PrincetonUniversity / PsyNeuLink / 25532019955

84%

master: 85%

Build Type

Pull #3535

github

Committed by younesStrittmatter

Commit Message

OneHot: fix PROB/PROB_INDICATOR sampling at the cumsum-rounding boundary

cum_sum[-1] of a softmax output can drift just below 1.0 due to float
rounding (e.g. np.cumsum([0.7, 0.2, 0.1])[-1] == 0.9999999999999999) while
random_state.uniform() draws from [0, 1). When the draw lands in the
(cum_sum[-1], 1.0) gap:

  * the Python path raised StopIteration from
        next(element for element in cum_sum if element > random_value)
  * the LLVM path silently produced an all-zero output (no bucket's
        sum_old <= random_draw < sum_new condition fired)

Both paths now fall back to the last index in that case. The Python path
also indexes by position rather than matching cum_sum entries by equality,
which fixes a separate bug where prob_dist values of 0 produced duplicate
cum_sum entries and selected multiple indices simultaneously.

Two SOFT_MAX MASK_THRESHOLD PROB test expectations encoded the equality-
match bug (two non-zero entries instead of one) and have been corrected
to a single chosen sample.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Pull Request Pull Request #3535: OneHot: fix PROB sampling at the cumsum-rounding boundary

Coverage Stats

10520 of 13698 branches covered (76.8%)

Branch coverage included in aggregate %.

6 of 6 new or added lines in 1 file covered. (100.0%)

2 existing lines in 1 file now uncovered.

36186 of 41762 relevant lines covered (86.65%)

0.87 hits per line