UNIFIED-IO: A Unified Model for Vision, Language, and Multi-modal Tasks.