I understood that LD is a technique that only uses logit, which is the head output, as the knowledge of the teacher instead of intermediate feature(e.g.neck output) . However, output_feature is set to True by default. If I try to use logit only by setting output_feature to False, the code don't work.
My question is, If LD sets output_feature to False, then logit mimicking, and if output_feature is set to True, does logit mimicking + feature imitation work?
And Can I get the performance in the paper if I delete the feature imitation part code and turn it around using logit only?
AP: 42.1 , AP50: 60.3 AP75: 45.6
I understood that LD is a technique that only uses logit, which is the head output, as the knowledge of the teacher instead of intermediate feature(e.g.neck output) . However, output_feature is set to True by default. If I try to use logit only by setting output_feature to False, the code don't work.
My question is, If LD sets output_feature to False, then logit mimicking, and if output_feature is set to True, does logit mimicking + feature imitation work? And Can I get the performance in the paper if I delete the feature imitation part code and turn it around using logit only? AP: 42.1 , AP50: 60.3 AP75: 45.6