CIFAR-10.2

Overview

CIFAR-10.2 reproduces the CIFAR-10 collection process to study distribution shift. It provides 10,000 training images (1,000 per class) and 2,000 test images (200 per class). There is no separate validation split.

All images are 32x32 RGB images with the same ten labels as CIFAR-10. The figure below shows one image from each class in the CIFAR-10.2 training set.

CIFAR-10.2 training images, one for each of the ten classes

Data Structure

Each example returned by ds[i] has these keys:

Key	Type	Description
`image`	`PIL.Image.Image`	RGB image; converting it to NumPy gives shape (32, 32, 3) and dtype uint8.
`label`	`int`	Class index from 0 to 9.

The label order is airplane, automobile, bird, cat, deer, dog, frog, horse, ship, and truck. The builder exposes the images and labels from the upstream NPZ files; it does not expose their Tiny Images indices or keywords.

Usage Example

import numpy as np

from stable_datasets.images import CIFAR10_2

train_ds = CIFAR10_2(split="train")
test_ds = CIFAR10_2(split="test")
print(len(train_ds), len(test_ds))  # 10000 2000

sample = train_ds[0]
image = np.asarray(sample["image"])
print(image.shape, image.dtype)  # (32, 32, 3) uint8
class_names = train_ds.info.features["label"].names
print(class_names[sample["label"]])

Train a classifier on the training split and report accuracy on the held-out test split. If model selection needs a validation set, create it from the training split.

License

The upstream repository does not include a license file or specify a data license in its README. The builder reports license="Unknown". The images originate from Tiny Images, but that provenance does not establish a license for this release.

References

CIFAR-10.2 repository
Harder or Different? A Closer Look at Distribution Shift in Dataset Reproduction, presented at the ICML 2020 Workshop on Uncertainty and Robustness in Deep Learning.

Citation

@inproceedings{lu2020harder,
  author={Lu, Shangyun and Nott, Bradley and Olson, Aaron and Todeschini, Alberto and
          Vahabi, Hossein and Carmon, Yair and Schmidt, Ludwig},
  title={Harder or Different? A Closer Look at Distribution Shift in Dataset Reproduction},
  booktitle={ICML Workshop on Uncertainty and Robustness in Deep Learning},
  year={2020},
  url={https://www.gatsby.ucl.ac.uk/~balaji/udl2020/accepted-papers/UDL2020-paper-101.pdf}
}