NumPy scalars, broadcasting and the __iadd__ method

Question

Any ideas on this oddity?

from numpy import *

a = array([1,2])
b = 1
b += a

gives array([2,3]), as you would expect. But

a = array([1,2])
b = array(1)
b += a

gives the error "non-broadcastable output operand with shape () doesn't match the broadcast shape (2)". At the same time

a = array([1,2])
b = array(1)
b = b + a

gives array([2,3]). Is this behaviour as odd as it looks at first glance?

Thanks in advance.

Answer 1

The += operator is taken to mean "in-place summation". Numpy imposes some constraints as to what in-place means: it cannot change the size or dtype of the array. When you do b = b + a there is no problem, because b now points to a new object resulting from adding b and a, which is a length 2 array. It is not surprise that b += a fails, since the length 2 array cannot be fitted into the length 1 array.

As for your first test case, my guess is that since Python ints are immutable objects, whenever you __iadd__ to one, you are effectively creating a new object and pointing to it, not modifying the object you had, so there is no reason why it shouldn't work with an array.