This is a Python3 / Pytorch implementation of TadGAN paper. The associated blog explaining the architecture details can be found here.

The TadGAN architecture can be used for detecting anomalies in time series data.

The trained model is saved in the Model directory. The training is incomplete and the model has to be retrained for other datasets.

The model implements an encoder and decoder as generator and two critics as discriminators as described in the paper. The loss function is wasserstein loss with gradient penalty.

1. Format of the dataset - The dataset should have a column name as signal containing the signals and a column with name anomaly containing the true labels (used during validation).

2. Delete the contents of the directory Model.

3. Change the file name exchange-2_cpc_results.csv in main.py to the name of your dataset.

This is an independent implementation and I am not related to the authors of the paper.