Sampling with blackjax - no divergence report

Questions v5
1

Hi all,

I’m running a pymc model using the blackjax sampler (like so: pm.sample(nuts_sampler="blackjax"). This seems to work fine, however, I am wondering about divergences as these seem to not be reported as is the case when using the standard nuts sampler. If there is no warning, can I assume that no divergences occurred? Or are these generally not reported?

Thanks!

2

Neither of the JAX samplers report divergences after they finish MCMC. You can check the idata by looking at idata.sample_stats.diverging (take the sum to get the total number of divergences, or you can look by chain), or by inspecting plots that report divergences (trace, pair…).

1 Like