This repository contains the code I wrote while working on my master's thesis project at the Department of Physics at Lund University, Sweden. My project explores the use of autoencoders (AEs), a type of artificial neural network (NN), for data compression of hadron jets in High Energy Physics (HEP).
The project can be divided in two parts. One where the jet 4-momentum is compressed to three variables using a 3D latent space AE and one where 27 different jet variables are compressed into a varying number of latent space dimension.
Due to size contraints the actual data files I used are not included in this GitHub repository.
The data was extracted from a ROOT file with process_ROOT_data.ipynb.
The data was extracted from a derived AOD with TLA data using process_AOD_data_all_jets.ipynb. The data was normalized in aod_custom_normalization.ipynb located in jet_by_jet_compression/aod_compression/.
The functions custom_normalization and custom_unnormalize in utils.py can be used to normalize and unnormalize the data according to my custom scheme.
The neural network architectures used throughout this project are defined as python classes in my_nn_modules.py. They all inherit from PyTorch's nn.Module class.
'utils.py' contains definitions of useful functions used for preprocessing data, evaluation etc.
The neural network architectures used throughout this project are defined as python classes in my_nn_modules.py. They all inherit from PyTorch's nn.Module class.
'utils.py' contains definitions of useful functions used for preprocessing data, evaluation etc.
These were mostly trained "by hand" using jupyter notebooks located in lth_thesis_project/jet_by_jet_compression/AE_3D_latent_space/, lth_thesis_project/jet_by_jet_compression/AE_2D_latent_space/ and lth_thesis_project/jet_by_jet_compression/fastai_AE_3D.
Some grid searches were also run. Those can be found in lth_thesis_project/jet_by_jet_compression/grid_search/.
These were mostly trained on LUNARC's Aurora and on HPC2N's Kebnekaise which are High Performance Computing (HPC) clusters located in Sweden. The scripts prepare_grid_search_aod.py and prepare_grid_search_aurora.py were used to create folders and training scripts with different combinations of hyperparameters as well as a bash script used to launch the the grid search on the HPC clusters. The scripts can be found in lth_thesis_project/jet_by_jet_compression/grid_search/
In the notebooks where networks were trained they were also often evaluated, producing plots of input and output distributions of all variables and their residuals.
A series of scripts were used to evaluate and summarize the results from hyperparameter grid searches. They can be found in lth_thesis_project/jet_by_jet_compression/grid_search/.
Scripts starting with 002 were used to produce plots of different kinds.
Scripts starting with 003 were used to read through the results from a grid search and produce summarizing txt files as well as pickled python dictionaries for use in further analysis.
Script(s) starting with 004 use the results produced by running the corresponding '003' script to produce grid search summary plots.
Scripts starting with 010 produce plots histogram plots with output and input data for the 27 variable input AE suitable for presentation slides.
These were mostly trained "by hand" using jupyter notebooks located in jet_by_jet_compression/AE_3D_latent_space/, jet_by_jet_compression/AE_2D_latent_space/ and jet_by_jet_compression/fastai_AE_3D.
The AE presented in my thesis was trained using the notebook jet_by_jet_compression/fastai_AE_3D/fastai_AE_3D_200_no1cycle.ipynb.
Some grid searches were also run. Those can be found in jet_by_jet_compression/grid_search/.
These were mostly trained on LUNARC's Aurora and on HPC2N's Kebnekaise which are High Performance Computing (HPC) clusters located in Sweden. The scripts prepare_grid_search_aod.py and prepare_grid_search_aurora.py were used to create folders and training scripts with different combinations of hyperparameters as well as a bash script used to launch the the grid search on the HPC clusters. The scripts can be found in jet_by_jet_compression/grid_search/
In the notebooks where networks were trained they were also often evaluated, producing plots of input and output distributions of all variables and their residuals.
A series of scripts were used to evaluate and summarize the results from hyperparameter grid searches. They can be found in jet_by_jet_compression/grid_search/.
Scripts starting with 002 were used to produce plots of different kinds.
Scripts starting with 003 were used to read through the results from a grid search and produce summarizing txt files as well as pickled python dictionaries for use in further analysis.
Script(s) starting with 004 use the results produced by running the corresponding '003' script to produce grid search summary plots.
Scripts starting with 010 produce plots histogram plots with output and input data for the 27 variable input AE suitable for presentation slides.
The script 012_time_all.py encodes and decodes a dataset 100 times with different AEs and computes the average encoding and decoding time per jet.
A signal MC sample of a Dark Matter (DM) mediator decaying into jets was used for independent evaluation of the AEs. For this, the process_MC_data.ipynb notebook was used to preprocess the data.
In jet_by_jet_compression/grid_search/AE_bn_LeakyReLU_25AOD_grid_search_custom_normalization_1500epochs/ the script 000_plot_mjj.py prduces plots showing the dijet mass distribution before and after being fed through the AEs.