DD2434-VAE-Project (Re-Autoencoding Variational Bayes)

Project in DD2434 Advanced Machine Learning to reproduce the paper title Auto-Encoding Variational Bayes. See the reproduced results in the report (some are also located at images/). The project was done by:

• Diogo Pinheiro
• Jakob Lindén
• Patrick Jonsson
• Márk Antal Csizmadia

Installation and Usage

In the CLI of your choice

git clone https://github.com/jakobGTO/DD2434-VAE-Project.git

Create a virtual environment (or in your base environment) with Anaconda or any virtual environment manager (see Anaconda example below):

conda create -n dd2434_vae_project python=3.8
source activate dd2434_vae_project
pip install -r requirements.txt

Use the config.yml configuration file to train a Variational Auto-Encoder (VAE), with desired encoder and decoder network parameters and decoder type (Bernoulli or Gaussian), training parameters (e.g.: learning rate) and the data set (MNIST or Frey face). Then run the main script VAE.py as follows:

cd DD2434-VAE-Project
python VAE.py

Results

Some of the results are shown below.

The two-dimensional latent space in the case of the MNIST dataset:

The two-dimensional latent space in the case of the Frey face dataset:

The variational lower bound in the case of the MNIST dataset with a 20-dimensional latent space:

The variational lower bound in the case of the Frey face dataset with a 20-dimensional latent space: