search

trackreco / mkFit

Vectorized, Parallelized Tracking
https://trackreco.github.io/
Apache License 2.0
17 stars 15 forks source link

Cluster size cuts #358

Closed cerati closed 3 years ago

cerati commented 3 years ago

Add option for cluster row size cut (iteration and layer-dependent), and turn it on for InitialStep.

Tested on initialStep within MTV: http://uaf-8.t2.ucsd.edu/~cerati/plots_Sep16_ttbar_clshcut_PR/plots_initialStep.html

A PR to CMSSW with JSON updates will also be needed.

cerati commented 3 years ago

More MTV plots.

Muon samples (very low, low, high pT): http://uaf-8.t2.ucsd.edu/~cerati/plots_Sep16_10muvlpt_clshcut_PR/plots_initialStep.html http://uaf-8.t2.ucsd.edu/~cerati/plots_Sep16_10mulowpt_clshcut_PR/plots_initialStep.html http://uaf-8.t2.ucsd.edu/~cerati/plots_Sep16_10mu_clshcut_PR/plots_initialStep.html

TTbar on top of PR357: http://uaf-8.t2.ucsd.edu/~cerati/plots_Sep16_ttbar_clshcut_PR-PR357/plots_building_initialStep/hitsLayers.pdf

cerati commented 3 years ago

Code updated to use WP99.

Tests have been re-made on top of updates to PR357: http://uaf-8.t2.ucsd.edu/~cerati/plots_Sep20_ttbar_clshcut_PR-PR357/plots_initialStep.html

Relevant plots for highPurityByOriginalAlgo: image image

WP99 was chosen taking into account the following values of "assocFraction" (faction of hits associated to the same simTrack), since WP95 does not really improve this metric (so the drop in number of hits is less motivated): PR357: 0.97264939 PR357+clshcut99: 0.97300457 PR357+clshcut95: 0.97309722

cerati commented 3 years ago

@slava77 I have limited time these days, please let me know what is needed from me before merging

slava77 commented 3 years ago

@slava77 I have limited time these days, please let me know what is needed from me before merging

we talked yesterday about merging this PR (@osschar or @mmasciov to push the button).

After that, if time allows the idea is to check this WP99 applied in the other iterations in our set of 7 planned for release.

cerati commented 3 years ago

I tried to apply this to the 7 iterations, which should have been straightforward, but looks like something is broken. I need to find the time to debug what is going on. I'd suggest to pause and not merge this yet. PR357 may probably go ahead first

cerati commented 3 years ago

Ok, the issue was actually with the reference. The result of using WP99 on the 7 iters compared to iter0 only is here: http://uaf-8.t2.ucsd.edu/~cerati/plots_Sep23_ttbar_clshcut99_7iter/ Note that is still before 357 and 360 were merged.

slava77 commented 3 years ago

Ok, the issue was actually with the reference. The result of using WP99 on the 7 iters compared to iter0 only is here: http://uaf-8.t2.ucsd.edu/~cerati/plots_Sep23_ttbar_clshcut99_7iter/ Note that is still before 357 and 360 were merged.

it looks like a good candidate to go in; apparently the outer iterations look better.

cerati commented 3 years ago

This is on top of current devel: http://uaf-8.t2.ucsd.edu/~cerati/plots_Sep24-devel_ttbar_clshcut99_7iter/ I am not sure it does much...

slava77 commented 3 years ago

did #360 somehow drop the large clusters already? I don't think #357 would affect the impact of this PR.

osschar commented 3 years ago

Yes, that's what I was saying on Wed, that most of large clusters likely come from tracks scraping the modules at low angles.

cerati commented 3 years ago

Closing as it does not bring improvements on top of other PRs.