Running Pathfinder on Turing.jl models

This tutorial demonstrates how Turing can be used with Pathfinder.

We'll demonstrate with a regression example.

using AbstractMCMC, AdvancedHMC, DynamicPPL, FlexiChains, Pathfinder, Random, Turing
Random.seed!(39)

@model function regress(x)
    α ~ Normal()
    β ~ Normal()
    σ ~ truncated(Normal(); lower=0)
    μ = α .+ β .* x
    y ~ product_distribution(Normal.(μ, σ))
end
x = 0:0.1:10
true_params = (; α=1.5, β=2, σ=2)
# simulate data
(; y) = rand(regress(x) | true_params)

model = regress(x) | (; y)
n_chains = 8

For convenience, pathfinder and multipathfinder can take Turing models as inputs and produce MCMCChains.Chains or FlexiChains.VNChain objects as outputs. To access this, we run Pathfinder normally; the chains representation of the draws (defaulting to Chains) is stored in draws_transformed.

result_single = pathfinder(model; ndraws=1_000)
Single-path Pathfinder result
  tries: 1
  draws: 1000
  fit iteration: 14 (total: 16)
  fit ELBO: -213.66 ± 0.09
  fit distribution: MvNormal{Float64, Pathfinder.WoodburyPDMat{Float64, LinearAlgebra.Diagonal{Float64, Vector{Float64}}, Matrix{Float64}, Matrix{Float64}, Pathfinder.WoodburyPDFactorization{Float64, LinearAlgebra.Diagonal{Float64, Vector{Float64}}, LinearAlgebra.QRCompactWYQ{Float64, Matrix{Float64}, Matrix{Float64}}, LinearAlgebra.UpperTriangular{Float64, Matrix{Float64}}}}, Vector{Float64}}(
dim: 3
μ: [1.6508971085750253, 1.9311753921107158, 0.5801261338729266]
Σ: [0.11087798427694032 -0.016527256928018597 -0.0014208078651489035; -0.016527256928018624 0.0034071472308229607 0.00019433068465847872; -0.0014208078651489287 0.00019433068465846744 0.004718299969699233]
)
result_single.draws_transformed
Chains MCMC chain (1000×6×1 Array{Float64, 3}):

Iterations        = 1:1:1000
Number of chains  = 1
Samples per chain = 1000
parameters        = α, β, σ
internals         = logprior, loglikelihood, logjoint

Use `describe(chains)` for summary statistics and quantiles.

To request a different chain type (e.g. VNChain), we can specify the chain_type directly.

pathfinder(model; ndraws=1_000, chain_type=VNChain).draws_transformed
FlexiChain (1000 iterations, 1 chain)
↓ iter=1:1000 | → chain=1:1

Parameter type   VarName
Parameters       α, β, σ
Extra keys       :logprior, :loglikelihood, :logjoint

Note that while Turing's sample methods default to initializing parameters from the prior with InitFromPrior, Pathfinder defaults to uniformly sampling them in the range [-2, 2] in unconstrained space (equivalent to Turing's InitFromUniform(-2, 2)). To use Turing's default in Pathfinder, specify init_sampler=InitFromPrior().

result_multi = multipathfinder(model, 1_000; nruns=n_chains, init_sampler=InitFromPrior())
Multi-path Pathfinder result
  runs: 8
  draws: 1000
  Pareto shape diagnostic: 0.57 (ok)

The Pareto shape diagnostic indicates that it is likely safe to use these draws to compute posterior estimates.

chns_pf = result_multi.draws_transformed
describe(chns_pf)
Chains MCMC chain (1000×6×1 Array{Float64, 3}):

Iterations        = 1:1:1000
Number of chains  = 1
Samples per chain = 1000
parameters        = α, β, σ
internals         = logprior, loglikelihood, logjoint

Summary Statistics

  parameters      mean       std      mcse    ess_bulk    ess_tail      rhat   ess_per_sec 
      Symbol   Float64   Float64   Float64     Float64     Float64   Float64       Missing 

           α    1.6431    0.3356    0.0108    954.2866   1023.6250    1.0002       missing
           β    1.9310    0.0587    0.0018   1053.3388    845.1344    0.9990       missing
           σ    1.8216    0.1289    0.0042    931.3650    804.3368    1.0001       missing


Quantiles

  parameters      2.5%     25.0%     50.0%     75.0%     97.5% 
      Symbol   Float64   Float64   Float64   Float64   Float64 

           α    0.9977    1.4224    1.6388    1.8658    2.3097
           β    1.8241    1.8886    1.9330    1.9678    2.0429
           σ    1.5913    1.7300    1.8096    1.8964    2.0972

We can also use these draws to initialize MCMC sampling with InitFromParams.

params = AbstractMCMC.to_samples(DynamicPPL.ParamsWithStats, chns_pf[1:n_chains, :, :])
initial_params = [InitFromParams(p.params) for p in vec(params)]
chns = sample(model, Turing.NUTS(), MCMCThreads(), 1_000, n_chains; initial_params, progress=false)
describe(chns)
┌ Warning: Only a single thread available: MCMC chains are not sampled in parallel
@ AbstractMCMC ~/.julia/packages/AbstractMCMC/oqm6Y/src/sample.jl:544
┌ Info: Found initial step size
  ϵ = 0.2
┌ Info: Found initial step size
  ϵ = 0.05
┌ Info: Found initial step size
  ϵ = 0.2
┌ Info: Found initial step size
  ϵ = 0.025
┌ Info: Found initial step size
  ϵ = 0.2
┌ Info: Found initial step size
  ϵ = 0.025
┌ Info: Found initial step size
  ϵ = 0.046875
┌ Info: Found initial step size
  ϵ = 0.046875
Chains MCMC chain (1000×17×8 Array{Float64, 3}):

Iterations        = 501:1:1500
Number of chains  = 8
Samples per chain = 1000
Wall duration     = 5.28 seconds
Compute duration  = 3.8 seconds
parameters        = α, β, σ
internals         = n_steps, is_accept, acceptance_rate, log_density, hamiltonian_energy, hamiltonian_energy_error, max_hamiltonian_energy_error, tree_depth, numerical_error, step_size, nom_step_size, logprior, loglikelihood, logjoint

Summary Statistics

  parameters      mean       std      mcse    ess_bulk    ess_tail      rhat   ess_per_sec 
      Symbol   Float64   Float64   Float64     Float64     Float64   Float64       Float64 

           α    1.6455    0.3393    0.0058   3386.1924   4204.2384    1.0013      892.2773
           β    1.9317    0.0595    0.0010   3435.4787   3999.7628    1.0014      905.2645
           σ    1.8160    0.1286    0.0018   4904.3235   4736.9706    1.0000     1292.3119


Quantiles

  parameters      2.5%     25.0%     50.0%     75.0%     97.5% 
      Symbol   Float64   Float64   Float64   Float64   Float64 

           α    0.9754    1.4161    1.6501    1.8703    2.3164
           β    1.8154    1.8926    1.9316    1.9717    2.0477
           σ    1.5875    1.7257    1.8068    1.8991    2.0920

We can use Pathfinder's estimate of the metric and only perform enough warm-up to tune the step size.

inv_metric = result_multi.pathfinder_results[1].fit_distribution.Σ
metric = Pathfinder.RankUpdateEuclideanMetric(inv_metric)
kernel = HMCKernel(Trajectory{MultinomialTS}(Leapfrog(0.0), GeneralisedNoUTurn()))
adaptor = StepSizeAdaptor(0.8, 1.0)  # adapt only the step size
nuts = AdvancedHMC.HMCSampler(kernel, metric, adaptor)

n_adapts = 50
n_draws = 1_000
chns = sample(
    model,
    externalsampler(nuts),
    MCMCThreads(),
    n_draws + n_adapts,
    n_chains;
    n_adapts,
    initial_params,
    progress=false,
)[n_adapts + 1:end, :, :]  # drop warm-up draws
describe(chns)
┌ Warning: Only a single thread available: MCMC chains are not sampled in parallel
@ AbstractMCMC ~/.julia/packages/AbstractMCMC/oqm6Y/src/sample.jl:544
[ Info: Found initial step size 1.6
[ Info: Found initial step size 3.2
[ Info: Found initial step size 1.6
[ Info: Found initial step size 3.2
[ Info: Found initial step size 1.6
[ Info: Found initial step size 1.7000000000000002
[ Info: Found initial step size 1.6125
[ Info: Found initial step size 1.6
Chains MCMC chain (1000×17×8 Array{Float64, 3}):

Iterations        = 51:1:1050
Number of chains  = 8
Samples per chain = 1000
Wall duration     = 1.64 seconds
Compute duration  = 1.45 seconds
parameters        = α, β, σ
internals         = n_steps, is_accept, acceptance_rate, log_density, hamiltonian_energy, hamiltonian_energy_error, max_hamiltonian_energy_error, tree_depth, numerical_error, step_size, nom_step_size, logprior, loglikelihood, logjoint

Summary Statistics

  parameters      mean       std      mcse     ess_bulk    ess_tail      rhat   ess_per_sec 
      Symbol   Float64   Float64   Float64      Float64     Float64   Float64       Float64 

           α    1.6439    0.3393    0.0034   10272.6852   6279.7259    1.0018     7084.6104
           β    1.9317    0.0591    0.0006   10322.0121   6023.3676    1.0018     7118.6290
           σ    1.8150    0.1258    0.0012   10411.2790   6212.8585    1.0013     7180.1924


Quantiles

  parameters      2.5%     25.0%     50.0%     75.0%     97.5% 
      Symbol   Float64   Float64   Float64   Float64   Float64 

           α    0.9618    1.4235    1.6464    1.8672    2.3049
           β    1.8163    1.8914    1.9310    1.9707    2.0494
           σ    1.5880    1.7259    1.8084    1.8946    2.0777

See Initializing HMC with Pathfinder for further examples.