where is the edge_norm argument of RGCNConv now ?

[RenzeLou]

Dear authors,

I found an argument of the RGCNConv class in the old version of PyG (v 1.3.0), namely edge_norm (as shown below):

I guess this argument should be the problem-specific normalization constant as claimed in equation (2) of the original paper:

However, I didn't find it in the latest version of PyG (v 2.0.2). It seems like this API has changed a lot.

So, how can I set the edge_norm now (i.e., the latest version of PyG)?

[rusty1s]

Yes, the edge_norm attribute was removed in favor of supporting different aggr methods (mean, add, max), which should model most of the use-cases of edge_norm. The goal was to make applying mean aggregation much more easier :)

[RenzeLou]

Thank you for your kind reply！

Nevertheless, I want to use the edge_norm to make the weighted aggrgation (i.e., set different normalization constants dynamically). Does the current version of PyG still support this operation?

[rusty1s]

No it does not, I'm sorry. We need to add it back in. Let me know if you have interest in working with us together on this.

[RenzeLou]

Ok, maybe I can try.

[sailist]

Hi, I am now reimplement a paper which use edge_norm attribute.

If given

out = self.conv1(x, edge_index, edge_type, edge_norm)

Is there any equal method I can workaround? Like

out = self.conv1(x, edge_index, edge_type)
out = norm(out, edge_norm)

[rusty1s]

Currently no, we will likely bring this feature back as part of our RGCN refactor. Please keep reminding me :)

[rusty1s]

Relevant dicussion: #5217