Repository navigation
numpy 1.11 has different floordiv behavior on infinity and 0 comparing to numpy 1.10 #7709
Description
I'm getting a different floordiv behavior when switching from 1.10 to 1.11. Is this intentional? It seems more inconsistent comparing to the normal division behavior:
# Old (numpy 1.10):
>>> import numpy
>>> x = numpy.float64(12.)
>>> x // 0.
inf
>>> x // numpy.inf
0.0
>>> x / 0.
inf
>>> x / numpy.inf
0.0
# New (numpy 1.11):
>>> import numpy
>>> x = numpy.float64(12.)
# Why nan?
>>> x // 0.
nan
# Why -0.0 when 12. is a positive number on a positive inf?
>>> x // numpy.inf
-0.0
>>> x / 0.
inf
>>> x / numpy.inf
0.0
This is on Linux 64-bit (numpy 1.11.0) from the conda environment with git_revision (4092a9e).
Activity
charris commentedon Jun 6, 2016
Floor division is now computed using
divmodso that it is compatible with%.divmodin turn is based aroundfmodas in python, indeed, the implementation is taken from python. See #7258. The nan is used so that the function can be used for arrays, whereas python raises aZeroDivisionError. We could possibly adjust the behavior for numpy scalars to raise a error, but it is probably better to remain compatible with arrays and array scalars. The-0.0looks like a bug.mhvk commentedon Jun 6, 2016
Mathematically, I guess
divmod(12., 0.)should beinf, nan, notnan, nan. In principle, since there is already a separate path forb=0, it would not take lead to any slowdown for the main path to assign a properly signed infinity.On the sign: almost certainly related to testing for sign of ratio as
(a / b) > 0.charris commentedon Jun 6, 2016
I thought about that, but as I was replacing a
ZeroDivisionError, nan, nan seemed more appropriate. It is an arguable point.njsmith commentedon Jun 10, 2016
In regular float division we replace
ZeroDivisionErrorwithinf, though (as required by IEEE754):So I guess I agree that returning
infmakes more sense, though I also agree that it isn't super clear...mhvk commentedon Jun 10, 2016
Yes, my logic was
1 / 0 = inf,1 // 0 = inf, and1 / 0 - 1 // 0 = inf - inf = nan.charris commentedon Jun 10, 2016
The argument for the other way round is that the divmod function determines the remainder first rather than doing a division. However, if
+/- infseems a better choice to most I'll happily change it before current behavior get too widespread. IIRC, the original changes broke some pandas tests which were fixed...charris commentedon Jun 10, 2016
@jreback Thoughts...
jreback commentedon Jun 10, 2016
This is the behavior in pandas master, which I think matches. On integers we do return a
NaNon//though.charris commentedon Jun 10, 2016
@jreback Any preference on how it ought to behave? Note that current behavior doesn't keep me up at nights, but if we want to change it, it should be before 1.12.
seberg commentedon Jun 11, 2016
np.floor(np.inf) == np.inf`, so I think I have a slight tendency to prefer infinity as well. Not sure how it plays with divmod though.
seberg commentedon Nov 19, 2022
This seems fixed now, giving
inffor floor division andinf+nanfor divmod.