search

cornellius-gp / gpytorch

A highly efficient implementation of Gaussian Processes in PyTorch
MIT License
3.53k stars 554 forks source link

PRs for custom kernels? #1750

Open jpchen opened 3 years ago

jpchen commented 3 years ago

We have a few additional custom kernels we've implemented for a project that I think would be better served upstream (eg white noise, changepoint, etc). Would you welcome PRs for these or would you prefer bespoke kernels to not clutter the kernels library?

Balandat commented 3 years ago

@jpchen I think generally the answer is yes, happy to consider if they are more generally useful. Can you point me to the implementations?

Re the white noise kernel, we used to have one but deprecated that in favor of having this be part of the likelihood - does this need to be a proper kernel for your use cases?

jpchen commented 2 years ago

They are internal atm, but I can open some PRs once I port tests over and defer to you on which are worth adding. We used a white noise kernel as a multiplicative component in a composite kernel.

Balandat commented 2 years ago

I can also look at the internal ones :)

jpchen commented 2 years ago

Will follow up over chat, probably won't get around to this in the next few weeks :)

GStechschulte commented 2 years ago

@jpchen Is a PR still in the works for the change point kernel? I have a project where this kernel could be of use and I will only develop myself if this PR won't happen in the foreseeable future. Thanks!

jpchen commented 2 years ago

Hi @GStechschulte, I have a version locally but it doesn't yet support batching (and thus probably won't pass the unit tests) and I havent had the bandwidth to do so. I'm happy to put up a PR you can add to or perhaps reference if that's helpful?

GStechschulte commented 2 years ago

Hi @GStechschulte, I have a version locally but it doesn't yet support batching (and thus probably won't pass the unit tests) and I havent had the bandwidth to do so. I'm happy to put up a PR you can add to or perhaps reference if that's helpful?

I appreciate the quick reply @jpchen. That would be great if you could put up a PR to which I can reference. Then, my plan would be to build off of your work and add batching. Thanks!