CODEWITHSUNDEEP
pythonmergekerasmodellstm
mobelahcen
mobelahcen

Reputation: 424

Merge different Keras models into ONE

I'm trying to forecast time series using LSTMs. In order to reduce variance, I tried to predict using 3 models and take the average of the 3 which sort of gave me better results. After training and validation,I now want to save my model for future forecasts. However, since I have 3 different models, I would like to know if it's possible to merge them into ONE model and then save/load it or if I should save all models one by one and load them later for future predictions?

# fit 3 models
   model1 = fit_lstm(train_scaled, batch_size,nb_epochs, nb_neurons)
   model2 = fit_lstm(train_scaled, batch_size,nb_epochs, nb_neurons)
   model3 = fit_lstm(train_scaled, batch_size,nb_epochs, nb_neurons)

# predict on test set using 3 models
   forecast1 = model1.predict(test_reshaped, batch_size=batch_size)
   forecast2 = model2.predict(test_reshaped, batch_size=batch_size)
   forecast3 = model3.predict(test_reshaped, batch_size=batch_size)

Upvotes: 1

Views: 496

Answers (1)

VnC
VnC

Reputation: 2016

You are after an ensemble model.

Something like the following:

from keras.models import load_model
models=[]
for i in range(numOfModels):

    modelTemp=load_model(path2modelx) # load model
    modelTemp.name="aUniqueModelName" # change name to be unique
    models.append(modelTemp)


def ensembleModels(models, model_input):
    # collect outputs of models in a list
    yModels=[model(model_input) for model in models] 
    # averaging outputs
    yAvg=layers.average(yModels) 
    # build model from same input and avg output
    modelEns = Model(inputs=model_input, outputs=yAvg,    name='ensemble')  

    return modelEns



model_input = Input(shape=models[0].input_shape[1:]) # c*h*w
modelEns = ensembleModels(models, model_input)
model.summary()

Save the ensemble model:

modelEns.save(<path_to_model>)

Load and predict:

modelEns=load_model(<path_to_model>)
modelEns.summary()
y=modelEns.predict(x)

Source

Check this article as well.

Upvotes: 1

Related Questions