Initial implementation of transition_hyperparameters #25
Conversation
| return log_numer - log_denom; | ||
| } | ||
| double logp_score() const { | ||
| double v1 = lbeta(s + alpha, N - s + beta); |
There was a problem hiding this comment.
Our current declaration / implementation of lbeta in util_math.hh and util_math.cc declare the arguments as ints. My compiler issued warnings that floating-point numbers were being coerced. I think now that we have hyperparameters that can be non-integer values, we'll want to change the argument types of lbeta (I think the implementation should stay the same).
| for (double alphat : ALPHA_GRID) { | ||
| alpha = alphat; | ||
| double lp = logp_score(); | ||
| if (!std::isnan(lp)) { |
There was a problem hiding this comment.
Do you have a sense of when nans are showing up?
There was a problem hiding this comment.
None in dirichlet_categorical, at least with the current unit test, but I did see them previously in BetaBernoulli's transition_hyperparamters.
Huh. When I rerun that now, I no longer see any Nan's. I wonder if the above lbeta change fixed them.
Only on the Distribution side (I.e., not hooked up to hirm.cc yet) and only for BetaBernoulli for now.