Fix the Kendalls Tau metric when used in graph mode

[nicolaspi]

Description

Blocked by #2740

Brief Description of the PR:

The metric was only working in eager mode. This fix makes it work in graph mode too.

Type of change

• Bug fix

Checklist:

• I've properly formatted my code according to the guidelines
  • By running Black + Flake8
  • By running pre-commit hooks
• This PR addresses an already submitted issue for TensorFlow Addons
• I have made corresponding changes to the documentation
• I have added tests that prove my fix is effective or that my feature works
• This PR contains modifications to C++ custom-ops

How Has This Been Tested?

The metric was giving NaNs when running in graph mode. I modified the test to check against the NaN values.

[bot-of-gabrieldemarmiesse]

@sorensenjs

You are owner of some files modified in this pull request.
Would you kindly review the changes whenever you have the time to?
Thank you very much.

[sorensenjs]

Looks great - thanks for this change.

[nicolaspi]

I removed a padding op that seemed unnecessary.

[bhack]

@theadactyl Just another TF ecosystem duplication reminder, we have this also in TFP:
tensorflow/probability#1417

[sorensenjs]

@theadactyl Just another TF ecosystem duplication reminder, we have this also in TFP: tensorflow/probability#1417

This is incorrect - the tfp implementation (1) can't be used as a keras metric (2) is not approximate. I'm the author of both, don't blame me for the complexities of TF packaging.

This particular implementation is an online O(n) approximate algorithm that is much better suited for use as a metric in machine learning. The TFP implementation is an O(n^2) algorithm.

[nicolaspi]

The failing check looks to come from a regression in tensorflow_docs tensorflow/docs@eaef2e3 ? (See #2740)

[bhack]

The failing check looks to come from a regression in tensorflow_docs tensorflow/docs@eaef2e3 ?

/cc @MarkDaoust

[bhack]

@theadactyl Just another TF ecosystem duplication reminder, we have this also in TFP: tensorflow/probability#1417

This is incorrect - the tfp implementation (1) can't be used as a keras metric (2) is not approximate. I'm the author of both, don't blame me for the complexities of TF packaging.

This particular implementation is an online O(n) approximate algorithm that is much better suited for use as a metric in machine learning. The TFP implementation is an O(n^2) algorithm.

I was exactly referring to the ecosystem boundaries and packaging.

See our original 2020 thread:
#2169 (comment)