32.pdf

    1. import  torch
    2. import  torch.nn as nn
    3. import  torch.nn.functional as F
    4. import  torch.optim as optim
    5. from    torchvision import datasets, transforms
    8. batch_size=200
    9. learning_rate=0.01
    10. epochs=10
    12. train_db = datasets.MNIST('../data', train=True, download=True,
    13.                    transform=transforms.Compose([
    14.                        transforms.ToTensor(),
    15.                        transforms.Normalize((0.1307,), (0.3081,))
    16.                    ]))
    17. train_loader = torch.utils.data.DataLoader(
    18.     train_db,
    19.     batch_size=batch_size, shuffle=True)
    21. test_db = datasets.MNIST('../data', train=False, transform=transforms.Compose([
    22.     transforms.ToTensor(),
    23.     transforms.Normalize((0.1307,), (0.3081,))
    24. ]))
    25. test_loader = torch.utils.data.DataLoader(test_db,
    26.     batch_size=batch_size, shuffle=True)
    29. print('train:', len(train_db), 'test:', len(test_db))
    30. train_db, val_db = torch.utils.data.random_split(train_db, [50000, 10000])
    31. print('db1:', len(train_db), 'db2:', len(val_db))
    32. train_loader = torch.utils.data.DataLoader(
    33.     train_db,
    34.     batch_size=batch_size, shuffle=True)
    35. val_loader = torch.utils.data.DataLoader(
    36.     val_db,
    37.     batch_size=batch_size, shuffle=True)
    42. class MLP(nn.Module):
    44.     def __init__(self):
    45.         super(MLP, self).__init__()
    47.         self.model = nn.Sequential(
    48.             nn.Linear(784, 200),
    49.             nn.LeakyReLU(inplace=True),
    50.             nn.Linear(200, 200),
    51.             nn.LeakyReLU(inplace=True),
    52.             nn.Linear(200, 10),
    53.             nn.LeakyReLU(inplace=True),
    54.         )
    56.     def forward(self, x):
    57.         x = self.model(x)
    59.         return x
    61. device = torch.device('cuda:0')
    62. net = MLP().to(device)
    63. optimizer = optim.SGD(net.parameters(), lr=learning_rate)
    64. criteon = nn.CrossEntropyLoss().to(device)
    66. for epoch in range(epochs):
    68.     for batch_idx, (data, target) in enumerate(train_loader):
    69.         data = data.view(-1, 28*28)
    70.         data, target = data.to(device), target.cuda()
    72.         logits = net(data)
    73.         loss = criteon(logits, target)
    75.         optimizer.zero_grad()
    76.         loss.backward()
    77.         # print(w1.grad.norm(), w2.grad.norm())
    78.         optimizer.step()
    80.         if batch_idx % 100 == 0:
    81.             print('Train Epoch: {} [{}/{} ({:.0f}%)]\tLoss: {:.6f}'.format(
    82.                 epoch, batch_idx * len(data), len(train_loader.dataset),
    83.                        100. * batch_idx / len(train_loader), loss.item()))
    86.     test_loss = 0
    87.     correct = 0
    88.     for data, target in val_loader:
    89.         data = data.view(-1, 28 * 28)
    90.         data, target = data.to(device), target.cuda()
    91.         logits = net(data)
    92.         test_loss += criteon(logits, target).item()
    94.         pred = logits.data.max(1)[1]
    95.         correct += pred.eq(target.data).sum()
    97.     test_loss /= len(val_loader.dataset)
    98.     print('\nVAL set: Average loss: {:.4f}, Accuracy: {}/{} ({:.0f}%)\n'.format(
    99.         test_loss, correct, len(val_loader.dataset),
    100.         100. * correct / len(val_loader.dataset)))
    104. test_loss = 0
    105. correct = 0
    106. for data, target in test_loader:
    107.     data = data.view(-1, 28 * 28)
    108.     data, target = data.to(device), target.cuda()
    109.     logits = net(data)
    110.     test_loss += criteon(logits, target).item()
    112.     pred = logits.data.max(1)[1]
    113.     correct += pred.eq(target.data).sum()
    115. test_loss /= len(test_loader.dataset)
    116. print('\nTest set: Average loss: {:.4f}, Accuracy: {}/{} ({:.0f}%)\n'.format(
    117.     test_loss, correct, len(test_loader.dataset),
    118.     100. * correct / len(test_loader.dataset)))