TensorFlow基础知识

1. 构造计算图

In [1]: import tensorflow as tf
H:\Software\Anaconda\lib\site-packages\h5py\__init__.py:36: FutureWarning: Conversion of the second argument of issubdtype from `float` to `np.floating` is deprecated. In future, it will be treated as `np.float64 == np.dtype(float).type`.
  from ._conv import register_converters as _register_converters

In [2]: a = tf.constant(1,dtype=tf.float32,shape=[1,1],name="a")    #‘a’,‘b’,‘c’等都是TensorFlow的（张量）对象，可以理解为一个一维数组

In [3]: print(a)
Tensor("a:0", shape=(1, 1), dtype=float32)

In [4]: b = tf.constant(1,dtype=tf.float32,shape=[1,1],name="b")    #constant函数的第一个参数表示生成常量的值，是必备参数，‘dtype’和‘shape’是非必备参数，前者表示数据类型，后者表示张量的维度，表示生成一个一行一列的矩阵常量

In [5]: print(b)
Tensor("b:0", shape=(1, 1), dtype=float32)

In [6]: c = tf.constant(1,dtype=tf.float32,shape=[1,1],name="c")    #一行一列矩阵相当于普通数字，但是TensorFlow里矩阵运算速度远高于数字

In [7]: print(c)
Tensor("c:0", shape=(1, 1), dtype=float32)

In [8]: #s=a+b

In [9]: s = tf.add(a,b,name="sum")

In [10]: print(s)
Tensor("sum:0", shape=(1, 1), dtype=float32)

In [11]: #re=c*s

In [12]: re = tf.multiply(c,s,name="multiply")

In [13]: print(re)
Tensor("multiply:0", shape=(1, 1), dtype=float32)

2. 执行定义好的计算图

为了执行计算图，需要创建session（会话）对象，通过session对象的run函数执行计算图上定义的运行，这一步骤也被称为数据的取回（fetches）

In [14]: #启动与外部运算的会话，通过‘run‘函数计算定义好的图

In [15]: sees = tf.Session()

In [16]: print(sees.run(s))
[[2.]]

In [17]: print(sees.run(re))
[[2.]]

3. TensorFlow提供的变量

除了constant（常量）外，TensorFlow还提供两种变量:Variable(可变变量)，placeholder（占位符）

In [3]: x = tf.Variable(0,name="counter")

In [4]: one = tf.constant(1)

In [5]: #对可变变量赋值

In [6]: update = tf.assign(x,tf.add(x,one))

In [7]: #对可变变量初始化

In [8]: init = tf.global_variables_initializer()

In [9]: sess = tf.Session()

In [10]: sess.run(init)

In [11]: print(sess.run(x))
0

In [12]: sess.run(update)
Out[12]: 1

In [13]: print(sess.run(x))
1

In [14]: sess.run(update)
Out[14]: 2

In [15]: print(sess.run(x))
2

In [16]: import numpy as np

In [17]: #定义占位符mat1和mat2

In [18]: mat1 = tf.placeholder(tf.float32,shape=[1,3],name="mat2")

In [19]: mat1 = tf.placeholder(tf.float32,shape=[3,1],name="mat1")

In [20]: mat1 = tf.placeholder(tf.float32,shape=[1,3],name="mat2")

In [21]: mat2 = tf.placeholder(tf.float32,shape=[3,1],name="mat1")

In [22]: output = tf.matmul(mat1,mat2)

In [23]: #将数据传给占位符，并进行计算

In [25]: print(sess.run([output],feed_dict={
    ...: mat1:np.array([1,2,3]).reshape(1,3),
    ...: mat2:np.array([4,5,6]).reshape(3,1)}))
[array([[32.]], dtype=float32)]