Stride tricks can be useful for local computations on arrays, when the computed value at a given position depends on the neighboring values. Examples include dynamical systems, digital filters, and cellular automata.
In this recipe, we will implement an efficient rolling average algorithm (a particular type of convolution-based linear filter) with NumPy stride tricks. A rolling average of a 1D vector contains, at each position, the average of the elements around this position in the original vector. Roughly speaking, this process filters out the noisy components of a signal so as to keep only the slower components.
The idea is to start from a 1D vector, and make a virtual 2D array where each line is a shifted version of the previous line. When using stride tricks, this process is very efficient as it does not involve any copy.
Let's generate a 1D vector:
>>> import numpy as np from numpy.lib.stride_tricks...