DraftDPSNoteMLPFACAT2021 < Main

Main Web>TWikiUsers>JoosepPata>DraftDPSNoteMLPFACAT2021 (2021-11-16, JoosepPata)

	One simulated ttbar event with pileup under Run 3 conditions, reconstructed with standard particle flow. The trajectories correspond to the particle flow candidates extrapolated to the ECAL surface, with candidates of different type shown in different colors. We also show the ECAL detector surface (cyan) and the muon stations (blue).

One simulated ttbar event with pileup under Run 3 conditions, reconstructed with standard particle flow. The trajectories correspond to the particle flow candidates extrapolated to the ECAL surface, with candidates of different type shown in different colors. We also show the ECAL detector surface (cyan) and the muon stations (blue).

	One simulated ttbar event with pileup under Run 3 conditions, reconstructed with machine learned particle flow. The trajectories correspond to the particle flow candidates extrapolated to the ECAL surface, with candidates of different type shown in different colors. We also show the ECAL detector surface (cyan) and the muon stations (blue).

One simulated ttbar event with pileup under Run 3 conditions, reconstructed with machine learned particle flow. The trajectories correspond to the particle flow candidates extrapolated to the ECAL surface, with candidates of different type shown in different colors. We also show the ECAL detector surface (cyan) and the muon stations (blue).

	Particle candidate multiplicity in transverse momentum from CMSSW reconstruction, shown for 1000 Run 3 ttbar events. We compare the distributions as reconstructed by standard PF to the ones reconstructed by machine-learned particle flow. Exactly the same events were used for both algorithms, thus the statistical errors are correlated.

Particle candidate multiplicity in transverse momentum from CMSSW reconstruction, shown for 1000 Run 3 ttbar events. We compare the distributions as reconstructed by standard PF to the ones reconstructed by machine-learned particle flow. Exactly the same events were used for both algorithms, thus the statistical errors are correlated.

	Particle candidate multiplicity in transverse momentum from CMSSW reconstruction, shown for 1000 Run 3 QCD events. We compare the distributions as reconstructed by standard PF to the ones reconstructed by machine-learned particle flow. Exactly the same events were used for both algorithms, thus the statistical errors are correlated.

Particle candidate multiplicity in transverse momentum from CMSSW reconstruction, shown for 1000 Run 3 QCD events. We compare the distributions as reconstructed by standard PF to the ones reconstructed by machine-learned particle flow. Exactly the same events were used for both algorithms, thus the statistical errors are correlated.

	Particle candidate multiplicity in pseudorapidity from CMSSW reconstruction, shown for 1000 Run 3 ttbar events. We compare the distributions as reconstructed by standard PF to the ones reconstructed by machine-learned particle flow. Exactly the same events were used for both algorithms, thus the statistical errors are correlated.

Particle candidate multiplicity in pseudorapidity from CMSSW reconstruction, shown for 1000 Run 3 ttbar events. We compare the distributions as reconstructed by standard PF to the ones reconstructed by machine-learned particle flow. Exactly the same events were used for both algorithms, thus the statistical errors are correlated.

	Particle candidate multiplicity in pseudorapidity from CMSSW reconstruction, shown for 1000 Run 3 QCD events. We compare the distributions as reconstructed by standard PF to the ones reconstructed by machine-learned particle flow. Exactly the same events were used for both algorithms, thus the statistical errors are correlated.

Particle candidate multiplicity in pseudorapidity from CMSSW reconstruction, shown for 1000 Run 3 QCD events. We compare the distributions as reconstructed by standard PF to the ones reconstructed by machine-learned particle flow. Exactly the same events were used for both algorithms, thus the statistical errors are correlated.

	Jet transverse momentum distributions from CMSSW reconstruction, shown for 1000 Run 3 ttbar events. We compare the distributions as reconstructed by standard PF to the ones reconstructed by machine-learned particle flow. Exactly the same events were used for both algorithms, thus the statistical errors are correlated. The charged hadron subtraction algorithm was used to remove charged pileup contributions from the jets in both cases.

Jet transverse momentum distributions from CMSSW reconstruction, shown for 1000 Run 3 ttbar events. We compare the distributions as reconstructed by standard PF to the ones reconstructed by machine-learned particle flow. Exactly the same events were used for both algorithms, thus the statistical errors are correlated. The charged hadron subtraction algorithm was used to remove charged pileup contributions from the jets in both cases.

	Jet transverse momentum distributions from CMSSW reconstruction, shown for 1000 Run 3 QCD events. We compare the distributions as reconstructed by standard PF to the ones reconstructed by machine-learned particle flow. Exactly the same events were used for both algorithms, thus the statistical errors are correlated. The charged hadron subtraction algorithm was used to remove charged pileup contributions from the jets in both cases.

Jet transverse momentum distributions from CMSSW reconstruction, shown for 1000 Run 3 QCD events. We compare the distributions as reconstructed by standard PF to the ones reconstructed by machine-learned particle flow. Exactly the same events were used for both algorithms, thus the statistical errors are correlated. The charged hadron subtraction algorithm was used to remove charged pileup contributions from the jets in both cases.

	Jet pseudorapidity distributions from CMSSW reconstruction, shown for 1000 Run 3 ttbar events. We compare the distributions as reconstructed by standard PF to the ones reconstructed by machine-learned particle flow. Exactly the same events were used for both algorithms, thus the statistical errors are correlated. The charged hadron subtraction algorithm was used to remove charged pileup contributions from the jets in both cases.

Jet pseudorapidity distributions from CMSSW reconstruction, shown for 1000 Run 3 ttbar events. We compare the distributions as reconstructed by standard PF to the ones reconstructed by machine-learned particle flow. Exactly the same events were used for both algorithms, thus the statistical errors are correlated. The charged hadron subtraction algorithm was used to remove charged pileup contributions from the jets in both cases.

	Jet pseudorapidity distributions from CMSSW reconstruction, shown for 1000 Run 3 QCD events. We compare the distributions as reconstructed by standard PF to the ones reconstructed by machine-learned particle flow. Exactly the same events were used for both algorithms, thus the statistical errors are correlated. The charged hadron subtraction algorithm was used to remove charged pileup contributions from the jets in both cases.

Jet pseudorapidity distributions from CMSSW reconstruction, shown for 1000 Run 3 QCD events. We compare the distributions as reconstructed by standard PF to the ones reconstructed by machine-learned particle flow. Exactly the same events were used for both algorithms, thus the statistical errors are correlated. The charged hadron subtraction algorithm was used to remove charged pileup contributions from the jets in both cases.

	Jet transverse momentum distributions from CMSSW reconstruction, shown for 1000 Run 3 ttbar events. We compare the distributions as reconstructed by standard PF to the ones reconstructed by machine-learned particle flow. Exactly the same events were used for both algorithms, thus the statistical errors are correlated. The PUPPI algorithm was used to remove charged pileup contributions from the jets in both cases.

Jet transverse momentum distributions from CMSSW reconstruction, shown for 1000 Run 3 ttbar events. We compare the distributions as reconstructed by standard PF to the ones reconstructed by machine-learned particle flow. Exactly the same events were used for both algorithms, thus the statistical errors are correlated. The PUPPI algorithm was used to remove charged pileup contributions from the jets in both cases.

	Jet transverse momentum distributions from CMSSW reconstruction, shown for 1000 Run 3 QCD events. We compare the distributions as reconstructed by standard PF to the ones reconstructed by machine-learned particle flow. Exactly the same events were used for both algorithms, thus the statistical errors are correlated. The PUPPI algorithm was used to remove charged pileup contributions from the jets in both cases.

Jet transverse momentum distributions from CMSSW reconstruction, shown for 1000 Run 3 QCD events. We compare the distributions as reconstructed by standard PF to the ones reconstructed by machine-learned particle flow. Exactly the same events were used for both algorithms, thus the statistical errors are correlated. The PUPPI algorithm was used to remove charged pileup contributions from the jets in both cases.

	Jet pseudorapidity distributions from CMSSW reconstruction, shown for 1000 Run 3 ttbar events. We compare the distributions as reconstructed by standard PF to the ones reconstructed by machine-learned particle flow. Exactly the same events were used for both algorithms, thus the statistical errors are correlated. The PUPPI algorithm was used to remove charged pileup contributions from the jets in both cases.

Jet pseudorapidity distributions from CMSSW reconstruction, shown for 1000 Run 3 ttbar events. We compare the distributions as reconstructed by standard PF to the ones reconstructed by machine-learned particle flow. Exactly the same events were used for both algorithms, thus the statistical errors are correlated. The PUPPI algorithm was used to remove charged pileup contributions from the jets in both cases.

	Jet pseudorapidity distributions from CMSSW reconstruction, shown for 1000 Run 3 QCD events. We compare the distributions as reconstructed by standard PF to the ones reconstructed by machine-learned particle flow. Exactly the same events were used for both algorithms, thus the statistical errors are correlated. The PUPPI algorithm was used to remove charged pileup contributions from the jets in both cases.

Jet pseudorapidity distributions from CMSSW reconstruction, shown for 1000 Run 3 QCD events. We compare the distributions as reconstructed by standard PF to the ones reconstructed by machine-learned particle flow. Exactly the same events were used for both algorithms, thus the statistical errors are correlated. The PUPPI algorithm was used to remove charged pileup contributions from the jets in both cases.

	The missing transverse energy (MET) kinematic distributions from CMSSW reconstruction, shown for 1000 Run 3 ttbar events. We compare the distributions as reconstructed by standard PF to the ones reconstructed by machine-learned particle flow. Exactly the same events were used for both algorithms, thus the statistical errors are correlated. The MET was reconstructed based on CHS inputs.

	The missing transverse energy (MET) kinematic distributions from CMSSW reconstruction, shown for 1000 Run 3 ttbar events. We compare the distributions as reconstructed by standard PF to the ones reconstructed by machine-learned particle flow. Exactly the same events were used for both algorithms, thus the statistical errors are correlated. The MET was reconstructed based on CHS inputs. We observe a misreconstructed high-MET tail in the QCD sample that was not used in training. This may possibly be mitigated with a per-event loss component, additional training samples and is left for a future study.

The missing transverse energy (MET) kinematic distributions from CMSSW reconstruction, shown for 1000 Run 3 ttbar events. We compare the distributions as reconstructed by standard PF to the ones reconstructed by machine-learned particle flow. Exactly the same events were used for both algorithms, thus the statistical errors are correlated. The MET was reconstructed based on CHS inputs. We observe a misreconstructed high-MET tail in the QCD sample that was not used in training. This may possibly be mitigated with a per-event loss component, additional training samples and is left for a future study.

	The cumulative missing transverse energy (MET) kinematic distributions from CMSSW reconstruction, shown for 1000 Run 3 ttbar events. We compare the distributions as reconstructed by standard PF to the ones reconstructed by machine-learned particle flow. Exactly the same events were used for both algorithms, thus the statistical errors are correlated. The MET was reconstructed based on CHS inputs.

	The cumulative missing transverse energy (MET) kinematic distributions from CMSSW reconstruction, shown for 1000 Run 3 ttbar events. We compare the distributions as reconstructed by standard PF to the ones reconstructed by machine-learned particle flow. Exactly the same events were used for both algorithms, thus the statistical errors are correlated. The MET was reconstructed based on CHS inputs. We observe a misreconstructed high-MET tail in the QCD sample that was not used in training. This may possibly be mitigated with a per-event loss component, additional training samples and is left for a future study.

The cumulative missing transverse energy (MET) kinematic distributions from CMSSW reconstruction, shown for 1000 Run 3 ttbar events. We compare the distributions as reconstructed by standard PF to the ones reconstructed by machine-learned particle flow. Exactly the same events were used for both algorithms, thus the statistical errors are correlated. The MET was reconstructed based on CHS inputs. We observe a misreconstructed high-MET tail in the QCD sample that was not used in training. This may possibly be mitigated with a per-event loss component, additional training samples and is left for a future study.

	The missing transverse energy (MET) kinematic distributions from CMSSW reconstruction, shown for 1000 Run 3 ttbar events. We compare the distributions as reconstructed by standard PF to the ones reconstructed by machine-learned particle flow. Exactly the same events were used for both algorithms, thus the statistical errors are correlated. The MET was reconstructed based on PUPPI inputs.

	The missing transverse energy (MET) kinematic distributions from CMSSW reconstruction, shown for 1000 Run 3 ttbar events. We compare the distributions as reconstructed by standard PF to the ones reconstructed by machine-learned particle flow. Exactly the same events were used for both algorithms, thus the statistical errors are correlated. The MET was reconstructed based on PUPPI inputs. We observe a misreconstructed high-MET tail in the QCD sample that was not used in training. This may possibly be mitigated with a per-event loss component, additional training samples and is left for a future study.

The missing transverse energy (MET) kinematic distributions from CMSSW reconstruction, shown for 1000 Run 3 ttbar events. We compare the distributions as reconstructed by standard PF to the ones reconstructed by machine-learned particle flow. Exactly the same events were used for both algorithms, thus the statistical errors are correlated. The MET was reconstructed based on PUPPI inputs. We observe a misreconstructed high-MET tail in the QCD sample that was not used in training. This may possibly be mitigated with a per-event loss component, additional training samples and is left for a future study.

	The cumulative missing transverse energy (MET) kinematic distributions from CMSSW reconstruction, shown for 1000 Run 3 ttbar events. We compare the distributions as reconstructed by standard PF to the ones reconstructed by machine-learned particle flow. Exactly the same events were used for both algorithms, thus the statistical errors are correlated. The MET was reconstructed based on PUPPI inputs.

	The cumulative missing transverse energy (MET) kinematic distributions from CMSSW reconstruction, shown for 1000 Run 3 ttbar events. We compare the distributions as reconstructed by standard PF to the ones reconstructed by machine-learned particle flow. Exactly the same events were used for both algorithms, thus the statistical errors are correlated. The MET was reconstructed based on PUPPI inputs. We observe a misreconstructed high-MET tail in the QCD sample that was not used in training. This may possibly be mitigated with a per-event loss component, additional training samples and is left for a future study.

The cumulative missing transverse energy (MET) kinematic distributions from CMSSW reconstruction, shown for 1000 Run 3 ttbar events. We compare the distributions as reconstructed by standard PF to the ones reconstructed by machine-learned particle flow. Exactly the same events were used for both algorithms, thus the statistical errors are correlated. The MET was reconstructed based on PUPPI inputs. We observe a misreconstructed high-MET tail in the QCD sample that was not used in training. This may possibly be mitigated with a per-event loss component, additional training samples and is left for a future study.

	The loss function for the model as a function of the training epoch. To check for stability, we perform 10 independent trainings and report the mean (solid line) and standard deviation (shaded area) of these trainings.

The loss function for the model as a function of the training epoch. To check for stability, we perform 10 independent trainings and report the mean (solid line) and standard deviation (shaded area) of these trainings.

	The GPU inference time on an 8GB NVIDIA RTX2060S device using ONNXRuntime 1.9.0. The average runtime per event for a typical simulated Run3-like event is ~10ms. We observe approximately linear scaling in the runtime with increasing event size. One execution stream (a single CPU thread) was used, thus it should not be interpreted as a production-like setup. The final performance numbers may vary depending on the computing optimizations and chosen hyperparameters. For context, on a single CPU thread (Intel i7-10700 @ 2.9GHz), the baseline PF requires approximately (9 ± 5) ms, the MLPF model approximately 320 ± 50 ms for Run 3 ttbar MC events.

The GPU inference time on an 8GB NVIDIA RTX2060S device using ONNXRuntime 1.9.0. The average runtime per event for a typical simulated Run3-like event is ~10ms. We observe approximately linear scaling in the runtime with increasing event size. One execution stream (a single CPU thread) was used, thus it should not be interpreted as a production-like setup. The final performance numbers may vary depending on the computing optimizations and chosen hyperparameters. For context, on a single CPU thread (Intel i7-10700 @ 2.9GHz), the baseline PF requires approximately (9 ± 5) ms, the MLPF model approximately 320 ± 50 ms for Run 3 ttbar MC events.

	The GPU memory usage on an 8GB NVIDIA RTX2060S device using ONNXRuntime 1.9.0. The average GPU memory usage for a typical simulated Run3-like event is ~800MB. We observe approximately linear scaling in memory usage with increasing event size. One execution stream (a single CPU thread) was used, thus it should not be interpreted as a production-like setup. The final performance numbers may vary depending on the computing optimizations and chosen hyperparameters.

The GPU memory usage on an 8GB NVIDIA RTX2060S device using ONNXRuntime 1.9.0. The average GPU memory usage for a typical simulated Run3-like event is ~800MB. We observe approximately linear scaling in memory usage with increasing event size. One execution stream (a single CPU thread) was used, thus it should not be interpreted as a production-like setup. The final performance numbers may vary depending on the computing optimizations and chosen hyperparameters.

-- JoosepPata - 2021-11-16

Attachments

Topic attachments
I	Attachment	History	Action	Size	Date	Who	Comment
png	ak4jet_chs_eta_qcd.png	r1	manage	67.7 K	2021-11-16 - 15:37	JoosepPata	AK4 jets with PF and MLPF
png	ak4jet_chs_eta_ttbar.png	r1	manage	66.7 K	2021-11-16 - 15:37	JoosepPata	AK4 jets with PF and MLPF
png	ak4jet_chs_pt_qcd.png	r1	manage	57.0 K	2021-11-16 - 15:37	JoosepPata	AK4 jets with PF and MLPF
png	ak4jet_chs_pt_ttbar.png	r1	manage	56.1 K	2021-11-16 - 15:37	JoosepPata	AK4 jets with PF and MLPF
png	ak4jet_puppi_eta_qcd.png	r1	manage	69.9 K	2021-11-16 - 15:37	JoosepPata	AK4 jets with PF and MLPF
png	ak4jet_puppi_eta_ttbar.png	r1	manage	68.5 K	2021-11-16 - 15:37	JoosepPata	AK4 jets with PF and MLPF
png	ak4jet_puppi_pt_qcd.png	r1	manage	57.1 K	2021-11-16 - 15:38	JoosepPata	AK4 jets with PF and MLPF
png	ak4jet_puppi_pt_ttbar.png	r1	manage	55.4 K	2021-11-16 - 15:38	JoosepPata	AK4 jets with PF and MLPF
png	bins_cg_0.png	r1	manage	973.6 K	2021-11-16 - 15:13	JoosepPata	CombinedGraph layer learned binning in the first two layers
png	bins_cg_1.png	r1	manage	1009.4 K	2021-11-16 - 15:13	JoosepPata	CombinedGraph layer learned binning in the first two layers
png	candidates_eta_single_qcd.png	r2 r1	manage	203.4 K	2021-11-16 - 15:46	JoosepPata	PFCandidates from PF and MLPF
png	candidates_eta_single_ttbar.png	r2 r1	manage	201.0 K	2021-11-16 - 15:46	JoosepPata	PFCandidates from PF and MLPF
png	candidates_pt_single_qcd.png	r2 r1	manage	278.1 K	2021-11-16 - 15:46	JoosepPata	PFCandidates from PF and MLPF
png	candidates_pt_single_ttbar.png	r2 r1	manage	250.0 K	2021-11-16 - 15:47	JoosepPata	PFCandidates from PF and MLPF
jpg	ev1_mlpf_3d.jpg	r1	manage	5867.2 K	2021-11-16 - 15:04	JoosepPata	Event visualizations
jpg	ev1_pf_3d.jpg	r1	manage	5816.6 K	2021-11-16 - 15:04	JoosepPata	Event visualizations
png	loss_curves_std.png	r1	manage	138.0 K	2021-11-16 - 15:08	JoosepPata	Loss curves.
png	memory_scaling.png	r2 r1	manage	231.5 K	2021-11-16 - 15:28	JoosepPata	Memory and runtime scaling of the MLPF algorithm.
png	pfmet_c_pt_qcd.png	r1	manage	55.1 K	2021-11-16 - 15:49	JoosepPata	MET with PF and MLPF
png	pfmet_c_pt_ttbar.png	r1	manage	54.7 K	2021-11-16 - 15:49	JoosepPata	MET with PF and MLPF
png	pfmet_pt_qcd.png	r1	manage	60.6 K	2021-11-16 - 15:49	JoosepPata	MET with PF and MLPF
png	pfmet_pt_ttbar.png	r1	manage	60.4 K	2021-11-16 - 15:49	JoosepPata	MET with PF and MLPF
png	pfmet_puppi_c_pt_qcd.png	r1	manage	55.3 K	2021-11-16 - 15:49	JoosepPata	MET with PF and MLPF
png	pfmet_puppi_c_pt_ttbar.png	r1	manage	55.4 K	2021-11-16 - 15:49	JoosepPata	MET with PF and MLPF
png	pfmet_puppi_pt_qcd.png	r1	manage	59.9 K	2021-11-16 - 15:49	JoosepPata	MET with PF and MLPF
png	pfmet_puppi_pt_ttbar.png	r1	manage	60.3 K	2021-11-16 - 15:49	JoosepPata	MET with PF and MLPF
png	runtime_scaling.png	r2 r1	manage	238.8 K	2021-11-16 - 15:28	JoosepPata	Memory and runtime scaling of the MLPF algorithm.

Topic revision: r1 - 2021-11-16 - JoosepPata

Main

Webs

Public webs

Welcome Guest

- Cern Search
- TWiki Search
- Google Search
Main All webs

Copyright &© 2008-2024 by the contributing authors. All material on this collaboration platform is the property of the contributing authors.
or Ideas, requests, problems regarding TWiki? use Discourse or Send feedback