_backselect function takes minute for a single image

[e7mul]

Hi! Thanks for providing code to reproduce your results!

I tried to run a simple experiment with the CIFAR-10 and ResNet18 model based on commands you provided in docstrings. While the training of the model went smoothly, I found that run_sis_on_cifar.py takes a lot of time. I managed to track the problem, which turned out to be the _backselect procedure. It takes approximately a minute to compute a mask for a single image with GPU. Is it expected? If so, how one can produce a masked version of the whole dataset?

Thanks for your help!