Compat for Turing v0.41 (#44)

* Compat for Turing v0.41

* Bump patch

* use original vi (shouldn't make a difference)

* Update comment

* bump deps in docs/ and test/ as well

* Format

* That was a mistake; use getlogdensity instead of hardcoding the function

* fix initial_params in tests

* Fix docs and format

* Fix more docs

* Fix docstring

Author: penelopeysm

.github/workflows/Docs.yml

@@ -24,6 +24,8 @@ jobs:
     steps:
       - name: Build and deploy Documenter.jl docs
         uses: TuringLang/actions/DocsDocumenter@main
+        with:
+          julia-version: 1.11
 
       - name: Run doctests
         shell: julia --project=docs --color=yes {0}

Project.toml

@@ -1,6 +1,6 @@
 name = "SliceSampling"
 uuid = "43f4d3e8-9711-4a8c-bd1b-03ac73a255cf"
-version = "0.7.8"
+version = "0.7.9"
 
 [deps]
 AbstractMCMC = "80f14c24-f653-4e6a-9b94-39d6b0f70001"
@@ -21,7 +21,7 @@ Distributions = "0.25"
 LinearAlgebra = "1"
 LogDensityProblems = "2"
 Random = "1"
-Turing = "0.40"
+Turing = "0.41"
 julia = "1.10"
 
 [extras]

docs/Project.toml

@@ -28,5 +28,5 @@ Random = "1"
 SliceSampling = "0.7.1"
 StableRNGs = "1"
 Statistics = "1"
-Turing = "0.37, 0.38, 0.39, 0.40"
+Turing = "0.41"
 julia = "1.10"

docs/src/gibbs_polar.md

@@ -63,8 +63,9 @@ end
 model = demo()
 
 n_samples = 1000
-latent_chain = sample(model, externalsampler(LatentSlice(10)), n_samples; initial_params=ones(10))
-polar_chain = sample(model, externalsampler(GibbsPolarSlice(10)), n_samples; initial_params=ones(10))
+initial_params = InitFromParams((x = ones(10),))
+latent_chain = sample(model, externalsampler(LatentSlice(10)), n_samples; initial_params=initial_params)
+polar_chain = sample(model, externalsampler(GibbsPolarSlice(10)), n_samples; initial_params=initial_params)
 
 l = @layout [a; b]
 p1 = Plots.plot(1:n_samples, latent_chain[:,1,:], ylims=[-10,10], label="LSS")

ext/SliceSamplingTuringExt.jl

@@ -38,33 +38,30 @@ end
 function SliceSampling.initial_sample(rng::Random.AbstractRNG, ℓ::Turing.LogDensityFunction)
     n_max_attempts = 1000
 
-    model  = ℓ.model
-    vi     = Turing.DynamicPPL.VarInfo(rng, model, Turing.SampleFromUniform())
-    vi_spl = last(Turing.DynamicPPL.evaluate_and_sample!!(rng, model, vi, Turing.SampleFromUniform()))
-    θ      = vi_spl[:]
-    ℓp     = LogDensityProblems.logdensity(ℓ, θ)
+    model, vi = ℓ.model, ℓ.varinfo
+    vi_spl = last(
+        Turing.DynamicPPL.init!!(rng, model, vi, Turing.DynamicPPL.InitFromUniform())
+    )
+    ℓp = ℓ.getlogdensity(vi_spl)
 
     init_attempt_count = 1
     for attempts in 1:n_max_attempts
         if attempts == 10
             @warn "Failed to find valid initial parameters after $(init_attempt_count) attempts; consider providing explicit initial parameters using the `initial_params` keyword"
         end
 
-        # NOTE: This will sample in the unconstrained space.
-        vi_spl = last(
-            Turing.DynamicPPL.evaluate_and_sample!!(
-                rng, model, vi, Turing.SampleFromUniform()
-            ),
-        )
+        # NOTE: This will sample in the unconstrained space if ℓ.varinfo is linked
+        vi_spl = last(Turing.DynamicPPL.init!!(rng, model, vi, Turing.InitFromUniform()))
+        ℓp = ℓ.getlogdensity(vi_spl)
         θ = vi_spl[:]
-        ℓp = LogDensityProblems.logdensity(ℓ, θ)
 
         if all(isfinite.(θ)) && isfinite(ℓp)
             return θ
         end
     end
 
     @error "Failed to find valid initial parameters after $(n_max_attempts) attempts; consider providing explicit initial parameters using the `initial_params` keyword"
+    θ = vi_spl[:]
     return θ
 end
 

src/multivariate/gibbspolar.jl

@@ -18,10 +18,7 @@ Gibbsian polar slice sampling algorithm by P. Schär, M. Habeck, and D. Rudolf [
     The initial window size `w` must be set at least an order of magnitude larger than what is sensible for other slice samplers. Otherwise, a large number of rejections might be experienced.
 
 !!! warning
-    When initializing the chain (*e.g.* the `initial_params` keyword arguments in `AbstractMCMC.sample`), it is necessary to inialize from a point \$\$x_0\$\$ that has a sensible norm \$\$\\lVert x_0 \\rVert > 0\$\$, otherwise, the chain will start from a pathologic point in polar coordinates. This might even result in the sampler getting stuck in an infinite loop. (This can be prevented by setting `max_proposals`.) If \$\$\\lVert x_0 \\rVert \\leq 10^{-5}\$\$, the current implementation will display a warning. 
-	
-!!! info
-    For Turing users: `Turing` might change `initial_params` to match the support of the posterior. This might lead to \$\$\\lVert x_0 \\rVert\$\$ being small, even though the vector you passed to`initial_params` has a sufficiently large norm. If this is suspected, simply try a different initialization value.
+    When initializing the chain (*e.g.* the `initial_params` keyword arguments in `AbstractMCMC.sample`), it is necessary to initialize from a point \$\$x_0\$\$ that has a sensible norm \$\$\\lVert x_0 \\rVert > 0\$\$, otherwise, the chain will start from a pathological point in polar coordinates. This might even result in the sampler getting stuck in an infinite loop. (This can be prevented by setting `max_proposals`.) If \$\$\\lVert x_0 \\rVert \\leq 10^{-5}\$\$, the current implementation will display a warning. 
 """
 struct GibbsPolarSlice{W<:Real} <: AbstractMultivariateSliceSampling
     w::W

test/Project.toml

@@ -18,5 +18,5 @@ MCMCTesting = "0.3"
 Random = "1"
 StableRNGs = "1"
 Test = "1"
-Turing = "0.37, 0.38, 0.39, 0.40"
+Turing = "0.41"
 julia = "1.10"

test/turing.jl

@@ -49,7 +49,7 @@
             model,
             externalsampler(sampler),
             n_samples;
-            initial_params=[1.0, 0.1],
+            initial_params=InitFromParams((s=1.0, m=0.1)),
             progress=false,
         )