ArcFace Loss Implementation

[aylinaydincs]

• This pull_request is a part of Google Summer of Code Project mentored bt @monatis at TensorFlow Organization.
• In this pull-request, I implement ArcFace Loss, and I also provide a sample notebook to show how to use it standalone and model.compile()
• I wrote a introductory medium story and a more comprehensive and advanced one will be coming soon
• I wrote unit tests for serializaiton and loss calculation

Ps. Previous Pull Request Deleted, due to the conflict of commits

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB

[owenvallis]

Thanks for the PR. Looks like there are duplicate files again for the loss and the tests, and I think you'll need to run the Black and Isort formatting (the imports in the init.py look out of order again).

Also let me know your thoughts regarding adding the sub-centers layer. I think the weights that map the embedding to the N classes (and optionally the K sub-centers) needs to be a Dense layer so that we can train the weights. The current implementation has the matrix as a variable in loss so I don't think the random values will be updated during training.