How to reproduce reported performance

Thanks for this nice code base.

Can you provided instructions on how to reproduce the reported performance on the val set?

I tried:

python test.py --cfg_file ${CONFIG_FILE} --batch_size 4 --ckpt ${CKPT}

with the pretrained model and I get:

Car_3d_easy = 54.303261947978974
Car_3d_moderate = 38.349761507380684
Car_3d_hard = 36.39773833709999

with a model trained with the default config:

Car_3d_easy = 76.20793977206878                                                
Car_3d_moderate = 60.98983744758463                                                          
Car_3d_hard = 57.58892404665682 

which is far from the reported values. Am I doing something wrong? Is the provided config the one with which you report the performance?

Thanks

[chreisinger]

Okay, I use Pytorch 1.3, the newest version of numpy and an unchanged version of spconv. By the way, I use a single 1080Ti.

[willyd]

The issue was with a corrupted KITTI dataset I got from someone else… Sorry for the noise and thanks for the support. I can reproduce the exact same numbers with the PointPillars pretrained model with pytorch 1.4.