[Paper Review] BLIP-2: Bootstrapping Language-Image Pre-trainingwith Frozen Image Encoders and Large Language Models