Import a Neural Network trained in PyTorch in Simulink

Hi everyone,

I would like to know if it is possible to import a neural network trained in PyTorch, into a Simulink model.

Specifically, in PyTorch I have trained a recurrent neural network in a parallel configuration (for simulation purposes), which identifies a dynamical black-box model. I would like to convert this network into a Simulink block, in order to fit it into a simulation model that marches through time.

Thank you in advance.

Marco

Answer 1

[Arkadiy Turevskiy]

Hi Marco,

At the moment the direct import of PyTorch models into MATLAB (and Simulink) is not supported. You can try exporting your PyTorch model to ONNX (open neural network exchange) format. Once the model is in ONNX, you can import it into MATLAB, and once the network is in MATLAB, you can add it to your Simulink model using Stateful Predict block from Deep Learning Toolbox Simulink library.

The other options you might consider:

1. Retrain/recreate the model in Deep Learning Toolbox, so you don't have to convert to ONNX and deal with model import into MATLAB.
2. If you can convert your model to TensrFlow, we support direct import from TensrFlow. Once the model is imported, again, use Stateful Predict block to bring it into Simulink.

Can I aks also: do you need to simulate only or is eventual goal code generation as well?

If none of these options are viable for you, please get in touch with me, and we'll try to support you.

Thanks

Arkadiy

[Marco Pizzoli]

Hi Arkadiy,

Thank you so much for the quick reply.

I was not aware of the existence of the Stateful Predict block.

I think I will try with the first solution you proposed and hope it will be ok.

If there are any problems, I will not hesitate to contact you.

Regarding your question, my only goal is to simulate a Simulink model, inside of which there is the correctly imported neural network interacting with another model (forming a closed-loop system).

Thank you again and best regards,

Marco

[Pierre Harouimi]

You can now import PyTorch model as MATLAB network with importNetworkFromPyTorch function. Note that for now, you can import only image classification models, but support for other model types will be added in future updates. You can follow activities on the Deep Learning Toolbox Converter for PyTorch Models File Exchange support package.

[Pierre Harouimi]

Another approach is to co-execute external code, e.g., call TensorFlow or PyTorch models built in Python, directly from AMTLAB & Simulink. A nice and simple example here:

Call Python Function Using MATLAB Function and MATLAB System Block