This package uses Gaussian Process Regression (GPR) as a foreground removal technique in the context single-dish 21cm intensity mapping. This user-friendly code shows you how to do this in the context of MeerKAT-like simulations, but any intensity mapping data in real space can be used. This is the accompaying code to the paper (https://arxiv.org/abs/2105.12665)
, where we look at how GPR performs as a foreground removal technique in our simulations in comparison with Principal Component Analysis.
To install this package, follow these instructions on a terminal:
pip install gpr4im
or if you prefer:
git clone https://github.com/paulassoares/gpr4im.git
cd gpr4im
pip install .
If using the second option, make sure you do pip install .
in the gpr4im folder, where the setup.py
file is.
Installing gpr4im
will also automatically install:
numpy
matplotlib
pandas
GPy
(see https://github.com/SheffieldML/GPy)scipy
getdist
astropy
jupyter
It will not install pymultinest
, which is required for the Nested sampling.ipynb
notebook. If you would like to run that notebook, please see http://johannesbuchner.github.io/PyMultiNest/install.html for details on installation.
An very quick example of how to run GPR foreground removal using our code is shown below, but please see the Jupyter notebooks for further explanation:
import pandas as pd
import GPy
from gpr4im import fg_tools as fg
data = pd.read_pickle('example_data.pkl')
dirty_map = data.beam.FGnopol_HI_noise
kern_fg = GPy.kern.RBF(1)
kern_fg.variance.constrain_bounded(1000,100000000)
kern_fg.lengthscale.constrain_bounded(200,10000)
kern_21 = GPy.kern.Exponential(1)
kern_21.variance.constrain_bounded(0.000001,0.5)
kern_21.lengthscale.constrain_bounded(0.01,15)
gpr_result = fg.GPRclean(dirty_map, data.freqs, kern_fg, kern_21,
NprePCA=0, num_restarts=10, noise_data=None,
heteroscedastic=False, zero_noise=True, invert=False)
cleaned_map = gpr_result.res
For a quick introduction on how to run the code, please see Running GPR.ipynb
. For a more thorough run through of how the code works, please see Understanding GPR.ipynb
. The Jupyter Notebooks folder contains other introductory notebooks for how all the aspects of our code and data work, and are all user friendly. These use the data set example_data.pkl
, which is described in the Data folder's README.
The Reproducible paper plots folder contains the notebooks showing how we obtained the analysis results for our companion paper (these are less introductory, but useful for those trying to understand how our analysis was done). The code here requires the multinest_results.pkl
file, as well as the full data used in our analysis, data.pkl
, which can be obtained from this link (but beware, it is 2.84 GB):
https://www.dropbox.com/sh/9zftczeypu7xgt3/AABiiBw_0SBPrLgSHsjiISz8a?dl=0
The Nested sampling.ipynb
notebook also uses this data, and requires pymultinest
to be installed.
If you make use of this code, please cite:
@article{2021,
title={Gaussian Process Regression for foreground removal in H i Intensity Mapping experiments},
volume={510},
ISSN={1365-2966},
url={http://dx.doi.org/10.1093/mnras/stab2594},
DOI={10.1093/mnras/stab2594},
number={4},
journal={Monthly Notices of the Royal Astronomical Society},
publisher={Oxford University Press (OUP)},
author={Soares, Paula S and Watkinson, Catherine A and Cunnington, Steven and Pourtsidou, Alkistis},
year={2021},
month={Sep},
pages={5872–5890}
}
This code is heavily based on the publicly available ps_eor
code (https://gitlab.com/flomertens/ps_eor), so if you use our code please also acknowledge:
@article{Mertens2018,
title={Statistical 21-cm Signal Separation via Gaussian Process Regression Analysis},
ISSN={1365-2966},
url={http://dx.doi.org/10.1093/mnras/sty1207},
DOI={10.1093/mnras/sty1207},
journal={Monthly Notices of the Royal Astronomical Society},
publisher={Oxford University Press (OUP)},
author={Mertens, F G and Ghosh, A and Koopmans, L V E},
year={2018},
month={May}
}