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。
import torchimport torch.nn as nnimport torch.nn.functional as Fimport torch.optim as optimfrom torchvision import datasets, transforms
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")
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。
# 下载训练集train_loader = torch.utils.data.DataLoader(datasets.MNIST('data', train = True, download = True,transform = pose([transforms.ToTensor(),transforms.Normalize((0.1037,), (0.3081,))])),batch_size = BATCH_SIZE, shuffle = True)# 测试集test_loader = torch.utils.data.DataLoader(datasets.MNIST('data', train = False, transform = pose([transforms.ToTensor(),transforms.Normalize((0.1037,), (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。
# 定义模型class ConvNet(nn.Module):def __init__(self):super().__init__()#1*1*28*28self.conv1 = nn.Conv2d(1, 10, 5) self.conv2 = nn.Conv2d(10, 20, 3) self.fc1 = nn.Linear(20 * 10 * 10, 500)self.fc2 = nn.Linear(500, 10)def forward(self, x):in_size = x.size(0)out= self.conv1(x) # 1* 10 * 24 *24out = F.relu(out)out = F.max_pool2d(out, 2, 2) # 1* 10 * 12 * 12out = self.conv2(out) # 1* 20 * 10 * 10out = F.relu(out)out = out.view(in_size, -1) # 1 * 2000out = self.fc1(out) # 1 * 500out = F.relu(out)out = self.fc2(out) # 1 * 10out = F.log_softmax(out, dim = 1)return out
#生成模型和优化器model = ConvNet().to(DEVICE)optimizer = optim.Adam(model.parameters())
# 定义训练函数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()))
# 定义测试函数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') # 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)))
# 最后开始训练和测试for epoch in range(1, EPOCHS + 1):train(model, DEVICE, train_loader, optimizer, epoch)test(model, DEVICE, test_loader)
Train Epoch: 20 [14848/60000 (25%)] Loss: 0.000499Train Epoch: 20 [30208/60000 (50%)] Loss: 0.002818
Train Epoch: 20 [45568/60000 (75%)] Loss: 0.011781
Test set: Average loss: 0.0249, Accuracy: 59469/60000 (99%)
