CODEWITHSUNDEEP
pythonkerasdeep-learningneural-networkclassification
francesco
francesco

Reputation: 57

What is the correct last layer for a concatenate multi-input deep neural network in Keras?

I am trying to implement with Keras a multi input model for a multiclass classification problem with 3 possible outputs, but i can't understand if it's correct to leave any layers as the last one, or if it should respect the restrictions of the number of class.
So what is the correct (if is possible to define a correct one) architecture between these 2 below?

1)

def createModel(numData, boolData, ordData):
    numIn = ks.Input(shape=numData.shape[1:3], name='numeric')
    x = ks.layers.Masking(mask_value=np.float64(0), input_shape=numData.shape[1:3])(numIn)
    mod1 = ks.layers.LSTM(128, return_sequences=True)(x)
    mod1 = ks.layers.LSTM(128)(mod1)
    model1 = ks.Model(numIn, mod1)

    boolIn = ks.Input(shape=boolData.shape[1], name='boolean')
    mod2 = ks.layers.Dense(128, activation='relu')(boolIn)
    mod2 = ks.layers.Dense(128, activation='relu')(mod2)
    model2 = ks.Model(boolIn, mod2)

    ordIn = ks.Input(shape=ordData.shape[1], name='ordinal')
    mod3 = ks.layers.Dense(128, activation='relu')(ordIn)
    mod3 = ks.layers.Dense(128, activation='relu')(mod3)
    model3 = ks.Model(ordIn, mod3)

    finMod = ks.layers.concatenate([model1.output, model2.output, model3.output])

    out = ks.layers.Dense(3, activation='softmax', name='out')(finMod)
    model = ks.Model(inputs=[model1.input, model2.input, model3.input], outputs=[out])
    return model

    def createModel(numData, boolData, ordData):
        numIn = ks.Input(shape=numData.shape[1:3], name='numeric')
        x = ks.layers.Masking(mask_value=np.float64(0), input_shape=numData.shape[1:3])(numIn)
        mod1 = ks.layers.LSTM(128, return_sequences=True)(x)
        mod1 = ks.layers.LSTM(128)(mod1)
        mod1 = ks.layers.Dense(3, activation='softmax')(mod1) #added layer
        model1 = ks.Model(numIn, mod1)

        boolIn = ks.Input(shape=boolData.shape[1], name='boolean')
        mod2 = ks.layers.Dense(128, activation='relu')(boolIn)
        mod2 = ks.layers.Dense(128, activation='relu')(mod2)
        mod2 = ks.layers.Dense(3, activation='softmax')(mod2) #added layer
        model2 = ks.Model(boolIn, mod2)

        ordIn = ks.Input(shape=ordData.shape[1], name='ordinal')
        mod3 = ks.layers.Dense(128, activation='relu')(ordIn)
        mod3 = ks.layers.Dense(128, activation='relu')(mod3)
        mod3 = ks.layers.Dense(3, activation='softmax')(mod3) #added layer
        model3 = ks.Model(ordIn, mod3)

        finMod = ks.layers.concatenate([model1.output, model2.output, model3.output])

        out = ks.layers.Dense(3, activation='softmax', name='out')(finMod)
        model = ks.Model(inputs=[model1.input, model2.input, model3.input], outputs=[out])
        return model

Obviously if there are other big mistakes please tell me because i am still learning Keras and deep learning and i may have misunderstood some basics.

Upvotes: 1

Views: 180

Answers (1)

Yoan B. M.Sc
Yoan B. M.Sc

Reputation: 1505

The intermediate layers do not have to output 3 categories (as in model #2). There's no "correct way" in this case it's just two different way.

In the case of model #2, the added layer is compressing the information of each channel into 3 cells. It could be interpreted as if each channel was asked to do their own prediction and the last one rely on the intermediate predictions to output the final class.

Whereas in the case of model #1, the final layer has much more connections with the channels layers but the overall model is less deeper.

But none of this can tell you in advance which one is the best in your case.

Upvotes: 1

Related Questions