HSVA: Hierarchical Semantic-Visual Adaptation for Zero-Shot Learning