Add weighting function for several scenarios

[veni-vidi-vici-dormivi]

Add a function go generate weights such that each scenario contributes equally to some fitting routine. This means that a member of a scenario with a lot of members weighs less than a member of a scenario with few members.

• Closes get_scenario_weights for new data structures #356
• Tests added
• Fully documented, including CHANGELOG.rst

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 78.10%. Comparing base (456776d) to head (693562e).
Report is 2 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #567      +/-   ##
==========================================
+ Coverage   77.93%   78.10%   +0.17%     
==========================================
  Files          49       49              
  Lines        2986     3010      +24     
==========================================
+ Hits         2327     2351      +24     
  Misses        659      659              

Flag	Coverage Δ
unittests	78.10% <100.00%> (+0.17%)	⬆️

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

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

[mathause]

Oh - I just realized that the text and code of Lea's weight generation does not align:

mesmer/mesmer/calibrate_mesmer/train_utils.py

Lines 14 to 15 in 456776d

	derive scenario weights such that each has equal weight, i.e., 1 / number of samples
	(= nr_runs * nr_ts)

mesmer/mesmer/calibrate_mesmer/train_utils.py

Line 39 in 456776d

weights.append(np.full(nr_samples_scen, 1 / nr_runs))

Let me check if I introduced this error or if this was always the case. -> See #569

---

Given your tests: should exclude be all dims that are not in ens_dim (i.e. remove it from the input params)? I am a bit afraid that the user will forget to add the correct exclude. Does it actually work if you don't exclude all the remaining dims?

[veni-vidi-vici-dormivi]

Hm given #569 we could also extend the weighting function to have several options, something like "by_ens", "by_ens_ts" or even pass a function to do it... I have to think about it some more.

[mathause]

Hm given #569 we could also extend the weighting function to have several options, something like "by_ens", "by_ens_ts" or even pass a function to do it... I have to think about it some more.

I would probably go via ens_dim or maybe dims by allowing to pass more than one dimension. I suggest to make our lives easier and keep the simple case for now.

[veni-vidi-vici-dormivi]

should exclude be all dims that are not in ens_dim

No, in practice all dims except for ens_dim and time should be excluded. The test does not reflect the use case in practice... I agree that if the outcome should always be the same we might as well do it internally. I wrote it because I was not sure if there might be another use case for the function but actually in find_localized_empirical_covariance takes the same weights and I am not sure yet how we will handle weights in the autoregression. I guess we can cross that bridge when we come to it.

[mathause]

Looks good, thanks. map_over_subtree skips empty nodes, right?

[veni-vidi-vici-dormivi]

skips empty nodes, right?

Yes.