f two stages that were trained with different optimization goals and controlled by the co