The problem of evaluation in voxelrcnn

[xxa783]

Hi,Recently, when I was training three classes of voxel-rcnn。The result of running locally is normal，but when running on the server, the accuracy of the pedestrian and cyclist categories is much worse than that of the local，The category of car is normal，The experimental parameter settings are exactly the same。How should i solve this problem，it bothered me for a long time，Hope to find help here！thanks

[xxa783]

Pedestrian AP R40 0.50，0.25，0.25:
bbox AP:66.9831，62.2371，58.6492
bev AP:71.4642，66.1941，62.7137
3d AP:71.3018，66.0044，62.0395
aos AP:63.38，58.13，54.34
Cyclist [email protected]，0.5，0.50:
bbox AP:61.9180，44.2106，44.0043
bev AP:60.6828，42.3162，42.1287
3d AP:59.9721，41.8940，41.4805
aos AP:61.20，43.58，43.30
Cyclist AP [email protected]，0.50，0.50:
bbox AP:60.7131，45.4572，43.1140
bev AP:59.6068，41.2654，39.0192
3d AP:56.9525，40.6235，38.4517
aos AP:59.99，44.54，42.20
Cyclist APa0.50，0.25，0.25:
bbox AP:61.9180，44.2106，44.0043
bev AP:61.2402，43.0665，42.6762
3d AP:61.2402，43.0665，42.6762
aos AP:61.20，43.58，43.30
Cyclist AP [email protected]，0.25，0.25:
bbox AP:60.7131，45.4572，43.1140
bev AP:60.1226，43.8350，41.5723
3d AP:60.1226，43.8350，41.5723
aos AP:59.99，44.54，42.20

[xxa783]

This is the result of running on the server，The two classes pedestrian and cyclist perform extremely worse than experiment on the local，What is the reason for this difficulty。The experimental setup is the same,batch_size is 2 ，epochs are 80.

[csjxchen]

@xxa783 Could you tell me the amount of your GPUs for training? I guess this is an ignored factor for reproducing the result.