Initial kronecker work

[thomaspinder]

Pull request type

Please check the type of change your PR introduces:

• [ ] Bugfix
• [x] Feature
• [ ] Code style update (formatting, renaming)
• [ ] Refactoring (no functional changes, no api changes)
• [ ] Build related changes
• [ ] Documentation content changes
• [ ] Other (please describe):

What is the current behavior?

Issue Number: N/A

What is the new behavior?

Other information

[codecov[bot]]

Codecov Report

Merging #87 (429f594) into master (246addf) will decrease coverage by 2.66%. The diff coverage is 76.47%.

:exclamation: Current head 429f594 differs from pull request most recent head d330769. Consider uploading reports for the commit d330769 to get more accurate results

@@            Coverage Diff             @@
##           master      #87      +/-   ##
==========================================
- Coverage   98.95%   96.29%   -2.67%     
==========================================
  Files          13       13              
  Lines         961     1080     +119     
==========================================
+ Hits          951     1040      +89     
- Misses         10       40      +30     

Flag	Coverage Δ
unittests	96.29% <76.47%> (-2.67%)	:arrow_down:

Flags with carried forward coverage won't be shown. Click here to find out more.

Impacted Files	Coverage Δ
gpjax/kernels.py	86.79% <74.40%> (-11.90%)	:arrow_down:
gpjax/gps.py	100.00% <100.00%> (ø)
gpjax/variational_families.py	100.00% <100.00%> (ø)
gpjax/variational_inference.py	97.67% <100.00%> (+0.02%)	:arrow_up:
gpjax/types.py	100.00% <0.00%> (ø)
gpjax/utils.py	100.00% <0.00%> (ø)
gpjax/config.py	100.00% <0.00%> (ø)
gpjax/likelihoods.py	100.00% <0.00%> (ø)
... and 5 more

:mega: We’re building smart automated test selection to slash your CI/CD build times. Learn more

[thomaspinder]

Comment from Dan's push:

• Minimal Kronecker kernel. The inverse is highly inefficient - this can be improved!
• Added white noise kernel.
• Added diagonal computation abstraction.
• Added minimal computation abstraction for kernel combinations (i.e., for * and +).

Main issues:

• Though we can replace the computations in the gps.py, we might miss out with caching stuff e.g., the Cholesky decomposition in the dense computation case. Is there a sensible way around this?

• As a further point to the preceding point, we should not be computing the gram matrix in general (yes I known we need to in the dense setting).

• Think about computation savings for KL divergences - how can we resolve these?

Minor issues:

• Need to discuss the best way to add some computation methods - it might be tidier than just adding a load of static methods altogether.
• Need tests.

[daniel-dodd]

Closed PR - Kronecker kernels will be added from a different branch, and not kernel_compute.