使用 pytorch 实现手写数字识别（GPU加速）

原文链接: MNIST Handwritten Digit Recognition in PyTorch

Sample Code

import torch
import random
import torchvision
from torch.utils.data import DataLoader
from torch import nn
import torch.nn.functional as F
import matplotlib.pyplot as plt

n_epochs = 10
batch_size_train = 64
batch_size_test = 1000
learning_rate = 0.01
momentum = 0.5
log_interval = 100
torch.backends.cudnn.enabled = False
random_seed = random.randint(1, 1000000)
torch.manual_seed(random_seed)

train_loader = torch.utils.data.DataLoader(
    torchvision.datasets.MNIST('./data/', train=True, download=True,
                               transform=torchvision.transforms.Compose([  # 归一化并转化为Tensor
                                   torchvision.transforms.ToTensor(),
                                   torchvision.transforms.Normalize(
                                       (0.1307,), (0.3081,))
                               ])),
    batch_size=batch_size_train, shuffle=True)
test_loader = torch.utils.data.DataLoader(
    torchvision.datasets.MNIST('./data/', train=False, download=True,
                               transform=torchvision.transforms.Compose([
                                   torchvision.transforms.ToTensor(),
                                   torchvision.transforms.Normalize(
                                       (0.1307,), (0.3081,))
                               ])),
    batch_size=batch_size_test, shuffle=True)


class Net(nn.Module):
    def __init__(self):
        super(Net, self).__init__()
        self.conv1 = nn.Conv2d(1, 10, kernel_size=3)
        self.conv2 = nn.Conv2d(10, 20, kernel_size=3)
        self.conv3 = nn.Conv2d(20, 20, kernel_size=3)
        self.conv4 = nn.Conv2d(20, 30, kernel_size=3)
        self.fc1 = nn.Linear(480, 100)
        self.fc2 = nn.Linear(100, 10)

    def forward(self, x):
        x = F.relu(self.conv1(x))
        x = F.relu(F.max_pool2d(self.conv2(x), 2))
        x = F.relu(self.conv3(x))
        x = F.relu(F.max_pool2d(self.conv4(x), 2))
        x = x.view(-1, 480)
        x = F.relu(self.fc1(x))
        x = F.dropout(x, training=self.training)
        x = self.fc2(x)
        return F.log_softmax(x, dim=1)


network = Net()
network = network.cuda()
optimizer = torch.optim.SGD(network.parameters(), lr=learning_rate, momentum=momentum)

train_dataset_size = len(train_loader.dataset)
train_size = len(train_loader)
test_dataset_size = len(test_loader.dataset)
test_size = len(test_loader)

train_losses = []
train_acc_counter = []
train_acc = []
train_counter = []
test_losses = []
test_acc = []
test_counter = [i * train_dataset_size for i in range(n_epochs + 1)]


def train(epoch):
    network.train()  # 训练时， 调整网络至训练状态，启用 BN 和 Dropout
    correct = 0
    for batch_idx, (data, target) in enumerate(train_loader):
        data = data.cuda()
        target = target.cuda()
        optimizer.zero_grad()  # 梯度归零，保证每个 batch 互不干扰
        output = network(data)
        pred = output.data.max(1, keepdim=True)[1]
        pred = pred.cuda()
        correct += pred.eq(target.data.view_as(pred)).sum()  # 计算正确答案数量
        loss = F.nll_loss(output, target)
        loss.backward()  # 反向传播计算梯度值
        optimizer.step()  # 梯度下降实现参数更新
        if batch_idx % log_interval == 0:
            print('Train Epoch: {} [{}/{} ({:.0f}%)]\tLoss: {:.6f}'.format(epoch, batch_idx * len(data),
                                                                           train_dataset_size,
                                                                           100. * batch_idx / train_size, loss.item()))
            train_losses.append(loss.item())  # 方便后续画图
            train_counter.append((batch_idx * batch_size_train) + ((epoch - 1) * train_dataset_size))
    correct = correct.cpu()
    train_acc_counter.append(epoch * train_dataset_size)
    train_acc.append(100. * correct / train_dataset_size)
    print('Train Epoch {} Accuracy: {}/{} ({:.2f}%)'.format(epoch, correct, train_dataset_size,
                                                            100. * correct / train_dataset_size))

def test():
    network.eval()  # 评价时，禁用 BN 和 Dropout
    test_loss = correct = 0
    with torch.no_grad():  # 保证 requires_grad = False
        for data, target in test_loader:
            data = data.cuda()
            target = target.cuda()
            output = network(data)
            test_loss += F.nll_loss(output, target, reduction='sum').item()
            pred = output.data.max(1, keepdim=True)[1]
            pred = pred.cuda()
            correct += pred.eq(target.data.view_as(pred)).sum()  # 计算正确答案数量
    test_loss /= test_dataset_size
    test_losses.append(test_loss)
    correct = correct.cpu()
    test_acc.append(100. * correct / test_dataset_size)
    print('\nTest set: Avg. loss: {:.4f}, Accuracy: {}/{} ({:.2f}%)\n'.format(
        test_loss, correct, test_dataset_size, 100. * correct / test_dataset_size))


test()
for i in range(1, n_epochs + 1):
    train(i)
    test()

fig = plt.figure()
plt.subplot(1, 2, 1)
plt.tight_layout()
plt.plot(train_counter, train_losses, color='blue')
plt.scatter(test_counter, test_losses, color='red')
plt.legend(['Train Loss', 'Test Loss'], loc='upper right')
plt.xlabel('number of training examples seen')
plt.ylabel('negative log likelihood loss')

plt.subplot(1, 2, 2)
plt.tight_layout()
plt.plot(train_acc_counter, train_acc, color = 'blue')
plt.scatter(test_counter, test_acc, color = 'red')
plt.legend(['Train Accuracy', 'Test Accuracy'], loc='upper right')
plt.xlabel('number of training examples seen')
plt.ylabel('Accuracy')
plt.show()

不同训练器下的训练效果（共 10 个epoch）

只用 Adam:

只用 SGDM:

前 2 个 epoch 用 Adam, 后 8 个 epoch 用 SGDM: