Rethinking the Truly Unsupervised Image-to-Image Translation