AbstractGPs.jl

AbstractGPs.jl copied to clipboard

Use non scalar method to compute logdet for cholesky and relax compat for PDMat

6

This enables the computation of `dtc` on GPU

sharanry

CompatHelper: bump compat for Soss to 0.21 for package ppl, (keep existing compat)

1

This pull request changes the compat entry for the `Soss` package from `0.19, 0.20` to `0.19, 0.20, 0.21` for package ppl. This keeps the compat entries for earlier versions. Note:...

github-actions[bot]

We should probably have some cross package tests with this SciML package?

theogf

Documentation build is ~ 55 min

3

That's slightly long! Maybe we can figure out which example is taking so much compute and reduce it somehow!

theogf

Maybe I am missing the point but it feels like there is a discrepancy between `LatentGP` and `AbstractGP` although they kinda behave the same. I don't understand why the noise...

theogf

VFE/DTC should be moved to ApproximateGPs.jl

4

theogf

The internal implementation of VFE/DTC are non-standard and need references to their derivations. Related literature: - VFE: M. K. Titsias. "Variational learning of inducing variables in sparse Gaussian processes". In:...

sharanry

Move AbstractGPs related test from GPLikelihoods here

9

I removed the tests using AbstactGPs from GPLikelihoods and saved them in [this gist](https://gist.github.com/theogf/ef57956dd052c39967c2697ec6214e09). They should probably be added here (or ApproximateGPs) to check compatibility.

theogf

**BREAKING** improve structs & constructors

9

Happy for others to contribute to this branch. Things that came to my mind: e.g. `FiniteGP` should check whether `length(x) == size(Sigma, 1) == size(Sigma, 2)`.

st--

Test VFE with the naive implementation

6

**Summary** We currently lack any test to confirm the predictive distribution is matching what is prescribed by the original papers - - VFE: M. K. Titsias. "Variational learning of inducing...

sharanry