convert keras or tensorflow or tensorflow-keras model to a predictable saved_model for tensorflow-serving and python predictor

export_model.py

import tensorflow as tf

def export_submodel(sess, inputs, outputs, output_dir):
    output_nodes = list(inputs.values()) + list(outputs.values())
    output_node_names = [node.name.replace(":0", "") for node in output_nodes]
    graph = tf.get_default_graph()
    graph_def = graph.as_graph_def()
    # sess = tf.get_default_session()
    output_graph_def = tf.graph_util.convert_variables_to_constants(
        sess,
        graph_def,
        output_node_names
    )
    with tf.Graph().as_default():
        with tf.Session().as_default():
            # 必须在一个新的session和新的graph里操作
            tf.import_graph_def(
                output_graph_def,
                input_map=None,
                return_elements=None,
                name="",
                op_dict=None,
                producer_op_list=None
            )
            builder = tf.saved_model.builder.SavedModelBuilder(output_dir)
            prediction_signature = tf.saved_model.signature_def_utils.build_signature_def(
                inputs={i: tf.saved_model.utils.build_tensor_info(j) for i, j in inputs.items()},
                outputs={i: tf.saved_model.utils.build_tensor_info(j) for i, j in outputs.items()},
                method_name=tf.saved_model.signature_constants.PREDICT_METHOD_NAME
            )
            builder.add_meta_graph_and_variables(
                sess, [tf.saved_model.tag_constants.SERVING],
                signature_def_map={
                    tf.saved_model.signature_constants.DEFAULT_SERVING_SIGNATURE_DEF_KEY: prediction_signature,
                })
            builder.save()