Consider the following numpy array array:
x = np.array([2]*4, dtype=np.uint8)
which is just an array of four 2's.
I want to perform a bitwise_and reduction of this array:
y = np.bitwise_and.reduce(x)
I expect the result to be:
2
because each element of the array is identical, so successive AND's should yield the same result, but instead I get:
0
Why the discrepancy?
In the reduce docstring, it is explained that the function is equivalent to
r = op.identity # op = ufunc
for i in range(len(A)):
r = op(r, A[i])
return r
The problem is that np.bitwise_and.identity is 1:
In [100]: np.bitwise_and.identity
Out[100]: 1
For the reduce method to work as you expect, the identity would have to be an integer with all bits set to 1.
The above code was run using numpy 1.11.2. The problem has been fixed in the development version of numpy:
In [3]: np.__version__
Out[3]: '1.13.0.dev0+87c1dab'
In [4]: np.bitwise_and.identity
Out[4]: -1
In [5]: x = np.array([2]*4, dtype=np.uint8)
In [6]: np.bitwise_and.reduce(x)
Out[6]: 2
