Question about the usability of our repo.

[jihanyang]

Hello,

I am the maintianer of OpenPCDet. I notice your nicework that intergrates spconv in the OpenPCDet. May I know the usability of this repo? For example, whether this repo support various pytorch version and their corresponding spconv version? Besides, whether intergrating spconv in OpenPCDet largely reduce the workload to setup OpenPCDet?

Best regards,
Jihan

[Divadi]

I sincerely apologize for the delayed response.

In theory, this repository should work with pytorch 1.3+ just like mmdetection3d, but I have only personally tested it on 1.8 & 1.9.

I have had issues in the past installing spconv from https://github.com/traveller59/spconv which is why I set up this repository (I found it especially difficult for scattered CUDA installations and/or not having sudo permissions). In that sense, I believe this repository makes setting up OpenPCDet much easier. In my experience, I have set up OpenPCDet many times on different machines, and most of the time is used installing spconv. Mmdetection3d, which the spconv + voxelization was taken from, seems to receive fewer compilation questions above spconv.

That said, I believe this modification of OpenPCDet requires more testing.

For one, I found that spconv 1.0, which is used here, suffers from high GPU memory issues - while I have not observed issues on KITTI, on Waymo, a 2080 Ti is unable to hold PV-RCNN with batch size 2. I see that OpenPCDet recommends usage of spconv 1.2.1 for torch 1.3+. Although I am not sure why, perhaps this GPU memory issue is a byproduct of using a high torch version (1.8) with spconv 1.0. I do hope to find some time to modify spconv 1.2.1 to be able to be installed in this easy one-step method, but I am not sure when I will have the time. Regardless, thank you for your interest in this repository.

Sincerely,
David

open-mmlab/mmdetection3d#639
open-mmlab/mmdetection3d#192 Note: Changing voxelization to cpu did not resolve the memory issue.

[Divadi]

In the latest commit here, I believe the GPU memory issue has been resolved.