1, keepdim=True)[1] # 找到概率最大的下标correct += pred.eq(target.view_as(pred)).sum().item()test_loss /= len(test_loader.dataset)print('\nTest set: Average loss: {:.4f}, Accuracy: {}/{} ({:.0f}%)\n'.format(test_loss, correct, len(test_loader.dataset),100. * correct / len(test_loader.dataset)))

下面开始训练，这里就体现出封装起来的好处了，只要写两行就可以了

for epoch in range(1, EPOCHS + 1):train(model, DEVICE, train_loader, optimizer, epoch)test(model, DEVICE, test_loader)

Train Epoch: 1 [14848/60000 (25%)] Loss: 0.272529
Train Epoch: 1 [30208/60000 (50%)]  Loss: 0.235455
Train Epoch: 1 [45568/60000 (75%)]  Loss: 0.101858Test set: Average loss: 0.1018, Accuracy: 9695/10000 (97%)Train Epoch: 2 [14848/60000 (25%)]  Loss: 0.057989
Train Epoch: 2 [30208/60000 (50%)]  Loss: 0.083935
Train Epoch: 2 [45568/60000 (75%)]  Loss: 0.051921Test set: Average loss: 0.0523, Accuracy: 9825/10000 (98%)Train Epoch: 3 [14848/60000 (25%)]  Loss: 0.045383
Train Epoch: 3 [30208/60000 (50%)]  Loss: 0.049402
Train Epoch: 3 [45568/60000 (75%)]  Loss: 0.061366Test set: Average loss: 0.0408, Accuracy: 9866/10000 (99%)Train Epoch: 4 [14848/60000 (25%)]  Loss: 0.035253
Train Epoch: 4 [30208/60000 (50%)]  Loss: 0.038444
Train Epoch: 4 [45568/60000 (75%)]  Loss: 0.036877Test set: Average loss: 0.0433, Accuracy: 9859/10000 (99%)Train Epoch: 5 [14848/60000 (25%)]  Loss: 0.038996
Train Epoch: 5 [30208/60000 (50%)]  Loss: 0.020670
Train Epoch: 5 [45568/60000 (75%)]  Loss: 0.034658Test set: Average loss: 0.0339, Accuracy: 9885/10000 (99%)Train Epoch: 6 [14848/60000 (25%)]  Loss: 0.067320
Train Epoch: 6 [30208/60000 (50%)]  Loss: 0.016328
Train Epoch: 6 [45568/60000 (75%)]  Loss: 0.017037Test set: Average loss: 0.0348, Accuracy: 9881/10000 (99%)Train Epoch: 7 [14848/60000 (25%)]  Loss: 0.022150
Train Epoch: 7 [30208/60000 (50%)]  Loss: 0.009608
Train Epoch: 7 [45568/60000 (75%)]  Loss: 0.012742Test set: Average loss: 0.0346, Accuracy: 9895/10000 (99%)Train Epoch: 8 [14848/60000 (25%)]  Loss: 0.010173
Train Epoch: 8 [30208/60000 (50%)]  Loss: 0.019482
Train Epoch: 8 [45568/60000 (75%)]  Loss: 0.012159Test set: Average loss: 0.0323, Accuracy: 9886/10000 (99%)Train Epoch: 9 [14848/60000 (25%)]  Loss: 0.007792
Train Epoch: 9 [30208/60000 (50%)]  Loss: 0.006970
Train Epoch: 9 [45568/60000 (75%)]  Loss: 0.004989Test set: Average loss: 0.0294, Accuracy: 9909/10000 (99%)Train Epoch: 10 [14848/60000 (25%)] Loss: 0.003764
Train Epoch: 10 [30208/60000 (50%)] Loss: 0.005944
Train Epoch: 10 [45568/60000 (75%)] Loss: 0.001866Test set: Average loss: 0.0361, Accuracy: 9902/10000 (99%)Train Epoch: 11 [14848/60000 (25%)] Loss: 0.002737
Train Epoch: 11 [30208/60000 (50%)] Loss: 0.014134
Train Epoch: 11 [45568/60000 (75%)] Loss: 0.001365Test set: Average loss: 0.0309, Accuracy: 9905/10000 (99%)Train Epoch: 12 [14848/60000 (25%)] Loss: 0.003344
Train Epoch: 12 [30208/60000 (50%)] Loss: 0.003090
Train Epoch: 12 [45568/60000 (75%)] Loss: 0.004847Test set: Average loss: 0.0318, Accuracy: 9902/10000 (99%)Train Epoch: 13 [14848/60000 (25%)] Loss: 0.001278
Train Epoch: 13 [30208/60000 (50%)] Loss: 0.003016
Train Epoch: 13 [45568/60000 (75%)] Loss: 0.001328Test set: Average loss: 0.0358, Accuracy: 9906/10000 (99%)Train Epoch: 14 [14848/60000 (25%)] Loss: 0.002219
Train Epoch: 14 [30208/60000 (50%)] Loss: 0.003487
Train Epoch: 14 [45568/60000 (75%)] Loss: 0.014429Test set: Average loss: 0.0376, Accuracy: 9896/10000 (99%)Train Epoch: 15 [14848/60000 (25%)] Loss: 0.003042
Train Epoch: 15 [30208/60000 (50%)] Loss: 0.002974
Train Epoch: 15 [45568/60000 (75%)] Loss: 0.000871Test set: Average loss: 0.0346, Accuracy: 9909/10000 (99%)Train Epoch: 16 [14848/60000 (25%)] Loss: 0.000618
Train Epoch: 16 [30208/60000 (50%)] Loss: 0.003164
Train Epoch: 16 [45568/60000 (75%)] Loss: 0.007245Test set: Average loss: 0.0357, Accuracy: 9905/10000 (99%)Train Epoch: 17 [14848/60000 (25%)] Loss: 0.001874
Train Epoch: 17 [30208/60000 (50%)] Loss: 0.013951
Train Epoch: 17 [45568/60000 (75%)] Loss: 0.000729Test set: Average loss: 0.0322, Accuracy: 9922/10000 (99%)Train Epoch: 18 [14848/60000 (25%)] Loss: 0.002581
Train Epoch: 18 [30208/60000 (50%)] Loss: 0.001396
Train Epoch: 18 [45568/60000 (75%)] Loss: 0.015521Test set: Average loss: 0.0389, Accuracy: 9914/10000 (99%)Train Epoch: 19 [14848/60000 (25%)] Loss: 0.000283
Train Epoch: 19 [30208/60000 (50%)] Loss: 0.001385
Train Epoch: 19 [45568/60000 (75%)] Loss: 0.011184Test set: Average loss: 0.0383, Accuracy: 9901/10000 (99%)Train Epoch: 20 [14848/60000 (25%)] Loss: 0.000472
Train Epoch: 20 [30208/60000 (50%)] Loss: 0.003306
Train Epoch: 20 [45568/60000 (75%)] Loss: 0.018017Test set: Average loss: 0.0393, Accuracy: 9899/10000 (99%)

我们看一下结果，准确率99%，没问题

如果你的模型连MNIST都搞不定，那么你的模型没有任何的价值

即使你的模型搞定了MNIST，你的模型也可能没有任何的价值

MNIST是一个很简单的数据集，由于它的局限性只能作为研究用途，对实际应用带来的价值非常有限。但是通过这个例子，我们可以完全了解一个实际项目的工作流程

我们找到数据集，对数据做预处理，定义我们的模型，调整超参数，测试训练，再通过训练结果对超参数进行调整或者对模型进行调整。

并且通过这个实战我们已经有了一个很好的模板，以后的项目都可以以这个模板为样例