Generating predictions from trained neural net in simulink

Hello,

Background: I have trained a neural network in python using tensorflow and keras, and saved the model in the h5 format. I have succesfully been able to load that model into the matlab environment and make the same predictions using a modified version of the following script:

% Load model into Matlab workspace

loaded_model = importKerasNetwork('Python_Trained_NN.h5');

% Normalize input data

X1_norm = X1 / max(X1);

X2_norm = X2 / max(X2);

X3_norm = X3 / max(X3);

X4_norm = X4 / max(X4);

% Create Input Vector

inputs = [X1_norm, X2_norm, X3_norm, X4_norm];

% Create zeros for predictions with length equal to the length of the time series data I am trying to make predictions on

test_preds = zeros(length(time), 1);

% % Make predictions using inputs

for i = 1:length(time)

test_preds(i) = predict(loaded_model, inputs(i, :));

end

I know there is probably a faster way to do this, but I am limited to matlab 2020a for my application and am unable to upgrade. Also, this works for now which was a good starting point at least. Please feel free to suggest a better way to do this.

Is there any way I can use a matlab function block to get the same behavior in simulink? Again, I am using matlab 2020a, so I do not have access to the full deep learning toolbox library. Right now I am trying the following.

function y = fcn(X1, X2, X3, X4)

%#codegen

inputs = [X1, X2, X3, X4];

y = predict(loaded_model, inputs);

But I get an error saying that 'loaded_model' is an undefined variable even though it is saved in the matlab workspace as a SeriesNetwork object.

Answer 1

[David Willingham]

Hi Mike,

Great to hear your exploring workflows that include TensorFlow and MATLAB/Simulink.

There are more options for running the model in Simulink as of 21a. For example there is a stateful predict for LSTMs: Stateful Predict.

[Mike Jadwin]

Thanks David, unfortunately I have no options for upgrading since my company is stuck on 2020a at least for sometime with no plans to upgrade in the near future. I will push for this upgrade for sure though!

[David Willingham]

Hi Mike,

As an FYI - What are you looking to do with the model in Simulink?

[Mike Jadwin]

My company prototypes our controls software in simulink, and we have many MIL type applications. Before I can get the approval to implement my new NN model into the production software, I will need to show it is capable of at least meeting the previous models performance. With each time step I need the model output to update based on the changing inputs.

[David Willingham]

I have checked back with R2020b, and this example highlights how you can load your own network using the command in the MATLAB function block:

coder.loadDeepLearningNetwork

Note: this example is different in the current release as there are now dedicated blocks.