The effect of phased recurrent units in the classification of multiple catalogs of astronomical lightcurves

This repository contains the implementation of the light curve classifier (L+P) described in C. Donoso-Oliva et al., 2021.

NOTE: A new version of the L+P is being developed at this link. It will contain new experimental features and code best practices. (not yet evaluated)

Requirements

The faster way to run the code is creating a conda environment as follows:

conda env create -f environment.yml

the first line on the YAML file defines the name of the environment. Feel free to change it.

Optionally, you can use the requirements.txt file to install dependencies using pip

pip install -r requirements.txt

Folders and files

./models/: This folder includes the LSTM and Phased LSTM model class.
- plstm.py: Phased LSTM classifier following the architecture explained in the paper.
- lstm.py: LSTM classifier following the architecture explained in the paper.
- ./layers/: Custom layers used on this work
  - phased.py: Phased LSTM unit. It consists in a LSTM + time gate
data.py: Contains relevant function for loading and creating records.
get_data.py: Script to download data
main.py: Main script that loads the data and instances the models (use --help to see running options) for training
predict: Prediction script which receive 4 (sys.argv) arguments:
- dataset : record path
- rnn_type : either plstm or lstm
- fold_n : according to our kfold
- norm : nomalization technique -i.e., n1 or n2
{train, test}_script.py: Code routines for hyperparameter tuning

Additional not included folders

For storage reasons we do not upload experiments and results folders which contain adjusted models and metrics/figures, respectively. Although they are not in this repo, you can download them using the experiments link or the results link. Once downloaded paste the folders into the root directory.

Downloading data

You can download data in their raw format or preprocessed record. Getting records implies using the same folds as the author.

Commands

python get_data.py --help: to see script options.
python get_data.py --dataset <name>: to download raw dataset.

e.g. Downloading LINEAR records,

  python get_data.py --dataset linear --record

Alternatively, you can download directly from google drive

Creating your own data

For custom training, please convert your data to tf.Record using the standard function create_record(light_curves, labels, path='') in ./data.py