Attempt at fixing unstable inference for `rho`
CHANGES: 1. Add InverseGamma prior to `sigma_space` 2. Add debugging `tf.print` statements to trace constrained and unconstrained parameter spaces. Result: `sigma_space` (unconstrained) hits a -nan as `rho` (unconstrained) gets large. The large value of `rho` causes a failure of the Cholesky decomposition.
Showing with 14 additions and 8 deletions