Bump AbstractMCMC major version

[torfjelde]

Supersedes #1574 and #1575

[devmotion]

Looks good, just have to make sure that it's actually tested and needs a version bump 👍

[devmotion]

The same strange errors in the Gibbs MH test. I'm even more surprised that the errors occur on Julia 1.3 🤔

[devmotion]

The failing tests were added recently in #1557.

[devmotion]

@cpfiffer Maybe we should just change the test in

Turing.jl/test/inference/mh.jl

Line 137 in 2eab753

MH((:σ, vc_σ)),

? To me it seems, it is likely to fail randomly (even with a small step size) since we use a random walk proposal for the standard deviation that has to be positive (otherwise the model errors)?

[torfjelde]

But it should be linking, no?

[torfjelde]

At least that's the intention:

Turing.jl/src/inference/mh.jl

Line 404 in 2eab753

maybe_link!(vi, spl, spl.alg.proposals)

[torfjelde]

I think I see the issue. https://github.com/TuringLang/Turing.jl/blob/master/src/inference/mh.jl#L404 passes spl.alg.proposals, but in this case the proposals are actually NamedTuple{sym, RandomWalkProposal} and so they hit the maybe_link! which does nothing.

[coveralls]

Pull Request Test Coverage Report for Build 736956259

• 0 of 0 changed or added relevant lines in 0 files are covered.
• No unchanged relevant lines lost coverage.
• Overall coverage remained the same at 78.167%

---

Totals
Change from base Build 736943876:	0.0%
Covered Lines:	1117
Relevant Lines:	1429

---

💛 - Coveralls

[codecov]

Codecov Report

Merging #1579 (35ea02b) into master (9104637) will not change coverage.
The diff coverage is n/a.

@@           Coverage Diff           @@
##           master    #1579   +/-   ##
=======================================
  Coverage   78.16%   78.16%           
=======================================
  Files          23       23           
  Lines        1429     1429           
=======================================
  Hits         1117     1117           
  Misses        312      312           

---

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 9104637...35ea02b. Read the comment docs.

[cpfiffer]

Suggested change

	AbstractMCMC = "2.1, 3"
	AbstractMCMC = "3"

This should cut out AbstractMCMC 2 because the internal API changes.

[devmotion]

But this is only relevant for users that use custom callbacks? I.e., for Turing it should be fine to support both versions since we do not depend on these internals, do we?

[cpfiffer]

Hm, I suppose not. IGNORE ME

[torfjelde]

Yeah, leaving it as is:)

[cpfiffer]

Suggested change

	AbstractMCMC = "2.1, 3"
	AbstractMCMC = "3"

Same here.