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:
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
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: