test pt3dnn dependence on pLS R error

[slava77]

I'm following a few failures using pLS with OT hits that happen in the pT3 DNN.

(on decent stats) comparing pLS fits using pixel/IT hits only vs the OT-extended: fitter pt errors x1.5 but rec-sim RMS goes down by /2.

With only two examples of the pt3dnn failures, it's unclear where the origins are: did the DNN learn some pt,eta-dependent biases to have good T3-pLS match or does it demote pLS with larger errors.

If the DNN demotes large error pLS, this test should show an efficiency drop in pTx reco.

[slava77]

/run all

[github-actions]

The PR was built and ran successfully in standalone mode. Here are some of the comparison plots.

The full set of validation and comparison plots can be found here.

Here is a timing comparison:

   Evt    Hits       MD       LS      T3       T5       pLS       pT5      pT3      TC       Reset    Event     Short             Rate
   avg     28.6    378.9    273.8    126.8     48.7    690.9     11.4    126.2    134.9    180.7      1.8    2002.7    1283.1+/- 301.1     616.7   explicit[s=4] (target branch)
   avg     28.3    374.6    271.3    125.3     47.1    683.6     11.4    118.8    130.6    178.8      1.5    1971.4    1259.4+/- 299.5     612.5   explicit[s=4] (this PR)

[slava77]

before and after:


clearly shows a loss of pt3s at low pt.

So, this confirms the guess that the pt3dnn demotes pLS with larger errors.

Retraining would be an option, but it's a complication on how to handle the model inputs: either something like in #143 or still compiled but a different function/inference call
@GNiendorf

[github-actions]

The PR was built and ran successfully with CMSSW. Here are some plots.

OOTB All Tracks

The full set of validation and comparison plots can be found here.