SplineCNN parameters in main paper

[mfarazi1991]

Hi Matthias,

I am doing a benchmark for shape correspondence and I am not sure how to convert the parameters in SplineCNN paper to SplineConv parameters. SplineConv(in_channels, out_channels, dim, kernel_size) is the PyG conv layer, but I am not sure how to define dim and kernel size (the rest is straightforward). Do you have the code for any experiments like graph node classification or others available? I want to see the network setup if possible. Thanks!

[rusty1s]

You can find an example of SplineCNN here: https://github.com/pyg-team/pytorch_geometric/blob/master/examples/faust.py

• dim defines the feature dimensionality of edge features. For example, in a point cloud setting with edge features denoting spatial relations, this is set to 3.
• kernel:size defines the number of parameters we want to learn in each spatial dimension. This is similar to the kernel size of traditional CNNs. If set to 5 with dim=3, one will apply a kernel window of 5x5x5 parameters.

[mfarazi1991]

Like always, thank you for your prompt and helpful response.