pyro-ppl / numpyro / 41208596076

93%

push

github

fix(distributions): use discrete thresholds in GammaPoisson CDF (#2373)

* fix(distributions): use discrete thresholds in GammaPoisson CDF

I floor the evaluation threshold before computing the GammaPoisson CDF and handle its support limits explicitly. I added SciPy comparisons for the negative-binomial parameterizations and checks for finite rate gradients at the limits. I also updated the CDF formula to match the discrete support.

Assisted-by: Claude
Assisted-by: GPT-6 Astra
Signed-off-by: mottopanikeiku <176798723+mottopanikeiku@users.noreply.github.com>

* test: trim the GammaPoisson CDF regression

I reduced the regression to a single GammaPoisson test with fractional, negative and infinite thresholds and the rate gradient at those points, as asked in review.

Assisted-by: Claude
Assisted-by: GPT-6 Astra
Signed-off-by: mottopanikeiku <176798723+mottopanikeiku@users.noreply.github.com>

* fix(distributions): clamp huge GammaPoisson CDF thresholds

I clamp very large finite thresholds before betainc, which returns NaN close to the largest float, and compute the support masks once. I added a parametrized test for integer and large thresholds, batched parameters with rate=inf, and the NegativeBinomial2 and NegativeBinomialLogits subclasses, and noted in the docstring that the CDF is not differentiable in concentration.

Assisted-by: Claude
Assisted-by: GPT-6 Astra
Signed-off-by: mottopanikeiku <176798723+mottopanikeiku@users.noreply.github.com>

---------

Signed-off-by: mottopanikeiku <176798723+mottopanikeiku@users.noreply.github.com>
Co-authored-by: mottopanikeiku <176798723+mottopanikeiku@users.noreply.github.com>

8 of 8 new or added lines in 1 file covered. (100.0%)

15820 of 17030 relevant lines covered (92.89%)

1.7 hits per line