NAG.select() reveals `torch-geometric>=2.4` backward compatibility error

[hpc100]

Hi Damien, thank's for sharing SPT.

Environnement : NVIDIA RTX A6000 and Torch 2.0.0 // CUDA 11.7

I try to use NAG.select() and I get : Code : n = NAG.load('Area_6_conferenceRoom_1.h5') ; n.select(2, 10)

Do you have any ideas ?

[hpc100]

It's work fine when I add "()" to :

• self.keys in edge_keys() line 126
• self.keys in v_edge_keys line 141

It's a good solution ? Thank's

[drprojects]

Hi @hpc100

This seems might be a change in syntax wrt the torch-geometric version I am using. After a quick check, I noticed that a breaking change was made to the library: https://github.com/pyg-team/pytorch_geometric/pull/7629

What is your torch_geometric version ? I am using torch-geometric versions [2.2.0, 2.3.0, 2.3.1] just fine on different machines. Can you please try downgrading your torch_geometric to one of those versions and see if it works ?

If this solves the issue, I will look into either limiting the torch-geometric version or adapting the code to the new syntax.

PS: if you are interested in this project, don't forget to give our project a ⭐, it matters to us !

[hpc100]

Thank you for the quick reply. I used 2.4.0 torch_geometric. I install torch_geometric 2.3.1 and it's work fine ! Thanks

[drprojects]

Cool, thanks for catching this issue ! I updated install.sh to prevent installing torch-geometric versions above 2.3 in the commit: https://github.com/drprojects/superpoint_transformer/commit/f6837bd3574c78c5d4ad800343a367e0a67c963b

Best, Damien