Jet Energy Corrections using DNN regression

The jet data is made up of variable-length unordered sets. To deal with this I use Deep Sets since the method has been used successfully before in Energy Flow Networks to discriminate quark jets from gluon jets.

In addition to that I experiment with ParticleNet which is based on Dynamic Graph Convolutional Neural Network. This architecture is also suitable for variable length sets.

Example model structures from a few runs:

Deep Sets MLP

Deep Sets ResNet

ParticleNet

Build the Docker image (essentially the dependencies in requirements.txt on top of tensorflow/tensorflow:latest-gpu).

docker build . -t jec-dnn

Enter a container with the current directory mounted using your user privileges.

./run_docker.sh

Re-enter stopped container

docker start -i <container id>

To train once you can edit config.yaml to your liking and then run something like:

python train.py -i data/test -o results/test-run -c config.yaml --gpus 1 3

Train using multiple configuration files:

nohup ./run_configs.sh -o results/resnet -c configs/resnet -d data/test -g 0 > resnet.txt

E.g.

python plot.py -i results/resnet -o figs/resnet