yaircarmon / semisup-adv Goto Github PK

Thanks for opensource the code!
Having a question on the memory requirement for running robust_self_training.py under default arguments. The readme says We tested this code on GPUs with 12GB of memory. , but I observed that it needs 16099MiB on P100 GPUs (tested after out of memory on 1080 Ti which has 11GB).
Wondering if this is normal?

Nice work! I was checking the pretrained checkpoint in your repo. I find the checkpoint is in Tensorflow (.ckpt) while the evaluation code is in Pytorch. Is there an easy way to get .pt checkpoint file instead? Thanks.

Is it possible for you to provide the pre-trained model used to generate the pseudo-labels ?

Thanks a lot for sharing.

I have problem in the procedure of evaluation.

Traceback (most recent call last):
File "attack_evaluation.py", line 243, in
eval_adv_test(model, device, test_loader, attack=args.attack,
File "attack_evaluation.py", line 85, in eval_adv_test
batch_linf_distances = cw(model, X, y,
File "/home/ubuntu/junyan/Python/adv_training/semisup-adv/attack_cw.py", line 32, in cw
raise ImportError('Could not import FoolBox')
ImportError: Could not import FoolBox

Check the package import.

Failure for " from foolbox.attacks.base import call_decorator", " from foolbox.utils import onehot_like"import
My foolbox version is 3.3.1.

What is the specified foolbox version to run CW attack?
Thanks & Regards!
Momo