In WAE, P(z) is supposed to be 2dim gaussian distribution and the model is trying to minimize the divergence between Q(z|x) and P(z).
In next step, you use GMM to model the latent space.So why not just use the 2dim gaussian distribution as the model of latent space.
Is it because the 2dim gaussian distribution is not so accurate to describe the latent space?
In WAE, P(z) is supposed to be 2dim gaussian distribution and the model is trying to minimize the divergence between Q(z|x) and P(z). In next step, you use GMM to model the latent space.So why not just use the 2dim gaussian distribution as the model of latent space. Is it because the 2dim gaussian distribution is not so accurate to describe the latent space?