针对MNIST的简单分类网络

from __future__ import print_function
import argparse
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
from torchvision import datasets, transforms
from torch.optim.lr_scheduler import StepLR
import matplotlib.pyplot as plt
import time
from PIL import Image


# Adjust the model to get a higher performance
class Net(nn.Module):
    def __init__(self):
        super(Net, self).__init__()
        self.conv1 = nn.Conv2d(1, 8, 3, 1)
        self.conv2 = nn.Conv2d(8, 16, 3, 1)
        self.dropout1 = nn.Dropout(0.25)
        self.dropout2 = nn.Dropout(0.5)
        self.fc1 = nn.Linear(2304, 64)
        self.fc2 = nn.Linear(64, 10)

    def forward(self, x):
        x = self.conv1(x)
        x = F.relu(x)
        x = self.conv2(x)
        x = F.relu(x)
        x = F.max_pool2d(x, 2)
        x = self.dropout1(x)
        x = torch.flatten(x, 1)
        x = self.fc1(x)
        x = F.relu(x)
        x = self.dropout2(x)
        x = self.fc2(x)
        return x


class Improved_Net(nn.Module):
    def __init__(self):
        super(Improved_Net, self).__init__()
        self.conv1 = nn.Conv2d(1, 8, 3, 1)
        self.conv2 = nn.Conv2d(8, 16, 3, 1)
        self.conv3 = nn.Conv2d(16, 16, 3, 1)
        self.dropout1 = nn.Dropout(0.25)
        self.dropout2 = nn.Dropout(0.5)
        self.block = nn.Sequential(
                        *[self.conv3 for _ in range(5)]
                    )
        self.fc = nn.Linear(64, 10)
        
    def forward(self, x):
        x = self.conv1(x)
        x = F.relu(x)
        x = self.conv2(x)
        x = F.relu(x)
        x = F.max_pool2d(x, 2)
        x = self.dropout1(x)
        x = self.block(x)
        x = F.relu(x)
        x = self.dropout2(x)
        x = torch.flatten(x, 1)
        x = self.fc(x)
        return x

def train(args, model, device, train_loader, optimizer, epoch):
    model.train()
    plt.figure()
    pic = None
    for batch_idx, (data, target) in enumerate(train_loader):
        if batch_idx in (1,2,3,4,5):
            if batch_idx == 1:
                pic = data[0,0,:,:]
            else:
                pic = torch.cat((pic,data[0,0,:,:]),dim=1)
        data, target = data.to(device), target.to(device)
        optimizer.zero_grad()
        output = model(data)
        loss = F.cross_entropy(output, target)
        # Calculate gradients
        loss.backward()
        # Optimize the parameters according to the calculated gradients
        optimizer.step()
        if batch_idx % args.log_interval == 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()))
            if args.dry_run:
                break
    # plt.imshow(pic.cpu(), cmap='gray')
    # plt.show()


def test(model, device, test_loader):
    start = time.time()
    model.eval()
    test_loss = 0
    correct = 0
    well_cls,mis_cls = 0,0
    with torch.no_grad():
        for data, target in test_loader:
            data, target = data.to(device), target.to(device)
            output = model(data)
            output = nn.Softmax(dim=1)(output)
            test_loss += F.cross_entropy(output, target, reduction='sum').item()  # sum up batch loss
            pred = output.argmax(dim=1, keepdim=True)  # get the index of the max log-probability
            probab = output.max(dim=1,keepdim=True)[0]
            is_well_cls = pred.eq(target.view_as(pred)).tolist()
            for i,k in enumerate(is_well_cls):
                if k[0]==True and well_cls < 5:
                    well_cls += 1
                    img = transforms.ToPILImage()(data[i])
                    save_img_path = 'well_cls_'+str(well_cls)+'.jpg'
                    save_probab_path = 'well_cls_'+str(well_cls)+'.txt'
                    img.save(save_img_path)
                    with open(save_probab_path, 'w') as f:
                        f.write('the digit is predicted to be {}\n'.format(str(pred[i][0].item())))
                        f.write('the digit of ground truth is {}\n'.format(str(target[i].item())))
                        f.write('the corresponding classification confidence is:{}'.format(str(probab[i][0].item())))

                if k[0]==False and mis_cls < 5:
                    mis_cls += 1
                    img = transforms.ToPILImage()(data[i])
                    save_img_path = 'mis_cls_'+str(mis_cls)+'.jpg'
                    save_probab_path = 'mis_cls_'+str(mis_cls)+'.txt'
                    img.save(save_img_path)
                    with open(save_probab_path, 'w') as f:
                        f.write('the digit is predicted to be {}\n'.format(str(pred[i][0].item())))
                        f.write('the digit of ground truth is {}\n'.format(str(target[i].item())))
                        f.write('the corresponding classification confidence is:{}'.format(str(probab[i][0].item())))
            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)))
    
    print('testing 10000 images takes:{:.4f}'.format(time.time()-start))


def main():
    # Training settings
    parser = argparse.ArgumentParser(description='PyTorch MNIST Example')
    parser.add_argument('--batch-size', type=int, default=64, metavar='N',
                        help='input batch size for training (default: 64)')
    parser.add_argument('--test-batch-size', type=int, default=1000, metavar='N',
                        help='input batch size for testing (default: 1000)')
    parser.add_argument('--epochs', type=int, default=14, metavar='N',
                        help='number of epochs to train (default: 14)')
    parser.add_argument('--lr', type=float, default=1.0, metavar='LR',
                        help='learning rate (default: 1.0)')
    parser.add_argument('--gamma', type=float, default=0.7, metavar='M',
                        help='Learning rate step gamma (default: 0.7)')
    parser.add_argument('--no-cuda', action='store_true', default=False,
                        help='disables CUDA training')
    parser.add_argument('--dry-run', action='store_true', default=False,
                        help='quickly check a single pass')
    parser.add_argument('--seed', type=int, default=1, metavar='S',
                        help='random seed (default: 1)')
    parser.add_argument('--log-interval', type=int, default=10, metavar='N',
                        help='how many batches to wait before logging training status')
    parser.add_argument('--save-model', action='store_true', default=False,
                        help='For Saving the current Model')
    parser.add_argument('--pt-path', default='mnist_cnn_original.pt',
                        help='For Loading the saved Model')   
                 
    args = parser.parse_args()
    use_cuda = not args.no_cuda and torch.cuda.is_available()

    torch.manual_seed(args.seed)

    device = torch.device("cuda" if use_cuda else "cpu")

    # batch_size is a crucial hyper-parameter
    train_kwargs = {'batch_size': args.batch_size}
    test_kwargs = {'batch_size': args.test_batch_size}
    if use_cuda:
        # Adjust num worker and pin memory according to your computer performance
        cuda_kwargs = {'num_workers': 1,
                       'pin_memory': True,
                       'shuffle': True}
        train_kwargs.update(cuda_kwargs)
        test_kwargs.update(cuda_kwargs)

    # Normalize the input (black and white image)
    transform=transforms.Compose([
        transforms.ToTensor(),
        transforms.Normalize((0.1307,), (0.3081,))
        ])

    # Make train dataset split
    dataset1 = datasets.MNIST('./data', train=True, download=True,
                       transform=transform)
    # Make test dataset split
    dataset2 = datasets.MNIST('./data', train=False,
                       transform=transform)

    # Convert the dataset to dataloader, including train_kwargs and test_kwargs
    train_loader = torch.utils.data.DataLoader(dataset1,**train_kwargs)
    test_loader = torch.utils.data.DataLoader(dataset2, **test_kwargs)

    # Put the model on the GPU or CPU
    # model = Net().to(device)
    model = Improved_Net().to(device)

    # the number of parameters of model
    total_num = sum(p.numel() for p in model.parameters())
    print('total number of parameters of model:',total_num)

    # Create optimizer
    optimizer = optim.Adadelta(model.parameters(), lr=args.lr)

    # Create a schedule for the optimizer
    scheduler = StepLR(optimizer, step_size=1, gamma=args.gamma)

    # Begin training and testing
    for epoch in range(1, args.epochs + 1):
        train(args, model, device, train_loader, optimizer, epoch)
        test(model, device, test_loader)
        scheduler.step()

    # Save the model
    if args.save_model:
        # torch.save(model.state_dict(), "mnist_cnn_original.pt")
        torch.save(model.state_dict(), "mnist_cnn.pt")


if __name__ == '__main__':
    main()

其中，Net网络在末尾连续使用两个全连接层，而Improved_Net将fc1这一全连接层替换为5个3*3的卷积层。Net的精度稍高，而Improved_Net的模型参数量小很多。根据需求选择想要的网络。（两种网络的分类精度都能达到98%以上）