[RFC] Changing time axis to be something else (iterations/comm. rounds/epochs) in benchopt reporting about benchopt

[mathurinm]

Hi @jeandut ! The WIP PR #479 aims at implementing this, your early feedback on it would help to merge it!

The link should point to Section E (I have fixed it, it will be updated in the final version), but it does not point to code so it would not have helped here.

[jeandut]

Note that apparently my question is already solved in one of benchopt's example as, citing Appendix C.3 of the article,:

C.3 Wall-clock time versus number of iterations
Measuring time or iteration are two alternatives that make sense in their respective contexts. Practitioners mostly care about the time it takes to solve their problem, while researchers in mathematical
optimization may want to abstract away the implementation and hardware details and only consider
iteration. The benchmarks we have presented showcase efficient implementations and are also
interested in hardware and implementation differences (e.g. CPU vs GPU solvers for ??, torch versus
tensorflow for Section F.4), hence our focus on time. However, Benchopt does not impose a choice
between the two measures: it is perfectly possible to create plots as a function of the number of
iterations as evidenced for example in ??.

However the hyperlink is broken (in the arXiv and OpenReview versions)

[jeandut]

I just did ; )

[jeandut]

#479 just got merged with the doc improvement this closes this issue !