Xtykutl

working in Keras 2.2.4

I have three data features (feature1, feature2, feature3) that correspond to a set of users and how they rank certain items. At present, I have modelled them all separately using a model architecture as follows:

input_vecs = add([feature1_vec, user_vec])

nn = Dropout(0.5)(Dense(128, activation='relu')(input_vecs))

nn = BatchNormalization()(nn)

nn = Dropout(0.5)(Dense(128, activation='relu')(nn))

nn = BatchNormalization()(nn)

nn = Dense(128, activation='relu')(nn)

feature1result = Dense(9, activation='softmax')(nn)



feature1model = Model([feature1_input, user_input], feature1result)

featuremodel.compile('adam', 'categorical_crossentropy')

Each of the models have similar architectures (each separately tuned), and the same style/shape of outputs. I would like to take the preliminary result of the three models and put those into a new layer and then create a final result.

I guess I could run the three models separately, take their output, and then put that output into an entirely new (likely Sequential) model using a structure like this:

model1 = load_model("feature1.h5")

model2 = load_model("feature2.h5")

model3 = load_model("feature3.h5")

merged_model = Sequential()

merged_model.add(merge([model1.layers[-1].output,model2.layers[-1].output,model3.layers[-1].output]))

merged_model.add(Dense(units = 9, activation='relu')) #or whatever units

merged_model.add(Dense(units = 12, activation='relu'))#or whatever units

merged_model.add(Dense(9, activation='softmax'))

However, I'd like to merge them beforehand as it will make managing the models easier later on (i.e., only updating one large model instead of updating 4 smaller models).

How can I do this before doing the compile() and fit() steps?

You need to use the functional API to use the merge layer. I hope this gives you the idea:

input = Input( ... ) 

model1 = load_model("feature1.h5")

model2 = load_model("feature2.h5")

model3 = load_model("feature3.h5")



m1 = model1(input)

m2 = model2(input)

m3 = model3(input)

merged_model = merge([m1,m2,m3]) 

...