KLMinRepGradDescent

This is a convenience constructor for ParamSpaceSGD with the RepGradELBO objective. This is equivalent to the algorithm commonly referred as automatic differentiation variational inference[KTRGB2017]. KLMinRepGradDescent is also an alias of ADVI .

AdvancedVI.KLMinRepGradDescent — Function
KLMinRepGradDescent(adtype; entropy, optimizer, n_samples, averager, operator)

KL divergence minimization by running stochastic gradient descent with the reparameterization gradient in the Euclidean space of variational parameters.

Arguments

  • adtype::ADTypes.AbstractADType: Automatic differentiation backend.

Keyword Arguments

  • entropy: Entropy gradient estimator to be used. Must be one of ClosedFormEntropy, StickingTheLandingEntropy, MonteCarloEntropy. (default: ClosedFormEntropy())
  • optimizer::Optimisers.AbstractRule: Optimization algorithm to be used. (default: DoWG())
  • n_samples::Int: Number of Monte Carlo samples to be used for estimating each gradient. (default: 1)
  • averager::AbstractAverager: Parameter averaging strategy.
  • operator::Union{<:IdentityOperator, <:ClipScale}: Operator to be applied after each gradient descent step. (default: ClipScale())

Requirements

  • The trainable parameters in the variational approximation are expected to be extractable through Optimisers.destructure. This requires the variational approximation to be marked as a functor through Functors.@functor.
  • The variational approximation $q_{\lambda}$ implements rand.
  • The target distribution and the variational approximation have the same support.
  • The target LogDensityProblems.logdensity(prob, x) must be differentiable with respect to x by the selected AD backend.
  • Additonal requirements on q may apply depending on the choice of entropy.
source
  • KTRGB2017Kucukelbir, A., Tran, D., Ranganath, R., Gelman, A., & Blei, D. M. (2017). Automatic differentiation variational inference. Journal of Machine Learning Research, 18(14), 1-45.