Problems with indices when working with multidimensional data

[duodenum96]

Hello,

It seems that when a variable is indexed as x[i, j], ParetoSmooth.jl is unable to parse the indices. Here are three MWEs:

Turing.@model function model(data)
    μ ~ Normal()
    for i in 1:size(data, 1)
        for j in 1:size(data, 2)
            data[i, j] ~ Normal(μ, 1)
        end
    end
end

data = randn(10, 10)
chain = sample(model(data), NUTS(), 1000)
psis_loo(model(data), chain)

This returns the error:

`ERROR: ArgumentError: invalid base 10 digit ',' in "1, 2"`

Stacktrace:
  [1] tryparse_internal(::Type{Int64}, s::SubString{String}, startpos::Int64, endpos::Int64, base_::Int64, raise::Bool)
    @ Base .\parse.jl:143
  [2] parse(::Type{Int64}, s::SubString{String}; base::Nothing)
    @ Base .\parse.jl:254
  [3] parse
    @ .\parse.jl:253 [inlined]
  [4] (::ParetoSmoothDynamicPPLExt.var"#ind_from_string#2")(x::String)
    @ ParetoSmoothDynamicPPLExt C:\Users\yasir\.julia\packages\ParetoSmooth\ILEFB\ext\ParetoSmoothDynamicPPLExt.jl:40
  [5] lt
    @ .\ordering.jl:120 [inlined]
  [6] _issorted
    @ .\sort.jl:1430 [inlined]
  [7] _sort!(v::Vector{…}, a::Base.Sort.StableCheckSorted{…}, o::Base.Order.By{…}, kw::@NamedTuple{…})
    @ Base.Sort .\sort.jl:1352
  [8] _sort!
    @ .\sort.jl:770 [inlined]
  [9] _sort!
    @ .\sort.jl:721 [inlined]
 [10] _sort!
    @ .\sort.jl:792 [inlined]
 [11] _sort!
    @ .\sort.jl:737 [inlined]
 [12] _sort!
    @ .\sort.jl:676 [inlined]
 [13] _sort!
    @ .\sort.jl:554 [inlined]
 [14] #sort!#32
    @ .\sort.jl:1694 [inlined]
 [15] sort!
    @ .\sort.jl:1687 [inlined]
 [16] #sort#33
    @ .\sort.jl:1720 [inlined]
 [17] pointwise_log_likelihoods(model::DynamicPPL.Model{…}, chains::Chains{…})
    @ ParetoSmoothDynamicPPLExt C:\Users\yasir\.julia\packages\ParetoSmooth\ILEFB\ext\ParetoSmoothDynamicPPLExt.jl:42
 [18] psis_loo(::DynamicPPL.Model{…}, ::Chains{…}; kwargs::@Kwargs{})
    @ ParetoSmoothDynamicPPLExt C:\Users\yasir\.julia\packages\ParetoSmooth\ILEFB\ext\ParetoSmoothDynamicPPLExt.jl:65
 [19] psis_loo(::DynamicPPL.Model{…}, ::Chains{…})
    @ ParetoSmoothDynamicPPLExt C:\Users\yasir\.julia\packages\ParetoSmooth\ILEFB\ext\ParetoSmoothDynamicPPLExt.jl:64
 [20] top-level scope
    @ c:\Users\yasir\Desktop\brain_stuff\mice_timescale_variability2\scripts\figure3_turing_proto.jl:61
Some type information was truncated. Use `show(err)` to see complete types.

Second example:

Turing.@model function model(data)
    μ ~ Normal()
    for i in 1:size(data, 1)
        data[i, :] ~ MvNormal(ones(size(data, 2)) * μ, 1.0 * I)
    end
end

data = randn(10, 10)
chain = sample(model(data), NUTS(), 1000)
psis_loo(model(data), chain)

Fails with the same error message.

Third example:

Turing.@model function model(data)
    μ ~ Normal()
    for i in 1:size(data, 1)
        data[i, :] .~ Normal(μ, 1)
    end
end

data = randn(10, 10)
chain = sample(model(data), NUTS(), 1000)
psis_loo(model(data), chain)

This one proceeds successfully.

[ParadaCarleton]

Generally, my recommendation is to go to Arviz.jl for LOO-related inference, as it provides more features. Otherwise, the issue here is closely related to #67: the problem is if we're working with a matrix of observations (instead of a vector), it's unclear what the individual (independent) observations are (is each row an observation, or each column)?

The error message should definitely be better-written, though. PRs welcome!