Even more fixes

Author: penelopeysm

ext/TuringDynamicHMCExt.jl

@@ -75,7 +75,9 @@ function DynamicPPL.initialstep(
 
     # Update the variables.
     vi = DynamicPPL.unflatten(vi, Q.q)
-    vi = DynamicPPL.setlogp!!(vi, Q.ℓq)
+    # TODO(DPPL0.37/penelopeysm): This is obviously incorrect. Fix this.
+    vi = DynamicPPL.setloglikelihood!!(vi, Q.ℓq)
+    vi = DynamicPPL.setlogprior!!(vi, 0.0)
 
     # Create first sample and state.
     sample = Turing.Inference.Transition(model, vi)

src/mcmc/mh.jl

@@ -194,7 +194,7 @@ end
 function LogDensityProblems.logdensity(f::LogDensityFunction, x::NamedTuple)
     vi = deepcopy(f.varinfo)
     set_namedtuple!(vi, x)
-    vi_new = last(DynamicPPL.evaluate!!(f.model, vi, f.context))
+    vi_new = last(DynamicPPL.evaluate!!(f.model, vi))
     lj = f.getlogdensity(vi_new)
     return lj
 end

test/mcmc/hmc.jl

@@ -171,25 +171,6 @@ using Turing
         @test Array(res1) == Array(res2) == Array(res3)
     end
 
-    # TODO(mhauru) Do we give up being able to sample from only prior/likelihood like this,
-    # or do we implement some way to pass `whichlogprob=:LogPrior` through `sample`?
-    @testset "prior" begin
-        # NOTE: Used to use `InverseGamma(2, 3)` but this has infinite variance
-        # which means that it's _very_ difficult to find a good tolerance in the test below:)
-        prior_dist = truncated(Normal(3, 1); lower=0)
-
-        @model function demo_hmc_prior()
-            s ~ prior_dist
-            return m ~ Normal(0, sqrt(s))
-        end
-        alg = NUTS(1000, 0.8)
-        gdemo_default_prior = DynamicPPL.contextualize(
-            demo_hmc_prior(), DynamicPPL.PriorContext()
-        )
-        chain = sample(gdemo_default_prior, alg, 5_000; initial_params=[3.0, 0.0])
-        check_numerical(chain, [:s, :m], [mean(prior_dist), 0]; atol=0.2)
-    end
-
     @testset "warning for difficult init params" begin
         attempt = 0
         @model function demo_warn_initial_params()

test/mcmc/mh.jl

@@ -262,26 +262,6 @@ GKernel(var) = (x) -> Normal(x, sqrt.(var))
         @test !DynamicPPL.islinked(vi)
     end
 
-    # TODO(mhauru) Do we give up being able to sample from only prior/likelihood like this,
-    # or do we implement some way to pass `whichlogprob=:LogPrior` through `sample`?
-    @testset "prior" begin
-        alg = MH()
-        gdemo_default_prior = DynamicPPL.contextualize(
-            gdemo_default, DynamicPPL.PriorContext()
-        )
-        burnin = 10_000
-        n = 10_000
-        chain = sample(
-            StableRNG(seed),
-            gdemo_default_prior,
-            alg,
-            n;
-            discard_initial=burnin,
-            thinning=10,
-        )
-        check_numerical(chain, [:s, :m], [mean(InverseGamma(2, 3)), 0]; atol=0.3)
-    end
-
     @testset "`filldist` proposal (issue #2180)" begin
         @model demo_filldist_issue2180() = x ~ MvNormal(zeros(3), I)
         chain = sample(