np.ma.argmax on masked array of unsigned integer dtype returns wrong result in numpy 1.11.0

Question

I stumbled over a strange fact concerning masked unsigned integer arrays and np.ma.argmax.

Consider the following array:

>>> marr = np.ma.array(np.array([[2,2,2], [3,3,3], [1,1,1]]), mask=False, dtype=np.uint16)
>>> marr
masked_array(data =
 [[2 2 2]
 [3 3 3]
 [1 1 1]],
             mask =
 [[False False False]
 [False False False]
 [False False False]],
       fill_value = 999999)

If I use np.ma.argmax the result is what I expected:

>>> print(np.ma.argmax(marr, axis=0))
[1 1 1]

If I now mask the last row the result is wrong:

>>> marr.mask[2] = True
>>> marr
masked_array(data =
 [[2 2 2]
 [3 3 3]
 [-- -- --]],
             mask =
 [[False False False]
 [False False False]
 [ True  True  True]],
       fill_value = 999999)
>>> print(np.ma.argmax(marr, axis=0))
[2 2 2]   # why?

It now thinks the masked row is the maximum? I even changed the fill_value to 0 but the result stays the same: It still thinks the masked row is the maximum. It seems like this only affects unsigned integer arrays.

However np.ma.MaskedArray.argmax and np.argmax return the correct result:

>>> print(marr.argmax(axis=0))
[1 1 1]
>>> print(np.argmax(marr, axis=0))
[1 1 1]

Why is the np.ma.argmax not doing the correct thing here? As far as I can see it is defined as the method itself.

Answer 1

Here's the code for np.ma.argmax (via ipython ?? magic) (version '1.11.0')

Definition:  np.ma.argmax(a, axis=None, fill_value=None)
Source:
def argmax(a, axis=None, fill_value=None):
    "Function version of the eponymous method."
    if fill_value is None:
        fill_value = default_fill_value(a)
        try:
            fill_value = -fill_value
        except:
            pass
    d = filled(a, fill_value)
    return d.argmax(axis=axis)

and for the method:

def argmax(self, axis=None, fill_value=None, out=None):
    if fill_value is None:
        fill_value = maximum_fill_value(self._data)
    d = self.filled(fill_value).view(ndarray)
    return d.argmax(axis, out=out)

The function uses a different fill value

In [180]: np.ma.maximum_fill_value(marr)
Out[180]: 0

In [181]: np.ma.maximum_fill_value(marr.astype(int))
Out[181]: -2147483648

In [182]: np.ma.default_fill_value(marr)
Out[182]: array(999999)

In [183]: -np.ma.default_fill_value(marr)
Out[183]: -999999

In [184]: np.ma.filled(marr,-np.ma.default_fill_value(marr))
Out[184]: 
array([[    2,     2,     2],
       [    3,     3,     3],
       [48577, 48577, 48577]], dtype=uint16)

In [186]: np.ma.filled(marr,np.ma.maximum_fill_value(marr))
Out[186]: 
array([[2, 2, 2],
       [3, 3, 3],
       [0, 0, 0]], dtype=uint16)

It's a bug in my version. There was a change in Feb, replacing the argmax (and argmin) with the method (argmax = _frommethod('argmax')).

https://github.com/numpy/numpy/commit/36f76ea2e6e91062df12d3a46ccaed7822bc82f2

So that correction isn't in my distribution - an presumably not yours.

So for now stick with method, or provide your own correct fill_value.

In [187]: np.ma.argmax(marr,axis=0,fill_value=0)
Out[187]: array([1, 1, 1], dtype=int32)