Less surprising model return values

[devmotion]

As can be seen in #50 (comment), the output of model() can be quite surprising if no return statements are specified in the model definition. One possible solution to the problem that is implemented in this PR is to always return nothing as a fallback. An alternative would be to return the VarInfo object as a fallback but that might be not completely user-friendly since it might not be clear how to handle or analyse it.

Additionally, currently logp is reset both in the default evaluation function and in runmodel!, so basically model(vi) and runmodel!(vi) are doing the same thing currently apart from adjusting eval_num. I removed the resetlogp! statement in the evaluation function to allow for greater flexibility and removed the runmodel! function. Instead now one can call model() and model(vi) - without vi a new VarInfo object is created but otherwise they are doing exactly the same thing.

[codecov]

Codecov Report

Merging #63 into master will decrease coverage by 0.04%.
The diff coverage is 100.00%.

@@            Coverage Diff             @@
##           master      #63      +/-   ##
==========================================
- Coverage   79.47%   79.42%   -0.05%     
==========================================
  Files          13       13              
  Lines         843      841       -2     
==========================================
- Hits          670      668       -2     
  Misses        173      173              

Impacted Files	Coverage Δ
src/DynamicPPL.jl	100.00% <ø> (ø)
src/compiler.jl	89.26% <100.00%> (-0.08%)	⬇️
src/model.jl	77.77% <100.00%> (-1.17%)	⬇️
src/varinfo.jl	88.27% <100.00%> (ø)

---

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 1c071df...18a0353. Read the comment docs.

[phipsgabler]

I, as a quite innocent user, would expect a model to behave like a function, and to return the value of the last line, with a tilde having the same semantics as an assignment.

nothing or VarInfo as automatic results unless there's a return is really counterintuitive, IMHO (nothing less so, but still).

So the easiest thing would be to define the semantics of tildes to be equivalent to = and .=, and just add the observed or sampled value to the generated code, after the accumulation of log-probabilities.

[devmotion]

Ah, yes I guess that's even better! Yeah the main motivation for nothing was to avoid the log probabilities but I wasn't completely satisfied with the solution in this PR 😄

[phipsgabler]

Beautyful.

[yebai]

I resolved some merge conflicts. It seems some tests are still broken.

[devmotion]

I removed two runmodel! imports that were reintroduced when merging master, let's see if the tests pass now.

[devmotion]

OK, now only code coverage fails.

[mohamed82008]

Nice work!

[yebai]

thanks @devmotion!