Generative Adversarial Networks for Simulation in Muon Tomography

Repository for code, data and results. It contains 3 subdirectories:

Paper

Code ready to evaluate the conditional model. Stuff related to the paper: code, datafiles and trained model.

Requirements

A list of the packages needed to run (with the version I use):

• python (3.9.16) any python 3.X will work
• tensorflow (2.10.0)
• keras (2.10.0)
• numpy (1.26.4)
• scikit-learn (1.4.0)
• scipy (1.12.0)
• pandas (2.2.0)
• joblib (1.3.2)
• matplotlib (3.8.2)

How I setup my conda environment

conda create --name=tf-root python=3.9
conda install -c conda-forge root
conda install -c conda-forge cudatoolkit=11.2.2 cudnn=8.1.0
python3 -m pip install tensorflow==2.10
pip install pandas scikit-learn scipy matplotlib

Specify the location of evaluation samples

In Paper/Common/Constants.pyit is necessary to specify the location of the .csv file with the samples used for evaluation. Currently, this file has a size of 493 MB.

Instructions to run

Generated ROOT files and Point of Closest Approach (PoCA) maps:

The generation of the PoCA maps for G4 samples and GAN samples is done via 3 steps:

1. generateSamplestoROOT.py uses the GAN model to generate muon events and save them in a .root file. The info saved in the .root is:

   • The thickness of the pipe corresponding to the event (R)
   • The first detector variables (p*1)
   • The second detector variables (G4) (p*2)
   • The second detector variables generated with GAN model (p*2_gan)

2. runPoCAfromROOT_GANsamples.pyuses the output of previous step to run the PoCA algorithm and save the info to a .root file.

3. plotPoCAmaps_GAN.py generates 2D hists os PoCA maps for G4 samples and GAN samples.

I already prerun the 3 steps, so to test you can skip any of them. The generated root files are in https://cernbox.cern.ch/s/l8RVUAVHXACFtmU (rootFilesGen/).

Old evaluation via Evaluatin.py

The script plotEvaluation.pycontains code to:

1. load the .h5 model placed in Paper/Common/Models/v1/muon_propagation_WGAN_model.h5
2. load the evaluation samples located in a .csv file.
3. evaluate numerically the model (mean, covariance matrix...)
4. produce plots of real and generated data to a .pdf file.

To evaluate the model, just run the following command:

python3 plotEvaluation.py

Pytorch [In Development]

Pytorch implementation of WGAN-GP (from https://github.com/EmilienDupont/wgan-gp).

Training

Old code I developed for TFM.