Deep-Autoencoders-Data-Compression-GSoC-2021

ML data compression of ATLAS trigger jet events using various deep autoencoders, with PyTorch and fastai python libraries.

This repository is developed by George Dialektakis, as a Google Summer of Code (GSoC) student

Setup

Running the code

Data extraction

Project Structure Description

Setup

First, clone the latest version of the project to any directory of your choice:

git clone https://github.com/Autoencoders-compression-anomaly/Deep-Autoencoders-Data-Compression-GSoC-2021.git

Install dependencies:

pip3 install -r requirements.txt

Running the code

usage: python main.py [--use_vae] [--use_sae] [--l1] [--epochs] [--custom_norm]
                      [--num_variables] [--plot]

optional arguments:
  --use_vae            whether to use Variational AE (default: False)
  --use_sae            whether to use Sparse AE (default: False)
  --l1                 whether to use L1 loss or KL-divergence in the Sparse AE (default: True)
  --epochs             number of epochs to train (default: 50)
  --custom_norm        whether to normalize all variables with min_max scaler or also use custom normalization for 4-momentum (default: False)
  --num_variables      number of variables we want to compress (either 19 or 24) (default: 24)
  --plot               whether to make plots (default: False)

Example:

python main.py --use_sae True --epochs 30 --num_variables 19 --plot True

The above command will train the Sparse Autoencoder for 30 epochs to compress the 19D data and will make plots of the input and preprocessed data.

Data extraction

The data that were used for this project can be downloaded from CERN Open Data Portal. The file that was used is: 00992A80-DF70-E211-9872-0026189437FE.root under the filename CMS_Run2012B_JetHT_AOD_22Jan2013-v1_20000_file_index.txt. The data can then be loaded with data_loader(), which produces a pandas dataframe from the ROOT file.

Project Structure Description

• data_loader.py loads the data from a ROOT file and creates a pandas dataframe.

• data_processing.py makes all the necessary preprocessing steps for our data (filtering, normalization, train-test split)

• create_plots.py holds all the necessary functions to plot the initial, preprocessed and reconstructed data

• autoencoders/ holds the implementation of three different Autoencoder types we considered (Standard AE, Sparse AE, Variational AE)

• evaluate.py performs the evaluation of the autoencoder on the test data in terms of MSE, RMSE loss, and residuals

• main.py is the main script which runs the whole code

You can find more details about this project as well as the experimental analysis and the results in report.pdf