44.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
    7. from visdom import Visdom
    9. batch_size=200
    10. learning_rate=0.01
    11. epochs=10
    13. train_loader = torch.utils.data.DataLoader(
    14.     datasets.MNIST('../data', train=True, download=True,
    15.                    transform=transforms.Compose([
    16.                        transforms.RandomHorizontalFlip(),
    17.                        transforms.RandomVerticalFlip(),
    18.                        transforms.RandomRotation(15),
    19.                        transforms.RandomRotation([90, 180, 270]),
    20.                        transforms.Resize([32, 32]),
    21.                        transforms.RandomCrop([28, 28]),
    22.                        transforms.ToTensor(),
    23.                        # transforms.Normalize((0.1307,), (0.3081,))
    24.                    ])),
    25.     batch_size=batch_size, shuffle=True)
    26. test_loader = torch.utils.data.DataLoader(
    27.     datasets.MNIST('../data', train=False, transform=transforms.Compose([
    28.         transforms.ToTensor(),
    29.         # transforms.Normalize((0.1307,), (0.3081,))
    30.     ])),
    31.     batch_size=batch_size, shuffle=True)
    35. class MLP(nn.Module):
    37.     def __init__(self):
    38.         super(MLP, self).__init__()
    40.         self.model = nn.Sequential(
    41.             nn.Linear(784, 200),
    42.             nn.LeakyReLU(inplace=True),
    43.             nn.Linear(200, 200),
    44.             nn.LeakyReLU(inplace=True),
    45.             nn.Linear(200, 10),
    46.             nn.LeakyReLU(inplace=True),
    47.         )
    49.     def forward(self, x):
    50.         x = self.model(x)
    52.         return x
    54. device = torch.device('cuda:0')
    55. net = MLP().to(device)
    56. optimizer = optim.SGD(net.parameters(), lr=learning_rate)
    57. criteon = nn.CrossEntropyLoss().to(device)
    59. viz = Visdom()
    61. viz.line([0.], [0.], win='train_loss', opts=dict(title='train loss'))
    62. viz.line([[0.0, 0.0]], [0.], win='test', opts=dict(title='test loss&acc.',
    63.                                                    legend=['loss', 'acc.']))
    64. global_step = 0
    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.         global_step += 1
    81.         viz.line([loss.item()], [global_step], win='train_loss', update='append')
    83.         if batch_idx % 100 == 0:
    84.             print('Train Epoch: {} [{}/{} ({:.0f}%)]\tLoss: {:.6f}'.format(
    85.                 epoch, batch_idx * len(data), len(train_loader.dataset),
    86.                        100. * batch_idx / len(train_loader), loss.item()))
    89.     test_loss = 0
    90.     correct = 0
    91.     for data, target in test_loader:
    92.         data = data.view(-1, 28 * 28)
    93.         data, target = data.to(device), target.cuda()
    94.         logits = net(data)
    95.         test_loss += criteon(logits, target).item()
    97.         pred = logits.argmax(dim=1)
    98.         correct += pred.eq(target).float().sum().item()
    100.     viz.line([[test_loss, correct / len(test_loader.dataset)]],
    101.              [global_step], win='test', update='append')
    102.     viz.images(data.view(-1, 1, 28, 28), win='x')
    103.     viz.text(str(pred.detach().cpu().numpy()), win='pred',
    104.              opts=dict(title='pred'))
    106.     test_loss /= len(test_loader.dataset)
    107.     print('\nTest set: Average loss: {:.4f}, Accuracy: {}/{} ({:.0f}%)\n'.format(
    108.         test_loss, correct, len(test_loader.dataset),
    109.         100. * correct / len(test_loader.dataset)))