variational autoencoder from scratch