Creating and using a Variational Auto Encoder (VAE) for generating realistic time-series data. This is part of a FZI seminar at the Karslruher Institute of Technology.
One major division in machine learning, which has gained an incredible amount of popularity in a short time, is the modeling. It is all about learning any true data distribution using unsupervised learning in order to easily generate new data with some variations.
The Variational Autoencoder (VAE) is such a deep generative model and it has already achieved an incredible amount of success in generating many kinds of rather complicated data, including handwritten digits, faces, CIFAR (Canadian Institute For Advanced Research) images and sounds. In this project the scope of the VAE is expanded, combining the generative modeling power with the importance of time series. The aim is to generate realistic time series using a VAE. The performance of the proposed VAE model is evaluated on an electro-cardiogram (ECG) dataset.
- Seminar_VAE contains all python files
- Files with COLAB should be executed on Google Colab since they are very CPU/GPU heavy. Use Hardware Accelerator on Colab Notebooks (Edit > Notebook settings > Hardware Accelerator > GPU)