Averaging over every n elements of a numpy array

If your array arr has a length divisible by 3:

np.mean(arr.reshape(-1, 3), axis=1)

Reshaping to a higher dimensional array and then performing some form of reduce operation on one of the additional dimensions is a staple of numpy programming.

For googlers looking for a simple generalisation for arrays with multiple dimensions: the function block_reduce in the scikit-image module (link to docs).

It has a very simple interface to downsample arrays by applying a function such as numpy.mean, but can also use others (maximum, median, ...). The downsampling can be done by different factors for different axes by supplying a tuple with different sizes for the blocks. Here's an example with a 2D array; downsampling only axis 1 by 5 using the mean:

import numpy as np
from skimage.measure import block_reduce

arr = np.stack((np.arange(1,20), np.arange(20,39)))

# array([[ 1,  2,  3,  4,  5,  6,  7,  8,  9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19],
#        [20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38]])

arr_reduced = block_reduce(arr, block_size=(1,5), func=np.mean, cval=np.mean(arr))

# array([[ 3. ,  8. , 13. , 17.8],
#        [22. , 27. , 32. , 33. ]])

As it was discussed in the comments to the other answer: if the array in the reduced dimension is not divisible by block size, padding values are provided by the argument cval (0 by default).