How to use keras for mutli label multiclass classification

Question

I have a data set with 6 possible labels of the type:

Class 1: Near-large 
Class 2: Far- Large 
Class 3: Near - Medium 
Class 4: Far - Medium 
Class 5: Near - small 
Class 6: far - small 

And I would like to modify the problem to separate the labels so each sample would be classified as far/near and small/medium/small independently, given different features for each classification as input.

My first idea was to train 2 different models for each sublabel and then make a custom function to join the predictions , but I wonder if there's a quicker way of doing it within the Keras framework.

I know I can use the functional API to create two models with independent inputs and two independent outputs. This would give me 2 predictions for 2 different sublabels. If I one hot encode the sublabels the output of those models would be something like this:

Model1.output = [ 0,1 ] or [1,0]  ( far/near) 
Model2.output = [ 1, 0, 0 ] or [0,1,0] or [0,0,1](small/medium/large)

But then how can I merge these two outputs to create a 6 dim vector for the complete labels ?

Model_merged.output = [1, 0, 0, 0, 0 ,0 ] , [010000], ...., [000001] (class1,... ,Class6) 

Answer 1

You can reshape model1 output to extend the axis, mulitply that with model2 output and flatten them both.

from keras.models import Model

reshaped = keras.layers.Reshape((2,-1))(model1.output)
combined = keras.layers.Multiply()([reshaped,model2.output])
flattened = keras.layers.Reshape((6,))(combined)


Combined_model = Model([model1.input,model2.input], flattened)

A simple numpy example of the above would be:

model1_output = np.array([0,1])[:,None] #Reshaped

#array([[0],
#       [1]])

model2_output = np.array([1,0,0])

# array([1, 0, 0])

combined = model1_output*model2_output

#array([[0, 0, 0],
#       [1, 0, 0]])

combined.ravel()

#array([0, 0, 0, 1, 0, 0])