cosmoGAN

This code is to accompany "Creating Virtual Universes Using Generative Adversarial Networks" manuscript arXiv:1706.02390. The architecture is an implementation of the DCGAN architecture (arXiv:1511.06434).

---

How to train:

git clone [email protected]:MustafaMustafa/cosmoGAN.git
cd cosmoGAN/networks
mkdir data
wget http://portal.nersc.gov/project/dasrepo/cosmogan/cosmogan_maps_256_8k_1.npy
cd ../

That will download sample data (8k maps) for testing. You can download more data from here. All of this data has been generated using our GAN and can be used to train your own. Original data is available upon request from the authors.

To run:

python run_dcgan.py

Load pre-trained weights:

First download the weights:

cd cosmoGAN/networks
wget http://portal.nersc.gov/project/dasrepo/cosmogan/cosmoGAN_pretrained_weights.tar
tar -xvf cosmoGAN_pretrained_weights.tar

Then take a look at networks/load_and_use_pretrained_weights.ipynb notebook for how to run.