Why is numpy.ndarray.T so much faster than numpy.transpose(numpy.ndarray)?

Question

Recently I came across someone using numpy.transpose instead of numpy.ndarray.T. I was curious so I timed it:

from timeit import timeit
import numpy as np

array1015 = np.random.rand(10,15)

def nptrans():
    np.transpose(array1015)

def npt():
    array1015.T

print(timeit(nptrans))
print(timeit(npt))

The results were:

np.transpose: 1.25864219666

np.ndarray.T: 0.720939874649

Why? Shouldn't they be doing the same thing under the hood? Maybe np.transpose is doing some sort of error checking or something that slows it down?

MSeifert · Accepted Answer · 2017-02-02 03:08:11Z

7

First, the operations are so fast it doesn't really matter if one optimizes there!

%timeit nptrans()  # 100000 loops, best of 3: 2.11 µs per loop
%timeit npt()      # 1000000 loops, best of 3: 905 ns per loop

Optimizing this doesn't make sense except you would be doing millions of transposes and nothing else. Even adding them is much, much slower:

%timeit array1015 + array1015  # 100000 loops, best of 3: 3.55 µs per loop

and addition should be really, really fast!

However there is some overhead in np.transpose that isn't present in np.ndarray.T:

np.transpose in the end calls the objects .transpose-method which means it has to look-up the method of the object and call it.
To avoid repeating their code the developers packed the function that actually calls the method in a seperate function. => One more function call.

So the overhead you see is the result of 2 function calls and a getattr-call. The np.transpose function is actually python so you can see the overhead quite easily (I removed comments):

def transpose(a, axes=None):
    return _wrapfunc(a, 'transpose', axes)  # extra function call

def _wrapfunc(obj, method, *args, **kwds):
    try:
        return getattr(obj, method)(*args, **kwds)  # here it finally calls ndarray.transpose()
    except (AttributeError, TypeError):
        return _wrapit(obj, method, *args, **kwds)

edited Feb 2, 2017 at 3:08

answered Feb 2, 2017 at 2:50

MSeifert

154k41 gold badges356 silver badges377 bronze badges

Sign up to request clarification or add additional context in comments.

1 Comment

wordsforthewise Over a year ago

I see. Well the code in question was a neural network for a Udacity project, so it might make some difference, I'll have to check.

Collectives™ on Stack Overflow

Why is numpy.ndarray.T so much faster than numpy.transpose(numpy.ndarray)?

1 Answer 1

1 Comment

Your Answer

Hot Network Questions

Collectives™ on Stack Overflow

1 Answer 1

1 Comment

Your Answer

Sign up or log in

Post as a guest

Related