Why does sum() operation on numpy masked_array change fill value to 1e20?

Question

Is this a feature or a bug? Can someone explain to me this behavior of a numpy masked_array? It seems to change the fill_value after applying the sum operation, which is confusing if you intend to use the filled result.

data=ones((5,5))
m=zeros((5,5),dtype=bool)

"""Mask out row 3"""
m[3,:]=True
arr=ma.masked_array(data,mask=m,fill_value=nan)

print arr
print 'Fill value:', arr.fill_value
print arr.filled()


farr=arr.sum(axis=1)
print farr
print 'Fill value:', farr.fill_value
print farr.filled()

"""I was expecting this"""
print nansum(arr.filled(),axis=1)

Prints output:

[[1.0 1.0 1.0 1.0 1.0]
 [1.0 1.0 1.0 1.0 1.0]
 [1.0 1.0 1.0 1.0 1.0]
 [-- -- -- -- --]
 [1.0 1.0 1.0 1.0 1.0]]
Fill value: nan
[[  1.   1.   1.   1.   1.]
 [  1.   1.   1.   1.   1.]
 [  1.   1.   1.   1.   1.]
 [ nan  nan  nan  nan  nan]
 [  1.   1.   1.   1.   1.]]
[5.0 5.0 5.0 -- 5.0]
Fill value: 1e+20
[  5.00000000e+00   5.00000000e+00   5.00000000e+00   1.00000000e+20
   5.00000000e+00]
[  5.   5.   5.  nan   5.]

Answer 1

The array returned by arr.sum is a new array which does not inherit the fill_value of arr (though I agree that might be a nice improvement to np.ma). As a workaround, you could use

In [18]: farr.filled(arr.fill_value)
Out[18]: array([  5.,   5.,   5.,  nan,   5.])