以下代码用于计算所有向量对之间的距离有什么问题？

Question

I'm trying to find manhattan distance between all pairs of vectors.

import numpy as np
import itertools

class vector:
    def __init__(self):
        self.a = 0
        self.b = 0

c = vector()
d = vector()
l = vector()
m = vector()

e = [c,d]
n = [l,m]
o = np.array(n)
f = np.array(e)
p = itertools.product(o,f)
p = list(p)
def comp(x):
    return (x[0].a-x[1].a) + (x[0].b-x[1].b)

g = np.vectorize(comp)
print g(p)

I get the error:

Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/usr/local/lib/python2.7/site-packages/numpy/lib/function_base.py", line 2207, in __call__
    return self._vectorize_call(func=func, args=vargs)
  File "/usr/local/lib/python2.7/site-packages/numpy/lib/function_base.py", line 2270, in _vectorize_call
    ufunc, otypes = self._get_ufunc_and_otypes(func=func, args=args)
  File "/usr/local/lib/python2.7/site-packages/numpy/lib/function_base.py", line 2232, in _get_ufunc_and_otypes
    outputs = func(*inputs)
  File "<stdin>", line 2, in comp
AttributeError: vector instance has no attribute '__getitem__'

Answer 1

I have to say I'd approach this differently. Numerical Python doesn't deal well with Python classes and such.

Your class

class vector:
    def __init__(self):
        self.a = 0
        self.b = 0

Is basically a length-2 vector. So, if you're going to operate on many length-2 vectors, I'd suggest something like this:

In [13]: p = np.array([[1, 2], [3, 4], [5, 6]])

In [14]: p
Out[14]: 
array([[1, 2],
       [3, 4],
       [5, 6]])

Each row is a length-2 vector. There are 3 such vectors. This is far far far more efficient than a Python list of Python classes.

Now your comp function

def comp(x):
    return (x[0].a-x[1].a) + (x[0].b-x[1].b)

is basically equivalent to

def comp(x):
    return (x[0].a+x[0].b) - (x[1].a+x[1].b)

ie, the component sum of the first vector, minus the component sum of the second vector. That being the case, you can efficiently calculate the pairwise outputs via

In [15]: q = p.sum(axis=1)

for calculating the component sum of each vector, followed by

In [16]: np.subtract.outer(q, q)
Out[16]: 
array([[ 0, -4, -8],
       [ 4,  0, -4],
       [ 8,  4,  0]])

Answer 2

The way you've written comp , it expects to be called with a two-tuple as an argument, but that's not what happens. p is a list of tuples. When you call a vectorized function on it, it is converted to a numpy array. The tuples are split into separate columns so you get a 4x2 array. Your function is then called on each cell of this array. So it gets called with just one vector object as an argument.

It's not really clear what you're trying to accomplish here. If your objects are not numbers, you won't gain anything by using things like np.vectorize on them; you should just call your function in a loop. If your objects are numbers, then just store them in an ordinary numpy array, and make use of better ways to compute such distances, like the pdist function in scipy .