Self Augmentation Loss

Can someone give an idea on how to implement self-augmentation loss in pytorch in my CNN?