ma.compress_nd behaves not as expected

[oldnaari]

Reproducing code example:

from numpy import ma
arr = ma.masked_all((3, 3, 2))
arr[0,0] = 1
out = ma.compress_nd(arr, axis=(0, 1))

This code sets out to the following value:

array([], shape=(0, 0, 2), dtype=float64)

The expected value is:

array([[1., 1.]])

Numpy/Python version information:

1.16.2 3.6.7 (default, Oct 22 2018, 11:32:17) 
[GCC 8.2.0]

[eric-wieser]

I think this is working as intended - the goal is to suppress slices that contain any masked elements, and in your array every slice along axes 0 and 1 contains a masked element.

There is however a bug here if you skip arr[0,0] = 1, we return an array of the wrong shape.

[oldnaari]

I don't get it. All elements for arr[0, 0, :] contain unmasked values. I have tried:

ma.compress_nd(arr, axis=(0, 1, 2))
ma.compress_nd(arr, axis=(0, 1))
ma.compress_nd(arr, axis=(1, 2))
ma.compress_nd(arr, axis=(0, 2))
ma.compress_nd(arr, axis=(0, ))
ma.compress_nd(arr, axis=(1, ))
ma.compress_nd(arr, axis=(2, ))

All of those give empty arrays. Shouldn't one of them be non empty? Or it works like

ma.compress_nd(arr, axis=(0, 1)) = ma.compress_nd(ma.compress_nd(arr, axis=0), axis=0)

?

[eric-wieser]

Or it works like ...

It should be the same as ma.compress_nd(ma.compress_nd(arr, axis=0), axis=1). The shape is wrong though, so you found another bug.

Turns out the two optimizations in that function are both wrong.

[ddasilva]

@eric-wieser Is this as simple as removing the following optimizations? I authored this function a number of years ago at the advice of a core developer (@charris I think?). Would be happy to fix.

    # Nothing is masked: return x
    if m is nomask or not m.any():
        return x._data
    # All is masked: return empty
    if m.all():
        return nxarray([])