Repository navigation
Add a test of Counts sampling with obs=None where observations are dense and the delay PMF is of length > 1 #660
Description
Activity
Tests added to
test/test_observation_counts.pyexercise this scenario: obs=None, dense, delay PMF longer than 1:- TestCountsBasics.test_delay_convolution (line 42) — 2-element delay PMF, obs=None, asserts predicted[:1] is NaN (init period), predicted[1:] is non-NaN, and observed >= 0 everywhere (line 60). If the NaN init positions were
leaking -1s or NaNs into the sampled observed, this assertion would fail. - TestCountsCornerCases.test_long_delay_distribution (line 232) — 10-element delay PMF (the long_delay_pmf fixture), obs=None, asserts both ~jnp.isnan(result.observed) and result.observed >= 0 (lines 244–245).
Combined with the current sample() logic in pyrenew/observation/count_observations.py:375 — safe_predicted = jnp.where(isnan, 1.0, predicted_counts) fed to noise.sample, so NaN-predicted positions sample from Poisson(1.0 + ε)
rather than propagating NaN or -1 — the NaN-padding/-1s issue is both fixed and guarded by tests.- TestCountsBasics.test_delay_convolution (line 42) — 2-element delay PMF, obs=None, asserts predicted[:1] is NaN (init period), predicted[1:] is non-NaN, and observed >= 0 everywhere (line 60). If the NaN init positions were
Not for this PR: a test of
Countssampling withobs=Nonewhere observations are dense and the delay PMF is of length > 1 would have revealed the issue with NaN padding and -1s. Should add.Originally posted by @dylanhmorris in #644 (comment)