tensorflow 三种模型:ckpt、pb、pb-savemodel

1、CKPT
　　目录结构
　　　　　 checkpoint：
　　　　　model.ckpt-1000.index

　　　　model.ckpt-1000.data-00000-of-00001
　　　　model.ckpt-1000.meta

　　特点：
　　　　首先这种模型文件是依赖 TensorFlow 的，只能在其框架下使用；
　　　　数据和图是分开的
　　　　这种在训练的时候用的比较多。
　　代码：就省略了

2、pb模型-只有模型
　　这种方式只保存了模型的图结构，可以保留隐私的公布到网上。
　　感觉一些水的论文会用这种方式。
　　代码：
　　　　thanks：https://www.jianshu.com/p/9221fbf52c55
　　·　　

 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

 #模型保存为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")

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

    #只需要修改这一段，定义输入输出，其他保持默认即可
     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))

 #加载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 = {}
     # 测试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") #加载ckpt模型
         export_model(session, m)

     load_pb()

　　

3、pb模型-Saved model
　　这是一种简单格式pb模型保存方式
　　目录结构
　　　　└── 1
　　　　···├── saved_model.pb
　　　　···└── variables
　　　　·········├── variables.data-00000-of-00001
　　　　·········└── variables.index
　　特点：
　　　　对于训练好的模型，我们都是用来进行使用的，也就是进行inference。
　　　　这个时候就不模型变化了。这种方式就将变量的权重变成了一个常亮。
　　　　这样方式模型会变小
　　　　在一些嵌入式吗，用C或者C++的系统中，我们也是常用.pb格式的。
　　代码：
　　　　thanks to https://www.jianshu.com/p/9221fbf52c55

 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

 #模型保存为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")

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

    #只需要修改这一段，定义输入输出，其他保持默认即可
     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))

 #加载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 = {}
     # 测试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") #加载ckpt模型
         export_model(session, m)

     load_pb()