Possible discrepancy between _NodeConv.forward() implementation and paper equation

[Snitro]

Hi all,

Thanks for the great work!

I noticed a potential inconsistency between the implementation of the _NodeConv class and the equation described in the paper.

Current implementation:

    def forward(self, graph, nfeat, efeat):
        h = (1 + self.eps) * nfeat
        h = self.gconv(graph, h, efeat)
        h = self.mlp(h)
        h = F.leaky_relu(self.batchnorm(h))
        return h

However, according to the equation in the paper:

It seems that the residual connection is added after the graph convolution, which might suggest that the code should be:

    def forward(self, graph, nfeat, efeat):
        h = self.gconv(graph, nfeat, efeat) + self.eps * nfeat
        h = self.mlp(h)
        h = F.leaky_relu(self.batchnorm(h))
        return h

Could you please confirm whether this is an intentional design choice or a possible deviation from the original formulation?

Thanks!

[Snitro]

Hi, sorry for the confusion earlier — I forgot to include the __init__ method in the previous snippet. In the full implementation, I do set residual=True when initializing NNConv. Here's the relevant part of the __init__:

self.gconv = NNConv(
    in_feats=node_feats,
    out_feats=out_feats,
    edge_func=nn.Linear(edge_feats, node_feats * out_feats),
    aggregator_type="sum",
    residual=True,  # <- residual is enabled here
    bias=False,
)

So the complete version includes the residual connection as intended.