search

xingyul / flownet3d

FlowNet3D: Learning Scene Flow in 3D Point Clouds (CVPR 2019)
MIT License
363 stars 83 forks source link

huge prediction value #38

Closed Yang-L1 closed 4 years ago

Yang-L1 commented 4 years ago

Hi, I download the pretrained model and the processed data. Then I evaluate the model using the given scripts.
I keep getting very huge loss like this:

loss value 1019623.6 pred_value [[[-1183.0692 -451.19238 -942.2431 ] [-1617.0183 -613.1068 -1288.0482 ] [-1646.4847 -628.52234 -1312.6896 ] ... [-1316.52 -497.82248 -1044.1621 ] [ -975.26605 -370.09656 -777.7711 ] [-1316.516 -497.82562 -1044.1562 ]]]

loss value 5635.0854 pred_value [[[-181.02733 78.97927 54.753517 ] [ -60.915703 37.60071 5.6472397] [-143.16997 68.90349 30.370804 ] ... [-141.26614 68.30685 29.489553 ] [-143.37962 68.98323 30.424229 ] [ -94.74317 50.511147 22.388628 ]]]

Do you have any idea about this? Thx.

xingyul commented 4 years ago

It's likely that the CUDA Op is not compiled correctly. Please check the arch flag in Makefiles.

Yang-L1 commented 4 years ago

Thanks. It works.