MATLAB: Exporting machine learning model to .NET Assembly

cclassificationmachine learningMATLABMATLAB Compilernet

I have created a .NET assembly using the Matlab Library Compiler and included it in a Visual Studio 2015 C# project. Everything works fine except for when I attempt to make a prediction using the model. The Matlab code looks like this:

load('C:\...\dispositionModel.mat', 'trainedModel1');   
result = trainedModel1.predictFcn(data);

I get the following error:

Warning: Could not find appropriate function on path loading function handle C:\Program Files\MATLAB\R2018b\toolbox\stats\mlearnapp\+mlearnapp\+internal\+model\DatasetSpecification.m>@(x)exportableModel.predictFcn(predictorExtractionFcn(x))

I'm not sure why it's calling functions on the R2018b path, instead of the MATLAB Runtime path. I have included a pragma in my function

%#function fitctree

Looking through other help documentation, I'm only seeing references to C/C++ for using machine learning models with generated code. Is it even possible with C#?

Best Answer

The answer was to call the prediction function differently.

ModelClassificationEnsemble = trainedModel1.ClassificationEnsemble;
result = predict(ModelClassificationEnsemble, data);

Related Solutions

MATLAB: Do I get “cannot load an object of class ‘compactclassificationECOC'” after compiling a standalone application of a machine learning model

During the compilation process, a dependency analysis is run to pick up files that the code relies on. In this case, the dependent functions are specified as a value in the trained model struct loaded through a MAT file. This causes the analysis to miss the necessary files.

To work around this, we can add a "function pragma" to the function to make the dependency explicit:

   % App designer method
          function myAppCallback(app, event)
              %#function classreg.learning.classif.CompactClassificationECOC
              [  additional code  ]

The "function pragma" was added based on the "ClassificationSVM" values in the trained model struct:

trainedModel = 
struct with fields:
predictFcn: @(x)exportableModel.predictFcn(predictorExtractionFcn(x))
ClassificationSVM: [1×1 classreg.learning.classif.CompactClassificationECOC]
About: 'This struct is a trained model exported from Classification Learner R2017b.'
HowToPredict: 'To make predictions on a new predictor column matrix, X, use: ...

Multiple functions can be added to the pragma with spaces:

%#function classreg.learning.classif.CompactClassificationECOC classreg.learning.regr.RegressionModel

Additional information about the "function pragma" can be found in the following link (see example 3 in particular):

https://www.mathworks.com/help/compiler_sdk/ml_code/function.html

Some of the warnings will remain, but those are known and expected and will not affect the functionality of the application.

MATLAB: How to eliminate, function not found warnings when using Matlab Runtime

From your description it sounds like you're loading an instance of an object from a MAT-file in your application.

If you're loading an object into your standalone application without giving any indication in the code of the application itself that the definition of the object will need to be included in the application, the dependency analysis may not include it in the application. [See the Tip in the "Fixing Callback Problems: Missing Functions" section on this documentation page, which though it's for the most recent release I believe is still applicable for older releases.]

You're probably going to need to use the %#function pragma or the -a option to include the definition of the class in your application, or you'd need to indicate in the code itself (say by calling the class constructor) something that informs the dependency analysis routine that it needs to include the class definition.

Best Answer

Related Solutions

MATLAB: Do I get “cannot load an object of class ‘compactcl​assificati​onECOC'” after compiling a standalone application of a machine learning model

MATLAB: How to eliminate, function not found warnings when using Matlab Runtime

Related Question

MATLAB: Do I get “cannot load an object of class ‘compactclassificationECOC'” after compiling a standalone application of a machine learning model