How do I export a Neural Network from MATLAB?

I have a Neural Network which I trained using MATLAB. I want to export the network so I can use it with other frameworks, for example caffe. How do I do that?

Answer 1

[MathWorks Support Team]

The recently released Neural Network Toolbox Converter for ONNX Model Format now allows one to export a trained Neural Network Toolbox™ deep learning network to the ONNX™ (Open Neural Network Exchange) model format. The ONNX model can then be imported into other deep learning frameworks, such as TensorFlow®, that support ONNX model import.

Alternatively, you could export via the MATLAB Compiler SDK.

Using the MATLAB Compiler SDK, you can save the trained network as a MAT file, and write a MATLAB function that loads the network from the file, performs the desired computation, and returns the Network's output.

You can then compile your MATLAB function into a shared library to be used in your C/C++, .NET, Java, or Python project.

You can find more information about MATLAB Compiler SDK in the following link:

https://uk.mathworks.com/help/compiler_sdk/ml_code/librarycompiler-app.html

Furthermore, the objects that MATLAB uses to represent Neural Networks are transparent, and you can therefore access all the information that describes your trained network.

For example, you will get an object of type SeriesNetwork, which is a trained Convolutional Neural Network. You can then see the weights and biases of the trained network:

convnet.Layers(2).Weights
convnet.Layers(2).Bias

Then, using for example caffe's MATLAB interface, you should be able to save a Convolutional Neural Network as a caffe model. The code for the MATLAB interface is in the following link:

https://github.com/BVLC/caffe/tree/master/matlab

and includes a classification demo that shows you how to use the interface.

Please note that the above code is not developed or supported by MathWorks Technical Support. If you have any questions about how to use the code, please contact the project's developers.

[michael scheinfeild]

should i compile to full caffe in c++ ?

Answer 2

[Maria Duarte Rosa]

The recently released https://www.mathworks.com/matlabcentral/fileexchange/67296-deep-learning-toolbox-converter-for-onnx-model-format now allows one to export a trained Neural Network Toolbox™ deep learning network to the ONNX™ (Open Neural Network Exchange) model format. The ONNX model can then be imported into other deep learning frameworks, such as TensorFlow®, that support ONNX model import.

[michael scheinfeild]

still i have no success to import it to c++ from onnx there are many issues of compilation

Answer 3

[michael scheinfeild]

after testing onnx i found that the output of convolutions is not the same as in matlab .

[Guillaume Vanoost]

Did you use TensorRT to import it to c++ ?

[Shweta Singh]

Hi Michael,

For any further questions or clarification, feel free to contact MathWorks Technical Support by using the following link:

https://www.mathworks.com/support/contact_us.html

Thanks,

Shweta Singh

[Vasil Ppov]

Yes, the output size is not the same! I'm trying to export ONNX YOLO model from Matlab to Python...succesfully! The expected output tensor should have size 14x14x12, however the size in Python is 2x6x196 (2x6=12 and 14*14=196). Could you tell me why and how to fix it ?