pyro-ppl / numpyro / 41439709888

93%

push

github

Add ICDF and CDF to BernoulliProbs (#1652)

* Add ICDF and CDF to BernoulliProbs

* Add tests for BernoulliProbs ICDF and CDF

* feat(distributions): add cdf and icdf to BernoulliProbs and BernoulliLogits

Replace the heaviside-based BernoulliProbs.cdf/icdf from this PR and add
the same methods to BernoulliLogits, as requested in review.

icdf is the generalized inverse min{k : F(k) >= q}, so the level
q = P(X = 0) maps to 0; the heaviside version mapped it to 1, which broke
icdf(cdf(0)) == 0. BernoulliLogits computes P(X = 0) as expit(-logits),
which keeps its precision when expit(logits) rounds to 1. NaN values and
parameters propagate as NaN.

Drop the Bernoulli special case in test_cdf_and_icdf, whose
cdf(icdf(q)) == q check cannot hold for a discrete distribution, in favor
of dedicated tests against scipy.stats.bernoulli that cover the jump
level and saturated logits.

Co-authored-by: Dan Manela <danielmanela@gmail.com>

---------

Co-authored-by: Juan Orduz <juanitorduz@gmail.com>

16 of 16 new or added lines in 1 file covered. (100.0%)

15837 of 17047 relevant lines covered (92.9%)

1.7 hits per line