TST: test all rvs, not just tfp supported distributions

[eigenfoo]

Closes #52

Building off of #49, this PR adds tests for all current distributions, not just those supported by tfp.

[coveralls]

Pull Request Test Coverage Report for Build 170

• 0 of 0 changed or added relevant lines in 0 files are covered.
• No unchanged relevant lines lost coverage.
• Overall coverage remained the same at ?%

---

Totals
Change from base Build 168:	0.0%
Covered Lines:
Relevant Lines:	0

---

💛 - Coveralls

[eigenfoo]

Well this caught a few bugs, mostly to do with using the wrong dtype (e.g. tf.float32 instead of tf.float64). I'll try and debug this sometime this weekend.

@canyon289 is it possible to get Coveralls to not comment like that? With PyMC3 it would should up in the same place that the Travis CI is.

[twiecki]

Disabled coverall comments.

[canyon289]

I think we need to split apart the tests for the tfp implemented distributions from the ones that pymc4 implements because for the pymc4 implemented distributions we do need to check for correctness, not just compilation

[twiecki]

@canyon289 I think we can keep these here but make sure we also add correctness tests in addition.

[eigenfoo]

Currently the following code snippet fails with a NotImplementedError for the inverse_log_det_jacobian, in bijector.py. I propose we just comment out the DiscreteUniform distribution (and its tests) until tfp implements the relevant functions. I'm working on making sure the other distributions are suitably tested.

In [1]: import tensorflow_probability as tfp
   ...: from tensorflow_probability import distributions as tfd
   ...:
   ...: x = tfd.TransformedDistribution(distribution=tfd.Categorical(probs=[0.5, 0.5]),
   ...:                                 bijector=tfp.bijectors.AffineScalar(shift=1),
   ...:                                 name='DiscreteUniform')
   ...:
   ...: x.log_prob(2)

[eigenfoo]

(Finally) ready for review.

I excluded the ZeroInflatedBinomial distribution from the regular tests for the same reason @canyon289 excluded the Binomial distribution: there's an upstream problem with tfp.

I've also commented out the DiscreteUniform distribution, and added TODOs at the appropriate places, as discussed above.

[twiecki]

Looks great, thanks!

We should open an issue with TFP about the problems we found.

[eigenfoo]

TFP issue filed here: tensorflow/probability#286

[twiecki]

Nice, looks like there is a simple solution available.