Tobi2K / SeminarNI

A project about autoencoders for a seminar paper at Ulm University

Seminar Neuroinformatik

This project was created as a part of a seminar work in my bachelor studies. It consists of two notebooks that train autoencoders on different criteria. Feel free to try it out.

Getting started

$ git clone https://github.com/Tobi2K/SeminarNI.git
$ cd SeminarNI
$ pip install -r requirements.txt

Configuring the trials

In the Autoencoder Notebook you can select which comparisons to run by setting the corresponding variables to True.

The possible comparisons are:

• difference in number of autoencoder runs aka. epochs
• the size of the dense layer
• the number of dense layers
• added noise vs no noise

In the Denoising Autoencoder Notebook you can run a set of denoising autoencoders.

Show loss trend

The loss of each epoch is tracked and can be displayed with TensorBoard.

# start TensorBoard
$ tensorboard --logdir=/path/to/project/SeminarNI/logs

The comparisons save their logs in separat folders, making separation easy:

# start TensorBoard but only show number of runs comparison
$ tensorboard --logdir=/path/to/project/SeminarNI/logs/epochs

# start TensorBoard but only show size of dense layer comparison
$ tensorboard --logdir=/path/to/project/SeminarNI/logs/layer_size

# start TensorBoard but only show number of dense layer comparison
$ tensorboard --logdir=/path/to/project/SeminarNI/logs/layer_count

# start TensorBoard but only show noise vs no noise comparison
$ tensorboard --logdir=/path/to/project/SeminarNI/logs/compare_noise

Preexisting data

There already are existing images, log-texts as well as TensorBoard logs contained in this repository.

When running the configurations, only TensorBoard logs are saved automatically. Images and log-texts have to be saved manually.