medium_tf_numerical_input_processor.py

def numerical_input_processor(inputs):
    if not inputs:
        return

    concat = None
    if len(inputs.values()) > 1:
        concat = tf.keras.layers.Concatenate()(list(inputs.values()))

    norm = tf.keras.layers.experimental.preprocessing.Normalization()

    for batch, _ in get_dataset(batch_size=DUMMY_BATCH_SIZE).take(1):
        data = []

        for k in inputs.keys():
            data.append(np.array(batch[k]))

        data = np.array(data)
        data = np.transpose(data)

        norm.adapt(data)

    for batch, _ in get_dataset(batch_size=BATCH_SIZE):
        data = []

        for k in inputs.keys():
            data.append(np.array(batch[k]))

        data = np.array(data)
        data = np.transpose(data)

        norm.adapt(data, reset_state=False)

    if concat is not None:
        numeric_layer = norm(concat)
    else:
        numeric_layer = norm(list(inputs.values())[0])

    return numeric_layer