How long should it be to process 70000 img?

[Junjie-Chu]

Hi, I am trying to use your code to produce an attack dataset (size 70000, hide 32*32 MNIST img in 178*228 CelebA img). And it will cost about 350 hours or more. Is this a normal timing performance?

[EQuiw]

In the class QuadrScaleAttack,
change
self.optimize_runtime: bool = False
to
self.optimize_runtime: bool = True

Location:

2019-scalingattack/scaleatt/attack/QuadrScaleAttack.py

Line 29 in 5633ac9

self.optimize_runtime: bool = False

This will make the attack considerably faster, as the optimization problem is solved more efficiently. In future versions, I will set this flag to True per default.
Apart from that, it is difficult to say if your reported time is normal, as it depends on a lot of factors. But the computation takes indeed some time.

[Junjie-Chu]

Thanks for your quick reply, I will try your advice.

By the way, it seems the repository is not compatible with macOS (Monterey 12.3 and 12.1, I try the code on two Mac). The jupyter kernel always dies when I try to run. After I switch to Ubuntu, everything works well.