Numpy - 如何按降序对值/键对数组进行排序

Question

我正在查看问题Fastest way to rank items with multiple values and weightings 并提出以下解决方案，但还有两个问题：

import numpy as np

# set up values
keys = np.array([
    ['key1'],
    ['key2'],
    ['key3']
])
values = np.matrix([
    [1.1, 1.2, 1.3, 1.4],
    [2.1, 2.2, 2.3, 2.4],
    [3.1, 3.2, 3.3, 3.4]
])
weights = np.matrix([10., 20., 30., 40.]).transpose()

# crunch the numbers
res = values * weights

# combine results with labels
items = np.hstack((np.array(res), keys))

# !First problem - .hstack has promoted the first column from float64 to S4:
# array([['130.', 'key1'],
#        ['230.', 'key2'],
#        ['330.', 'key3']], 
#       dtype='|S4')
# How can I force it to stay numeric?

items.sort(reverse=True)   # doesn't work, no 'reverse' argument

# !Second problem - how to sort the array in descending order?

Answer 1

您可以将res 和keys 合并为一个结构化数组：

import numpy.lib.recfunctions as recfunctions
items = recfunctions.merge_arrays([res,keys])

由于np.sort 没有reverse=True 标志，我认为你能做的最好的就是反转返回的数组，（例如items[::-1]）或者取res 的负数：

import numpy as np
import numpy.lib.recfunctions as recfunctions

# set up values
keys = np.array([
    ['key1'],
    ['key2'],
    ['key3']
])
values = np.matrix([
    [1.1, 1.2, 1.3, 1.4],
    [2.1, 2.2, 2.3, 2.4],
    [3.1, 3.2, 3.3, 3.4]
])
weights = np.matrix([10., 20., 30., 40.]).transpose()

# crunch the numbers
res = values * weights

# combine results with labels
res = np.asarray(-res)
items = recfunctions.merge_arrays([res,keys])
items.dtype.names = ['res', 'key']
items.sort(order=['res'])
print(items)

产量

[(-330.0, 'key3') (-230.0, 'key2') (-130.0, 'key1')]

请注意，refunctions.merge_arrays 只是一个 Python 便利函数。它使用zip 和np.fromiter。避免加入res 和keys 并使用argsort 查找排序res 的索引并使用这些索引重新排序keys 肯定会更快：

import numpy as np

# set up values
keys = np.array([
    ['key1'],
    ['key2'],
    ['key3']
])
values = np.matrix([
    [1.1, 1.2, 1.3, 1.4],
    [2.1, 2.2, 2.3, 2.4],
    [3.1, 3.2, 3.3, 3.4]
])
weights = np.matrix([10., 20., 30., 40.]).transpose()

# crunch the numbers
res = values * weights

# combine results with labels
res = np.squeeze(np.asarray(res))
idx = np.argsort(res)[::-1]
print(keys[idx])
print(res[idx])

产量

[['key3']
 ['key2']
 ['key1']]
[ 330.  230.  130.]

Answer 2

您可以使用 numpy 数组的 argsort 方法对具有索引的键进行排序，从而对其他数组进行排序。

import numpy as np

# set up values
keys = np.array([
    ['key1'],
    ['key2'],
    ['key3']
])
values = np.array([
    [1.1, 1.2, 1.3, 1.4],
    [2.1, 2.2, 2.3, 2.4],
    [3.1, 3.2, 3.3, 3.4]
])
weights = np.array([10., 20., 30., 40.])

# crunch the numbers
res = np.dot(values, weights)

sortedkeys = keys[res.argsort(axis=0)[::-1]]

Answer 3

感谢@ondro 和@unutbu，这就是我最终得到的结果：

import numpy as np

# set up values
keys = np.array(['key1', 'key2', 'key3'])
values = np.array([
    [1.1, 1.2, 1.3, 1.4],    # values1_x
    [2.1, 2.2, 2.3, 2.4],    # values2_x
    [3.1, 3.2, 3.3, 3.4]     # values3_x
])
weights = np.array([10., 20., 30., 40.])

# crunch the numbers
res = np.dot(values, -weights)   # negative of weights!

order = res.argsort(axis=0)  # sorting on negative value gives
                             # same order as reverse-sort; there does
                             # not seem to be any way to reverse-sort
                             # directly
sortedkeys = keys[order].tolist()

返回['key3', 'key2', 'key1']（键，按值和权重的点积倒序排序）。