Tim Tim - 2 months ago 18
Python Question

How to efficiently convert Matlab engine arrays to numpy ndarray?

I am currently working on a project where I need do some steps of processing with legacy Matlab code (using the Matlab engine) and the rest in Python (numpy).

I noticed that converting the results from Matlab's

matlab.mlarray.double
to numpy's
numpy.ndarray
seems horribly slow.

Here is some example code for creating an ndarray with 1000 elements from another ndarray, a list and an mlarray:

import timeit
setup_range = ("import numpy as np\n"
"x = range(1000)")
setup_arange = ("import numpy as np\n"
"x = np.arange(1000)")
setup_matlab = ("import numpy as np\n"
"import matlab.engine\n"
"eng = matlab.engine.start_matlab()\n"
"x = eng.linspace(0., 1000.-1., 1000.)")
print 'From other array'
print timeit.timeit('np.array(x)', setup=setup_arange, number=1000)
print 'From list'
print timeit.timeit('np.array(x)', setup=setup_range, number=1000)
print 'From matlab'
print timeit.timeit('np.array(x)', setup=setup_matlab, number=1000)


Which takes the following times:

From other array
0.00150722111994
From list
0.0705359556928
From matlab
7.0873282467


The conversion takes about 100 times as long as a conversion from list.

Is there any way to speed up the conversion?

Tim Tim
Answer

Moments after posting the question I found the solution.

For one-dimensional arrays, access only the _data property of the Matlab array.

import timeit
print 'From list'
print timeit.timeit('np.array(x)', setup=setup_range, number=1000)
print 'From matlab'
print timeit.timeit('np.array(x)', setup=setup_matlab, number=1000)
print 'From matlab_data'
print timeit.timeit('np.array(x._data)', setup=setup_matlab, number=1000)

prints

From list
0.0719847538787
From matlab
7.12802865169
From matlab_data
0.118476275533

For multi-dimensional arrays you need to reshape the array afterwards. In the case of two-dimensional arrays this means calling

np.array(x._data).reshape(x.size[::-1]).T
Comments