ASTROMER

Welcome to the main repository of the ASTROMER project. This is the stable version 0 where you can find all the resources associated to the article: ASTROMER: A transformer-based embedding for the representation of light curves.

ASTROMER is a transformer-based model that learns light curves representations using millions of light curves.

Representation are then used for extracting useful embeddings that we can use for training another downstream tasks.

Features

• Pre-trained weights from MACHO R-band light curves
• Visualization jupyter notebooks (/presentation/notebooks/*)
• Scripts for training, finetuning and classify using ASTROMER (/presentation/scripts/*)
• Model implementation (/core/astromer.py and related)
• Data preprocessing, saving and reading tf.Records (/core/data.py)
• Dockerfile and scripts for building (build_container.sh) and run (run_container.sh) the ASTROMER container

Get started

We recomend to use Docker since it provides a kernel-isolated and identical environment to the one used by the authors

The Dockerfile contains all the configuration for running ASTROMER model. No need to touch it, build_container.sh and run_container.sh make the work for you 🙂

The first step is to build the container,

  bash build_container.sh

It creates a "virtual machine", named astromer, containing all the dependencies such as python, tensorflow, among others.

The next and final step is running the ASTROMER container,

  bash run_container.sh

The above script looks for the container named astromer and run it on top of your kernel. Automatically, the script recognizes if there are GPUs, making them visible inside the container.

By default the run_container.sh script opens the ports 8888 and 6006 for jupyter notebook and tensorboard, resepectively. To run them, use the usal commands but adding the following lines:

For Jupyter Notebook

jupyter notebook --ip 0.0.0.0

(Optionally) You can add the --no-browser tag in order to avoid warnings.

For Tensorboard

tensorboard --logdir <my-logs-folder> --host 0.0.0.0

Finally, if you do not want to use Docker the requirements.txt file contains all the packages needed to run ASTROMER. Use pip install -r requirements.txt on your local python to install them.

Usage/Examples

We recomend to save data in tf.Records format. For creating records jump to the create records tutorial

Otherwise, if you have numpy-based light curves, use load_numpy() function.

For pre-training run train.py

python -m presentation.scripts.train --data ./data/records/macho

To see the --tag options use --help. For example,

python -m presentation.scripts.finetuning --help

The pre-training, finetuning, and classification scripts work in the same way.

Contributing

Contributions are always welcome!

Issues and featuring can be directly published in this repository via Pull Request. Similarly, New pre-trained weights must be uploaded to the weights repo using the same mechanism.

Look at this tutorial for more information about pull requests