PyTorch Geometric Implementation of PointGNN

[andreasbinder]

Dear Weijing,
I enjoyed your paper very much and based on your research, I am trying to create an implementation in PyTorch Geometric. While doing so, some (small) questions came up:

1. How did you identify the final radius for the cutoff for creating the graph? Did you apply any normalisation on the PointCloud dataset (I did not see information on that in the paper)
2. By looking at the message passing update, I figure that the state dimensions as input are the same as for the output as you add s_i^t in the final step, is that correct?

Thank you so much!