Releases: cms-egamma/ID-Trainer
Releases · cms-egamma/ID-Trainer
Release for general use! v1.4.1
| Currently supports |
|
| Binary-classification (currently using XGBoost and DNN) |
Examples: DY vs ttbar, DY prompt vs DY fake, good electrons vs bad electrons |
| Multi-sample classification (currently using XGBoost and DNN) |
Examples: DY vs (ttbar and QCD) |
| Multi-class classification (currently using XGBoost and DNN) |
Examples: DY vs ttbar vs QCD, good photons vs bad photons vs very bad photons |
Release for general use!
| Currently supports |
|
| Binary-classification (currently using XGBoost and DNN) |
Examples: DY vs ttbar, DY prompt vs DY fake, good electrons vs bad electrons |
| Multi-sample classification (currently using XGBoost and DNN) |
Examples: DY vs (ttbar and QCD) |
| Multi-class classification (currently using XGBoost and DNN) |
Examples: DY vs ttbar vs QCD, , good photons vs bad photons |
v1.3 of ID-Trainer
ID-Trainer
A simple config-based tool for high-energy-physics machine learning tasks.
Full documentation and instructions to use are available here: https://akapoorcern.github.io/ID-Trainer/
| Currently supports |
|
| Binary-classification (currently using XGBoost and DNN) |
Examples: DY vs ttbar, DY prompt vs DY fake, good electrons vs bad electrons |
| Multi-sample classification (currently using XGBoost and DNN) |
Examples: DY vs (ttbar and QCD) |
| Multi-class classification (currently using XGBoost and DNN) |
Examples: DY vs ttbar vs QCD, , good photons vs bad photons |
| Salient features: |
| Parallel reading of root files (using DASK) |
| Runs on flat ntuples (even NanoAODs) out of the box |
| Adding multiple MVAs is very trivial (Subject to available computing power) |
| Cross-section and pt-eta reweighting can be handled together |
| Multi-Sample training possible |
| Multi-Class training possible |
| Ability to customize thresholds |
| What will be the output of the trainer: |
| Feature distributions |
| Statistics in training and testing |
| ROCs, loss plots, MVA scores |
| Confusion Matrices |
| Correlation plots |
| Trained models (h5/pb for DNN / pkl for XGBoost) |
Optional outputs
- Threshold values of scores for chosen working points
- Efficiency vs pT and Efficiency vs eta plots for all classes
- Reweighting plots for pT and eta
- Comparison of new ID performance with benchmark ID flags
A simple config-based tool for machine learning tasks
ID-Trainer
A simple config-based tool for high-energy-physics machine learning tasks.
Full documentation and instructions to use are available here: https://akapoorcern.github.io/ID-Trainer/
| Currently supports |
|
| Binary-classification (currently using XGBoost and DNN) |
Examples: DY vs ttbar, DY prompt vs DY fake, good electrons vs bad electrons |
| Multi-sample classification (currently using XGBoost and DNN) |
Examples: DY vs (ttbar and QCD) |
| Multi-class classification (currently using XGBoost and DNN) |
Examples: DY vs ttbar vs QCD, , good photons vs bad photons |
| Salient features: |
| Parallel reading of root files (using DASK) |
| Runs on flat ntuples (even NanoAODs) out of the box |
| Adding multiple MVAs is very trivial (Subject to available computing power) |
| Cross-section and pt-eta reweighting can be handled together |
| Multi-Sample training possible |
| Multi-Class training possible |
| Ability to customize thresholds |
| What will be the output of the trainer: |
| Feature distributions |
| Statistics in training and testing |
| ROCs, loss plots, MVA scores |
| Confusion Matrices |
| Correlation plots |
| Trained models (h5 for DNN / pkl for XGBoost) |
Optional outputs
- Threshold values of scores for chosen working points
- Efficiency vs pT and Efficiency vs eta plots for all classes
- Reweighting plots for pT and eta
- Comparison of new ID performance with benchmark ID flags
Releasing v1.1 of EGamma ID-Trainer
###############Do not use v1.1 anymore############## Depreciated#################
Clone
git clone --branch v1.1 https://github.com/cms-egamma/ID-Trainer.git
Setup
source /cvmfs/sft.cern.ch/lcg/views/LCG_97python3/x86_64-centos7-gcc8-opt/setup.sh
Create a new config (Just copy the default one and start editing on top of it)
cp Tools/TrainConfig.py MyTrainConfig.py
More information on how to edit the config is in the attached pdf.
All you need to do is to edit the NewTrainConfig.py with the settings for your analysis and then run
python Trainer.py MyTrainConfig
The Trainer.py will read the settings from the config file and run training
Suggestion: Do not remove or touch the original Tools/TrainConfig.py (Keep it for reference)
Releasing v1 of EGamma ID-Trainer
Clone
git clone --branch v1 https://github.com/cms-egamma/ID-Trainer.git
Setup
source /cvmfs/sft.cern.ch/lcg/views/LCG_97python3/x86_64-centos7-gcc8-opt/setup.sh
Create a new config (Just copy the default one and start editing on top of it)
cp Tools/TrainConfig.py Tools/NewTrainConfig.py
All you need to do is to edit the NewTrainConfig.py with the settings for your analysis and then run
python Trainer.py Tools/NewTrainConfig
The Trainer.py will read the settings from the config file and run training
Suggestion: Do not remove or touch the original Tools/TrainConfig.py (Keep it for reference)
More information in the attached pdf.
