Support more distributions as random variables

[eigenfoo]

Filing this as an issue so we don't forget: #41 (comment)

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:

• Flat
• HalfFlat
• SkewNormal
• HalfStudentT
• Weibull
• Wald (a.k.a. InverseGaussian)
• ExGaussian
• LogitNormal
• Interpolated

Discrete:

• BetaBinomial
• ZeroInflatedBinomial
• ZeroInflatedPoisson
• ZeroInflatedNegativeBinomial
• DiscreteUniform
• Constant (EDIT: actually, tfp does have this, but it's called tfp.distributions.Deterministic... which may be problematic for us 😝)
• DiscreteWeibull
• OrderedLogistic

Multivariate:

• MatrixNormal
• KroneckerNormal
• LKJCholeskyCov (EDIT: the tfp "LKJ" distribution is the same as "LKJCorr", though)

Timeseries:

• All Timeseries distributions

[junpenglao]

LKJ in tfp is the same as LKJCorr.

[eigenfoo]

Edited!

[eigenfoo]

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]

@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]

FYI: LKJ in tfp now has Cholesky parameterization tensorflow/probability@9b007d8

[tblazina]

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]

@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 😄

[tblazina]

@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]

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]

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 😄, but I do like the idea of opening smaller issues and referencing those in the PR.