gengli 0.0.1

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
• Colour the glitch with user-defined PSD, as well as with one of the analytical PSDs available in lalsuite
• 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.

How to install

Eventually a PyPI relase of the package will be made and the installation will be as simple as:

pip install gengli

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.

How to generate a glitch

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

import gengli
g = 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_raw_glitch()

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

The function get_glitch provides an easy to use API to the post-processing operations described above. For instance, to generate 4 glitches, coloured with the Einstein Telescope analytical PSD with an SNR of 10 and evaluate at a high sampling rate, you can just type:

g_coloured = g.get_glitch(4,
	srate = 16384,
	psd = 'EinsteinTelescopeP1600143',
	SNR = 10)

The generated glitches will look quite different from before:

If you want to use your favourite PSD, you can load it from a csv or xml file with:

from gengli.noise import read_psd
srate = 4096.
psd = read_psd('my_favourite_psd.xml', srate, g.get_len_glitch(srate),
	asd = False, ifo = 'H1', flow = 20.)

Once this is loaded you can pass it to the psd argument of get_glitch. Remember that the sampling rate should match that of the required glitches!

If you want to get an anomalous glitch, you can play with the confidence_interval parameter. For instance:

g_anomalous = g.get_glitch(10,
	confidence_interval = (90,100),
	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. Here there are the bibtex entries...