Regression model fails on Julia 1.6 beta1

[itsdfish]

Hi all-

I want to report a model that fails on Julia 1.6 beta1. I tried to run a multilevel model described here. I simplified the model into a linear regression, which runs on Julia 1.5.3

Model:

using Turing
using Distributions
import Random

@model regression(y, X) = begin

  sigma ~ LogNormal()

  beta ~ filldist(Normal(), 1)

  intercept ~ Normal()

  s2 ~ InverseGamma(3, 2)
  s = sqrt(s2)
  
  y .~ Normal.(intercept .+ X * beta, sigma)
end

# Generate data.
Random.seed!(0)
N = 30
X = randn(N, 1)
beta = 1.0
intercept = 1
sigma = .1
y = intercept .+ X * beta + randn(N) * sigma

# Sample via NUTS.
chain = sample(regression(y, X), NUTS(), 2000, progress=true)

Error:

ERROR: LoadError: MethodError: no method matching ~(::Float64, ::Normal{Float64})
Stacktrace:
  [1] _broadcast_getindex_evalf
    @ ./broadcast.jl:648 [inlined]
  [2] _broadcast_getindex
    @ ./broadcast.jl:621 [inlined]
  [3] getindex
    @ ./broadcast.jl:575 [inlined]
  [4] copy
    @ ./broadcast.jl:922 [inlined]
  [5] materialize
    @ ./broadcast.jl:883 [inlined]
  [6] #5
    @ ~/Desktop/temp.jl:89 [inlined]
  [7] (::var"#5#6")(_rng::Random._GLOBAL_RNG, _model::DynamicPPL.Model{var"#5#6", (:y, :X), (), (), Tuple{Matrix{Float64}, Matrix{Float64}}, Tuple{}}, _varinfo::DynamicPPL.ThreadSafeVarInfo{DynamicPPL.UntypedVarInfo{DynamicPPL.Metadata{Dict{DynamicPPL.VarName, Int64}, Vector{Distribution}, Vector{DynamicPPL.VarName}, Vector{Real}, Vector{Set{DynamicPPL.Selector}}}, Float64}, Vector{Base.RefValue{Float64}}}, _sampler::DynamicPPL.SampleFromUniform, _context::DynamicPPL.DefaultContext, y::Matrix{Float64}, X::Matrix{Float64})
    @ Main ./none:0
  [8] macro expansion
    @ ~/.julia/packages/DynamicPPL/u14IH/src/model.jl:0 [inlined]
  [9] _evaluate
    @ ~/.julia/packages/DynamicPPL/u14IH/src/model.jl:154 [inlined]
 [10] evaluate_threadsafe(rng::Random._GLOBAL_RNG, model::DynamicPPL.Model{var"#5#6", (:y, :X), (), (), Tuple{Matrix{Float64}, Matrix{Float64}}, Tuple{}}, varinfo::DynamicPPL.UntypedVarInfo{DynamicPPL.Metadata{Dict{DynamicPPL.VarName, Int64}, Vector{Distribution}, Vector{DynamicPPL.VarName}, Vector{Real}, Vector{Set{DynamicPPL.Selector}}}, Float64}, sampler::DynamicPPL.SampleFromUniform, context::DynamicPPL.DefaultContext)
    @ DynamicPPL ~/.julia/packages/DynamicPPL/u14IH/src/model.jl:144
 [11] Model
    @ ~/.julia/packages/DynamicPPL/u14IH/src/model.jl:94 [inlined]
 [12] VarInfo
    @ ~/.julia/packages/DynamicPPL/u14IH/src/varinfo.jl:126 [inlined]
 [13] VarInfo
    @ ~/.julia/packages/DynamicPPL/u14IH/src/varinfo.jl:125 [inlined]
 [14] step(rng::Random._GLOBAL_RNG, model::DynamicPPL.Model{var"#5#6", (:y, :X), (), (), Tuple{Matrix{Float64}, Matrix{Float64}}, Tuple{}}, spl::DynamicPPL.Sampler{NUTS{Turing.Core.ForwardDiffAD{40}, (), AdvancedHMC.DiagEuclideanMetric}}; resume_from::Nothing, kwargs::Base.Iterators.Pairs{Symbol, Int64, Tuple{Symbol}, NamedTuple{(:nadapts,), Tuple{Int64}}})
    @ DynamicPPL ~/.julia/packages/DynamicPPL/u14IH/src/sampler.jl:73
 [15] macro expansion
    @ ~/.julia/packages/AbstractMCMC/Nw3Wn/src/sample.jl:78 [inlined]
 [16] macro expansion
    @ ~/.julia/packages/ProgressLogging/BBN0b/src/ProgressLogging.jl:328 [inlined]
 [17] macro expansion
    @ ~/.julia/packages/AbstractMCMC/Nw3Wn/src/logging.jl:8 [inlined]
 [18] mcmcsample(rng::Random._GLOBAL_RNG, model::DynamicPPL.Model{var"#5#6", (:y, :X), (), (), Tuple{Matrix{Float64}, Matrix{Float64}}, Tuple{}}, sampler::DynamicPPL.Sampler{NUTS{Turing.Core.ForwardDiffAD{40}, (), AdvancedHMC.DiagEuclideanMetric}}, N::Int64; progress::Bool, progressname::String, callback::Nothing, discard_initial::Int64, thinning::Int64, chain_type::Type, kwargs::Base.Iterators.Pairs{Symbol, Int64, Tuple{Symbol}, NamedTuple{(:nadapts,), Tuple{Int64}}})
    @ AbstractMCMC ~/.julia/packages/AbstractMCMC/Nw3Wn/src/sample.jl:76
 [19] sample(rng::Random._GLOBAL_RNG, model::DynamicPPL.Model{var"#5#6", (:y, :X), (), (), Tuple{Matrix{Float64}, Matrix{Float64}}, Tuple{}}, sampler::DynamicPPL.Sampler{NUTS{Turing.Core.ForwardDiffAD{40}, (), AdvancedHMC.DiagEuclideanMetric}}, N::Int64; chain_type::Type, resume_from::Nothing, progress::Bool, nadapts::Int64, discard_adapt::Bool, discard_initial::Int64, kwargs::Base.Iterators.Pairs{Union{}, Union{}, Tuple{}, NamedTuple{(), Tuple{}}})
    @ Turing.Inference ~/.julia/packages/Turing/O1Pn0/src/inference/hmc.jl:140
 [20] #sample#2
    @ ~/.julia/packages/Turing/O1Pn0/src/inference/Inference.jl:143 [inlined]
 [21] #sample#1
    @ ~/.julia/packages/Turing/O1Pn0/src/inference/Inference.jl:133 [inlined]
 [22] top-level scope
    @ ./timing.jl:206 [inlined]
 [23] top-level scope
    @ ~/Desktop/temp.jl:0
in expression starting at /home/dfish/Desktop/temp.jl:102

Version:

Julia 1.6.0 beta1
Turing v0.15.8

[devmotion]

I assume it is caused by JuliaLang/julia#37583 and works if you replace

  y .~ Normal.(intercept .+ X * beta, sigma)

with e.g.

  y ~ arraydist(Normal.(intercept .+ X * beta, sigma))

[devmotion]

The version with arraydist seems to work for me. An alternative would be

    @. y ~ Normal(intercept + $(X * beta), sigma)

which seems to work as well.

[itsdfish]

Thank you for your solutions. I will leave this open in case the documentation requires an update.

[devmotion]

These are just temporary workarounds. AFAICT one just has to fix DynamicPPL.getargs_dottilde.

Edit: I fixed it locally, I'll open a PR later today.

[devmotion]

The issue is fixed by TuringLang/DynamicPPL.jl#212.

BTW note that Libtask_jll does not support Julia 1.6 yet and hence an older version of Libtask is installed which does not use the JLL package (but of course the older version of Libtask does not work with Julia 1.6 either). I noticed that also an older version of Turing is installed on Julia 1.6 (0.15.1 instead of 0.15.8).

[itsdfish]

Thanks for fixing that. You are correct. I was using .15.1 on Julia 6. When I realized the versions were different, I checked whether the example worked with Turing .15.1 on Julia 1.5.3. Sorry for the oversight.