Updated HMC implementation for new AHMC version (#1660)

* updated HMC implementation according to new AHMC interface

* bump compat bound for AdvancedHMC

* bumped patch version

* disable GMM Gibbs conditional test to see if it fixes CI

* include tests again

* dont test non-AD samplers for every AD backend

* added back a test

* added back a test

* removed some redundant tests and fixed a typo

* added macro timed_testset

* upper-bound Distributions.jl apparently fixes the test-freeze

* hyphen compat specifies arent compatible with Julia 1.3

* removed the timed_testset stuff

Author: torfjelde

Project.toml

@@ -1,6 +1,6 @@
 name = "Turing"
 uuid = "fce5fe82-541a-59a6-adf8-730c64b5f9a0"
-version = "0.16.5"
+version = "0.16.6"
 
 [deps]
 AbstractMCMC = "80f14c24-f653-4e6a-9b94-39d6b0f70001"
@@ -34,7 +34,7 @@ ZygoteRules = "700de1a5-db45-46bc-99cf-38207098b444"
 
 [compat]
 AbstractMCMC = "3.2"
-AdvancedHMC = "0.2.24"
+AdvancedHMC = "0.3.0"
 AdvancedMH = "0.6"
 AdvancedPS = "0.2.4"
 AdvancedVI = "0.1"

src/inference/gibbs.jl

@@ -126,7 +126,7 @@ function gibbs_state(
     state.z.θ .= θ_old
     z = state.z
 
-    return HMCState(varinfo, state.i, state.traj, hamiltonian, z, state.adaptor)
+    return HMCState(varinfo, state.i, state.kernel, hamiltonian, z, state.adaptor)
 end
 
 """

src/inference/hmc.jl

@@ -4,14 +4,14 @@
 
 struct HMCState{
     TV<:AbstractVarInfo,
-    TTraj<:AHMC.AbstractTrajectory,
+    TKernel<:AHMC.HMCKernel,
     THam<:AHMC.Hamiltonian,
     PhType<:AHMC.PhasePoint,
     TAdapt<:AHMC.Adaptation.AbstractAdaptor,
 }
     vi::TV
     i::Int
-    traj::TTraj
+    kernel::TKernel
     hamiltonian::THam
     z::PhType
     adaptor::TAdapt
@@ -190,18 +190,18 @@ function DynamicPPL.initialstep(
         ϵ = spl.alg.ϵ
     end
 
-    # Generate a trajectory.
-    traj = gen_traj(spl.alg, ϵ)
+    # Generate a kernel.
+    kernel = make_ahmc_kernel(spl.alg, ϵ)
 
     # Create initial transition and state.
     # Already perform one step since otherwise we don't get any statistics.
-    t = AHMC.step(rng, hamiltonian, traj, z)
+    t = AHMC.transition(rng, hamiltonian, kernel, z)
 
     # Adaptation
     adaptor = AHMCAdaptor(spl.alg, hamiltonian.metric; ϵ=ϵ)
     if spl.alg isa AdaptiveHamiltonian
-        hamiltonian, traj, _ =
-            AHMC.adapt!(hamiltonian, traj, adaptor,
+        hamiltonian, kernel, _ =
+            AHMC.adapt!(hamiltonian, kernel, adaptor,
                         1, nadapts, t.z.θ, t.stat.acceptance_rate)
     end
 
@@ -215,7 +215,7 @@ function DynamicPPL.initialstep(
     end
 
     transition = HMCTransition(vi, t)
-    state = HMCState(vi, 1, traj, hamiltonian, t.z, adaptor)
+    state = HMCState(vi, 1, kernel, hamiltonian, t.z, adaptor)
 
     return transition, state
 end
@@ -234,16 +234,16 @@ function AbstractMCMC.step(
     # Compute transition.
     hamiltonian = state.hamiltonian
     z = state.z
-    t = AHMC.step(rng, hamiltonian, state.traj, z)
+    t = AHMC.transition(rng, hamiltonian, state.kernel, z)
 
     # Adaptation
     i = state.i + 1
     if spl.alg isa AdaptiveHamiltonian
-        hamiltonian, traj, _ =
-            AHMC.adapt!(hamiltonian, state.traj, state.adaptor,
+        hamiltonian, kernel, _ =
+            AHMC.adapt!(hamiltonian, state.kernel, state.adaptor,
                         i, nadapts, t.z.θ, t.stat.acceptance_rate)
     else
-        traj = state.traj
+        kernel = state.kernel
     end
 
     # Update variables
@@ -255,7 +255,7 @@ function AbstractMCMC.step(
 
     # Compute next transition and state.
     transition = HMCTransition(vi, t)
-    newstate = HMCState(vi, i, traj, hamiltonian, t.z, state.adaptor)
+    newstate = HMCState(vi, i, kernel, hamiltonian, t.z, state.adaptor)
 
     return transition, newstate
 end
@@ -459,9 +459,9 @@ function gen_metric(dim::Int, spl::Sampler{<:AdaptiveHamiltonian}, state)
     return AHMC.renew(state.hamiltonian.metric, AHMC.getM⁻¹(state.adaptor.pc))
 end
 
-gen_traj(alg::HMC, ϵ) = AHMC.StaticTrajectory(AHMC.Leapfrog(ϵ), alg.n_leapfrog)
-gen_traj(alg::HMCDA, ϵ) = AHMC.HMCDA(AHMC.Leapfrog(ϵ), alg.λ)
-gen_traj(alg::NUTS, ϵ) = AHMC.NUTS(AHMC.Leapfrog(ϵ), alg.max_depth, alg.Δ_max)
+make_ahmc_kernel(alg::HMC, ϵ) = AHMC.StaticTrajectory(AHMC.Leapfrog(ϵ), alg.n_leapfrog)
+make_ahmc_kernel(alg::HMCDA, ϵ) = AHMC.HMCDA(AHMC.Leapfrog(ϵ), alg.λ)
+make_ahmc_kernel(alg::NUTS, ϵ) = AHMC.NUTS(AHMC.Leapfrog(ϵ), alg.max_depth, alg.Δ_max)
 
 ####
 #### Compiler interface, i.e. tilde operators.
@@ -584,14 +584,14 @@ function HMCState(
         ϵ = spl.alg.ϵ
     end
 
-    # Generate a trajectory.
-    traj = gen_traj(spl.alg, ϵ)
+    # Generate a kernel.
+    kernel = make_ahmc_kernel(spl.alg, ϵ)
 
     # Generate a phasepoint. Replaced during sample_init!
     h, t = AHMC.sample_init(rng, h, θ_init) # this also ensure AHMC has the same dim as θ.
 
     # Unlink everything.
     invlink!(vi, spl)
 
-    return HMCState(vi, 0, 0, traj, h, AHMCAdaptor(spl.alg, metric; ϵ=ϵ), t.z)
+    return HMCState(vi, 0, 0, kernel.τ, h, AHMCAdaptor(spl.alg, metric; ϵ=ϵ), t.z)
 end

test/Project.toml

@@ -34,7 +34,7 @@ AdvancedPS = "0.2"
 AdvancedVI = "0.1"
 Clustering = "0.14"
 CmdStan = "6.0.8"
-Distributions = "0.23.8, 0.24, 0.25"
+Distributions = "< 0.25.11"
 DistributionsAD = "0.6.3"
 DynamicHMC = "2.1.6, 3.0"
 DynamicPPL = "0.12"

test/inference/Inference.jl

@@ -1,4 +1,4 @@
-@testset "io.jl" begin
+@testset "inference.jl" begin
     # Only test threading if 1.3+.
     if VERSION > v"1.2"
         @testset "threaded sampling" begin

test/inference/gibbs.jl

@@ -50,19 +50,10 @@
         chain = sample(gdemo(1.5, 2.0), alg, 5_000)
         check_numerical(chain, [:s, :m], [49/24, 7/6], atol=0.1)
 
-        setadsafe(true)
-
         Random.seed!(200)
         gibbs = Gibbs(PG(15, :z1, :z2, :z3, :z4), HMC(0.15, 3, :mu1, :mu2))
         chain = sample(MoGtest_default, gibbs, 5_000)
         check_MoGtest_default(chain, atol=0.15)
-
-        setadsafe(false)
-
-        Random.seed!(200)
-        gibbs = Gibbs(PG(15, :z1, :z2, :z3, :z4), ESS(:mu1), ESS(:mu2))
-        chain = sample(MoGtest_default, gibbs, 5_000)
-        check_MoGtest_default(chain, atol=0.1)
     end
 
     @turing_testset "transitions" begin

test/runtests.jl

@@ -46,22 +46,27 @@ include("test_utils/AllUtils.jl")
         include("core/ad.jl")
     end
 
+    @testset "samplers (without AD)" begin
+        include("inference/AdvancedSMC.jl")
+        include("inference/emcee.jl")
+        include("inference/ess.jl")
+        include("inference/is.jl")
+    end
+
     Turing.setrdcache(false)
     for adbackend in (:forwarddiff, :tracker, :reversediff)
         Turing.setadbackend(adbackend)
+        @info "Testing $(adbackend)"
+        start = time()
         @testset "inference: $adbackend" begin
             @testset "samplers" begin
                 include("inference/gibbs.jl")
                 include("inference/gibbs_conditional.jl")
                 include("inference/hmc.jl")
-                include("inference/is.jl")
-                include("inference/mh.jl")
-                include("inference/ess.jl")
-                include("inference/emcee.jl")
-                include("inference/AdvancedSMC.jl")
                 include("inference/Inference.jl")
                 include("contrib/inference/dynamichmc.jl")
                 include("contrib/inference/sghmc.jl")
+                include("inference/mh.jl")
             end
         end
 
@@ -72,6 +77,11 @@ include("test_utils/AllUtils.jl")
         @testset "modes" begin
             include("modes/ModeEstimation.jl")
         end
+
+        # Useful for
+        # a) discovering performance regressions,
+        # b) figuring out why CI is timing out.
+        @info "Tests for $(adbackend) took $(time() - start) seconds"
     end
     @testset "variational optimisers" begin
         include("variational/optimisers.jl")