Introduce Unified Particle Transformer v2 #47173
Conversation
adjust the LT fix for v2 model
|
cms-bot internal usage |
|
-code-checks Logs: https://cmssdt.cern.ch/SDT/code-checks/cms-sw-PR-47173/43412 Code check has found code style and quality issues which could be resolved by applying following patch(s)
|
|
+code-checks Logs: https://cmssdt.cern.ch/SDT/code-checks/cms-sw-PR-47173/43413 |
|
A new Pull Request was created by @AlexDeMoor for master. It involves the following packages:
@cmsbuild, @jfernan2, @mandrenguyen can you please review it and eventually sign? Thanks. cms-bot commands are listed here |
|
test parameters:
|
|
please test |
|
+1 Size: This PR adds an extra 28KB to repository 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:
|
|
@AlexDeMoor do I understand correctly from your PR description that performance should be improved with this PR? The rate ev/s/thd is decreased for all the NANO workflows tested |
|
assign xpog |
|
New categories assigned: xpog @ftorrresd,@hqucms you have been requested to review this Pull request/Issue and eventually sign? Thanks |
I think there are some fluctuations in the NANO tests -- let's run again and see. |
|
please test |
@hqucms Just to add that the changes in this PR should not directly affect NanoAOD workflows computing time because the inference is done at PAT level. |
|
+1 Size: This PR adds an extra 16KB to repository 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:
|
| unsigned int n_vtx = features.sv_features.size(); | ||
|
|
||
| // Use actual sizes for dynamic axes version | ||
| n_cpf_ = std::max((unsigned int)1, n_cpf); |
There was a problem hiding this comment.
The part could be simplified using std::clamp, e.g.,
n_cpf_ = std::clamp(features.c_pf_features.size(), 1, 29);| @@ -76,6 +77,8 @@ UnifiedParticleTransformerAK4ONNXJetTagsProducer::UnifiedParticleTransformerAK4O | |||
| : src_(consumes<TagInfoCollection>(iConfig.getParameter<edm::InputTag>("src"))), | |||
| flav_names_(iConfig.getParameter<std::vector<std::string>>("flav_names")), | |||
| input_names_(iConfig.getParameter<std::vector<std::string>>("input_names")), | |||
| use_dynamic_axes_(iConfig.getParameter<edm::FileInPath>("model_path").fullPath().find("v2.onnx") != | |||
| std::string::npos), | |||
There was a problem hiding this comment.
I would suggest adding a configurable parameter for use_dynamic_axes (with a default value to true), otherwise we are stuck with the naming of v2.onnx.
| @@ -89,7 +92,7 @@ void UnifiedParticleTransformerAK4ONNXJetTagsProducer::fillDescriptions(edm::Con | |||
| desc.add<edm::InputTag>("src", edm::InputTag("pfUnifiedParticleTransformerAK4TagInfos")); | |||
| desc.add<std::vector<std::string>>( | |||
| "input_names", {"input_1", "input_2", "input_3", "input_4", "input_5", "input_6", "input_7", "input_8"}); | |||
| desc.add<edm::FileInPath>("model_path", edm::FileInPath("RecoBTag/Combined/data/UParTAK4/PUPPI/V00/UParTAK4.onnx")); | |||
| desc.add<edm::FileInPath>("model_path", edm::FileInPath("RecoBTag/Combined/data/UParTAK4/PUPPI/V1/UParTAK4_v2.onnx")); | |||
There was a problem hiding this comment.
For consistency, I suggest to rename V1 to V01.
|
@AlexDeMoor It seems |
cmsbuild
added
tests-pending
code-checks-pending
and removed
tests-approved
code-checks-approved
labels
|
+code-checks Logs: https://cmssdt.cern.ch/SDT/code-checks/cms-sw-PR-47173/43463 |
|
Pull request #47173 was updated. @cmsbuild, @ftorrresd, @hqucms, @jfernan2, @mandrenguyen can you please check and sign again. |
PR description:
This PR opens the update of the Unified Particle Transformer bringing substantial performance improvement for flavour tagging performance and model inference time. Link of the XPOG meeting : https://indico.cern.ch/event/1504557/#3-upart-training-updates
This PR has to be tested with the following PR of the model :cms-data/RecoBTag-Combined#64
Please note the final model is being finalized for training and validation.