Anshuman Suri

@zziz Vision conferences such as WACV would be a great addition as well

This PR fixes this issue : https://github.com/zachwill/dom/pull/16

@zhongyy @wangdi19941224 @ryokamoi did any of you manage to fix the NaN blowing up issue? I face the same whenever I encase it with a `torch.no_grad()`

@ryokamoi it's VGG19. I did try increasing "scale" to 500. I got rid of the NaNs (for now). Is there a good heuristic/estimate to see what lowest (or ballpark) value...

Right. Thanks a lot, @ryokamoi :D

The code tries to access 'contextss' key from configuration files at several places On Fri 4 Aug, 2017, 5:04 AM Blacksoil, wrote: > Not sure if the you are mentioning...

MadryEtal uses [ProjectedGradientDescent](https://cleverhans.readthedocs.io/en/latest/_modules/cleverhans/attacks/projected_gradient_descent.html#ProjectedGradientDescent), which further uses FGM to compute the perturbation using the gradient. I could not find any step in the FGSM implementation that normalizes gradients (I may have...

@michaelshiyu if you have a look [here](https://github.com/MadryLab/robustness/blob/master/robustness/attack_steps.py#L127) (the step() function is called inside the main attack), they normalize the gradient first. The 'clip_eta' step is for limiting the perturbation to...

Page 18 of [this paper](https://arxiv.org/pdf/1905.02175.pdf) by the same research lab references the Madry attack, saying that "we normalize the gradient at each step of PGD". If you look at their...

I think they are aware of that fact, and that it may not be interference according to them. They normalize the gradients first and then do other projection steps. As...