Closed ZheningHuang closed 1 year ago
loicland
commented
2 years ago Hi,
Parallelizing the current partition code with CUDA would not be easy at all.
We are working on an accelerated version, which we should release this fall and is roughly x10 faster.
Otherwise, you could use another CUDA-friendly partition scheme such as mean-field and keep the rest as is.
ZheningHuang
commented
2 years ago Hi Loic,
Thank you so much for your reply. Could you please point me in the direction of the mean-field partition scheme which is more CUDA-friendly? Maybe get me a link to the GitHub Repo or the name of the paper.
I am testing different partition schemes, including yours for a paper and it would be really appreciated if you could help.
Best regards, Zhening
loicland
commented
2 years ago You can try this paper and associated repo:
it's not mean field but some sort of soft-kmeans that is GPU friendly, provided that the input graphs are not too large.
As for parallelized cut pursuit, we will release the new version this fall.
loicland
commented
1 year ago Hi!
We are releasing a new version of SuperPoint Graph called SuperPoint Transformer (SPT). It is better in any way:
| ✨ SPT in numbers ✨ |
|---|
| 📊 SOTA results: 76.0 mIoU S3DIS 6-Fold, 63.5 mIoU on KITTI-360 Val, 79.6 mIoU on DALES |
| 🦋 212k parameters only! |
| ⚡ Trains on S3DIS in 3h on 1 GPU |
| ⚡ Preprocessing is x7 faster than SPG! |
| 🚀 Easy install (no more boost!) |
If you are interested in lightweight, high-performance 3D deep learning, you should check it out. In the meantime, we will finally retire SPG and stop maintaining this repo.
Do you think it is possible to use Cuda to accelerate the speed of handcrafted partition?
Currently, it takes 3-5 seconds to process a 0.04 voxelized file in the S3DIS dataset and I am thinking of accelerating this with CUDA. Will be appreciated if you could give us know your thoughts on this. Is it feasible and if so what would be a good way to do this.
Best, Zhening