Introduce Unified Particle Transformer AK4 jet tagger #44641
Conversation
|
cms-bot internal usage |
|
-code-checks Logs: https://cmssdt.cern.ch/SDT/code-checks/cms-sw-PR-44641/39837
Code check has found code style and quality issues which could be resolved by applying following patch(s)
|
cmsbuild
added
code-checks-rejected
code-checks-pending
and removed
code-checks-pending
code-checks-rejected
labels
|
+code-checks Logs: https://cmssdt.cern.ch/SDT/code-checks/cms-sw-PR-44641/39839
|
|
A new Pull Request was created by @AlexDeMoor for master. It involves the following packages:
@hqucms, @vlimant, @cmsbuild, @jfernan2, @mandrenguyen can you please review it and eventually sign? Thanks. cms-bot commands are listed here |
|
please test with cms-data/RecoBTag-Combined#57 |
|
-1 Failed Tests: UnitTests Unit TestsI found 1 errors in the following unit tests: ---> test runtestPhysicsToolsPatAlgos had ERRORS Comparison SummarySummary:
|
| desc.add<edm::InputTag>("puppi_value_map", edm::InputTag("puppi")); | ||
| desc.add<edm::InputTag>("secondary_vertices", edm::InputTag("inclusiveCandidateSecondaryVertices")); | ||
| desc.add<edm::InputTag>("jets", edm::InputTag("ak4PFJetsCHS")); | ||
| desc.addUntracked<edm::InputTag>("unsubjet_map", {}); |
There was a problem hiding this comment.
| desc.addUntracked<edm::InputTag>("unsubjet_map", {}); | |
| desc.add<edm::InputTag>("unsubjet_map", {}); |
I think it should be "tracked", see #44591.
There was a problem hiding this comment.
Thanks I will adjust it 👍
cmsbuild
added
tests-pending
code-checks-pending
and removed
tests-rejected
requires-external
code-checks-approved
labels
|
Again the failed unittest in |
|
ignore tests-rejected with ib-failure |
|
+1 |
|
@antoniovilela I think we need to merge cms-data/RecoBTag-Combined#57 which is required by this PR. |
The model was updated after this PR was tested. Shouldn't we relaunch the tests? |
@mandrenguyen Nice catch -- indeed we should re-run the tests. BTW we were informed by the developers that they would probably update the model with some minor fixes. Maybe @AlexDeMoor could just add some more info here to bring everyone up to date? |
|
please test |
|
-1 Failed Tests: UnitTests 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: Unit TestsI found 1 errors in the following unit tests: ---> test runtestPhysicsToolsPatAlgos had ERRORS Comparison SummarySummary:
NANO Comparison SummarySummary:
Nano size comparison Summary:
|
|
Sorry, I lost track of the external. |
|
type btv, jetmet, tau |
This PR introduce UnifiedParticleTransformerAK4 a novel inclusive tagger for jet. This network will perform an inclusive tagging combining b/c/tau/lep tagging with jet energy regression (both the regression and the resolution estimation via quantile regression). The model is trained with a specific robust training combining improved adversarial training and domain adaptation for reducing the impact of the data/MC disagreement on the final performance. The output nodes of the domains are also kept for exploring the possibility of efficiency mapping and their impact later.
An overview of the method can be seen at: https://indico.cern.ch/event/1368069/contributions/5793148/
A focus on the novel adversarial training is described here: https://indico.cern.ch/event/1372038/#3-adversarial-training-for-par
The preliminary results of the model were shown in the following meeting: https://indico.cern.ch/event/1397392/#17-preliminary-results-of-part
The final results will be shared this Monday: https://indico.cern.ch/event/1403350/#3-part-2024-final-results-and
This PR requires the associated ONNX model which has been submitted in the adequate RecoBTag-Combined repo: cms-data/RecoBTag-Combined#57
For your information: a last training is ongoing trying to improve the current performance via an enriched dataset. A modification of the final model could occur. This will only affect the RecoBTag-Combined PR, not this one.
This pull request is for the master branch but will be backported for the 2024 data-taking. If anyone could specify in which release I should backport, their information is welcome.