[FEATURE REQUEST]: Allow custom kernels in `SingleTaskMultiFidelityGP`

[TobyBoyne]

Motivation

I have a BO problem that is multifidelity, and requires the use of a custom kernel. I therefore cannot easily use a SingleTaskMultiFidelityGP, since it only supports using the default get_covar_module_with_dim_scaled_prior for the non-fidelity dims.

Describe the solution you'd like to see implemented in BoTorch.

I would like to add a cont_kernel_factory argument to SingleTaskMultiFidelityGP, similar to the same argument in SingleTaskMixedGP.

As an aside, I also wonder if we should remove the ScaleKernel, since this has also been removed across most other models? Although maybe it's required since the DownsamplingKernel is not bounded to [0, 1] compared to, for example, the RBFKernel.

Describe any alternatives you've considered to the above solution.

One could simply construct the kernel first, and pass it as the covar_module argument into SingleTaskGP. This works, but would require duplicating a lot of the logic in _setup_multifidelity_covar_module, as well as missing the setup logic in SingleTaskMultiFidelityGP.__init__

Is this related to an existing issue in BoTorch or another repository? If so please include links to those Issues here.

No response

Pull Request

Yes

Code of Conduct

• I agree to follow BoTorch's Code of Conduct