search

JuliaGaussianProcesses / ApproximateGPs.jl

Approximations for Gaussian processes: sparse variational inducing point approximations, Laplace approximation, ...
https://juliagaussianprocesses.github.io/ApproximateGPs.jl/dev
Other
35 stars 6 forks source link

`SparseVariationalApproximation` breaks with PDMats.jl v0.11.8 #131

Closed simsurace closed 2 years ago

simsurace commented 2 years ago

When running the example, the optimization does not work properly, i.e. the optimal covariance matrix of the sparse approximation remains diagonal.

simsurace commented 2 years ago

So I guess ApproximateGPs.jl depends on some internals of PDMats.jl then.

devmotion commented 2 years ago

Indeed, it seems SparseVariationalApproximation uses the unexported PDMats.chol_lower: https://github.com/JuliaGaussianProcesses/ApproximateGPs.jl/blob/dec97855a780cea0ede48fd85b5106722ab9f159/src/SparseVariationalApproximationModule.jl#L12

(BTW it also uses some other internals such as https://github.com/JuliaGaussianProcesses/ApproximateGPs.jl/blob/master/src/SparseVariationalApproximationModule.jl#L26)

st-- commented 2 years ago

(BTW it also uses some other internals such as https://github.com/JuliaGaussianProcesses/ApproximateGPs.jl/blob/master/src/SparseVariationalApproximationModule.jl#L26)

these are actually defined in https://github.com/JuliaGaussianProcesses/ApproximateGPs.jl/blob/master/src/utils.jl

devmotion commented 2 years ago

Fixed by #133.