1konny / fgsm Goto Github PK

I notice that your code only have un-target attack part. But target-attack, where we can specify the target class mannually is loss/missing?

Hi @1Konny Can you please give me an example how to run this code? I am confused what should I put into [NAME]? Can you provide one example. Thanks.

python main.py --mode train --env_name **[NAME]**

Isn't it supposed to be # x_adv = x_adv + eps*x_adv.grad?

python main.py --mode generate --iteration 1 --epsilon 0.03 --load_ckpt best_acc.tar
Traceback (most recent call last):
File "main.py", line 75, in
main(args)
File "main.py", line 39, in main
iteration=args.iteration)
File "G:\FGSM-master\solver.py", line 213, in generate
pad_value=0.5)
File "F:\anaconda\envs\cw\lib\site-packages\torchvision\utils.py", line 105, in save_image
im.save(filename)
File "F:\anaconda\envs\cw\lib\site-packages\PIL\Image.py", line 2237, in save
fp = builtins.open(filename, "w+b")
OSError: [Errno 22] Invalid argument: 'output\main\legitimate(t:-1,e:0.03,i:1).jpg'

do you know if we can get the dimensions automatically based on the input and output shapes? It is really hard to understand what these are doing. I am trying to use the same code for attacks on different datasets and would like to have a neat way to infer these indices/repeats from the input shapes.

It would be wonderful if you could explain the four lines related to changed and where(), thanks!