作者你好，这个FLMR和官方版本的Retrieval-Augmented-Visual-Question-Answering区别在哪？FLMR是不是没用blip？

[zzk2021]

作者你好，这个FLMR和官方版本的Retrieval-Augmented-Visual-Question-Answering区别在哪？FLMR是不是没用blip？

[LinWeizheDragon]

This is also official. It is difficult to set up the entire environment with the original RA-VQA codebase. We provide an implementation based on Huggingface-transformers. This implementation provides support for inference with FLMR and PreFLMR, fine-tuning scripts, and evaluation scripts. You can easily run inference with this implementation.
If you want to fine-tune a blip based on the retrieved documents of PreFLMR, you can run inference, collect the retrieved documents, and write your own code to fine-tune a blip model.

[zzk2021]

thanks for your reply!