gengli 0.2.0

Glitch generation with Generative Adversarial Network

gengli

gengli (generating glitches) is a package to provide an handy an easy to use interface to CTGAN, a generative adversarial neural network to generate glitches. A "glitch" is a non stationary noise transient recorded in Gravitational Waves (GW) interferometer. As it resembles some GW signals, their appearece hinders the searches for such signal, downgrading the instrument sensitivity. A nice characterization of glitches can enable lots of exciting GW science!

This network is trained with real (although heavily preprocessed) glitch and when given some random noise, it outputs a whithened glitch. gengli offers a simple API to access the network, enabling some post-processing of the generated glitch that can be useful for downstream analysis. In more, details it allow the user to:

• Resample the glitch
• Lowpass (i.e. removing spurious high frequency content) and window
• Scale it to achieve a target SNR
• Select among random glitches only those within a given percentile w.r.t. to some anomaly score (more on the paper). This enables the user to get only "normal" glitches and/or "atypical" glitches.

You can read the full documentation here!

How to install

To install the latest released version of gengli, available on PyPI:

pip install gengli

From the repo

If you want to install the code from this repo, you can use the Makefile provided:

git clone https://git.ligo.org/melissa.lopez/gengli.git
cd gengli
make install

If you want to build a local version of the documention, just type make docs. This will create the folder docs/__build/ where all the relevant html docs will be located.

With a conda environment

If you don't want to mess up with your existing conda installations, this will create a fresh new conda environment that only contains gengli and its dependencies:

conda create --name gengli_env python=3.9
conda activate gengli_env
git clone https://git.ligo.org/melissa.lopez/gengli.git
cd gengli
python setup.py sdist
pip install dist/gengli-0.1.0.tar.gz 

A ready made conda environment is also provided in env.yml. To use it just type:

conda env create --file env.yml

This will create an environment with all the dependencies (and gengli installed)

How to generate a glitch

To generate a glitch, you will need to instantiate a glitch_generator object:

import gengli
g = gengli.glitch_generator('L1')

You should provide the generator the weights of the generator network. Some default (and reviewed) weights are relaesed with this package: if you want to use them, you should just pass the string L1 or H1 to specify which interferometer you want to consider and the weights will be loaded authomatically.

To generate the raw glitch (as output by the network):

g_raw = g.get_glitch()

The glitch is whithened and evaluted on a default time grid sampled at 4096Hz. It has a characteristic shape:

The function get_glitch also provides an easy to use API to the post-processing operations described above. For instance, to generate 10 glitches, filtered and windowed with an SNR of 15 and evaluated at a lower sampling rate, you can just type:

g_processed = g.get_glitch(10,
	srate = 2048,
	snr = 15,
	alpha = 0.2,
	fhigh = 250)

The generated glitches will look quite different from before:

If you want to get three anomalous glitches, you can play with the percentiles parameter of get_glitch_confidence_interval. For instance:

g.get_glitch_confidence_interval(percentiles = (0,80),
	n_glitches = 3,
	snr = 20)

This will take a while to run as it will call internally (only once) glitch_generator.initialize_benchmark_set, which creates a set of benchmark glitches to compute the "anomaly score" against. To have more controls of the parameters of initialize_benchmark_set, you can call it from your script: you will note that, once such initialization is done, get_glitch will run much faster!

You can take a look at some examples in a dedicated folder.

For more information, you can read the docs.

About

This project is developed in the GW group at Utrecht Univerisity. Many people were involved in this project:

• Vincent Boudart
• Kerwin Buijsman
• Sarah Caudill
• Melissa Lopez (corresponding author: [e-mail])
• Amit Reza
• Stefano Schmidt (package maintener: [e-mail])

If you want to have more information, or just to say hello, please feel free to contact any of us!

Publications

If you find our work useful, please consider reading and citing our publications.

The main work, detailing the method and its performances, is the following:

@article{Lopez:2022lkd,
    author = "Lopez, Melissa and Boudart, Vincent and Buijsman, Kerwin and Reza, Amit and Caudill, Sarah",
    title = "{Simulating transient noise bursts in LIGO with generative adversarial networks}",
    eprint = "2203.06494",
    archivePrefix = "arXiv",
    primaryClass = "astro-ph.IM",
    doi = "10.1103/PhysRevD.106.023027",
    journal = "Phys. Rev. D",
    volume = "106",
    number = "2",
    pages = "023027",
    year = "2022"
}

A follow up paper describes gengli and its features, providing more details on the network architecture and training:

@article{Lopez:2022dho,
    author = "Lopez, Melissa and Boudart, Vincent and Schmidt, Stefano and Caudill, Sarah",
    title = "{Simulating Transient Noise Bursts in LIGO with gengli}",
    eprint = "2205.09204",
    archivePrefix = "arXiv",
    primaryClass = "astro-ph.IM",
    month = "5",
    year = "2022"
}