hjens / lasso_spectra

Functions for predicting properties from galaxy spectra, using Lasso regression

Package overview

lasso_spectra is a package for fitting Lasso regression models to data, specifically galaxy spectra. It contains two different classes for performing the actual model fitting. GeneralizedLasso is a tensorflow implementation of Lasso regression, which includes the ability to use link functions. SKLasso is a wrapper around the scikit-learn Lasso implementation intended to give the same syntax as GeneralizedLasso. It is much faster and more reliable, but does not support generalized linear models. Use SKLasso if unsure.

To get started, read this file and look at the included example.py. To learn more about the theory behind the algorithm, see the Theory section of Jensen et al 2016.

Package installation

Download the package and put it in your python path. Make sure you can import it:

import lasso_spectra

You need to have numpy, scipy and scikit-learn installed. For GeneralizedLasso you also need tensorflow (https://www.tensorflow.org).

Usage

The file example.py shows an example of how to use the package.

In short, there are three steps to using lasso_spectra:

1. Load your data
2. Fit a model
3. Use the model to predict values from new data

The first step can be done quickly using the classes SpectrumTableNarrow and SpectrumTableWide. These classes read csv files containing galaxy spectra and convert them into the format required for the model fitting. Use SpectrumTableNarrow if your data is in a narrow format, with one column per variable. Use SpectrumTableWide if your data contains one column for each wavelength bin (or other feature).

If you have data in other formats, you can subclass the class DataSource, or write your own data loading code from scratch. See the documentation for SKLasso to find out the correct values format of the data.

For the second step, you first need to create a model object, either of the class SKLasso or GeneralizedLasso. Then, you typically use the function fit_CV, passing in the training data and targets from your data source. This function will try many different values of the regularization parameter (called lambda in Jensen et al 2016, but referred to as alpha here to keep with the syntax of scikit-learn). Make sure to pass a wide range of alphas to the function to make sure it can find the one that gives the smallest cross-validation error. After the fitting is done, you may save the model using pickling.

Finally, you can use the model to predict values of your target (for example fesc) on new data. This is done using the predict function.