Turing.jl icon indicating copy to clipboard operation
Turing.jl copied to clipboard

Published 20 hours ago • verified publisher icon TuringLang

Optimizing model with vector-valued observations produces error when data are a LinearAlgebra.Adjoint

Open ElOceanografo opened this issue 3 years ago • 2 comments

This confused me for a while. The following MWE, based on a more elaborate model I've been working on, runs fine sampling with NUTS:

using Turing, Optim
@model function test(x)
    n = length(x)
    σ ~ Exponential()
    x ~ MvNormal(zeros(n), σ)
end
x = randn(1, 10)
m = test(x')
c = sample(m, NUTS(), 100)

...but when trying to optimize it, it throws the following error:

julia> opt = optimize(m, MLE())
ERROR: MethodError: no method matching +(::ReverseDiff.TrackedReal{Float64,Float64,ReverseDiff.TrackedArray{Float64,Float64,1,Array{Float64,1},Array{Float64,1}}}, ::ReverseDiff.TrackedArray{Float64,Float64,1,Array{Float64,1},Array{Float64,1}})
For element-wise addition, use broadcasting with dot syntax: scalar .+ array
Closest candidates are:
  +(::Any, ::Any, ::Any, ::Any...) at operators.jl:538
  +(::ChainRulesCore.DoesNotExist, ::Any) at C:\Users\sam.urmy\.julia\packages\ChainRulesCore\l6bZW\src\differential_arithmetic.jl:23
  +(::ChainRulesCore.One, ::Any) at C:\Users\sam.urmy\.julia\packages\ChainRulesCore\l6bZW\src\differential_arithmetic.jl:94
  ...
Stacktrace:
 [1] acclogp!(::DynamicPPL.ThreadSafeVarInfo{DynamicPPL.VarInfo{NamedTuple{(:σ, :μ),Tuple{DynamicPPL.Metadata{Dict{DynamicPPL.VarName{:σ,Tuple{}},Int64},Array{Exponential{Float64},1},Array{DynamicPPL.VarName{:σ,Tuple{}},1},ReverseDiff.TrackedArray{Float64,Float64,1,Array{Float64,1},Array{Float64,1}},Array{Set{DynamicPPL.Selector},1}},DynamicPPL.Metadata{Dict{DynamicPPL.VarName{:μ,Tuple{}},Int64},Array{DistributionsAD.TuringScalMvNormal{Array{Float64,1},Float64},1},Array{DynamicPPL.VarName{:μ,Tuple{}},1},ReverseDiff.TrackedArray{Float64,Float64,1,Array{Float64,1},Array{Float64,1}},Array{Set{DynamicPPL.Selector},1}}}},ReverseDiff.TrackedReal{Float64,Float64,ReverseDiff.TrackedArray{Float64,Float64,1,Array{Float64,1},Array{Float64,1}}}},Array{Base.RefValue{ReverseDiff.TrackedReal{Float64,Float64,ReverseDiff.TrackedArray{Float64,Float64,1,Array{Float64,1},Array{Float64,1}}}},1}}, ::ReverseDiff.TrackedArray{Float64,Float64,1,Array{Float64,1},Array{Float64,1}}) at C:\Users\sam.urmy\.julia\packages\DynamicPPL\wcVCr\src\threadsafe.jl:19
 [2] tilde_observe(::Turing.OptimizationContext{DynamicPPL.LikelihoodContext{Nothing}}, ::DynamicPPL.SampleFromPrior, ::DistributionsAD.TuringScalMvNormal{ReverseDiff.TrackedArray{Float64,Float64,1,Array{Float64,1},Array{Float64,1}},ReverseDiff.TrackedReal{Float64,Float64,Nothing}}, ::LinearAlgebra.Adjoint{Float64,Array{Float64,2}}, ::DynamicPPL.VarName{:x,Tuple{}}, ::Tuple{}, ::DynamicPPL.ThreadSafeVarInfo{DynamicPPL.VarInfo{NamedTuple{(:σ, :μ),Tuple{DynamicPPL.Metadata{Dict{DynamicPPL.VarName{:σ,Tuple{}},Int64},Array{Exponential{Float64},1},Array{DynamicPPL.VarName{:σ,Tuple{}},1},ReverseDiff.TrackedArray{Float64,Float64,1,Array{Float64,1},Array{Float64,1}},Array{Set{DynamicPPL.Selector},1}},DynamicPPL.Metadata{Dict{DynamicPPL.VarName{:μ,Tuple{}},Int64},Array{DistributionsAD.TuringScalMvNormal{Array{Float64,1},Float64},1},Array{DynamicPPL.VarName{:μ,Tuple{}},1},ReverseDiff.TrackedArray{Float64,Float64,1,Array{Float64,1},Array{Float64,1}},Array{Set{DynamicPPL.Selector},1}}}},ReverseDiff.TrackedReal{Float64,Float64,ReverseDiff.TrackedArray{Float64,Float64,1,Array{Float64,1},Array{Float64,1}}}},Array{Base.RefValue{ReverseDiff.TrackedReal{Float64,Float64,ReverseDiff.TrackedArray{Float64,Float64,1,Array{Float64,1},Array{Float64,1}}}},1}}) at C:\Users\sam.urmy\.julia\packages\DynamicPPL\wcVCr\src\context_implementations.jl:90
 [3] (::var"#51#52")(::Random._GLOBAL_RNG, ::DynamicPPL.Model{var"#51#52",(:x,),(),(),Tuple{LinearAlgebra.Adjoint{Float64,Array{Float64,2}}},Tuple{}}, ::DynamicPPL.ThreadSafeVarInfo{DynamicPPL.VarInfo{NamedTuple{(:σ, :μ),Tuple{DynamicPPL.Metadata{Dict{DynamicPPL.VarName{:σ,Tuple{}},Int64},Array{Exponential{Float64},1},Array{DynamicPPL.VarName{:σ,Tuple{}},1},ReverseDiff.TrackedArray{Float64,Float64,1,Array{Float64,1},Array{Float64,1}},Array{Set{DynamicPPL.Selector},1}},DynamicPPL.Metadata{Dict{DynamicPPL.VarName{:μ,Tuple{}},Int64},Array{DistributionsAD.TuringScalMvNormal{Array{Float64,1},Float64},1},Array{DynamicPPL.VarName{:μ,Tuple{}},1},ReverseDiff.TrackedArray{Float64,Float64,1,Array{Float64,1},Array{Float64,1}},Array{Set{DynamicPPL.Selector},1}}}},ReverseDiff.TrackedReal{Float64,Float64,ReverseDiff.TrackedArray{Float64,Float64,1,Array{Float64,1},Array{Float64,1}}}},Array{Base.RefValue{ReverseDiff.TrackedReal{Float64,Float64,ReverseDiff.TrackedArray{Float64,Float64,1,Array{Float64,1},Array{Float64,1}}}},1}}, ::DynamicPPL.SampleFromPrior, ::Turing.OptimizationContext{DynamicPPL.LikelihoodContext{Nothing}}, ::LinearAlgebra.Adjoint{Float64,Array{Float64,2}}) at .\none:9
 [4] macro expansion at C:\Users\sam.urmy\.julia\packages\DynamicPPL\wcVCr\src\model.jl:0 [inlined]
 [5] _evaluate at C:\Users\sam.urmy\.julia\packages\DynamicPPL\wcVCr\src\model.jl:154 [inlined]
 [6] evaluate_threadsafe(::Random._GLOBAL_RNG, ::DynamicPPL.Model{var"#51#52",(:x,),(),(),Tuple{LinearAlgebra.Adjoint{Float64,Array{Float64,2}}},Tuple{}}, ::DynamicPPL.VarInfo{NamedTuple{(:σ, :μ),Tuple{DynamicPPL.Metadata{Dict{DynamicPPL.VarName{:σ,Tuple{}},Int64},Array{Exponential{Float64},1},Array{DynamicPPL.VarName{:σ,Tuple{}},1},ReverseDiff.TrackedArray{Float64,Float64,1,Array{Float64,1},Array{Float64,1}},Array{Set{DynamicPPL.Selector},1}},DynamicPPL.Metadata{Dict{DynamicPPL.VarName{:μ,Tuple{}},Int64},Array{DistributionsAD.TuringScalMvNormal{Array{Float64,1},Float64},1},Array{DynamicPPL.VarName{:μ,Tuple{}},1},ReverseDiff.TrackedArray{Float64,Float64,1,Array{Float64,1},Array{Float64,1}},Array{Set{DynamicPPL.Selector},1}}}},ReverseDiff.TrackedReal{Float64,Float64,ReverseDiff.TrackedArray{Float64,Float64,1,Array{Float64,1},Array{Float64,1}}}}, ::DynamicPPL.SampleFromPrior, ::Turing.OptimizationContext{DynamicPPL.LikelihoodContext{Nothing}}) at C:\Users\sam.urmy\.julia\packages\DynamicPPL\wcVCr\src\model.jl:144
 [7] Model at C:\Users\sam.urmy\.julia\packages\DynamicPPL\wcVCr\src\model.jl:94 [inlined]
 [8] Model at C:\Users\sam.urmy\.julia\packages\DynamicPPL\wcVCr\src\model.jl:98 [inlined]
 [9] (::Turing.Core.var"#f#24"{DynamicPPL.VarInfo{NamedTuple{(:σ, :μ),Tuple{DynamicPPL.Metadata{Dict{DynamicPPL.VarName{:σ,Tuple{}},Int64},Array{Exponential{Float64},1},Array{DynamicPPL.VarName{:σ,Tuple{}},1},Array{Float64,1},Array{Set{DynamicPPL.Selector},1}},DynamicPPL.Metadata{Dict{DynamicPPL.VarName{:μ,Tuple{}},Int64},Array{DistributionsAD.TuringScalMvNormal{Array{Float64,1},Float64},1},Array{DynamicPPL.VarName{:μ,Tuple{}},1},Array{Float64,1},Array{Set{DynamicPPL.Selector},1}}}},Float64},DynamicPPL.Model{var"#51#52",(:x,),(),(),Tuple{LinearAlgebra.Adjoint{Float64,Array{Float64,2}}},Tuple{}},DynamicPPL.SampleFromPrior,Turing.OptimizationContext{DynamicPPL.LikelihoodContext{Nothing}}})(::ReverseDiff.TrackedArray{Float64,Float64,1,Array{Float64,1},Array{Float64,1}}) at C:\Users\sam.urmy\.julia\packages\Turing\B6rlT\src\core\compat\reversediff.jl:30
 [10] ReverseDiff.GradientTape(::Turing.Core.var"#f#24"{DynamicPPL.VarInfo{NamedTuple{(:σ, :μ),Tuple{DynamicPPL.Metadata{Dict{DynamicPPL.VarName{:σ,Tuple{}},Int64},Array{Exponential{Float64},1},Array{DynamicPPL.VarName{:σ,Tuple{}},1},Array{Float64,1},Array{Set{DynamicPPL.Selector},1}},DynamicPPL.Metadata{Dict{DynamicPPL.VarName{:μ,Tuple{}},Int64},Array{DistributionsAD.TuringScalMvNormal{Array{Float64,1},Float64},1},Array{DynamicPPL.VarName{:μ,Tuple{}},1},Array{Float64,1},Array{Set{DynamicPPL.Selector},1}}}},Float64},DynamicPPL.Model{var"#51#52",(:x,),(),(),Tuple{LinearAlgebra.Adjoint{Float64,Array{Float64,2}}},Tuple{}},DynamicPPL.SampleFromPrior,Turing.OptimizationContext{DynamicPPL.LikelihoodContext{Nothing}}}, ::Array{Float64,1}, ::ReverseDiff.GradientConfig{ReverseDiff.TrackedArray{Float64,Float64,1,Array{Float64,1},Array{Float64,1}}}) at C:\Users\sam.urmy\.julia\packages\ReverseDiff\NoIPU\src\api\tape.jl:199
 [11] GradientTape at C:\Users\sam.urmy\.julia\packages\ReverseDiff\NoIPU\src\api\tape.jl:198 [inlined]
 [12] tape at C:\Users\sam.urmy\.julia\packages\Turing\B6rlT\src\core\compat\reversediff.jl:41 [inlined]
 [13] taperesult at C:\Users\sam.urmy\.julia\packages\Turing\B6rlT\src\core\compat\reversediff.jl:43 [inlined]
 [14] gradient_logp(::Turing.Core.ReverseDiffAD{false}, ::Array{Float64,1}, ::DynamicPPL.VarInfo{NamedTuple{(:σ, :μ),Tuple{DynamicPPL.Metadata{Dict{DynamicPPL.VarName{:σ,Tuple{}},Int64},Array{Exponential{Float64},1},Array{DynamicPPL.VarName{:σ,Tuple{}},1},Array{Float64,1},Array{Set{DynamicPPL.Selector},1}},DynamicPPL.Metadata{Dict{DynamicPPL.VarName{:μ,Tuple{}},Int64},Array{DistributionsAD.TuringScalMvNormal{Array{Float64,1},Float64},1},Array{DynamicPPL.VarName{:μ,Tuple{}},1},Array{Float64,1},Array{Set{DynamicPPL.Selector},1}}}},Float64}, ::DynamicPPL.Model{var"#51#52",(:x,),(),(),Tuple{LinearAlgebra.Adjoint{Float64,Array{Float64,2}}},Tuple{}}, ::DynamicPPL.SampleFromPrior, ::Turing.OptimizationContext{DynamicPPL.LikelihoodContext{Nothing}}) at C:\Users\sam.urmy\.julia\packages\Turing\B6rlT\src\core\compat\reversediff.jl:33
 [15] gradient_logp(::Array{Float64,1}, ::DynamicPPL.VarInfo{NamedTuple{(:σ, :μ),Tuple{DynamicPPL.Metadata{Dict{DynamicPPL.VarName{:σ,Tuple{}},Int64},Array{Exponential{Float64},1},Array{DynamicPPL.VarName{:σ,Tuple{}},1},Array{Float64,1},Array{Set{DynamicPPL.Selector},1}},DynamicPPL.Metadata{Dict{DynamicPPL.VarName{:μ,Tuple{}},Int64},Array{DistributionsAD.TuringScalMvNormal{Array{Float64,1},Float64},1},Array{DynamicPPL.VarName{:μ,Tuple{}},1},Array{Float64,1},Array{Set{DynamicPPL.Selector},1}}}},Float64}, ::DynamicPPL.Model{var"#51#52",(:x,),(),(),Tuple{LinearAlgebra.Adjoint{Float64,Array{Float64,2}}},Tuple{}}, ::DynamicPPL.SampleFromPrior, ::Turing.OptimizationContext{DynamicPPL.LikelihoodContext{Nothing}}) at C:\Users\sam.urmy\.julia\packages\Turing\B6rlT\src\core\ad.jl:83
 [16] (::Turing.OptimLogDensity{DynamicPPL.Model{var"#51#52",(:x,),(),(),Tuple{LinearAlgebra.Adjoint{Float64,Array{Float64,2}}},Tuple{}},Turing.OptimizationContext{DynamicPPL.LikelihoodContext{Nothing}},DynamicPPL.VarInfo{NamedTuple{(:σ, :μ),Tuple{DynamicPPL.Metadata{Dict{DynamicPPL.VarName{:σ,Tuple{}},Int64},Array{Exponential{Float64},1},Array{DynamicPPL.VarName{:σ,Tuple{}},1},Array{Float64,1},Array{Set{DynamicPPL.Selector},1}},DynamicPPL.Metadata{Dict{DynamicPPL.VarName{:μ,Tuple{}},Int64},Array{DistributionsAD.TuringScalMvNormal{Array{Float64,1},Float64},1},Array{DynamicPPL.VarName{:μ,Tuple{}},1},Array{Float64,1},Array{Set{DynamicPPL.Selector},1}}}},Float64}})(::Float64, ::Array{Float64,1}, ::Nothing, ::Array{Float64,1}) at C:\Users\sam.urmy\.julia\packages\Turing\B6rlT\src\modes\ModeEstimation.jl:160
 [17] (::NLSolversBase.var"#71#72"{NLSolversBase.InplaceObjective{Nothing,Nothing,Turing.OptimLogDensity{DynamicPPL.Model{var"#51#52",(:x,),(),(),Tuple{LinearAlgebra.Adjoint{Float64,Array{Float64,2}}},Tuple{}},Turing.OptimizationContext{DynamicPPL.LikelihoodContext{Nothing}},DynamicPPL.VarInfo{NamedTuple{(:σ, :μ),Tuple{DynamicPPL.Metadata{Dict{DynamicPPL.VarName{:σ,Tuple{}},Int64},Array{Exponential{Float64},1},Array{DynamicPPL.VarName{:σ,Tuple{}},1},Array{Float64,1},Array{Set{DynamicPPL.Selector},1}},DynamicPPL.Metadata{Dict{DynamicPPL.VarName{:μ,Tuple{}},Int64},Array{DistributionsAD.TuringScalMvNormal{Array{Float64,1},Float64},1},Array{DynamicPPL.VarName{:μ,Tuple{}},1},Array{Float64,1},Array{Set{DynamicPPL.Selector},1}}}},Float64}},Nothing,Nothing},Float64})(::Array{Float64,1}, ::Array{Float64,1}) at C:\Users\sam.urmy\.julia\packages\NLSolversBase\QPnui\src\objective_types\incomplete.jl:55
 [18] value_gradient!!(::OnceDifferentiable{Float64,Array{Float64,1},Array{Float64,1}}, ::Array{Float64,1}) at C:\Users\sam.urmy\.julia\packages\NLSolversBase\QPnui\src\interface.jl:82
 [19] initial_state(::LBFGS{Nothing,LineSearches.InitialStatic{Float64},LineSearches.HagerZhang{Float64,Base.RefValue{Bool}},Optim.var"#15#17"}, 
::Optim.Options{Float64,Nothing}, ::OnceDifferentiable{Float64,Array{Float64,1},Array{Float64,1}}, ::Array{Float64,1}) at C:\Users\sam.urmy\.julia\packages\Optim\auGGa\src\multivariate\solvers\first_order\l_bfgs.jl:165
 [20] optimize(::OnceDifferentiable{Float64,Array{Float64,1},Array{Float64,1}}, ::Array{Float64,1}, ::LBFGS{Nothing,LineSearches.InitialStatic{Float64},LineSearches.HagerZhang{Float64,Base.RefValue{Bool}},Optim.var"#15#17"}, ::Optim.Options{Float64,Nothing}) at C:\Users\sam.urmy\.julia\packages\Optim\auGGa\src\multivariate\optimize\optimize.jl:33
 [21] #optimize#85 at C:\Users\sam.urmy\.julia\packages\Optim\auGGa\src\multivariate\optimize\interface.jl:142 [inlined]
 [22] optimize at C:\Users\sam.urmy\.julia\packages\Optim\auGGa\src\multivariate\optimize\interface.jl:141 [inlined]
 [23] _optimize(::DynamicPPL.Model{var"#51#52",(:x,),(),(),Tuple{LinearAlgebra.Adjoint{Float64,Array{Float64,2}}},Tuple{}}, ::Turing.OptimLogDensity{DynamicPPL.Model{var"#51#52",(:x,),(),(),Tuple{LinearAlgebra.Adjoint{Float64,Array{Float64,2}}},Tuple{}},Turing.OptimizationContext{DynamicPPL.LikelihoodContext{Nothing}},DynamicPPL.VarInfo{NamedTuple{(:σ, :μ),Tuple{DynamicPPL.Metadata{Dict{DynamicPPL.VarName{:σ,Tuple{}},Int64},Array{Exponential{Float64},1},Array{DynamicPPL.VarName{:σ,Tuple{}},1},Array{Float64,1},Array{Set{DynamicPPL.Selector},1}},DynamicPPL.Metadata{Dict{DynamicPPL.VarName{:μ,Tuple{}},Int64},Array{DistributionsAD.TuringScalMvNormal{Array{Float64,1},Float64},1},Array{DynamicPPL.VarName{:μ,Tuple{}},1},Array{Float64,1},Array{Set{DynamicPPL.Selector},1}}}},Float64}}, ::Array{Float64,1}, ::LBFGS{Nothing,LineSearches.InitialStatic{Float64},LineSearches.HagerZhang{Float64,Base.RefValue{Bool}},Optim.var"#15#17"}, ::Optim.Options{Float64,Nothing}; kwargs::Base.Iterators.Pairs{Union{},Union{},Tuple{},NamedTuple{(),Tuple{}}}) at C:\Users\sam.urmy\.julia\packages\Turing\B6rlT\src\modes\ModeEstimation.jl:420
 [24] _optimize at C:\Users\sam.urmy\.julia\packages\Turing\B6rlT\src\modes\ModeEstimation.jl:411 [inlined] (repeats 2 times)
 [25] #_optimize#35 at C:\Users\sam.urmy\.julia\packages\Turing\B6rlT\src\modes\ModeEstimation.jl:387 [inlined]
 [26] _optimize(::DynamicPPL.Model{var"#51#52",(:x,),(),(),Tuple{LinearAlgebra.Adjoint{Float64,Array{Float64,2}}},Tuple{}}, ::Turing.OptimLogDensity{DynamicPPL.Model{var"#51#52",(:x,),(),(),Tuple{LinearAlgebra.Adjoint{Float64,Array{Float64,2}}},Tuple{}},Turing.OptimizationContext{DynamicPPL.LikelihoodContext{Nothing}},DynamicPPL.VarInfo{NamedTuple{(:σ, :μ),Tuple{DynamicPPL.Metadata{Dict{DynamicPPL.VarName{:σ,Tuple{}},Int64},Array{Exponential{Float64},1},Array{DynamicPPL.VarName{:σ,Tuple{}},1},Array{Float64,1},Array{Set{DynamicPPL.Selector},1}},DynamicPPL.Metadata{Dict{DynamicPPL.VarName{:μ,Tuple{}},Int64},Array{DistributionsAD.TuringScalMvNormal{Array{Float64,1},Float64},1},Array{DynamicPPL.VarName{:μ,Tuple{}},1},Array{Float64,1},Array{Set{DynamicPPL.Selector},1}}}},Float64}}, ::Optim.Options{Float64,Nothing}) at C:\Users\sam.urmy\.julia\packages\Turing\B6rlT\src\modes\ModeEstimation.jl:387
 [27] _mle_optimize(::DynamicPPL.Model{var"#51#52",(:x,),(),(),Tuple{LinearAlgebra.Adjoint{Float64,Array{Float64,2}}},Tuple{}}, ::Optim.Options{Float64,Nothing}; kwargs::Base.Iterators.Pairs{Union{},Union{},Tuple{},NamedTuple{(),Tuple{}}}) at C:\Users\sam.urmy\.julia\packages\Turing\B6rlT\src\modes\ModeEstimation.jl:316
 [28] _mle_optimize at C:\Users\sam.urmy\.julia\packages\Turing\B6rlT\src\modes\ModeEstimation.jl:315 [inlined]
 [29] #optimize#24 at C:\Users\sam.urmy\.julia\packages\Turing\B6rlT\src\modes\ModeEstimation.jl:295 [inlined]
 [30] optimize at C:\Users\sam.urmy\.julia\packages\Turing\B6rlT\src\modes\ModeEstimation.jl:295 [inlined] (repeats 2 times)
 [31] top-level scope at none:1

After much puzzlement, I figured out that instantiating the model with x' (i.e., a LinearAlgebra.Adjoint) was causing the problem. Changing the eighth line to

m = test(vec(x))

allows the optimization to work as expected. The error doesn't depend on AD or the optimization method; gradient-free optimizers fail also. It seems that optimization fails whenever x is a non-Vector, since test(collect(x')) also produces the same error.

ElOceanografo avatar Dec 02 '20 01:12 ElOceanografo

I dunno if this is a Turing error per se, does anyone with more experience on the AD side agree?

cpfiffer avatar Dec 02 '20 21:12 cpfiffer

I would guess (without having checked the details) that the problem are some of the tilde dispatches in src/modes or DynamicPPL. Both x' and collect(x') are matrices, and hence x ~ MvNormal(...) in the model is interpreted as multiple multivariate observations. The log pdf computation then results in a vector of values which can't be accumulated to varinfo.logp.

devmotion avatar Dec 02 '20 21:12 devmotion