Closed kangqiyue closed 1 year ago
The feature size of feats
and new_feats
are not necessarily the same. Therefore we use a linear layer to project the input features.
@mufeili ok, I understand, thanks!
In addition, I want to remind you that the linear layer was adopted in the class of GCNLayer, therefore one linear layer is created for each layer of the final GCN model. That is what is confusing me, because I think one linear layer for each GCNLayer means adding lots of extra parameters. However, in my case, I just remove the linear layer, considering that I have "fixed" dimensions.
Thanks again!
Glad to know that!
Hello, I recently used the GCN model in DGL-LifeSci, and I noticed that "residual connection" is implemented. However, I have some questions on the code implemented in this link:
https://github.com/awslabs/dgl-lifesci/blob/f854adf6f41ec5825f76783fb64c2b59ab520ef5/python/dgllife/model/gnn/gcn.py#L85
The residual connection was performed for the features produced after GraphConv layer. However, I think the residual connection should be performed to the original features (named as "feats" in the above code). I think the code should probably be written as follows (maybe):
Could you check the code or give out the reason why "residual connection" was implemented as the first style, instead of the second style? Thanks!