Implement GibbsConditional #2647

AoifeHughes · 2025-08-07T08:15:00Z

Copilot

Pull Request Overview

This PR implements a GibbsConditional sampler component for Turing.jl that allows users to provide analytical conditional distributions for variables in Gibbs sampling. The implementation enables mixing user-defined conditional distributions with other MCMC samplers within the Gibbs framework.

Key changes:

Added GibbsConditional struct and supporting functions for analytical conditional sampling
Comprehensive test coverage for the new functionality
Added example test file demonstrating usage

Reviewed Changes

Copilot reviewed 5 out of 6 changed files in this pull request and generated 5 comments.

Show a summary per file

File	Description
`src/mcmc/gibbs_conditional.jl`	Core implementation of GibbsConditional sampler with step functions and variable handling
`src/mcmc/Inference.jl`	Added GibbsConditional export and module inclusion
`test/mcmc/gibbs.jl`	Added comprehensive test suite for GibbsConditional functionality
`test_gibbs_conditional.jl`	Example/demo file showing GibbsConditional usage
`HISTORY.md`	Version history update (unrelated to main feature)

src/mcmc/gibbs_conditional.jl

HISTORY.md

test_gibbs_conditional.jl

github-actions · 2025-08-07T13:36:34Z

Turing.jl documentation for PR #2647 is available at:
https://TuringLang.github.io/Turing.jl/previews/PR2647/

codecov · 2025-08-07T13:40:14Z

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 86.80%. Comparing base (19bf7d6) to head (744d254).

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #2647      +/-   ##
==========================================
+ Coverage   86.47%   86.80%   +0.32%     
==========================================
  Files          21       22       +1     
  Lines        1420     1455      +35     
==========================================
+ Hits         1228     1263      +35     
  Misses        192      192

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

🚀 New features to boost your workflow:

❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

coveralls · 2025-08-07T14:13:49Z

Pull Request Test Coverage Report for Build 18095923386

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

0 of 69 (0.0%) changed or added relevant lines in 1 file are covered.
52 unchanged lines in 4 files lost coverage.
Overall coverage decreased (-7.0%) to 76.573%

Changes Missing Coverage	Covered Lines	Changed/Added Lines	%
src/mcmc/gibbs_conditional.jl	0	69	0.0%

Files with Coverage Reduction	New Missed Lines	%
src/mcmc/particle_mcmc.jl	2	88.4%
src/mcmc/hmc.jl	5	93.65%
src/mcmc/repeat_sampler.jl	8	56.52%
src/mcmc/gibbs.jl	37	69.59%

Totals
Change from base Build 16937761215:	-7.0%
Covered Lines:	1144
Relevant Lines:	1494

💛 - Coveralls

mhauru

As discussed on Slack, rather than doing a full review, I'm just going to give some high level comments and pointers for where to find more details on some of the relevant context.

src/mcmc/gibbs_conditional.jl

test_gibbs_conditional.jl

src/mcmc/gibbs_conditional.jl

…gibbs-sampler

mhauru

I left a couple of new comments.

Also, a few of the comments from my first review remain to be addressed and all the newly added tests are still failing.

src/mcmc/gibbs_conditional.jl

…ution and improve context variable retrieval

…actMCMC.step functions

Red-Portal · 2025-11-16T16:40:11Z

Hi, I think this PR needs to be worked on as quickly as possible given that lots of people (#2547 and in slack) have been requesting this feature to be added back. I could take this over if nobody has bandwidth to work on this. However, I am a bit less familiar with the relevant parts in Turing. @mhauru @penelopeysm could you take a look if you have any suggestions?

mhauru · 2025-11-17T10:05:18Z

Aoife is no longer working on Turing, and I said I would take over this one, but just haven't found time. Let me try to see what the state of this is and report back.

Copilot · 2025-11-17T17:44:09Z

@mhauru I've opened a new pull request, #2716, to work on those changes. Once the pull request is ready, I'll request review from you.

mhauru · 2025-11-17T17:45:12Z

I fixed and improved the implementation and added a lot of tests. I think this may be done now, but I'll reread it myself tomorrow before requesting reviews.

mhauru · 2025-11-18T11:48:42Z

Julia package registry seems to be having trouble, will rerun CI in a moment. This is ready for review though. @Red-Portal, would you be happy to take an overall look? @penelopeysm could I ask you to take a look at the implementation and HISTORY.md entry at the least? Up to you if you want to also read the tests.

src/mcmc/gibbs_conditional.jl

penelopeysm · 2025-11-18T12:30:33Z

src/mcmc/gibbs_conditional.jl

+    m = c[@varname(m)]
+    x = c[@varname(x)]


I feel that the c[@varname(x)] here is deceptively simple... it will work under the following scenarios

x is passed as a single vector to the model arguments

x is not in the arguments, but is conditioned on as a vector i.e. model() | (; x = vec)

it will fail in the case where

x is not in the arguments, but is conditioned on as individual elements i.e. model() | Dict(@varname(x[1]) => 1.0, @varname(x[2]) => 2.0)

Notice that because the model has x[i] ~ Normal, all three cases will work correctly with plain model evaluation. But one of the cases will fail with GibbsConditional.

I don't really know how to fix this, and I don't know whether it should even be fixed, but it makes me quite uncomfortable.

I've fixed all the other points you raised. This will have to wait untill tomorrow.

As a simple enough solution, could we maybe take the model arguments out of the dictionary, and instead give c a field called c.model_args = deepcopy(model.args)? Then people can use c.model_args.x. Deepcopy would be needed to avoid accidental aliasing. (Or we could not copy, and leave it to the user to copy if they need it.)

The outcome of this is:

x is passed as a single vector to the model arguments

OK, people can use c.model_args.x, clear.

x is not in the arguments, but is conditioned on as a vector i.e. model() | (; x = vec)

OK, people can use c[@varname(x)] because that's what they conditioned on.

x is not in the arguments, but is conditioned on as individual elements i.e. model() | Dict(@varname(x[1]) => 1.0, @varname(x[2]) => 2.0)

This will still fail but at least there's a good explanation for it, it's because they conditioned on x[1] and x[2] rather than x.

Looking at my previous comment, actually the original implementation is also explainable along the same lines. I guess the tldr is basically, if you supply x as a single thing, that's fine. So actually I'd be OK with leaving it as is.

The root cause of this is essentially the same as why you can't do

@model function f() x ~ MvNormal() end m = condition(f(), Dict(@varname(x[1]) => 1.0))

it's just that in that case you request the value of the conditioned x in the model, whereas here you request it in the conditionals function. So in some sense this is a bigger problem that needs to be fixed lower down (https://github.com/TuringLang/DynamicPPL.jl/issues/11480). So I'd be okay with leaving this be too.

The one thing that I was wondering about is that ConditionContext internally uses getvalue. If the user did the same in the conditionals function then at least they could get x1 = getvalue(c, @varname(x[1])) to work even if they conditioned on x as a whole. Thus I'm now thinking that maybe the best thing to do here would be to set a good example in the docstring and use getvalue rather than getindex, and leave it at that.

As a silver lining, if this goes wrong, at least the user gets a clear "key not found" error, and the error comes from code they wrote, in quite a traceable way.

Yeah, that sounds good to me! Agree that getvalue is better, albeit a tiny bit more annoying since it has to be imported from AbstractPPL

src/mcmc/gibbs_conditional.jl

penelopeysm · 2025-11-18T12:38:17Z

src/mcmc/gibbs_conditional.jl

+    # TODO(mhauru) Can we avoid invlinking all the time? Note that this causes a model
+    # evaluation, which may be expensive.


I think for typed VarInfo it shouldn't need to evaluate the model. Obviously it still has a cost, just not as much as model evaluation.

julia> @model function f() @info "hi" x ~ Normal() end f (generic function with 2 methods) julia> model = f(); v = VarInfo(model); [ Info: hi julia> v2 = DynamicPPL.link!!(v, model); v3 = DynamicPPL.invlink!!(v, model);

Adjusted the comment to reflect this.

penelopeysm · 2025-11-18T12:46:34Z

src/mcmc/gibbs_conditional.jl

+            (DynamicPPL.VarName{sym}() => val for (sym, val) in pairs(cond_nt))...,
+            (DynamicPPL.VarName{sym}() => val for (sym, val) in pairs(fixed_nt))...,
+            (DynamicPPL.VarName{sym}() => val for (sym, val) in pairs(model.args))...,


cond_nt and fixed_nt might not be NamedTuples, they might be dicts, in which case this will fail in a very weird manner.

julia> VarName{@varname(x)}() x julia> VarName{@varname(x)}() == @varname(x) # not the same thing false

I think you have to convert them to dicts first. DynamicPPL.to_varname_dict will do this (albeit inefficiently TuringLang/DynamicPPL.jl#1134).

Nice catch. I fixed this and added a test that would have caught this.

Making so many Dicts is sad, but for now I just want to get this working, and worry about performance later.

test/mcmc/gibbs_conditional.jl

penelopeysm · 2025-11-18T12:51:07Z

test/mcmc/gibbs_conditional.jl

+            prior_var = 100.0  # 10^2
+            post_var = 1 / (1 / prior_var + n / var)
+            post_mean = post_var * (0 / prior_var + sum(x) / var)


Suggested change

prior_var = 100.0 # 10^2

post_var = 1 / (1 / prior_var + n / var)

post_mean = post_var * (0 / prior_var + sum(x) / var)

prior_var = 100.0 # 10^2

post_var = 1 / (1 / prior_var + n / var)

post_mean = post_var * (sum(x) / var)

again - are there formulas for this somewhere? Maybe a statistician looks at this and goes 'ah yes of course', but I feel quite uncomfortable.

Will add the link. I didn't write this part, but I think it's deliberately in this format to match the formula I found e.g. on wikipedia. The 0 I think is the mean of the prior.

Ah okay in that case it can stay!

test/mcmc/gibbs_conditional.jl

src/mcmc/gibbs_conditional.jl

Co-authored-by: Penelope Yong <[email protected]>

penelopeysm

Thanks, great work!

AoifeHughes added 2 commits August 7, 2025 09:08

Add GibbsConditional sampler and corresponding tests

c0158ea

clarified comment

a972b5a

AoifeHughes requested a review from Copilot August 7, 2025 08:15

AoifeHughes self-assigned this Aug 7, 2025

Copilot AI reviewed Aug 7, 2025

View reviewed changes

AoifeHughes changed the base branch from breaking to main August 7, 2025 08:16

Merge branch 'main' into gibbs-sampler

bdb7f73

mhauru requested changes Aug 7, 2025

View reviewed changes

AoifeHughes added 4 commits August 8, 2025 11:11

add MHs suggestions

c3cc773

formatter

714c1e8

Merge branch 'gibbs-sampler' of github.com:TuringLang/Turing.jl into …

97c571d

…gibbs-sampler

fixed exporting thing

94b723d

mhauru self-requested a review August 8, 2025 13:49

Merge branch 'main' into gibbs-sampler

891ac14

mhauru requested changes Aug 18, 2025

View reviewed changes

src/mcmc/gibbs_conditional.jl Outdated Show resolved Hide resolved

src/mcmc/gibbs_conditional.jl Outdated Show resolved Hide resolved

src/mcmc/gibbs_conditional.jl Outdated Show resolved Hide resolved

Aoife added 4 commits September 23, 2025 13:09

Refactor Gibbs sampler to use inverse of parameters for Gamma distrib…

2058ae5

…ution and improve context variable retrieval

removed file added by mistake

b0812a3

Add safety checks and error handling in find_global_varinfo and Abstr…

d910312

…actMCMC.step functions

imports?

4b1dc2f

Red-Portal requested review from mhauru and penelopeysm November 16, 2025 16:36

mhauru added 4 commits November 17, 2025 10:54

Merge remote-tracking branch 'origin/main' into gibbs-sampler

1e84309

Fixes and improvements for GibbsConditional

a33d8a9

Move GibbsConditional tests to their own file

f41fc6e

More GibbsConditional tests

bd5ff0b

mhauru added 2 commits November 17, 2025 17:41

Bump patch version to 0.41.2, add HISTORY.md entry

34acad7

Remove spurious change

4786a59

Copilot AI mentioned this pull request Nov 17, 2025

[WIP] WIP Address feedback from review on Gibbs sampler implementation #2716

Closed

mhauru added 2 commits November 18, 2025 11:12

Code style and documentation

8951d98

Add one test_throws, tweak test thresholds and dimensions

45ab5f8

mhauru closed this Nov 18, 2025

mhauru reopened this Nov 18, 2025

mhauru requested a review from Red-Portal November 18, 2025 11:46

mhauru changed the title ~~Gibbs sampler~~ Implement GibbsConditional Nov 18, 2025

penelopeysm reviewed Nov 18, 2025

View reviewed changes

src/mcmc/gibbs_conditional.jl Outdated Show resolved Hide resolved

mhauru and others added 6 commits November 18, 2025 16:54

Apply suggestions from code review

805bc60

Co-authored-by: Penelope Yong <[email protected]>

Add links for where to get analytical posteriors

d0c3cf4

Update TODO note

98f4213

Fix a GibbsConditional bug, add a test

c74b0a0

Set seeds better

4a7d08c

Use getvalue in docstring

744d254

mhauru requested a review from penelopeysm November 19, 2025 16:47

penelopeysm approved these changes Nov 19, 2025

View reviewed changes

		# TODO(mhauru) Can we avoid invlinking all the time? Note that this causes a model
		# evaluation, which may be expensive.

Implement GibbsConditional #2647

Are you sure you want to change the base?

Implement GibbsConditional #2647

Uh oh!

Conversation

AoifeHughes commented Aug 7, 2025

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull Request Overview

Reviewed Changes

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

github-actions bot commented Aug 7, 2025

Uh oh!

codecov bot commented Aug 7, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Codecov Report

Uh oh!

coveralls commented Aug 7, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Pull Request Test Coverage Report for Build 18095923386

Warning: This coverage report may be inaccurate.

Details

💛 - Coveralls

Uh oh!

mhauru left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

mhauru left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Red-Portal commented Nov 16, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

mhauru commented Nov 17, 2025

Uh oh!

Copilot AI commented Nov 17, 2025

Uh oh!

mhauru commented Nov 17, 2025

Uh oh!

mhauru commented Nov 18, 2025

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

penelopeysm Nov 19, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

penelopeysm Nov 19, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

codecov bot commented Aug 7, 2025 •

edited

Loading

coveralls commented Aug 7, 2025 •

edited

Loading

Red-Portal commented Nov 16, 2025 •

edited

Loading

penelopeysm Nov 19, 2025 •

edited

Loading

penelopeysm Nov 19, 2025 •

edited

Loading