Skip to content

Differentiable strong lens modelling on CPUs and GPUs

License

Notifications You must be signed in to change notification settings

Herculens/herculens

Repository files navigation

Herculens logo

Herculens: differentiable gravitational lensing

PyPi python support Tests arXiv License

Analysis of strong lensing imaging data

The primary purpose of herculens is to provide flexible modeling methods to model current and future observations of strong gravitational lenses. Currently, it supports various degrees of model complexity, ranging from standard smooth analytical profiles to pixelated models combined with machine learning approaches.

Currently, herculens supports several of the most widely-used analytical profiles, as well as multi-scale pixelated models regularized with wavelets. Future updates will include the support of point source modeling, new regularization techniques, and more expressive models based on neural networks.

JAX-based automatic differentiation and code compilation

herculens is based on the powerful framework of differentiable programming. The code is entirely based on the automatic differentiation and compilation features of JAX. This simply means that you have access, analytically, to all partial derivatives of your model with respect to any of its parameters. This enables faster convergence to the solution, more efficient exploration of the parameter space including the sampling of posterior distributions, and new ways to mitigate degeneracies that affect gravitational lensing.

This highly modular framework offers a way to merge all modeling paradigms explored in the literature, into a single tool:

  • analytical: model components are described by analytical functions with few parameters and clear physical meaning, but that may be insufficient to fit all observations;
  • pixelated: regular or irregular grid of pixels are used as individual parameters, which offer higher flexibility, but requires well-motivated regularization strategies;
  • deep learning: neural networks (among others) are by construction fully differentiable, regardless of being pre-trained or not. It is therefore effortless to plug-in any deep learning-based model component to herculens.

Example notebooks

Below you can find a few examples of synthetic problems that can be tackled with herculens:

New notebooks will continuously be added in order to showcase other features of the package and apply those in different situations.

Installation

Manual installation

The package will be soon available through PyPi directly, but it is as easy to install it manually. It has been texted against Python 3.7, but should work with Python 3.8 or more recent versions.

Good practice is to create a new python environment:

conda create -n herculens-env python=3.7
conda activate herculens-env

Download the package cd into the directory. Then install the local herculens package (or use -e for a development install) as follows:

pip install (-e) .

The following dependencies will be installed automatically.

External package requirements

The requirements.txt file lists all required and optional package dependencies, along with their specific versions.

Attribution

Citation

If you make use of herculens, please cite Galan et al. 2022. This work presents the global architecture of the tool, and applies it to reconstruct a pixelated lens potential with wavelets. See CITATION.md for the bibtex entry.

Special mention regarding Lenstronomy

Part of the herculens code originates from the open-source lens modeling software package lenstronomy, described in Birrer et al. 2021 (and references therein). In every source file, proper credits are given to the specific developers and contributors to both the original lenstronomy (up to version 1.9.3) and herculens.

Contributors

The list of people that contributed to Herculens and credits to original lenstronomy contributors, is in this document.

About

Differentiable strong lens modelling on CPUs and GPUs

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages