Bipartite mappings with pytorch-geometric

[mishooax]

Hi,

Is there a pyg-tutorial example that shows how to create bipartite graph mappings between two data domains? I'm looking to recreate the model described in this paper: https://arxiv.org/abs/2202.07575. Specifically the mappings shown in Figure 2 of said paper:

E.g. the "encoder" (leftmost graph) represents a mapping between the latitude/longitude data grid (blueish dots) and a coarser, icosahedron grid (gray nodes). Each gray node is connected to its 12 nearest neighbors on the lat/lon mesh.

I'm aware of this implementation: https://github.com/openclimatefix/graph_weather/blob/main/graph_weather/models/layers/encoder.py but i'm having trouble understanding some parts of the implementation and would like to work my way up from a simpler example, if possible.

Many thanks!

[rusty1s]

You can take a look at radius and knn functions in PyG, which can create a mapping between two sets of points. Let me know if this is helpful!

[mishooax]

Cheers - can I use these functions with the Haversine (great circle) distance? My graph nodes (coordinates == lat, lon in radians) lie on the sphere.