Kaggle Competition: G2Net Detecting Continuous Gravitational Waves

Problem Statement

The goal of this competition is to develop a model capable of detecting weak and long-lived continuous gravitational wave signals emitted by rapidly rotating neutron stars in noisy data. The contest aims to help scientists detect a second class of gravitational waves, which could lead to further understanding of the structure of the most extreme stars in the Universe. The contest is designed to help detect signals from a class of gravitational waves that have not yet been detected and that could potentially provide new insights in the field.

Data

You can get the download data from here. Also if you want to run my code you need to put the raw unpacked data in data/raw directory

• [train/|test/] - folders containing the training and test files, files are presented in hdf5, and contain SFT (Short-time Fourier Transforms), spectrograms obtained from LIGO Livingston and LIGO Hanford
• train_labels.csv - a file containing the target labels. 1 if the data contains the presence of a gravitational wave, 0 otherwise. Target label - 1 was ignored, because the files with this label are just a passcheck from the authors of the competition.
• sample_submission.csv - a sample submission file in the correct format.

Approach

As a baseline project I used Basic spectrogram image classification, and the main idea was to use the generation of new simulated data.

Usage

Clone repo

git clone https://github.com/bezbahen0/g2net

Install requirements

pip install -r requirements.txt

And run generation new data, training, and inference.

snakemake --cores all

To run another experiment, you can replace the path of the configuration file in Snakefile, or change the configuration file located in the configs directory

Results

#	Experiment	Coment	Backend	Input size	Private LB	Public LB
1	baseline	V3 data gen amplitued 20	tf_efficientnet_b7_ns	128*2	0.726	0.707
2	spectorgram	augmentations, amplitued 20	tf_efficientnet_b7_ns	128*2	0.745	0.721
3	spectrogram	amplitued 30, augmentations	tf_efficientnet_b7_ns	128*2	0.748	0.732
4	spectrogram	linear layer, amplitude 30, dropout-0.25, lr-0.00056	tf_efficientnet_b7_ns	128*2	0.750	0.721