numpy masked array regression in 1.13.0rc1

[jswhit]

With version 1.12.1 the following script

from numpy import ma, __version__
print __version__
x = ma.array(1, mask=True)
print x[...],type(x[...])

produced

1.12.1
-- <class 'numpy.ma.core.MaskedConstant'>

With 1.13.0rc1 I get

1.13.0rc1
-- <class 'numpy.ma.core.MaskedArray'>

This means that x[...] is ma.masked returns False in 1.13, when it used to return True.

So it appears that the behaviour of scalar masked arrays has changed. This is causing test failures in netcdf4-python.

[ahaldane]

This was an intentional change in #8905, but we didn't note it in the release notes. (Probably we should).

I think the new behavior is more correct because in general indexing with [...] gives back an array, and converts scalars to arrays:

>>> np.array(3)[...]
array(3)
>>> np.int64(3)[...]
array(3)
>>> np.arange(2)[...]
array([0, 1])

if you goal is to convert a 0d array to a scalar, I think you should replace x[...] in your code by x[()]. Indexing with () is the usual way of converting 0d-arrays to scalars, and in a sense is the opposite of indexing with ....

>>> x = np.ma.array(1, mask=True)
>>> x[()] is np.ma.masked
True

[jswhit]

OK - thanks for the clarification. Close if you wish.

[charris]

@ahaldane Probably worth a note in the 1.9.0 release notes under compatibility.