Merge branch 'master' of https://github.com/FredericWantiez/AdvancedPS.jl into feature/traced_rng

Author: FredericWantiez

Project.toml

@@ -1,7 +1,7 @@
 name = "AdvancedPS"
 uuid = "576499cb-2369-40b2-a588-c64705576edc"
 authors = ["TuringLang"]
-version = "0.2.2"
+version = "0.2.4"
 
 [deps]
 AbstractMCMC = "80f14c24-f653-4e6a-9b94-39d6b0f70001"
@@ -13,6 +13,6 @@ StatsFuns = "4c63d2b9-4356-54db-8cca-17b64c39e42c"
 [compat]
 AbstractMCMC = "2, 3"
 Distributions = "0.23, 0.24, 0.25"
-Libtask = "0.5"
+Libtask = "0.5.3"
 StatsFuns = "0.9"
 julia = "1.3"

src/container.jl

@@ -330,11 +330,16 @@ The resampling steps use the given `resampler`.
 Del Moral, P., Doucet, A., & Jasra, A. (2006). Sequential monte carlo samplers.
 Journal of the Royal Statistical Society: Series B (Statistical Methodology), 68(3), 411-436.
 """
-function sweep!(rng::Random.AbstractRNG, pc::ParticleContainer, resampler)
+function sweep!(
+    rng::Random.AbstractRNG,
+    pc::ParticleContainer,
+    resampler,
+    ref::Union{Particle,Nothing}=nothing,
+)
     # Initial step:
 
     # Resample and propagate particles.
-    resample_propagate!(rng, pc, resampler)
+    resample_propagate!(rng, pc, resampler, ref)
 
     # Compute the current normalizing constant ``Z₀`` of the unnormalized logarithmic
     # weights.
@@ -355,7 +360,7 @@ function sweep!(rng::Random.AbstractRNG, pc::ParticleContainer, resampler)
     # For observations ``y₂, …, yₜ``:
     while !isdone
         # Resample and propagate particles.
-        resample_propagate!(rng, pc, resampler)
+        resample_propagate!(rng, pc, resampler, ref)
 
         # Compute the current normalizing constant ``Z₀`` of the unnormalized logarithmic
         # weights.

src/smc.jl

@@ -118,7 +118,9 @@ function AbstractMCMC.step(
     particles = ParticleContainer(x, rng)
 
     # Perform a particle sweep.
-    logevidence = sweep!(particles.rng, particles, sampler.resampler)
+    logevidence = sweep!(
+        particles.rng, particles, sampler.resampler, particles.vals[nparticles]
+    )
 
     # Pick a particle to be retained.
     newtrajectory = rand(rng, particles)

test/container.jl

@@ -51,6 +51,19 @@
         )
         @test AdvancedPS.logZ(pc) ≈ log(sum(exp, 2 .* logps))
 
+        # Resample and propagate particles with reference particle
+        particles_ref = [AdvancedPS.Trace(fpc(logp)) for logp in logps]
+        pc_ref = AdvancedPS.ParticleContainer(particles_ref)
+        AdvancedPS.resample_propagate!(
+            Random.GLOBAL_RNG, pc_ref, AdvancedPS.resample_systematic, particles_ref[end]
+        )
+        @test pc_ref.logWs == zeros(3)
+        @test AdvancedPS.getweights(pc_ref) == fill(1 / 3, 3)
+        @test all(AdvancedPS.getweight(pc_ref, i) == 1 / 3 for i in 1:3)
+        @test AdvancedPS.logZ(pc_ref) ≈ log(3)
+        @test AdvancedPS.effectiveSampleSize(pc_ref) == 3
+        @test pc_ref.vals[end] === particles_ref[end]
+
         # Resample and propagate particles.
         AdvancedPS.resample_propagate!(Random.GLOBAL_RNG, pc)
         @test pc.logWs == zeros(3)