Reduce number of trees in RF regressors

[moralejo]

Reduced to 150 to 50 trees, physics performance is not impacted, while memory needs will decrease (and speed increase)

[morcuended]

I think it would be good to upload the comparison in the wiki for bookkeeping

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 73.50%. Comparing base (e4a5b7a) to head (f154d2e).

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #1294      +/-   ##
==========================================
+ Coverage   73.49%   73.50%   +0.01%     
==========================================
  Files         134      134              
  Lines       14210    14210              
==========================================
+ Hits        10443    10445       +2     
+ Misses       3767     3765       -2     

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

[moralejo]

Some comparisons here: RF_tree_reduction.pdf

[vuillaut]

why not changing the classifiers parameters as well?

[moralejo]

why not changing the classifiers parameters as well?

The tests for that (since they involve background, ideally from real data) are more complicated.
The main goal of this is to reduce the memory footprint of the RF training & application, and the biggest RFs are those for regression. Currently it is 67% for regression, 33% for classification (according to models' .sav files). With this change the total RF size will become ~55% of what it is now, with a share of 60% for classification, 40% for regression. If we aim for a further reduction we should indeed target the classifiers.

[morcuended]

there is something wrong with mamba setup step in the CI (is taking > 50 min)

[moralejo]

there is something wrong with mamba setup step in the CI (is taking > 50 min)

Can we ignore it?

[morcuended]

Tests will not run either until we merge #4563 or the problem with mamba is solved. Anyway, I think the changes here are not tested. I'd say it can be merged