How to run batches with HypergraphConv

[FernandezR]

Hello,

I am using the HypergraphConv in my algorithm, but it doesn't seem to handle batch learning. Is there some unique way of doing batches with HypergraphConv that I'm not seeing or is it not possible?

Thanks

[rusty1s]

HypergraphConv should support batch learning by diagonal stacking hypergraphs. Importantly, if you apply hypergraph learning in a mini-batch scenario, you need to adjust Data.__inc__ to incorporate for the bipartite incidence matrix. See here on how to do so.

[FernandezR]

Can you give an example of how I'm supposed to stack the hypergraphs diagonally?

Also, the bipartite edge index example doesn't help me with the hypergraph. I don't have source and target nodes so I'm not sure how that is supposed to be modified to work with hypergraphs.

[FernandezR]

Also, how do we pass the data loader object to the layers during training? I haven't been able to find an example.

Is the purpose of the dataloader to merge all the data together into one tensor for data and one tensor for hyperedge index?

[FernandezR]

So this is what I need? And then I just pass in the stacked batch tensors created by the loader to the HyGNN layer?

class HypergraphData(Data):
    def __inc__(self, key: str, value: Any, *args, **kwargs) -> Any:
        if 'batch' in key:
            return int(value.max()) + 1
        elif 'index' in key or key == 'face':
            return torch.tensor([[self.num_nodes], [self.num_edges]])
        else:
            return 0

[rusty1s]

Yes, this looks nearly perfect. Not sure if self.num_edges as it probably refers to the number of elements in edge_index, not necessarily the number of hyperedges (the second dimension of the incidence matrix).

[FernandezR]

Yes, I found that while testing, you need to add a redefinition of the num_edges function. Though, for some reason, the num edges function doesn't carry over to the batch data that is created. Though, it doesn't affect the conversion back to a data list from the batch data object. I can't see how to fix that in the batch data object, as some dynamic inheritance is used to create the batch data object type.

class HypergraphData(Data):
    def __inc__(self, key: str, value: Any, *args, **kwargs) -> Any:
        if 'batch' in key:
            return int(value.max()) + 1
        elif 'index' in key or key == 'face':
            return torch.tensor([[self.num_nodes], [self.num_edges]])
        else:
            return 0

    @property
    def num_edges(self) -> int:
        return int(self.hyperedge_index[1][-1]) + 1