[Feature] Additive Margin Softmax

[Atharva-Phatak]

🚀 Feature Request

I would like to add Additive Margin Softmax to its amazing list of various 'Face' losses.

Motivation

Catalyst has nice support for metric learning and I would like catalyst modules to be expanded more and possibly have most SOTA implementations of Face metrics.

Proposal

Just like how ArcFace and CosFace are implemented in the catalyst library, I would like to add AM-Softmax to this list as well. Possibly other metric learning modules if provided the opportunity.

Additional context

Additive margin softmax was used in 2nd place solution of a Kaggle competition.

Checklist

• feature proposal description.
• motivation.
• extra proposal context / proposal alternatives review.
• I know, that I could join Catalyst slack (#__questions channel) for issue discussion.

Note

This goes without saying I want to implement it

[Scitator]

Hi,

the idea looks great, especially if you are interested in the implementation contribution :)

I also think @ditwoo would be interested in too.

[Atharva-Phatak]

Hi, yes I have a working implementation ready, I can write up some tests for it and create a pr.

[Scitator]

That's would be great,
btw, you could check current tests for the "Faces" family here

[Atharva-Phatak]

I will take look at it for sure.

[Atharva-Phatak]

@Scitator I took a look at the tests, the tests seems to be well designed. So no worries. I will start with implementation and create PR soon.

[Atharva-Phatak]

@Scitator can you give me some tips on how to pass the codestyle ?
Here are the steps I followed.

1. I forked the latest repo and cloned it
2. installed the dependencies for requirements.txt and requirements-dev.txt
3. I run catalyst-make-codestyle first and the catalyst-check-codestyle
4. Files seem to be reformatted but flake8 errors persist. Also, there are no errors in the file which I added.(Last time also the same thing happened with codestyle, hence I want to do it perfectly this time.)

Hence I wanted some help to make codestyle do not throw any error.

[Scitator]

sorry, but could not reproduce the issue: https://colab.research.google.com/drive/1JCGTVvWlrIsLXMPRRRSWiAstSLic4nbA?usp=sharing

[Atharva-Phatak]

@Scitator Thank you now my codestyle is compatible. But when I run tests using pytest .. I get the following error
ImportError: cannot import name 'SummaryWriter' from 'tensorboardX' (unknown location)

Here are the specs of my env
platform linux 20.04 -- Python 3.8.5, pytest-5.3.1, py-1.10.0, pluggy-0.13.1

Any tips to resolve it ?

[Scitator]

I think, you still have to reinstall the requirements TensorboardX is must-have one :)
https://github.com/catalyst-team/catalyst/blob/master/requirements/requirements.txt#L13

[Scitator]

looks like merged ;)