Predictions are wrong after model.load_weights()

[chirico85]

Hello @rdisipio Thank you a lot for the work.

I am trying to reproduce your work with the VQC. First, training and val accuracy did not enhance during training. After adapting the learning rate, they did.

  # optimizer = tf.keras.optimizers.Adam(learning_rate=1e-3)
  lr_schedule = tf.keras.optimizers.schedules.ExponentialDecay(
      initial_learning_rate=1e-2,
      decay_steps=10000,
      decay_rate=0.9)
  optimizer = tf.keras.optimizers.Adam(learning_rate=lr_schedule)

Now I have a problem with model.predict(txt_embeddings). After saving the model with model.save_weights(modelname), the predictions are not right. On the other hand, when I train the model and predict classes right after model.evaluate(), the predictions are right.

The problem is how I load the model and the saved weights. My approach is:

model = make_model_quantum(parameters) Parameters are equal to the ones used for training. model.load_weights(modelname) model.predict(txt_embeddings)

Can you help me out?

Regards

[rdisipio]

Hi,

thanks for looking into this. Indeed, saving and load models is always a pain in the neck. Have you tried to pickle the whole model? Something like

import pickle

with open('model.pkl', 'wb') as f:
    pickle.dump(model, f)

Also, if your model is based on TensorFlow, you may want to try model.save("models/quantum_model") or model.save("models/quantum_model.h5") or tf.saved_model.save(model, "models/quantum_model").

There are probably other ways, but try these first.

[chirico85]

Hi @rdisipio , Thanks for your answer and sorry for the late reply. All your suggestions did not help, unfortunately. However, I found out that the best way for me to save hybrid models remains to save the weights: model.save_weights(modelname)

You can load the model weights by:

• creating first the same model used for training
• compiling the model
• evaluating the model (I do it, but do not know if necessary)

and finally

model.load_weights(modelname)

Regards