Bijectors.ordered and MvLogNormal interaction .. only supported for unconstrained distributions.

[jae0]

Hi, Working through the pPCA Tutorial.

(https://turing.ml/v0.21/tutorials/11-probabilistic-pca)

Everything working but the last part,

Section: Rotation Invariant Householder Parameterization for Bayesian PCA

In the Turing model "pPCA_householder", the sampling fails with error:

ERROR: ArgumentError: ordered transform is currently only supported for unconstrained distributions.
Stacktrace:  
[1] ordered(d::MvLogNormal{Float64, PDMats.PDiagMat{Float64, Vector{Float64}}, FillArrays.Zeros{Float64, 1, Tuple{Base.OneTo{Int64}}}})
@ Bijectors ~/.julia/packages/Bijectors/QhObI/src/bijectors/ordered.jl:21

The key part seems to be:

sigma ~ Bijectors.ordered(MvLogNormal(MvNormal(ones(K))))

Maybe due to call to PDMats.PDiagMat? Any way around this? Thx!

Julia Version 1.9.4 (official release; 8e5136fa297)
[76274a88] + Bijectors v0.13.7
[fce5fe82] Turing v0.29.3
OS: Linux (x86_64-linux-gnu)

[torfjelde]

Ah, yes we need to update this; thank you!

As a heads up: the tutorials are only guaranteed to be runnable for the specific package versions specified in the tutorial itself, e.g. https://turing.ml/v0.21/tutorials/11-probabilistic-pca#appendix . But we're of course aiming to keep them up to do date, so will address this 👍