This pull contains a working script for PG model building (without affine/residual)
PG scaling is done differently from John's code. It seems to work (it improves the lower bound), but the resulting subspace and precision matrix are a bit strange in the case where a Wishart prior is used for Az.
I am still using an ML update for the template. I need to change it for a MAP one.
This pull contains a working script for PG model building (without affine/residual)