Slicing each row in a different way

Question

I have a numpy.array a with the shape (8, 16000, 2) which represents 8 stereo audios each 1s long. I want to sample a 0.5s subsample from each recording. Thus, I sample start and end indices in the following way:

a = np.random.rand(8, 16000, 2)
possible_starts = np.arange(8000)
start_idxs = np.random.choice(possible_starts, size=a.shape[0])
end_idxs = start_idxs + 8000

How to extract these different slices for each row in numpy efficienty? I obviously can do it in loop:

b = np.zeros((8, 8000, 2))
for i in range(a.shape[0]):
    b[i] = a[i, start_idxs[i]: end_idxs[i], :]

but I want to do it in the vectorised way.

@fountainhead I said it has a shape (8, 16000, 2). It can be any such np.array. — Fallen Apart, Oct 27 '20 at 15:10
You'll get more help if you provide code to generate example data, so that we can copy, paste, and directly test the code. Think of it this way: we want to help but we are also lazy! — senderle, Oct 27 '20 at 15:17
@Divakar I was surprised that there is no any simpler solution and I ended up using loop, which turned out to be quick enough. I will yet investigate it later. — Fallen Apart, Oct 28 '20 at 13:07

Divakar · Answer 1 · 2020-10-27T16:00:49.933

1

We can leverage np.lib.stride_tricks.as_strided based scikit-image's view_as_windows to get sliding windowed views and thus solve our problem here in a vectorized way.

from skimage.util.shape import view_as_windows

w = view_as_windows(a, (1,8000,1))[:,:,:,0,:,0]
b_out = w[np.arange(len(start_idxs)), start_idxs].swapaxes(1,2)

Or use np.take_along_axis(w, start_idxs, axis=1).swapaxes(1,2) for better readability as suggested by @Daniel F.

More info on use of as_strided based view_as_windows.

Given that your loop is iterating for just 8 times, you should benchmark.

edited Oct 27 '20 at 16:00

answered Oct 27 '20 at 15:33

Divakar

218,885
19
262
358

`b_out = np.take_along_axis(w, start_idxs, 1).swapaxes(1,2)` might be a bit more readable? – Daniel F Oct 27 '20 at 15:56

Slicing each row in a different way

1 Answers1