Setting bounds on kernel length scales

[kaushikns]

I am struggling to find a way to set bounds on kernel length scales. Is there a recommended way to implement upper and lower bounds for kernel lengthscales, or could this functionality be considered for future releases?

[AdrianSosic]

Hi @kaushikns. Currently, we're not offering any interface for this, unfortunately. But it's clearly planned to make the model specification much more flexible in the future (also talking about active dimensions etc).

While I can't currently make promises about timelines regarding the interface of the BayBE Kernel objects, I see a possible shortcut how I might ship the functionality earlier via low-level input: one of my next workitems is to improve and generalize the interface of the GP model itself. Potentially, I can make it also accept gpytorch objects directly as a low-level alternative to our BayBE wrappers. This would come at the cost of reduced support for high-level convenience features like serialization etc, but would give you a way to manually construct any kernel you want. Would that help?

As I said, no promises at this point, but I'll definitely check if it's doable. If that's the case, I could ping you to test it before officially released, once I have a draft version.

[kaushikns]

Yes, that would be quite good to have :)

[AdrianSosic]

@kaushikns: I'm currently working on the support for GPyTorch objects. I think I might have a first draft available this week. My plan is to enable GPyTorch support for:

• Kernels
• Mean Functions
• Likelihoods

With that, users should have full (low-level) configurability of the GP model. Is there anything else you'd wish to have that is not on my radar yet?

[AdrianSosic]

@kaushikns: I have a dev version ready now. In addition to the above-mentioned points it also ships the possibility to conveniently specify active kernel dimensions via parameter names. Let me know if you are willing to beta-test it and I'll point you to the right place. Otherwise, we'll finish the regular review process and will probably include it in the next release.