Update to new DPPL version

[torfjelde]

This PR makes Turing compatible with [email protected].

EDIT: This should go after #1633 .

[yebai]

I'm a bit surprised that we need to overload tilde_* functions for ESS and MAP. It seems these algorithms only require a subset of three operations on the model, i.e. logprior, loglikelihood and logdensity/logjoint. Since these operations are already defined in DynamicPPL using contexts, is there a way to remove such overloads, or only overload sampler-facing functions like assume/observe?

[torfjelde]

Likely yeah, but IMO this should be done in a separate PR? This PR should be "minimal change necessary to make Turing.jl compatible with new DPPL".

[torfjelde]

Also, it seems like SMC samplers are now broken? o.O Do any of you @yebai @devmotion have any idea as to why this is? I'm not at all familiar with APS so it's a bit difficult to debug.

EDIT: Just discovered there's a TracedModel. Think I'm on to something.

EDIT 2: Figured it out. observe(::Samler, ...) is never hit because DynamicPPL.tilde_observe(::DefaultContext, sampler, ...) = observe(...), i.e. requires a change in DPPL.

[yebai]

Likely yeah, but IMO this should be done in a separate PR? This PR should be "minimal change necessary to make Turing.jl compatible with new DPPL".

Sounds good.

EDIT 2: Figured it out. observe(::Samler, ...) is never hit because DynamicPPL.tilde_observe(::DefaultContext, sampler, ...) = observe(...), i.e. requires a change in DPPL.

Great - I'll take a look at the DPPL PR.

[codecov]

Codecov Report

Merging #1636 (d90ed39) into master (9f52d75) will increase coverage by 0.29%.
The diff coverage is 66.66%.

❗ Current head d90ed39 differs from pull request most recent head 46a00a8. Consider uploading reports for the commit 46a00a8 to get more accurate results

@@            Coverage Diff             @@
##           master    #1636      +/-   ##
==========================================
+ Coverage   79.32%   79.61%   +0.29%     
==========================================
  Files          23       23              
  Lines        1422     1413       -9     
==========================================
- Hits         1128     1125       -3     
+ Misses        294      288       -6     

Impacted Files	Coverage Δ
src/modes/ModeEstimation.jl	84.29% <57.14%> (+3.52%)	⬆️
src/inference/ess.jl	98.00% <85.71%> (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 9f52d75...46a00a8. Read the comment docs.

[coveralls]

Pull Request Test Coverage Report for Build 925470820

Warning: This coverage report may be inaccurate.

This pull request's base commit is no longer the HEAD commit of its target branch. This means it includes changes from outside the original pull request, including, potentially, unrelated coverage changes.

• For more information on this, see Tracking coverage changes with pull request builds.
• To avoid this issue with future PRs, see these Recommended CI Configurations.
• For a quick fix, rebase this PR at GitHub. Your next report should be accurate.

Details

• 14 of 21 (66.67%) changed or added relevant lines in 2 files are covered.
• No unchanged relevant lines lost coverage.
• Overall coverage increased (+0.3%) to 79.618%

---

Changes Missing Coverage	Covered Lines	Changed/Added Lines	%
src/inference/ess.jl	6	7	85.71%
src/modes/ModeEstimation.jl	8	14	57.14%

Totals
Change from base Build 925007883:	0.3%
Covered Lines:	1125
Relevant Lines:	1413

---

💛 - Coveralls

[torfjelde]

Ready with the 👍 once the last test passes?

[devmotion]

Looks good, just bump the Turing version number? Do we want it to be a breaking release as well?

[torfjelde]

Good quesh. Technically nothing that we export is breaking, buuuut we're exporting DynamicPPL so I don't know how that works.

[torfjelde]

Maybe make it breaking just on the off-chance someone is actually overloading Turing.tilde, etc.?

EDIT: I misremebered, Turing.tilde isn't available, but Turing.assume is. Buuuut Turing.assume isn't changed, or rather than changes we've made doesn't break the current assume. Soooo non-breaking?

[devmotion]

As long as we do not export the methods that were broken in DynamicPPL, technically it is non-breaking. E.g., if we would reexport DynamicPPL or any of these methods it would be clear. It is a bit unclear how to deal with an exported module such as DynamicPPL... But generally if methods are only available internally as Turing.tilde etc. it does not require a breaking release.

[torfjelde]

The boundary is blurred further by the fact that most people don't about DPPL but are using stuff from there directly through Turing.whatever.

BUT I was mistaken; we don't have Turing.tilde, etc. available, so we should be good 👍