graph from grid

[jakubMitura14]

Hello, thanks for sharing your fantastic work in this repository.
I have a 3d array (medical imaging), and I wonder if it is possible to create a graph preserving the possibility of backward differentiation.
I want to create super voxels (from those physically close and similar in feature space), then, based on features in those super voxels, I want to get the node features and graph edge weights based on proximity.

The problem is that the graph is a discrete structure, and I do not know how to approach generating a graph and keeping differentiability.

I considered perturbing the graph creation process and running it a couple of times, each time getting a slightly different graph so that the perturbed graph creation function would be differentiable.

As your tool seems close to my idea, I was wondering whether you would share your thoughts on the presented concepts. Is it achievable?

Thanks !!