Dev (#332)

* update Project.toml

* s/logpdf/logdensity/

* update deps

* scratchpad

* version bump

* representative => rootmeasure

* update dependency versions

* reduce dependencies

* update deps

* scratchpad

* representative => rootmeasure

* update dependency versions

* reduce dependencies

* `as` methods for `xform`

* cleanup

* require latest MeasureTheory

* dorp old distributions code

* drop old iid code

* drop extra space

* limit deps to three newest releases

* update dynamichmc

* add Aqua

* bump version

* Better `predict` method

* withmeasures(::ConditionalModel)

* update dependencies

* updating symbolics

* some updates to symbolics

* hmm example

* update MeasureBase bound

* better dispatch for `predict`

* drop redundant method

* remove whitespace

* bump version

* update for upcoming MeasureTheory release

* small change to toposort

* have `iid` use powermeasure

* tests passing!

* iid(n::Integer...)

* update dependencies

* start logdensity with partialstatic

* update test to account for partialstatic

* add `insupport`

* clean up insupport

* speed up model building

* updates

* starting updates for new MeasureTheory stuff

* get tests passing

* Kleisli => TransitionKernel

* bump MB dependency version

* Kleisli => TransitionKernel

* tests passing

* update dependencies

* update CI

* Should only need BayesianLinearRegression for testing

* bump version

.github/workflows/ci.yml

@@ -17,7 +17,8 @@ jobs:
       fail-fast: false
       matrix:
         version:
-          - '1.5'
+          - '1.6'
+          - '1.7'
           - '1'
         os:
           - ubuntu-latest

Project.toml

@@ -1,15 +1,16 @@
 name = "Soss"
 uuid = "8ce77f84-9b61-11e8-39ff-d17a774bf41c"
 author = ["Chad Scherrer <[email protected]>"]
-version = "0.20.9"
+version = "0.21.0"
 
 [deps]
 ArrayInterface = "4fba245c-0d91-5ea0-9b3e-6abc04ee57a9"
+DensityInterface = "b429d917-457f-4dbc-8f4c-0cc954292b1d"
 DiffResults = "163ba53b-c6d8-5494-b064-1a9d43ac40c5"
 Distributions = "31c24e10-a181-5473-b8eb-7969acd0382f"
 FillArrays = "1a297f60-69ca-5386-bcde-b61e274b549b"
 GeneralizedGenerated = "6b9d7cbe-bcb9-11e9-073f-15a7a543e2eb"
-IRTools = "7869d1d1-7146-5819-86e3-90919afe41df"
+IfElse = "615f187c-cbe4-4ef1-ba3b-2fcf58d6d173"
 JuliaVariables = "b14d175d-62b4-44ba-8fb7-3064adc8c3ec"
 LinearAlgebra = "37e2e46d-f89d-539d-b4ee-838fcccc9c8e"
 MLStyle = "d8e11817-5142-5d16-987a-aa16d5891078"
@@ -30,45 +31,46 @@ SimpleGraphs = "55797a34-41de-5266-9ec1-32ac4eb504d3"
 SimplePartitions = "ec83eff0-a5b5-5643-ae32-5cbf6eedec9d"
 SimplePosets = "b2aef97b-4721-5af9-b440-0bad754dc5ba"
 SpecialFunctions = "276daf66-3868-5448-9aa4-cd146d93841b"
+Static = "aedffcd0-7271-4cad-89d0-dc628f76c6d3"
 Statistics = "10745b16-79ce-11e8-11f9-7d13ad32a3b2"
 StatsBase = "2913bbd2-ae8a-5f71-8c99-4fb6c76f3a91"
 StatsFuns = "4c63d2b9-4356-54db-8cca-17b64c39e42c"
-SymbolicCodegen = "fc9b0551-4f1b-42e1-8440-e8a535f5a551"
 SymbolicUtils = "d1185830-fcd6-423d-90d6-eec64667417b"
 TransformVariables = "84d833dd-6860-57f9-a1a7-6da5db126cff"
 TupleVectors = "615932cf-77b6-4358-adcd-5b7eba981d7e"
 
 [compat]
-ArrayInterface = "3"
+ArrayInterface = "5,6"
+DensityInterface = "0.4"
 DiffResults = "1"
 Distributions = "0.23, 0.24, 0.25"
-FillArrays = "0.10, 0.11, 0.12"
+FillArrays = "0.11, 0.12, 0.13"
 GeneralizedGenerated = "0.3"
-IRTools = "0.4"
+IfElse = "0.1"
 JuliaVariables = "0.2"
 MLStyle = "0.3,0.4"
 MacroTools = "0.5"
 MappedArrays = "0.3, 0.4"
-MeasureBase = "0.5"
-MeasureTheory = "0.13"
-NamedTupleTools = "0.12, 0.13"
-NestedTuples = "0.3"
-RecipesBase = "0.7,0.8, 1"
+MeasureBase = "0.9"
+MeasureTheory = "0.16"
+NamedTupleTools = "0.12, 0.13, 0.14"
+NestedTuples = "0.3.9"
+RecipesBase = "1"
 Reexport = "1"
 Requires = "1"
 RuntimeGeneratedFunctions = "0.5"
 SampleChains = "0.5"
 SimpleGraphs = "0.5, 0.6, 0.7"
 SimplePartitions = "0.2, 0.3"
-SimplePosets = "0.0, 0.1"
-SpecialFunctions = "0.9, 0.10, 1"
+SimplePosets = "0.1"
+SpecialFunctions = "1, 2"
+Static = "0.5, 0.6"
 StatsBase = "0.33"
-StatsFuns = "0.9"
-SymbolicCodegen = "0.2"
-SymbolicUtils = "0.15, 0.16, 0.17"
-TransformVariables = "0.4"
+StatsFuns = "0.9, 1"
+SymbolicUtils = "0.17, 0.18, 0.19"
+TransformVariables = "0.5, 0.6"
 TupleVectors = "0.1"
-julia = "1.5"
+julia = "1.6"
 
 [extras]
 Aqua = "4c88cf16-eb10-579e-8560-4a9242c79595"

demos/boundmodels.jl

@@ -10,7 +10,7 @@ m(μ=1.0)
 
 rand(m(μ=1.0))
 
-logdensity(m(μ=1.0),(x=0.4,))
+logdensityof(m(μ=1.0),(x=0.4,))
 # rand(m1)
 
 weightedSample(m(μ=1.0),(x=3,))

demos/hmm.jl

@@ -53,7 +53,7 @@ end
     end
 end
 
-# function Distributions.logdensity(c::Chain, xs)
+# function Distributions.logdensityof(c::Chain, xs)
 
 c = Chain(s0, step)
 

demos/repeated-measurements.jmd

@@ -92,8 +92,8 @@ for t in symlogdensity(m()).evalf(3).args
 
 <!--
 using BenchmarkTools
-@btime logdensity(m(),truth)
-@btime logdensity(m(),truth, codegen)
+@btime logdensityof(m(),truth)
+@btime logdensityof(m(),truth, codegen)
 
 f1 = Soss._codegen(m, true);
 f2 = Soss._codegen(m,false);
@@ -105,7 +105,7 @@ codegen(m(),truth)
 
 
 
-logdensity(m(), merge(truth, (p_bad=shuffle(truth.p_bad),)), codegen)
+logdensityof(m(), merge(truth, (p_bad=shuffle(truth.p_bad),)), codegen)
 
 
 @time result = dynamicHMC(m(), (y=truth.y,), codegen) ;

demos/repeated-measurements.md

@@ -192,8 +192,8 @@ _j1 = 1 _j2 = 1
 
 <!--
 using BenchmarkTools
-@btime logdensity(m(),truth)
-@btime logdensity(m(),truth, codegen)
+@btime logdensityof(m(),truth)
+@btime logdensityof(m(),truth, codegen)
 
 f1 = Soss._codegen(m, true);
 f2 = Soss._codegen(m,false);
@@ -205,7 +205,7 @@ codegen(m(),truth)
 
 
 
-logdensity(m(), merge(truth, (p_bad=shuffle(truth.p_bad),)), codegen)
+logdensityof(m(), merge(truth, (p_bad=shuffle(truth.p_bad),)), codegen)
 
 
 @time result = dynamicHMC(m(), (y=truth.y,), codegen) ;

docs/src/misc.md

@@ -65,7 +65,7 @@ function sourceWeightedSample(_data)
         proc(_m, st :: LineNumber) = nothing
 
         function proc(_m, st :: Sample)
-            st.x ∈ _datakeys && return :(_ℓ += logdensity($(st.rhs), $(st.x)))
+            st.x ∈ _datakeys && return :(_ℓ += logdensity_def($(st.rhs), $(st.x)))
             return :($(st.x) = rand($(st.rhs)))
         end
 

scratchpad/MCM.jl

@@ -7,7 +7,7 @@ function expect(x,ℓ)
 end
 
 x = Particles(5000,TDist(3))
-ℓ = logpdf(Normal(),x) - logpdf(TDist(3), x)
+ℓ = logdensityof(Normal(),x) - logdensityof(TDist(3), x)
 
 expect(x,ℓ)
 expect(x^2,ℓ)
@@ -26,7 +26,7 @@ expect(z^4,ℓ)
 # And now a new weighted variable
 
 y = Particles(5000,TDist(3))
-ℓ += logpdf(Normal(),y) - logpdf(TDist(3), y)
+ℓ += logdensityof(Normal(),y) - logdensityof(TDist(3), y)
 
 expect(y,ℓ)
 expect(y^2,ℓ)

scratchpad/SampleChainsDynamicHMC-demo.jl

@@ -44,8 +44,8 @@ plot(p, exp.([ℓ((p=pj,)) for pj in p]))
 
 using TransformVariables
 
-ℓ(x) = logdensity(Beta(4,3), x.p)
-t = xform(post)
+ℓ(x) = logdensity_def(Beta(4,3), x.p)
+t = as(post)
 chain = initialize!(DynamicHMCChain, ℓ, t)
 drawsamples!(chain, 10000)
 plot(ash(chain.p))

scratchpad/advancedhmc.jl

@@ -6,7 +6,7 @@ D = 10; initial_θ = rand(D)
 dist = Normal() ^ D
 
 # Define the target distribution
-ℓπ(θ) = logdensity(dist, θ)
+ℓπ(θ) = logdensity_def(dist, θ)
 
 # Set the number of samples to draw and warmup iterations
 n_samples, n_adapts = 2_000, 1_000

scratchpad/bouncy.jl

@@ -17,9 +17,9 @@ Returns a `Trace` object.
 """
 function bouncy(m::ConditionalModel, T = 1000.0;  c=10.0, λref=0.1, ρ=0.0, adapt=false) where {A,B}
 
-    ℓ(pars) = logdensity(m, pars)
+    ℓ(pars) = logdensity_def(m, pars)
 
-    t = xform(m)
+    t = as(m)
 
     function f(x)
         (θ, logjac) = transform_and_logjac(t, x)

scratchpad/codegen.jl

@@ -8,7 +8,7 @@ m = @model σ begin
 
 x = rand(m(σ=3.0));
 
-sourceSymlogdensity(m(σ=3.0) | (;x))
+sourceSymlogdensityof(m(σ=3.0) | (;x))
 
 s = symlogdensity(m(σ=3.0))
 

scratchpad/composition.jl

@@ -33,17 +33,17 @@ dynamicHMC(post)
 
 
 
-logdensity(post, truth)
+logdensity_def(post, truth)
 
 speed = truth.speed
 x = truth.x
 v = truth.v
 
 _ℓ = 0.0
-_ℓ += logdensity(Normal(0.0, 100.0), speed)
+_ℓ += logdensity_def(Normal(0.0, 100.0), speed)
 speed = predict(Normal(0.0, 100.0), speed)
-_ℓ += logdensity(gps(; speed, n, t), x)
+_ℓ += logdensity_def(gps(; speed, n, t), x)
 x = predict(gps(; speed, n, t), x)
-_ℓ += logdensity(radar(; speed, n), v)
+_ℓ += logdensity_def(radar(; speed, n), v)
 v = predict(radar(; speed, n), v)
 _ℓ

scratchpad/composition2.jl

@@ -29,7 +29,7 @@ rand(m())
 
 x = randn(2)
 
-xform(m() | (x=x, μ = (z = 1.0,)))
+as(m() | (x=x, μ = (z = 1.0,)))
 
 dynamicHMC(m() | (x=x, μ = (z = 1.0,)))
 

scratchpad/example-linear-regression.jl

@@ -130,7 +130,7 @@ y_true - particles(y_ppc)
 
  # `rand`  turns each `v ~ dist` into `v = rand(dist)`, finally outputting the `NamedTuple` of all values it has seen.
 
-# `logdensity` steps through the same program, but instead accumulates a log-density. It begins by initializing `_ℓ = 0.0`. Then at each step, it turns `v ~ dist` into `_ℓ += logdensity(dist, v)`, before finally returning `_ℓ`.
+# `logdensity` steps through the same program, but instead accumulates a log-density. It begins by initializing `_ℓ = 0.0`. Then at each step, it turns `v ~ dist` into `_ℓ += logdensity_def(dist, v)`, before finally returning `_ℓ`.
 
 # Note that I said "turns into" instead of "interprets". Soss uses [`GG.jl`](https://github.com/thautwarm/GG.jl) to generate specialized code for a given model, inference primitive (like `rand` and `logdensity`), and type of data.
 

scratchpad/flux.jl

@@ -5,7 +5,7 @@
 #     t = getTransform(model)
 #     z = randn(t.dimension) |> Flux.param
 
-#     fpre = @eval $(logdensity(model))
+#     fpre = @eval $(logdensity_def(model))
 #     f(par, data) = Base.invokelatest(fpre, par, data)
 
 #     loss(data) = -f(transform(t, z), data)

scratchpad/gps.jl

@@ -8,7 +8,7 @@ y = sinpi.(x) + 0.3*randn(100)
 
 import TransformVariables
 const TV = TransformVariables 
-Soss.xform(gp::AbstractGPs.FiniteGP, _data=NamedTuple()) = TV.as(Array, TV.asℝ,size(gp)...)
+Soss.as(gp::AbstractGPs.FiniteGP, _data=NamedTuple()) = as(Array, asℝ,size(gp)...)
 
 sqexpkernel(alpha::Real, rho::Real) = 
 alpha^2 * transform(SEKernel(), 1/(rho*sqrt(2)))

scratchpad/hmm.jl

@@ -1,16 +1,41 @@
 using MeasureTheory
 using Base.Iterators
 using Statistics
-
+using Soss
 using Random
-rng = Random.Xoshiro(3)
 
-x = Chain(Normal()) do xj Normal(μ=xj) end
-xobs = rand(rng, x)
-y = For(xobs) do xj Poisson(logλ=xj) end
-yobs = rand(rng, y)
-xv = take(xobs, 10) |> collect
-yv = take(yobs, 10) |> collect
+# 5 7 10
+
+rng = Random.Xoshiro(12)
+
+m = @model begin
+    latent ~ Chain(Normal(μ=1)) do x Normal(μ=x, σ=0.2) end
+    observed ~ For(latent) do x Poisson(logλ=x) end
+end
+
+truth = rand(rng, m())
+
+xv = take(truth.latent, 100) |> collect
+yv = take(truth.observed, 100) |> collect
+
+
+using Plots
+
+plt = scatter(yv, label="observations")
+plot!(exp.(xv), lw=3, label="latent process")
+
+using Statistics
+
+scatter(exp.(xv), poiscdf.(exp.(xv), yv), label=false)
+
+x = Chain(Normal(μ=1)) do xj Normal(μ=xj, σ=0.2) end
+
+
+xvals = rand(rng, x)
+y = For(xobs) do xj Poisson(logλ= xj) end
+yvals = rand(rng, y)
+xv = take(xobs, 100) |> collect
+yv = take(yobs, 100) |> collect
 
 take(xobs.parent, 10) |> collect
 take(yobs.parent, 10) |> collect
@@ -19,7 +44,6 @@ take(yobs.parent, 10) |> collect
 exp.(xv)
 yv
 
-# using Plots
 
 # plt = scatter(normcdf.(xv, 1, yv), label=false)
 
@@ -44,5 +68,5 @@ truth = rand(rng, m())
 xobs = take(truth.x, 10) |> collect
 yobs = take(truth.y, 10) |> collect
 
-logdensity(m(), (x=xobs, y=yobs))
+logdensity_def(m(), (x=xobs, y=yobs))
 

scratchpad/repeated-measurements-example.jl

@@ -77,7 +77,7 @@ function elbo(x)
 
     qλ = q(r=5,s=14, λα=λα)
     r = rand(qλ)
-    logdensity(posterior, r) - logdensity(qλ, r)
+    logdensity_def(posterior, r) - logdensity_def(qλ, r)
 end
 
 elbo(x)
@@ -139,7 +139,7 @@ x0 = zeros(48)
 
 
 
-logdensity(p(r=5, s=14) | (;y), rand(q(r=5,s=14, λ=λ.λ)))
+logdensity_def(p(r=5, s=14) | (;y), rand(q(r=5,s=14, λ=λ.λ)))
 
 
 
@@ -202,7 +202,7 @@ logdensity(p(r=5, s=14) | (;y), rand(q(r=5,s=14, λ=λ.λ)))
 
 function elbo(p, args, obs, q, λ)
     qargs = merge(args, λ)
-    mean(logpdf(p(args), merge(particles(q(qargs)), obs))) + entropy(q(args))
+    mean(logdensityof(p(args), merge(particles(q(qargs)), obs))) + entropy(q(args))
 end
 
 
@@ -239,7 +239,7 @@ using AdvancedVI
 
 bad = rand(Bernoulli(0.5) |> iid(14))
 
-tr = xform(p(r=5,s=14), (y=y,));
+tr = as(p(r=5,s=14), (y=y,));
 d = tr.dimension
 
 getq(λ) = MvNormal(λ[1:d], exp.(λ[d .+ (1:d)]))
@@ -249,7 +249,7 @@ advi = ADVI(10, 10_000)
 
 
 function logπ(θ)
-    logpdf(p(r=5,s=14),  merge((y=y,), tr(θ)))
+    logdensityof(p(r=5,s=14),  merge((y=y,), tr(θ)))
 end
 
 logπ(randn(24))