Running inference in FP16

[YoushaaMurhij]

Nice work!
I tried to run the model in inference mode. Obviously, it is faster than CP-VoxelNet.
I am interested in running inference in FP16. I am struggled with converting input data to Float 16.
Could you please tell me which input variables should be converted to half precious beside the model itself?

Thanks,
Youshaa

[WWDCRobot]

The lossless pillarization still need be changed, you can modify the related CUDA code for your need.