Numpy - Use values as index of another arrays

Question

I face a problem with Numpy, I try to use the values of each row (of B) as the indexes of another multidimensional array (A):

>>> A
array([[0, 1, 2],
       [3, 4, 5],
       [6, 7, 8]])
>>> B
    array([[ 0,  1,  2,  3],
           [ 1,  5,  6,  7],
           [ 8,  9, 10, 11]])
>>> np.clip(B, 0, 2)
    array([[0, 1, 2, 2],
           [1, 2, 2, 2],
           [2, 2, 2, 2]])

Here is the expected result:

array([[0, 1, 2, 2],
       [4, 5, 5, 5],
       [8, 8, 8, 8]])

Any idea ? Thanks a lot for your help.

np.clip just clips (i.e. clamps) the values of each element of an array, so in your case it would clip the values to between 0 and 2. Why do you expect that result from np.clip(...)? — XPhyro
– XPhyro, Commented Apr 1, 2020 at 9:13
Hi XPhyro. Because length of rows in 'A' is 3, so indexes need to be between 0 and 2. So i supposed that np.clip is the first step to answer my question. — Hantlowt
– Hantlowt, Commented Apr 1, 2020 at 9:19
What you're doing here is just clipping the values of the elements of B so that they are in the closed range [0, 2]; therefore, it is impossible for np.clip to return an array that has elements valued outside this range, which is what you expect it to do (looking by the values 4, 5 and 8 in the last two rows of the expected result). I suspect np.clip is not what you want. Can you elaborate what exactly you're trying to do here, perhaps with steps? I could not understand much just from the question itself. — XPhyro
– XPhyro, Commented Apr 1, 2020 at 9:30
My question is perhaps ambiguous indeed, i'm sorry ... Let us consider the result of np.clip as being already 'B'. I want to replace the values in each line by their correspondence in A. For example, the last line is filled with '2', so we replace them with '8', the value at index 2 of the last line of A. — Hantlowt
– Hantlowt, Commented Apr 1, 2020 at 9:35

hilberts_drinking_problem · Accepted Answer · 2020-04-01 10:11:48Z

3

It seems that you are looking for np.take_along_axis:

A = np.array([
       [0, 1, 2],
       [3, 4, 5],
       [6, 7, 8]]
)
B = np.array([
           [ 0,  1,  2,  3],
           [ 1,  5,  6,  7],
           [ 8,  9, 10, 11]]
)
C = np.clip(B, 0, 2)

res = np.take_along_axis(A, C, 1)
print(res)

answered Apr 1, 2020 at 10:11

hilberts_drinking_problem

11.6k3 gold badges25 silver badges55 bronze badges

Sign up to request clarification or add additional context in comments.

Comments

XPhyro · Accepted Answer · 2020-04-03 08:36:49Z

0

Try this:

result = np.clip(B, 0, 2)
for i, row in enumerate(result):
    for j, elem in enumerate(row):
        result[i][j] = A[i][elem]

After @hilberts_drinking_problem's answer, I have done some benchmarks using timeit. The answers, when tested with 10,000,000 loops took 103s and 82s, the latter being mine. Therefore, if your data is small but if you are going to be looping over it quite a lot, I would recommend you to use my answer. If such is not the case, I believe the convenience of hilberts_drinking_problem's answer makes it superior. For more information, see the comments to this answer.

Note that in both cases, I tested just the parts after clipping, meaning the preceding preparation part does not interfere with the benchmark.

edited Apr 3, 2020 at 8:36

answered Apr 1, 2020 at 9:46

XPhyro

5481 gold badge7 silver badges18 bronze badges

6 Comments

Hantlowt Over a year ago

Probably a working answer, but i was looking for a "vectorised" solution :/ There is no possibilities ?

XPhyro Over a year ago

By a vectorised solution, do you mean that the result should be in a vector form, or that the result should be computed via a series of vector operations?

Hantlowt Over a year ago

Vectorized operations, so by using numpy functions only

hilberts_drinking_problem Over a year ago

As a side comment on the benchmark, I would expect better performance from numpy as A and B get larger. For example, with n = 100; A = B = np.random.randint(0, 100, (n, n)), I see a 70x speedup for numpy. If A and B remain small, while the number of iterations becomes large, Python for loops may be viable. At this point, one may consider using numba.

Hantlowt Over a year ago

The numpy solution was the best for me because i have 250 millions rows with 6 columns.

|

Collectives™ on Stack Overflow

Numpy - Use values as index of another arrays

2 Answers 2

Comments

6 Comments

Your Answer

Hot Network Questions

Collectives™ on Stack Overflow

2 Answers 2

Comments

6 Comments

Your Answer

Sign up or log in

Post as a guest

Related