About quality score

[lon0862]

Hi, I found that although you predict the quality score for different refine_iter's result. But you don't use them to determine use which refine_iter's as the final result like paper says. You just use the last one.
I try to build the algo by myself, but get worse evaluation metrics than original.
Does anyone successfully use quality score to get better result?

[youngzhou1999]

Hi, thanks for your interest.
The code uses the last prediction for smooth batch evaluation.
The algorithm for adaptive inference is simple and clear, but not convenient for evaluation. you can try the first rule referred to in the paper (use the quality score to determine whether multi-iter refinement is needed) to test adaptive inference.