SVHN data grayscale and resize

1

Hi, sorry for yet another SVHN grayscale and resize problem permutation. After following lots of advice here on the forum I have a solution that renders some output. However when plotting a sample image it just shows a distorted color image, instead of the expected grayscale house number.

EDIT: I want a SVHN dataset with images in grayscale and size 28 x 28 in order to train the dataset on my MNIST CNN routine. Is this a correct method for fetching the dataset?

import torch
import torchvision
import torchvision.transforms as transforms
import torchvision.datasets as datasets
import matplotlib.pyplot as plt

class GetDataset(datasets.SVHN):
    def __init__(self, root, split='train',
                 transform=None, target_transform=None, download=True):
        super(GetDataset, self).__init__(
            root, split, transform, target_transform, download)

    def __getitem__(self, index):
        """
        Args:
            index (int): Index
        Returns:
            tuple: (image, target) where target is index of the target class.
        """
        img, target = self.data[index], int(self.labels[index])

        if self.transform is not None:
            img = self.transform(img)

        if self.target_transform is not None:
            target = self.target_transform(target)

        return img, target


svhn_transform = transforms.Compose([
    transforms.ToPILImage(),
    transforms.Resize((28, 28)),
    transforms.Grayscale(num_output_channels=1),
    transforms.ToTensor(),
    # transforms.Normalize((0.5, 0.5), (0.5, 0.5))
    transforms.Normalize([0.5], [0.5])
])


def main():
    svhn_training_dataset = GetDataset(
        root='./data',
        transform=svhn_transform
    )
    svhn_training_set_loader = torch.utils.data.DataLoader(svhn_training_dataset,
                                                           batch_size=128,
                                                           shuffle=True,
                                                           num_workers=1)
    test_img, test_lb = next(iter(svhn_training_set_loader))
    plt.imshow(test_img[0, 0], cmap='gray')
    plt.show()


if __name__ == '__main__':
    main()
Summary

This text will be hidden

2

After a bit of fiddling I find that setting transforms.Resize(28) returns an image of shape (28, 298). Could my use of the Grayscale() function be wrong?

weird_svhn

3

You are missing the permutation inside your __getitem__:

img, target = self.data[index], int(self.labels[index])
img = np.transpose(img, (1, 2, 0))
if self.transform is not None:
    ...

Since you are loading each sample as a numpy array, you would have to permute it manually.

1 Like
4

@ptrblck thanks so much for reaching out! Your answer certainly solved the dimensional weirdness.

Unfortunately there is still an issue with the grayscale. A sample image test_img[0] has the shape (1, 28, 28). If i try plt.imshow(test_img[0].permute(1, 2, 0) that is one dimension too many for imshow(). If I try plt.imshow(test_img[0].squeeze() I get the following picture. That sample image seem neither grayscaled nor appropriately resized. Any ideas?

squeeze

5

matplotlib uses the viridis colormap by default.
In your first code snippet you were setting the color map to gray via:

plt.imshow(test_img[0, 0], cmap='gray')
6

@ptrblck sorry for that, of course you are right. I also realize that I was fooled by the high resolution of my screen. Thank you so much for helping, solution accepted.

1 Like
7

Haha, there is no reason to be sorry. :wink:
I’m glad it’s working now.