Question

我有两个大数组可供使用。但是，让我们看看下面的简化示例来获得这个想法：

我想查找data1中的元素是否与data2中的元素匹配，如果匹配，则返回data1和data2中的数组索引以新数组[index of data1, index of data2]的形式找到。例如，使用以下data1和data2集合，程序将返回：

data1 = [[1,1],[2,5],[623,781]] 
data2 = [[1,1], [161,74],[357,17],[1,1]]
expected_output = [[0,0],[0,3]]

我目前的代码如下：

result = []
for index, item in enumerate(data1):
    for index2,item2 in enumerate(data2):
        if np.array_equal(item,item2):
            result.append([index,index2])
>>> result
[[0, 0], [0, 3]]

这很好用。但是，我正在处理的实际两个数组各有60万个项目。上面的代码会非常慢。有没有办法加快这个过程？

Answer 1

可能不是最快，但容易且相当快：使用KDTrees：

>>> data1 = [[1,1],[2,5],[623,781]] 
>>> data2 = [[1,1], [161,74],[357,17],[1,1]]
>>>
>>> from operator import itemgetter
>>> from scipy.spatial import cKDTree as KDTree
>>>
>>> def intersect(a, b):
...     A = KDTree(a); B = KDTree(b); X = A.query_ball_tree(B, 0.5)
...     ai, bi = zip(*filter(itemgetter(1), enumerate(X)))
...     ai = np.repeat(ai, np.fromiter(map(len, bi), int, len(ai)))
...     bi = np.concatenate(bi)
...     return ai, bi
... 
>>> intersect(data1, data2)
(array([0, 0]), array([0, 3]))

两个假数据集1,000,000对每个需要3秒：

>>> from time import perf_counter
>>> 
>>> a = np.random.randint(0, 100000, (1000000, 2))
>>> b = np.random.randint(0, 100000, (1000000, 2))
>>> t = perf_counter(); intersect(a, b); s = perf_counter()
(array([   971,   3155,  15034,  35844,  41173,  60467,  73758,  91585,
        97136, 105296, 121005, 121658, 124142, 126111, 133593, 141889,
       150299, 165881, 167420, 174844, 179410, 192858, 222345, 227722,
       233547, 234932, 243683, 248863, 255784, 264908, 282948, 282951,
       285346, 287276, 302142, 318933, 327837, 328595, 332435, 342289,
       344780, 350286, 355322, 370691, 377459, 401086, 412310, 415688,
       442978, 461111, 469857, 491504, 493915, 502945, 506983, 507075,
       511610, 515631, 516080, 532457, 541138, 546281, 550592, 551751,
       554482, 568418, 571825, 591491, 594428, 603048, 639900, 648278,
       666410, 672724, 708500, 712873, 724467, 740297, 740640, 749559,
       752723, 761026, 777911, 790371, 791214, 793415, 795352, 801873,
       811260, 815527, 827915, 848170, 861160, 892562, 909555, 918745,
       924090, 929919, 933605, 939789, 940788, 940958, 950718, 950804,
       997947]), array([507017, 972033, 787596, 531935, 590375, 460365,  17480, 392726,
       552678, 545073, 128635, 590104, 251586, 340475, 330595, 783361,
       981598, 677225,  80580,  38991, 304132, 157839, 980986, 881068,
       308195, 162984, 618145,  68512,  58426, 190708, 123356, 568864,
       583337, 128244, 106965, 528053, 626051, 391636, 868254, 296467,
        39446, 791298, 356664, 428875, 143312, 356568, 736283, 902291,
         5607, 475178, 902339, 312950, 891330, 941489,  93635, 884057,
       329780, 270399, 633109, 106370, 626170,  54185, 103404, 658922,
       108909, 641246, 711876, 496069, 835306, 745188, 328947, 975464,
       522226, 746501, 642501, 489770, 859273, 890416,  62451, 463659,
       884001, 980820, 171523, 222668, 203244, 149955, 134192, 369508,
       905913, 839301, 758474, 114597, 534015, 381467,   7328, 447698,
       651929, 137424, 975677, 758923, 982976, 778075,  95266, 213456,
       210555]))
>>> print(s-t)
2.98617472499609

Answer 2

因为你的数据都是整数，你可以使用字典（哈希表），与保罗答案相同的数据的时间是0.55秒。这不一定会找到a和b之间的所有配对副本（即如果a和b本身包含重复项），但很容易将其修改为那个或之后进行第二次传递（仅在匹配的项目上）以检查数据中其他出现的那些向量。

import numpy as np

def intersect1(a, b):
    a_d = {}
    for i, x in enumerate(a):
        a_d[x] = i
    for i, y in enumerate(b):
        if y in a_d:
            yield a_d[y], i

from time import perf_counter
a = list(tuple(x) for x in list(np.random.randint(0, 100000, (1000000, 2))))
b = list(tuple(x) for x in list(np.random.randint(0, 100000, (1000000, 2))))
t = perf_counter(); print(list(intersect1(a, b))); s = perf_counter()
print(s-t)

相比之下，保罗在我的机器上需要2.46秒。

Answer 3

注意 其他答案，使用字典（用于检查完全匹配）或KDTree（用于epsilon-close匹配），要比这更好 - 更快更多存储器效率。

使用scipy.spatial.distance.cdist。如果您的两个数据阵列各有N个和M个条目，那么它将成为N个M成对距离数组。如果你可以将它放在RAM中，那么很容易找到匹配的索引：

import numpy as np
from scipy.spatial.distance import cdist

# Generate some data that's very likely to have repeats    
a = np.random.randint(0, 100, (1000, 2))
b = np.random.randint(0, 100, (1000, 2))

# `cityblock` is likely the cheapest distance to calculate (no sqrt, etc.)
c = cdist(a, b, 'cityblock')

# And the indexes of all the matches:
aidx, bidx = np.nonzero(c == 0)

# sanity check:
print([(a[i], b[j]) for i,j in zip(aidx, bidx)])

以上打印出来：

[(array([ 0, 84]), array([ 0, 84])),
 (array([50, 73]), array([50, 73])),
 (array([53, 86]), array([53, 86])),
 (array([96, 85]), array([96, 85])),
 (array([95, 18]), array([95, 18])),
 (array([ 4, 59]), array([ 4, 59])), ... ]

枚举两个大数组的快速方法？

3 个答案: