Multigraph Data Structure coding for GDAS NAS

[buttercutter]

how to code the following graph structure ?
which kind of data structure is suitable in this case for both forward inference and backward propagation ?

D-X-Y/AutoDL-Projects#99 (comment)

https://gist.github.com/promach/b6f526c56e20f029d68e6f9041c3f5c0#file-gdas-py-L106-L107

[rusty1s]

Representing multigraphs is well supported by simply duplicating edges in edge_index. Does that answer your question?

[buttercutter]

However, does the multigraph have support for both forward inference and backward propagation operations ?

[rusty1s]

edge_index is a LongTensor, so it cannot track any gradients. If you want to do that, I suggest to let gradients flow through an additional edge_weight tensor that determines the strengths of each edge.