search

MadryLab / constructed-datasets

Datasets for the paper "Adversarial Examples are not Bugs, They Are Features"
http://gradientscience.org/adv
185 stars 23 forks source link

Image order in constructed datasets #12

Open kiranchari opened 3 years ago

kiranchari commented 3 years ago

Hi,

I am trying to view the constructed datasets (alongside standard CIFAR10) image by image i.e. simultaneously viewing all versions of a particular image. But it appears the constructed datasets contain the images in different orders so I am not able to pull up the same image across different datasets.

Any suggestions?

Thanks, Kiran

kiranchari commented 3 years ago

Hi can you let me know how I can retrieve the ids of the images in the constructed datasets? Thanks

andrewilyas commented 3 years ago

Hi! Unfortunately we have stored all the data from our experiments in archives that are very hard to retrieve. I would recommend running (l2) nearest neighbors in order to match images back to indices.

kiranchari commented 3 years ago

Thanks for the reply @andrewilyas

The datasets appear to be a single file without a train/test split. Do these datasets have a test test that I can run the NN on? (running on the train set would take much longer).

Thanks!

Asteriajojo commented 3 years ago

@andrewilyas Hi,could you please tell me how to running (l2) nearest neighbors in order to match images back to indices? If we want to find out in which image the robust feature belongs to,with which two datasets should we compare?Should we do it in original cifar training datasets and robust datasets?Thanks a lot!