how to random rotate or flip input during the training and do the same operation on output

Question

I am trying to implement this paper.

It proposes a new loss, which forces the predicted masks of the geometric transformed (horizontal flip and rotation)input image to be