mahyarnajibi / freeadversarialtraining Goto Github PK

this question is about train_fgsm.py
I modified function "get_loaders" and net's num_classes to let program can run with cifar100. But pgd acc is only 3%. And it's normal with cifar10.

The modified part is as follows：
(1)net

(2) get_loaders

Has anyone encountered the same problem as me?

I am facing memory leak issue due to the gradient calculation of the noise_batch tensor. Is there any specific PyTorch version required to avoid this issue?

The existed code do not normalize both train and valid data.

in paper, g_{adv} is x's gradient, but in code, it's delta's ?

the code is

Hi, your work is great! I have tried to run Free-8 with WideResNet34-10 on CIFAR10 using PyTorch.
My experiment settings: m=8, clip_epsilon=8.0 and fgsm_step=2.0. I trained 240 epochs. Initial learning rate is 0.1 and learning rate decays 10 times at 100//8 epoch and 150//8 epoch. SGD optimizer, momentum 0.9 and weight_decay 2e-4. And I set torch.backends.cudnn.benchmark = False and torch.backends.cudnn.deterministic = True.
The experiment results: clean validation accuracy is about 88.5, but adversarial validation accuracy is only about 43. Is there something wrong with my settings? Could you please provide more detail information?

is in1.clamp_(0, 1.0) should be in1.clamp_(0, 255.0) ? As "in1" has not been normalized yet.

Hi,

Thank you for providing the code. I modified the code a little by only changing the dataset/dataloader to CIFAR-10 and network architecture to WideResNet32-10/ResNet-50 to reproduce the free adversarial training results on CIFAR-10. Unfortunately, I failed to get the expected results. The top1 test accuracy of natural images and the top1 accuracy againt PGD attacks are far away from that the article reported, less than 20% and 10% respectively.

Here is the modified code and the training logs are attached.
resnet50_log.txt
WRN_32_10_log.txt

Is there any way to access pre-trained models?

Thanks for your great work.
I notice the training script mainfree.py does not support multi-gpus training, however the TensorFlow version does.

Is there any technical issue?