Hi,
I see that you mentioned that you have used three dataset. However , the preprocess script is for COCO VQAV2 and Captions only. Further, Is there any training/pre-training required using caption dataset as well??

I was running the script only for Visual QA and realised that single A100/80GB does not support training with the given parameters. I am able to run it with my set of parameters. Please confirm if there are multiple GPUs used at your end for running those experiments and related configurations.