ml4astro / galaxy2galaxy

Library of models, datasets, and utilities to build generative models for astronomical images.
MIT License
27 stars 7 forks source link

Galaxy2Galaxy Build Status Documentation Status Join the chat at https://gitter.im/ml4astro/galaxy2galaxy

Galaxy2Galaxy, or G2G for short, is a library of models, datasets, and utilities to build generative models for astronomical images, based on the Tensor2Tensor library. Similarly to T2T, the goal of this project is to accelerate research in machine learning models applied to astronomical image processing problems.

Current features:

Note that G2G is still under development, additional feature/dataset suggestions are most welcome, all contributions are most welcome. Don't hesitate to join the gitter room for any questions!

Install

We recommend users create a conda environment before installing galaxy2galaxy. This makes installing tensorflow and galsim very easy:

$ conda install tensorflow-gpu==1.15
$ conda install -c conda-forge galsim

G2G can then easily be installed using pip inside the environment:

$ pip install git+https://github.com/ml4astro/pixel-cnn.git 
$ pip install git+https://github.com/ml4astro/GalFlow.git
$ pip install galaxy2galaxy

Usage

To generate the COSMOS 25.2 sample at native pixel scale and stamp size:

$ g2g-datagen --problem=img2img_cosmos --data_dir=data/img2img_cosmos

This uses GalSim to draw postage stamps and save them in TFRecord format which can then be used for training. This assumes that you have downloaded the GalSim COSMOS sample, if that's not the case, you can dowload it with: galsim_download_cosmos -s 25.2

To train an autoencoder with this data:

$ g2g-trainer --data_dir=data/img2img_cosmos --output_dir=training/cosmos_ae   --problem=img2img_cosmos --model=continuous_autoencoder_basic  --train_steps=2000  --eval_steps=100 --hparams_set=continuous_autoencoder_basic

Publications making use of Galaxy2Galaxy

If you make use of Galaxy2Galaxy for your work, don't hesitate to let us know, and we will gladly add you to this list.