search

open-mmlab / OpenPCDet

OpenPCDet Toolbox for LiDAR-based 3D Object Detection.
Apache License 2.0
4.58k stars 1.28k forks source link

Question about the 'one_cycle' strategy in training stage #248

Closed yifanzhang713 closed 4 years ago

yifanzhang713 commented 4 years ago

Hi, Shaoshuai,

I found that all models in OpenPCDet used the one_cycle learning rate delay strategy only related to total epoch setting, which is really different with the original SECOND and PointPillars. And peformances of SECOND and PointPillars in OpenPCDet are also 1~2 points higher than original versions. So I wonder if this learning rate delay strategy is better and has made the main contributions to the AP promotion based both two original versions'.

Best regards.

sshaoshuai commented 4 years ago

Actually the official second codes also use one-cycle decay strategy, and it is important for achieving the reported performance in current settings. But I think the multi-step decay strategy could also achieve similar performance but the related hyper-parameters may need to be tuned. Since the one-cycle decay strategy is simple and has few hyper-parameters, so I set it as default decay strategy.