I have a regression problem where I want to penalize more on top of a normal MSELoss if the sign of y_true & y_pred is not the same. I am thinking of something like this
