MLJ.jl <> Turing.jl

Interface for Turing models with MLJ.

Usage

For example, to create a machine for a linear regression, use something like:

@model function linear_regression(X, y)
    [...]
end

model = linear_regression
n_samples = 100
sampler = NUTS()
tm = TuringModel(model, n_samples, sampler)
mach = machine(tm, X, y)

where X and y are your features and predictor variables respectively. Next, this machine can be used like normally in MLJ.

See the tests for more detailed examples.

Named data

Usually, MLJ interface models can take any object which satisfies the Tables.jl interface. Turing.jl does so too, but this will usually not work out when the Turing model uses broadcasting.

To work around this, this package allows defining a renamer::Dict which is a mapping. When this renamer is set, X is transformed to a Matrix before being passed to Turing. After fitting the model, the parameter names are updated based on the renames in renamer.

For example, when fitting a linear model with something like

@model function linear_regression(X::Matrix, y)
    σ₂ ~ truncated(Normal(0, 100), 0, Inf)

    intercept ~ Normal(0, 0.4)

    n_features = size(X, 2)
    coef ~ MvNormal(n_features, 0.4)

    mu = intercept .+ X * coef
    y ~ MvNormal(mu, sqrt(σ₂))
end;

then the output contains coef[1], coef[2], ..., coef[n].

To rename this to the keys of a namedtuple, use:

[...]
renamer = Dict(["coef[$i]" => "coef[$key]" for (i, key) in enumerate(keys(X))])
tm = TuringModel(model, n_samples, sampler; renamer)