Maximum likelihood for Rainbow

[karpov-sv]

Switch Rainbow to maximum likelihood based cost function. It allows to correctly implement upper limits (using Tobit model for censored data) and additional barrier terms that would prevent the parameters from wandering too close to their limits (as then Hessian estimation would be wrong, giving extremely large errors, due to internal non-linear transformations of the parameters inside iminuit).
Improve initial values estimation and make them a bit more error-prone.
Optimize some computations inside model function (zeros_like is much slower than plain zeros)
Also, slightly improve code readability by generalizing parameter unscaling. It allows to correctly return the covariance matrix.

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 82.41%. Comparing base (c504d8c) to head (6250e66).
Report is 9 commits behind head on master.

Additional details and impacted files

@@            Coverage Diff             @@
##           master     #407      +/-   ##
==========================================
- Coverage   82.45%   82.41%   -0.04%     
==========================================
  Files           9        9              
  Lines        2570     2565       -5     
==========================================
- Hits         2119     2114       -5     
  Misses        451      451              

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[codspeed-hq]

CodSpeed Performance Report

Merging #407 will not alter performance

_{Comparing karpov-sv:rainbow_ml (6250e66) with master (c504d8c)}

Summary

✅ 99 untouched benchmarks

[hombit]

Thank you, @karpov-sv, it looks really nice. I'd like @erusseil to review it before the merge.

@karpov-sv, do you think it may go to a patch (e.g. 0.9.4) release?

[karpov-sv]

@karpov-sv, do you think it may go to a patch (e.g. 0.9.4) release?

Yes, it would be ok, as the code is declared experimental anyway. The changes from the point of view of "primary" API should be minimal (just a bit more stable).

[erusseil]

Thanks a lot Sergey for the work, it looks promising! I made some tests and everything appears very robust to me.

[hombit]

@karpov-sv @erusseil v0.9.4 has been released, and you both are now listed as co-authors of the package (as I should do earlier). Thank you!