Problem compiling an app with Turing

[joscani]

Hi. I try to compile an app to predict bayesian model . I saved posterior chain, and model specification. My precomp_file is something like that.

precomp_file.jl

using Turing
using Distributions
using LinearAlgebra
using DataFrames
using CSV
using StatsBase
using Tracker
using MKL
using Bijectors

import Logging
Logging.disable_logging(Logging.Warn)

# El objetivo es precompilar las funciones para
# que la app vaya rápido

## Leer cadena y especificación modelo
chain_path = "/home/jose/Julia_projects/bayesian_app/resources/cadena_refactor.jls"
# al hacer include dentro de la app mira por defecto en src
esp_modelo_path = "/home/jose/Julia_projects/bayesian_app/src/especificacion_modelo_refactor.jl"
df_path = "/home/jose/Julia_projects/bayesian_app/resources/Advertising.csv"
chain = read(chain_path, Chains)

df = CSV.read(df_path, DataFrame, ntasks=10)

include(esp_modelo_path)

## predecir 
predicciones = predict(model_turing(df ), chain)

pred_df = DataFrame(predicciones)
pred_df = pred_df[!, Not([:iteration, :chain])]
medias  = mean.(eachcol(pred_df))
mediana = quantile.(eachcol(pred_df), 0.5)
desv_ti = std.(eachcol(pred_df))
q01     = quantile.(eachcol(pred_df), 0.1)
q90     = quantile.(eachcol(pred_df), 0.9)

pred_interval = DataFrame(hcat(medias, mediana, desv_ti, q01, q90), :auto)

rename!(pred_interval,[:means,:median , :std, :q01, :q90])

CSV.write("data/predicciones.csv", pred_interval) 

where especificacion_modelo_refactor.jl is a julia script only with model turing specification

 @model function model_turing(df)
    TV    = df.TV
    radio = df.radio
    newspaper = df.newspaper
    # Prior coeficientes
    a ~ Normal(0, 0.5)
    tv_coef  ~ Normal(0, 0.5)
    radio_coef  ~ Normal(0, 0.5)
    newspaper_coef  ~ Normal(0, 0.5)
    σ₁ ~ Gamma(1, 1)

    mu = a .+ tv_coef .* TV .+ radio_coef .* radio .+ newspaper_coef .* newspaper 
    sales ~ MvNormal(mu, σ₁^2 * I)
end

My app file is something like that.

module bayesian_app

using Turing
using Distributions
using LinearAlgebra
using DataFrames
using CSV
using StatsBase
using Tracker
using MKL

function julia_main()::Cint
    try
        real_main()
    catch
        Base.invokelatest(Base.display_error, Base.catch_stack())
        return 1
    end
    return 0
end

function real_main()
    if length(ARGS) == 0
        error("pass arguments")
    end

# Read chain
    chain = read(ARGS[1], Chains)
# Read model_espec in julia script
   include(ARGS[2])    

# read data    
    df = CSV.read(ARGS[3], DataFrame, ntasks=10)
# Predict    
    predicciones = predict(model_turing(df), chain ) 

    pred_df = DataFrame(predicciones)
    pred_df = pred_df[!, Not([:iteration, :chain])]
    medias  = mean.(eachcol(pred_df))
    mediana = quantile.(eachcol(pred_df), 0.5)
    desv_ti = std.(eachcol(pred_df))
    q01     = quantile.(eachcol(pred_df), 0.1)
    q90     = quantile.(eachcol(pred_df), 0.9)

    pred_interval = DataFrame(hcat(medias, mediana, desv_ti, q01, q90), :auto)

    rename!(pred_interval,[:means,:median , :std, :q01, :q90])
    CSV.write("pred_interval_80.csv", pred_interval)
    CSV.write("pred_posterior.csv", pred_df)

end

end # module

And I try to compile using

create_app("../bayesian_app", "../bayesian_bin_predict", precompile_execution_file="../bayesian_app/src/precomp_file.jl", force=true, filter_stdlibs = false)

And get the compile, once I install some dpendencies like Tracker.jl an others.

I try to test my app passing a mcmc chain, a julia file where I wrote the model especification , but I get weird error message

../bayesian_bin_predict/bin/bayesian_app cadena_refactor.jls especificacion_modelo_refactor.jl Advertising.csv 
┌ Warning: Package Bijectors does not have ChainRulesCore in its dependencies:
│ - If you have Bijectors checked out for development and have
│   added ChainRulesCore as a dependency but haven't updated your primary
│   environment's manifest file, try `Pkg.resolve()`.
│ - Otherwise you may need to report an issue with Bijectors
└ Loading ChainRulesCore into Bijectors from project dependency, future warnings for Bijectors are suppressed.

┌ Warning: Error requiring `Tracker` from `Bijectors`

But when I see Manifest and Project I view Tracker.jl and Bijectors.

Any idea?

Thanks