seq_serving.py

import os

import tensorflow as tf


def build_serving_inputs(num_steps, verbose=False):

    if verbose:
        print("About to create inputs using num_steps:", num_steps)  

    serialized_input = tf.placeholder(tf.string, name='serialized_input')
    feature_configs = dict()
    names = []

    if verbose:
        print("About to create feature dict")
    # Create the feature config dict
    # which includes the <insert thing here>
    for i in range(num_steps):  
        name = "encoder{0}".format(i)
        names.append(name)
        shape = []
        dtype = tf.int64
        #feature_configs[name] = tf.FixedLenFeature(shape=shape, dtype=dtype)
        #FixedLenSequenceFeature.
        feature_configs[name] = tf.FixedLenFeature(shape=shape, dtype=dtype)
    
    if verbose:
        print("Feature dict creates")
    
    parsed_example = tf.parse_example(serialized_input, feature_configs)
    
    if verbose:
        print("About to assign name to inputs")
    # use tf.identity() to assign name
    encoder_inputs = []
    for name in names:
        encoder_inputs.append(tf.identity(parsed_example[name], name=name))

    if verbose:
        print("Returning inputs")
    return serialized_input, encoder_inputs


def export_model_to_serving(sess, export_path_base, version, serialized_tf_example, inputs, outputs):
    
    export_path = os.path.join(tf.compat.as_bytes(export_path_base), tf.compat.as_bytes(str(version)))

    print('Exporting trained model to', export_path)

    builder = tf.saved_model.builder.SavedModelBuilder(export_path)

    # Build the signature_def_map.

    prediction_inputs = tf.saved_model.utils.build_tensor_info(serialized_tf_example)
    prediction_outputs =  tf.saved_model.utils.build_tensor_info(outputs)

    prediction_signature = tf.saved_model.signature_def_utils.build_signature_def(
            inputs={
                tf.saved_model.signature_constants.PREDICT_INPUTS: prediction_inputs
                },
            outputs={
                tf.saved_model.signature_constants.PREDICT_OUTPUTS: prediction_outputs
                },
            method_name=tf.saved_model.signature_constants.PREDICT_METHOD_NAME)

    tensor_infos_input = dict()
    for input_ in inputs:
        tensor_infos_input[input_.name] = tf.saved_model.utils.build_tensor_info(input_)
    
    tensor_infos_output = dict()
    for output in outputs:
        tensor_infos_output[output.name] = tf.saved_model.utils.build_tensor_info(output)


    add_signature = tf.saved_model.signature_def_utils.build_signature_def(inputs=tensor_infos_input,
            outputs=tensor_infos_output,
            method_name=tf.saved_model.signature_constants.PREDICT_METHOD_NAME)

    legacy_init_op = tf.group(tf.tables_initializer(), name='legacy_init_op')
    builder.add_meta_graph_and_variables(sess,
            [tf.saved_model.tag_constants.SERVING],
            signature_def_map={'joke': add_signature,
            tf.saved_model.signature_constants.DEFAULT_SERVING_SIGNATURE_DEF_KEY: prediction_signature},
            legacy_init_op=legacy_init_op)

    builder.save()

    print('Done exporting!')