unp.isnan fails when input is empty array

[MichaelTiemannOSC]

np.isnan correctly returns an empty array when given an empty array as an argument.

unp.isnan goes straight to its wrapper, which attempts to vectorize the operation:

import numpy as np
print(np.isnan([]))
# array([], dtype=bool)
print(np.isnan([np.nan]))
# array([ True])
from uncertainties import unumpy as unp
print( unp.isnan([]))

Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/Users/michael/opt/miniconda3/envs/pandas/lib/python3.9/site-packages/numpy/lib/function_base.py", line 2328, in __call__
    return self._vectorize_call(func=func, args=vargs)
  File "/Users/michael/opt/miniconda3/envs/pandas/lib/python3.9/site-packages/numpy/lib/function_base.py", line 2406, in _vectorize_call
    ufunc, otypes = self._get_ufunc_and_otypes(func=func, args=args)
  File "/Users/michael/opt/miniconda3/envs/pandas/lib/python3.9/site-packages/numpy/lib/function_base.py", line 2362, in _get_ufunc_and_otypes
    raise ValueError('cannot call `vectorize` on size 0 inputs '
ValueError: cannot call `vectorize` on size 0 inputs unless `otypes` is set

[lebigot]

Indeed, this makes sense. Well spotted! Let me check how to fix this in a proper way.

[MichaelTiemannOSC]

Adding a note that when called with the singleton argument np.nan, it goes to a lot of trouble to create a vectorized call for that one element. Should there be a fast path for singletons vs arrays?

[lebigot]

I would say that the simplicity/uniformity of the code is a positive point and accelerating an already fast case (singletons) is not worth breaking a more maintainable code, at this stage.

Also, one could think of similarly accelerating the 2-element case, etc. This is a rabbit hole!

[MichaelTiemannOSC]

Fair enough. I'm looking at ways to avoid the overhead a different way (using PintArray to give me an array I can dequantify and pass to unp.isnan as a proper array rather than using map to call unp.isnan on elements one-by-one).