The score of dataset v2.

[attackon1point6]

Hello, I find that the method proposed in the paper performs well on the cp-v2 dataset but shows poor results on the v2 dataset (with a score around 26 when not modifying hyperparameters). Should I consider adjusting hyperparameters specifically for these two datasets?What is the best score you achieved on the v2 dataset?

[guoyang9]

I haven't done experiments on the original VQA v2 dataset since it is more biased.
But you can try to tune the entropy value to see whether it affects the final model performance.

[attackon1point6]

Oh!Thank you for your reply!I'll try it.