MixUp augmentation forms a new training sample by taking a convex combination of two images and their labels using a mixing weight [0, 1].
