Question about the requirement of GPUs

[Hugo-cell111]

Hi! How many GPUs and what kind of GPUs are needed in semi-supervised point-cloud segmentation task?

[ldkong1205]

Hi @Hugo-cell111, thanks for your interest in this work!

For your question:

• It depends on which point cloud segmentation backbone you are going to use.
• As mentioned in the paper, our approach consumes 2x memory compared to the fully-supervised methods, and the training speed will not be affected.
• Since our codebase is based on MMDetection3D, you can refer to the training time of different fully-supervised LiDAR segmentation models from the benchmark. Here I list some statistics for you:

Method	Backend	Memory	Training Time	FPS
MinkUNet18-W16	torchsparse	3.4 GB	-	-
MinkUNet34-W32	minkowski engine	11.5 GB	6.5 hours	12.2
MinkUNet34-W32	spconv	6.7 GB	2.0 hours	14.6
Cylinder3D	spconv	10.2 GB	-	-

• The experiments use 8 GPUs. You can adopt the number of machines to use based on these statistics.

[ldkong1205]

Hi @Hugo-cell111, hope the above answer resolves your question. Free feel to let us know if you have further questions!