Tensorflow 使用pb文件保存(恢复)模型计算图和参数实例详解

Nadia ·

更新时间:2024-09-21

· 712 次阅读

一、保存：

graph_util.convert_variables_to_constants 可以把当前session的计算图串行化成一个字节流（二进制），这个函数包含三个参数：参数1：当前活动的session，它含有各变量

参数2：GraphDef 对象，它描述了计算网络

参数3：Graph图中需要输出的节点的名称的列表

返回值：精简版的GraphDef 对象，包含了原始输入GraphDef和session的网络和变量信息，它的成员函数SerializeToString()可以把这些信息串行化为字节流，然后写入文件里：


constant_graph = graph_util.convert_variables_to_constants( sess, sess.graph_def , ['sum_operation'] )
with open( pbName, mode='wb') as f:
f.write(constant_graph.SerializeToString())

需要指出的是，如果原始张量（包含在参数1和参数2中的组成部分）不参与参数3指定的输出节点列表所指定的张量计算的话，这些张量将不会存在返回的GraphDef对象里，也不会被串行化写入pb文件。

二、恢复：

恢复时，创建一个GraphDef，然后从上述的文件里加载进来，接着输入到当前的session：


    graph0 = tf.GraphDef()
    with open( pbName, mode='rb') as f:
      graph0.ParseFromString( f.read() )
      tf.import_graph_def( graph0 , name = '' )

三、代码：


import tensorflow as tf 
from tensorflow.python.framework import graph_util
pbName = 'graphA.pb'
def graphCreate() :
  with tf.Session() as sess :
    var1 = tf.placeholder ( tf.int32 , name='var1' ) 
    var2 = tf.Variable( 20 , name='var2' )#实参name='var2'指定了操作名，该操作返回的张量名是在
                       #'var2'后面:0 ,即var2:0 是返回的张量名，也就是说变量
                       # var2的名称是'var2:0'
    var3 = tf.Variable( 30 , name='var3' )
    var4 = tf.Variable( 40 , name='var4' )
    var4op = tf.assign( var4 , 1000 , name = 'var4op1' )
    sum = tf.Variable( 4, name='sum' )
    sum = tf.add ( var1 , var2, name = 'var1_var2' ) 
    sum = tf.add( sum , var3 , name='sum_var3' )
    sumOps = tf.add( sum , var4 , name='sum_operation' )
    oper = tf.get_default_graph().get_operations()
    with open( 'operation.csv','wt' ) as f:
      s = 'name,type,output\n'
      f.write( s ) 
      for o in oper:
        s = o.name
        s += ','+ o.type 
        inp = o.inputs
        oup = o.outputs
        for iip in inp :
          s #s += ','+ str(iip)
        for iop in oup :
          s += ',' + str(iop)
        s += '\n'
        f.write( s ) 
      for var in tf.global_variables():
        print('variable=> ' , var.name) #张量是tf.Variable/tf.Add之类操作的结果，
                        #张量的名字使用操作名加:0来表示
    init = tf.global_variables_initializer()
    sess.run( init )
    sess.run( var4op )
    print('sum_operation result is Tensor ' , sess.run( sumOps , feed_dict={var1:1}) )
    constant_graph = graph_util.convert_variables_to_constants( sess, sess.graph_def , ['sum_operation'] )
    with open( pbName, mode='wb') as f:
      f.write(constant_graph.SerializeToString())
def graphGet() :
  print("start get:" )
  with tf.Graph().as_default():
    graph0 = tf.GraphDef()
    with open( pbName, mode='rb') as f:
      graph0.ParseFromString( f.read() )
      tf.import_graph_def( graph0 , name = '' )
    with tf.Session() as sess :
      init = tf.global_variables_initializer()
      sess.run(init)
      v1 = sess.graph.get_tensor_by_name('var1:0' )
      v2 = sess.graph.get_tensor_by_name('var2:0' )
      v3 = sess.graph.get_tensor_by_name('var3:0' )
      v4 = sess.graph.get_tensor_by_name('var4:0' )
      sumTensor = sess.graph.get_tensor_by_name("sum_operation:0")
      print('sumTensor is : ' , sumTensor )
      print( sess.run( sumTensor , feed_dict={v1:1} ) ) 
graphCreate()
graphGet()

四、保存pb函数代码里的操作名称/类型/返回的张量：

operation name	operation type	output
var1	Placeholder	Tensor("var1:0"	dtype=int32)
var2/initial_value	Const	Tensor("var2/initial_value:0"	shape=()	dtype=int32)
var2	VariableV2	Tensor("var2:0"	shape=()	dtype=int32_ref)
var2/Assign	Assign	Tensor("var2/Assign:0"	shape=()	dtype=int32_ref)
var2/read	Identity	Tensor("var2/read:0"	shape=()	dtype=int32)
var3/initial_value	Const	Tensor("var3/initial_value:0"	shape=()	dtype=int32)
var3	VariableV2	Tensor("var3:0"	shape=()	dtype=int32_ref)
var3/Assign	Assign	Tensor("var3/Assign:0"	shape=()	dtype=int32_ref)
var3/read	Identity	Tensor("var3/read:0"	shape=()	dtype=int32)
var4/initial_value	Const	Tensor("var4/initial_value:0"	shape=()	dtype=int32)
var4	VariableV2	Tensor("var4:0"	shape=()	dtype=int32_ref)
var4/Assign	Assign	Tensor("var4/Assign:0"	shape=()	dtype=int32_ref)
var4/read	Identity	Tensor("var4/read:0"	shape=()	dtype=int32)
var4op1/value	Const	Tensor("var4op1/value:0"	shape=()	dtype=int32)
var4op1	Assign	Tensor("var4op1:0"	shape=()	dtype=int32_ref)
sum/initial_value	Const	Tensor("sum/initial_value:0"	shape=()	dtype=int32)
sum	VariableV2	Tensor("sum:0"	shape=()	dtype=int32_ref)
sum/Assign	Assign	Tensor("sum/Assign:0"	shape=()	dtype=int32_ref)
sum/read	Identity	Tensor("sum/read:0"	shape=()	dtype=int32)
var1_var2	Add	Tensor("var1_var2:0"	dtype=int32)
sum_var3	Add	Tensor("sum_var3:0"	dtype=int32)
sum_operation	Add	Tensor("sum_operation:0"	dtype=int32)