icgy96/apr

[ICCV 2021] Amplitude-Phase Recombination: Rethinking Robustness of Convolutional Neural Networks in Frequency Domain

apr's Issues

randomness in apr-p and apr-s

Hi

I see that in both APR-P and APR-S there is some randomness in the augmentations. For APR-P, the data is mixed only 40% of the time

APR/core/train.py

Line 16 in ca61e35

if p > prob:

and in APR-S, the data is augmented 50% of the time

APR/datasets/APR.py

Line 25 in ca61e35

if p > 0.5:

with additional randomness in the choice of image whose phase is returned

APR/datasets/APR.py

Line 46 in ca61e35

if p > 0.5:

Were these steps used to obtain the results in your paper, as I do not find this randomness mentioned in the paper?

Also, could you briefly describe how APR-SP was implemented - did you first apply APR-S and then apply APR-P?

Thank you,
KC

Can you share your ResNet-50 model checkpoint?

I want to use your pretrain model to do other experiments.
Can you share your ResNet-50 model checkpoint?

Thanks

Can you share the implementation codes for AutoAttack? (FSGM)

I read your paper very impressive using Fourier transform.
I wish I could have the implementations to run the AutoAttack experiments.
Can you share your code for FSGM training/eval?

Question about the section 5.3 in your paper.

Hi @iCGY96! Thanks for your insightful work! I found it is hard to understand in Section 5.3: Labeled by Amplitude or Phase? In this section, the authors investigated the use of labels amplitude and phase separately. "The model is trained for the sample x combined by the phase spectrum Pxi and the amplitude spectrum Axj by optimizing Eq(7)". How to understand x combined by the phase spectrum P and the amplitude spectrum A. Is it means that you first decouple the x into A and P, then input the A and P to the networks with two FC
layers, or input [X, P] and [X, A] into the networks?

dafault path of data_dir is wrong in ./autoattacks/train_fgsm.py

if you have tried ./main.py and prepared datasets in ./data, and you want to train the code in ./autoattacks like me, attention here.
in line 26 of train_fgsm.py:
parser.add_argument('--data-dir', default='../data', type=str)
should be modifieed to:
parser.add_argument('--data-dir', default='../data/cifar10', type=str)

Delete

Hi, I meet the 404 error when download the pretrained models

the shared link of APR-SP model seems broken, and yields a 404 error