A pytorch implementation of Time-series Generative Adversarial Networks (https://github.com/jsyoon0823/TimeGAN)
The Goal was to create smoothed time series data via a GAN. This should be achieved via a combination of https://papers.nips.cc/paper/2019/file/c9efe5f26cd17ba6216bbe2a7d26d490-Paper.pdf and https://arxiv.org/pdf/2106.10414.pdf in PyTorch.
In this updated Version I was able to achieve a better recreation of the original time GAN implementation. With increasing size of epochs, the quality (evauluated on the third method of the time gan paper, namely pca and tsne plots) seems to increase. With epochs = 20 I was able to achieve these results, where we can clearly see that the spread out of snythetic and original data overlaps:
The demo notebook contains less epochs due to memory reasons.
Also an Ada FNN Layer was added to the Generator to try if the generated time Series will smooth out. This might (depending on epochs; higher number of epochs tend to produce smoother curves) be the case, however during some test runs without the ada fnn lyer included, smooth(er) curves were also generated, so it is hard to say how beneficial the ada fnn layer is in fact. See the demo notebook for visualizations.
General Note: The Training is quite unstable and results seem to be impacted quite strongly by the initialization as well as learning rate, drop out rate and number of epochs.
requirements.txt contains all dependencies and can be run with pip:
pip install -r requirements.txt
utils.py contains all helper functions, mostly from the original repository
modules_and_training contains the main implementation of TimeGAN. The Network blocks are defined there, aswell as a function that runs training and returns the trained networks.
Demo_Time_GAN_and_ADA_fnn includes a demo run of the time gan without ada fnn layer as well as a demo run with a ada fnn layer.
Tests.py contains test that were written to ensure correct functioning of each module.