Pointnet2_classification.py each eposch training time

[utkarsh0902311047]

Hello
I am trying the Pointnet2_classification.py example and its each epoch training takes approx 2min and 10 sec.
I am using cuda 10.1 with GPU NVIDIA T1000 8gb.

Is it too slow or its the appropriate training time?

[rusty1s]

A single epoch + testing takes around 17 seconds for me. Which PyG and PyTorch version are you using?

[utkarsh0902311047]

Cuda:10.1
Pytorch: 1.8.1
Pyg: 2.2.0

[rusty1s]

I think there was a bug in BatchNorm for older PyTorch versions. You could try to remove the BatchNorm and see if that speeds things up.

[utkarsh0902311047]

self.mlp = MLP([1024, 512, 256, 10], dropout=0.5, norm=None)
Removing BatchNorm is done by setting norm=None in Pointnet2_classification.py line 56, or is there any other way?
The above code line already has norm=None

[rusty1s]

You would need to specify norm=None in every initialization of MLP, e.g., in SAModule.

[utkarsh0902311047]

Ok now it is taking around 30sec to train every epoch+testing. thank you so much.
What should be the appropriate pytorch, PyG, and Cuda versions to make it work without any bugs such that also it can be run in editable mode?

[rusty1s]

I think it was fixed around PyTorch 1.10, but I am not totally sure. I suggest to always ensure you are running on either one of the two most recent PyTorch versions.