@teoliphant wrote on 2008-04-03

Interesting suggestion, but this is a new feature. I'm moving it to 1.1

Milestone changed to 1.1 by @teoliphant on 2008-04-03

Milestone changed to Unscheduled by @cournape on 2009-03-02

@mwiebe wrote on 2011-03-23

This requires adding an object-array loop to the 'isnan' ufunc, which would then return False for anything that's not a float, and run the regular isnan on floats.

In the meantime, what is the recommended solution, do we check for types before we throw things at numpy.isnan?
My current workaround is
import numpy as np def np_isnan(inputVal): try: return np.isnan(inputVal) except TypeError: return False

Example please.

Here's the rather elaborate workaround I use in patsy:
https://github.com/pydata/patsy/blob/master/patsy/util.py#L510

On Wed, Jul 17, 2013 at 3:53 AM, Charles Harris [email protected]:

Example please.

—
Reply to this email directly or view it on GitHubhttps://github.com//issues/1307#issuecomment-21089024
.

Still open and worth considering.

Although this seems to have been partially fixed on #14831, isnan still raises when called on an array with mixed column types. For example:

# This raises TypeError
np.isnan(
  np.array([
    [5],
    [np.datetime64('2010-01-01')]
  ])
)

# Both of these work fine
np.isnan(
  np.array([
    [np.datetime64('2010-01-01')]
  ])
)

np.isnan(
  np.array([
    [5]
  ])
)

I'm using numpy 1.18.2

By "mixed column types" you mean "when dtype is object", so the comment above is still true:

This requires adding an object-array loop to the 'isnan' ufunc, which would then return False for anything that's not a float, and run the regular isnan on floats.

This would start with adding a loop function to numpy/core/src/umath/loops.c.src, and then adding that function via TD(O, f='fancy_object_isnan') to numpy/core/code_generators/generate_umath.py which generates the code to call the loop for object arrays. Then add a few tests, adjust the test suite for any new failures, and add a release note.

Closing. If this is needed for your workflow please reopen or open a new issue.