空间变换网络教程
作者: Ghassen HAMROUNI
在本教程中，您将学习如何使用一种称为空间变换网络（spatial transformer networks）的视觉注意力机制来增强您的网络。您可以在DeepMind论文中了解更多关于空间变换网络的信息。
空间变换网络是可微分注意力机制在任意空间变换上的泛化。空间变换网络（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。由于卷积神经网络（CNN）对旋转、缩放以及更一般的仿射变换不具备不变性，因此空间变换网络可以作为一种有用的机制。
STN 最棒的一点是，它能够以极少的修改直接集成到任何现有的 CNN 中。
# License: BSD
# Author: Ghassen Hamrouni
importtorch
importtorch.nnasnn
importtorch.nn.functionalasF
importtorch.optimasoptim
importtorchvision
fromtorchvisionimport datasets, transforms
importmatplotlib.pyplotasplt
importnumpyasnp
plt.ion() # interactive mode
<contextlib.ExitStack object at 0x7fc7aeef67a0>
加载数据
在这篇文章中，我们使用经典的MNIST数据集进行实验。我们采用了一个标准的卷积网络，并增加了空间变换网络来增强其性能。
fromsix.movesimport urllib
opener = urllib.request.build_opener()
opener.addheaders = [('User-agent', 'Mozilla/5.0')]
urllib.request.install_opener(opener)
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
# Training dataset
train_loader = torch.utils.data.DataLoader(
datasets.MNIST(root='.', train=True, download=True,
transform=transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.1307,), (0.3081,))
])), batch_size=64, shuffle=True, num_workers=4)
# Test dataset
test_loader = torch.utils.data.DataLoader(
datasets.MNIST(root='.', train=False, transform=transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.1307,), (0.3081,))
])), batch_size=64, shuffle=True, num_workers=4)
0%| | 0.00/9.91M [00:00<?, ?B/s]
100%|##########| 9.91M/9.91M [00:00<00:00, 138MB/s]
0%| | 0.00/28.9k [00:00<?, ?B/s]
100%|##########| 28.9k/28.9k [00:00<00:00, 32.1MB/s]
0%| | 0.00/1.65M [00:00<?, ?B/s]
100%|##########| 1.65M/1.65M [00:00<00:00, 366MB/s]
0%| | 0.00/4.54k [00:00<?, ?B/s]
100%|##########| 4.54k/4.54k [00:00<00:00, 17.2MB/s]
空间变换网络的描述
空间变换网络可以归结为三个主要组件：
-
定位网络是一个普通的卷积神经网络（CNN），用于回归变换参数。变换并不是从数据集中显式学习的，而是网络自动学习能够提高全局准确性的空间变换。
-
网格生成器在输入图像中生成一个坐标网格，该网格对应于输出图像中的每个像素。
-
采样器使用变换参数并将其应用到输入图像上。
我们需要包含 affine_grid 和 grid_sample 模块的最新版本 PyTorch。
classNet(nn.Module):
def__init__(self):
super(Net, self).__init__()
self.conv1 = nn.Conv2d(1, 10, kernel_size=5)
self.conv2 = nn.Conv2d(10, 20, kernel_size=5)
self.conv2_drop = nn.Dropout2d()
self.fc1 = nn.Linear(320, 50)
self.fc2 = nn.Linear(50, 10)
# Spatial transformer localization-network
self.localization = nn.Sequential(
nn.Conv2d(1, 8, kernel_size=7),
nn.MaxPool2d(2, stride=2),
nn.ReLU(True),
nn.Conv2d(8, 10, kernel_size=5),
nn.MaxPool2d(2, stride=2),
nn.ReLU(True)
)
# Regressor for the 3 * 2 affine matrix
self.fc_loc = nn.Sequential(
nn.Linear(10 * 3 * 3, 32),
nn.ReLU(True),
nn.Linear(32, 3 * 2)
)
# Initialize the weights/bias with identity transformation
self.fc_loc[2].weight.data.zero_()
self.fc_loc[2].bias.data.copy_(torch.tensor([1, 0, 0, 0, 1, 0], dtype=torch.float))
# Spatial transformer network forward function
defstn(self, x):
xs = self.localization(x)
xs = xs.view(-1, 10 * 3 * 3)
theta = self.fc_loc(xs)
theta = theta.view(-1, 2, 3)
grid = F.affine_grid(theta, x.size())
x = F.grid_sample(x, grid)
return x
defforward(self, x):
# transform the input
x = self.stn(x)
# Perform the usual forward pass
x = F.relu(F.max_pool2d(self.conv1(x), 2))
x = F.relu(F.max_pool2d(self.conv2_drop(self.conv2(x)), 2))
x = x.view(-1, 320)
x = F.relu(self.fc1(x))
x = F.dropout(x, training=self.training)
x = self.fc2(x)
return F.log_softmax(x, dim=1)
model = Net().to(device)
训练模型
现在，让我们使用 SGD 算法来训练模型。网络正在以监督的方式学习分类任务。同时，模型以端到端的方式自动学习 STN。
optimizer = optim.SGD(model.parameters(), lr=0.01)
deftrain(epoch):
model.train()
for batch_idx, (data, target) in enumerate(train_loader):
data, target = data.to(device), target.to(device)
optimizer.zero_grad()
output = model(data)
loss = F.nll_loss(output, target)
loss.backward()
optimizer.step()
if batch_idx % 500 == 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()))
#
# A simple test procedure to measure the STN performances on MNIST.
#
deftest():
with torch.no_grad():
model.eval()
test_loss = 0
correct = 0
for data, target in test_loader:
data, target = data.to(device), target.to(device)
output = model(data)
# sum up batch loss
test_loss += F.nll_loss(output, target, size_average=False).item()
# get the index of the max log-probability
pred = output.max(1, keepdim=True)[1]
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)))
可视化 STN 结果
现在，我们将检查我们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，以便在训练过程中可视化这些变换。
defconvert_image_np(inp):
"""Convert a Tensor to numpy image."""
inp = inp.numpy().transpose((1, 2, 0))
mean = np.array([0.485, 0.456, 0.406])
std = np.array([0.229, 0.224, 0.225])
inp = std * inp + mean
inp = np.clip(inp, 0, 1)
return inp
# We want to visualize the output of the spatial transformers layer
# after the training, we visualize a batch of input images and
# the corresponding transformed batch using STN.
defvisualize_stn():
with torch.no_grad():
# Get a batch of training data
data = next(iter(test_loader))[0].to(device)
input_tensor = data.cpu()
transformed_input_tensor = model.stn(data).cpu()
in_grid = convert_image_np(
torchvision.utils.make_grid(input_tensor))
out_grid = convert_image_np(
torchvision.utils.make_grid(transformed_input_tensor))
# Plot the results side-by-side
f, axarr = plt.subplots(1, 2)
axarr[0].imshow(in_grid)
axarr[0].set_title('Dataset Images')
axarr[1].imshow(out_grid)
axarr[1].set_title('Transformed Images')
for epoch in range(1, 20 + 1):
train(epoch)
test()
# Visualize the STN transformation on some input batch
visualize_stn()
plt.ioff()
plt.show()
/usr/local/lib/python3.10/dist-packages/torch/nn/functional.py:5082: UserWarning:
Default grid_sample and affine_grid behavior has changed to align_corners=False since 1.3.0. Please specify align_corners=True if the old behavior is desired. See the documentation of grid_sample for details.
/usr/local/lib/python3.10/dist-packages/torch/nn/functional.py:5015: UserWarning:
Default grid_sample and affine_grid behavior has changed to align_corners=False since 1.3.0. Please specify align_corners=True if the old behavior is desired. See the documentation of grid_sample for details.
Train Epoch: 1 [0/60000 (0%)] Loss: 2.315648
Train Epoch: 1 [32000/60000 (53%)] Loss: 1.073955
/usr/local/lib/python3.10/dist-packages/torch/nn/_reduction.py:51: UserWarning:
size_average and reduce args will be deprecated, please use reduction='sum' instead.
Test set: Average loss: 0.2694, Accuracy: 9233/10000 (92%)
Train Epoch: 2 [0/60000 (0%)] Loss: 0.537731
Train Epoch: 2 [32000/60000 (53%)] Loss: 0.396136
Test set: Average loss: 0.1439, Accuracy: 9595/10000 (96%)
Train Epoch: 3 [0/60000 (0%)] Loss: 0.323309
Train Epoch: 3 [32000/60000 (53%)] Loss: 0.227351
Test set: Average loss: 0.1749, Accuracy: 9418/10000 (94%)
Train Epoch: 4 [0/60000 (0%)] Loss: 0.522717
Train Epoch: 4 [32000/60000 (53%)] Loss: 0.207835
Test set: Average loss: 0.1005, Accuracy: 9713/10000 (97%)
Train Epoch: 5 [0/60000 (0%)] Loss: 0.218420
Train Epoch: 5 [32000/60000 (53%)] Loss: 0.179662
Test set: Average loss: 0.1658, Accuracy: 9498/10000 (95%)
Train Epoch: 6 [0/60000 (0%)] Loss: 0.298722
Train Epoch: 6 [32000/60000 (53%)] Loss: 0.103238
Test set: Average loss: 0.0810, Accuracy: 9764/10000 (98%)
Train Epoch: 7 [0/60000 (0%)] Loss: 0.059216
Train Epoch: 7 [32000/60000 (53%)] Loss: 0.151852
Test set: Average loss: 0.0685, Accuracy: 9800/10000 (98%)
Train Epoch: 8 [0/60000 (0%)] Loss: 0.235398
Train Epoch: 8 [32000/60000 (53%)] Loss: 0.124425
Test set: Average loss: 0.1000, Accuracy: 9702/10000 (97%)
Train Epoch: 9 [0/60000 (0%)] Loss: 0.099140
Train Epoch: 9 [32000/60000 (53%)] Loss: 0.110606
Test set: Average loss: 0.0861, Accuracy: 9738/10000 (97%)
Train Epoch: 10 [0/60000 (0%)] Loss: 0.116240
Train Epoch: 10 [32000/60000 (53%)] Loss: 0.265121
Test set: Average loss: 0.0826, Accuracy: 9772/10000 (98%)
Train Epoch: 11 [0/60000 (0%)] Loss: 0.215463
Train Epoch: 11 [32000/60000 (53%)] Loss: 0.065376
Test set: Average loss: 0.0635, Accuracy: 9819/10000 (98%)
Train Epoch: 12 [0/60000 (0%)] Loss: 0.147400
Train Epoch: 12 [32000/60000 (53%)] Loss: 0.126500
Test set: Average loss: 0.0558, Accuracy: 9846/10000 (98%)
Train Epoch: 13 [0/60000 (0%)] Loss: 0.101385
Train Epoch: 13 [32000/60000 (53%)] Loss: 0.099800
Test set: Average loss: 0.0617, Accuracy: 9828/10000 (98%)
Train Epoch: 14 [0/60000 (0%)] Loss: 0.070670
Train Epoch: 14 [32000/60000 (53%)] Loss: 0.140083
Test set: Average loss: 0.0527, Accuracy: 9845/10000 (98%)
Train Epoch: 15 [0/60000 (0%)] Loss: 0.036850
Train Epoch: 15 [32000/60000 (53%)] Loss: 0.137506
Test set: Average loss: 0.0530, Accuracy: 9852/10000 (99%)
Train Epoch: 16 [0/60000 (0%)] Loss: 0.060140
Train Epoch: 16 [32000/60000 (53%)] Loss: 0.240648
Test set: Average loss: 0.0525, Accuracy: 9849/10000 (98%)
Train Epoch: 17 [0/60000 (0%)] Loss: 0.271172
Train Epoch: 17 [32000/60000 (53%)] Loss: 0.158772
Test set: Average loss: 0.0592, Accuracy: 9823/10000 (98%)
Train Epoch: 18 [0/60000 (0%)] Loss: 0.088480
Train Epoch: 18 [32000/60000 (53%)] Loss: 0.097523
Test set: Average loss: 0.0471, Accuracy: 9864/10000 (99%)
Train Epoch: 19 [0/60000 (0%)] Loss: 0.133166
Train Epoch: 19 [32000/60000 (53%)] Loss: 0.119862
Test set: Average loss: 0.0511, Accuracy: 9855/10000 (99%)
Train Epoch: 20 [0/60000 (0%)] Loss: 0.102579
Train Epoch: 20 [32000/60000 (53%)] Loss: 0.063040
Test set: Average loss: 0.0651, Accuracy: 9809/10000 (98%)