Efficient Numpy 2D array construction from 1D array

Question

I have an array like this:

A = array([1,2,3,4,5,6,7,8,9,10])

And I am trying to get an array like this:

B = array([[1,2,3],
          [2,3,4],
          [3,4,5],
          [4,5,6]])

Where each row (of a fixed arbitrary width) is shifted by one. The array of A is 10k records long and I'm trying to find an efficient way of doing this in Numpy. Currently I am using vstack and a for loop which is slow. Is there a faster way?

Edit:

width = 3 # fixed arbitrary width
length = 10000 # length of A which I wish to use
B = A[0:length + 1]
for i in range (1, length):
    B = np.vstack((B, A[i, i + width + 1]))

Answer 1

Actually, there's an even more efficient way to do this... The downside to using vstack etc, is that you're making a copy of the array.

Incidentally, this is effectively identical to @Paul's answer, but I'm posting this just to explain things in a bit more detail...

There's a way to do this with just views so that no memory is duplicated.

I'm directly borrowing this from Erik Rigtorp's post to numpy-discussion, who in turn, borrowed it from Keith Goodman's Bottleneck (Which is quite useful!).

The basic trick is to directly manipulate the strides of the array (For one-dimensional arrays):

import numpy as np

def rolling(a, window):
    shape = (a.size - window + 1, window)
    strides = (a.itemsize, a.itemsize)
    return np.lib.stride_tricks.as_strided(a, shape=shape, strides=strides)

a = np.arange(10)
print rolling(a, 3)

Where a is your input array and window is the length of the window that you want (3, in your case).

This yields:

[[0 1 2]
 [1 2 3]
 [2 3 4]
 [3 4 5]
 [4 5 6]
 [5 6 7]
 [6 7 8]
 [7 8 9]]

However, there is absolutely no duplication of memory between the original a and the returned array. This means that it's fast and scales much better than other options.

For example (using a = np.arange(100000) and window=3):

%timeit np.vstack([a[i:i-window] for i in xrange(window)]).T
1000 loops, best of 3: 256 us per loop

%timeit rolling(a, window)
100000 loops, best of 3: 12 us per loop

If we generalize this to a "rolling window" along the last axis for an N-dimensional array, we get Erik Rigtorp's "rolling window" function:

import numpy as np

def rolling_window(a, window):
   """
   Make an ndarray with a rolling window of the last dimension

   Parameters
   ----------
   a : array_like
       Array to add rolling window to
   window : int
       Size of rolling window

   Returns
   -------
   Array that is a view of the original array with a added dimension
   of size w.

   Examples
   --------
   >>> x=np.arange(10).reshape((2,5))
   >>> rolling_window(x, 3)
   array([[[0, 1, 2], [1, 2, 3], [2, 3, 4]],
          [[5, 6, 7], [6, 7, 8], [7, 8, 9]]])

   Calculate rolling mean of last dimension:
   >>> np.mean(rolling_window(x, 3), -1)
   array([[ 1.,  2.,  3.],
          [ 6.,  7.,  8.]])

   """
   if window < 1:
       raise ValueError, "`window` must be at least 1."
   if window > a.shape[-1]:
       raise ValueError, "`window` is too long."
   shape = a.shape[:-1] + (a.shape[-1] - window + 1, window)
   strides = a.strides + (a.strides[-1],)
   return np.lib.stride_tricks.as_strided(a, shape=shape, strides=strides)

So, let's look into what's going on here... Manipulating an array's strides may seem a bit magical, but once you understand what's going on, it's not at all. The strides of a numpy array describe the size in bytes of the steps that must be taken to increment one value along a given axis. So, in the case of a 1-dimensional array of 64-bit floats, the length of each item is 8 bytes, and x.strides is (8,).

x = np.arange(9)
print x.strides

Now, if we reshape this into a 2D, 3x3 array, the strides will be (3 * 8, 8), as we would have to jump 24 bytes to increment one step along the first axis, and 8 bytes to increment one step along the second axis.

y = x.reshape(3,3)
print y.strides

Similarly a transpose is the same as just reversing the strides of an array:

print y
y.strides = y.strides[::-1]
print y

Clearly, the strides of an array and the shape of an array are intimately linked. If we change one, we have to change the other accordingly, otherwise we won't have a valid description of the memory buffer that actually holds the values of the array.

Therefore, if you want to change both the shape and size of an array simultaneously, you can't do it just by setting x.strides and x.shape, even if the new strides and shape are compatible.

That's where numpy.lib.as_strided comes in. It's actually a very simple function that just sets the strides and shape of an array simultaneously.

It checks that the two are compatible, but not that the old strides and new shape are compatible, as would happen if you set the two independently. (It actually does this through numpy's __array_interface__, which allows arbitrary classes to describe a memory buffer as a numpy array.)

So, all we've done is made it so that steps one item forward (8 bytes in the case of a 64-bit array) along one axis, but also only steps 8 bytes forward along the other axis.

In other words, in case of a "window" size of 3, the array has a shape of (whatever, 3), but instead of stepping a full 3 * x.itemsize for the second dimension, it only steps one item forward, effectively making the rows of new array a "moving window" view into the original array.

(This also means that x.shape[0] * x.shape[1] will not be the same as x.size for your new array.)

At any rate, hopefully that makes things slightly clearer..

Answer 2

This solution is not efficiently implemented by a python loop since it comes with all kinds of type-checking best avoided when working with numpy arrays. If your array is exceptionally tall, you will notice a large speed up with this:

newshape = (4,3)
newstrides = (A.itemsize, A.itemsize)
B = numpy.lib.stride_tricks.as_strided(A, shape=newshape, strides=newstrides)

This gives a view of the array A. If you want a new array you can edit, do the same but with .copy() at the end.

Details on strides:

The newstrides tuple in this case will be (4,4) because the array has 4-byte items and you want to continue to step thru your data in single-item steps in the i-dimension. The second value '4' refers to the strides in the j-dimension (in a normal 4x4 array it would be 16). Because in this case you want to also also increment your read from the buffer in 4-byte steps in the j-dimension.

Joe give a nice, detailed description and makes things crystal-clear when he says that all this trick does is change strides and shape simultaneously.

Answer 3

Which approach are you using?

import numpy as np
A = np.array([1,2,3,4,5,6,7,8,9,10])
width = 3

np.vstack([A[i:i-len(A)+width] for i in xrange(len(A)-width)])
# needs 26.3µs

np.vstack([A[i:i-width] for i in xrange(width)]).T
# needs 13.2µs

If your width is relatively low (3) and you have a big A (10000 elements), then the difference is even more important: 32.4ms for the first and 44µs for the second.