Implementing softmax regression with a latent Gaussian variable

[junpenglao]

@grayfall

I suppose, specifying a shape in latent means that any future input must follow this shape. Is this the case?

Yes

@grayfall

I use an isometry to map probability vectors onto an orthonormal basis (and model them as latent Gaussians) and back to the simplex

As long as the mapping doesnt introduce unidentifiability, your model should run. In the FAQ there is tips on how to deal with bad energy error you can take a look at.