numpy data type casting behaves differently in x=x+a and x+=a

Question

I noticed some differences between the operations between x=x+a and x+=a when manipulating some numpy arrays in python.

What I was trying to do is simply adding some random errors to an integer list, like this:

x=numpy.arange(12)
a=numpy.random.random(size=12)
x+=a

but printing out x gives an integer list [0,1,2,3,4,5,6,7,8,9,10,11].

It turns out that if I use x=x+a instead, it works as expected.

Is that something we should be aware of, I mean it behaves so differently. I used to think that it is totally equivalent between x+=a and x=x+a and I have been using them interchangeably without paying attention all the time. Now I am so concerned and anxious about all the computations I have done so far. Who knows when and where this has been creating a problem and I have to go through everything to double check.

Is this a bug in numpy? I have tested in numpy version 1.2.0 and 1.6.1 and they both did this.

Answer 1

No, this is not a bug, this is intended behavior. += does an in-place addition, so it can't change the data type of the array x. When the dtype is integral, that means the floating-point temporaries resulting from adding in the elements of a get truncated to integers. Since np.random.random returns floats in the range [0, 1), the result is always truncated back to the values in x.

By contrast, x + a needs to allocate a new array anyway, and upcasts the dtype of that new array to float when one argument is float and the other is integral.

The best way to avoid this problem is to be explicit about the required dtype in the arange call:

x = np.arange(12, dtype=float)
x += np.random.random(size=12)

(Note that x += a and x = x + a are seldom equivalent in Python, since the latter typically modifies the object pointed to by x. E.g. with pure Python lists:

a = []
b = a
a += [1]

modifies b as well, while a = a + [1] would leave b untouched.)

Answer 2

x += a modifies x in-place: data will be cast to int on assignment. x = x + a will assign the result of x + a to the label x, and in this case x + a will promote to a float.