mPLUG-2: A Modularized Multi-modal Foundation Model Across Text, Image and Video论文笔记