Matthias Fey

This is similar to any GNN layer. Only the aggregation of neighbors and root node is a bit different (although it has the same expressivity as `SAGEConv` since `- \alpha...

I agree that we need better documentation around these special arguments that are available as function parameters in `message` and `aggregate`. Happy to answer any questions in the meantime.

You can either do ``` from torch_geometric.utils import subgraph ``` or do ``` import torch_geometric.utils._subgraph ``` The file was re-named from `subgraph` to `_subgraph`.

You can find the `torch.nn.Module` implementation [here](https://github.com/pyg-team/pytorch_geometric/blob/master/torch_geometric/nn/conv/spline_conv.py), which should work fine both for CPU and CUDA usage. For batching, we follow PyG's mini-batch structure, see [here](https://pytorch-geometric.readthedocs.io/en/latest/advanced/batching.html) for an explanation.

Thanks for creating this issue. I think this would be a pretty cool feature. Two comments: * Currently, `collate` operates on a per attribute level (it computes the mini-batch feature...

You can also use `Data.subgraph()` or `HeteroData.subgraph()`.

Yes, part of the reason I am not necessarily eager to use this is since we are no longer able to group features of multiple PRs into a single bullet...

Thanks, looks good. There is one last problem with `to_hetero` (which seems to be a bug in `ToHeteroTransformer`).

You are right that heterogeneous GNN models currently do not support TorchScript. This is mostly due to a limitation by TorchScript, in which Tuples can not be used as keys...

I'm not yet sure TBH. It's sounds pretty complex to me. I think we could first start to work on making `HeteroConv` jittable, which seems easier to do than `to_hetero()`....