Use imported keras network to predict in Simulink

Hi,

I am trying to imported my trained series network from python into Simulink to do prediction following this previous link:https://www.mathworks.com/matlabcentral/answers/592243-import-keras-tensorflow-model-into-simulink. However I am still getting some errors. Can you help me to look into why? I have tested the imported keras network to do prediction on a single sample in a separate m file (PT_NN_Prediction.m). It works fine. However, when I use above link to create a matlab function to call the imported keras network to do prediction in Simulink. It does not work with some errors poped up.

PT_NN_modeling.slx is the simulink model to test the imported keras network prediction.

PTNN.m function is the function that is used to generate the series network.

2020-11-16_17h59_55.png is a screenshot of the errors I got.

And I am using Matlab2019b.

Thanks,

Chuan Yu

[Sourabh Kondapaka]

Hi Chuan Yu,

Can you upload the model1.json and model1.h5 files for me to check ?

[Chuan Yu]

Hi, I have attached one more zip file with those two files zipped in. Let me know if you need more information. Thanks!

[Chuan Yu]

Hi,

It has been some time since I posted my question here, can someone help to look into my question and give some answers?

Thanks,

Chuan

Answer 1

[Peter Man]

Hi Chuan Yu,

In 19b, there isn't a layer specifically for vector inputs, which is why the imported network contains its input layer as imageInputLayer. That means the network is expecting an image - which is an array of dimensions height x weight x channels, (channels is 3 for rgb and 1 for greyscale). You can emulate a vector input by making your vector input be essentially a height=1 and channel=1 image i.e. the input to the network can be [1 2 3] for example. I see that you have essentially this in your attached file PT_NN_Prediction.m.

Now, as for your Simulink model, your model function is expecting an image input too. That means instead of passing in 3 separate scalar inputs, you need a single vector input. Furthermore, you need to ensure that vector input is the correct orientation (i.e. [1,2,3] not [1;2;3]) which means your MATLAB function probably needs to do a transpose on the input e.g.

function y = NNPrediction(x)

persistent mynet;

if isempty(mynet)

mynet = coder.loadDeepLearningNetwork('PTNN','myPTNN');

end

disp(x);

y = predict(mynet,x'); % x might be passed in as [1;2;3] rather than [1,2,3]

end

If you upgrade to a later version (R2021a), if you import your network, the input layer would instead be a featureInputLayer rather than an imageInputLayer. The featureInputLayer is designed for vectorial inputs.