Conditional Wasserstein GAN with gradient penalty for the generation of synthetic time series.
Propriatary dataset was used to train the conditional WGAN with gradient penalty. Any dataset with shape (num_samples, num_features) will work.
The code was based on generating synthetic images, so whenever word 'images' appears, it should be interpreted as 'time-series' instead. I will fix this soon.
Run the cells in sequence in cwgan-gp_time_series.ipynb
jupyter notebook. Final cell contains code to create synthetic
image conditioned on a label.
Code in part based on:
- https://keras.io/examples/generative/conditional_gan/
- https://keras.io/examples/generative/wgan_gp/
- https://keras.io/examples/generative/dcgan_overriding_train_step/
- tensorflow
- numpy
- matplotlib
See requirements.txt
file.
Free to use for any purpose