Towards Learning a Self-inverse Network for Bidirectional Image-to-image Translation