Regression: np.ma.masked_values(..., shrink=True) no longer works

[lpsinger]

The documentation for np.ma.masked_values() explains the role of the shrink option (which defaults to True) as follows:

Whether to collapse a mask full of False to nomask.

In Numpy <= 1.13.3, this works correctly:

$ python -c 'import numpy as np; print(repr(np.ma.masked_values(np.arange(10.0), -1)))'
masked_array(data = [ 0.  1.  2.  3.  4.  5.  6.  7.  8.  9.],
             mask = False,
       fill_value = -1.0)

In Numpy >= 1.14.0 and on master, it is broken:

$ python -c 'import numpy as np; print(repr(np.ma.masked_values(np.arange(10.0), -1)))'
masked_array(data=[0.0, 1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0, 9.0],
             mask=[False, False, False, False, False, False, False, False,
                   False, False],
       fill_value=-1.0)

[charris]

@eric-wieser This looks like yours (and mine :), #10232.

[eric-wieser]

This looks like a bug in np.ma.array(..., shrink=True) that was never previously exposed.

[eric-wieser]

So, the shrink argument to np.ma.array is documented as:

Whether to force compression of an empty mask. Default is True.

This is only used in exactly one case - when shrink=True, keep_mask=False, it means "don't expand nomask to an array of zeros"

This is super-specific, and not a reasonable interpretation of the documentation.

So I think that we can probably add back the shrinking to masked_values, but we should deprecate the argument to MaskedArray.__new__, since it's probably used only ever in error.

[charris]

We need to make a decision here. Just marking for future reference.

[charris]

@eric-wieser Could you make a PR with your suggestion here?

[charris]

Looks like this changed because the keepdims default value is True