May I ask if there is no need for pre training methods. Does this mean that the projection layer and classification layer are two branches?