Bi-directional Distribution Alignment for Transductive Zero-Shot Learning