importance sampling

Project.toml

@@ -1,7 +1,7 @@
 name = "Soss"
 uuid = "8ce77f84-9b61-11e8-39ff-d17a774bf41c"
 author = ["Chad Scherrer <chad.scherrer@gmail.com>"]
-version = "0.15.3"
+version = "0.15.4"
 
 [deps]
 AdvancedHMC = "0bf59076-c3b1-5ca4-86bd-e02cd72cde3d"
@@ -20,6 +20,7 @@ LazyArrays = "5078a376-72f3-5289-bfd5-ec5146d43c02"
 LogDensityProblems = "6fdf6af0-433a-55f7-b3ed-c6c6e0b8df7c"
 MLStyle = "d8e11817-5142-5d16-987a-aa16d5891078"
 MacroTools = "1914dd2f-81c6-5fcd-8719-6d5c9610ff09"
+MappedArrays = "dbb5928d-eab1-5f90-85c2-b9b0edb7c900"
 MonteCarloMeasurements = "0987c9cc-fe09-11e8-30f0-b96dd679fdca"
 NamedTupleTools = "d9ec5142-1e00-5aa0-9d6a-321866360f50"
 Printf = "de0858da-6303-5e67-8744-51eddeeeb8d7"

src/Soss.jl

@@ -57,12 +57,10 @@ include("inference/rejection.jl")
 include("inference/dynamicHMC.jl")
 include("inference/advancedhmc.jl")
 
-
-# include("weighted.jl")
 #
 # # include("graph.jl")
 # # # include("optim.jl")
-# include("importance.jl")
+include("importance.jl")
 #
 # # # include("sobols.jl")
 # # # include("fromcube.jl")

src/core/utils.jl

@@ -226,7 +226,7 @@ end
 
 getntkeys(::NamedTuple{A,B}) where {A,B} = A
 getntkeys(::Type{NamedTuple{A,B}}) where {A,B} = A
-
+getntkeys(::Type{NamedTuple{A}}) where {A} = A
 
 # These macros quickly define additional methods for when you get tired of typing `NamedTuple()`
 macro tuple3args(f)

src/core/weighted.jl

@@ -6,6 +6,6 @@ end
 
 using Printf
 function Base.show(io::IO, ℓx::Weighted)
-    @printf io "Weighted(%g.4\n" (ℓx.ℓ)
-    println(",", ℓx.val)
+    @printf io "Weighted(%.4g" (ℓx.ℓ)
+    println(io, ", ", ℓx.val)
 end

src/importance.jl

@@ -1,6 +1,30 @@
 using Distributions
 using MonteCarloMeasurements
 
+"""
+    importanceSample(p(p_args), q(q_args), observed_data)
+
+Sample from `q`, and weight the result to behave as if the sample were taken from `p`. For example,
+
+```
+julia> p = @model begin
+    x ~ Normal()
+    y ~ Normal(x,1) |> iid(5)
+end;
+
+julia> q = @model μ,σ begin
+    x ~ Normal(μ,σ)
+end;
+
+julia> y = rand(p()).y;
+
+julia> importanceSample(p(),q(μ=0.0, σ=0.5), (y=y,))
+Weighted(-7.13971.4
+,(x = -0.12280566635062592,)
+````
+"""
+function importanceSample end
+
 export importanceSample
 @inline function importanceSample(p::JointDistribution, q::JointDistribution, _data)
     return _importanceSample(getmoduletypencoding(p.model), p.model, p.args, q.model, q.args, _data)
@@ -9,33 +33,39 @@ end
 @gg M function _importanceSample(_::Type{M}, p::Model, _pargs, q::Model, _qargs, _data) where M <: TypeLevel{Module}
     p = type2model(p)
     q = type2model(q)
-
+        
     Expr(:let,
         Expr(:(=), :M, from_type(M)),
-        sourceImportanceSample()(p,q) |> loadvals(_qargs, _data) |> loadvals(_pargs, NamedTuple()))
+        sourceImportanceSample(_data)(p,q) |> loadvals(_qargs, _data) |> loadvals(_pargs, NamedTuple()) |> merge_pqargs)
+
+
 end
 
-sourceImportanceSample(p::Model,q::Model) = sourceImportanceSample()(p::Model,q::Model)
+sourceImportanceSample(p::Model,q::Model,_data) = sourceImportanceSample(_data)(p::Model,q::Model)
 
 export sourceImportanceSample
-function sourceImportanceSample()
+function sourceImportanceSample(_data)
     function(p::Model,q::Model)
         p = canonical(p)
         q = canonical(q)
         m = merge(p,q)
 
+        _datakeys = getntkeys(_data)
+
         function proc(m, st::Sample) 
+            st.x ∈ _datakeys && return :(_ℓ += logpdf($(st.rhs), $(st.x)))
+
             if hasproperty(p.dists, st.x)
                 pdist = getproperty(p.dists, st.x)
-                qdist = st.rhs
+                qdist = getproperty(q.dists, st.x)
                 @gensym ℓx
                 result = @q begin
                     $ℓx = importanceSample($pdist, $qdist, _data)
                     _ℓ += $ℓx.ℓ
                     $(st.x) = $ℓx.val
                 end
                 return flatten(result)
-            else return :($(st.x) = $(st.rhs))
+            else return :($(st.x) = rand($(st.rhs)))
             end
             return flatten(result)
         end
@@ -45,11 +75,11 @@ function sourceImportanceSample()
 
         body = buildSource(m, proc) |> flatten
 
-        kwargs = freeVariables(q) ∪ arguments(p)
+        kwargs = arguments(p) ∪ arguments(q)
         kwargsExpr = Expr(:tuple,kwargs...)
 
         stochExpr = begin
-            vals = map(sampled(m)) do x Expr(:(=), x,x) end
+            vals = map(sampled(q)) do x Expr(:(=), x,x) end
             Expr(:tuple, vals...)
         end
 
@@ -64,7 +94,7 @@ function sourceImportanceSample()
 end
 
 @inline function importanceSample(p, q, _data)
-    x = merge(rand(q), _data)
+    x = rand(q)
     ℓ = logpdf(p,x) - logpdf(q,x)
     Weighted(ℓ,x)
 end
@@ -139,3 +169,10 @@ end
 #     ))
 
 # end
+
+function merge_pqargs(src)
+    @q begin
+        _args = merge(_pargs, _qargs)
+        $src
+    end |> flatten
+end

src/particles.jl

@@ -33,6 +33,7 @@ end
 
 export parts
 
+using MappedArrays, FillArrays
 
 # Just a little helper function for particles
 # https://github.com/baggepinnen/MonteCarloMeasurements.jl/issues/22
@@ -43,12 +44,9 @@ parts(x::Integer, N::Int=DEFAULT_SAMPLE_SIZE) = parts(float(x))
 parts(x::Real, N::Int=DEFAULT_SAMPLE_SIZE) = parts(repeat([x],N))
 parts(x::AbstractArray, N::Int=DEFAULT_SAMPLE_SIZE) = Particles(x)
 parts(p::Particles, N::Int=DEFAULT_SAMPLE_SIZE) = p
-parts(d::For, N::Int=DEFAULT_SAMPLE_SIZE) = parts.(d.f.(d.θ...), N)
-function parts(d::For{F,Tuple{I}}, N::Int=DEFAULT_SAMPLE_SIZE) where {F <: Function, I <: Integer}
-    parts.(d.f.(Base.OneTo.(d.θ)...), N)
-end
+parts(d::For, N::Int=DEFAULT_SAMPLE_SIZE) = parts.(mappedarray(d.f, CartesianIndices(d.θ)), N)
 
-parts(d::iid, N::Int=DEFAULT_SAMPLE_SIZE) = map(1:d.size) do j parts(d.dist, N) end
+parts(d::iid, N::Int=DEFAULT_SAMPLE_SIZE) = parts.(fill(d.dist, d.size))
 # size
 # dist