This Python 2.7 package contains the (1) spatial generative adversarial network (SGAN) and (2) DREAM_(ZS) Markov chain Monte Carlo (MCMC) sampler used for multiple-point statistics simulation and inversion by Laloy et al. (2018)

Now with:

- a Pytorch 1.1.0 / Python 3.6 GAN-based MCMC inversion example for the 2D binary channelized aquifer case

- a Pytorch 1.1.0 / Python 3.6 version of the 2D braided river case (with Wasserstein GAN training)

Because of Github space limits, the trained SGAN models (needed to reproduce the results presented in the paper) are not provided herein. But these are of course available. Please drop me an email if you want to get these files.

We ported the original 2D SGAN code by Jetchev et al. (2016) to 3D. The codes for training the SGAN with Theano/Lasagne are contained in the SGAN folder. Also, 2D and 3D geologic model realizations generated by the trained models can be produced by running the generation2D.py and generation3D.py scripts.

Note that we have a Pytorch 1.1.0 / Python 3.6 example version available (see this repo for the GAN-based MCMC inversion and the gan_for_gradient_based_inv repo for the training code).

To perform the MCMC inversion within the SGAN latent space using DREAM_(ZS), use the run_mcmc.py script from the inversion folder.

Laloy, E., Hérault, R., Jacques, D., & Linde, N. (2018). Training-image based geostatistical inversion using a spatial generative adversarial neural network. Water Resources Research, 54. https://doi.org/10.1002/2017WR022148

The Theano/Lasagne and Pytorch codes for the SGAN are under MIT license while the MCMC inversion code is under GPL license. See the corresponding folders for details.

Eric Laloy ([email protected])