Unify Turing Transitions, fix some tests (#2651)

* Unify `Transition` methods

* Add tests

* Add same test for SGLD/SGHMC

* Refactor so that it's nice and organised

* Fix failing test on 1.10

* just increase the atol

* Make addlogprob test more robust

* Remove stray `@show`

Co-authored-by: Markus Hauru <[email protected]>

---------

Co-authored-by: Markus Hauru <[email protected]>

Author: penelopeysm

src/mcmc/Inference.jl

@@ -124,19 +124,16 @@ end
 # Default Transition #
 ######################
 getstats(::Any) = NamedTuple()
+getstats(nt::NamedTuple) = nt
 
-# TODO(penelopeysm): Remove this abstract type by converting SGLDTransition,
-# SMCTransition, and PGTransition to Turing.Inference.Transition instead.
-abstract type AbstractTransition end
-
-struct Transition{T,F<:AbstractFloat,N<:NamedTuple} <: AbstractTransition
+struct Transition{T,F<:AbstractFloat,N<:NamedTuple}
     θ::T
     logprior::F
     loglikelihood::F
     stat::N
 
     """
-        Transition(model::Model, vi::AbstractVarInfo, sampler_transition; reevaluate=true)
+        Transition(model::Model, vi::AbstractVarInfo, stats; reevaluate=true)
 
     Construct a new `Turing.Inference.Transition` object using the outputs of a
     sampler step.
@@ -146,8 +143,10 @@ struct Transition{T,F<:AbstractFloat,N<:NamedTuple} <: AbstractTransition
     have junk contents. The role of this method is to re-evaluate `model` and
     thus set the accumulators to the correct values.
 
-    `sampler_transition` is the transition object returned by the sampler
-    itself and is only used to extract statistics of interest.
+    `stats` is any object on which `Turing.Inference.getstats` can be called to
+    return a NamedTuple of statistics. This could be, for example, the transition
+    returned by an (unwrapped) external sampler. Or alternatively, it could
+    simply be a NamedTuple itself (for which `getstats` acts as the identity).
 
     By default, the model is re-evaluated in order to obtain values of:
       - the values of the parameters as per user parameterisation (`vals_as_in_model`)
@@ -167,8 +166,11 @@ struct Transition{T,F<:AbstractFloat,N<:NamedTuple} <: AbstractTransition
     must be set up to track `x := y` statements.
     """
     function Transition(
-        model::DynamicPPL.Model, vi::AbstractVarInfo, sampler_transition; reevaluate=true
+        model::DynamicPPL.Model, vi::AbstractVarInfo, stats; reevaluate=true
     )
+        # Avoid mutating vi as it may be used later e.g. when constructing
+        # sampler states.
+        vi = deepcopy(vi)
         if reevaluate
             vi = DynamicPPL.setaccs!!(
                 vi,
@@ -187,7 +189,7 @@ struct Transition{T,F<:AbstractFloat,N<:NamedTuple} <: AbstractTransition
         loglikelihood = DynamicPPL.getloglikelihood(vi)
 
         # Get additional statistics
-        stats = getstats(sampler_transition)
+        stats = getstats(stats)
         return new{typeof(vals_as_in_model),typeof(logprior),typeof(stats)}(
             vals_as_in_model, logprior, loglikelihood, stats
         )
@@ -196,17 +198,14 @@ struct Transition{T,F<:AbstractFloat,N<:NamedTuple} <: AbstractTransition
     function Transition(
         model::DynamicPPL.Model,
         untyped_vi::DynamicPPL.VarInfo{<:DynamicPPL.Metadata},
-        sampler_transition;
+        stats;
         reevaluate=true,
     )
         # Re-evaluating the model is unconscionably slow for untyped VarInfo. It's
         # much faster to convert it to a typed varinfo first, hence this method.
         # https://github.com/TuringLang/Turing.jl/issues/2604
         return Transition(
-            model,
-            DynamicPPL.typed_varinfo(untyped_vi),
-            sampler_transition;
-            reevaluate=reevaluate,
+            model, DynamicPPL.typed_varinfo(untyped_vi), stats; reevaluate=reevaluate
         )
     end
 end
@@ -318,7 +317,7 @@ getlogevidence(transitions, sampler, state) = missing
 # Default MCMCChains.Chains constructor.
 # This is type piracy (at least for SampleFromPrior).
 function AbstractMCMC.bundle_samples(
-    ts::Vector{<:Union{AbstractTransition,AbstractVarInfo}},
+    ts::Vector{<:Union{Transition,AbstractVarInfo}},
     model::AbstractModel,
     spl::Union{Sampler{<:InferenceAlgorithm},SampleFromPrior,RepeatSampler},
     state,
@@ -381,7 +380,7 @@ end
 
 # This is type piracy (for SampleFromPrior).
 function AbstractMCMC.bundle_samples(
-    ts::Vector{<:Union{AbstractTransition,AbstractVarInfo}},
+    ts::Vector{<:Union{Transition,AbstractVarInfo}},
     model::AbstractModel,
     spl::Union{Sampler{<:InferenceAlgorithm},SampleFromPrior,RepeatSampler},
     state,

src/mcmc/particle_mcmc.jl

@@ -135,23 +135,6 @@ function SMC(threshold::Real)
     return SMC(AdvancedPS.resample_systematic, threshold)
 end
 
-struct SMCTransition{T,F<:AbstractFloat} <: AbstractTransition
-    "The parameters for any given sample."
-    θ::T
-    "The joint log probability of the sample (NOTE: does not work, always set to zero)."
-    lp::F
-    "The weight of the particle the sample was retrieved from."
-    weight::F
-end
-
-function SMCTransition(model::DynamicPPL.Model, vi::AbstractVarInfo, weight)
-    theta = getparams(model, vi)
-    lp = DynamicPPL.getlogjoint_internal(vi)
-    return SMCTransition(theta, lp, weight)
-end
-
-getstats_with_lp(t::SMCTransition) = (lp=t.lp, weight=t.weight)
-
 struct SMCState{P,F<:AbstractFloat}
     particles::P
     particleindex::Int
@@ -228,7 +211,8 @@ function DynamicPPL.initialstep(
     weight = AdvancedPS.getweight(particles, 1)
 
     # Compute the first transition and the first state.
-    transition = SMCTransition(model, particle.model.f.varinfo, weight)
+    stats = (; weight=weight, logevidence=logevidence)
+    transition = Transition(model, particle.model.f.varinfo, stats)
     state = SMCState(particles, 2, logevidence)
 
     return transition, state
@@ -246,7 +230,8 @@ function AbstractMCMC.step(
     weight = AdvancedPS.getweight(particles, index)
 
     # Compute the transition and the next state.
-    transition = SMCTransition(model, particle.model.f.varinfo, weight)
+    stats = (; weight=weight, logevidence=state.average_logevidence)
+    transition = Transition(model, particle.model.f.varinfo, stats)
     nextstate = SMCState(state.particles, index + 1, state.average_logevidence)
 
     return transition, nextstate
@@ -300,32 +285,28 @@ Equivalent to [`PG`](@ref).
 """
 const CSMC = PG # type alias of PG as Conditional SMC
 
-struct PGTransition{T,F<:AbstractFloat} <: AbstractTransition
-    "The parameters for any given sample."
-    θ::T
-    "The joint log probability of the sample (NOTE: does not work, always set to zero)."
-    lp::F
-    "The log evidence of the sample."
-    logevidence::F
-end
-
 struct PGState
     vi::AbstractVarInfo
     rng::Random.AbstractRNG
 end
 
 get_varinfo(state::PGState) = state.vi
 
-function PGTransition(model::DynamicPPL.Model, vi::AbstractVarInfo, logevidence)
-    theta = getparams(model, vi)
-    lp = DynamicPPL.getlogjoint_internal(vi)
-    return PGTransition(theta, lp, logevidence)
-end
-
-getstats_with_lp(t::PGTransition) = (lp=t.lp, logevidence=t.logevidence)
-
-function getlogevidence(samples, sampler::Sampler{<:PG}, state::PGState)
-    return mean(x.logevidence for x in samples)
+function getlogevidence(
+    transitions::AbstractVector{<:Turing.Inference.Transition},
+    sampler::Sampler{<:PG},
+    state::PGState,
+)
+    logevidences = map(transitions) do t
+        if haskey(t.stat, :logevidence)
+            return t.stat.logevidence
+        else
+            # This should not really happen, but if it does we can handle it
+            # gracefully
+            return missing
+        end
+    end
+    return mean(logevidences)
 end
 
 function DynamicPPL.initialstep(
@@ -357,7 +338,7 @@ function DynamicPPL.initialstep(
 
     # Compute the first transition.
     _vi = reference.model.f.varinfo
-    transition = PGTransition(model, _vi, logevidence)
+    transition = Transition(model, _vi, (; logevidence=logevidence))
 
     return transition, PGState(_vi, reference.rng)
 end
@@ -397,7 +378,7 @@ function AbstractMCMC.step(
 
     # Compute the transition.
     _vi = newreference.model.f.varinfo
-    transition = PGTransition(model, _vi, logevidence)
+    transition = Transition(model, _vi, (; logevidence=logevidence))
 
     return transition, PGState(_vi, newreference.rng)
 end

src/mcmc/sghmc.jl

@@ -184,23 +184,6 @@ function SGLD(;
     return SGLD(stepsize, adtype)
 end
 
-struct SGLDTransition{T,F<:Real} <: AbstractTransition
-    "The parameters for any given sample."
-    θ::T
-    "The joint log probability of the sample."
-    lp::F
-    "The stepsize that was used to obtain the sample."
-    stepsize::F
-end
-
-function SGLDTransition(model::DynamicPPL.Model, vi::AbstractVarInfo, stepsize)
-    theta = getparams(model, vi)
-    lp = DynamicPPL.getlogjoint_internal(vi)
-    return SGLDTransition(theta, lp, stepsize)
-end
-
-getstats_with_lp(t::SGLDTransition) = (lp=t.lp, SGLD_stepsize=t.stepsize)
-
 struct SGLDState{L,V<:AbstractVarInfo}
     logdensity::L
     vi::V
@@ -220,13 +203,13 @@ function DynamicPPL.initialstep(
     end
 
     # Create first sample and state.
-    sample = SGLDTransition(model, vi, zero(spl.alg.stepsize(0)))
+    transition = Transition(model, vi, (; SGLD_stepsize=zero(spl.alg.stepsize(0))))
     ℓ = DynamicPPL.LogDensityFunction(
         model, DynamicPPL.getlogjoint_internal, vi; adtype=spl.alg.adtype
     )
     state = SGLDState(ℓ, vi, 1)
 
-    return sample, state
+    return transition, state
 end
 
 function AbstractMCMC.step(
@@ -245,8 +228,8 @@ function AbstractMCMC.step(
     vi = DynamicPPL.unflatten(vi, θ)
 
     # Compute next sample and state.
-    sample = SGLDTransition(model, vi, stepsize)
+    transition = Transition(model, vi, (; SGLD_stepsize=stepsize))
     newstate = SGLDState(ℓ, vi, state.step + 1)
 
-    return sample, newstate
+    return transition, newstate
 end

test/mcmc/gibbs.jl

@@ -598,8 +598,8 @@ end
             means = Dict(:b => 0.5, "x[1]" => 1.0, "x[2]" => 2.0)
             stds = Dict(:b => 0.5, "x[1]" => 1.0, "x[2]" => 1.0)
             for vn in keys(means)
-                @test isapprox(mean(skipmissing(chain[:, vn, 1])), means[vn]; atol=0.1)
-                @test isapprox(std(skipmissing(chain[:, vn, 1])), stds[vn]; atol=0.1)
+                @test isapprox(mean(skipmissing(chain[:, vn, 1])), means[vn]; atol=0.15)
+                @test isapprox(std(skipmissing(chain[:, vn, 1])), stds[vn]; atol=0.15)
             end
         end
 
@@ -651,7 +651,7 @@ end
             chain = sample(
                 StableRNG(468),
                 model,
-                Gibbs(:b => PG(10), :x => ESS()),
+                Gibbs(:b => PG(20), :x => ESS()),
                 2000;
                 discard_initial=100,
             )

test/mcmc/particle_mcmc.jl

@@ -1,10 +1,11 @@
 module ParticleMCMCTests
 
 using ..Models: gdemo_default
-#using ..Models: MoGtest, MoGtest_default
+using ..SamplerTestUtils: test_chain_logp_metadata
 using AdvancedPS: ResampleWithESSThreshold, resample_systematic, resample_multinomial
 using Distributions: Bernoulli, Beta, Gamma, Normal, sample
 using Random: Random
+using StableRNGs: StableRNG
 using Test: @test, @test_throws, @testset
 using Turing
 
@@ -49,6 +50,10 @@ using Turing
         @test_throws ErrorException sample(fail_smc(), SMC(), 100)
     end
 
+    @testset "chain log-density metadata" begin
+        test_chain_logp_metadata(SMC())
+    end
+
     @testset "logevidence" begin
         Random.seed!(100)
 
@@ -65,7 +70,10 @@ using Turing
         chains_smc = sample(test(), SMC(), 100)
 
         @test all(isone, chains_smc[:x])
+        # the chain itself has a logevidence field
         @test chains_smc.logevidence ≈ -2 * log(2)
+        # but each transition also contains the logevidence
+        @test chains_smc[:logevidence] ≈ fill(chains_smc.logevidence, 100)
     end
 end
 
@@ -88,6 +96,10 @@ end
         @test s.resampler === resample_systematic
     end
 
+    @testset "chain log-density metadata" begin
+        test_chain_logp_metadata(PG(10))
+    end
+
     @testset "logevidence" begin
         Random.seed!(100)
 
@@ -105,6 +117,7 @@ end
 
         @test all(isone, chains_pg[:x])
         @test chains_pg.logevidence ≈ -2 * log(2) atol = 0.01
+        @test chains_pg[:logevidence] ≈ fill(chains_pg.logevidence, 100)
     end
 
     # https://github.com/TuringLang/Turing.jl/issues/1598
@@ -114,6 +127,24 @@ end
         @test length(unique(c[:s])) == 1
     end
 
+    @testset "addlogprob leads to reweighting" begin
+        # Make sure that PG takes @addlogprob! into account. It didn't use to:
+        # https://github.com/TuringLang/Turing.jl/issues/1996
+        @model function addlogprob_demo()
+            x ~ Normal(0, 1)
+            if x < 0
+                @addlogprob! -10.0
+            else
+                # Need a balanced number of addlogprobs in all branches, or
+                # else PG will error
+                @addlogprob! 0.0
+            end
+        end
+        c = sample(StableRNG(468), addlogprob_demo(), PG(10), 100)
+        # Result should be biased towards x > 0.
+        @test mean(c[:x]) > 0.7
+    end
+
     # https://github.com/TuringLang/Turing.jl/issues/2007
     @testset "keyword arguments not supported" begin
         @model kwarg_demo(; x=2) = return x

test/mcmc/sghmc.jl

@@ -2,6 +2,7 @@ module SGHMCTests
 
 using ..Models: gdemo_default
 using ..NumericalTests: check_gdemo
+using ..SamplerTestUtils: test_chain_logp_metadata
 using DynamicPPL.TestUtils.AD: run_ad
 using DynamicPPL.TestUtils: DEMO_MODELS
 using DynamicPPL: DynamicPPL
@@ -32,6 +33,10 @@ using Turing
         chain = sample(rng, gdemo_default, alg, 10_000)
         check_gdemo(chain; atol=0.1)
     end
+
+    @testset "chain log-density metadata" begin
+        test_chain_logp_metadata(SGHMC(; learning_rate=0.02, momentum_decay=0.5))
+    end
 end
 
 @testset "Testing sgld.jl" begin
@@ -46,6 +51,7 @@ end
         sampler = DynamicPPL.Sampler(alg)
         @test sampler isa DynamicPPL.Sampler{<:SGLD}
     end
+
     @testset "sgld inference" begin
         rng = StableRNG(1)
 
@@ -59,6 +65,10 @@ end
         @test s_weighted ≈ 49 / 24 atol = 0.2
         @test m_weighted ≈ 7 / 6 atol = 0.2
     end
+
+    @testset "chain log-density metadata" begin
+        test_chain_logp_metadata(SGLD(; stepsize=PolynomialStepsize(0.25)))
+    end
 end
 
 end