Unable to import an ONNX network in complied application / issues loading DAG networks from MAT files / cannot include support package in compiled application

I'm having an issue with a deployed application and a DAG/ONNX network:

• The network was originally trained in pytorch and written to an ONNX file.
• I subsequently produced a .mat file with the following routine <updateCnnMat>
• I then run <testCnn> in both deployed mode and in a native environment: it runs successfully in the latter and fails in the former.
• Variable names and file names have been changed for this copy because of IP reasons
• I checked <mccExcludedFiles.log>, which does not indicate that anything important is being excluded
• I confirmed that the file is being included in my path
• I do not get the support package section in the compiler window as suggested by https://www.mathworks.com/help/compiler/manage-support-packages.html
• Note that I'm using R2020a - changing this could be a little painful, but I could do it if really needed.

Any ideas? I feel like I've turned over every stone at this point.

$ cat mccExcludedFiles.log
The List of Excluded Files
Excluded files	 Exclusion Message ID	 Reason For Exclusion	 Exclusion Rule

function updateCnnMat(in,out)

model =importONNXNetwork(in,'outputlayertype','regression');
save(out,'model','-v7.3');

Log output:

~
  DAGNetwork with no properties.

-
mat fail!
  MException with properties:

    identifier: 'MATLAB:structRefFromNonStruct'
       message: 'Dot indexing is not supported for variables of this type.'
         cause: {}
         stack: [4x1 struct]
    Correction: []

-----------
coder fail!
  MException with properties:

    identifier: 'MATLAB:undefinedVarOrClass'
       message: 'Unable to resolve the name coder.internal.loadDeepLearningNetwork.'
         cause: {}
         stack: [3x1 struct]
    Correction: []

-----------
onnx fail!
  MException with properties:

    identifier: 'nnet_cnn:supportpackages:InstallRequired'
       message: 'importONNXNetwork requires the Deep Learning Toolbox Converter for ONNX Model Format support package. To install this support package, use the <a href="matlab: matlab.addons.supportpackage.internal.explorer.showSupportPackages('ONNXCONVERTER', 'tripwire')">Add-On Explorer</a>.'
         cause: {}
         stack: [4x1 struct]
    Correction: []

-----------

function testCnn(din)
 % modelDir is already specified
load cnn.mat model
disp('if you got here, you were able to read the mat file...')
try
    load cnn.mat model
    dout = predict(model, din);
	
    disp('mat win!')
catch e
    disp('mat fail!')
    disp(e)
end
%     model=model0.model;
try
    modelCoder=coder.loadDeepLearningNetwork(fullfile(modelDir, 'cnn.mat'));
    dout = predict(modelCoder, din);
    
    disp('coder win!')
catch e
    disp('coder fail!')
    disp(e)
end

try
    modelOnnx=importONNXNetwork(fullfile(modelDir, 'cnn.onnx'),'outputlayertype','regression');
    dout = predict(modelOnnx, din);
    disp('onnx win!')
catch e
    disp('onnx fail!')
    disp(e)
end

Answer 1

[Pierre Harouimi]

Hi

Could you share the ONNX file, so that I can reproduce it?

It seems there is a dynamic licensing issue...

[Ryan Monroe]

Hi Pierre, I'm DMing you the file, but I have solved this problem through a horrifying combination of hacking the toolbox code, so no need for further effort unless you're simply curious.

EDIT: looks like matlab answers does not have a DM function. If you email me, I can hook you up.

[Pierre Harouimi]

Sure, please give me your email address, thanks.

[Ryan Monroe]

(email redacted since Pierre made contact)

[Xiaoya Chen]

Hi Ryan, I met with the problem too. Could you please tell me how to solve it? Many thanks.

[Pierre Harouimi]

There is a “workaround”: you have to overload the DL objects (SeriesNetwork & DAGNetwork) and add it to your compiled function after.

[Mirko Prezioso]

Hi Pierre, I am having the same problem. Could you simplify your workaround in English? :)

What does overload means, and how I can add the overloaded functions after?

[Federico]

can anyone share me the workaround?

Answer 2

[Federico]

I have the same problem, anyone solved the problem?