tf.Graph()

1. 它可以通过tensorboard用图形化界面展示出来流程结构

2. 它可以整合一段代码为一个整体存在于一个图中

 

声明情况大体有三种

1. tensor:通过张量本身直接出graph

# -*- coding: utf-8 -*-  
import tensorflow as tf
 
c = tf.constant(4.0)
 
sess = tf.Session()
sess.run(tf.global_variables_initializer())
c_out = sess.run(c)
print(c_out)
print(c.graph == tf.get_default_graph())
print(c.graph)
print(tf.get_default_graph())

输出

4.0
True
<tensorflow.python.framework.ops.Graph object at 0x7f382f9ef110>
<tensorflow.python.framework.ops.Graph object at 0x7f382f9ef110>

2.通过声明一个默认的，然后定义张量内容，在后面可以调用或保存

# -*- coding: utf-8 -*-  
import tensorflow as tf
 
g = tf.Graph()
with g.as_default():
    c = tf.constant(4.0)
 
sess = tf.Session(graph=g)
c_out = sess.run(c)
print(c_out)
print(g)
print(tf.get_default_graph())

输出

4.0
<tensorflow.python.framework.ops.Graph object at 0x7f65f1cb2fd0>
<tensorflow.python.framework.ops.Graph object at 0x7f65de447c90>

3.通过多个声明，在后面通过变量名来分别调用

# -*- coding: utf-8 -*-  
import tensorflow as tf
 
g1 = tf.Graph()
with g1.as_default():
    c1 = tf.constant(4.0)
 
g2 = tf.Graph()
with g2.as_default():
    c2 = tf.constant(20.0)
 
with tf.Session(graph=g1) as sess1:
    print(sess1.run(c1))
with tf.Session(graph=g2) as sess2:
    print(sess2.run(c2))

输出

4.0
20.0

对graph的操作大体有三种

1.保存

# -*- coding: utf-8 -*-  
import tensorflow as tf
 
g1 = tf.Graph()
with g1.as_default():
    # 需要加上名称，在读取pb文件的时候，是通过name和下标来取得对应的tensor的
    c1 = tf.constant(4.0, name='c1')
 
g2 = tf.Graph()
with g2.as_default():
    c2 = tf.constant(20.0)
 
with tf.Session(graph=g1) as sess1:
    print(sess1.run(c1))
with tf.Session(graph=g2) as sess2:
    print(sess2.run(c2))
 
# g1的图定义，包含pb的path, pb文件名，是否是文本默认False
tf.train.write_graph(g1.as_graph_def(),'.','graph.pb',False)

输出

4.0
20.0

并且在当前文件夹下面生成graph.pb文件

2.从pb文件中调用

# -*- coding: utf-8 -*-  
import tensorflow as tf
from tensorflow.python.platform import gfile
 
#load graph
with gfile.FastGFile("./graph.pb",'rb') as f:
    graph_def = tf.GraphDef()
    graph_def.ParseFromString(f.read())
    tf.import_graph_def(graph_def, name='')
 
sess = tf.Session()
c1_tensor = sess.graph.get_tensor_by_name("c1:0")
c1 = sess.run(c1_tensor)
print(c1)

输出
4.0

3.穿插调用

# -*- coding: utf-8 -*-  
import tensorflow as tf
 
g1 = tf.Graph()
with g1.as_default():
    # 声明的变量有名称是一个好的习惯，方便以后使用
    c1 = tf.constant(4.0, name="c1")
 
g2 = tf.Graph()
with g2.as_default():
    c2 = tf.constant(20.0, name="c2")
 
with tf.Session(graph=g2) as sess1:
    # 通过名称和下标来得到相应的值
    c1_list = tf.import_graph_def(g1.as_graph_def(), return_elements = ["c1:0"], name = '')
    print(sess1.run(c1_list[0]+c2))

输出
24.0