feat: Machine learning based Ambiguity Solver

[Corentin-Allaire]

This MR depends on #1868.

Implement a machine learning based ambiguity solver. If Acts is compiled with the ACTS_PLUGIN_ONNX option, it can be used in the odd full chain by simply adding the --MLSolver argument.

In addition to the implementation of the solver, this PR adds an option for the regular solver to only work on track with a certain amount of hits (to obtain fairer comparison).

This also add some small change to the Csv multitrajectory writing, it now also write the list of measurements associated with each track.

Finally, the python script to train the network, test the network and compute its performances are made available with the onnx file of the model.

I haven't written the documentation for this solver, yet it will be part of a future MR after this gets merged in.

[asalzburger]

This is on top of the batch input for ONNX #1868 ?

[Corentin-Allaire]

This is on top of the batch input for ONNX #1868 ?

Yes, the batch input caused a factor 7 speed up so I thought it should go in first. But I guess it is close to be merged

[asalzburger]

Alright, I have seen that you had addressed @benjaminhuth s comments, so I merged #1868

[github-actions]

📊 Physics performance monitoring for 05f12d8

Full report
Seeding: seeded, truth estimated, orthogonal
CKF: seeded, truth smeared, truth estimated, orthogonal
IVF: seeded, truth smeared, truth estimated, orthogonal
Ambiguity resolution: seeded, orthogonal
Truth tracking
Truth tracking (GSF)

Vertexing

Vertexing vs. mu

IVF seeded

IVF truth_smeared

IVF truth_estimated

IVF orthogonal

Seeding

Seeding seeded

Seeding truth_estimated

Seeding orthogonal

CKF

CKF seeded

CKF truth_smeared

CKF truth_estimated

CKF orthogonal

Ambiguity resolution

seeded

Truth tracking (Kalman Filter)

Truth tracking

Truth tracking (GSF)

Truth tracking

[codecov]

Codecov Report

Merging #1877 (05f12d8) into main (f35d49b) will not change coverage.
The diff coverage is n/a.

@@           Coverage Diff           @@
##             main    #1877   +/-   ##
=======================================
  Coverage   49.48%   49.48%           
=======================================
  Files         408      408           
  Lines       22704    22704           
  Branches    10363    10363           
=======================================
  Hits        11236    11236           
  Misses       4257     4257           
  Partials     7211     7211           

📣 We’re building smart automated test selection to slash your CI/CD build times. Learn more

[Corentin-Allaire]

The jobs that don't pass the CI are related to the fact I have change the default solver to only consider track with at least 7 hits. I can put the threshold to 0 but honestly I would be concern by the quality of a track with < 7 hits.
If anybody has an opinion on this...

[asalzburger]

e CI are rel

In that case, the PR should include the updated reference, what do you think? @paulgessinger

[paulgessinger]

Is it obvious why the non-ambi outputs also change @Corentin-Allaire?

[Corentin-Allaire]

Is it obvious why the non-ambi outputs also change @Corentin-Allaire?

I the odd-full-chain the vertexing uses the output of the Solver as an input. If you remove some tracks (the one with < 7 measurements) then it will affect the vertexing. What was a bit surprising is that it appear to improve the vertexing performance.

[andiwand]

in general the python scripts could share some common code

left a few other comments. nothing critical - as always feel free to ignore

[Corentin-Allaire]

All of the comment should have been addressed now ! I made one tiny change to the file name for the Csv multitrajectory writing, it use to always be CKFtracks.csv now the naming depend on the algorithm that called the writer (so it is easier to write the output for the CKF and the solver at the same time) .

[andiwand]

great! thank you!