python - Odd behavior of numpy.all with object dtypes

Question

Given an array of dtype=object, numpy.all/any return the last object. For example:

>>> from string import ascii_lowercase
>>> x = np.array(list(ascii_lowercase), dtype=object)
>>> x.all()
'z'

In researching this issue, I couldn't find much except for this seemingly unrelated SO post which led me to find out that this is an open bug in numpy (as of March 2015): first report and more relevant issue. Posting this so others grappling with this can find this information out more efficiently.

Answer 1

In numpy version 1.8.2, np.any and np.all behave as classic short circuit logical and/or functions. LISP behavor comes to mind. Python's and and or operators do this.

Some examples:

In [203]: np.all(np.array([[1,2],1,[],[1,2,3]],dtype=object))
Out[203]: []

In [204]: np.any(np.array([[1,2],1,[],[1,2,3]],dtype=object))
Out[204]: [1, 2]

In [205]: np.any(np.array([0,[],[1,2],1,[],[1,2,3]],dtype=object))
Out[205]: [1, 2]

In [206]: np.all(np.array([True,False,[1,2],1,[],[1,2,3]],dtype=object))
Out[206]: False

np.all returns the first item that is logically False; else the last item. np.any the first item that is logically True; else the last item.

In the LISP world this is regarded as a useful feature. Not only does it stop evaluating elements as soon as the result is clear, but the identity of that return value can be used.

Is there a way of replicating this behavior using the and/or operators and some sort of map or reduce?

In [8]: 0 or [] or [1,2] or 1 or [1,2,3]
Out[8]: [1, 2]

???([0,[],[1,2],1,[1,2,3]])

---

as suggested in the comment:

In [26]: reduce(lambda a,b:a and b, np.array([1,2,3,[1,2,3]],dtype=object))
Out[26]: [1, 2, 3]

This might not actually short circuit the whole loop. Rather it short circuits each step, and propagates that value forward. Using lambda a,b:b and a returns the 1st item in the list, not the last. Timings could be used to test whether it is looping through the whole array (or not).

---

np.all is a ufunc that is defined as np.logical_and.reduce.

https://github.com/numpy/numpy/blob/master/numpy/core/_methods.py

umr_all = um.logical_and.reduce
def _all(a, axis=None, dtype=None, out=None, keepdims=False):
    return umr_all(a, axis, dtype, out, keepdims)

logical_and for dtype=object is defined in c source

https://github.com/numpy/numpy/blob/master/numpy/core/src/umath/funcs.inc.src

/* Emulates Python's 'a and b' behavior */
static PyObject *
npy_ObjectLogicalAnd(PyObject *i1, PyObject *i2)

similarly for np.any. Numeric dtype versions are defined else where.

There's a patch that forces np.all/any to return dtype=bool. But by calling np.logical_all directly you can control this yourself.

In [304]: np.logical_or.reduce(np.array([0,[1,2,3],4],dtype=object))
Out[304]: [1, 2, 3]

In [305]: np.logical_or.reduce(np.array([0,[1,2,3],4],dtype=object),dtype=bool)
Out[305]: True