Dataloader loading color distorted image

vision
1

Quick show and tell will be the easiest way to tell you the problem.

# Define the dataset and the loader as usual
dataset = dset.ImageFolder(root=datapath,
                           transform=transforms.Compose([
                                transforms.Resize(image_size),
                                #transforms.CenterCrop(image_size),
                                transforms.ToTensor(),
                                transforms.Normalize((0.5414, 0.5333, 0.5338),(0.1897, 0.1911, 0.1932))]))

dataloader = torch.utils.data.DataLoader(dataset, batch_size=64,
                                         shuffle=False, num_workers=workers)

#load the image
data = next(iter(dataloader))

#save the image
dlX = data[0].to(device)
vutils.save_image(dlX[0], 'test.jpg')

Result of save_image
test

Am I missing something? or doing something wrong?
I have trained a DCGAN with this dataset with results look like this but can only post a single image)

I am attaching a link to my dropbox where you can see the result, original, and the training results
https://www.dropbox.com/sh/apbfcuezzlj1kbf/AAB_-fzU1_8oLRorZ9kWF0fAa?dl=0

2

Take a look at torchvision.transforms.Normalize. Your picture is being normalized, which results in this weird picture. If you remove transforms.Normalize((0.5414, 0.5333, 0.5338),(0.1897, 0.1911, 0.1932))], the image should look normal.

3

Thank you for the reply.

As far I as I understand normalizing, I thought it was crucial that I normalize the dataset accordlingly to it’s avg and std, thus those are the calculated value.

Then, to get results more resembling the original, is it okay to skip normalization?

4

Yes, normalization should be done. I wouldn’t skip it, as normalizing data improves performance.
If you want the original image back and you know the mean and std, you could simply reverse the normalization:

normalized_image = (original_image - mean) / std
original_image = (normalized_image * std) + mean
5

Thank you for the reply, I had not been able to come back to this issue for a bit while.
I will check within next week to see if it work, which I highly think would.

Thanks for your time