How to adjust SuperPoint number to avoid CUDA memory full?

[glennliu]

Hi

I'm running GeoTransformer on fused indoor point cloud (from ScanNet). I run it using demo.py. My GPU is Nvidia3090 with 24GB memory. But the program frequently fails due to "not enough CUDA memory". I downsample the point cloud with 2.5cm voxel size if it is >30000.

I understand the key part is reduce the superpoint number. So I follow #16 to adjust the parameters

backbone.num_stages=5,
neighbor_limits=[38,36,36,38,38]

But it stopped at RPEMultiHeadAttention and it shows,

Exception has occurred: RuntimeError
einsum(): subscript n has size 390 for operand 1 which does not broadcast with previously seen size 2418

Is any suggestion on how to adjust the parameter properly?
Thanks

[glennliu]

I believe the error should comes from the KPConvFPN backbone.
It is fixed at 4 stages and not adjusted with changed parameters.

I can modify the backbone to more stages. But it requires to re-train the network.
So I find an easier way to test the pre-trained weight. I adjust the backbone.voxel_size=0.05 to make it generate less superpoints. It works fine now without memory full issue.

But it may be less accurate than the original setting.