[WIP] distribution classes

[pavanramkumar]

closes #350

TODOs

• make it work for all distribution classes
• add tests for BaseDistribution class
• make tests and examples pass locally
• remove unused hidden methods from pyglmnet.py (currently commented out)
• migrate simulate_glm() functionality into simulate() methods of respective classes
• migrate metrics definitions that depend on log_likelihood into distribution classes (pseudo_R2, deviance)
• add an example to show how users should roll out their own distribution class (e.g., using jax numpy)
• ensure Probit and Binomial work correctly
• check if all docstrings still make sense
• update API reference
• update what's new

[geektoni]

💃 💃

[jasmainak]

you probably don't need to redefine it since you're inheriting from BaseDistribution ?

[jasmainak]

exciting stuff! I think it significantly improves our repository :-D

I would love to see an example with autograd or something like that

[geektoni]

I wonder if we can change this huge if-else to something in which we dynamically instantiate the correct class given the string. I was thinking of something like this https://softwareengineering.stackexchange.com/a/351394, in which we have a dictionary which is holding all the required classes.

[jasmainak]

@pavanramkumar I'm not sure this is legit. The log_likelihood function for Binomial does not call self.mu or does it? Is the log likelihood the log likelihood taking into account the link function?

[geektoni]

Quick question. If I wanted to contribute on this one here, should I fork your repo @pavanramkumar or do you know if there is any way to push directly to this PR?

[jasmainak]

I think forking might be the easiest. But maybe @pavanramkumar should confirm first if he has any local changes that have not been pushed?

[pavanramkumar]

Thanks for picking this up @geektoni ! Please go ahead and fork.