当top_k<<axis_length时，它优于argsort。

import numpy as np

def get_sorted_top_k(array, top_k=1, axis=-1, reverse=False):
    if reverse:
        axis_length = array.shape[axis]
        partition_index = np.take(np.argpartition(array, kth=-top_k, axis=axis),
                                  range(axis_length - top_k, axis_length), axis)
    else:
        partition_index = np.take(np.argpartition(array, kth=top_k, axis=axis), range(0, top_k), axis)
    top_scores = np.take_along_axis(array, partition_index, axis)
    # resort partition
    sorted_index = np.argsort(top_scores, axis=axis)
    if reverse:
        sorted_index = np.flip(sorted_index, axis=axis)
    top_sorted_scores = np.take_along_axis(top_scores, sorted_index, axis)
    top_sorted_indexes = np.take_along_axis(partition_index, sorted_index, axis)
    return top_sorted_scores, top_sorted_indexes

if __name__ == "__main__":
    import time
    from sklearn.metrics.pairwise import cosine_similarity

    x = np.random.rand(10, 128)
    y = np.random.rand(1000000, 128)
    z = cosine_similarity(x, y)
    start_time = time.time()
    sorted_index_1 = get_sorted_top_k(z, top_k=3, axis=1, reverse=True)[1]
    print(time.time() - start_time)

这将比完整排序更快，这取决于原始数组的大小和选择的大小:

>>> A = np.random.randint(0,10,10)
>>> A
array([5, 1, 5, 5, 2, 3, 2, 4, 1, 0])
>>> B = np.zeros(3, int)
>>> for i in xrange(3):
...     idx = np.argmax(A)
...     B[i]=idx; A[idx]=0 #something smaller than A.min()
...     
>>> B
array([0, 2, 3])

当然，这涉及到对原始数组的篡改。你可以修复(如果需要)通过复制或替换回原始值. ...对你的用例来说，哪个更便宜。

对于多维数组，可以使用axis关键字，以便沿着预期的轴应用分区。

# For a 2D array
indices = np.argpartition(arr, -N, axis=1)[:, -N:]

对于抓取物品:

x = arr.shape[0]
arr[np.repeat(np.arange(x), N), indices.ravel()].reshape(x, N)

但请注意，这不会返回一个排序的结果。在这种情况下，你可以沿着预期的轴使用np.argsort():

indices = np.argsort(arr, axis=1)[:, -N:]

# Result
x = arr.shape[0]
arr[np.repeat(np.arange(x), N), indices.ravel()].reshape(x, N)

这里有一个例子:

In [42]: a = np.random.randint(0, 20, (10, 10))

In [44]: a
Out[44]:
array([[ 7, 11, 12,  0,  2,  3,  4, 10,  6, 10],
       [16, 16,  4,  3, 18,  5, 10,  4, 14,  9],
       [ 2,  9, 15, 12, 18,  3, 13, 11,  5, 10],
       [14,  0,  9, 11,  1,  4,  9, 19, 18, 12],
       [ 0, 10,  5, 15,  9, 18,  5,  2, 16, 19],
       [14, 19,  3, 11, 13, 11, 13, 11,  1, 14],
       [ 7, 15, 18,  6,  5, 13,  1,  7,  9, 19],
       [11, 17, 11, 16, 14,  3, 16,  1, 12, 19],
       [ 2,  4, 14,  8,  6,  9, 14,  9,  1,  5],
       [ 1, 10, 15,  0,  1,  9, 18,  2,  2, 12]])

In [45]: np.argpartition(a, np.argmin(a, axis=0))[:, 1:] # 1 is because the first item is the minimum one.
Out[45]:
array([[4, 5, 6, 8, 0, 7, 9, 1, 2],
       [2, 7, 5, 9, 6, 8, 1, 0, 4],
       [5, 8, 1, 9, 7, 3, 6, 2, 4],
       [4, 5, 2, 6, 3, 9, 0, 8, 7],
       [7, 2, 6, 4, 1, 3, 8, 5, 9],
       [2, 3, 5, 7, 6, 4, 0, 9, 1],
       [4, 3, 0, 7, 8, 5, 1, 2, 9],
       [5, 2, 0, 8, 4, 6, 3, 1, 9],
       [0, 1, 9, 4, 3, 7, 5, 2, 6],
       [0, 4, 7, 8, 5, 1, 9, 2, 6]])

In [46]: np.argpartition(a, np.argmin(a, axis=0))[:, -3:]
Out[46]:
array([[9, 1, 2],
       [1, 0, 4],
       [6, 2, 4],
       [0, 8, 7],
       [8, 5, 9],
       [0, 9, 1],
       [1, 2, 9],
       [3, 1, 9],
       [5, 2, 6],
       [9, 2, 6]])

In [89]: a[np.repeat(np.arange(x), 3), ind.ravel()].reshape(x, 3)
Out[89]:
array([[10, 11, 12],
       [16, 16, 18],
       [13, 15, 18],
       [14, 18, 19],
       [16, 18, 19],
       [14, 14, 19],
       [15, 18, 19],
       [16, 17, 19],
       [ 9, 14, 14],
       [12, 15, 18]])

我能想到的最简单的是:

>>> import numpy as np
>>> arr = np.array([1, 3, 2, 4, 5])
>>> arr.argsort()[-3:][::-1]
array([4, 3, 1])

这涉及到一个完整的数组。我想知道numpy是否提供了一种内置的方法来进行部分排序;到目前为止我还没有找到。

如果这个解决方案太慢(特别是对于小n)，那么可能值得考虑用Cython编写一些东西。

较新的NumPy版本(1.8及以上)有一个名为argpartition的函数。要得到四个最大元素的索引，请执行以下操作

>>> a = np.array([9, 4, 4, 3, 3, 9, 0, 4, 6, 0])
>>> a
array([9, 4, 4, 3, 3, 9, 0, 4, 6, 0])

>>> ind = np.argpartition(a, -4)[-4:]
>>> ind
array([1, 5, 8, 0])

>>> top4 = a[ind]
>>> top4
array([4, 9, 6, 9])

与argsort不同，这个函数在最坏的情况下以线性时间运行，但返回的索引没有排序，这可以从求值a[ind]的结果中看出。如果你也需要，那就把它们分类:

>>> ind[np.argsort(a[ind])]
array([1, 8, 5, 0])

以这种方式获得排在前k位的元素需要O(n + k log k)时间。

使用argpartition的向量化2D实现:

k = 3
probas = np.array([
    [.6, .1, .15, .15],
    [.1, .6, .15, .15],
    [.3, .1, .6, 0],
])

k_indices = np.argpartition(-probas, k-1, axis=-1)[:, :k]

# adjust indices to apply in flat array
adjuster = np.arange(probas.shape[0]) * probas.shape[1]
adjuster = np.broadcast_to(adjuster[:, None], k_indices.shape)
k_indices_flat = k_indices + adjuster

k_values = probas.flatten()[k_indices_flat]

# k_indices:
# array([[0, 2, 3],
#        [1, 2, 3],
#        [2, 0, 1]])
# k_values:
# array([[0.6 , 0.15, 0.15],
#        [0.6 , 0.15, 0.15],
#       [0.6 , 0.3 , 0.1 ]])