Add more tests for single distributions

[xukai92]

• Uni
• Multi
• Matrix

[willtebbutt]

I appreciate that this is a WIP, but I've added a few comments anyway. Is there an obvious way to construct some logpdf tests? I'm not really sure what the easiest way to go about this is, but it would be cool if we could do something.

[xukai92]

By logpdf tests do you refer to more complex models or single distributions?

[willtebbutt]

By logpdf tests do you refer to more complex models or single distributions?

Fair point. If we're only considering simple distributions in this PR then we are essentially just trying to ensure that Turing hasn't screwed anything up and, therefore, can just compare against the Distributions.jl logpdf for the component distributions.

[xukai92]

Yes. Actually I'd prefer to add them into this https://github.com/TuringLang/Turing.jl/blob/master/test/ad.jl/AD_compatibility_with_distributions.jl file, as they didn't need any sampling. Also, maybe we can leverage https://github.com/JuliaStats/ConjugatePriors.jl to construct tests more than single distributions - I will try to do this as well.

[xukai92]

@mohamed82008 I kept seeing numerical errors from Bijectors.jl (https://travis-ci.org/TuringLang/Turing.jl/jobs/474636293#L873) when sampling for single Dirichlet distribution with NUTS. I didn't manage to figure out what happens here - any idea?

[xukai92]

I guess it may be related to this test case https://github.com/TuringLang/Bijectors.jl/blob/master/test/transform.jl#L51, for which in Turing.jl we don't have a special rand for those dimensions belonging to Dirichlet distributions.

Sorry this is irrelevant.

[xukai92]

I found a possible cause and created an issue in Bijectors.jl (see TuringLang/Bijectors.jl#12).

[xukai92]

@yebai @willtebbutt As the numerical issue on Dirichlet is a separate issue, can we merge this PR if there is no objection?

[yebai]

@mohamed82008 Could you take a look at the numerical issues identified by this PR? Ideally, we should fix them before merging this PR.

[mohamed82008]

Sure thing. I was planning to take a look a while back but forgot! I will see if I can reproduce the error first.

[mohamed82008]

All tests pass on my machine.

[yebai]

All tests pass on my machine.

I've seenBeta and Gamma tests randomly break on Travis, see

https://travis-ci.org/TuringLang/Turing.jl/jobs/475012349

[xukai92]

I guess it's because of the end-of-window update that tunes the step size to a very high value (https://travis-ci.org/TuringLang/Turing.jl/jobs/475012349#L780). Let me cap that as well.

[yebai]

@xukai92 all the tests are now passing. Do you want to resolve Mohamed's comment before this PR gets merged?

[xukai92]

@yebai Thanks for the reminding - resolved.

[willtebbutt]

@xukai92 I don't want to too much additional work as this PR looks great, but there do appear to be a few distributions missing from this list e.g. InverseWishart. Is there a particular reason for this?

[xukai92]

@willtebbutt I added more dists now!

[xukai92]

Samples from VonMises(0, 1) seems to be biased (2 runs of tests on my local fails).

[mohamed82008]

Rebased on master.

[xukai92]

Let's merge this one if not objections

[mohamed82008]

@yebai It would be nice to merge this before the other small PRs which are also ready to merge.

[xukai92]

@mohamed82008 I merged it :)