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