How to load machine learning model in python 3.6 that is trained on python 3.5?

Question

I have trained a machine learning model on Python 3.5, and now I switched to Google Colab, which uses Python 3.6, and when I try to load the model that I have trained on Python 3.5, it gives this error:

SystemError: unknown opcode.

After googling, I found that this error occurs because of the environment change, then I cross-checked my python version, and both Python version were different. How can I load my model on Python 3.6?

Answer 1

You shouldn't.

Even if you get it to run without errors / warnings, there might be slight changes under the hood that change the behavior / performance of the model.

You should either retrain the model on Python 3.6 or create a virtual environment that runs Python 3.5 for your model to ensure it performs as expected. Also always ensure that the actual libraries (e.g. keras...) have the same version.

Answer 2

I ran into same problem as you did. I trained my model on a GCP on python 3.5 and move it to colab to continue with evaluation which is python 3.6.

What I did is to reinstantiate the exact model from code, and then call load_weights:

model = create_my_model()

model.load_weights('my_model_trained_with_py_35.h5')

model.save('my_model_py36.h5')

For my case, I don't have lot of custom code other than a Lambda with:

def abs_diff(x):
  return tf.abs(x[0] - x[1])

Since your model could be arbitrarily more complex, this may or may not work, but it is worth a try esp. if re-training is too expensive. As usual, evaluate the model with the same data and ensure nothing is strange.