Light Curve Classifier Library

Installing late_classifier

git clone https://github.com/alercebroker/lc_classifier.git
python -m pip install -r requirements.txt
python -m pip install -e .

Some packages like matplotlib and tqdm are necessary to run the examples.

How to use the library?

Check the available Jupyter notebooks in the examples directory.

• Feature extraction with ZTF light curves.

Functionalities

Feature computation

This library provides an extensive number of feature extractors for astronomical light curves, including period computation, autoregresive models, parametric models, statistical features, etc. We also provide tools to transform your data into the format that this library expects (Pandas dataframes).

Augmentation

If you want more samples you can use our ShortTransientAugmenter class. More data augmentation techniques will be implemented in further releases.

Classifier

Two classifiers are available: A traditional Random Forest model, and a hierarchical model made from 4 Random Forest classifiers.

Preprocessing for ZTF data:

Before computing features, we preprocess the time series with filters and boundary conditions:

• Drop duplicate observations.
• Discard noisy detections.
• Discard bogus.
• Filter time series with more than 5 detections.
• Discard invalid values (like nans and infinite).

How can I add my own feature extractors to the library?

Feature extractors extend the following classes:

• FeatureExtractor
• FeatureExtractorSingleBand. This type of feature extractor runs independently on each available band.

Check out the existent feature extractors in the directory lc_classifier/features/extractors.

Profile functionalities

The easiest way to profile a step is using cProfile, for this we just have to run the step with the following command:

python -m cProfile -o <outputfile> profiling/<functionality>.py

After that you can run snakeviz (first install it).

snakeviz <outputfile>

Test functionalities

You must first install the following packages:

pip install coverage pytest

All scripts of tests must be in tests folder. For run all tests:

coverage run --source lc_classifier -m pytest -x -s tests/

If you want run a specify functionality you can run:

coverage run --source lc_classifier -m pytest -x -s tests/<functionality>

After that you can see a report of tests:

coverage report

Run a container

This repository comes with a Dockerfile to test the model.

To build the image run

docker build -t alerce/lc_classifier

Then run the container

docker run --rm -p 8888:8888 alerce/lc_classifier

The container comes with a jupyter notebook and some examples in http://localhost:8888

Reference

If you use this library, please cite our work:

@inproceedings{sanchez2020alert,
  title={Alert Classification for the ALeRCE Broker System: The Light Curve Classifier},
  author={S{\'a}nchez-S{\'a}ez, P and Reyes, I and Valenzuela, C and F{\"o}rster, F and Eyheramendy, S and Elorrieta, F and Bauer, FE and Cabrera-Vives, G and Est{\'e}vez, PA and Catelan, M and others},
  year={2020}
}