Make default model evaluation independent of Turing

[phipsgabler]

The following simple evaluation does not currently work without Turing being loaded, since the tilde and assume/observe functions for the contexts provided by DynamicPPL are actually implemented in Turing:

using DynamicPPL@model coinflip(y) = begin
    p ~ Beta(1, 1)
    N = length(y)
    for i = 1:N
        y[i] ~ Bernoulli(p)
    end
end
model = coinflip([1,1,0])
vi = VarInfo()
model(vi)

As @mohamed82008 suggested, I basically just moved the "fallback methods" from Inference.jl over here. This required some additional movements:

• I prefixed the tilde metafunctions in compiler.jl with "generate", to avoid name confusion
• Some code had had to be split out into separate files and moved, to make imports work
• The pseudo-distributions NoDist and NamedDist had to be moved to this package, since the assume/observe code depends on them (crucially, LikelihoodContext doesn't work without NoDist).

For testing, I didn't really now what to do and where to put it, except checking it does compile now. So I just created a new file for now.

I'd appreciate feedback on especially imports and other things I broke :) The Turing counterpart (removing the fallback methods) will follow soon.

[trappmartin]

Look good to me. Thanks.

[phipsgabler]

One thing I was worried about is removing NamedDist from Turing. Is that (supposed to be) used anywhere by externals?

[trappmartin]

Not that I'm aware of.

[mohamed82008]

NamedDist is a feature. We should export it from Turing.

[mohamed82008]

NamedDist is used by DiffEqBayes.

[phipsgabler]

It will still be reexported from Turing. Same for assume/observe and their dot-variants.

[phipsgabler]

The Turing side of the PR is now out as well.

I wonder how to test all this in a reasonable way. The Turing integration tests over here have certainly not caught a lot of the namespace errors I made initially -- those I only discovered by testing Turing itself with this branch of DynamicPPL.

[devmotion]

There's some discussion about the test setup in #18. I guess ideally DynamicPPL should be tested without depending on Turing.

[phipsgabler]

I just saw that this relates to #19 as well.

[trappmartin]

@mohamed82008 can this be merged?

[devmotion]

I guess it would be nice if the tests pass. The Libtask error should be fixed by now, I guess, but I assume one has to update the Turing folder in the test directory with the changes in TuringLang/Turing.jl#1151 (which should maybe be updated with master since there might have been some changes to Turing master in the mean time).

[mohamed82008]

I will review this tonight.

[phipsgabler]

@devmotion Travis is still failing solely due to the Libtask error. On my system, all DynamicPPL tests pass with Libtask fixed to the older version. From the other side, also Turing's test pass for TuringLang/Turing.jl#1151 being used together with this branch.

I would have liked to update the tests already, of course, but right now I consider that all a bit too fragile for me to touch it.

[devmotion]

Hmm it seems that the tests still used Libtask 0.3.2, I thought the problem should be fixed with Libtask 0.3.3?

[phipsgabler]

Right -- appearently those test results are still from 5 days ago. Is there a way to trigger Travis manually? Found it, I wasn't signed in.

[phipsgabler]

Now it's just Travis killing the job because it takes so long...

[devmotion]

No, the problem is that the logs are too long. The problem is that the Turing test folder in DynamicPPL has not been updated with the latest changes in Turing, in particular TuringLang/Turing.jl#1156.

[mohamed82008]

Why not update it?

[mohamed82008]

LGTM. But let's make sure the tests pass.

[phipsgabler]

Alright, the Turing test folder is now updated to the state of the https://github.com/phipsgabler/Turing.jl/tree/context_independence, which is up to date with the current Turing master. All tests pass on my machine, let's see what Travis says :)

[codecov]

Codecov Report

Merging #43 into master will decrease coverage by 7.33%.
The diff coverage is 59.72%.

@@            Coverage Diff            @@
##           master     #43      +/-   ##
=========================================
- Coverage   88.24%   80.9%   -7.34%     
=========================================
  Files           9      12       +3     
  Lines         723     880     +157     
=========================================
+ Hits          638     712      +74     
- Misses         85     168      +83

Impacted Files	Coverage Δ
src/DynamicPPL.jl	100% <ø> (ø)	⬆️
src/compiler.jl	94.28% <ø> (+0.39%)	⬆️
src/varinfo.jl	87.8% <100%> (-0.04%)	⬇️
src/utils.jl	32.35% <25%> (-4.02%)	⬇️
src/distribution_wrappers.jl	25% <25%> (ø)
src/context_implementations.jl	50.75% <50.75%> (ø)
src/varname.jl	92.85% <92.85%> (ø)
... and 1 more

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update c59f519...ee9b693. Read the comment docs.

[mohamed82008]

Thanks @phipsgabler , great work!