Self-supervised learning for image-to-image translation in the small data regime