Question

In both MATLAB and Numpy, arrays can be indexed by arrays. However, the behavior is different. Let me explain this by an example.

MATLAB:

>> A = rand(5,5)

A =

    0.1622    0.6020    0.4505    0.8258    0.1067
    0.7943    0.2630    0.0838    0.5383    0.9619
    0.3112    0.6541    0.2290    0.9961    0.0046
    0.5285    0.6892    0.9133    0.0782    0.7749
    0.1656    0.7482    0.1524    0.4427    0.8173

>> A([1,3,5],[1,3,5])

ans =

    0.1622    0.4505    0.1067
    0.3112    0.2290    0.0046
    0.1656    0.1524    0.8173

Numpy:

In [2]: A = arange(25).reshape((5,5))

In [3]: A
Out[3]: 
array([[ 0,  1,  2,  3,  4],
       [ 5,  6,  7,  8,  9],
       [10, 11, 12, 13, 14],
       [15, 16, 17, 18, 19],
       [20, 21, 22, 23, 24]])

In [6]: A[[0,2,4], [0,2,4]]
Out[6]: array([ 0, 12, 24])

In words: MATLAB selects rows and columns, Numpy “zips” the two index arrays and uses the tuples to point to entries.

How can I get the MATLAB behavior with Numpy?

Answer 1

You can do this:

A[[0,2,4],:][:,[0,2,4]]

which will give the MATLAB-like result you want.

It’s worth being aware that, rather inconsistently, if you use slices for indexing then you get MATLAB-like results without any such hackery:

>>> A[1:3,1:3]
array([[ 6, 7],
       [11,12]])

In numpy, unlike MATLAB, 1:3 is not just an abbreviation for [1,2] or anything of the kind. (At which point I feel obliged to mention something you surely know already, namely that Python’s 1:3 is kinda like [1,2] whereas MATLAB’s is kinda like [1,2,3]: the right-hand endpoint is included in MATLAB and excluded in Python.)