Pytorch implementation of the method used in "Generation of non-stationary stochastic fields using Generative Adversarial Networks" https://arxiv.org/abs/2205.05469.
The work is an extenstion of the previous work https://github.com/Alhasan-Abdellatif/cGANs where spatial conditioning is implemented using the SPADE algorithm https://github.com/NVlabs/SPADE to generate geological samples conditioned on soft data.
- Python 3.8.10
- PyTorch 1.12.1
- NumPy
- matplotlib
- torchvision
To run conditional GAN using SPADE algorithm on images in datasets/images/ and their labels (soft probabilities e.g., 4x4) save models in results :
python train.py --data_path datasets/images --labels_path datasets/labels/ --data_ext txt --img_ch 1 --zdim 128 --spec_norm_D --x_fake_GD --y_real_GD --n_cl 1 --cgan --G_cond_method conv1x1 --D_cond_method conv1x1 --batch_size 32 --epochs 100 --smooth --save_rate 10 --ema --dev_num 1 --att --fname results
The sample notebook provides some examples on how to use the trained models to genereate different geological facies controlled spatailly by a soft map input.