Create summary table of all GP benchmarks

[luiarthur]

The GP models benchmark for a given sample size can be summarized in the table below, with the cells being compile times (if applicable) and execution times. Timings for forward passes and gradient computations of the models will also be computed if high-level API is available and exposed. The appropriate accompanying data visualizations (line plots, surface plots) will be included to make the results more interpretable. Links will be included to MWEs.

• [ ] Make a dynamic table which changes N. (e.g. N=50, 100, 200, 400, 800).
• [ ] Make interactive line plots where
  • x-axis is N, and y-axis is execution time or compile time
  • be able to select which implementations / PPL to show on the plot
• [ ] Demo how GPs can be integrated & used as a building block within a probabilistic model.

Model (PPL), sample size = N	Collapsed Gibbs	SMC	Particle Gibbs	HMC/NUTS (SB)	ADVI (SB)
GP (Turing)	[ ] TODO	[ ] TODO	[ ] TODO	[ ] TODO	[ ] TODO
GP (STAN)	[ ] TODO	[ ] TODO	[ ] TODO	[ ] TODO	[ ] TODO
GP (Nimble)	[ ] TODO	[ ] TODO	[ ] TODO	[ ] TODO	[ ] TODO
GP (Pyro)	[ ] TODO	[ ] TODO	[ ] TODO	[ ] TODO	[ ] TODO
GP (TFP)	NA	NA	NA	[ ] TODO	[ ] TODO
LVGP (Turing)	[ ] TODO	[ ] TODO	[ ] TODO	[ ] TODO	[ ] TODO
LVGP (STAN)	[ ] TODO	[ ] TODO	[ ] TODO	[ ] TODO	[ ] TODO
LVGP (Nimble)	[ ] TODO	[ ] TODO	[ ] TODO	[ ] TODO	[ ] TODO
LVGP (Pyro)	[ ] TODO	[ ] TODO	[ ] TODO	[ ] TODO	[ ] TODO
LVGP (TFP)	NA	NA	NA	[ ] TODO	[ ] TODO

Notes:

• Prefer doing the experiments by column as opposed to by row in this table. This way, if we are run out of time, we have benchmarks across PPLs, for a few implementations; and not benchmarks across a few implementations for one or two PPLs.
• LVGP refers to latent variable GP. GP refer to vanilla (full-rank) Gaussian process.