pyro-ppl / numpyro / 42599001613

93%

push

github

fix(diagnostics): summarize boolean samples as 0/1 values (#2414)

`summary`, `print_summary` and `hpdi` raised `TypeError: numpy boolean
subtract` on boolean samples, so `MCMC.print_summary(exclude_deterministic=False)`
crashed for any model with an indicator site such as
`numpyro.deterministic("theta_positive", theta > 0)`. Once past the HPDI,
the median formula would also add the two middle draws as a logical or and
report 0.5 instead of 1.0.

Cast boolean arrays to int32 before computing the statistics, so they are
summarized as their 0/1 values in both the unweighted and the weighted path.

5 of 5 new or added lines in 1 file covered. (100.0%)

15949 of 17157 relevant lines covered (92.96%)

1.7 hits per line