Wrapper Prior / Distribution class

[wd60622]

Personally don't like writing the priors in big dictionaries. Inspired here by the bambi API

mu = Prior("Normal")
sigma = Prior("HalfNormal")
hierarchical_parameter = Prior("Normal", mu=mu, sigma=sigma, dims="channel", centered=False)

Honestly think this will be way shorted than dictionaries plus bring in goodies like:

• initialization check for distribution name in pm module
• make use of pymc distributions having defaults
• dims will broadcast / are possible
• quick distribution creation used under hood or for checks:

  with pm.Model(coords={"channel": ["A", "B", "C"]}):
      alpha = hierarchical_parameter.create_variable("alpha")

This could be a documented submodule and would help demystify what goes in dictionary instead of any type of key!

This can also support post creation transformations. For instance:

positive_hierarchical_parameter = Prior("Normal", mu=mu, sigma=sigma, dims="channel", centered=False, transform="exp")

And some support for to and from dictionary for the serialization requirement.

Just some thoughts after typing out many large dictionaries in #747
CC: @carlosagostini @juanitorduz

[juanitorduz]

Yes! I love this idea! These dictionaries are growing up a lot and becoming hard to maintain! (also ... Pydantic haha)

[wd60622]

Yes! I love this idea! These dictionaries are growing up a lot and becoming hard to maintain! (also ... Pydantic haha)

Big fan of pydantic. How does kwargs in intialization method work though?

Some initialization keys would be same:

• distribution
• dims
• centered
• transform

Then parameter specific would change and might be handled with **kwargs

[wd60622]

Any thoughts here @ColtAllen

[wd60622]

Think it would be interesting to also wrap it into #706

mu = Prior("Normal", mu=0, sigma=1).constrain(lower=5, upper=10)

mu.kwargs # {"mu": 7.5, "sigma": 1.25}

[ColtAllen]

Any thoughts here @ColtAllen

It's certainly more convenient than dealing with a bunch of nested dicts in a custom model config. Is this an issue unique to pymc-marketing? If not it'd make more sense as a pymc addition.

[wd60622]

It's certainly more convenient than dealing with a bunch of nested dicts in a custom model config. Is this an issue unique to pymc-marketing? If not it'd make more sense as a pymc addition.

I think it could definitely be used elsewhere. I think it can start here, can be pitched to larger PyMC group, then go from there