20 classes train, 10 classes val, and unlabeled images for semi-supervised learning.

original resolution