This dataset seems to have 10 or 50 images per class of the imagenet. I’m wondering is there any requirement of cherrypicking the images per class? Or they’re selected randomly?