Faster filldist() (#227)

* fix testset_zygote_broken()

define vars used by error()

* logpdf(arraydist): use mapreduce

* logpdf(filldist): use mapreduce

* remove filldist(Zygote) from broken

* improve mapreduce

Co-authored-by: David Widmann <[email protected]>

* improve mapreduce() invocation

Co-authored-by: David Widmann <[email protected]>

* tests: exclude Chernoff from Zygote filldist tests

* simplify mapreduce -> sum

Co-authored-by: David Widmann <[email protected]>

* explicitly broadcast

since it looks like `mapreduce()` still allocates

Co-authored-by: David Widmann <[email protected]>

* require ChainRulesTestUtils >= 1.9.2

some graident tests require test_approx(::Array{<:Array}, ::Zero)

* _flat_logpdf(): explicit lazy broadcasting

* filldist tests: enable Skellam

* use product_distribution() to fix deprecation

* eliminate unnecessary intermediate var

* replace some anonymous funcs with Base.Fix1

* replace sum(lambda, zip(...)) with lazy broadcast

* Update src/arraydist.jl

* Update test/ad/distributions.jl

Co-authored-by: David Widmann <[email protected]>
Co-authored-by: Hong Ge <[email protected]>

Author: 3 people

src/arraydist.jl

@@ -21,15 +21,14 @@ function arraydist(dists::AbstractMatrix{<:UnivariateDistribution})
     return MatrixOfUnivariate(dists)
 end
 function Distributions._logpdf(dist::MatrixOfUnivariate, x::AbstractMatrix{<:Real})
-    # return sum(((d, xi),) -> logpdf(d, xi), zip(dist.dists, x))
-    # Broadcasting here breaks Tracker for some reason
-    return sum(map(logpdf, dist.dists, x))
+    # Lazy broadcast to avoid allocations and use pairwise summation
+    return sum(Broadcast.instantiate(Broadcast.broadcasted(logpdf, dist.dists, x)))
 end
 function Distributions.logpdf(dist::MatrixOfUnivariate, x::AbstractArray{<:AbstractMatrix{<:Real}})
-    return map(x -> logpdf(dist, x), x)
+    return map(Base.Fix1(logpdf, dist), x)
 end
 function Distributions.logpdf(dist::MatrixOfUnivariate, x::AbstractArray{<:Matrix{<:Real}})
-    return map(x -> logpdf(dist, x), x)
+    return map(Base.Fix1(logpdf, dist), x)
 end
 
 function Distributions.rand(rng::Random.AbstractRNG, dist::MatrixOfUnivariate)
@@ -52,16 +51,16 @@ function arraydist(dists::AbstractVector{<:MultivariateDistribution})
 end
 
 function Distributions._logpdf(dist::VectorOfMultivariate, x::AbstractMatrix{<:Real})
-    return sum(((di, xi),) -> logpdf(di, xi), zip(dist.dists, eachcol(x)))
+    return sum(Broadcast.instantiate(Broadcast.broadcasted(logpdf, dist.dists, eachcol(x))))
 end
 function Distributions.logpdf(dist::VectorOfMultivariate, x::AbstractArray{<:AbstractMatrix{<:Real}})
-    return map(x -> logpdf(dist, x), x)
+    return map(Base.Fix1(logpdf, dist), x)
 end
 function Distributions.logpdf(dist::VectorOfMultivariate, x::AbstractArray{<:Matrix{<:Real}})
-    return map(x -> logpdf(dist, x), x)
+    return map(Base.Fix1(logpdf, dist), x)
 end
 
 function Distributions.rand(rng::Random.AbstractRNG, dist::VectorOfMultivariate)
     init = reshape(rand(rng, dist.dists[1]), :, 1)
-    return mapreduce(i -> rand(rng, dist.dists[i]), hcat, 2:length(dist); init = init)
+    return mapreduce(Base.Fix1(rand, rng), hcat, view(dist.dists, 2:length(dist)); init = init)
 end

src/filldist.jl

@@ -30,21 +30,19 @@ end
 function _flat_logpdf(dist, x)
     if toflatten(dist)
         f, args = flatten(dist)
-        return sum(f.(args..., x))
+        # Lazy broadcast to avoid allocations and use pairwise summation
+        return sum(Broadcast.instantiate(Broadcast.broadcasted(xi -> f(args..., xi), x)))
     else
-        return sum(map(x) do x
-            logpdf(dist, x)
-        end)
+        return sum(Broadcast.instantiate(Broadcast.broadcasted(Base.Fix1(logpdf, dist), x)))
     end
 end
 
 function _flat_logpdf_mat(dist, x)
     if toflatten(dist)
         f, args = flatten(dist)
-        return vec(sum(f.(args..., x), dims = 1))
+        return vec(mapreduce(xi -> f(args..., xi), +, x, dims = 1))
     else
-        temp = map(x -> logpdf(dist, x), x)
-        return vec(sum(temp, dims = 1))
+        return vec(mapreduce(Base.Fix1(logpdf, dist), +, x; dims = 1))
     end
 end
 

test/Project.toml

@@ -17,7 +17,7 @@ Zygote = "e88e6eb3-aa80-5325-afca-941959d7151f"
 
 [compat]
 ChainRulesCore = "1"
-ChainRulesTestUtils = "1"
+ChainRulesTestUtils = "1.9.2"
 Combinatorics = "1.0.2"
 Distributions = "0.25.15"
 FiniteDifferences = "0.11.3, 0.12"

test/ad/distributions.jl

@@ -408,9 +408,7 @@
             # PoissonBinomial fails with Zygote
             # Matrix case does not work with Skellam:
             # https://github.com/TuringLang/DistributionsAD.jl/pull/172#issuecomment-853721493
-            filldist_broken = if D <: Skellam
-                ((d.broken..., :Zygote, :ReverseDiff), (d.broken..., :Zygote, :ReverseDiff))
-            elseif D <: PoissonBinomial
+            filldist_broken = if D <: PoissonBinomial
                 ((d.broken..., :Zygote), (d.broken..., :Zygote))
             elseif D <: Chernoff
                 # Zygote is not broken with `filldist`

test/ad/utils.jl

@@ -396,12 +396,16 @@ function testset_zygote(distspec, unpack_x_θ, args...; kwargs...)
     end
 end
 
-function testset_zygote_broken(args...; kwargs...)
+function testset_zygote_broken(distspec, args...; kwargs...)
     # don't show test errors - tests are known to be broken :)
     testset = suppress_stdout() do
-        testset_zygote(args...; kwargs...)
+        testset_zygote(distspec, args...; kwargs...)
     end
 
+    f = distspec.f
+    θ = distspec.θ
+    x = distspec.x
+
     # change errors and fails to broken results, and count number of errors and fails
     efs = errors_to_broken!(testset)