Transplant is an easy way of calling Matlab from Python.
import transplant
matlab = transplant.Matlab()
# call Matlab functions:
length = matlab.numel([1, 2, 3])
magic = matlab.magic(2)
spectrum = matlab.fft(numpy.random.randn(100))
# inject variables into Matlab:
matlab.signal = numpy.zeros(100)
Python lists are converted to cell arrays in Matlab, dicts are converted to Maps, and Numpy arrays are converted do native Matlab matrices.
All Matlab functions and objects can be accessed from Python.
- Fixes for finding libzmq on Windows (Thank you, hardmstar)
- Now correctly encodes bool ndarrays as logical arrays (thank you, Júlio)
- Fixes working with Matlab packages (Thank you, dani-l)
- Fixes recursion at Matlab shutdown (Thank you, dani-l)
- Should now reliably raise an error if Matlab dies unexpectedly.
- Keyword arguments are now automatically translated to string-value pairs in Matlab.
close
was renamedexit
. Even though Python typically usesclose
to close files and connections, this conflicts with Matlab's ownclose
function.- Matlab will now start Matlab at the current working directory.
- Transplant can now be installed through
pip install transplant
. - You can now use
jvm=False
anddesktop=False
to auto-supply common command line arguments for Matlab.
matlab = transplant.Matlab()
Will start a Matlab session and connect to it. This will take a few
seconds while Matlab starts up. All of Matlab's output will go to the
standard output and will appear interspersed with Python output.
Standard input is suppressed to make REPLs work, so Matlab's input
function will not work.
By default, this will try to call matlab
on the command line. If
you want to use a different version of Matlab, or matlab
is not in
PATH, use Matlab(executable='/path/to/matlab')
.
By default, Matlab is called with -nodesktop
and -nosplash
(and -minimize
on Windows), so no IDE or splash screen show up.
You can change this by setting desktop=True
.
You can start Matlab without loading the Java-based GUI system
('-nojvm'
) by setting jvm=False
. This will speed up startup
considerably, but you won't be able to open figures any more.
If you want to start Matlab with additional command line arguments,
you can supply them like this: Matlab(arguments=['-c licensefile'])
.
By default, Matlab will be started on the local machine. To start
Matlab on a different computer, supply the IP address of that
computer: Matlab(address='172.168.1.5')
. This only works if that
computer is reachable through ssh
, Matlab is available on the
other computer's command line, and transplant is in the other Matlab's
path.
Note that due to a limitation of Matlab on Windows, command line output from Matlab running on Windows isn't visible to Transplant.
matlab.disp("Hello, World")
Will call Matlab's disp
function with the argument 'Hello, World'
.
It is equivalent to disp('Hello, World')
in Matlab. Return values
will be returned to Python, and errors will be converted to Python
errors (Matlab stack traces will be given, too!).
Input arguments are converted to Matlab data structures:
Python Argument | Matlab Type |
---|---|
str |
char vector |
float |
double scalar |
int |
an int{8,16,32,64} scalar |
True /False |
logical scalar |
None |
[] |
list |
cell |
dict |
containers.Map |
transplant.MatlabStruct(dict) |
struct |
numpy.ndarray |
double matrix |
scipy.sparse |
sparse matrix |
proxy object | original object |
proxy function | original function |
Return values are treated similarly:
Matlab Return Value | Python Type |
---|---|
char vector |
str |
numeric scalar | number |
logical scalar |
True /False |
[] |
None |
cell |
list |
struct or containers.Map |
dict |
numeric matrix | numpy.ndarray |
sparse matrix | scipy.sparse |
function | proxy function |
object | proxy object |
If the function returns a function handle or an object, a matching Python functions/objects will be created that forwards every access to Matlab. Objects can also be handed back to Matlab and will work as intended.
f = matlab.figure() # create a Figure object
f.Visible = 'off' # modify a property of the Figure object
matlab.set(f, 'Visible', 'on') # pass the Figure object to a Matlab function
In Matlab, some functions behave differently depending on the number
of output arguments. By default, Transplant uses the Matlab function
nargout
to figure out the number of return values for a function.
If nargout
can not determine the number of output arguments
either, Matlab functions will return the value of ans
after the
function call.
In some cases, nargout
will report a wrong number of output
arguments. For example nargout profile
will say 1
, but x =
profile('on')
will raise an error that too few output arguments were
used. To fix this, every function has a keyword argument nargout
,
which can be used in these cases: matlab.profile('on', nargout=0)
calls profile on
with no output arguments. s, f, t, p =
matlab.spectrogram(numpy.random.randn(1000), nargout=4)
returns all
four output arguments of spectrogram
.
All other keyword arguments are transparently translated to key-value
pairs in Matlab, i.e. matlab.struct(a=1, b=2)
is another way of
writing matlab.struct('a', 1, 'b', 2)
.
When working with plots, note that the Matlab program does not wait
for drawing on its own. Use matlab.drawnow()
to make figures
appear.
Note that functions are not called in the base workspace. Functions
that access the current non-lexical workspace (this is very rare) will
therefore not work as expected. For example, matlab.truth = 42
,
matlab.exist('truth')
will not find the truth
variable. Use
matlab.evalin('base', "exist('truth')", nargout=1)
instead in this
case.
If you hit Ctrl-C, the KeyboardInterrupt
will be applied to both
Python and Matlab, stopping any currently running function. Due to a
limitation of Matlab, the error and stack trace of that function will
be lost.
The way multidimensional arrays are indexed in Matlab and Python are fundamentally different. Thankfully, the two-dimensional case works as expected:
Python | Matlab --------------------------+------------------------ array([[ 1, 2, 3], | 1 2 3 [ 10, 20, 30]]) | 10 20 30
In both languages, this array has the shape (2, 3)
.
With higher-dimension arrays, this becomes harder. The next array is again identical:
Python | Matlab --------------------------+------------------------ array([[[ 1, 2], | (:,:,1) = [ 3, 4]], | 1 3 | 10 30 [[ 10, 20], | 100 300 [ 30, 40]], | (:,:,2) = | 2 4 [[100, 200], | 20 40 [300, 400]]]) | 200 400
Even though they look different, they both have the same shape (3,
2, 2)
, and are indexed in the same way. The element at position a,
b, c
in Python is the same as the element at position a+1, b+1,
c+1
in Matlab (+1
due to zero-based/one-based indexing).
You can think about the difference in presentation like this: Python
displays multidimensional arrays as [n,:,:]
, whereas Matlab
displays them as (:,:,n)
.
Matlab processes end when the Matlab
instance goes out of scope or
is explicitly closed using the exit
method. Alternatively, the
Matlab
class can be used as a context manager, which will properly
clean up after itself.
If you are not using the context manager or the exit
method, you
will notice that some Matlab processes don't die when you expect them
to die. If you are running the regular python
interpreter, chances
are that the Matlab process is still referenced to in
sys.last_traceback
, which holds the value of the last exception
that was raised. Your Matlab process will die once the next exception
is raised.
If you are running ipython
, though, all bets are off. I have
noticed that ipython
keeps all kinds of references to all kinds of
things. Sometimes, %reset
will clear them, sometimes it won't.
Sometimes they only go away when ipython
quits. And sometimes,
even stopping ipython
doesn't kill it (how is this even
possible?). This can be quite annoying. Use the exit
method or the
context manager to make sure the processes are stopped correctly.
- Install the zeromq library on your computer and add it to your
PATH. Alternatively, Transplant automatically uses
conda
's zeromq if you use conda. - Install Transplant using
pip install transplant
. This will installpyzmq
,numpy
andmsgpack
as dependencies.
If you want to run Transplant over the network, the remote Matlab has to have access to ZMQ.m and transplant_remote.m and the zeromq library and has to be reachable through SSH.
- Install the latest version of zeromq through your package manager. Install version 4 (often called 5).
- Make sure that Matlab is using the system's version of libstdc++.
If it is using an incompatible version, starting Transplant might
fail with an error like
GLIBCXX_3.4.21 not found
. If you experience this, disable Matlab's own libstdc++ either by removing/renaming $MATLABROOT/sys/os/glnxa64/libstdc++, or by installingmatlab-support
(if you are running Ubuntu).
- Install the latest version of zeromq from here: http://zeromq.org/distro:microsoft-windows OR through conda.
- Install a compiler. See here for a list of supported compilers:
http://uk.mathworks.com/support/compilers/R2017a/ Matlab needs a
compiler in order to load and use the ZeroMQ library using
loadlibrary
.
Transplant opens Matlab as a subprocess (optionally over SSH), then connects to it via 0MQ in a request-response pattern. Matlab then runs the transplant remote and starts listening for messages. Now, Python can send messages to Matlab, and Matlab will respond. Roundtrip time for sending/receiving and encoding/decoding values from Python to Matlab and back is about 2 ms.
All messages are Msgpack-encoded or JSON-encoded objects. You can
choose between Msgpack (faster) and JSON (slower, human-readable)
using the msgformat
attribute of the Matlab
constructor. There
are seven messages types used by Python:
set_global
andget_global
set and retrieve a global variable.del_proxy
removes a cached object.call
calls a Matlab function with some function arguments and returns the result.die
tells Matlab to shut down.
Matlab can then respond with one of three message types:
ack
for successful execution.value
for return values.error
if there was an error during execution.
In addition to the regular Msgpack/JSON data types, _transplant_ uses
specially formatted Msgpack/JSON arrays for transmitting numerical
matrices as binary data. A numerical 2x2 32-bit integer matrix
containing [[1, 2], [3, 4]]
would be encoded as ["__matrix__",
"int32", [2, 2], "AQAAAAIAAAADAAAABAAAA==\n"]
, where "int32"
is
the data type, [2, 2]
is the matrix shape and the long string is
the base64-encoded matrix content. This allows for efficient data
exchange and prevents rounding errors due to JSON serialization. In
Msgpack, the data is not base64-encoded.
When Matlab returns a function handle, it is encoded as
["__function__", func2str(f)]
. When Matlab returns an object, it
caches its value and returns ["__object__", cache_idx]
. These
arrays are translated back to their original Matlab values if passed
to Matlab.
Note that this project includes a Msgpack serializer/parser, a JSON serializer/parser, and a Base64 encoder/decoder in pure Matlab.
- I get errors with integer numbers
Many Matlab functions crash if called with integers. Convert your
numbers to
float
in Python to fix this problem. - How do I pass structs to Matlab?
Since Matlab structs can't use arbitrary keys, all Python
dictionaries are converted to Matlab
containers.Map
instead of structs. Wrap your dicts intransplant.MatlabStruct
in Python to have them converted to structs. Note that this will change all invalid keys to whatever Matlab thinks is an appropriate key name usingmatlab.lang.makeValidName
. - I get errors like
GLIBCXX_3.4.21 not found
Matlab's version of libstdc++ is incompatible with your OS's version. See INSTALLATION GUIDE FOR LINUX for details. - Does Transplant work in Python 2.7? No, it does not.
- How to integrate Transplant with Jupyter?
Use the provided
transplant_magic.py
, to get %%matlab cell magic.
I know of two programs that try to do similar things as Transplant:
- Mathwork's own MATLAB Engine API for Python provides a CPython
extension for calling Matlab code from some versions of Python. In
my experience, it is significantly slower than Transplant, less
feature-complete (no support for non-scalar structs, objects,
methods, packages, numpy), and more cumbersome to use (all arguments
and return values need to be wrapped in a
matlab.double
instead of Numpy Arrays). For a comparison of the two, here are two blog posts on the topic: Intro to Transplant, Transplant speed. - Oct2Py calls Octave from Python. It is very similar to Transplant, but uses Octave instead of Matlab. This has huge benefits in startup time, but of course doesn't support all Matlab code.
Transplant is a side project of mine that I use for running cross-language experiments on a small compute cluster. As such, my usage of Transplant is very narrow, and I do not see bugs that don't happen in my typical usage. That said, I have used Transplant for hundreds of hours, and hundreds of Gigabytes of data without errors.
If you find a bug, or would like to discuss a new feature, or would like to contribute code, please open an issue on Github.
I do not have a Windows machine to test Transplant. Windows support might contain bugs, but at least one user has used it on Windows in the past. If you are hitting problems on Windows, please open an issue on Github.
Running Transplant over the network might contain bugs. If you are hitting problems, please open an issue on Github.
Finally, I would like to remind you that I am developing this project for free, and in my spare time. While I try to be as accomodating as possible, I can not guarantee a timely response to issues. Publishing Open Source Software on Github does not imply an obligation to fix your problem right now. Please be civil.