SparseGaussianProcesses.jl
SparseGaussianProcesses.jl copied to clipboard
aterenin
Depreciation of CuArrays prevents installation
Hi,
I'd love to use this implementation for a project, but I'm having some trouble with installation. I believe it's because the old version of Flux uses CuArrays (instead of CUDA like the new versions) and so it can't be installed. Unfortunately I'm not sure I know how to fix the problem, sorry!
Small update, I have tried to install the package with a newer version of Flux (0.12). Because the new version of Flux doesn't use CuArrays, I have removed the parts of the code that deal with the GPU. The good news is the sine example does run, it just doesn't actually fit the data. About half the tests pass and half fail. Not sure how helpful any of this is!
Hi and thanks for your interest!
Unfortunately, this package should be viewed as mostly-deprecated: I originally built it as a technical demonstration of how pathwise conditioning might look from a package point of view, with the intent of developing it further, but ended up instead focusing on other research matters.
Most of the functionality within it is has now become available elsewhere in the Julia ecosystem.
Within the next few months, as part of an ongoing research project, I will probably simplify the package down to a more minimalist version which includes only the functionality of key interest and less fluff, which in turn will make it easier to keep maintained as Julia evolves.
It should not take much to fix the key issues and get the sine example working again. If you are interested in this, I am happy to take a PR with these fixes! 😊
Thanks for getting back to me, I totally understand your reasoning for not maintaining it.
I have been looking at the other GPs packages in the Julia ecosystem, but so far I couldn't find any that implement pathwise conditioning explicitly. Is there any chance you could point me in the right direction?
I very much a Julia beginner, so I can't immediately see what needs fixing, but I may well have a go in the future, or try to build the same functionality off one of the other GP packages. Cheers!