Empty model in predict about soss.jl

[jtrakk]

Got it, thanks!

[cscherrer]

Can you give me code to replicate the error?

[jtrakk]

This reproduces ArgumentError: reducing over an empty collection is not allowed.

Is it clear what I'm doing wrong?

using Soss, SampleChainsDynamicHMC, TypedTables, Random, MeasureTheory

m = @model (d) begin
    β1 ~ Normal(μ=0., σ=0.05)
    rs = rows(d)
    y ~ let rs = rs
            For(rs) do r
                v = β1 * r.x1
                p = logistic(v)
                Bernoulli(p)
        end
    end
end

pred = predictive(m)
n = 100
tbl = Table(x1 = rand(1:10, n), res=rand(Bool, n))
condmod =  m(d=collect(tbl)) | (; y=tbl.res)
multichain = Soss.sample(Random.GLOBAL_RNG, condmod, dynamichmc(), 10, 1)
mapped = map(θ->pred(;d=tbl,), multichain);
simulate.(mapped)

[cscherrer]

julia> predictive(m)
@model begin
    end

Do you mean to say

julia> predictive(m, :β1)
@model (d, β1) begin
        rs = rows(d)
        y ~ let rs = rs
                For(rs) do r
                    v = β1 * r.x1
                    p = logistic(v)
                    Bernoulli(p)
                end
            end
    end

Then maybe

mapped = map(multichain) do sample
    rand(pred(d=d, β1=sample.β1))
end

[jtrakk]

Thanks, that fixes my issue for simulate and rand. Still getting that error for predict though, and the non-broadcasted version hangs:

using Soss, SampleChainsDynamicHMC, TypedTables, Random, MeasureTheory, DataFrames

m = @model (d) begin
    β1 ~ Normal(μ=0., σ=0.05)
    rs = rows(d)
    y ~  For(rs) do r
            v = β1 * r.x1
            p = logistic(v)
            Bernoulli(p)
    end
end

n = 100
tbl = DataFrame(x1 = rand(Float64.(1:10), n), res=rand(Bool, n))
condmod =  m(d=tbl) | (; y=tbl.res)
multichain = Soss.sample(Random.GLOBAL_RNG, condmod, dynamichmc(), 10, 1)
pred = predictive(m, :β1)
mapped = [pred(;d=tbl,β1=θ.β1) for θ in multichain];
simulate.(mapped)
rand.(mapped)
predict.(mapped) # ArgumentError: reducing over an empty collection is not allowed
predict(mapped) # hangs

[cscherrer]

It's the same problem. predictive makes a new model using the names of variables you'd like to turn into arguments. predict is basically predictive followed by rand. The code for it is pretty short, and may be the easiest way to see what it does:
https://github.com/cscherrer/Soss.jl/blob/master/src/transforms/predict.jl