Example showing the a multi-input node Keras model that's trained on toy data and saved as a TensorFlow Serving model artifact.

multi_input_keras_model_tf_serving_save.py

import os
import sys

import numpy as np

import tensorflow as tf
from tensorflow.python.keras import Input, Model
from tensorflow.python.keras.layers import Dense, concatenate


def single_feat_input_model(n_output_per: int) -> Model:
    inputNode = Input(shape=(1,))
    m = Model(
        inputs=inputNode, outputs=Dense(n_output_per, activation="relu")(inputNode)
    )
    return m


def make_multi_single_feat(
    n_input: int, n_output_per: int, n_output_final: int = 1
) -> Model:
    inputs = [single_feat_input_model(n_output_per) for _ in range(n_input)]
    combined = concatenate(list(map(lambda i_layer: i_layer.output, inputs)))
    final_layer = Dense(n_output_final, activation="softmax")(combined)
    model = Model(
        inputs=list(map(lambda i_layer: i_layer.input, inputs)), outputs=final_layer
    )
    return model


def remove(path: str) -> None:
    if os.path.isfile(path):
        os.remove(path)
    elif os.path.isdir(path):
        os.rmdir(path)


if __name__ == "__main__":
    export_path = sys.argv[1] if len(sys.argv) > 1 else "./model_export"
    remove(export_path)

    N_INPUT = 5
    X = np.array(
        [
            [[1], [2], [3], [4], [5]],
            [[3], [1], [5], [17], [18]],
            [[0], [0], [0], [0], [0]],
        ],
        dtype=float,
    )
    X_multi_input_examples = [X[:, i, :] for i in range(N_INPUT)]
    Y = np.array([1, 1, 0], dtype=int)

    model = make_multi_single_feat(n_input=N_INPUT, n_output_per=3, n_output_final=1)
    model.compile(optimizer="adadelta", loss="binary_crossentropy")
    history = model.fit(X_multi_input_examples, Y)

    with tf.keras.backend.get_session() as sess:
        tf.saved_model.simple_save(
            sess,
            export_path,
            inputs={i.name: i for i in model.input},
            outputs={t.name: t for t in model.outputs},
        )