search

Megvii-BaseDetection / DenseTeacher

DenseTeacher: Dense Pseudo-Label for Semi-supervised Object Detection
Apache License 2.0
120 stars 12 forks source link

Could not get results in paper #18

Closed fu-7 closed 1 year ago

fu-7 commented 1 year ago

I run coco-p10 without modification and get AP33.11 which is much lower than 35.11 in paper. I want to know the reason. By the way, when i check the code in runner.py get_distill_loss: loss_deltas = (iou_loss( student_deltas[b_mask], teacher_deltas[b_mask], box_mode="ltrb", loss_type='giou', reduction="none", ) * teacher_quality[b_mask]).mean() Here why is teacher_quality[b_mask] not teacher_quality[b_mask].sigmoid()? It is supposed to represent probability so why not use sigmoid?

ZRandomize commented 1 year ago

please check whether you are evaluating the TEACHER model, 33.11 mAP looks like the performance of the student. The second question is a bug, the code is removed in latest code and seems like it won't affect final performance..

fu-7 commented 1 year ago

Thanks for your reply, the performance of teacher model is 35. Since i get the performance by using "tensorboard --logdir=output". Does the events of tensorboard save the performance of student model？

ZRandomize commented 1 year ago

yes, you can modify the "self._last_eval_results" here to choose which result to show in tensorboard