CINIC-10

Overview

CINIC-10 (Darlow et al., 2018) extends CIFAR-10 to 270,000 32×32 RGB images over the same ten classes: all 60,000 CIFAR-10 images plus 210,000 down-sampled ImageNet images. Each split (train, validation, test) has 90,000 class-balanced images, 20,000 of them from CIFAR-10.

Data Structure

When accessing an example using ds[i], you will receive a dictionary with the following keys:

Key	Type	Description
`image`	`PIL.Image.Image`	32×32×3 RGB image
`label`	int	Class label (0-9), same classes and order as CIFAR-10: airplane, automobile, bird, cat, deer, dog, frog, horse, ship, truck
`origin`	int	Image source, read from the release file names: `0` = `cifar10`, `1` = `imagenet`

About 0.1% of the ImageNet-derived images are grayscale in the release; the loader converts them to RGB.

Usage Example

from stable_datasets.images import CINIC10

ds_train = CINIC10(split="train")
ds_test = CINIC10(split="test")

sample = ds_train[0]
print(sample.keys())  # {"image", "label", "origin"}

References

Dataset record (Edinburgh DataShare): https://datashare.ed.ac.uk/handle/10283/3192 (DOI 10.7488/ds/2448)
Paper: https://arxiv.org/abs/1810.03505
License: CC BY 4.0

Citation

@article{darlow2018cinic,
  title   = {CINIC-10 is not ImageNet or CIFAR-10},
  author  = {Darlow, Luke N and Crowley, Elliot J and Antoniou, Antreas and Storkey, Amos J},
  journal = {arXiv preprint arXiv:1810.03505},
  year    = {2018}
}