Code repository for the SFU CMPT726 project in Spring 2020. An exploration of the DeepLIFT (Deep Learning Important FeaTures) algorithm in three steps.
mnist_original_cnn.py: trains a CNN with parameters identical to original DeepLIFT paper and saves model into mnist_my_cnn_model.h5
mist_modified_cnn.py: trains a CNN with modified architecture (see report) and saves model into mnist_my_cnn_model.h5 (included)
mnist_test.py: runs DeepLIFT on the model in mnist_my_cnn_model.h5 and produces plots as in original DeepLIFT paper
TwoMoonsModel.h5: trained model for Two Moons classfication (see report)
two_moons.py: traines a simple NN for Two Moons classification and runs DeepLIFT. Produces plots in report.
TwoMoonsTest.ipynb: notebook version
For running the analysis of the ATLAS Open Data a docker image needs to be downloaded:
# In the main DeepLIFTforHEP directory:
source setup.sh # defines a few aliases and env variables
# pull docker image uploaded docker image
pullimg # pull latest container from gitlab with ATLAS Open Data included (takes a while)
# this runs the alias: 'docker pull gitlab-registry.cern.ch/bejaeger/deepliftforhep'
# you might need to execute docker as superuser with sudo: 'sudo docker ...'
runcont # runs the container
# this runs the alias: 'docker run --rm -it -v $PWD:/home/DeepLIFTforHEP gitlab-registry.cern.ch/bejaeger/deepliftforhep:latest /bin/bash'
# inside container you have to manually source the setup.sh again
source setup.sh # manually sourcing is currently still needed
cd ATLAS #
./visualize.py # run script to produce plots as an examplevisualize.py: produces plots of the input variable distributions in a new directory called 'output/'
train.py: trains an NN to classify signal and background events
analyze.py: applies DeepLIFT to the trained NN and saves plots of importances scores
config.cfg: settings for the main scripts
plot.py: plot helpers
helper.py: miscellaneous helper functions
loaddata.py: functions for I/O
variables.py: class and functions to modify/constructinput variables