Mario Geiger

The following code (not tested) is inefficient and works only if the `cutoff` is small enough. ```python lattice = torch.randn(3, 3) # 3 vectors forming a parallelogram in 3d pos...

Yes I will try to make something. I plan to look at it this week or maybe the week after. I'd love to contribute

No we give up the idea of implementing it in cuda. We [use](https://github.com/mir-group/nequip/blob/ef4eadd1e2f53796ec5b80b6909b56c60941974c/nequip/data/AtomicData.py#L412) [ase](https://wiki.fysik.dtu.dk/ase/ase/neighborlist.html)

@keano130 Why is ase implementation not good enough for your usage?

Thanks for the complete explanation of how to do the indexing properly!

I have no idea

Ring like kernels giving better performance than near identity kernel is only an observation done with shrec17 experiment. It might be that there is an even better kernel form for...

I realized the same issue

Hi, I'm sorry because this repository is not maintained since a long time. I have the plan to reproduce the architecture and experiments of this repository within the e3nn library...

Here is an implementation using e3nn: https://github.com/e3nn/e3nn/tree/main/examples/s2cnn/mnist There still might be an issue with this code, I remember it was not training properly last time I tried. The fast Fourier...