Open yippp opened 4 years ago
You are right that we currently do not provide any graph generation models. That is a current limitation.
The encoder-decoder scheme of the GraphVAE should be quite easy to implement with the help of PyTorch Geometric. In the end, this is just a stacking of GNN layers, global pooling, and MLPs. The hard part is to implement its reconstruction loss where you need to find the binary assignment matrix X
first.
Thanks for your hint. Hope you can consider to add example for GraphRNN or GraphVAE. Nodes generation is really important for graph generation task. I think an example will inspire people to develop related model much easier.
❓ Questions & Help
I have viewed the example codes in
example/
, the examples for graph autoencoders are about edges generation for link prediction. It seems that there is no example for nodes generation. I have no idea how to implement both nodes and edges generation network like molecule generation. Could you provide some relative exmaple, such as implementation of GraphRNN and GraphVAE? Also, I will be pleasure if you could give me some hint usingtorch_geometric
to implement them.