# Turing.jl: DomainError with -960.1841700397322:

**URL:** <https://discourse.julialang.org/t/turing-jl-domainerror-with-960-1841700397322/44829>\
**Category:** Machine Learning\
**Tags:** turing\
**Created:** [August 12, 2020, 9:50pm UTC](https://discourse.julialang.org/t/turing-jl-domainerror-with-960-1841700397322/44829 "2020-08-12T21:50:02Z")\
**Posts on this page:** 4\
**Page:** 1

<div class="post-metadata">

**Author:** ![vkv](https://avatars.discourse-cdn.com/v4/letter/v/d9b06d/32.png) [@vkv](https://discourse.julialang.org/u/vkv)\
**Post date:** [August 12, 2020, 9:50pm UTC](https://discourse.julialang.org/t/turing-jl-domainerror-with-960-1841700397322/44829/1 "2020-08-12T21:50:02Z")

</div>

Running the code results in the following error

```julia
using Turing, ReverseDiff, Memoization
Turing.setadbackend(:reversediff)
Turing.setrdcache(true)

y1 = rand(10)
A1 = rand(10, 10)

function logsumexp(mat) 
    max_ = maximum(mat, dims=1)
    exp_mat = exp.(mat .- max_) .- (mat .== max_)
    sum_exp_ = sum(exp_mat, dims=1)
    res = log1p.(sum_exp_) .+ max_
end

@model function mwe(y, A, ::Type{T} = Vector{Float64}) where {T}
    n,m = size(A)
    scale = rand(10)
    vec_raw = T(undef, n)
    for i in eachindex(vec_raw)
        vec_raw[i] ~ truncated(Laplace(0, 1), ((10-30)./scale[i]), ((100-30)./scale[i]))
    end
    vec = vec_raw .* scale .+ 30
    μ = logsumexp(A .- vec_raw)[1,:]
    y .~ Normal.(μ, 1)
    return vec
end

model = mwe(y1, A1);
chain = sample(model, NUTS(.65), 300);

```

Error:

```julia
DomainError with -9.892202109140928:
log will only return a complex result if called with a complex argument. Try log(Complex(x)).

```

Full Stacktrace: [DomainError with -9.892202109140928:log will only return a complex result if c - Pastebin.com](https://pastebin.com/GzfY3Qfu)

If I change

`vec_raw[i] ~ truncated(Laplace(0, 1), ((10-30)./scale[i]), ((100-30)./scale[i]))`

to

`vec_raw[i] ~ truncated(Laplace(0, 1), 5, 6) # any number for bounds are fine`

the model starts sampling but, this is not how I want to paremeterize the model.

How to fix this?

---

<div class="post-metadata">

**Author:** ![vkv](https://avatars.discourse-cdn.com/v4/letter/v/d9b06d/32.png) [@vkv](https://discourse.julialang.org/u/vkv)\
**Post date:** [August 13, 2020, 9:08am UTC](https://discourse.julialang.org/t/turing-jl-domainerror-with-960-1841700397322/44829/2 "2020-08-13T09:08:13Z")

</div>

The sampling works if I make `scale = ones(10) * any integer`. The error resumes if I change to `scale = rand(10:15, 10)`. This seems like a bug to me.

---

<div class="post-metadata">

**Author:** ![mohamed82008](https://sea2.discourse-cdn.com/julialang/user_avatar/discourse.julialang.org/mohamed82008/32/18171_2.png) [@mohamed82008](https://discourse.julialang.org/u/mohamed82008)\
**Post date:** [August 13, 2020, 9:13am UTC](https://discourse.julialang.org/t/turing-jl-domainerror-with-960-1841700397322/44829/3 "2020-08-13T09:13:39Z")

</div>

Just a guess but try without the caching. The Laplace distribution’s logpdf uses the absolute value so I suspect it is not compatible with ReverseDiff caching.

Edit: I tried it and it works. Use more adaptation steps though to ensure it finds a good starting point.

---

<div class="post-metadata">

**Author:** ![devmotion](https://sea2.discourse-cdn.com/julialang/user_avatar/discourse.julialang.org/devmotion/32/9201_2.png) [@devmotion](https://discourse.julialang.org/u/devmotion)\
**Post date:** [August 13, 2020, 9:46am UTC](https://discourse.julialang.org/t/turing-jl-domainerror-with-960-1841700397322/44829/4 "2020-08-13T09:46:58Z")

</div>

I think your implementation of `logsumexp` is incorrect (see my comment on Github [https://github.com/TuringLang/Turing.jl/issues/1374#issuecomment-673375285](https://github.com/TuringLang/Turing.jl/issues/1374#issuecomment-673375285)).

BTW I’m a bit confused where to reply, just posting here as well in case you missed my reply on Github.