XTagger

Training setup & tools

Setup environment

The following scripts will download miniconda environment and install the required packages for training networks with Keras, Tensorflow & ROOT.

for CPU usage only: source Env/setupEnvCPU.sh <installdir>
for CPU & GPU usage: source Env/setupEnvFull.sh <installdir>

After the installation the environments can be used with source Env/env_cpu.sh or source Env/env_gpu.sh respectively.

Custom operations

A set of custom operation modules for Tensorflow can be installed using cmake as following. These allow to train on ROOT trees directly and to perform preprocessing of the training data such as resampling.

PROJECTDIR=$PWD
cd Ops
cmake -DCMAKE_INSTALL_PREFIX=$PROJECTDIR/Ops/release .
make
make install
export PYTHONPATH=$PROJECTDIR/Ops/release/lib/python2.7/site-packages

Standalone evaluation:

python Training/evaluation_test.py --name test_eval --test hdf5_batches/hdf5_list.txt --overwrite

Two batches are provided.

Training using nanoX unpacked samples:

Training is be performed by unpacking jets in extended nanoAOD (NANOX) samples.

python Training/training.py --gpu -b 10000 --train samples/nanox_ctau_10_train.txt --test samples/nanox_ctau_10_test.txt -e 100 --name ctau_10 -c -n 10 --name ctau_10

This would start training using GPUs if available, with a batch size of 10000 and with specified training and testing samples. Furthermore, the training would be performed for 100 epochs and achieving balance of all classes as well as kinematic resampling. The output folder will be created as output/ctau_10. To load a different model, the -m parameter can be used.

Training using hdf5 files (skipping ROOT-to-TF input pipeline)