通常訓練模型的時候是保存ckpt方便接著訓練，但是上線可以保存為pb模型，加載的時候不需要重新定義模型，只用輸入輸出來調用模型。

import os
import tensorflow as tf
from tensorflow.python.saved_model import builder as saved_model_builder
from tensorflow.python.saved_model import (signature_constants, signature_def_utils, tag_constants, utils)

class model():
    def __init__(self):
        self.a = tf.placeholder(tf.float32, [None])
        self.w = tf.Variable(tf.constant(2.0, shape=[1]), name="w")
        b = tf.Variable(tf.constant(0.5, shape=[1]), name="b")
        self.y = self.a * self.w + b

#模型保存為ckpt
def save_model():
    graph1 = tf.Graph()
    with graph1.as_default():
        m = model()
    with tf.Session(graph=graph1) as session:
        session.run(tf.global_variables_initializer())
        update = tf.assign(m.w, [10])
        session.run(update)
        predict_y = session.run(m.y,feed_dict={m.a:[3.0]})
        print(predict_y)

        saver = tf.train.Saver()
        saver.save(session,"model_pb/model.ckpt")


#保存為pb模型
def export_model(session, m):


   #只需要修改這一段，定義輸入輸出，其他保持默認即可
    model_signature = signature_def_utils.build_signature_def(
        inputs={"input": utils.build_tensor_info(m.a)},
        outputs={
            "output": utils.build_tensor_info(m.y)},

        method_name=signature_constants.PREDICT_METHOD_NAME)

    export_path = "pb_model/1"
    if os.path.exists(export_path):
        os.system("rm -rf "+ export_path)
    print("Export the model to {}".format(export_path))

    try:
        legacy_init_op = tf.group(
            tf.tables_initializer(), name='legacy_init_op')
        builder = saved_model_builder.SavedModelBuilder(export_path)
        builder.add_meta_graph_and_variables(
            session, [tag_constants.SERVING],
            clear_devices=True,
            signature_def_map={
                signature_constants.DEFAULT_SERVING_SIGNATURE_DEF_KEY:
                    model_signature,
            },
            legacy_init_op=legacy_init_op)

        builder.save()
    except Exception as e:
        print("Fail to export saved model, exception: {}".format(e))

#加載pb模型
def load_pb():
    session = tf.Session(graph=tf.Graph())
    model_file_path = "pb_model/1"
    meta_graph = tf.saved_model.loader.load(session, [tf.saved_model.tag_constants.SERVING], model_file_path)

    model_graph_signature = list(meta_graph.signature_def.items())[0][1]
    output_tensor_names = []
    output_op_names = []
    for output_item in model_graph_signature.outputs.items():
        output_op_name = output_item[0]
        output_op_names.append(output_op_name)
        output_tensor_name = output_item[1].name
        output_tensor_names.append(output_tensor_name)
    print("load model finish!")
    sentences = {}
    # 測試pb模型
    for test_x in [[1],[2],[3],[4],[5]]:
        sentences["input"] = test_x
        feed_dict_map = {}
        for input_item in model_graph_signature.inputs.items():
            input_op_name = input_item[0]
            input_tensor_name = input_item[1].name
            feed_dict_map[input_tensor_name] = sentences[input_op_name]
        predict_y = session.run(output_tensor_names, feed_dict=feed_dict_map)
        print("predict pb y:",predict_y)

if __name__ == "__main__":

    save_model()

    graph2 = tf.Graph()
    with graph2.as_default():
        m = model()
        saver = tf.train.Saver()
    with tf.Session(graph=graph2) as session:
        saver.restore(session, "model_pb/model.ckpt") #加載ckpt模型
        export_model(session, m)

    load_pb()

save_model 和load_model兩個函數要分開執(zhí)行，第一次注釋掉load，只save，第二次load的時候注釋掉save。因為聲明模型的時候都是用默認圖，變量命名會依次是0,1，load的時候名字對應不上。

保存好的pb模型路徑文件格式為

保存好的pb模型

還有一種保存pb模型的方法，保存的東西只有model.pb沒有variables。

#保存為pb模型，只有model.pb 沒有variables
def export_model_one(session, m, graph):

    output_names = [m.y.op.name] # 定義模型輸出

    input_graph_def = graph.as_graph_def()
    print(m.y.op.name, m.a.op.name)
    output_graph_def = convert_variables_to_constants(session, input_graph_def, output_names)
    output_graph = 'pb_model/model.pb'  # 保存地址
    with tf.gfile.GFile(output_graph, 'wb') as f:
        f.write(output_graph_def.SerializeToString())

#加載pb模型, 只有model.pb 沒有variables
def load_pb_one():
    graph = tf.Graph()
    with graph.as_default():
        output_graph_def = tf.GraphDef()

        with open( 'pb_model/model.pb' , "rb") as f:
            output_graph_def.ParseFromString(f.read())
            tensors = tf.import_graph_def(output_graph_def, name="")
        sess = tf.Session()
        init = tf.global_variables_initializer()
        sess.run(init)

        input = sess.graph.get_tensor_by_name("Placeholder:0")
        output = sess.graph.get_tensor_by_name("add:0")

        # 測試pb模型
        for test_x in [[1], [2], [3], [4], [5]]:
            predict_y = sess.run(output, feed_dict={input: test_x})
            print("predict pb y:", predict_y)

如果需要發(fā)布模型成服務，請看：http://www.itdecent.cn/p/5b74f1bc0178