Fix for HMCs dot_assume

[torfjelde]

Currently dot_assume for HMC is broken, which this PR fixes.

This case is covered by tests, but (AFAIK) only in DynamicPPL's integration tests, not in Turing's tests.

Should I copy-paste these or is there a way we can just run the integration tests from DPPL with this Turing version?

[devmotion]

Should I copy-paste these or is there a way we can just run the integration tests from DPPL with this Turing version?

I think the test setup is simpler and more reliable if you just copy the tests (I guess we could use the test models but probably that's what's done already in DynamicPPL?).

[torfjelde]

Eeeh weird. I can't reproduce the errors locally 😳

[torfjelde]

Aaaaight so I finally figured out what's going wrong:

julia> using Tracker, Distributions

julia> m = first(param(zeros(1)))
0.0 (tracked)

julia> s = first(param(ones(1)))
1.0 (tracked)

julia> typeof(Normal.(m, s))
Tracker.Tracked{Normal{Float64}}

And DistributionsAD.jl doesn't address this issue:

julia> using DistributionsAD

julia> typeof(Normal.(m, s))
Tracker.Tracked{Normal{Float64}}

It comes down to Tracker not handling broadcast properly. I'm just going to mark this test as broken for now (don't think anyone uses Tracker with Turing anyways) and raise an issue over at Tracker.jl.

[devmotion]

Maybe related to or even the same as FluxML/Tracker.jl#65?

[torfjelde]

Possibly! I've made a new issue as it's not completely obvious to me whether it's the same or not atm: FluxML/Tracker.jl#119

[torfjelde]

Now running into #1756

[torfjelde]

Renamed Core to TuringCore. Should we maybe bump the minor version? Seems like this could break people's code (e.g. I know lots of my snippets will break).

[torfjelde]

This should be good to go once #1762 has gone through 👍

[codecov]

Codecov Report

Merging #1758 (1d8eb1e) into master (9087412) will increase coverage by 0.63%.
The diff coverage is 100.00%.

❗ Current head 1d8eb1e differs from pull request most recent head b1a8195. Consider uploading reports for the commit b1a8195 to get more accurate results

@@            Coverage Diff             @@
##           master    #1758      +/-   ##
==========================================
+ Coverage   81.23%   81.87%   +0.63%     
==========================================
  Files          24       24              
  Lines        1471     1473       +2     
==========================================
+ Hits         1195     1206      +11     
+ Misses        276      267       -9     

Impacted Files	Coverage Δ
src/inference/hmc.jl	79.39% <100.00%> (+3.03%)	⬆️
src/inference/AdvancedSMC.jl	97.47% <0.00%> (+0.04%)	⬆️
src/essential/ad.jl	86.07% <0.00%> (+5.06%)	⬆️

---

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 9087412...b1a8195. Read the comment docs.

[coveralls]

Pull Request Test Coverage Report for Build 1698293351

• 1 of 1 (100.0%) changed or added relevant line in 1 file are covered.
• No unchanged relevant lines lost coverage.
• Overall coverage increased (+0.6%) to 81.874%

---

Totals
Change from base Build 1696660269:	0.6%
Covered Lines:	1206
Relevant Lines:	1473

---

💛 - Coveralls