Adaptive Compression of the Latent Space in Variational Autoencoders