This repository is made as part of an assignment for the "Bayesian Learning" class of the University of Lille's Msc. in Data Science taught by Remi Bardenet.
Here, I am studying the pionneering paper of VAEs called "Auto-Encoding Variational Bayes" written by Diederik P. Kingma and Max Welling
In this paper, the authors propose a 'stochastic variational inference and learning algorithm that scales to large dataset' which corresponds to the first VAE model. In practice, the algorithm described in this paper can be applied to tasks of different kinds such as denoising, impainting and super-resolution. In this work, I treat exclusively the generative purpose since it's the most famous application of VAEs and it's also the kind of task chosen by the authors for their experiments.
This repo is made of 3 parts :
-
The article studied in a .pdf format
-
A short report discussing about the strategy proposed in the paper for solving the generation problem with some additional information enable to better understand it.
-
An illustrative notebook in which I propose an experiment enabling to save time in the design process of a VAE for a particular problem. This experiment is detailed in the report and the notebook.
If you want to test my implementation for the MAB-VAE you'll have to install the requirements listed in requirements.txt.
pip install -r requirements.txtA part of the code is inspired by the VAEs presented in this repo.
If you want to see more impressive implementations of VAEs in pytorch-lightning, I advise you to check it =)
Exemple of satisfying results obtained for the MNIST Dataset and the FashionMNIST Dataset :
