[Numpy-discussion] overhead in np.matrix

[Pieter Eendebak]

The overhead of the np.matrix class is quite high for small matrices. See
for example the following code:

import time
import math
import numpy as np

def rot2D(phi):
    c=math.cos(phi);
    return np.matrix(c)

_b=np.matrix(np.zeros( (1,)))
def rot2Dx(phi):
    global _b
    r=_b.copy()
    c=math.cos(phi);
    r.itemset(0, c)
    return r

phi=.023

%timeit rot2D(phi)
%timeit rot2Dx(phi)

The second implementation performs much better by using a copy instead of a
constructor. Is there a way to efficiency create a new np.matrix object?
For other functions in my code I do not have the option to copy an existing
matrix, but I need to construct a new object or perform a cast from
np.array to np.matrix.

I am already aware of two alternatives:

- Using the new multiplication operator (
https://www.python.org/dev/peps/pep-0465/). This is a good solution, but
only python 3.5
- Using the .dot functions from np.array. This works, but personally I like
the notation using np.matrix much better.

I also created an issue on github:
https://github.com/numpy/numpy/issues/6186

With kind regards,
Pieter Eendebak
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.python.org/pipermail/numpy-discussion/attachments/20150811/a370c669/attachment.html>