RobDHess / Steerable-E3-GNN

E(3) Steerable Graph Neural Network
MIT License
102 stars 19 forks source link

Question about Visualizing Network Filters #5

Open amorehead opened 2 years ago

amorehead commented 2 years ago

Hello.

I have been experimenting with this network on a new dataset, and I was curious if anyone knows of a simple way to visualize the filters of the network once it has been trained. I have seen filter visualizations for other 3D graph neural networks before (e.g., SphereNet, DimeNet), however, I was curious if such visualization techniques would be applicable to this network architecture.

Thank you in advance for your time and help.

RobDHess commented 10 months ago

Hi,

Sorry that it took me so long to respond. In the case that this is still useful to you:

Visualising our filters would work a little bit differently than for those networks. Our method is not a convolution-style method, so there isn't a real filter to speak. It is a message passing framework that uses two steerable MLPs to send messages and update node features.

For each path in the tensor product, you could visualise the relative position spherical harmonic and include the weight of that path, but I do not think it would be what you are looking for.

Kind regards,

Rob