This is a reference implementation of a style-based generative adversarial network (GAN) designed to generate geophysical fields. This code supports a paper published in the Proceedings of the 9th International Workshop on Climate Informatics: CI 2019. The paper was selected for the Best Paper Award at the meeting.

You might also be interested in this paper (code) or in my post on GANs in the atmospheric sciences.

Download the data from https://doi.org/10.7910/DVN/ZDWWMG and follow the instructions there.

Get the trained weights from this release. Unzip the file, preferably into the models directory.

For training, you'll want a machine with a GPU and around 32 GB of memory (the training procedure for the radar dataset loads the entire dataset into memory). Running the pre-trained model should work just fine on a CPU.

You may want to work with the code interactively; in this case, just start a Python shell in the geogan directory.

If you want the simplest way to run the code, the following two options are available. You may also want to look at what main.py does in order to get an idea of how the training flow works.

You can replicate the plots in the paper by going to the geogan directory and using

python main.py plot --data_file=<data_file> --weights_root=<weights_root>

where <data_file> is the path to the training data and <weights_root> is the path and prefix of the stored weights. For example, if you unzipped the pre-trained weights to the models directory, you should use --weights_root=../models/mch_gan.

Run the following to start the training:

python main.py train --data_file=<data_file> --weights_root=<weights_root>

where the parameters are as above. This will run the training loop until terminated (with e.g. ctrl-C) and save the weights after each 100 batches (overwriting the original weights, so save the weights under a different name if you want to avoid this).