67

I'm looking for a one line solution that would help me do the following.

Suppose I have

array = np.array([10, 20, 30, 40, 50])

I'd like to rearrange it based upon an input ordering. If there were a numpy function called arrange, it would do the following:

newarray = np.arrange(array, [1, 0, 3, 4, 2])
print newarray

    [20, 10, 40, 50, 30]

Formally, if the array to be reordered is m x n, and the "index" array is 1 x n, the ordering would be determined by the array called "index".

Does numpy have a function like this?

Jonathan Hall
  • 75,165
  • 16
  • 143
  • 189
hlin117
  • 20,764
  • 31
  • 72
  • 93

4 Answers4

92

You can simply use your "index" list directly, as, well, an index array:

>>> arr = np.array([10, 20, 30, 40, 50])
>>> idx = [1, 0, 3, 4, 2]
>>> arr[idx]
array([20, 10, 40, 50, 30])

It tends to be much faster if idx is already an ndarray and not a list, even though it'll work either way:

>>> %timeit arr[idx]
100000 loops, best of 3: 2.11 µs per loop
>>> ai = np.array(idx)
>>> %timeit arr[ai]
1000000 loops, best of 3: 296 ns per loop
DSM
  • 342,061
  • 65
  • 592
  • 494
  • 6
    Thanks! What if the "index" is 2d array? For example I want to turn `[[1,2], ["a", "b"]` into `[[2,1], ["a", "b"]]` by using the index `[[1,0], [0,1]]`. I know an ugly way to do is using for-loop to re-arrange each row of the array using the corresponding row of index array. But there gotta be a faster way. – Skywalker326 Jun 29 '16 at 08:04
  • 1
    Note: the indexing array `idx` can not be a tuple. This was somehow counter-intuitive, but your solution works perfectly. Thanks – patzm Oct 02 '18 at 16:34
  • I know this is old but I am looking for the same solution. The line arr[ai] seems to have no effect at all. python version 3.7.3 – user1941319 Dec 25 '21 at 03:39
  • Check [Dong Dong Justin's answer](https://stackoverflow.com/a/64638032/7058363) for those who it isn't working. I had to convert It using `np.argsort()` – Genarito Mar 07 '23 at 18:47
9

for those whose index is 2d array, you can use map function. Here is an example:

a = np.random.randn(3, 3)
print(a)
print(np.argsort(a))

print(np.array(list(map(lambda x, y: y[x], np.argsort(a), a))))

the output is 

[[-1.42167035  0.62520498  2.02054623]
 [-0.17966393 -0.01561566  0.24480554]
 [ 1.10568543  0.00298402 -0.71397599]]
[[0 1 2]
 [0 1 2]
 [2 1 0]]
[[-1.42167035  0.62520498  2.02054623]
 [-0.17966393 -0.01561566  0.24480554]
 [-0.71397599  0.00298402  1.10568543]]
Jiaming Huang
  • 101
  • 1
  • 6
8

For those who have the same confusion, I am actually looking for a slightly different version of "rearrange array based upon index". In my situation, the index array is indexing the target array instead of the source array. In other words, I am try to rearrange an array based on its position in the new array.

In this case, simply apply an argsort before indexing. E.g.

>>> arr = np.array([10, 20, 30, 40, 50])
>>> idx = [1, 0, 3, 4, 2]
>>> arr[np.argsort(idx)]
array([20, 10, 50, 30, 40])

Note the difference between this result and the desired result by op.

One can verify back and forth

>>> arr[np.argsort(idx)][idx] == arr
array([ True,  True,  True,  True,  True])
>>> arr[idx][np.argsort(idx)] == arr
array([ True,  True,  True,  True,  True])
Dong Justin
  • 127
  • 1
  • 3
1

If you want to sort it but descending:

a = np.array([1,2,3,4,5])
np.argsort(a)
> array([0, 1, 2, 3, 4])
np.argsort(-a)
> array([4, 3, 2, 1, 0])
Pau
  • 536
  • 5
  • 16