Repository navigation
Object-type masked array special-cases elements that are ndarray #18981
Description
The MaskedArray class has unexpected special-casing of ndarray objects within an object array. I expect an object array to be a generic container of whatever is in each cell, and if that element is masked then accessing it should always return np.ma.masked. Instead MaskedArray appears to turn the ndarray into a fully-masked MaskedArray.
Reproducing code example:
import numpy as np
>>> m = np.ma.array([None, [1, 2], np.array([1, 2])], dtype=object)
>>> m
masked_array(data=[None, list([1, 2]), array([1, 2])],
mask=False,
fill_value='?',
dtype=object)
>>> m.mask = [True, True, True]
>>> m[0]
masked
>>> m[1]
masked
>>> m[2] # Accessing the ndarray has auto-magically converted it to masked_array
masked_array(data=[--, --],
mask=[ True, True],
fill_value=999999,
dtype=int64)
>>> m.data[2] # The data element under the mask is still the original ndarray
array([1, 2])Expected:
>>> m[2]
masked
Error message:
N/A
NumPy/Python version information:
1.19.5 3.7.9 (default, Aug 31 2020, 07:22:35)
[Clang 10.0.0 ]
Activity
seberg commentedon May 12, 2021
I expect this needs a check for
objectdtype (maybe together with a check forndim == 1. Although, IIRC that logic is definitely pretty holey... It was not a regression in 1.19.x, right?fengluoqiuwu commentedon Oct 26, 2024
It also looks like there could be a conflict with another issue from about 10 years ago.
If we want to fix both of it, we must change the basic structure of masked......