TadGAN: Time Series Anomaly Detection Using Generative Adversarial Networks

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

Data:

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

Pretrained Model:

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

Architecture:

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.

Usage:

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.

Note:

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