In this report we will see the effect of training the student network with a batch of original CIFAR100 images combined with a MixUp set of images from the same batch. The teacher network is trained on pure images and the targets for the MixUp images are created on the fly by using the separate images used to create the MixUp.

Runs to see the performance of knowledge distillation with and without CutMix augmentation on Teacher: ResNet 32x4 Student: ResNet 8x4