How do I operate fitnet function of Matlab by Python?

Hi. I'm a Japanese university student.

I'm researching AI and will be doing neural network analysis using MATLAB and Python.

Previously, We had been manually analyzing hundreds of data one by one using MATLAB's Neural Net Fitting APP (nftool), but now I would like to automate the process of starting MATLAB and analyzing the data using Python.

I would like to execute the following code in Python.

net = fitnet(10,'trainlm');
net = train(net, input, target);
output = net(input);
R = corrcoef(output, target);
R = (1,2)

I've written this in Python like below

import matlab.engine
eng = matlab.engine.start_matlab('-desktop')
eng.workspace['net'] = eng.fitnet(10.)
eng.net = eng.train(eng.net,input,target)
eng.workspace['output'] = eng.net(input)
R = eng.corrcoef(eng.output,eng,target)

Necessary data are loaded at workspace in Matlab by using another Python code.

At Running it, this error occurred.

---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
Input In [78], in <cell line: 3>()
      1 import matlab.engine
      2 eng = matlab.engine.start_matlab('-desktop')
----> 3 eng.workspace['net'] = eng.fitnet(10.)

File ~\anaconda3\envs\python_ex\lib\site-packages\matlabengineforpython-r2021a-py3.8.egg\matlab\engine\matlabengine.py:70, in MatlabFunc.__call__(self, *args, **kwargs)
     68     return FutureResult(self._engine(), future, nargs, _stdout, _stderr, feval=True)
     69 else:
---> 70     return FutureResult(self._engine(), future, nargs, _stdout,
     71                         _stderr, feval=True).result()

File ~\anaconda3\envs\python_ex\lib\site-packages\matlabengineforpython-r2021a-py3.8.egg\matlab\engine\futureresult.py:67, in FutureResult.result(self, timeout)
     64     if timeout < 0:
     65         raise TypeError(pythonengine.getMessage('TimeoutCannotBeNegative'))
---> 67 return self.__future.result(timeout)

File ~\anaconda3\envs\python_ex\lib\site-packages\matlabengineforpython-r2021a-py3.8.egg\matlab\engine\fevalfuture.py:82, in FevalFuture.result(self, timeout)
     79 if not result_ready:
     80     raise TimeoutError(pythonengine.getMessage('MatlabFunctionTimeout'))
---> 82 self._result = pythonengine.getFEvalResult(self._future,self._nargout, None, out=self._out, err=self._err)
     83 self._retrieved = True
     84 return self._result

ValueError:  MATLAB can return only 1-by-N and N-by-1 cell arrays.

This's the similar situation as this questioner.

https://stackoverflow.com/questions/31276805/call-matlab-neural-network-function-from-matlab-python-engine

According to this thread, fitnet returns a variable that Python cannot read, so it cannot process it and is causing this error, and I think so.

My Matlab's varsion is R2021a.

How should I code it to make it work correctly?

Answer 1

[David Willingham]

Hi,

Here is code in python that will call the neural network training:

import matlab.engine
eng = matlab.engine.start_matlab('-desktop')

data = eng.simplefit_dataset(nargout=2)
x = data[0][0]
t = data[1][0]
 
eng.workspace["x"] = x
eng.workspace["t"] = t
eng.evalc("net = fitnet(10.0);")
eng.evalc("net = train(net,x,t);")
eng.evalc("out = net(x);")
out = eng.workspace["out"]
R = eng.corrcoef(out,t)

If you want to test it from MATLAB you can run (where the code above is saved as 'mypythonscript.py':

pyrunfile('mypythonscript.py')

[David Willingham]

Another option is to create functions in MATLAB (attached) which will minimise calls to evalc and then use the following python code to call them:

import matlab.engine
eng = matlab.engine.start_matlab('-desktop')
data = eng.simplefit_dataset(nargout=2)
x = data[0][0]
t = data[1][0]
hiddenSizes = 10.0
modelFilename = eng.myNNTrain(hiddenSizes, x, t)
out = eng.myNNPredict(x)
R = eng.corrcoef(out,t)

Answer 2

[David Willingham]

Try changing this line in python:

eng.workspace['net'] = eng.fitnet(10.)

to

a = matlab.double(10)

eng.workspace['net'] = eng.fitnet(a)

Does this solve your error?

[威吹 塚本]

Thank you for your response.

I tried the method you provided and the following error occurred.

---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
<ipython-input-2-4731bf8d7b8a> in <module>
----> 1 a = matlab.double(10)
      2 eng.workspace['net'] = eng.fitnet(a)

~\anaconda3\Lib\site-packages\matlab\mlarray.py in __init__(self, initializer, size, is_complex)
     49             super(double, self).__init__('d', initializer, size, is_complex)
     50         except Exception as ex:
---> 51             raise ex
     52 
     53 

~\anaconda3\Lib\site-packages\matlab\mlarray.py in __init__(self, initializer, size, is_complex)
     47         """
     48         try:
---> 49             super(double, self).__init__('d', initializer, size, is_complex)
     50         except Exception as ex:
     51             raise ex

~\anaconda3\Lib\site-packages\matlab\_internal\mlarray_sequence.py in __init__(self, typecode, initializer, size, is_complex)
     39         self._python_type = python_type[typecode]
     40         if initializer is not None:
---> 41             init_dims = _get_size(initializer)
     42             try:
     43                 self._size = _normalize_size(size, init_dims)

~\anaconda3\Lib\site-packages\matlab\_internal\mlarray_utils.py in _get_size(initializer)
     74     is_rect, depth = _is_rectangular(initializer)
     75     if not is_rect:
---> 76         raise ValueError("initializer must be a rectangular nested sequence")
     77 
     78     dims = []

ValueError: initializer must be a rectangular nested sequence

[威吹 塚本]

it works correctly! Thank you for your help.