Using spline_conv to approximate gradients on a graph about pytorch_spline_conv HOT 3 CLOSED

I'm not sure why you think one needs sort a node's neighborhood. If I understand you correctly, you want to implement a SplineConv layer where B-spline weights models a Sobel-like kernel. You can do this by customizing the SplineConv.weight, which is a tensor of shape [num_spline_weights, in_channels, out_channels]. Therefore, for in_channels = out_channels = 1, that would be something like

conv.weight = torch.tensor([-1, 0, 1, -2, 0, 2, -1, 0, 1]).view(-1, 1, 1)

from pytorch_spline_conv.

Thanks for your prompt response. I tried what you suggested, but without the intended results.
My goal is to have a SplineConv layer that is given a feature on a graph and returns the feature's derivative along a specific spatial direction.
In a sense, this would be analogous to a finite difference scheme, but generalized to unstructured meshes. This is why I thought that one needs to sort each node's neighborhood (related paper).

Thanks again for the support.

from pytorch_spline_conv.

I managed to do what I wanted using RBFs. I eventually implemented a RBF-FD network from scratch, but your repo proved to be useful in the process.
Thanks for the support anyway.

from pytorch_spline_conv.