tensorflow1.14（一、初识）

默认图与自定义图

import tensorflow as tf
a = tf.constant(1)
b = tf.constant(3)
c = tf.constant(5,name="name1")
d = tf.constant(7,name="name1")   # name：操作名称，可自定义
e = a+b+c+d
print(a,a.graph)
print(b,b.graph)
print(c,c.graph)
print(d,d.graph)
print(e,e.graph)

with tf.Session() as sess:
    e_value = sess.run(e)
    print(e_value)
    print(sess.graph)

# 自定义图------------------------------------------------------------
my_g = tf.Graph()
with my_g.as_default():
    my_a = tf.constant(10)
    my_b = tf.constant(30)
    my_c = tf.constant(50, name="name1")
    my_d = tf.constant(70, name="name1")  # name：操作名称，可自定义
    my_e = my_a + my_b + my_c + my_d
    print(my_a, my_a.graph)
    print(my_b, my_b.graph)
    print(my_c, my_c.graph)
    print(my_d, my_d.graph)
    print(my_e, my_e.graph)

with tf.Session(graph=my_g) as my_sess:
    my_e_value = my_sess.run(my_e)
    print(my_e_value)
    print(my_sess.graph)

Tensor("Const:0", shape=(), dtype=int32)
Tensor("Const_1:0", shape=(), dtype=int32)
Tensor("name1:0", shape=(), dtype=int32)
Tensor("name1_1:0", shape=(), dtype=int32)
Tensor("add_2:0", shape=(), dtype=int32)
16

Tensor("Const:0", shape=(), dtype=int32)
Tensor("Const_1:0", shape=(), dtype=int32)
Tensor("name1:0", shape=(), dtype=int32)
Tensor("name1_1:0", shape=(), dtype=int32)
Tensor("add_2:0", shape=(), dtype=int32)
160

结论：

1. sess.graph和a.graph均可以查询图程序的地址

2. 不同图有不同的命名空间（不同图的操作名称可以重复，同一图程序则不能重复）

操作函数与操作对象

操作函数：tf.constant(Tensor对象)  输入一个Tensor对象，输出一个Const对象

操作对象：Const

一个操作函数会返回操作对象

a = tf.constant(5,name="name1")
print(a)
……
e = a+b+c+d
print(e)

Tensor("name1:0", shape=(), dtype=int32) 

创建了一个名为name1的tf.Operation并返回了一个名为"name1:0"的tf.Tensor

Tensor("add_2:0", shape=(), dtype=int32)  add:指令名称 0:输出序号

创建了一个名为add_2的tf.Operation并返回了一个名为"add_2:0"的tf.Tensor

TensorBoard基础操作

# 将图写入本地生成event文件
tf.summary.FileWriter("../summary",graph=sess.graph)  #一定是运行的时候才能写吗

(tf1) G:\Pyfor4\Tensorflor114_prj>tensorboard --logdir="./summary"

浏览器输入：127.0.0.1:6006

两种会话

tf.InteractiveSession()与a.eval()的配合：用于交互式上下文中的TensorFlow，例如shell

>>>tf.InteractiveSession()
 
>>>a = tf.constant(3)
>>>a
Out[5]: 
>>>a.eval()
Out[6]: 3

tf.Session()用于完整的程序当中sess.run()等同于eva() ：运行图程序并获取张量值

    #试图运行自定义图
    with tf.Session(graph=my_g) as my_sess:   #not an element of this graph.
        # my_C = my_sess.run(c_new)
        # print(my_C)
        print(c_new.eval())  # 等同于my_sess.run(c_new)

*传统会话（需要手动close） 

# 传统的会话方式，为了方便回收通常使用上下文管理器
my_sess = tf.Session(graph=my_g)
my_e_value = my_sess.run(my_e)
print(my_e_value)
my_sess.close()

*tf.Session的两个参数：关于远程调试 

tf.Session(graph=my_g,config=tf.ConfigProto(allow_soft_placement=True,
                                                      log_device_placement=True))

2022-11-18 07:09:01.925172: I tensorflow/core/common_runtime/direct_session.cc:296] Device mapping:

2022-11-18 07:09:01.925773: I tensorflow/core/common_runtime/placer.cc:54] Add: (Add)/job:localhost/replica:0/task:0/device:CPU:0
2022-11-18 07:09:01.925896: I tensorflow/core/common_runtime/placer.cc:54] Const: (Const)/job:localhost/replica:0/task:0/device:CPU:0
2022-11-18 07:09:01.926035: I tensorflow/core/common_runtime/placer.cc:54] Const_1: (Const)/job:localhost/replica:0/task:0/device:CPU:0
2022-11-18 07:09:01.926153: I tensorflow/core/common_runtime/placer.cc:54] name1: (Const)/job:localhost/replica:0/task:0/device:CPU:0
2022-11-18 07:09:01.926272: I tensorflow/core/common_runtime/placer.cc:54] name1_1: (Const)/job:localhost/replica:0/task:0/device:CPU:0
 

会话的run()

sess = tf.Session()
# 计算C的值
print(sess.run(c))
print(c.eval(session=sess))

 同时查看多个Operation：用列表的形式传入

my_eabc_value = my_sess.run([my_e,my_a,my_b,my_c])

[90, 10, 30, 50]

会话中占位符的处理：

占位符：tf.placeholder()

sess.run(sum_ab,feed_dict={a:3.0,b:5.0}) 必须有feed_dict参数

a = tf.placeholder(tf.float32)
b = tf.placeholder(tf.float32)
sum_ab = tf.add(a,b)

with tf.Session() as sess:
    print(sess.run(sum_ab,feed_dict={a:3.0,b:5.0}))  #当叶子为占位符、必须传递feed_dict

会话报错：

张量（Tensor）

n维数组、类似numpy.ndarray的操作

创建：

# 常规
a = tf.constant(10)    # shape=()
b = tf.constant([12,34,45])   # shape=(3,)
c = tf.constant([[12,34,45],
                 [23,45,65]])   # shape=(2, 3)
# 零张量与壹张量
tf.zeros(shape = [3,4])         # dtype=float32
tf.ones(shape = [3,4])          # dtype=float32
# 随机张量
tf.random_normal(shape = [3,4])
tf.random_normal(shape = [3,4], mean=1.75, stddev=0.12)
# 其他……

张量的变换：

1. dtype转变

2.更新静态形状

3. 更新动态形状

#dtype转变 
a = tf.constant([[1,2,3,4],[5,6,7,8]])
print(a)   # Tensor("Const:0", shape=(2, 4), dtype=int32)
res_a = tf.cast(a,dtype=tf.float32)   # 类型转换：并不会改变原始的type ，返回新的张量
print(a)   # Tensor("Const:0", shape=(2, 4), dtype=int32)
print(res_a)  # Tensor("Cast:0", shape=(2, 4), dtype=float32)

# 更新静态形状（在形状没有完全固定下来的占位符）
a_p = tf.placeholder(dtype=tf.float32, shape=[None,None])
b_p = tf.placeholder(dtype=tf.float32, shape=[None,10])
c_p = tf.placeholder(dtype=tf.float32, shape=[3,2])
print(a_p)    # Tensor("Placeholder:0", shape=(?, ?), dtype=float32)
print(b_p)    # Tensor("Placeholder_1:0", shape=(?, 10), dtype=float32)
print(c_p)    # Tensor("Placeholder_2:0", shape=(3, 2), dtype=float32)
a_p.set_shape([2,3])   # 要兼容原有的shape，只能更新没有完全定下来的部分
b_p.set_shape([2,10])
print(a_p)    # Tensor("Placeholder:0", shape=(2, 3), dtype=float32)
print(b_p)    # Tensor("Placeholder_1:0", shape=(2, 10), dtype=float32)

# 更新动态形状 tf.reshape()：可以跨阶，但是不能修改元素的个数
a_p_321 = tf.reshape(a_p, shape=[3,2,1])
print("reshape后的a_p:",a_p_321)   # reshape后的a_p: Tensor("Reshape:0", shape=(3, 2, 1), dtype=float32)

张量的运算

降维运算tf.reduce_all(x)

def math_reduce():
    x = tf.constant([[True,True],
                     [False,False]])
    print(x)
    a1 = tf.reduce_all(x)
    print(a1)
    a2 = tf.reduce_all(x,0)    # 纵轴
    print(a2)
    a3 = tf.reduce_all(x,1)    # 横轴
    print(a2)
    with tf.Session() as sess:
        print(sess.run([x,a1,a2,a3]))

Tensor("Const:0", shape=(2, 2), dtype=bool)
Tensor("All:0", shape=(), dtype=bool)
Tensor("All_1:0", shape=(2,), dtype=bool)
Tensor("All_1:0", shape=(2,), dtype=bool)
[array([[ True,  True],
           [False, False]]), False, array([False, False]), array([ True, False])]

变量OP   

• 存储持久化
• 可修改值
• 可指定被训练

创建、初始化变量

def variable_demo():
    # 定义变量
    a = tf.Variable(initial_value=50)
    b = tf.Variable(initial_value=40)
    c = tf.add(a,b)
    print("a:\n",a)    #  
    print("b:\n",b)    #  
    print("c:\n",c)    #  Tensor("Add:0", shape=(), dtype=int32)

    #初始化变量
    init = tf.global_variables_initializer()

    #开启会话
    with tf.Session() as sess:
        sess.run(init)
        print(sess.run([a,b,c]))   # [50, 40, 90]

    return None

定义新的命名空间 ：tf.variable_scope("my_scope")

• 使得结构更加清晰

    with tf.variable_scope("my_scope"):
        a = tf.Variable(initial_value=50)
        b = tf.Variable(initial_value=40)
    with tf.variable_scope("your_scope"):
        c = tf.add(a,b)
    print("a:\n",a)    # 
    print("b:\n",b)    # 
    print("c:\n",c)    # Tensor("your_scope/Add:0", shape=(), dtype=int32)

其他基础API

高级API