Equality tests on floats: add support for math.isclose() and numpy.isclose()

[alexandrehuat]

Feature request

When testing the equality of two floating point numbers, it’s good practices to rely on approximations (e.g., abs(a - b) < epsilon) rather than on exact/bit-to-bit comparison.

In Python we can use math.isclose or numpy.isclose to do such tests. These functions are a bit different but both can be equally useful. Unlike the built-in math.isclose, numpy.isclose is not symmetric in a and b – it assumes b is the reference value – so that isclose(a, b) might be different from isclose(b, a).

• Floating point comparison is a common enough operation so that Numba can support it (for both math and numpy libraries).
• Moreover, all arguments of each function should be supported (relative tolerance, absolute tolerance, and equal_nan if existing) because each use case can have specific parameters.

---

I could not find an issue about that when searching issues with keywords "isclose", "float comparison", "float equality".

[stuartarchibald]

Thanks for the suggestion. I've added this as a feature request. If you'd like to contribute, the high level Numba extension API @overload decorator should get you on the way to implementing this for Numba. Docs are here. A lot of example uses are in 'numba/np/arraymath.py'. Thanks again.

[esc]

@alexandrehuat It looks like no one is answering this question 😞 . If you would still like an answer, Rafia at codeFair knows Python and can help! Fill out this form if you're interested: http://tinyurl.com/2uttspwt.

This comment appears to be spam since it is out of context and doesn't make sense in the current conversation. My advice is: DO NOT CLICK THE LINK since it may very well redirect you to malware. I have reported the user @codeFairOfficial to the GitHub abuse program as outlined here: https://docs.github.com/en/github/building-a-strong-community/reporting-abuse-or-spam

[codeFairOfficial]

This is not spam, but we are so sorry to intrude! We actually are company looking to find new ways to get issues resolved quickly. We currently have a team of programmers finding issues that they could help get fixed. Our appologies for coming off as spam/malware, we will take this into account next time. I have deleted the original post

[esc]

@codeFairOfficial thank you for following up on this. We (the maintainers) always do welcome new members to our community, especially when it revolves around offering commercial services related to Numba. To avoid any false-positive classifications of your submissions, I would suggest introducing yourselves, your company and maybe your programmers on our Community forum at: https://numba.discourse.group/ -- this would also allow you to discuss with our community, your approach to finding issues to resolve in order to find consensus on a good process. For example, it will make sense to discuss with users about their needs in the issue itself rather than linking to an external service since "random hyperlinks" are always a cause of suspicion.

[codeFairOfficial]

@esc thank you for the feedback, we couldn't agree more. We will reach out to your community forum and discuss before posting any further.

[esc]

@codeFairOfficial thank you!

[alexandrehuat]

Hi, I also prefer to see a long URL rather than a short one. At least we know what domain we are reaching. Le ven. 5 févr. 2021 à 15:51, esc <notifications@github.com> a écrit :

…

@codeFairOfficial <https://github.com/codeFairOfficial> thank you! — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#5977 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AFTXC55KEVCHCSRTJA6OHJLS5QAV7ANCNFSM4OWSHSGQ> .

[jpivarski]

Similar to #6857, I think np.isclose has been implemented:

>>> import numpy as np
>>> import numba as nb
>>> nb.__version__
'0.55.1'
>>> def g(a, b):
...     return np.isclose(a, b)
... 
>>> g(np.array([1.1, 2.2, 3.3]), np.array([1.1+1e-12, 2.0, 3.3+1e-12]))
array([ True, False,  True])

Similarly for math.isclose:

>>> import math
>>> def h(a, b):
...     return math.isclose(a, b)
... 
>>> h(1.1, 1.1+1e-12)
True
>>> h(1.1, 1.0)
False

So I think this issue can be closed.

[esc]

@jpivarski thank you for the heads up!

[jpivarski]

Sorry—I have to take that back!

np.isclose and math.isclose are not implemented in Numba.

In my tests, I forgot to add the @nb.njit decorator.

[jpivarski]

This issue should be reopened!

[esc]

This issue should be reopened!

Of course, no problem! Errare humanum est! 😘

[renatobellotti]

I'd like to have this feature as well.