This PR adds support for calling Pathfinder on MeasureTheory.AbstractMeasures. This indirectly should add support for Soss.jl and Tilde.jl, which both express conditional models as subtypes of AbstractMeasure

However, this code doesn't currently work for Soss, which uses an outdated version of MeasureTheory. In particular:

Soss uses MeasureTheory.logdensity, which is not defined in more recent versions of MeasureTheory.
Soss uses xform to get transformations, instead of TransformVariables.as, which MeasureTheory now uses.

cc @cscherrer

Example:

julia> using Pathfinder, MeasureTheory

julia> pathfinder(Beta(α=2, β=5))
[ Info: Optimized for 5 iterations. Maximum ELBO of 0.12 reached at iteration 5.
(Distributions.MvNormal{Float64, Pathfinder.WoodburyPDMat{Float64, LinearAlgebra.Diagonal{Float64, Vector{Float64}}, Matrix{Float64}, Matrix{Float64}, LinearAlgebra.Diagonal{Float64, Vector{Float64}}, LinearAlgebra.QRCompactWYQ{Float64, Matrix{Float64}}, LinearAlgebra.UpperTriangular{Float64, Matrix{Float64}}}, Vector{Float64}}(
dim: 1
μ: [-0.9162907318741551]
Σ: [0.6999999602078875;;]
)
, [-2.491897025639392 -2.268906498193299 … -1.9451470847676866 -1.2163230050919374], [-2.51384055715131, -2.0474364008828, -0.8489568418465779, -1.4967049292861692, -0.8049005828762684])

Apr 29 '22 01:04 sethaxen

Codecov Report

Merging #46 (3f9a048) into main (3516702) will decrease coverage by 5.98%. The diff coverage is 2.70%.

@@            Coverage Diff             @@
##             main      #46      +/-   ##
==========================================
- Coverage   92.89%   86.91%   -5.99%     
==========================================
  Files          13       14       +1     
  Lines         521      558      +37     
==========================================
+ Hits          484      485       +1     
- Misses         37       73      +36

Impacted Files	Coverage Δ
src/integration/measuretheory.jl	`0.00% <0.00%> (ø)`
src/Pathfinder.jl	`100.00% <100.00%> (ø)`

Continue to review full report at Codecov.

Legend - Click here to learn more Δ = absolute <relative> (impact), ø = not affected, ? = missing data Powered by Codecov. Last update 3516702...3f9a048. Read the comment docs.

Apr 29 '22 01:04 codecov[bot]

Very nice!! I'll find a couple of examples to try out. @mschauer this might also be useful for you, if your PDMP sampler needs a warm-up

Apr 29 '22 13:04 cscherrer

It gets better. Because DynamicHMC's API supports all initialization approaches we might use Pathfinder for, and because SampleChainsDynamicHMC exposes those configuration options untouched, we will be able to do this, which bypasses the Stan-style warm-up entirely and uses Pathfinder's estimate of the metric.

using Soss, SampleChains, SampleChainsDynamicHMC

model = ... # define Soss.ConditionalModel

result_pf = pathfinder(model)

config = dynamichmc(
    initialization=(;
        q=result_pf[2][:, 1], κ=GaussianKineticEnergy(result_pf[1].Σ)
    ),
    warmup_stages=default_warmup_stages(; middle_steps=0, doubling_stages=0),
    reporter=ProgressMeterReport(),
)

post = sample(mod, config, 1_000)

Apr 29 '22 13:04 sethaxen

This is so cool. Looks like I was missing some as methods, but adding those in lets us do things like

julia> using MeasureTheory

julia> using Pathfinder

julia> pathfinder(Normal() ⊗ Gamma(2,3))
[ Info: Optimized for 7 iterations. Maximum ELBO of -0.02 reached at iteration 1.
(Distributions.MvNormal{Float64, Pathfinder.WoodburyPDMat{Float64, LinearAlgebra.Diagonal{Float64, Vector{Float64}}, Matrix{Float64}, Matrix{Float64}, LinearAlgebra.Diagonal{Float64, Vector{Float64}}, LinearAlgebra.QRCompactWYQ{Float64, Matrix{Float64}, Matrix{Float64}}, LinearAlgebra.UpperTriangular{Float64, Matrix{Float64}}}, Vector{Float64}}(
dim: 2
μ: [0.037802, 1.71518]
Σ: [0.876813 -0.0536098; -0.0536098 0.536895]
)
, [-0.956011 -0.853106 … 0.10737 0.0554134; 1.2142 1.18968 … 1.85447 0.846168], [-2.317, -2.2259, -1.54494, -1.48017, -2.16414])

julia> pathfinder(For(rand(3)) do x InverseGaussian(x,5) end)
[ Info: Optimized for 9 iterations. Maximum ELBO of 0.12 reached at iteration 3.
(Distributions.MvNormal{Float64, Pathfinder.WoodburyPDMat{Float64, LinearAlgebra.Diagonal{Float64, Vector{Float64}}, Matrix{Float64}, Matrix{Float64}, LinearAlgebra.Diagonal{Float64, Vector{Float64}}, LinearAlgebra.QRCompactWYQ{Float64, Matrix{Float64}, Matrix{Float64}}, LinearAlgebra.UpperTriangular{Float64, Matrix{Float64}}}, Vector{Float64}}(
dim: 3
μ: [-1.85876, -1.2139, -0.466543]
Σ: [0.0316455 -0.00529572 -0.00216225; -0.00529572 0.0465599 -0.00650248; -0.00216225 -0.00650248 0.121807]
)
, [-2.02251 -1.48321 … -1.56437 -1.93967; -1.03717 -1.19963 … -1.41102 -1.77201; -0.843345 -0.597198 … -0.844978 -0.741263], [0.348766, -0.759287, 1.31935, -0.638448, -2.66604])

It will be really great to get it working with Soss and Tilde. Thanks for putting this together!

Apr 29 '22 14:04 cscherrer

It will be really great to get it working with Soss and Tilde. Thanks for putting this together!

You're welcome! Currently it returns the variational approximation and draws only in the unconstrained space. This is more useful for initializing MCMC, since the initial point and metric can be passed directly to dynamichmc. But it's not as useful for diagnosing issues, since they're easily user-interpretable. I would like to also return a transformed measure and transformed draws. I just need to think through a good design for this.

Apr 29 '22 14:04 sethaxen

@cscherrer it seems SuperpositionMeasure does not have a rand method:

julia> μ = Dirichlet(10, 1.0)
Dirichlet(α = Fill(1.0, 10),)

julia> result = multipathfinder(μ, 1_000; nruns=20)
┌ Warning: Pareto shape k = 0.8 > 0.7. Resulting importance sampling estimates are likely to be unstable.
└ @ PSIS ~/.julia/packages/PSIS/Rf0S3/src/core.jl:268
Multi-path Pathfinder result
  runs: 20
  draws: 9
  Pareto shape diagnostic: 0.8 (bad)

julia> result.fit_distribution_transformed |> typeof
Pushforward{TransformVariables.CallableTransform{TransformVariables.UnitSimplex}, SuperpositionMeasure{Vector{MvNormal{(:μ, :σ), Tuple{Vector{Float64}, Pathfinder.WoodburyPDMat{Float64, LinearAlgebra.Diagonal{Float64, Vector{Float64}}, Matrix{Float64}, Matrix{Float64}, LinearAlgebra.Diagonal{Float64, Vector{Float64}}, LinearAlgebra.QRCompactWYQ{Float64, Matrix{Float64}}, LinearAlgebra.UpperTriangular{Float64, Matrix{Float64}}}}}}}, Static.True}

julia> result.fit_distribution_transformed |> rand
ERROR: MethodError: no method matching rand(::Random._GLOBAL_RNG, ::Type{Float64}, ::SuperpositionMeasure{Vector{MvNormal{(:μ, :σ), Tuple{Vector{Float64}, Pathfinder.WoodburyPDMat{Float64, LinearAlgebra.Diagonal{Float64, Vector{Float64}}, Matrix{Float64}, Matrix{Float64}, LinearAlgebra.Diagonal{Float64, Vector{Float64}}, LinearAlgebra.QRCompactWYQ{Float64, Matrix{Float64}}, LinearAlgebra.UpperTriangular{Float64, Matrix{Float64}}}}}}})
...

julia> result.draws_transformed
1000-element Vector{Vector{Float64}}:
 [0.12646688194244224, 0.025977504820487007, 0.012354054767482447, 0.1334736170622469, 0.03246415592570656, 0.11141269659298245, 0.024034133186899103, 0.24475990778479786, 0.1342558624335393, 0.15480118548341593]
 [0.0169086648766402, 0.09911231516043292, 0.34416889221132596, 0.10180645019569193, 0.14439126397502686, 0.11713472197731346, 0.08444424836455564, 0.07675136145856229, 0.004312850832038056, 0.010969230948412829]
 [0.204562486395576, 0.10418880831568512, 0.14665258035543918, 0.03270907277182203, 0.025198320818369802, 0.08760170721317947, 0.22661983988004777, 0.0454094719713609, 0.08236705364451188, 0.04469065863400787]
 [0.01816521614098672, 0.08744121201467416, 0.08194354696600549, 0.21065542194335074, 0.05446274719421834, 0.2462246106035293, 0.08504411121447199, 0.09401263968332611, 0.05085735912825384, 0.07119313511118336]
 ⋮
 [0.05276387864922493, 0.06859177668105768, 0.12000714757911174, 0.013786343525336437, 0.0936220599997711, 0.014944597056716297, 0.15220938553755983, 0.32035991301423006, 0.07951434052610347, 0.08420055743088842]
 [0.037353097348125405, 0.1003300343889233, 0.014071365428468915, 0.19619179988878918, 0.002071999284771993, 0.15604889644757544, 0.394789935358081, 0.017367043688102463, 0.02483761912275614, 0.05693820904440616]
 [0.092177514123393, 0.03885924399407327, 0.11357006993749118, 0.040281419867287165, 0.0017297788961319048, 0.30301174382535745, 0.08705530131888517, 0.10436044162668617, 0.12720672435099528, 0.09174776205969937]
 [0.1036299013468673, 0.0237354567117644, 0.00808528404000608, 0.08094192248237413, 0.012319402714347172, 0.17985596931554573, 0.044975357793939554, 0.09246673694172126, 0.33590939793731694, 0.11808057071611726]

May 02 '22 11:05 sethaxen

Pausing development of this until Soss works with HMC again, see https://github.com/cscherrer/Soss.jl/issues/337

May 31 '22 15:05 sethaxen

Add MeasureTheory support

Codecov Report