This repository contains a python package and scripts accompanying the paper "Weak-lensing shear measurement with machine learning: Teaching artificial neural networks about feature noise" (Tewes et al. 2018, arXiv:1807.02120, paper website). Together with the neural network package tenbilac (and a decent computing server) this code allows to reproduce the results from our paper. Note that to experiement with MomentsML on your own data or simulations, much more coding would certainly be required, as the present software really is a demonstration toolbox and not a pipeline.
The code depends on:
- python 2.7, with numpy, scipy, and matplotlib
- GalSim
- Astropy
- tenbilac (for the Machine Learning part)
After installing the above dependencies, we recommend to simply add the location of your clone of this directory to your PYTHONPATH.
To do so, if you use bash, add this line to your .bash_profile or .profile or equivalent file:
export PYTHONPATH=${PYTHONPATH}:/path/to/momentsml/
You should then be able to import momentsml.
- momentsml: the python package
- scripts: scripts and configurations to reproduce the results from our paper
- wrappers: standalone packages providing interfaces to specific datasets
- other: additional figures and demonstration scripts
The CPU and disk space requirements vary widely between the different applications described in the paper. For first experiments similar to what is shown in Section 6 of the paper, a machine with 16 or more cores and a few hundred GB of disk space should be sufficient.
To begin, add the tenbilac and momentsml packages to your PYTHONPATH as described above (for momentsml) and here (for tenbilac). If you plan to process GREAT3 data, do the same with the momentsmlgreat3 wrapper, as described in wrappers/.
You could check that import momentsml and import tenbilac succeed.
The scripts and configuration files for the different sections of the paper are in scripts/ (click to see an overview). The structure of these scripts for the different sections is very similar, but we strongly suggest to start with the "fiducial" scripts, as those are the most documented.
Make a personal copy of the directory scripts/fiducial/ somewhere else on your system (so that you can adapt paths and settings), and proceed by following the included README.