EleutherAI/cifarnet

CIFARNet contains 200K images sampled from ImageNet-21K (Winter 2019 release), resized to 64x64, using coarse-grained labels that roughly match those of CIFAR-10. The exact ImageNet synsets used were:

{
  "n02691156": 0,  # airplane
  "n02958343": 1,  # automobile
  "n01503061": 2,  # bird
  "n02121620": 3,  # cat
  "n02430045": 4,  # deer
  "n02083346": 5,  # dog
  "n01639765": 6,  # frog
  "n02374451": 7,  # horse
  "n04194289": 8,  # ship
  "n04490091": 9,  # truck
}

The classes are balanced, and the dataset is pre-split into a training set of 190K images and a validation set of 10K images.