论文阅读图片和文本联合训练：IMAGEBERT: CROSS-MODAL PRE-TRAINING WITH LARGE-SCALE WEAK-SUPERVISED IMAGE-TEXT DATA