ENH: Add an "isclose" function for element-wise approximate floating point comparison of two arrays.

[joferkington]

Following the discussion on the mailing list, it seems like a floating point comparison function that returns a boolean array would be useful.

[charris]

Might be worth making this NA aware.

[rgommers]

Docstring should be removed, otherwise np.test(verbose=2) won't show the arguments to yield (IIRC). Same for other test methods.

[rgommers]

Looks good, nice test coverage. It would be useful to test against allclose to make sure that isclose().all() == allclose().

[rgommers]

I do confess to not liking the name isclose, although it's hard to come up with a better one. I'd expect any name starting with "is" or "has" to return a single boolean.

[rgommers]

allclose behavior after fixing http://projects.scipy.org/numpy/ticket/1905 that is. If you could review and add the proposed patch in that ticket too, that would be great.

[joferkington]

I've made things MA-aware, and added isclose().all() == allclose() tests. I completely agree on the name implying a boolean return, but "approximately_equal" is a bit too long, and I can't think of anything better either.

[charris]

Actually, I was thinking of NA, where NA means missing data:

In [1]: a = arange(5, maskna=1)

In [2]: b = arange(5)

In [3]: a[1] = NA

In [4]: isclose(a,b)
---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
/home/charris/<ipython-input-4-512760b5689e> in <module>()
----> 1 isclose(a,b)

/home/charris/.local/lib/python2.7/site-packages/numpy/core/numeric.pyc in isclose(a, b, rtol, atol, equal_nan)
   2091     xfin = isfinite(x)
   2092     yfin = isfinite(y)
-> 2093     if all(xfin) and all(yfin):
   2094         return within_tol(x, y, atol, rtol)
   2095     else:

ValueError: numpy.NA represents an unknown missing value, so its truth value cannot be determined

There is a skipna flag in all that you might want to use in some spots. Maybe it should be an option in isclose also. You can get the mask using isna(a), and for efficiency you can also check it the array is maskna

In [6]: a.flags.maskna
Out[6]: True

.Edit: There is also some trailing whitespace.

[joferkington]

Ah! Sorry, I was confused. 408f94c should make things NA-aware. Would it be useful to go ahead and make allclose NA-aware as well (and add a skipna kwarg, a la all), or should that be left for another time?

[charris]

This can be simplified also:

In [9]: a = arange(5, maskna=1)

In [10]: a[1] = NA

In [12]: isfinite(a) & isfinite(a)
Out[12]: array([ True, NA,  True,  True,  True], dtype=bool)

In [12]: isfinite(a) & isfinite(a)
Out[12]: array([ True, NA, True, True, True], dtype=bool)

[joferkington]

Thanks! Done in 3d28404 along with your other suggestions.

[charris]

Hmm, that is strange. There may indeed be a bug somewhere. I just pushed this PR after squashing it a bit and adding a release note. But I'll leave it open for a while. If you can produce a simple failing example it would be helpful.

[charris]

Oh, and how current is your numpy fork?

[joferkington]

The "bug" I was running into turned out to just be because ma masked arrays don't have a skipna option in their allmethod. (Which makes sense, in retrospect.) I didn't realize that the error was being thrown with a masked array, and so I couldn't reproduce it by trying a "normal" ndarray. I probably should have added another comment instead of editing the one in question.

[joferkington]

I think it's safe to close the pull request. There's no bug, I'm just easily confused. :) Thanks!

[charris]

OK, that is a bug in ma, please open a ticket for that. Or you could make a PR with a fix ;) I'll go ahead and close this PR.