Transformed parameters are not part of random sampling, only mcmc.
If you want you can either use mcmc to do prior sampling (remove observations from your model and call pm.sample) or compute them as deterministic.
But maybe it helps to first know why you want those?