Benchmark datasets for the paper "Look Beyond Bias with Entropic Adversarial Data Augmentation", published in ICPR 2022.

You can download the datasets at the following link: https://drive.google.com/drive/folders/1jH9NVWgvYq6dLWRjKNml1SYsfL5vCNxf?usp=sharing

File Format

The file is a pickled dictionary with the following keys: 'train', 'val', 'test'. The value for each key is also a dictionary with the keys 'images' and 'labels', corresponding to, respectively, the images and the labels of the corresponding dataset. Inside these sub-dictionaries, images is a tensor of dimension N x 3 x 32 x 32, where N is the number of images in the dataset, and labels is a tensor of length N containing the classes from 0 to 9. The range value of the images is [-1, 1]. The file can be read with the torch.load() function.

Example for the located CIFAR-10 dataset

The train (train) dataset contains CIFAR-10 images that were strongly biased by the addition of a patch in a class-correlated position. It is made from the original CIFAR-10 training set.

The validation (val) dataset contains CIFAR-10 images with the same patch bias but coming from the original CIFAR-10 test set.

The test (test) dataset contains CIFAR-10 images from the original test set. The added shortcut (or bias) is the average of all the patches (resulting in diffuse patches in all positions).

Misc

The colors used for the training set (for both the colored MNIST and CIFAR-10) :

{0:(255, 127, 127), 1:(0, 127, 127), 2:(127, 255, 127), 3:(127, 0, 127), 4:(127, 127, 255), 5:(127, 127, 0), 6:(0, 255, 255), 7:(255, 0, 0), 8:(255, 0, 255), 9:(0, 255, 0)}