Access multiple elements of an array

Question

Is there a way to get array elements in one operation for known rows and columns of those elements? In each row I would like to access elements from col_start to col_end (ea

Answer 1

A = np.arange(40).reshape(4,10)*.1
startend = [[2,5],[3,6],[4,7],[5,8]]
index_list = [np.arange(v[0],v[1]) + i*A.shape[1] 
                 for i,v in enumerate(startend)]
# [array([2, 3, 4]), array([13, 14, 15]), array([24, 25, 26]), array([35, 36, 37])]
A.flat[index_list]

producing

array([[ 0.2,  0.3,  0.4],
       [ 1.3,  1.4,  1.5],
       [ 2.4,  2.5,  2.6],
       [ 3.5,  3.6,  3.7]])

This still has an iteration, but it's a rather basic one over a list. I'm indexing the flattened, 1d, version of A. np.take(A, index_list) also works.

If the row intervals differ in size, I can use np.r_ to concatenate them. It's not absolutely necessary, but it is a convenience when building up indices from multiple intervals and values.

A.flat[np.r_[tuple(index_list)]]
# array([ 0.2,  0.3,  0.4,  1.3,  1.4,  1.5,  2.4,  2.5,  2.6,  3.5,  3.6, 3.7])

---

The idx that ajcr uses can be used without choose:

idx = [np.arange(v[0], v[1]) for i,v in enumerate(startend)]
A[np.arange(A.shape[0])[:,None], idx]

idx is like my index_list except that it doesn't add the row length.

np.array(idx)

array([[2, 3, 4],
       [3, 4, 5],
       [4, 5, 6],
       [5, 6, 7]])

Since each arange has the same length, idx can be generated without iteration:

col_start = np.array([2,3,4,5])
idx = col_start[:,None] + np.arange(3)

The first index is a column array that broadcasts to match this idx.

np.arange(A.shape[0])[:,None] 
array([[0],
       [1],
       [2],
       [3]])

With this A and idx I get the following timings:

In [515]: timeit np.choose(idx,A.T[:,:,None])
10000 loops, best of 3: 30.8 µs per loop

In [516]: timeit A[np.arange(A.shape[0])[:,None],idx]
100000 loops, best of 3: 10.8 µs per loop

In [517]: timeit A.flat[idx+np.arange(A.shape[0])[:,None]*A.shape[1]]
10000 loops, best of 3: 24.9 µs per loop

The flat indexing is faster, but calculating the fancier index takes up some time.

For large arrays, the speed of flat indexing dominates.

A=np.arange(4000).reshape(40,100)*.1
col_start=np.arange(20,60)
idx=col_start[:,None]+np.arange(30)

In [536]: timeit A[np.arange(A.shape[0])[:,None],idx]
10000 loops, best of 3: 108 µs per loop

In [537]: timeit A.flat[idx+np.arange(A.shape[0])[:,None]*A.shape[1]]
10000 loops, best of 3: 59.4 µs per loop

The np.choose method runs into a hardcoded limit: Need between 2 and (32) array objects (inclusive).

---

What out of bounds idx?

col_start=np.array([2,4,6,8])
idx=col_start[:,None]+np.arange(3)
A[np.arange(A.shape[0])[:,None], idx]

produces an error because the last idx value is 10, too large.

You could clip idx

idx=idx.clip(0,A.shape[1]-1)

producing duplicate values in the last row

[ 3.8,  3.9,  3.9]

You could also pad A before indexing. See np.pad for more options.

np.pad(A,((0,0),(0,2)),'edge')[np.arange(A.shape[0])[:,None], idx]

Another option is to remove out of bounds values. idx would then become a ragged list of lists (or array of lists). The flat approach can handle this, though the result will not be a matrix.

startend = [[2,5],[4,7],[6,9],[8,10]]
index_list = [np.arange(v[0],v[1]) + i*A.shape[1] 
                 for i,v in enumerate(startend)]
# [array([2, 3, 4]), array([14, 15, 16]), array([26, 27, 28]), array([38, 39])]

A.flat[np.r_[tuple(index_list)]]
# array([ 0.2,  0.3,  0.4,  1.4,  1.5,  1.6,  2.6,  2.7,  2.8,  3.8,  3.9])