空间转换器网络教程

作者：Ghassen HAMROUNI

译者：片刻

校验：片刻

https://pytorch.org/tutorials/_images/FSeq.png

在本教程中，您将学习如何使用称为空间变换器网络的视觉注意力机制来扩充网络。 您可以在DeepMind论文中阅读有关空间变换器网络的更多信息。

空间变换器网络是对任何空间变换的可区别关注的概括。空间变换器网络(简称STN）允许神经网络学习如何对输入图像执行空间变换，以增强模型的几何不变性。例如，它可以裁剪感兴趣的区域，缩放并校正图像的方向。这可能是一个有用的机制，因为CNN不会对旋转和缩放以及更一般的仿射变换保持不变。

关于STN的最好的事情之一就是能够将其简单地插入到任何现有的CNN中，而无需进行任何修改。

  1. # License: BSD
  2. # Author: Ghassen Hamrouni
  4. from __future__ import print_function
  5. import torch
  6. import torch.nn as nn
  7. import torch.nn.functional as F
  8. import torch.optim as optim
  9. import torchvision
  10. from torchvision import datasets, transforms
  11. import matplotlib.pyplot as plt
  12. import numpy as np
  14. plt.ion()   # interactive mode

加载数据

在本文中，我们将尝试使用经典的MNIST数据集。使用标准卷积网络和空间变换器网络。

  1. device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
  3. # Training dataset
  4. train_loader = torch.utils.data.DataLoader(
  5.     datasets.MNIST(root='.', train=True, download=True,
  6.                    transform=transforms.Compose([
  7.                        transforms.ToTensor(),
  8.                        transforms.Normalize((0.1307,), (0.3081,))
  9.                    ])), batch_size=64, shuffle=True, num_workers=4)
  10. # Test dataset
  11. test_loader = torch.utils.data.DataLoader(
  12.     datasets.MNIST(root='.', train=False, transform=transforms.Compose([
  13.         transforms.ToTensor(),
  14.         transforms.Normalize((0.1307,), (0.3081,))
  15.     ])), batch_size=64, shuffle=True, num_workers=4)

Out:

  1. Downloading http://yann.lecun.com/exdb/mnist/train-images-idx3-ubyte.gz to ./MNIST/raw/train-images-idx3-ubyte.gz
  2. Extracting ./MNIST/raw/train-images-idx3-ubyte.gz to ./MNIST/raw
  3. Downloading http://yann.lecun.com/exdb/mnist/train-labels-idx1-ubyte.gz to ./MNIST/raw/train-labels-idx1-ubyte.gz
  4. Extracting ./MNIST/raw/train-labels-idx1-ubyte.gz to ./MNIST/raw
  5. Downloading http://yann.lecun.com/exdb/mnist/t10k-images-idx3-ubyte.gz to ./MNIST/raw/t10k-images-idx3-ubyte.gz
  6. Extracting ./MNIST/raw/t10k-images-idx3-ubyte.gz to ./MNIST/raw
  7. Downloading http://yann.lecun.com/exdb/mnist/t10k-labels-idx1-ubyte.gz to ./MNIST/raw/t10k-labels-idx1-ubyte.gz
  8. Extracting ./MNIST/raw/t10k-labels-idx1-ubyte.gz to ./MNIST/raw
  9. Processing...
  10. Done!

描绘空间变换器网络

空间转换器网络可归结为三个主要组成部分：

  • 定位网络是常规的CNN，可以对转换参数进行回归。永远不会从该数据集中显式学习变换，而是网络会自动学习增强全局精度的空间变换。
  • 网格生成器在输入图像中生成与来自输出图像的每个像素相对应的坐标网格。
  • 采样器使用转换的参数，并将其应用于输入图像。

https://pytorch.org/tutorials/_images/stn-arch.png

Note 我们需要包含affine_grid和grid_sample模块的最新版本的PyTorch。

  1. class Net(nn.Module):
  2.     def __init__(self):
  3.         super(Net, self).__init__()
  4.         self.conv1 = nn.Conv2d(1, 10, kernel_size=5)
  5.         self.conv2 = nn.Conv2d(10, 20, kernel_size=5)
  6.         self.conv2_drop = nn.Dropout2d()
  7.         self.fc1 = nn.Linear(320, 50)
  8.         self.fc2 = nn.Linear(50, 10)
  10.         # Spatial transformer localization-network
  11.         self.localization = nn.Sequential(
  12.             nn.Conv2d(1, 8, kernel_size=7),
  13.             nn.MaxPool2d(2, stride=2),
  14.             nn.ReLU(True),
  15.             nn.Conv2d(8, 10, kernel_size=5),
  16.             nn.MaxPool2d(2, stride=2),
  17.             nn.ReLU(True)
  18.         )
  20.         # Regressor for the 3 * 2 affine matrix
  21.         self.fc_loc = nn.Sequential(
  22.             nn.Linear(10 * 3 * 3, 32),
  23.             nn.ReLU(True),
  24.             nn.Linear(32, 3 * 2)
  25.         )
  27.         # Initialize the weights/bias with identity transformation
  28.         self.fc_loc[2].weight.data.zero_()
  29.         self.fc_loc[2].bias.data.copy_(torch.tensor([1, 0, 0, 0, 1, 0], dtype=torch.float))
  31.     # Spatial transformer network forward function
  32.     def stn(self, x):
  33.         xs = self.localization(x)
  34.         xs = xs.view(-1, 10 * 3 * 3)
  35.         theta = self.fc_loc(xs)
  36.         theta = theta.view(-1, 2, 3)
  38.         grid = F.affine_grid(theta, x.size())
  39.         x = F.grid_sample(x, grid)
  41.         return x
  43.     def forward(self, x):
  44.         # transform the input
  45.         x = self.stn(x)
  47.         # Perform the usual forward pass
  48.         x = F.relu(F.max_pool2d(self.conv1(x), 2))
  49.         x = F.relu(F.max_pool2d(self.conv2_drop(self.conv2(x)), 2))
  50.         x = x.view(-1, 320)
  51.         x = F.relu(self.fc1(x))
  52.         x = F.dropout(x, training=self.training)
  53.         x = self.fc2(x)
  54.         return F.log_softmax(x, dim=1)
  57. model = Net().to(device)

训练模式

现在，让我们使用SGD算法训练模型。网络正在以监督方式学习分类任务。同时，该模型以端到端的方式自动学习STN。

  1. optimizer = optim.SGD(model.parameters(), lr=0.01)
  4. def train(epoch):
  5.     model.train()
  6.     for batch_idx, (data, target) in enumerate(train_loader):
  7.         data, target = data.to(device), target.to(device)
  9.         optimizer.zero_grad()
  10.         output = model(data)
  11.         loss = F.nll_loss(output, target)
  12.         loss.backward()
  13.         optimizer.step()
  14.         if batch_idx % 500 == 0:
  15.             print('Train Epoch: {} [{}/{} ({:.0f}%)]\tLoss: {:.6f}'.format(
  16.                 epoch, batch_idx * len(data), len(train_loader.dataset),
  17.                 100. * batch_idx / len(train_loader), loss.item()))
  18. #
  19. # A simple test procedure to measure STN the performances on MNIST.
  20. #
  23. def test():
  24.     with torch.no_grad():
  25.         model.eval()
  26.         test_loss = 0
  27.         correct = 0
  28.         for data, target in test_loader:
  29.             data, target = data.to(device), target.to(device)
  30.             output = model(data)
  32.             # sum up batch loss
  33.             test_loss += F.nll_loss(output, target, size_average=False).item()
  34.             # get the index of the max log-probability
  35.             pred = output.max(1, keepdim=True)[1]
  36.             correct += pred.eq(target.view_as(pred)).sum().item()
  38.         test_loss /= len(test_loader.dataset)
  39.         print('\nTest set: Average loss: {:.4f}, Accuracy: {}/{} ({:.0f}%)\n'
  40.               .format(test_loss, correct, len(test_loader.dataset),
  41.                       100. * correct / len(test_loader.dataset)))

可视化STN结果

现在，我们将检查学习到的视觉注意力机制的结果。 我们定义了一个小的辅助函数，以便在训练时可视化转换。

  1. def convert_image_np(inp):
  2.     """Convert a Tensor to numpy image."""
  3.     inp = inp.numpy().transpose((1, 2, 0))
  4.     mean = np.array([0.485, 0.456, 0.406])
  5.     std = np.array([0.229, 0.224, 0.225])
  6.     inp = std * inp + mean
  7.     inp = np.clip(inp, 0, 1)
  8.     return inp
  10. # We want to visualize the output of the spatial transformers layer
  11. # after the training, we visualize a batch of input images and
  12. # the corresponding transformed batch using STN.
  15. def visualize_stn():
  16.     with torch.no_grad():
  17.         # Get a batch of training data
  18.         data = next(iter(test_loader))[0].to(device)
  20.         input_tensor = data.cpu()
  21.         transformed_input_tensor = model.stn(data).cpu()
  23.         in_grid = convert_image_np(
  24.             torchvision.utils.make_grid(input_tensor))
  26.         out_grid = convert_image_np(
  27.             torchvision.utils.make_grid(transformed_input_tensor))
  29.         # Plot the results side-by-side
  30.         f, axarr = plt.subplots(1, 2)
  31.         axarr[0].imshow(in_grid)
  32.         axarr[0].set_title('Dataset Images')
  34.         axarr[1].imshow(out_grid)
  35.         axarr[1].set_title('Transformed Images')
  37. for epoch in range(1, 20 + 1):
  38.     train(epoch)
  39.     test()
  41. # Visualize the STN transformation on some input batch
  42. visualize_stn()
  44. plt.ioff()
  45. plt.show()

https://pytorch.org/tutorials/_images/sphx_glr_spatial_transformer_tutorial_001.png

Out:

  1. Train Epoch: 1 [0/60000 (0%)]   Loss: 2.290877
  2. Train Epoch: 1 [32000/60000 (53%)]      Loss: 0.910913
  4. Test set: Average loss: 0.2449, Accuracy: 9312/10000 (93%)
  6. Train Epoch: 2 [0/60000 (0%)]   Loss: 0.489534
  7. Train Epoch: 2 [32000/60000 (53%)]      Loss: 0.296471
  9. Test set: Average loss: 0.1443, Accuracy: 9563/10000 (96%)
  11. Train Epoch: 3 [0/60000 (0%)]   Loss: 0.410248
  12. Train Epoch: 3 [32000/60000 (53%)]      Loss: 0.355454
  14. Test set: Average loss: 0.1019, Accuracy: 9687/10000 (97%)
  16. Train Epoch: 4 [0/60000 (0%)]   Loss: 0.217658
  17. Train Epoch: 4 [32000/60000 (53%)]      Loss: 0.185522
  19. Test set: Average loss: 0.0818, Accuracy: 9751/10000 (98%)
  21. Train Epoch: 5 [0/60000 (0%)]   Loss: 0.471464
  22. Train Epoch: 5 [32000/60000 (53%)]      Loss: 0.591574
  24. Test set: Average loss: 0.0770, Accuracy: 9760/10000 (98%)
  26. Train Epoch: 6 [0/60000 (0%)]   Loss: 0.119462
  27. Train Epoch: 6 [32000/60000 (53%)]      Loss: 0.093015
  29. Test set: Average loss: 0.0817, Accuracy: 9744/10000 (97%)
  31. Train Epoch: 7 [0/60000 (0%)]   Loss: 0.074523
  32. Train Epoch: 7 [32000/60000 (53%)]      Loss: 0.414406
  34. Test set: Average loss: 0.0944, Accuracy: 9714/10000 (97%)
  36. Train Epoch: 8 [0/60000 (0%)]   Loss: 0.100317
  37. Train Epoch: 8 [32000/60000 (53%)]      Loss: 0.114539
  39. Test set: Average loss: 0.1519, Accuracy: 9510/10000 (95%)
  41. Train Epoch: 9 [0/60000 (0%)]   Loss: 0.205053
  42. Train Epoch: 9 [32000/60000 (53%)]      Loss: 0.135724
  44. Test set: Average loss: 0.0892, Accuracy: 9749/10000 (97%)
  46. Train Epoch: 10 [0/60000 (0%)]  Loss: 0.213368
  47. Train Epoch: 10 [32000/60000 (53%)]     Loss: 0.208627
  49. Test set: Average loss: 0.0634, Accuracy: 9813/10000 (98%)
  51. Train Epoch: 11 [0/60000 (0%)]  Loss: 0.078725
  52. Train Epoch: 11 [32000/60000 (53%)]     Loss: 0.099131
  54. Test set: Average loss: 0.0580, Accuracy: 9834/10000 (98%)
  56. Train Epoch: 12 [0/60000 (0%)]  Loss: 0.133572
  57. Train Epoch: 12 [32000/60000 (53%)]     Loss: 0.213358
  59. Test set: Average loss: 0.0506, Accuracy: 9854/10000 (99%)
  61. Train Epoch: 13 [0/60000 (0%)]  Loss: 0.289802
  62. Train Epoch: 13 [32000/60000 (53%)]     Loss: 0.165571
  64. Test set: Average loss: 0.0542, Accuracy: 9842/10000 (98%)
  66. Train Epoch: 14 [0/60000 (0%)]  Loss: 0.219281
  67. Train Epoch: 14 [32000/60000 (53%)]     Loss: 0.284233
  69. Test set: Average loss: 0.0505, Accuracy: 9856/10000 (99%)
  71. Train Epoch: 15 [0/60000 (0%)]  Loss: 0.218599
  72. Train Epoch: 15 [32000/60000 (53%)]     Loss: 0.055698
  74. Test set: Average loss: 0.0507, Accuracy: 9848/10000 (98%)
  76. Train Epoch: 16 [0/60000 (0%)]  Loss: 0.048718
  77. Train Epoch: 16 [32000/60000 (53%)]     Loss: 0.093410
  79. Test set: Average loss: 0.0502, Accuracy: 9855/10000 (99%)
  81. Train Epoch: 17 [0/60000 (0%)]  Loss: 0.071185
  82. Train Epoch: 17 [32000/60000 (53%)]     Loss: 0.053381
  84. Test set: Average loss: 0.0587, Accuracy: 9829/10000 (98%)
  86. Train Epoch: 18 [0/60000 (0%)]  Loss: 0.127790
  87. Train Epoch: 18 [32000/60000 (53%)]     Loss: 0.169319
  89. Test set: Average loss: 0.0484, Accuracy: 9863/10000 (99%)
  91. Train Epoch: 19 [0/60000 (0%)]  Loss: 0.224094
  92. Train Epoch: 19 [32000/60000 (53%)]     Loss: 0.175750
  94. Test set: Average loss: 0.0628, Accuracy: 9817/10000 (98%)
  96. Train Epoch: 20 [0/60000 (0%)]  Loss: 0.251131
  97. Train Epoch: 20 [32000/60000 (53%)]     Loss: 0.024119
  99. Test set: Average loss: 0.0445, Accuracy: 9869/10000 (99%)

脚本的总运行时间： (1分钟44.448秒）

Download Python source code: spatial_transformer_tutorial.py

Download Jupyter notebook: spatial_transformer_tutorial.ipynb