Small gap between the reproduced and the reported results

[yukang2017]

Dear authors,

Thanks for your good work. I have reproduced the FSA results on KITTI val split with 40 recall positions across PointPillars, SECOND, Point-RCNN, and PV-RCNN. I use 4 GPUs for training and use the training cfg files provided. Below is the results.

---	3D	BEV
PointPillars (Report)	79.04	88.47
PointPillars (Reproduce)	78.64 (-0.4)	88.11
SECOND (Report)	81.86	90.01
SECOND (Reproduce)	81.48 (-0.38)	90.20
Point R-CNN (Report)	82.10	88.37
Point R-CNN (Reproduce)	81.86 (-0.24)	88.29
PV-RCNN (Report)	84.95	90.92
PV-RCNN (Reproduce)	84.68 (-0.27)	91.01

Would you please provide some suggestions to remove the gap? Thank you very much in advance.

Regards,
Yukang Chen

[nightfuryyy]

@yukang2017 @AutoVision-cloud hi, I want to train PV-RCNN_FSA model on v100 - 16gb on colab pro.
But I can't train because of out of memory error even with batch size = 1.
When I start training, memory keep increasing.
How can I fix that problem ?
Thank you so much.