how to do model prediction using python in matlab

this is the python code, and it works so well in python

import keras
from keras.models import load_model
from keras.preprocessing import image
import json
import numpy as np
filepath = 'model2.h5'
test_img = 'falling14145853.png'
model = keras.models.load_model(filepath)
img = image.load_img(test_img, target_size=(150, 150))
x = image.img_to_array(img)
x = np.expand_dims(x, axis = 0)

images = np.vstack([x])
classes = model.predict(images, batch_size = 10)

_mapping = ["falling", "sitting", "standing"]
a = np.concatenate(classes, axis=None).astype(int).tolist()
idx = np.argmax(a[0])
prediction = _mapping[idx]
response_json = {"Prediction" : prediction}
data_in_json = json.dumps(response_json)
with open("prediction.json", "w") as out:
    out.write(data_in_json)
    

and this is in matlab (I'm using R2019b)

    system('test_model.py'); %doing prediction

this is the matlab output and i've tried to run pyhton to make a capture using cv and works so well in matlab but when I tried to run python code which contain of keras, its showing this error

>> bismillahbisaaaa

Traceback (most recent call last):

File "E:\angie\test_model.py", line 1, in <module>

import keras

ImportError: No module named keras

[Mudit Chaturvedi]

Hello!

I understand you are trying to do model prediction using python in Matlab and are facing an error while trying to import keras module.

The import statement does not have the same functionality in Matlab as in Python. Please refer to the Understanding Python and MATLAB import Commands documentation for more details.

To call Python libraries from MATLAB, install a supported version of the reference implementation (CPython) for Python. MATLAB supports versions 2.7, 3.6, and 3.7. Refer to this R2019b documentation link for more information.

[Eugenia Angela Salsabillah]

Thank you for your suggestion, I would like to try to follow that documentation first. Also I'm using python 3.7 already. And I have tried to running cv2 using python in matlab and it works so well without adding py. like in the documentation

Answer 1

[Peter O]

So it looks like it's launching into Python fine. A couple things to try:

1. Standard Python or Anaconda? Sometimes Anaconda doesn't play so well. Make sure you're launching matlab from the Anaconda prompt if you go this route, and sometimes the python libraries that depend on dll's will still have some trouble. If OpenCV is playing nicely in a python call through MATLAB, then chances are you're OK there.
2. Is it a case of environments? Is the activated environment missing keras? Is it a module import problem? Does numpy or another library import without issue if you do that first?
3. Try calling the python file through the system call as 'python test_model.py'. This should be no different than calling the file, but stranger things have happened.

[Eugenia Angela Salsabillah]

I'm using standard python, not anaconda. And I just realize the error comes when using import keras or tensorflow only. It's just fine when I called numpy and etc. Also it works so well when I tried to calling the python file through the system call as 'python test_model.py'. So, any other suggestion due to this is very important for my thesis

[Peter O]

Are you using a GPU component for them? My hunch is that some hook file is getting activated with the interpreter's startup routine (maybe an environment variable pointing to a required DLL or it using a different environment) when you use

system("python test_model.py")

which doesn't happen when you just pass it the file directly using

system("test_model.py")

which relies on the formulation from the shell command. Maybe the python run from the shell has a different permission set than the full interpreter, particularly if it gets called from inside MATLAB? Keras stores a configuration file in your home directory -- maybe it fails quietly with insufficient privileges?

This is about the extent of my knowledge on Python environment/startup configuration.

Is there any reason the former invocation doesn't work for your purposes? Seems like it would occupy the same amount of code and overhead in the m-file, and there's no difference in the output returned. If you need to send it multiple files you can always concatenate them at run-time:

files_to_process = ["a.py", "b.py"];
for ix=1:numel(files_to_process)
    my_cmd = "python " + files_to_process(ix);
    system(my_cmd)
end

[Eugenia Angela Salsabillah]

Thank you so much. I'm so excited because the problem is solved by adding 'python' and just it.

So, the code becomes

system('python test_model.py');