Describir: Multi-modality multi-label ocular abnormalities detection with transformer-based semantic dictionary learning.