search

A03ki / f-AnoGAN

Implementation of f-AnoGAN with PyTorch
MIT License
131 stars 35 forks source link

Need absolute value for image subtraction #19

Open jianzhang96 opened 1 year ago

jianzhang96 commented 1 year ago

Thanks for your excellent work! Recently I use the code to train my custom dataset, and I try to calculate the AUROC of localization performance. In the Line 25 of file fanogan/save_compared_images.py, there is code for image subtraction: compared_images[2::3] = real_img - fake_img But there are much white pixels in the compared_images. I find that the output images will be more reasonable by using absolute values of the compared_images: compared_images[2::3] = torch.abs(real_img - fake_img) Thanks a lot!

A03ki commented 1 year ago

Hi, @jianzhang96. Thanks for your comment and using this repository.

I tried that with the MNIST dataset.
The left figure shows the difference real_img - fake_img and the right figure shows the absolute difference torch.abs(real_img - fake_img). 000010000010_abs Indeed, using the absolute difference seems to make the output images reasonable.

I will consider adding an option to select the absolute difference. Thanks again for your comment.