PyTorch 空間變壓器網(wǎng)絡(luò)教程

2020-09-07 17:25 更新
原文：https://pytorch.org/tutorials/intermediate/spatial_transformer_tutorial.html

作者： Ghassen HAMROUNI

../_images/FSeq.png

在本教程中，您將學(xué)習(xí)如何使用稱(chēng)為空間變換器網(wǎng)絡(luò)的視覺(jué)注意力機(jī)制來(lái)擴(kuò)充網(wǎng)絡(luò)。 您可以在 DeepMind 論文中詳細(xì)了解空間變壓器網(wǎng)絡(luò)。

空間變換器網(wǎng)絡(luò)是對(duì)任何空間變換的可區(qū)別關(guān)注的概括。 空間變換器網(wǎng)絡(luò)(簡(jiǎn)稱(chēng) STN）允許神經(jīng)網(wǎng)絡(luò)學(xué)習(xí)如何對(duì)輸入圖像執(zhí)行空間變換，以增強(qiáng)模型的幾何不變性。 例如，它可以裁剪感興趣的區(qū)域，縮放并校正圖像的方向。 這可能是一個(gè)有用的機(jī)制，因?yàn)?CNN 不會(huì)對(duì)旋轉(zhuǎn)和縮放以及更一般的仿射變換保持不變。

關(guān)于 STN 的最好的事情之一就是能夠?qū)⑺?jiǎn)單地插入到任何現(xiàn)有的 CNN 中。

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

加載數(shù)據(jù)

在本文中，我們將嘗試使用經(jīng)典的 MNIST 數(shù)據(jù)集。 使用標(biāo)準(zhǔn)卷積網(wǎng)絡(luò)和空間變換器網(wǎng)絡(luò)。

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

得出：

  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!

描述空間變壓器網(wǎng)絡(luò)

空間變壓器網(wǎng)絡(luò)可歸結(jié)為三個(gè)主要組成部分：

  • 本地化網(wǎng)絡(luò)是常規(guī)的 CNN，可以對(duì)轉(zhuǎn)換參數(shù)進(jìn)行回歸。 永遠(yuǎn)不會(huì)從此數(shù)據(jù)集中顯式學(xué)習(xí)變換，而是網(wǎng)絡(luò)會(huì)自動(dòng)學(xué)習(xí)增強(qiáng)全局精度的空間變換。
  • 網(wǎng)格生成器在輸入圖像中生成與來(lái)自輸出圖像的每個(gè)像素相對(duì)應(yīng)的坐標(biāo)網(wǎng)格。
  • 采樣器使用轉(zhuǎn)換的參數(shù)，并將其應(yīng)用于輸入圖像。

../_images/stn-arch.png

注意：

我們需要包含 affine_grid 和 grid_sample 模塊的最新版本的 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)
  11.         # Spatial transformer localization-network
  12.         self.localization = nn.Sequential(
  13.             nn.Conv2d(1, 8, kernel_size=7),
  14.             nn.MaxPool2d(2, stride=2),
  15.             nn.ReLU(True),
  16.             nn.Conv2d(8, 10, kernel_size=5),
  17.             nn.MaxPool2d(2, stride=2),
  18.             nn.ReLU(True)
  19.         )
  22.         # Regressor for the 3 * 2 affine matrix
  23.         self.fc_loc = nn.Sequential(
  24.             nn.Linear(10 * 3 * 3, 32),
  25.             nn.ReLU(True),
  26.             nn.Linear(32, 3 * 2)
  27.         )
  30.         # Initialize the weights/bias with identity transformation
  31.         self.fc_loc[2].weight.data.zero_()
  32.         self.fc_loc[2].bias.data.copy_(torch.tensor([1, 0, 0, 0, 1, 0], dtype=torch.float))
  35.     # Spatial transformer network forward function
  36.     def stn(self, x):
  37.         xs = self.localization(x)
  38.         xs = xs.view(-1, 10 * 3 * 3)
  39.         theta = self.fc_loc(xs)
  40.         theta = theta.view(-1, 2, 3)
  43.         grid = F.affine_grid(theta, x.size())
  44.         x = F.grid_sample(x, grid)
  47.         return x
  50.     def forward(self, x):
  51.         # transform the input
  52.         x = self.stn(x)
  55.         # Perform the usual forward pass
  56.         x = F.relu(F.max_pool2d(self.conv1(x), 2))
  57.         x = F.relu(F.max_pool2d(self.conv2_drop(self.conv2(x)), 2))
  58.         x = x.view(-1, 320)
  59.         x = F.relu(self.fc1(x))
  60.         x = F.dropout(x, training=self.training)
  61.         x = self.fc2(x)
  62.         return F.log_softmax(x, dim=1)
  65. model = Net().to(device)

訓(xùn)練模型

現(xiàn)在，讓我們使用 SGD 算法訓(xùn)練模型。 網(wǎng)絡(luò)正在以監(jiān)督方式學(xué)習(xí)分類(lèi)任務(wù)。 同時(shí)，該模型以端到端的方式自動(dòng)學(xué)習(xí) 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)
  10.         optimizer.zero_grad()
  11.         output = model(data)
  12.         loss = F.nll_loss(output, target)
  13.         loss.backward()
  14.         optimizer.step()
  15.         if batch_idx % 500 == 0:
  16.             print('Train Epoch: {} [{}/{} ({:.0f}%)]\tLoss: {:.6f}'.format(
  17.                 epoch, batch_idx * len(data), len(train_loader.dataset),
  18.                 100\. * batch_idx / len(train_loader), loss.item()))
  19. ## ## A simple test procedure to measure STN the performances on MNIST.
  20. ## 
  21. def test():
  22.     with torch.no_grad():
  23.         model.eval()
  24.         test_loss = 0
  25.         correct = 0
  26.         for data, target in test_loader:
  27.             data, target = data.to(device), target.to(device)
  28.             output = model(data)
  31.             # sum up batch loss
  32.             test_loss += F.nll_loss(output, target, size_average=False).item()
  33.             # get the index of the max log-probability
  34.             pred = output.max(1, keepdim=True)[1]
  35.             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)))

可視化 STN 結(jié)果

現(xiàn)在，我們將檢查學(xué)習(xí)到的視覺(jué)注意力機(jī)制的結(jié)果。

我們定義了一個(gè)小的輔助函數(shù)，以便在訓(xùn)練時(shí)可視化轉(zhuǎn)換。

  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
  11. ## We want to visualize the output of the spatial transformers layer
  12. ## after the training, we visualize a batch of input images and
  13. ## the corresponding transformed batch using STN.
  16. def visualize_stn():
  17.     with torch.no_grad():
  18.         # Get a batch of training data
  19.         data = next(iter(test_loader))[0].to(device)
  22.         input_tensor = data.cpu()
  23.         transformed_input_tensor = model.stn(data).cpu()
  26.         in_grid = convert_image_np(
  27.             torchvision.utils.make_grid(input_tensor))
  30.         out_grid = convert_image_np(
  31.             torchvision.utils.make_grid(transformed_input_tensor))
  34.         # Plot the results side-by-side
  35.         f, axarr = plt.subplots(1, 2)
  36.         axarr[0].imshow(in_grid)
  37.         axarr[0].set_title('Dataset Images')
  40.         axarr[1].imshow(out_grid)
  41.         axarr[1].set_title('Transformed Images')
  44. for epoch in range(1, 20 + 1):
  45.     train(epoch)
  46.     test()
  49. ## Visualize the STN transformation on some input batch
  50. visualize_stn()
  53. plt.ioff()
  54. plt.show()

../_images/sphx_glr_spatial_transformer_tutorial_001.png

得出:

  1. Train Epoch: 1 [0/60000 (0%)]   Loss: 2.312544
  2. Train Epoch: 1 [32000/60000 (53%)]      Loss: 0.865688
  5. Test set: Average loss: 0.2105, Accuracy: 9426/10000 (94%)
  8. Train Epoch: 2 [0/60000 (0%)]   Loss: 0.528199
  9. Train Epoch: 2 [32000/60000 (53%)]      Loss: 0.273284
  12. Test set: Average loss: 0.1150, Accuracy: 9661/10000 (97%)
  15. Train Epoch: 3 [0/60000 (0%)]   Loss: 0.312562
  16. Train Epoch: 3 [32000/60000 (53%)]      Loss: 0.496166
  19. Test set: Average loss: 0.1130, Accuracy: 9661/10000 (97%)
  22. Train Epoch: 4 [0/60000 (0%)]   Loss: 0.346181
  23. Train Epoch: 4 [32000/60000 (53%)]      Loss: 0.206084
  26. Test set: Average loss: 0.0875, Accuracy: 9730/10000 (97%)
  29. Train Epoch: 5 [0/60000 (0%)]   Loss: 0.351175
  30. Train Epoch: 5 [32000/60000 (53%)]      Loss: 0.388225
  33. Test set: Average loss: 0.0659, Accuracy: 9802/10000 (98%)
  36. Train Epoch: 6 [0/60000 (0%)]   Loss: 0.122667
  37. Train Epoch: 6 [32000/60000 (53%)]      Loss: 0.258372
  40. Test set: Average loss: 0.0791, Accuracy: 9759/10000 (98%)
  43. Train Epoch: 7 [0/60000 (0%)]   Loss: 0.190197
  44. Train Epoch: 7 [32000/60000 (53%)]      Loss: 0.154468
  47. Test set: Average loss: 0.0647, Accuracy: 9791/10000 (98%)
  50. Train Epoch: 8 [0/60000 (0%)]   Loss: 0.121149
  51. Train Epoch: 8 [32000/60000 (53%)]      Loss: 0.288490
  54. Test set: Average loss: 0.0583, Accuracy: 9821/10000 (98%)
  57. Train Epoch: 9 [0/60000 (0%)]   Loss: 0.244609
  58. Train Epoch: 9 [32000/60000 (53%)]      Loss: 0.023396
  61. Test set: Average loss: 0.0685, Accuracy: 9778/10000 (98%)
  64. Train Epoch: 10 [0/60000 (0%)]  Loss: 0.256878
  65. Train Epoch: 10 [32000/60000 (53%)]     Loss: 0.091626
  68. Test set: Average loss: 0.0684, Accuracy: 9783/10000 (98%)
  71. Train Epoch: 11 [0/60000 (0%)]  Loss: 0.181910
  72. Train Epoch: 11 [32000/60000 (53%)]     Loss: 0.113193
  75. Test set: Average loss: 0.0492, Accuracy: 9856/10000 (99%)
  78. Train Epoch: 12 [0/60000 (0%)]  Loss: 0.081072
  79. Train Epoch: 12 [32000/60000 (53%)]     Loss: 0.082513
  82. Test set: Average loss: 0.0670, Accuracy: 9800/10000 (98%)
  85. Train Epoch: 13 [0/60000 (0%)]  Loss: 0.180748
  86. Train Epoch: 13 [32000/60000 (53%)]     Loss: 0.194512
  89. Test set: Average loss: 0.0439, Accuracy: 9874/10000 (99%)
  92. Train Epoch: 14 [0/60000 (0%)]  Loss: 0.099560
  93. Train Epoch: 14 [32000/60000 (53%)]     Loss: 0.084377
  96. Test set: Average loss: 0.0416, Accuracy: 9880/10000 (99%)
  99. Train Epoch: 15 [0/60000 (0%)]  Loss: 0.070021
  100. Train Epoch: 15 [32000/60000 (53%)]     Loss: 0.241336
  103. Test set: Average loss: 0.0588, Accuracy: 9820/10000 (98%)
  106. Train Epoch: 16 [0/60000 (0%)]  Loss: 0.060536
  107. Train Epoch: 16 [32000/60000 (53%)]     Loss: 0.053016
  110. Test set: Average loss: 0.0405, Accuracy: 9877/10000 (99%)
  113. Train Epoch: 17 [0/60000 (0%)]  Loss: 0.207369
  114. Train Epoch: 17 [32000/60000 (53%)]     Loss: 0.069607
  117. Test set: Average loss: 0.1006, Accuracy: 9685/10000 (97%)
  120. Train Epoch: 18 [0/60000 (0%)]  Loss: 0.127503
  121. Train Epoch: 18 [32000/60000 (53%)]     Loss: 0.070724
  124. Test set: Average loss: 0.0659, Accuracy: 9814/10000 (98%)
  127. Train Epoch: 19 [0/60000 (0%)]  Loss: 0.176861
  128. Train Epoch: 19 [32000/60000 (53%)]     Loss: 0.116980
  131. Test set: Average loss: 0.0413, Accuracy: 9871/10000 (99%)
  134. Train Epoch: 20 [0/60000 (0%)]  Loss: 0.146933
  135. Train Epoch: 20 [32000/60000 (53%)]     Loss: 0.245741
  138. Test set: Average loss: 0.0346, Accuracy: 9892/10000 (99%)

腳本的總運(yùn)行時(shí)間：(2 分鐘 3.339 秒）

Download Python source code: spatial_transformer_tutorial.py Download Jupyter notebook: spatial_transformer_tutorial.ipynb

由獅身人面像畫(huà)廊生成的畫(huà)廊


以上內(nèi)容是否對(duì)您有幫助：
在線(xiàn)筆記
App下載
App下載

掃描二維碼

下載編程獅App

公眾號(hào)
微信公眾號(hào)

編程獅公眾號(hào)