Error while calling a Keras Model from Matlab

Basically, I have regression neural network saved in .H5 format. I want to use the NN model's prediction in my simulink model. I am trying to call the python where the NN model is imported using Keras.load_model. When I call the python with NN model imported I am getting an error but if I call it without the NN model its working fine.

Matlab function to call python:

function y = fcn(in1,in2)

y = 0; % Has to be preassigned, otherwise Simulink throws an error

coder.extrinsic('py.final10.test') % Python functions have to be run extrinsically, meaning no C code generated

y = double(py.final10.test(in1,in2));

end

Python :final10.py

from tensorflow import keras

from keras.models import load_model

import numpy as np

model = load_model('model.h5')

def test(in1,in2):

x1 = in1

x2 = in2

a = np.asarray(x1)

b = np.asarray(x2)

G = np.asarray([a,b])

x3 = G.reshape(1,2)

x4 = model.predict(x3)

print(x4)

y = x1+x2;

return y

If I didn't include the load_model line, I am able to call the python from Matlab. Can anyone help me to solve this issue.

Below is the error, I am getting:

Undefined function 'py.finalv10.test' for input arguments of type 'double'. Error in 'active_vehicleV1/Vehicle/MATLAB Function' (line 4) y = double(py.finalv10.test(in1,in2));

Answer 1

[surya venu]

Hi,

The issue arises because Simulink attempts to generate C code for the py.final10.test function when the load_model line is included. However, C code generation fails for functions involving external libraries like TensorFlow.

Here's how you can solve the problem:

Python Script for Model Loading (model_loader.py):

from tensorflow import keras
from keras.models import load_model

model = load_model('model.h5')

def get_model():
    global model
    return model

Python Script for Prediction (final10.py):

import numpy as np

def test(in1, in2, loaded_model):
    x1 = in1
    x2 = in2
    a = np.asarray(x1)
    b = np.asarray(x2)
    G = np.asarray([a, b])
    x3 = G.reshape(1, 2)
    x4 = loaded_model.predict(x3)
    # print(x4)  # Optional for debugging
    y = x1 + x2
    return y

# Load the model outside the function to avoid reloading on every call
loaded_model = get_model()

MATLAB Function (fcn.m):

function y = fcn(in1, in2)
    y = 0; % Preassign

    coder.extrinsic('py.final10.test', 'py.model_loader.get_model');

    % Load the model in the function for better memory management
    loaded_model = py.model_loader.get_model();
    y = double(py.final10.test(in1, in2, loaded_model));
end

Hope it helps