search

xingyul / flownet3d

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

pre-trained model performance #25

Closed ziruiw-dev closed 5 years ago

ziruiw-dev commented 5 years ago

Hi xingyul,

Thanks for the code and the pre-trained model. I run the evaluation.py over the pre-trained model you provided but I got performance output below:

eval mean loss: 0.404094
eval mean EPE 3D: 0.790930
eval mean ACC 3D: 0.000854
eval mean ACC 3D 2: 0.006859

maybe it's an untrained model? or did I do it in the wrong way?

Best, Zirui

xEnVrE commented 5 years ago

Hi @Ryan-Zirui-Wang

I had similar issues as yours due to wrong sm architecture in the Makefile of custom tensorflow operators. As an example, consider the following

https://github.com/xingyul/flownet3d/blob/3a4b3b6377c18cc5d32c7bdf0b2a0833df809aed/tf_ops/3d_interpolation/Makefile#L12

where sm_61 is used while my GPU required sm_60.

After rebuild of operators with the correct architecture I had the right results. Hope this helps.

ziruiw-dev commented 5 years ago

Hi @xEnVrE,

Many thanks for your help. Mine is sm_61 and I did try to re-compile operators but the problem remains... so it must be something wrong at my side if the pre-trained model works on your machine.

Best, Zirui