Closed eddiezha closed 1 year ago
Hello, currently the DIG is based on the torch_geometric
and the explainability relies on the MessagePassing class. It's a pity that the dig explainability can't be applied to dgl model directly.
One possible way is to transfer your dgl model into torch_geometric model. That's to say, write the model with the torch_geometric framework and then load the parameters from the dgl model.
For the second question, I think some model-agnostic explainability method such as SubgraphX
can be applied to the Relational Graph Convolutional Network.
Thanks, transfer to torch_geometric seems difficult, I am wondering if there is any quick way to transfer dig explainability to dgl? Thanks.
I think your implementation cannot be used for RGCN right? It does not include edge type.
Yes, I think it is a hard work to transfer dig explainability to dgl, and current implementation is not for GNN considering edge features like RGCN
.
Hi, my GNN is trained on DGL, it seems your lib does not support this. Is it possible to add this feature? Or you mind providing a quick way to do it? Thanks. (Specifically for explainability.) BTW, a mode general question, do these methods support Relational Graph Convolutional Network?