JuliaBUGS support

[miguelbiron]

@serenlee I've created this new PR with a more structured approach for giving support for JuliaBUGS models. It is based on Julia extension mechanisms, which allows us to keep precompilation times manageable for most users.
Now you can still keep working on your own PRs. But once you are comfortable with the code, we should migrate it to this PR.

Also, beyond the extension boilerplate code, I added a function that can automatically produce a prior model to use as default reference. It works by manipulating the underlying graph of the target, removing nodes related to observations. I think it is 80% there but needs testing.

[codecov]

Codecov Report

Attention: Patch coverage is 89.09091% with 6 lines in your changes missing coverage. Please review.

Project coverage is 87.30%. Comparing base (15ba905) to head (4d73dc3).

Files with missing lines	Patch %	Lines
ext/PigeonsJuliaBUGSExt/utils.jl	91.30%	2 Missing ⚠️
ext/PigeonsJuliaBUGSExt/interface.jl	94.11%	1 Missing ⚠️
ext/PigeonsJuliaBUGSExt/invariance_test.jl	90.90%	1 Missing ⚠️
src/Pigeons.jl	0.00%	1 Missing ⚠️
src/explorers/invariance_test.jl	0.00%	1 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #303      +/-   ##
==========================================
+ Coverage   87.23%   87.30%   +0.06%     
==========================================
  Files         103      107       +4     
  Lines        2570     2623      +53     
==========================================
+ Hits         2242     2290      +48     
- Misses        328      333       +5     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[alexandrebouchard]

Awesome!! Glad to hear the prior can be automatically extracted after all...

[serenlee]

@miguelbiron Thanks!

I have also been interacting with the JuliaBUGS team. I will work on the JuliaBUGS target and push it to this PR branch.

[sunxd3]

Hi guys, thanks for working on this. Please just let me know if anything is needed, happy to make additions on the JuliaBUGS end!

[miguelbiron]

Thank you @sunxd3! Actually I'd like your opinion on our approach to evaluating the prior. Context: this is an important requirement for PT methods. From what we understand of your code, there is no parallel in JuliaBUGS to the PriorContext construct in DynamicPPL. Therefore, there is no way of using evaluate!! to compute just the prior part (i.e., with no likelihood terms).

My solution was to automatically extract a prior JuliaBUGS model from the joint model. This is achieved by pruning non-parameter nodes from the full graph. Then we use the smaller graph to build the prior model, with which we can easily get prior logpdfs.

Question: do you think this is a reasonable approach? And do you think JuliaBUGS will in the future support prior logpdf computation?

Edit: Actually it would be even better to have access to a TemperedContext that would let you evaluated the logpdf of the tempered posterior for any beta>=0. I.e., logpdf_prior(x) + beta*loglikelihood(x).

[sunxd3]

TuringLang/JuliaBUGS.jl#247

I am not taking the contextual evaluation design for JuliaBUGS now because I felt the complexity is not worth it yet. So I just adde a function in the above PR. (at the cost of duplicated code)

The interface will very likely to change, but I'll make sure to update here once the potential happen.

[miguelbiron]

Oh wow thank you so much! I'll test the new interface as soon as possible.

[miguelbiron]

Ok, switched to the new interface and things seem to be proceeding smoothly! Thanks again @sunxd3 .

On a related note @sunxd3, I was thinking you could avoid having to maintain duplicated code by redefining evaluate!! to just call the tempered version with temp = 1.

[miguelbiron]

Need to make initialization more stable (currently depends on global RNG); might explain this sole fail

https://github.com/Julia-Tempering/Pigeons.jl/actions/runs/12261789152/job/34209633693?pr=303#step:8:1128