Multi-inputs for HeteroGraphConv?

Questions
,
#1

Hi, I have a problem that if the neural networks in mods of dglnn.HeteroGraphConv require two more inputs, such as node_features and edge_features. For example,

mods = {
    'A-B': dglnn.CFConv(3, 3, 3, 3), 
    'B-C': dglnn.CFConv(3, 3, 3, 3), 
    'A-C': dglnn.CFConv(3, 3, 3, 3)
}

model = dglnn.HeteroGraphConv(mods=mods, aggregate='sum')

dglnn.HeteroGraphConv.forward receive g and feats, but dglnn.CFConv receive g, node_feats and edge_feats. How to handle with this condition? Thank you.

#2

Currently you will need to write a module yourself to support that. Note that CFConv currently only supports graphs of a single node type and you may need to modify that as well.

#3

HeteroGraphConv's forward has two additional argument mod_args and mod_kwargs. You can give a dictionary whose keys are edge types and values are tuples for mod_args (or dicts for mod_kwargs).

However as far as I know CFConv does not work on heterogeneous graph, so you will need to write your own module for that.

#4

Thank you very much. Is there an example in dgl that implement a GCN which supports multi node types?

#5

Sorry for the late reply, see this PR.

#6

Thank you very much. I’ve fix this problem. :rofl:

#7