Save normalized tensor to png in a loop using pytorch

Question

I am working on a GAN and cannot make it work to save images that I transformed into tensors back to “normal” pngs within a loop. The same goes for the tensors that are generated by the Generator.

I applied the following transformation for the original images I am using for the training in the GAN ( I hope i did it the right way):

transform = transforms.Compose(
[
transforms.ToPILImage(),
transforms.Resize(img_size),
transforms.CenterCrop(img_size),
transforms.ToTensor(),
transforms.Normalize([0.5, 0.5, 0.5], [0.5, 0.5, 0.5]),
]
)

When trying to save the tensors as png images in a loop with the following code they do not come out the right way:

real_samples = next(iter(train_loader))
for i in range(4):
torchvision.utils.save_image(real_samples[i, :, :, :],
‘Real_Images/real_image{}.png’.format(i))

On the left is an example of the original image after transformation and on the right an example of the “wrongly” saved ones:

Can anyone please help me out with saving the images in the right way?

Answer 1

You apply normalization with mean 0.5 and std 0.5, so your images are transformed from range (0., 1.) to (-1., 1.). You should denormalize them and bring back to the original range before saving them.

In your case, simply doing

real_samples = real_samples * 0.5 + 0.5

before saving should work.