Skip to content

TDist - cdf and quantile function - Auto-Differentiability #152

Description

@paschermayr

Hi there,

Thank you for all your work! I have seen that recently, pull request #147 was closed in favor of #149.

I believe the former issue tried to make the cdf and quantile function of the TDistribution auto-differentiable, but the latter closed it and did not alleviate this issue. MWE from a fresh project with all up-to-date libraries:

using Distributions, DistributionsAD, StatsBase
using ForwardDiff, ReverseDiff
using StatsFuns

function mytargetfunction(data::AbstractVector)
    function obtaingradient::AbstractVector{R}) where {R<:Real}
        nu = θ[1]
        distr = TDist(nu)
        data_uniform = [cdf(distr, data[iter]) for iter in eachindex(data)]
        data_real = [quantile(distr, data_uniform[iter]) for iter in eachindex(data_uniform)]
        return sum( logpdf(distr, data_real[iter]) for iter in eachindex(data_real) )
    end
end

#working
ν = [3.0]
data = randn(1000)
target = mytargetfunction(data)
target(ν)
#not working
ForwardDiff.gradient(target, ν) #MethodError: no method matching _beta_inc(::ForwardDiff.Dual
ReverseDiff.gradient(target, ν) #MethodError: no method matching _beta_inc(::ReverseDiff.TrackedReal

It seems like the beta_inc function is from the Specialfunctions.jl package and requires Float64 as arguments instead of just reals. Is there a reason for that? I believe I should probably open an issue there as well?

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions