Pytorch implementation of Learning Latent Subspaces in Variational Autoencoders