Wrap tfp distributions as pm RVs

[eigenfoo]

(Following up @twiecki's suggestion)

This PR wraps tfp.distributions as pm.RandomVariables. We enter the list of tfp distributions we want as strings in __all__, and the for loop at the end of the module should define the necessary classes.

Some notes:

• I assumed that we would name our RVs the exact same name as the tfp distributions (e.g., it's now Chi2, not ChiSquared).
• I made the list of distributions by just comparing what RVs we have in PyMC3 right now. I'll double check this list later today.

[ColCarroll]

Cool! I guess since you set __all__, this allows tab completion?

I might use

setattr(sys.modules[__name__], dist_name, getattr(tfp.distributions, dist_name))

because I like setattr and getattr better. Would that even work? This suggests they are similar, and gives no prohibition against either. I guess take this as an observation that there's another way to do it, then!

Looks good to merge, but I'll give it a few more hours in case there's input.

[eigenfoo]

@ColCarroll That does indeed work! I like getattr and setattr better: I didn't know that they worked at the module level. I've changed it to do that.

And no, I manually typed out the distributions, since I use vim with very few plugins 😛

[eigenfoo]

Looking carefully, tfp does not implement the following distributions, which are currently supported in PyMC3. There are some fairly important ones here: we'll need to decide what we want to do there (i.e. implement the RVs ourselves, submit PRs to tfp, or drop support for the obscure distributions). We should go ahead and merge this PR, and deal with the rest of these distributions in subsequent PRs.

Continuous:

• Flat
• HalfFlat
• SkewNormal
• HalfStudentT
• Weibull
• Wald
• 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 and LKJCorr (there's just an "LKJ" distribution, though)

Timeseries:

• All Timeseries distributions

[ColCarroll]

This will be nice, not adding them piecemeal. Thanks, @eigenfoo!

[majidaldo]

What happens if there is a distribution that can't be based of a (existing) TFP distribution? TFD Distributions have so many methods. In PyMC3 you could at least attempt to provide your own .log_p() and sample().