demo_inception.py can't be used to generate a universal perturbation

[HotSammiches]

In demo_inception.py:
path_train_imagenet = '/datasets2/ILSVRC2012/train'

is passed into create_imagenet_npy(). This path doesn't exist, so the rest of the function doesn't run. What is the best way to get the data so the universal perturbation vector can be created?

[husseinfawzi]

You need to download the data from http://image-net.org/download-images, and specify the path of the folder containing the images.
Alternatively, you can use the pre-computed perturbations under the "precomputed" folder.

[HotSammiches]

Thanks! Another question - how many passes does it usually take to compute a universal perturbation?

Due to memory limitations I needed to drastically decrease the number of images used (from 10,000 to 500) and I'm running CPU-only. So it's taking much longer than it would on a GPU (which I expect) but I'm wondering how many passes should be necessary.

[HotSammiches]

@husseinfawzi - Do you know the answer to my question above? Thanks for your help.

[husseinfawzi]

Usually 5 passes.
As shown in the paper (Fig. 6), you should only expect however to achieve roughly 30% misclassification with 500 images.