Differentiable Group Normalization.

[wsad1]

🚀 Feature

This paper proposes a new normalization technique. Where the nodes in a group are normalized independently and the group assignment is learnable.

It also proposes two metrics to measure over smoothing.

Motivation

The results look good in comparison to no norm(NN), pair-norm(PN) and batch norm(BN). But the paper doesn't compare results of other normalizations techniques.

Further the metrics for over smoothing might be useful to analyze gcn's.

Additional Context

torch-geometric has a few normalization layers, so it might not be too useful to add this. But i feel the over smoothing metrics could be a good addition. Do let me know your thoughts on this.

[rusty1s]

Interesting, are you interested in contributing?

[wsad1]

Yes, i was planning to contribute this feature.