如何在PyTorch数据加载器中将RGB图像转换为灰度图像？

Question

我已经从MNIST数据集中以.jpg格式下载了一些样本图像。现在，我正在加载这些图像以测试我的预训练模型。

# transforms to apply to the data
trans = transforms.Compose([transforms.ToTensor(), transforms.Normalize((0.1307,), (0.3081,))])

# MNIST dataset
test_dataset = dataset.ImageFolder(root=DATA_PATH, transform=trans)

# Data loader
test_loader = DataLoader(dataset=test_dataset, batch_size=batch_size, shuffle=False)

这里DATA_PATH包含带有示例图像的子文件夹。

这是我的网络定义

# Convolutional neural network (two convolutional layers)
class ConvNet(nn.Module):
    def __init__(self):
        super(ConvNet, self).__init__()
        self.network2D = nn.Sequential(
           nn.Conv2d(1, 32, kernel_size=5, stride=1, padding=2),
           nn.ReLU(),
           nn.MaxPool2d(kernel_size=2, stride=2),
           nn.Conv2d(32, 64, kernel_size=5, stride=1, padding=2),
           nn.ReLU(),
           nn.MaxPool2d(kernel_size=2, stride=2))
        self.network1D = nn.Sequential(
           nn.Dropout(),
           nn.Linear(7 * 7 * 64, 1000),
           nn.Linear(1000, 10))

    def forward(self, x):
        out = self.network2D(x)
        out = out.reshape(out.size(0), -1)
        out = self.network1D(out)
        return out

这是我的推论部分

# Test the model
model = torch.load("mnist_weights_5.pth.tar")
model.eval()

for images, labels in test_loader:
   outputs = model(images.cuda())

运行此代码时，出现以下错误：

RuntimeError: Given groups=1, weight of size [32, 1, 5, 5], expected input[1, 3, 28, 28] to have 1 channels, but got 3 channels instead

我了解这些图像将以3个通道（RGB）的形式加载。那么，如何在中将它们转换为单通道dataloader？

更新：我更改transforms为包括Grayscale选项

trans = transforms.Compose([transforms.ToTensor(), transforms.Normalize((0.1307,), (0.3081,)), transforms.Grayscale(num_output_channels=1)])

但是现在我得到这个错误

TypeError: img should be PIL Image. Got <class 'torch.Tensor'>

Answer 1

当使用ImageFolderclass并且没有自定义加载器时，pytorch使用PIL加载图像并将其转换为RGB。如果Torchvision图像后端为PIL，则为默认加载程序：

def pil_loader(path): with open(path, 'rb') as f: img = Image.open(f) return img.convert('RGB')

您可以在转换中使用Torchvision的Grayscale功能。它将3通道RGB图像转换为1通道灰度。在https://pytorch.org/docs/stable/torchvision/transforms.html#torchvision.transforms.Grayscale中找到有关此内容的更多信息。

下面是示例代码，

import torchvision as tv
import numpy as np
import torch.utils.data as data
dataDir         = 'D:\\general\\ML_DL\\datasets\\CIFAR'
trainTransform  = tv.transforms.Compose([tv.transforms.Grayscale(num_output_channels=1),
                                    tv.transforms.ToTensor(), 
                                    tv.transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))])
trainSet        = tv.datasets.CIFAR10(dataDir, train=True, download=False, transform=trainTransform)
dataloader      = data.DataLoader(trainSet, batch_size=1, shuffle=False, num_workers=0)
images, labels  = iter(dataloader).next()
print (images.size())

Answer 2

我找到了解决这个问题的一个非常简单的方法。张量所需的维度为[1,1,28,28]，而输入张量的形式为[1,3,28,28]。所以我只需要从中读取 1 个通道

images = images[:,0,:,:]

这给了我一个形式的张量[1,28,28]。现在我需要将其转换为形式的张量[1,1,28,28]。可以这样做

images = images.unsqueeze(0)

所以把上面两行放在一起，预测部分的代码可以这样写

for images, labels in test_loader:
   images = images[:,0,:,:].unsqueeze(0) ## Extract single channel and reshape the tensor
   outputs = model(images.cuda())