Kernel size on Spline Conv?

[hkim716]

Hi Matt,
I have two questions.

1. What does the kernel_size mean here self.conv1 = SplineConv(1, 64, dim=2, kernel_size=5)?
   I don't see any detailed explanation on the SplineCNN paper
   Is it similar to k in k-hop embedding?

2. I also have a question about the mechanism of mini-batch training in SplineCNN.
   I think it should be a little bit different from typical GNNs. In a typical GNN, M nodes are selected, and node embeddings are calculated for mini-batch training right?
   Then how does it work for SplineCNN? (Mini-batch handling part is a little bit confusing to me :<)

[rusty1s]

1. The kernel_size means how many control points/parameters there will be in each dimension. A kernel_size=5 across two dimensions will utilize 5x5=25 parameters per kernel (similar to a 5x5 kernel in CNNs).
2. We haven't applied SplineCNN in large-graphs with sampling techniques yet, but it should be similar to apply than any other GNN layer in PyG. Note that we distinguish between mini-batching of single graphs (in graph datasets) and mini-batching of sub-graphs (in single large graphs). I can try to explain it in more detail in case you specify your issues more concretely :)