Add numpy like percentile and quantile

[Dhruvanshu-Joshi]

Description

This pr aims to add support for numpy like percentile and quantile functions.

Related Issue

• Closes Implement equivalent numpy median and quantile / percentile #53
• Related to #

Checklist

• Checked that the pre-commit linting/style checks pass
• Included tests that prove the fix is effective or that the new feature works
• Added necessary documentation (docstrings and/or example notebooks)
• If you are a pro: each commit corresponds to a relevant logical change

Type of change

• New feature / enhancement
• Bug fix
• Documentation
• Maintenance
• Other (please specify):

[codecov]

Codecov Report

Attention: Patch coverage is 95.91837% with 2 lines in your changes missing coverage. Please review.

Project coverage is 81.61%. Comparing base (c6d85d1) to head (77d8604).
Report is 9 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #942      +/-   ##
==========================================
+ Coverage   81.38%   81.61%   +0.22%     
==========================================
  Files         172      178       +6     
  Lines       46868    47285     +417     
  Branches    11423    11497      +74     
==========================================
+ Hits        38145    38592     +447     
+ Misses       6542     6504      -38     
- Partials     2181     2189       +8     

Files	Coverage Δ
pytensor/tensor/math.py	91.47% <95.91%> (+0.19%)	⬆️

... and 25 files with indirect coverage changes

[ricardoV94]

This is a but irrelevant, but does it make sense to define quantile first and then percentile from quantile. So the other way around?

Quantile seems more fundamental

[Dhruvanshu-Joshi]

Yes makes sense. I have pushed this.

[ricardoV94]

This won't work with symbolic q. Can we have a vectorized implementation?

[Dhruvanshu-Joshi]

Hi. Sorry this completely skipped me. I'll get back to it and see how to vectorise this and also look into the failing tests.

[ricardoV94]

This code could look a lot like the median one, with the exception that the k in x_sorted[..., k] will be an array if q is an array as well. And we can't use ifelse if q is an array, because then oddness/eveness may depend on the quantile (whether size / k is an integer). We have to use switch in that case.

We should be able to re-implement the median as partial(quantile, q=0.5) once this is done.

[Dhruvanshu-Joshi]

Hi @ricardoV94

I have added some changes. However I need some help in some parts.

For "ndim, axis, q" equal to (3, (1, 2), [0.1, 0.9]), the tests fail due to some shapes issues:

For percentile, numpy output:
[[0.14843302]
 [0.65699253]]

Our output:
[[0.14843302 0.65699253]]

Also, I am confused on how to introduce symbolic checks to ensure the all of the q is between 0 to 1 in case of quantile and 0 to 100 in case of percentile.

[ricardoV94]

Also, I am confused on how to introduce symbolic checks to ensure the all of the q is between 0 to 1 in case of quantile and 0 to 100 in case of percentile.

We can be a bit permissive and consider it undefined behavior if it makes our code simpler. Otherwise there's a RaiseAndCheck Op that can be used for symbolic checks

[ricardoV94]

For "ndim, axis, q" equal to (3, (1, 2), [0.1, 0.9]), the tests fail due to some shapes issues:

I'll try to take a look when I get some free time