Use of HEATConv with HeteroGraph

[SimonCrouzet]

Hello everyone, and thanks for your work on the package,

I am encountering some troubles regarding the implementation of a model using HEATConv layers to perform message passing on an HeteroGraph. The layers are defined as Heterogeneous Edge-enhanced Graph Attentional operators, and should be useful on heterogeneous graph, but the expected input format make it really hard to be used from an heterograph: format() takes as argument forward(x: [Tensor], edge_index: [Union][Tensor, SparseTensor], node_type: [Tensor], edge_type: [Tensor], edge_attr: [Optional][Tensor]) while when using HeteroGraphs x has the format of a dict {node_type => x}, same for edge_index and edge_attr.

Is there any simple way to use HEATConv with such graph? Or am I just missing the point and we should implement GATConv directly followed by the use of to_hetero(model, metadata, aggregation_type) as HEATConv is designed to replicate an attentional mechanisms on graphs originally not heterogeneous?

Thank you for your help,
Best,

[rusty1s]

You can use hetero_data.to_homogeneous() to merge your heterogeneous graph into a single edge_index and x representation with corresponding node_type and edge_type vectors.