Questions about the radius_graph

[StefanIsSmart]

The radius_graph in pyg could calculate the point that is in a distance (r) right?
However, I am wondering how to calculate the distance in implementation.
The Euclidean Distance？

[rusty1s]

Yes, you can just do

edge_index = radius_graph(x, r=...)
distance = (x[edge_index[0]] - x[edge_index[1]]).norm(dim=-1)

to get the final distance.

[StefanIsSmart]

In fact, before asking this question, I always thought this function could return the paired nodes that their Euclidean Distance in 3 Angstrom. But one time I tried to set the r as 1e10 (the x is the xyz coordinates of a real molecule.), which means the graph would be fully connected, but the result was not as I expected.

[rusty1s]

Yes, radius_graph is based on Euclidean distance. Currently, there is an upper limit on the number of neighbors returned, denoted by the max_num_neighbors argument in radius_graph.

[StefanIsSmart]

Thank you !
So by now, my understanding is:
The radius_graph and the knn_graph are all calculating the distance based on Euclidean distance.
The difference point is the knn_graph will return the nearest k nodes, and the raduis_graph will return all the nodes in a distance cutoff.
If I set the distance cutoff as positive infinite and set the max_num_neighbors as k, the radius graph will == the knn_graph.
Right?

[rusty1s]

Yes, this is mostly correct. However, distance cutoff==infinity + max_num_neighbors=k is not guaranteed to return the closest k neighbors. It will randomly pick k neighbors that fall into the distance cutoff (at least on GPU).

[StefanIsSmart]

I see, thank you!
I am a new fish in PyG, it is very helpful for me!