Open juesato opened 5 years ago
(I think the failing Travis builds are a Python 3.5 issue and are not related to this PR, since this only updates the README.)
@nottombrown Do you know if an earlier version of the dataset had 40K images in the extras split? I've checked 0.0.3 and 0.0.2 (the first public version), and they both have 25K or 27.5K images in extras.
Related question: the leaderboard says defenses will be evaluated on 1000 images. Is this based on an earlier version of the dataset, using the 1000 image test set, or is this a separate hold-out set?