Finding performance writer: comments + fake tracks

[SylvainJoube]

While writing a ROOT documentation for the finding and ambiguity resolution algorithms, I thought it would be interesting to have a visualization for the fake tracks rates. The greedy ambiguity resolution algorithm successfully managed to remove every track duplicates while retaining reasonable efficiency, but I wanted to make sure that it would also greatly reduce the number of fake tracks, at least on the data generated by the toy detector (mostly 100GeV muons).

The fake_tracks_plot_tool.hpp file is a duplication of the code from the duplication_plot_tool.hpp file. I think it would be better to merge the two codes, so I will do it if you think it’s a good idea.

If you like the idea, I will write the corresponding documentation and append it to the doc/root/README.md file from PR 535

For example, with toy detector generated data and 100GeV muons, the mean value for “fake tracks” goes from 1.707 with the finding algorithm down to 0.0155 after the ambiguity resolution algorithm (a x110 decrease!):

After the finding algorithm:

After the ambiguity resolution algorithm:

I used the following code to generate the ROOT data with traccc_seeding_example:

# Generate toy detector data
./bin/traccc_simulate_toy_detector --gen-vertex-xyz-mm=0:0:0 --gen-vertex-xyz-std-mm=0:0:0 --gen-mom-gev=100:100 --gen-phi-degree=0:360 --gen-events=10 --gen-nparticles=2000 --output-directory=detray_simulation/toy_detector/n_particles_2000/ --gen-eta=-3:3 --constraint-step-size-mm=1

# …copy data were needed…

# Run the seeding example
./bin/traccc_seeding_example --input-directory=detray_simulation/toy_detector/n_particles_2000/ --detector-file=toy_detector_geometry.json --material-file=toy_detector_homogeneous_material.json --grid-file=toy_detector_surface_grids.json --input-event=1 --track-candidates-range=3:30 --constraint-step-size-mm=1000 --check-performance

[SylvainJoube]

If I’m not mistaken, we can do the following math.
5296 tracks were found and fitted, 1780 remained after ambiguity resolution. So only 1 out of (about) 3 tracks were kept, whereas only 1 out of 110 fake tracks were kept. All duplicated tracks were removed, meaning the ambiguity resolution algorithm disproportionately affected fake and duplicated tracks (which is great!). The efficiency graphs seem to be largely unaffected by ambiguity resolution (but I can’t compare the exact numbers as they are not mentioned in the ROOT plots for some reason I don’t understand).

==> Statistics ... 
- read    7830 spacepoints
- read    7830 measurements
- created (cpu)  4794 seeds
- created (cpu)  5296 found tracks
- created (cpu)  5296 fitted tracks
- created (cpu)  1780 ambiguity free tracks

[beomki-yeo]

Great I assigned myself for review

[SylvainJoube]

Thanks a lot Beom Ki! :)

But let's wait the performance measurement code reorganization by Attila, as he mentioned on PR 578

[beomki-yeo]

Looks good to me but I might need to understand the logic clearly. Actually I don't know the exact definition of fake rate. Could you find its definition?

[beomki-yeo]

Looks good to me. I will update the branch and merge

[SylvainJoube]

Okay, thanks!