[bug] Side effect of Attack.save() with int return type

[Martin Gubri]

I am a bit confused about the management of return types of the base class torchattacks.attack.Attack.

1. If the return type is set to int (_return_type=='int'), calling Attack.save() resets the return type to float (see line below). This may be a hard-to-detect side effect.
   

   adversarial-attacks-pytorch/torchattacks/attack.py

   Line 175 in 7121441

    self._return_type = 'float' 

2. The metrics in Attack.save() are computed on float adversarial examples and not on integer adversarial examples. It might lead to unexpected conclusions when comparing the printed accuracy and the accuracy of saved adversarial examples.

My proposition

I can create a PR with the following modifications:

• Attack.save() does not modify _return_type of the class
• Attack.save() computes metrics (L2 norm and accuracy) on scaled adversarial examples (adv_images/255) when _return_type=='int'.

@Harry24k Does it look good to you? Thanks for your time :)

[Hoki Kim]

Thank you for your comment :)
I fully understood what can be the problem, and I removed the modification during saving.
It will be updated in this commit.

[Martin Gubri]

Great! It looks good. Thanks for your time :)