Using np.lib.stride_tricks.as_stride one can very efficiently create a sliding window that segments an array as a preprocessing step for vectorized applications. For example a moving average of a window length 3, stepsize 1:

a = numpy.arange(10)
a_strided = numpy.lib.stride_tricks.as_strided(
    a, shape=(8, 3), strides=(8, 8)
)
print numpy.mean(a_strided, axis=1)

This is very performant but very hard to do, as the shape and strides parameters are very hard to understand.

I suggest the implementation of a "simple sliding_window function" that does the job of figuring out those two parameters for you.

The implementation I have been using for years allows the above to be replaced by

a = numpy.arange(10)
a_strided = sliding_window(a, size=3, stepsize=1)
print numpy.mean(a_strided, axis=1)

@teoliphant also has an implementation that would change the above to

a = numpy.arange(10)
a_strided = array_for_sliding_window(a, 3)
print numpy.mean(a_strided, axis=1)

both making it much more readable.

Seeing it is a common usecase in vectorized computing I suggest we put a similar function into NumPy itself.

Regarding to which implementation to follow, they are both assume different things but allow you to do the same thing eventually:

sliding_window

• slides over one axis only
• allows setting windowsize and stepsize
• returns array with dimension n+1
• sliding over several axes requires two calls (which come for free as there is no memory reordered)
• has a superfluous copy parameter that can be removed and replaced by appending .copy() after the call

array_for_sliding_window

• slides over all axes simultaneously, window lengths are given as tuple parameter
• Assumes a stepsize one in all directions
• returns array with dimension n*2
• stepsize not equal to one requires slicing of output data (unsure if this implies copying data)
• Disabling sliding over axis[n] requires you set argument wshape[n] = 1 or wshape[n] = a.shape[n]

This means for flexible stepsize the following are equivalent (with some minor bug in sliding_window there):

a = numpy.arange(10)
print sliding_window(a, size=3, stepsize=2)

a = numpy.arange(10)
print array_for_sliding_window(a, 3)[::2, :] # Stepsize 2 by dropping every 2nd row

for sliding over one axis the following are equivalent (with some transposing and squeezing):

a = numpy.arange(25).reshape(5, 5)
print sliding_window(a, size=3, axis=1)

a = numpy.arange(25).reshape(5, 5)
print array_for_sliding_window(a, (1, 3))

and for sliding over two axis the following are equivalent:

a = numpy.arange(25).reshape(5, 5)
print sliding_window(sliding_window(a, size=3, axis=0), size=2, axis=1)

a = numpy.arange(25).reshape(5, 5)
print array_for_sliding_window(a, (3, 2))

This issue is about sparking discussion about

• Do we need such a function?
• Which features are required?
• Which interface should we persue?

After discussion I am willing to draft up a pull request with an implementation we agreed on.