How to retain original image size?

[artyres1]

@LukasStruppek is there any option to retain the original image size because when I run the attacks, the result is seems to be a resized image?!?

[LukasStruppek]

Hey,

yes, there should be an option included in the code by calling a script using the option --optimize_original. This should optimize the image before resizing.

Best,
Lukas

[artyres1]

@LukasStruppek That option is not included in adv1_collision_attack.py file

[LukasStruppek]

Oh, I see. We only implemented this option for adv2_evasion_attack.py. However, it should be easy to add, just remove the resizing option from the preprocessing step and instead add it before feeding the image to the model. Since resizing operations are differentiable in PyTorch, the optimization should still run as usual. Best way is have a look at the adv2_evasion_attack.py script and how it is implemented there.

Best,
Lukas

[artyres1]

I added it like exactly you said but the resulting image kinda gives an opacity over top of image

https://gist.github.com/artyres1/6d0516d2ebc916c8ad84a742679f7534

This is the code..I use Mac M1 MPS engine..instead cuda..

[LukasStruppek]

What exactly do you mean by "an opacity over top of image"? You should make sure that the image is correctly saved in the end + that the updates are also correctly applied to the correct image tensor. Maybe check which tensor is optimized by the optimizer and use a debugger for localizing the problem. I currently do not have the capacity to implement this functionality into our code, I am sorry.