dmaust / rounding

Stochastic Rounding

rounding

Rounding methods

StandardRound rounds to the nearest, and in the event of a tie, rounds up.

RoundToEven rounds to the nearest, but in the event of a tie, rounds toward the nearest even number.

StochasticRound provides a mechanism to eliminate accumulated roundoff error in the presence of a distribution where for individual samples, the roundoff error is skewed. This is typically caused by small values. This may be a common problem when dealing with applying a function to many small integer values as the number of discrete inputs is small.

One drawback to stochastic rounding is the output is non-deterministic, but this can be avoided by providing a custom deterministic generator, or invoking providing the random number generator with a fixed seed.

r = random.Random()
r.seed(123)
sr = StochasticRound(precision=0, random_generator=r)