remove LogDensityProblemsAD #2490

penelopeysm · 2025-02-20T13:57:43Z

this PR removes LogDensityProblemsAD in favour of using the new interface in TuringLang/DynamicPPL.jl#806

codecov · 2025-02-20T13:59:45Z

Codecov Report

Attention: Patch coverage is 4.91803% with 58 lines in your changes missing coverage. Please review.

Project coverage is 29.77%. Comparing base (52b2105) to head (7c32e3e).
Report is 1 commits behind head on mhauru/dppl-0.35.

Files with missing lines	Patch %	Lines
src/optimisation/Optimisation.jl	0.00%	28 Missing ⚠️
ext/TuringOptimExt.jl	0.00%	15 Missing ⚠️
src/mcmc/hmc.jl	0.00%	9 Missing ⚠️
src/mcmc/abstractmcmc.jl	60.00%	2 Missing ⚠️
src/mcmc/sghmc.jl	0.00%	2 Missing ⚠️
ext/TuringDynamicHMCExt.jl	0.00%	1 Missing ⚠️
src/mcmc/gibbs.jl	0.00%	1 Missing ⚠️

Additional details and impacted files

@@                 Coverage Diff                  @@
##           mhauru/dppl-0.35    #2490      +/-   ##
====================================================
+ Coverage             26.93%   29.77%   +2.83%     
====================================================
  Files                    21       21              
  Lines                  1407     1387      -20     
====================================================
+ Hits                    379      413      +34     
+ Misses                 1028      974      -54

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

penelopeysm

I removed the entire test/ad.jl file, my reasoning is explained below

penelopeysm · 2025-02-20T13:58:46Z

test/essential/ad.jl

-        @test l ≈ logp
-        @test sort(∇E) ≈ grad_FWAD atol = 1e-9
-    end


this checks that the gradient obtained with LogDensityProblemsAD is equal to the gradient obtained directly with ForwardDiff - this isn't Turing's job to test hence removed

penelopeysm · 2025-02-20T14:00:14Z

test/essential/ad.jl

-        @test zygoteℓ.ℓ === ℓ
-        ∇E2 = LogDensityProblems.logdensity_and_gradient(zygoteℓ, x)[2]
-        @test sort(∇E2) ≈ grad_FWAD atol = 1e-9


This is checking that Zygote gives the same gradient as ForwardDiff does.

Firstly Zygote isn't formally supported so I don't think we need to test its correctness on this one particular model, and secondly if we did want to support it we should add this to the DPPL test suite

penelopeysm · 2025-02-20T14:01:10Z

test/essential/ad.jl

-        @test sort(∇E2) ≈ grad_FWAD atol = 1e-9
-    end
-
-    @testset "general AD tests" begin


this whole testset is just testing AD correctness on various models, and should all be moved into DynamicPPL, if it's not already there

It's there already, in test/compat/ad.jl.

Okay what that's so weird. Doubly happy to delete it then

penelopeysm · 2025-02-20T14:01:31Z

test/essential/ad.jl

-
-        test_model_ad(wishart_ad(), logp3, [:v])
-    end
-    @testset "Simplex Zygote and ReverseDiff (with and without caching) AD" begin


same with this one, it's not even testing correctness, it's just testing that AD can run

penelopeysm · 2025-02-20T14:02:03Z

test/essential/ad.jl

-        sample(invwishart(), HMC(0.01, 1; adtype=AutoReverseDiff(; compile=false)), 1000)
-        sample(invwishart(), HMC(0.01, 1; adtype=AutoZygote()), 1000)
-    end
-    @testset "Hessian test" begin


same with this one, it's testing correctness of Hessians against an analytic value, it's not really our job imo

penelopeysm · 2025-02-20T14:03:11Z

test/essential/ad.jl

-        @test H_f == H_r
-    end
-
-    @testset "memoization: issue #1393" begin


i don't think this testset is relevant anymore because the memoisation code has been long removed #1414

penelopeysm · 2025-02-20T14:03:29Z

test/essential/ad.jl

-        end
-    end
-
-    @testset "ReverseDiff compiled without linking" begin


this testset is fully contained in DynamicPPL already

ext/TuringOptimExt.jl

src/mcmc/Inference.jl

src/optimisation/Optimisation.jl

test/ext/OptimInterface.jl

test/optimisation/Optimisation.jl

mhauru

LGTM. One question mostly out of curiosity.

mhauru · 2025-02-25T14:16:43Z

test/test_utils/ad_utils.jl

@@ -202,7 +202,7 @@ end
 All the ADTypes on which we want to run the tests.
 """
 adbackends = [
-    Turing.AutoForwardDiff(; chunksize=0),
+    Turing.AutoForwardDiff(),


What's the significance of not setting chunksize?

if chunksize isn't set it will default to nothing, and both 0 and nothing will get 'optimised' into the same thing by this function in DynamicPPLForwardDiffExt https://github.com/TuringLang/DynamicPPL.jl/blob/90c7b26c852b7d0ae87dee0a5a0010b097a0c1d3/ext/DynamicPPLForwardDiffExt.jl#L10-L40

but also technically chunksize=0 is 'wrong' and not supported in DifferentiationInterface, the fact that it works in Turing is solely because of that extension, which I'd honestly like to get rid of but have mainly kept around for compatibility - see TuringLang/DynamicPPL.jl#806 (comment)

if the extension isn't present, then chunksize=0 actually errors, and I believe it was the source of this error #2369 (it's because ForwardDiff tries to count the number of chunks by dividing by the chunksize). I suspect that that error appeared because Optimisation.jl started using DifferentiationInterface under the hood

* Progress towards compat with DPPL v0.35 * More fixing of DPPL v0.35 stuff * Fix LogDensityFunction argument order * More minor bugfixes * [TEMP] Commit Manifest pointing to DynamicPPL#release-0.35 * remove LogDensityProblemsAD (#2490) * Remove LogDensityProblemsAD, part 1 * update Optimisation code to not use LogDensityProblemsAD * Fix field name change * Don't put chunksize=0 * Remove LogDensityProblemsAD dep * Improve OptimLogDensity docstring * Remove unneeded model argument to _optimize * Fix more tests * Remove essential/ad from the list of CI groups * Fix HMC function * more test fixes (#2491) * Remove LogDensityProblemsAD, part 1 * update Optimisation code to not use LogDensityProblemsAD * Fix field name change * Don't put chunksize=0 * Remove LogDensityProblemsAD dep * Improve OptimLogDensity docstring * Remove unneeded model argument to _optimize * Fix more tests * Remove essential/ad from the list of CI groups * Fix HMC function * More test fixes * Remove Manifest * More fixes for DynamicPPL 0.35 (#2494) * Remove test/dynamicppl/compiler.jl * Remove old regression tests * Remove vdemo2 * Fix last test * Add HISTORY.md entry about DPPL 0.35 * Allow ESS to sample variables with different symbols * Update a TODO note --------- Co-authored-by: Penelope Yong <[email protected]>

* Bump minor version to v0.37 * Remove selector/space stuff (#2458) * Remove selector stuff from ESS * Remove selector stuff from MH * Remove selector stuff from HMC * Remove selector stuff from Emcee * Remove selector stuff from IS * Add missing getspace methods * Remove selector stuff for particle methods * Fix an HMC selector bug * Code style * Fix Emcee selector bug * Fix typo in ESS tests * Fix some constructor overwrites * Remove unnecessary tests * Remove selector stuff from SGHMC * Remove drop_space and other non-longer-necessary deprecation measures * Bump minor version 0.37. Add a HISTORY.md entry * Apply suggestions from code review Co-authored-by: Penelope Yong <[email protected]> * Remove unnecessary type parameters Co-authored-by: Penelope Yong <[email protected]> * Simplify constructors in particle_mcmc.jl * Remove calls to setgid and updategid --------- Co-authored-by: Penelope Yong <[email protected]> * Bump Mooncake compat to 0.4.95 * Support for DynamicPPL v0.35 (#2488) * Progress towards compat with DPPL v0.35 * More fixing of DPPL v0.35 stuff * Fix LogDensityFunction argument order * More minor bugfixes * [TEMP] Commit Manifest pointing to DynamicPPL#release-0.35 * remove LogDensityProblemsAD (#2490) * Remove LogDensityProblemsAD, part 1 * update Optimisation code to not use LogDensityProblemsAD * Fix field name change * Don't put chunksize=0 * Remove LogDensityProblemsAD dep * Improve OptimLogDensity docstring * Remove unneeded model argument to _optimize * Fix more tests * Remove essential/ad from the list of CI groups * Fix HMC function * more test fixes (#2491) * Remove LogDensityProblemsAD, part 1 * update Optimisation code to not use LogDensityProblemsAD * Fix field name change * Don't put chunksize=0 * Remove LogDensityProblemsAD dep * Improve OptimLogDensity docstring * Remove unneeded model argument to _optimize * Fix more tests * Remove essential/ad from the list of CI groups * Fix HMC function * More test fixes * Remove Manifest * More fixes for DynamicPPL 0.35 (#2494) * Remove test/dynamicppl/compiler.jl * Remove old regression tests * Remove vdemo2 * Fix last test * Add HISTORY.md entry about DPPL 0.35 * Allow ESS to sample variables with different symbols * Update a TODO note --------- Co-authored-by: Penelope Yong <[email protected]> * Fix call to DynamicPPL.initialize_parameters!! * Remove `Zygote` (#2505) * Remove `Zygote`; fix #2504 * Update test/test_utils/ad_utils.jl Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> * Add HISTORY.md entry about removing support for Zygote --------- Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> Co-authored-by: Markus Hauru <[email protected]> Co-authored-by: Penelope Yong <[email protected]> Co-authored-by: Markus Hauru <[email protected]> * Fix a Gibbs test * Clean up exports (#2474) * Regroup exports by package * Export DynamicPPL.returned and DynamicPPL.prefix * Stop exporting @logprob_str and @prob_str * Remove DynamicPPL module export * Remove DynamicPPL.LogDensityFunction re-export * Remove BernoulliLogit, drop support for Distributions < 0.25.77 * Stop blanket re-exporting Libtask and Bijectors * Manually specify AbstractMCMC exports * Format Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> * Remove Bijectors.ordered export * Re-export LinearAlgebra.I * Replace Turing.Model -> DynamicPPL.Model * Format * Keep exporting LogDensityFunction * Add note in docs * Align Turing exports with docs API page * Fix things like `predict` on docs API page * Fix a Gibbs test * Format * Fix missing Bijectors import * Update docs/src/api.md Co-authored-by: Markus Hauru <[email protected]> * Update docs/src/api.md Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> * Clean up broken docs links, remove unneeded deps * Format * Add changelog entry for exports * Clean up exports in essential/Essential * Apply suggestions from code review Co-authored-by: Markus Hauru <[email protected]> --------- Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> Co-authored-by: Markus Hauru <[email protected]> Co-authored-by: Markus Hauru <[email protected]> --------- Co-authored-by: Penelope Yong <[email protected]> Co-authored-by: Hong Ge <[email protected]> Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>

Remove LogDensityProblemsAD, part 1

b21bc4b

penelopeysm commented Feb 20, 2025

View reviewed changes

update Optimisation code to not use LogDensityProblemsAD

fe46018

penelopeysm requested a review from mhauru February 20, 2025 15:12

mhauru reviewed Feb 20, 2025

View reviewed changes

penelopeysm added 6 commits February 20, 2025 17:56

Fix field name change

f83d695

Don't put chunksize=0

68ceb39

Remove LogDensityProblemsAD dep

71547c9

Improve OptimLogDensity docstring

62ab4b1

Remove unneeded model argument to _optimize

48dac18

Fix more tests

56997c1

penelopeysm force-pushed the py/no-ldp-ad branch from 70bb546 to 56997c1 Compare February 20, 2025 18:44

penelopeysm changed the title ~~remove LogDensityProblemsAD part 1~~ remove LogDensityProblemsAD Feb 20, 2025

penelopeysm added 2 commits February 20, 2025 18:51

Remove essential/ad from the list of CI groups

ef7328f

Fix HMC function

7c32e3e

penelopeysm marked this pull request as ready for review February 21, 2025 18:55

mhauru approved these changes Feb 25, 2025

View reviewed changes

penelopeysm merged commit ee2b148 into mhauru/dppl-0.35 Feb 25, 2025
18 of 52 checks passed

penelopeysm deleted the py/no-ldp-ad branch February 25, 2025 14:39

penelopeysm mentioned this pull request Mar 27, 2025

Clean up LogDensityFunctions interface code + setADtype #2473

Closed

remove LogDensityProblemsAD #2490

remove LogDensityProblemsAD #2490

Uh oh!

Conversation

penelopeysm commented Feb 20, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

codecov bot commented Feb 20, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Codecov Report

Uh oh!

penelopeysm left a comment

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!

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!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

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!

Choose a reason for hiding this comment

Uh oh!

penelopeysm Feb 25, 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!

Uh oh!

Uh oh!

penelopeysm commented Feb 20, 2025 •

edited

Loading

codecov bot commented Feb 20, 2025 •

edited

Loading

penelopeysm Feb 25, 2025 •

edited

Loading