carl
is a toolbox for likelihood-free inference in Python.
The likelihood function is the central object that summarizes the information from an experiment needed for inference of model parameters. It is key to many areas of science that report the results of classical hypothesis tests or confidence intervals using the (generalized or profile) likelihood ratio as a test statistic. At the same time, with the advance of computing technology, it has become increasingly common that a simulator (or generative model) is used to describe complex processes that tie parameters of an underlying theory and measurement apparatus to high-dimensional observations. However, directly evaluating the likelihood function in these cases is often impossible or is computationally impractical.
In this context, the goal of this package is to provide tools for the likelihood-free setup, including likelihood (or density) ratio estimation algorithms, along with helpers to carry out inference on top of these. It currently supports:
- Composition and fitting of distributions;
- Likelihood-free inference from classifiers;
- Parameterized supervised learning;
- Calibration tools.
This project is still in its early stage of development. Join us if you feel like contributing!
-
Extensive details regarding likelihood-free inference with calibrated classifiers can be found in the companion paper "Approximating Likelihood Ratios with Calibrated Discriminative Classifiers", Kyle Cranmer, Juan Pavez, Gilles Louppe. https://arxiv.org/abs/1506.02169
The following dependencies are required:
- Numpy >= 1.11
- Scipy >= 0.17
- Scikit-Learn >= 0.18-dev
- Theano >= 0.8
Once satisfied, carl
can be installed from source using the following
commands:
git clone https://github.com/diana-hep/carl.git
cd carl
python setup.py install
See CONTRIBUTING.md for setup instructions to start
developing and contributing to carl
.
@misc{carl,
author = {Gilles Louppe and Kyle Cranmer and Juan Pavez},
title = {carl: a likelihood-free inference toolbox},
month = mar,
year = 2016,
doi = {10.5281/zenodo.47798},
url = {https://dx.doi.org/10.5281/zenodo.47798}
}