vi(model, alg::VariationalInference)
vi(model, alg::VariationalInference, q::VariationalPosterior)
vi(model, alg::VariationalInference, getq::Function, θ::AbstractArray)

Constructs the variational posterior from the model and performs the optimization following the configuration of the given VariationalInference instance.

Arguments

• model: Turing.Model or Function z ↦ log p(x, z) where x denotes the observations
• alg: the VI algorithm used
• q: a VariationalPosterior for which it is assumed a specialized implementation of the variational objective used exists.
• getq: function taking parameters θ as input and returns a VariationalPosterior
• θ: only required if getq is used, in which case it is the initial parameters for the variational posterior

struct ADVI{AD} <: VariationalInference{AD}

Automatic Differentiation Variational Inference (ADVI) with automatic differentiation backend AD.

Fields

• samples_per_step::Int64: Number of samples used to estimate the ELBO in each optimization step.

• max_iters::Int64: Maximum number of gradient steps.

• adtype::Any: AD backend used for automatic differentiation.