Integration of extended ParticleNet trainings for simultaneous jet flavor tagging, pT regression, and tau ID + reconstruction #40745
Conversation
|
-code-checks Logs: https://cmssdt.cern.ch/SDT/code-checks/cms-sw-PR-40745/34157
Code check has found code style and quality issues which could be resolved by applying following patch(s)
|
|
@scooperstein As it says above, you can need to apply the code format patch with |
|
+code-checks Logs: https://cmssdt.cern.ch/SDT/code-checks/cms-sw-PR-40745/34184 |
|
A new Pull Request was created by @scooperstein for master. It involves the following packages:
@cmsbuild, @mandrenguyen, @clacaputo can you please review it and eventually sign? Thanks. cms-bot commands are listed here |
Hi @mandrenguyen, yes sorry about that. I ran the code formatting and I have updated the PR. |
|
test parameters: |
|
@scooperstein can you please update the PR title to be a bit more informative? |
scooperstein
changed the title
done, hopefully this is better :) Just to add, if it useful I can also add links to various presentations outlining the object performance deliverables from these trainings. The implementation of these networks has also given the go-ahead by the Tau POG and JME conveners. |
|
@scooperstein Yes, please link any material in the PR description. It doesn't appear that this code is classification is actually being executed in any workflow. Shouldn't we test this? |
|
type jetmet, btv |
|
Pull request #40745 was updated. @swertz, @vlimant, @clacaputo, @cmsbuild, @simonepigazzini, @mandrenguyen can you please check and sign again. |
|
please test Another test after rebasing, to sort out these differences in jet variables |
|
+1 Summary: https://cmssdt.cern.ch/SDT/jenkins-artifacts/pull-request-integration/PR-470863/31621/summary.html The following merge commits were also included on top of IB + this PR after doing git cms-merge-topic:
You can see more details here: Comparison SummarySummary:
NANO Comparison SummarySummary:
Nano size comparison Summary:
|
|
+xpog All the changes in jet and fatJet discriminators are understood. We think the changes flagged for other jet variables are spurious. |
|
+reconstruction |
cmsbuild
added
reconstruction-approved
fully-signed
and removed
reconstruction-pending
pending-signatures
labels
|
This pull request is fully signed and it will be integrated in one of the next master IBs (tests are also fine). This pull request will now be reviewed by the release team before it's merged. @perrotta, @dpiparo, @rappoccio (and backports should be raised in the release meeting by the corresponding L2) |
|
+1 |
|
this required cms-data/RecoBTag-Combined#50 to go in to |
This PR adds a set of ParticleNet-based networks and infers them on the relevant jet collections. These networks are significant extensions of the ParticleNet jet flavor classifiers that have been included in CMSSW so far. The new AK4 network performs simultaneously jet flavor classification, jet pT regression, and hadronic tau ID and reconstruction (charge, DM). The new AK8 network performs jet classification and mass regression simultaneously, while also including resonance decays to merged hadronic taus for the first time.
The AK4 networks are split by jet eta (central vs. forward) and by jet source (CHS vs. PUPPI). The CHS training is needed for studies of using ParticleNet for hadronic tau ID and reconstruction (TAU), while the PUPPI training is intended for the jet classification and pT regression (JME, BTV). We intend to unify these tasks to run on only one jet source (PUPPI) in the future, but this solution was agreed upon with the L2 POG conveners in the interim because current PUPPI tunes are inefficient for seeding hadronic taus. The networks are also split by jet eta because the scope of the forward network is highly reduced, allowing for a lighter and faster network architecture for the forward jet inferences. The AK8 network takes PUPPI jets as a source and is only inferred for central AK8 jets.
Note that the input features for these networks are dependent on miniAOD input collections, therefore they can only be inferred on the miniAOD.
This commit is dependent on the model files as committed in this PR: cms-data/RecoBTag-Combined#50. This PR has been tested by running the added tasks on a single tt file and checking the outputs of the network evaluations.
Additional information on these networks available in recent presentations:
at JME: https://indico.cern.ch/event/1220368/contributions/5141019/attachments/2548768/4389765/cooperstein_ParticleNet_nov162022.pdf
and at the Tau POG: https://indico.cern.ch/event/1223165/contributions/5158297/attachments/2561047/4414677/cooperstein_ParticleNetPlusTau_dec62022.pdf
In addition, an analysis note AN-22-094 (http://cms.cern.ch/iCMS/jsp/openfile.jsp?tp=draft&files=AN2022_094_v2.pdf) describes these networks in detail.
@rgerosa