Selecting specific rows and columns from NumPy array

Fancy indexing requires you to provide all indices for each dimension. You are providing 3 indices for the first one, and only 2 for the second one, hence the error. You want to do something like this:

>>> a[[[0, 0], [1, 1], [3, 3]], [[0,2], [0,2], [0, 2]]]
array([[ 0,  2],
       [ 4,  6],
       [12, 14]])

That is of course a pain to write, so you can let broadcasting help you:

>>> a[[[0], [1], [3]], [0, 2]]
array([[ 0,  2],
       [ 4,  6],
       [12, 14]])

This is much simpler to do if you index with arrays, not lists:

>>> row_idx = np.array([0, 1, 3])
>>> col_idx = np.array([0, 2])
>>> a[row_idx[:, None], col_idx]
array([[ 0,  2],
       [ 4,  6],
       [12, 14]])

USE:

 >>> a[[0,1,3]][:,[0,2]]
array([[ 0,  2],
   [ 4,  6],
   [12, 14]])

OR:

>>> a[[0,1,3],::2]
array([[ 0,  2],
   [ 4,  6],
   [12, 14]])

As Toan suggests, a simple hack would be to just select the rows first, and then select the columns over that.

>>> a[[0,1,3], :]            # Returns the rows you want
array([[ 0,  1,  2,  3],
       [ 4,  5,  6,  7],
       [12, 13, 14, 15]])
>>> a[[0,1,3], :][:, [0,2]]  # Selects the columns you want as well
array([[ 0,  2],
       [ 4,  6],
       [12, 14]])

[Edit] The built-in method: np.ix_

I recently discovered that numpy gives you an in-built one-liner to doing exactly what @Jaime suggested, but without having to use broadcasting syntax (which suffers from lack of readability). From the docs:

Using ix_ one can quickly construct index arrays that will index the cross product. a[np.ix_([1,3],[2,5])] returns the array [[a[1,2] a[1,5]], [a[3,2] a[3,5]]].

So you use it like this:

>>> a = np.arange(20).reshape((5,4))
>>> a[np.ix_([0,1,3], [0,2])]
array([[ 0,  2],
       [ 4,  6],
       [12, 14]])

And the way it works is that it takes care of aligning arrays the way Jaime suggested, so that broadcasting happens properly:

>>> np.ix_([0,1,3], [0,2])
(array([[0],
        [1],
        [3]]), array([[0, 2]]))

Also, as MikeC says in a comment, np.ix_ has the advantage of returning a view, which my first (pre-edit) answer did not. This means you can now assign to the indexed array:

>>> a[np.ix_([0,1,3], [0,2])] = -1
>>> a    
array([[-1,  1, -1,  3],
       [-1,  5, -1,  7],
       [ 8,  9, 10, 11],
       [-1, 13, -1, 15],
       [16, 17, 18, 19]])