Hey, I’m using NUTS to sample a model with ~30 parameters, and I find that I get good convergence when I reparametrize the model with
TransformReparam for Gaussian and Uniform prior distributions, respectively.
I would like to extract the potential energy of each sample in the (original) chain. However, I find that if I ask for
extra_fields=('potential_energy',), this value is different from the one I can compute manually by adding the log prior and the log likelihood for the posterior samples. I have set
Is the reparametrization changing the value of the potential energy? Or are there two sets of potential energy values, one for the “base” parameters and one for the “original” ones?
Many thanks for your help!