Description of the theory of variational autoencoders as described in the paper Auto-Encoding Variational Bayes [1] and implementation on the CelebA dataset [2].
The notebook includes:
- Useful resources about variational autoencoders
- A refresher on information theory and more especially Kullback-Leibler divergence, needed to develop the variational bound
- A description of the problem and an expression of the variational bound following the notations of the aforementioned paper
- A hands-on implementation and training of a variational autoencoder on the CelebA dataset
- Example of interpolation in the latent space
[1] Auto-Encoding Variational Bayes, Diederik P Kingma, Max Welling, 2014.
[2] Large-scale CelebFaces Attributes (CelebA) Dataset, Ziwei Liu, Ping Luo, Xiaogang Wang, Xiaoou Tang.