np.mean() vs np.average() in Python NumPy?

Question

np.average takes an optional weight parameter. If it is not supplied they are equivalent. Take a look at the source code: Mean, Average

np.mean:

try:
    mean = a.mean
except AttributeError:
    return _wrapit(a, 'mean', axis, dtype, out)
return mean(axis, dtype, out)

np.average:

...
if weights is None :
    avg = a.mean(axis)
    scl = avg.dtype.type(a.size/avg.size)
else:
    #code that does weighted mean here

if returned: #returned is another optional argument
    scl = np.multiply(avg, 0) + scl
    return avg, scl
else:
    return avg
...

np.mean always computes an arithmetic mean, and has some additional options for input and output (e.g. what datatypes to use, where to place the result).

np.average can compute a weighted average if the weights parameter is supplied.

In some version of numpy there is another imporant difference that you must be aware:

average do not take in account masks, so compute the average over the whole set of data.

mean takes in account masks, so compute the mean only over unmasked values.

g = [1,2,3,55,66,77]
f = np.ma.masked_greater(g,5)

np.average(f)
Out: 34.0

np.mean(f)
Out: 2.0

In addition to the differences already noted, there's another extremely important difference that I just now discovered the hard way: unlike np.mean, np.average doesn't allow the dtype keyword, which is essential for getting correct results in some cases. I have a very large single-precision array that is accessed from an h5 file. If I take the mean along axes 0 and 1, I get wildly incorrect results unless I specify dtype='float64':

>T.shape
(4096, 4096, 720)
>T.dtype
dtype('<f4')

m1 = np.average(T, axis=(0,1))                #  garbage
m2 = np.mean(T, axis=(0,1))                   #  the same garbage
m3 = np.mean(T, axis=(0,1), dtype='float64')  # correct results

Unfortunately, unless you know what to look for, you can't necessarily tell your results are wrong. I will never use np.average again for this reason but will always use np.mean(.., dtype='float64') on any large array. If I want a weighted average, I'll compute it explicitly using the product of the weight vector and the target array and then either np.sum or np.mean, as appropriate (with appropriate precision as well).