remove find_contrained_prior

[aloctavodia]

find_contrained_prior was deprecated in v5.17.0 (https://github.com/pymc-devs/pymc/pull/7458), released on Oct 3, 2024. This PR removes it completely.

---

📚 Documentation preview 📚: https://pymc--7680.org.readthedocs.build/en/7680/

[codecov[bot]]

Codecov Report

All modified and coverable lines are covered by tests :white_check_mark:

Project coverage is 92.66%. Comparing base (358b825) to head (c66901d). Report is 2 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #7680      +/-   ##
==========================================
- Coverage   92.70%   92.66%   -0.05%     
==========================================
  Files         107      106       -1     
  Lines       18391    18279     -112     
==========================================
- Hits        17050    16938     -112     
  Misses       1341     1341              

Files with missing lines	Coverage Δ
pymc/__init__.py	100.00% <ø> (ø)

... and 2 files with indirect coverage changes

[twiecki]

Can you link the issue about the deprecation?

[ricardoV94]

How long ago was v5.17 released? Time is more important than version numbers

[aloctavodia]

Comment updated

[twiecki]

I now I should have commented on the original issue but can't we add the max entropy condition to find_constrained_prior? I'm not even sure what the alternative workflow is to do this with preliz but it seems to require it's own, different set of distributions.

[aloctavodia]

PreliZ's distributions have a to_pymc method

[twiecki]

How difficult would it be to port max_ent to find_constrained_prior? Can you share example code of how this would look like in preliz?

[aloctavodia]

If you want to rewrite find_constrained_prior to behave like maxent, then entropy needs to be added to distributions and then I would recommend to also use an initialization routine like the one used in maxent so users don't have to guess initial values themselves.

This will compute the maxentropy Gamma distribution that has 90% of the mass between 1 and 10, and it will return it as a PyMC distribution, that you can then use as usual.

pz.maxent(pz.Gamma(), 1, 10, 0.9).to_pymc()

You can also do other stuff like fix some of the parameters (mu in this example)

pz.maxent(pz.Gamma(mu=4), 1, 10, 0.9)

Or stats like median or mode,

pz.maxent(pz.Beta(), 0.1, 0.7, 0.94, fixed_stat=("mode", 0.3))

More examples here. https://preliz.readthedocs.io/en/latest/examples/gallery/direct_elicitation_1D.html

PreliZ's distributions follow the same parameterization used by PyMC. For instance, Gamma can be defined using, alpha-beta, or mu-sigma.

[twiecki]

Would it be too hacky to replace find_constrained_prior with a call to preliz? I assume since you have code that translates preliz distribution to pymc you could have a function to tranlsate a pymc distribution into preliz.

[aloctavodia]

We did that with arviz for a long time, so as hacky as that. There is no from_pymc method, but I guess it should not be that hard.

[twiecki]

I think that could be worth considering, just because find_constrained_prior is quite useful and also by now used in quite some training materials. And just dropping it without giving users an obvious replacement doesn't seem like a great experience.