search

pymc-devs / pymc4

Experimental PyMC interface for TensorFlow Probability. Official work on this project has been discontinued.
Apache License 2.0
713 stars 113 forks source link

Support more distributions as random variables #44

Open eigenfoo opened 5 years ago

eigenfoo commented 5 years ago

Filing this as an issue so we don't forget: https://github.com/pymc-devs/pymc4/pull/41#issuecomment-453775302

I think we can achieve most of the important distributions using tfp.distributions.Mixture and tfp.distributions.Bijector (e.g. the zero-inflated distributions are just mixtures with the zero distribution).

@ColCarroll I'm assuming there will be a problem with having a distribution/RV named Deterministic? 😄 That's what tfp calls the thing that PyMC calls the Constant distribution. If so, I can write something to remap names.

Continuous:

Discrete:

Multivariate:

Timeseries:

junpenglao commented 5 years ago

LKJ in tfp is the same as LKJCorr.

eigenfoo commented 5 years ago

Edited!

eigenfoo commented 5 years ago

Update: #46 merges in some more random variables (those that can be created via tfp's bijectors and transformed variables). There are still quite a few random variables outstanding, though.

ericmjl commented 5 years ago

@eigenfoo your PR gives me a template to work off! I'll see how fast I can get setup and running for development at work.

junpenglao commented 5 years ago

FYI: LKJ in tfp now has Cholesky parameterization https://github.com/tensorflow/probability/commit/9b007d8e52074fe973cf06290cf5411274636834

tblazina commented 5 years ago

I haven't contributed yet but would be interested to have a go with helping with some of these. Looking at the code base it seems a lot of these still are not implemented, so it is still open, correct?

ericmjl commented 5 years ago

@tblazina yes, still open! I recently learned a new pattern at this year's PyCon, which helps with tracking progress. Was wondering what the other devs think of this - when putting in a PR, put in also an issue that we can close, and reference it in the branch name (e.g. distribution-flat-#93)? I think there's github magic that we can take advantage of :smile:

tblazina commented 5 years ago

@ericmjl, ok great, thanks for quick response! I'll try to get dev environment all set up and see if I can get plugging away at some of this. I think your idea about tracking progress sounds to me and was along the lines of my next question, is that for a big issue like this with lots of different parts, is the best practice in this project to submit a PR for each distribution that we implement and simply reference this issue in the PR?

ericmjl commented 5 years ago

is the best practice in this project to submit a PR for each distribution that we implement and simply reference this issue in the PR?

That is how I would have done it before this year's PyCon. This kind of issue is like an over-arching one, which multiple other smaller issues can reference. Hence, I think it's more appropriate to reference the smaller issues, so that you can take advantage of the auto-closing keywords (e.g. "resolves #391")

tblazina commented 5 years ago

Ok, well for now (until there is a bit of consensus), I proceed as was done before, referencing the issue in the PR. Don't want to step on any toes if I'm just starting to contribute :smile:, but I do like the idea of opening smaller issues and referencing those in the PR.