20% or 100% of Waymo training datasets used in PillarNet?

[Coincidance-Levi]

I noticed that in the paper, PillarNet followed the same training schedule as CenterPoint, which decleared in the paper that it followed PV-RCNN. However, PV-RCNN didn't point out the way it made use of the Waymo training datasets, which usually meant it used the full set of training datasets. On the other hand, the code you offered in det3d and OpenPCDet indicated that 20% is the right option. For further evidence, could you tell me the answer?