foolbox
foolbox copied to clipboard
Parallelize BrendelBethgeAttack
Currently the BB attack only uses on CPU for finding optimal perturbation directions, which can severly bottleneck the attack speed. The solution proposed here doesn't work because jitclass does not release GIL, but the other backends of joblib fail due to serilization errors.
@wielandbrendel Can you please explain the current state of this PR? Is it correct that the idea you had to speed things up doesn't work or do you think this is still fixable?