keras-training

Installation

Install miniconda2 by sourcing install_miniconda.sh in your home directory. Log out and log back in after this.

cp install_miniconda.sh ~/
cd ~
source install_miniconda.sh

Install the rest of the dependencies:

cd ~/keras-training
source install.sh

Each time you log in set things up:

source setup.sh

Conversion of data

All of the data ntuple files are available here: https://cernbox.cern.ch/index.php/s/AgzB93y3ac0yuId

To add the truth values and flatten the trees (you can skip this step)

cd ~/keras-training/convert
python addTruth.py -t t_allpar \
../data/processed-pythia82-lhc13-*-pt1-50k-r1_h022_e0175_t220_nonu.root

To hadd these files and convert from TTree to numpy array with random shuffling (you can skip this step)

hadd -f \
../data/processed-pythia82-lhc13-all-pt1-50k-r1_h022_e0175_t220_nonu_truth.root \
../data/processed-pythia82-lhc13-*-pt1-50k-r1_h022_e0175_t220_nonu_truth.root
python convert.py -t t_allpar_new \
../data/processed-pythia82-lhc13-all-pt1-50k-r1_h022_e0175_t220_nonu_truth.root

You can also copy this file directly from: https://cernbox.cern.ch/index.php/s/jvFd5MoWhGs1l5v

Training and evaluation

To run a simple training with L1 regularization (lamba = 1e-4):

cd ~/keras-training/train
python train.py -t t_allpar_new \
	-i ../data/processed-pythia82-lhc13-all-pt1-50k-r1_h022_e0175_t220_nonu_truth.z \
	-c train_config_threelayer.yml \
	-o train_simple_l10p0001/

and evaluate the training:

python eval.py -t t_allpar_new \
	-i ../data/processed-pythia82-lhc13-all-pt1-50k-r1_h022_e0175_t220_nonu_truth.z \
	-m train_simple/KERAS_check_best_model.h5 \
	-c train_config_threelayer.yml \
	-o eval_simple_l10p0001/

Pruning and retraining

To prune the trained model by removing weights below a certain percentile (relative weight percentile < 32.7%):

mkdir -p prune_simple_l10p0001_33perc
python prune.py -m train_simple_l10p0001/KERAS_check_best_model.h5 \
	-o prune_simple_l10p0001_33perc/trained_model_33perc.h5 \
	--relative-weight-percentile 32.7

and evaluate the pruned model:

python eval.py -t t_allpar_new \
	-i ../data/processed-pythia82-lhc13-all-pt1-50k-r1_h022_e0175_t220_nonu_truth.z \
	-m prune_simple_l10p0001_33perc/pruned_model.h5 \
	-c train_config_threelayer.yml \
	-o eval_simple_l10p0001_33perc/

To retrain the pruned model (keeping the pruned weights fixed to 0) with L1 regularization (lamba = 1e-4):

python retrain.py -t t_allpar_new \
	-i ../data/processed-pythia82-lhc13-all-pt1-50k-r1_h022_e0175_t220_nonu_truth.z \
	-o retrain_simple_l10p0001_33perc \
	-m prune_simple_l10p0001_33perc/pruned_model.h5 \
	-c train_config_threelayer.yml \
	-d prune_simple_l10p0001_33perc/pruned_model_drop_weights.h5

and prune again (this time 48 percent of the weights):

mkdir -p prune_simple_l10p0001_48perc
python prune.py -m retrain_simple_l10p0001_33perc/KERAS_check_best_model.h5 \
       --relative-weight-percentile 47.5 \
       -o prune_simple_l10p0001_48perc/pruned_model.h5

and evalute the pruned model (2nd iteration):

python eval.py -t t_allpar_new \
       -i ../data/processed-pythia82-lhc13-all-pt1-50k-r1_h022_e0175_t220_nonu_truth.z \
       -m prune_simple_l10p0001_48perc/pruned_model.h5 \
       -c train_config_threelayer.yml \
       -o eval_simple_l10p0001_48perc/

This procedure can be repeated as done in train/train_prune_eval_retrain.sh.