End to end reconstruction of neutral pions produced in hadronic tau decays for analysis of the CP structure of the Higgs boson.This repository currently includes two training configurations: DeepPi_v1
for decay mode classification (counting the number of pi0s), and DeepPi_v2
for regression of kinematic properties of DM1/11 taus. This code uses low level input features, Reduced ECAL RecHits, which are available in the AOD data tier.
General guidelines for different areas:
- Production of RHTree ROOT tuples in
Production
(run withcmsenv
) - Tools to convert to dataframes/images in
Analysis
(best to run withtau-ml
) - Neural Network Training (and DataLoading) in
Training
(run withtau-ml
) - Neural Network Evaluation in
Evaluation
(run withtau-ml
)
When using cmsenv
for tuple production, it is recommended to only compile from the Production
folder as python files in other areas may not be compatible with Python 2 and can cause compilation errors.
To install/activate tau-ml
conda environment run source env.sh conda
from the central DeepPi folder.
Parts of the code that you will find here were adapted from MLAnalyzer (Micheal Andrews), TauMLTools (CMS Tau POG developers) and the Imperial College Higgs->TauTau repository, many thanks to all who contributed.
The first step towards producing input images for the neural network is generating ROOT ntuples using the DeepPi produced in the Production
area.
To have access the the BDT scores previously used for DM classification by the HIG-20-006
Analysis, you need to install a custom CMSSW environment. This code is designed to run in a modified version of CMSSW_10_6_19
, first setup your environment:
scramv1 project CMSSW CMSSW_10_6_19
Navigate to the src
directory:
cd CMSSW_10_6_19/src/
Enter the `cmsenv environment:
cmsenv
Intialise the are for git:
git cms-addpkg FWCore/Version
This command will have created two remote repositories, official-cmssw
and my-cmssw
. It will also have created and switched to a new branch, from-CMSSW_10_6_19
.
Add and pull the remote branch from the pre-configured ICHiggsToTauTau
github:
git remote add ic-cmssw git@github.com:gputtley/cmssw.git
git pull ic-cmssw from-CMSSW_10_6_19
You must then compile the code:
scram b -j8
Your custom CMSSW environment is now set up!
Finally clone this repository into the src folder:
git clone git@github.com:lucasrussell01/DeepPi.git
For this step you should be working with the cmsenv
environment.
Compile code from within the production folder:
cd Production
scram b -j8
This configuration file is used to produce ntuples:
python/higgstautau_cfg_aod_inc_hps_2018.py
To run it locally for individual files, use cmsRun
.
To produce ntuples for all samples, jobs can be submitted using crab:
cd python
python ../crab/crab_DeepPi.py --output_folder=DetectorImages_MVA --year=2018 --mc
where the output_folder
argument specifies where in your personal dcache space you want to files to be outputted to.
Once ROOT tuples are produced, they must be processed to extract true taus, and form the pickle
dataframes that are used as inputs for training. For maximal efficiency, this step should also be performed with the cmsenv
environemnt active.
Switch to the Analysis/python
area:
cd ../../Analysis/python
After crab jobs have finished get the list of the files in the dcache area:
../scripts/get_filelists.sh store/user/dwinterb/DetectorImages_MVA_Oct06_MC_106X_2018/ HPS_centre
you must change the name of the directory to the dcache directory where you stored the output from the crab jobs. HPS_centre
is an example run name used for the file list naming convention, you can pick whatever you like but will need to specify it during the next image generation step.
Image of taus are generated from ROOT files using GenerateImagesFile.py
for each ROOT file. To complete this step quickly, jobs should be submitted for each sample type (eg ggH, VBF...) using BatchSubmission.py
:
python BatchSubmission.py --sample=GluGluHToTauTau_M-125 --path_to_list=/vols/cms/lcr119/CMSSW_10_6_19/src/DeepPi/Analysis/python --d_cache=/store/user/dwinterb/DetectorImages_MVA_Oct06_MC_106X_2018/ --run_name=HPS_centre --save_path=/vols/cms/lcr119/Images/HPSCentering/
Here sample
is the sample type to process, path_to_list
is where the filelists you just generated are stored, d_cache
is the path to the crab output in the d cache storage system, run_name
is what you specified when generating the file list, and save-path
is where you would like the images to be saved.
NB: Each sample will created and submit 100-1000 jobs, which take 3-7 minutes each.
You must move the images you created into to separate folders, one for training/validation and one for evaluation. The paths to these folders will need to be specified in the training configuration file later. A Training/Validation and Evaluation split of around 85/15 is what was used previously here.
For neural network training you must work in the tau-ml
environemnt, to install this, go to the DeepPi
directory and source the environment:
cd ..
source env.sh conda
Next, navigate to the Training/python
folder:
cd Training/python
The file to configure training is: Training/configs/training.yaml
You can either update options in this file directly or specify them as command line arguments when runnign training (e.g change input file directory).
Note that to choose which model to train (kinematic regression or DM classification) a model_name
is specified. At the moment there are 2 options:
-
'DeepPi_v1' = DM classification
-
'DeepPi_v2' = kinematic regression
To run with all of the options as specified in training.yaml
:
python TrainingNN.py experiment_name=training
where experiment_name
is used to define an experiment ID, which is useful for keeping track of models/grouping several similar runs together.
To run with changes to options in training.yaml
:
python TrainingNN.py experiment_name=training training_cfg.Setup.n_tau=50
where in this example the number of taus per batch was modified to 50.
ssh lxgpu00
then source conda again and follow above instructions to run TrainingNN.py
again.
Use scripts/training_gpu.sh
.
You need to change the first cd
command in this file by hand to match where your DeepPi
folder is. The script will then source the environment and run all the steps for training listed above. You can modify the training command in the same manner to specify options etc.
To submit the training to the batch system:
qsub -q gpu.q -l h_rt=23:45:45 ../scripts/training_gpu.sh
Running training will create an mlruns
folder within Training/python
with an experiment ID allocated to the experiment name (e.g 0,1,2,...), this will also create a model with a run-ID (hash code) attributed to it e.g mlruns/2/f27c2b7c4c7343e7aa0db74aeec1b924
To evaluate the model or do anything else with it (eg continue trainign), you need to specify the experiment ID and the run-ID
To continue training a model (for example extending the number of epochs), modify the config file: hydra_train.yaml
.
Comment the pretrained=null
line, and uncomment the lines below, modifying the experiemnt_id
and run_id
options as needed:
# pretrained:
# run_id : f27c2b7c4c7343e7aa0db74aeec1b924
# experiment_id: 2
# starting_model: DeepPi_v1_step_final.tf
Models are saved after each epoch, with DeepPi_v1
for DM, and DeepPi_v2
for kinematic regression. In the example above the final model is used but you can also use another epoch e.g by changing DeepPi_v1_step_final.tf
to e.g DeepPi_v1_step_e4.tf
Loss functions are defined in Training/python/losses.py
:
DecayMode_loss
is loss for Decay mode classificationKinematic_loss
is loss for kinematic regression
If you modify these definitions a different loss function will be applied during training.
You can modify input variables inside DataLoader.py
. Note that you need to make sure the information you want to use is stored in the dataframe within the pkl file produced by GenerateImages.py
.
Inside TrainingNN.py: Specify the input_shape and input_types variables if you have modified the shapes or types of the inputs e.g:
input_shape = (((33, 33, 5), 31), 3, None)
input_types = ((tf.float32, tf.float32), tf.float32, tf.float32)
in this example the shapes "(33, 33, 5)" are the images, "31" is the high-level variables, 3 is the kinematic truth shape, and None is the shape of the weights (not used in this example), for types: "(tf.float32, tf.float32)" are the types of the images and the high-level variables respetivly, and "tf.float32, tf.float32" are the targets and the weights.
You must also modify the network architecture to handle the input/output shapes. The "create_vX_model" functions have input_layers - specify a shape that matches the shape of the inputs.
If you want to change the targets that you are attemptin to predict, it is easiest to add a new generator in DataLoader.py
e.g "get_generator_v3" - you could follow "get_generator_v2" as an example.
Inside TrainingNN.py
there is a "run_training" function that calls the "get_generator_vX function" - modify name here
Then apply the above instructions for modifying the shapes and types of inputs, but this time modifying the targets. You will also need to modify the output layers in the network architecture e.g "outputKin" to number of nodes you want in create_vX_model in TrainingNN.py
. You can then modify the loss function as you like, taking into account the different outputs.
Evaluation tools can be found in the Evaluation
directory:
cd Evaluation/python
The following scripts are provided:
apply_DM_training.py
: applies the DM classification model to a evaluation sampleapply_Kin_training.py
: applies the kinematic models to a evaluation samplekinematic_predictions.py
: plots the evaluated kinematic models e.g plots of eta, phi, and momentum distributionsmonitor_metrics.py
: plots the metrics e.g loss, accuracy etc..confusion_matrix.py
: plots efficiency and purity matrices for SM classification
At the moment need to modify the directory by hand inside apply\*training.py
scripts by changing "input_dir".
Additionally, if the shapes or types of inputs and outputs have been modified also need to change these by hand.
There are two steps in evaluating models:
- Apply training (will evaluate the model on a given number of taus)
- Plot Results (plot distributions, confusion matrices etc...)
To, for example, apply the model trained above to 10k taus:
python apply_DM_training.py --expID=2 --runID=f27c2b7c4c7343e7aa0db74aeec1b924 --n_tau=10000
python apply_Kin_training.py --expID=2 --runID=f27c2b7c4c7343e7aa0db74aeec1b924 --n_tau=10000
For now use the respective juypter notebook inside Evaluation/notebooks
.
NB: To start a notebook session, make sure port forwarding is enabled when you connected to the cluster via ssh, specifying a port e.g:
ssh lcr119@lx03.hep.ph.ic.ac.uk -Y -L 1879:localhost:1879
Then source the environement, and run:
jupyter-notebook --port=1879 --no-browser