如何从 python 访问 matlab/octave 模块？

Question

I am looking for a way to access a matlab module from python. My current situation is this:

• I have a python code that does numerical computations by calling Lapack routines while the memory is allocated as ctypes and passed as pointers to the Lapack routines.
• I also have a matlab module, which is compatible with octave, that does some mathematical tricks I want to use.

My question now is this:
What is an efficient way to keep all the main work in python while at the same time exploit the possibilities that matlab/octave modules offer. Also it would be kind of nice, if my ctype arrays do not have to be converted into some other object in order to run octave. However, I can see that that last point is hard to accomplish.

My current research shows me two possible options:

1. Pytave : However it seems that this packages is kind of pre alpha?!
2. Go this humpy road: ctypes -> *.mat file (via numpy) -> octave -> *.mat file -> ctypes (via numpy)

Answer 1

You can use oct2py , which IIUC was started by its author because pytave didn't work on win32. It is successfully used in IPython through its octavemagic extension and I can tell it is easy to use on its own, the code is maintained (I reported a little Unicode bug and the author fixed it in a day) and works well. Most of the times is as simple as:

>>> from oct2py import octave
>>> octave.run("cos(pi / 3)")
'ans =  0.50000'
>>> octave.call("foo", a, b)  # Function in foo.m

For further examples you can check this blog article .

Answer 2

Have you considered using OMPC, http://ompc.juricap.com/ ? I have used it with great success when not wishing to re-write some numerical linear algebra routines. I can imagine that the more esoteric the Matlab commands, the harder it would be to translate... but it might be worth a try. In the end, you're going to want to convert your Matlab code to Python because it will be a bottleneck on speed and performance. The only reason to leave the Matlab code in Matlab format is if it would be an enormous up-front cost to translate it all, which OMPC should mitigate somewhat. Otherwise, it's almost always worth that up-front cost to completely rid yourself of Matlab/Octave dependence.

Answer 3

I had some trouble getting OMPC to work because (I) the md5 module is deprecated, (II) Python 2.6 and later no longer accept arguments for Object.__new__() or Object.__init__(), and (III) the byteplay.py script needed to be updated.

To solve issue (I), I changed line 74 of yacc.py found in the ompc/ directory. This line imports md5 among other things. I deleted the md5 module and added the line below:

from hashlib import md5

Later in the yacc.py script, at line 1160, I changed,

Signature = md5.new()

to the following,

Signature = md5()

To run the code generated by ompcply.py, add 'from ompc import *' to the beginning of the file and then run it with an earlier version of Python, as:

$ python2.5 ompcply.py script.m > newscript.pym
$ python2.5 newscript.pym

Using a version of Python later than 2.5 will give you the following error:

/home/connor/downloads/OMPC/ompclib/ompclib_numpy.py:66: DeprecationWarning: object.__new__() takes no parameters
  a = super(mvar, cls).__new__(cls, *args, **kwargs)

To solve issue (III) I googled byteplay, and replaced the existing script with the newer version.