New Gibbs sampler using condition (#2099)

* initial work on the new gibbs sampler

* added tests for the new Gibbs sampler

* added tests for new Gibbs

* new Gibbs is now sampling (correctly) sequentially

* let's not overload merge just yet

* export GibbsV2 + added more samplers to the tests

* added TODO comment

* removed lots of varinfo related merging functionality that is now
available in DynamicPPL

* shifting some code around

* removed redundant constructor for GibbsV2

* added GibbsContext which is similar to FixContext but also computes
the log-prob of the fixed variables

* adopted the rerun mechanism in Gibbs for GibbsV2, thus fixing the
issues with some of the tests for GibbsV2

* broken tests are no longer broken

* fix issues with dot_tilde_* impls for GibbsContext

* fix for dot_tilde_assume when using GibbsContext

* fixed re-running of models for Gibbs sampling properly this time

* added new gibbs to tests

* added some further comments on why we need `GibbsContext`

* went back to using `DynamicPPL.condition` rather than using custom
`GibbsContext` while we wait for
TuringLang/DynamicPPL.jl#563 to be merged

* add concrete comment about reverting changes for `gibbs_condition`

* Update test/mcmc/gibbs_new.jl

Co-authored-by: Hong Ge <[email protected]>

* fixed recursive definition of `condition` varinfos

* use `fix` instead of `condition`

* Revert "use `fix` instead of `condition`"

This reverts commit f87e2d1.

* rmeoved unnused symbol

* Revert "went back to using `DynamicPPL.condition` rather than using custom"

This reverts commit 53bd707.

* bump compat entry of DynamicPPL so we can overload acclogp!

* update assume for SMC samplers to make use of new `acclogp!`

* added proper impl of acclogp!! for SMC samplers + made accessing
task local varinfo and rng a bit nicer

* added experimental module and moved gibbs to it

* fixed now-inccorect references in new gibbs file

* updated gibbs tests

* moved experimental gibbs tests

* updated tests to include experiemntal tests

* removed refrences to previews tests of experimental Gibbs sampler

* removed solved TODO

* added a comment on `reconstruct_getvalue` usage

* bump patch version

* added comments on future work

* Update test/experimental/gibbs.jl

* fixed bug where particle samplers didn't properly account for
weightings of logpdf, etc.

* relax atol for a numerical test with Gibbs a bit

* fixed bug with `AbstractDict` constructor for experimental `Gibbs`

* aaaalways link the varinfo in the new Gibbs sampler, just to be sure

* add test to cover recent improvement to `DynamicPPL.subset`

ref: TuringLang/DynamicPPL.jl#587

* bump compat entry for DynamicPPL

* added some docstrings

* fixed test

* fixed import

* another attempt at fixing tests

* another attempt at fixing tests

* attempt at fix tests

* forgot something in previos commit

* cleaned up the experimental Gibbs sampler a bit

* removed accidentaly psuedocode inclusion

* Apply suggestions from code review

* relaxed olerance in one MH test a bit

* bump patch version

---------

Co-authored-by: Hong Ge <[email protected]>

Project.toml

@@ -1,6 +1,6 @@
 name = "Turing"
 uuid = "fce5fe82-541a-59a6-adf8-730c64b5f9a0"
-version = "0.30.8"
+version = "0.30.9"
 
 [deps]
 ADTypes = "47edcb42-4c32-4615-8424-f2b9edc5f35b"
@@ -58,7 +58,7 @@ Distributions = "0.23.3, 0.24, 0.25"
 DistributionsAD = "0.6"
 DocStringExtensions = "0.8, 0.9"
 DynamicHMC = "3.4"
-DynamicPPL = "0.24.7"
+DynamicPPL = "0.24.10"
 EllipticalSliceSampling = "0.5, 1, 2"
 ForwardDiff = "0.10.3"
 Libtask = "0.7, 0.8"

src/Turing.jl

@@ -48,6 +48,7 @@ using .Variational
 include("optimisation/Optimisation.jl")
 using .Optimisation
 
+include("experimental/Experimental.jl")
 include("deprecated.jl") # to be removed in the next minor version release
 
 ###########

src/experimental/Experimental.jl

@@ -0,0 +1,16 @@
+module Experimental
+
+using Random: Random
+using AbstractMCMC: AbstractMCMC
+using DynamicPPL: DynamicPPL, VarName
+using Setfield: Setfield
+
+using DocStringExtensions: TYPEDFIELDS
+using Distributions
+
+using ..Turing: Turing
+using ..Turing.Inference: gibbs_rerun, InferenceAlgorithm
+
+include("gibbs.jl")
+
+end