sam-may / deepisolation

Deep neural network for identifying prompt vs. fake leptons

DeepIsolation

Deep neural network for identifying prompt vs. fake leptons

Instructions

1. git clone https://github.com/sam-may/DeepIsolation --recursive ; cd DeepIsolation
2. Make baby root files that can be used to train BDT and DNN. See instructions in BabyMaker directory.
3. Train BDT (see instructions in BDT directory) and/or DNN (see instructions in NeuralNet directory).
4. Evaluate results (see instructions in NeuralNet directory).

Singularity shell command

singularity shell --bind /usr/lib64/nvidia:/host-libs /cvmfs/singularity.opensciencegrid.org/opensciencegrid/tensorflow-gpu:latest

deepisolation's Issues

Implement preprocessing for features

Transform each input feature such that it has zero mean and unit width.

Implement selective sampling such that pT and eta distributions are equal for signal/background

Selectively sample signal events such that the pT and eta distributions of signal/background are identical.

Implement batch training

Already nearing GPU memory limit when training with 100k events. Need to implement batch training so that NN can be trained with many more events (DeepJet is trained with 10's of millions).