Turkie.jl

A Julia package for visualizing dynamically sampling and statistics of Bayesian models ***

Installation

Turkie is a registered package and is symply installed by running

pkg> add Turkie

Basic example with Turing

Right now Turkie only works with Turing.jl but it should be compatible with any sampling algorithm following the AbstractMCMC.jl interface

Here is a simple example to start right away :

using Turing
using Turkie
using Makie # You could also use CairoMakie or another backend
@model function demo(x) # Some random Turing model
    m0 ~ Normal(0, 2)
    s ~ InverseGamma(2, 3)
    m ~ Normal(m0, √s)
    for i in eachindex(x)
        x[i] ~ Normal(m, √s)
    end
end

xs = randn(100) .+ 1 # Create some random data
m = demo(xs) # Create the model
cb = TurkieCallback(m) # Create a callback function to be given to the sample
chain = sample(m, NUTS(0.65), 300; callback = cb) # Sample and plot at the same time

You should observe something like this :

The example in details

If we look at the last 2 steps:

c = TurkieCallback(m)

will create a Makie window and store the names of all the (non-observed) variables of m. By default the quantities looked at for each of the variables are :

• A combination of an histogram and a KDE approximation
• The sample mean
• The sample variance
• The auto-covariance of the samples

Next :

chain = sample(m, NUTS(0.65), 300; callback = cb) 

This is your typical posterior sampling with Turing.jl. While sampling the callback object cb will be called and the statistics will be updated live.

Tuning the quantities

Of course the default is not always desirable. You can chose what variables and what quantities are shown by giving a NamedTuple to TurkieCallback instead of a model. For example,

cb = TurkieCallback(
                (v = [:trace, :mean],
                s = [:autocov, :var])
            )

will only show the trace and the sample mean of v and the auto-covariance and variance of s. Pairs should be of the type {Symbol,AbstractVector}. In these vectors you can either throw a symbol from the following list:

• :mean : The sample mean of the variable
• :var : The sample variance of the variable
• :trace : The trace of the variable (every value)
• :hist : An histogram of the variable
• :kde : A KDE estimation of the variable
• :histkde : A KDE estimation combined with an histogram
• :autocov : The sample auto-covariance

You can also pass an OnlineStat object from OnlineStats.jl. By default, it will plot the value of the OnlineStat at every iteration as a trace. If you want a specific implementation of a certain stat please open an issue.

Recording the sampling

If you want to make a cool animation you can use the built-in recording features of Makie. Here is the simple example, using the Turing example from above:

record(cb.figure, joinpath(@__DIR__, "video.webm")) do io
    addIO!(cb, io)
    sample(m,  NUTS(0.65), 300; callback = cb)
end