[Feature] Generalized Mean Pooling

[Atharva-Phatak]

🚀 Feature Request

I would catalyst to have an implementation of GeM(Generalized Mean Pooling) as proposed in this paper

Motivation

Well I like to use catalyst's contrib features very much and it has an amazing set of pooling methods available(GlobalMaxPool2d ,
GlobalConcatPool2d ,etc). I think that an implementation of GeM pooling will be a nice addition to the pooling module of catalyst.

Note

Goes without saying that I will work on this implementation myself

[github-actions]

Hi! Thank you for your contribution! Please re-check all issue template checklists - unfilled issues would be closed automatically. And do not forget to join our slack for collaboration.

[Scitator]

Hi,

The idea with the contrib pooling looks great for me (especially for our representation learning track). In this case, if you have time and motivation – it would be a valuable contribution. As far as there is already an implementation from the authors, I hope it would be a bit easier to add to the contrib. Nevertheless, ensure please that you have added docs and a few tests for new functionality. I know, that is not the most interesting part of the contribution, but the most long-term one for the user adoption.

[Atharva-Phatak]

Definetly, I will create a pull request including documentation and the required tests along with accepted implementations

Also since you mentioned representation learning for catalyst, would like to add more modules/Loss functions ? Like arcface , cosface, etc?

[Scitator]

The interesting thing... we already have many arcfaces implemented and tested. The same for metric learning.

That's why all modules/losses/approaches for representation and metric learning are always welcome in the Catalyst ;)