Deep Representations and Codes for Image Auto-Annotation