Multi Input Keras Model with Tfdata

Question

I have a model I am trying to set up with 8 inputs. The first 7 are length 1 IDs that are each fed into embedding layers and these outputs are concatenated with a set of 4 numeric variables.

So in the model definition includes:

input_A = keras.Input(shape=(1,))
input_B = keras.Input(shape=(1,))
input_C = keras.Input(shape=(1,))    
input_D = keras.Input(shape=(1,))    
input_E = keras.Input(shape=(1,)) 
input_F = keras.Input(shape=(1,))
input_G = keras.Input(shape=(1,))

input_nums = keras.Input(shape=(4,))

embed_A = keras.layers.Embedding(1223 + 1, 50)(input_A)
embed_B = keras.layers.Embedding(50 + 1, 25)(input_B)
embed_C = keras.layers.Embedding(1259 + 1, 50)(input_C)
embed_D = keras.layers.Embedding(3995 + 1, 50)(input_D)
embed_E = keras.layers.Embedding(2040 + 1, 50)(input_E)
embed_F = keras.layers.Embedding(174 + 1, 50)(input_F)
embed_G = keras.layers.Embedding(227 + 1, 50)(input_G)

embed_A = keras.layers.Flatten()(embed_A)
embed_B = keras.layers.Flatten()(embed_B)
embed_C = keras.layers.Flatten()(embed_C)
embed_D = keras.layers.Flatten()(embed_D)
embed_E = keras.layers.Flatten()(embed_E)
embed_F = keras.layers.Flatten()(embed_F)
embed_G = keras.layers.Flatten()(embed_G)

x = keras.layers.concatenate([embed_A,embed_B,embed_C,embed_D,embed_E,embed_F,embed_G,input_nums])

Then the model is constructed:

model = keras.Model(inputs=[input_A, input_B, input_C, input_D, input_E, input_F, input_G, input_nums], outputs = [out])

In the tfdataset map function I tried to structure the input data like this but fitting the model produces an error:

# keras needs:  Should return a tuple of either (inputs, targets) or (inputs, targets, sample_weights).
return  (
        (example["A"],example["B"],example["C"],
         example["D"],example["E"],example["F"],
         example["G"],
         
         (example[‘num_A'],example[' num_B '],example[' num_C'],example[' num_D '])
         ),
    
         label)


ValueError: Layer model expects 8 input(s), but it received 11 input tensors

How can I set up the map function of tfdataset to work with this model?

That is pretty difficult with my current data. I was hoping this was pretty straightforward and someone might suggest how to structure what comes out of map that can work with a multi-input network of this type — B_Miner, May 29 '22 at 15:07
Here is one [example](https://keras.io/examples/nlp/multimodal_entailment/) regarding the multi-input and tf.data API. — Innat, May 29 '22 at 15:14

score 0 · Accepted Answer · answered May 30 '22 at 23:50

I found this works as the return out of the map function:

return   (example["A"],example["B"],example["C"],
         example["D"],example["E"],example["F"],
         example["G"],
         
         [example[‘num_A'],example[' num_B '],example[' num_C'],example[' num_D ']]
         ),
    
         label

Multi Input Keras Model with Tfdata

1 Answers1