Using the Explainer model with our own models

[alzaia]

Hello, I'm having a bit of trouble using the Explainer (e.g. GNNExplainer model) on top of my PyG homemade GNN model. The issue is that the Explainer assumes a certain input args in the forward method of the model:

pytorch_geometric/torch_geometric/nn/models/explainer.py

Line 237 in 48ae505

out = self.model(x, edge_index, **kwargs)

But the args in my user-defined model are different (I just pass the batch and then use whatever I need from that (e.g. batch.x, batch.edge_index, batch.edge_attr, ...). I assume this can also cause issues if used with some of your pre-defined models. For example:

pytorch_geometric/torch_geometric/nn/models/lightgcn.py

Line 101 in 48ae505

def forward(self, edge_index: Adj,

Is this a limitation with the current implementation of the Explainer?

[alzaia]

I guess this is directly related to this discussion: #5814

[rusty1s]

Yes, this is a current limitation. We can either use the first two arguments to expect node features and edge indices. Alternatively, we can fix name convention such that x and edge_index needs to be present in the input arguments but in any order.

I think it is generally bad practice to use Data as input to model.forward though, especially since it is not well-defined which attributes in Data exist to make your model run successfully. I think something like

def forward(self, x, edge_index, ...):
    pass

model(data.x, data.edge_index)

is easier to understand.

[alzaia]

Thanks a lot for the feedback!