why using output of relu for rebuilding knn graph

[chenchao_xu]

I have a question about rebuilding knn graph in gcn-v inference.
Why using the output of relu layer, not that of linear layer to get feat for rebuilding knn graph, since there are lots of zeros after passing feat map through the relu layer.

[Lei Yang]

I think ArcFace reports some experiments about how to choose the output of feature extractor as face features.