Detecting structured signals in radio telescope data using RKHS

Data

Download the single beam data from SPARKESX > Files > sbeam > all folders.

Running the code

1. We use an open source implementation of kernel logistic regression (klr) https://github.com/RussellTsuchida/klr. Unzip this directory, cd into it and install using python -m pip install .
2. Install the rest of the dependencies using python -m pip install -r requirements.txt.
3. Depending on your setup, you may wish to run the code differently.
4. We ran our expierments on a cluster that uses the SLURM job manager. To run code using SLURM, point to the location of the data on line 14 of run.sh. Then us sbatch run.sh.
5. If you want to manage individual python instances separately for each .sf file, you can run python runalgo_klr.py FNAME, where FNAME is the absolute path to the .sf file.

The outputs of the program

The verbosity of results that are saved is controlled by the VERBOSITY parameter on line 24 of runalgo_klr.py.

1. If VERBOSITY <= 0. Results are written to a scores.npy file for each .sf file. These scores represent the power of each chunk as a function of time, and are also plotted in scores.png.
2. If VERBOSITY <= 1. The function f of the RKHS (i.e. the logits of the bernoulli distribution) are plotted for each chunk of data. These are called _t.png, where t is the chunk index.
3. If VERBOSITY <= 2. The raw data itself is plotted in t.png, where t is the chunk index.

Collecting the outputs of the program

1. The file post_process.py and associated SLURM script post_process.sh put the scores in a predictions.csv file. This file associates to every time chunk a 0 or a 1, depending on whether an event is observed or not.
2. Given such post-processed predictions.csv files, collate_results.py produces the figures given in the paper.