Sampler occasionally gets static in pymc3.1, but not in pymc3.0

Questions
9

Just a quick post, I’ll have another look tomorrow:

  • Paper about horseshoe: http://biomet.oxfordjournals.org/content/early/2010/04/28/biomet.asq017
    it also mentions a couple of similar priors with more or less shrinkage. The tau, theta and eta vars build a horseshoe prior together.
  • The 0.2 in subject_sd scales the HalfCauchy. This should be equivalent to using pm.HalfCauchy('subject_sd', beta=0.2).
  • subject_sd is the same as your hypersigma_intercept. intercept is the same as your hypermu_intercept, and subject is your intercept (in noncentered parametrization). I think that is the same as in your version, but maybe I’m missing something.

About the traces: Yes, that looks like trouble again. You should also get warnings about divergences. In a proper sampler run you don’t want any of those. Even a single divergence is a clear sign of trouble. You could try to increase target_accept to maybe 0.99 and see if that helps.