How to initialize edge feature/edge index tensors in Heterogeneous Graphs for edges with no features?

[stefanschutz]

Hi everyone, I have some questions regarding creating heterogeneous graphs. In the case that I have only one node type and multiple edge types, and only 1 edge type has edge features, should I initialize the empty tensors for those edge types that don't have edge features? For example:

data = HeteroData()
data["node"].x = ... # [num_nodes, num_node_feats], only 1 type

data["node", "type1", "node"].edge_index = ... #[2, num_edge_type1]
data["node", "type2", "node"].edge_index = ... #[2, num_edge_type2]

data["node", "type1", "node"].edge_attr = ... #[num_edge_type1, num_edge_type1_feat]
data["node", "type2", "node"].edge_attr = ... # zeros-tensor or no init?

Another question popping up is some graphs might not have a specific edge type, which leads to the edge_index of that type having empty tensor. In this case, using utils.to_dgl to convert pyg object to dgl will have an error since it assumes edge_index of all edge types are not empty. What could be a good solution for handling this case for learning on GNNs?

data["node", "type3", "node"].edge_index = ... # [0]
dgl_data = to_dgl(data) # ValueError: not enough values to unpack (expected 2, got 0)

Thank you in advance!

[rusty1s]

In general, node types/ edge types / their features need to be consistent across graphs. If an edge type of all graphs does not have edge features, then you don't need to specify them.

[stefanschutz]

Thanks for the quick reply! And sorry for my sudden update, could you also answer the second question about the index as well? I believe this would cause some inconsistency for the edge index if converting to dgl.

[rusty1s]

The answer would be the same, right? edge_index would need to be defined for the edge types that exist in the union of graphs.

data[edge_type].edge_index = torch.empty(2, 0, dtype=torch.long)

Maybe I am misunderstanding the question.

[stefanschutz]

Initializing with the size torch.empty(2, 0, dtype=torch.long) works for me. I left the empty edge_index with long tensor([0]), which causes the mismatched size when converting. Thank you!