Tensorflow actual combat Google deep learning framework second edition learning summary tensorflow introduction

The content of this paper mainly makes a knowledge summary for the third chapter
summary ：TensorFlow The working principle of is divided into calculation model , Data model and operation model .
1.TensorFlow Calculation model - Calculation chart
1.1 Computational graph model
tensorflow A programming system that expresses computation in the form of a computation diagram , One of those tensorflow Each calculation in is a node on the calculation graph , And the edge between nodes describes the dependency between calculations . The following figure is a typical calculation diagram model

1.2 The calculation chart uses
tensorflow The procedure is generally divided into two stages , The first stage is to define all the calculations in the calculation diagram , In the second stage, the calculation is performed. The default calculation graph can be obtained through the default calculation graph function , The line and surface code demonstrates how to get the default calculation diagram and how to view the calculation diagram ：

print(a.grapy is tf.get_default>graph)

In addition to the default calculation chart , Can pass tf.Graph Generate a new calculation chart . Be careful , The tensor on the newly generated computational graph （ It can be understood as data ） And operations are not shared . For example, the final output results of the following code are 0 and 1.

import tensorflow as tf
g1 = tf.Graph()
with g1.as_default():
    v = tf.get_variable("v", [1], initializer = tf.zeros_initializer()) #  Set the initial value to 0
    g2 = tf.Graph()
with g2.as_default():
    v = tf.get_variable("v", [1], initializer = tf.ones_initializer())  #  Set the initial value to 1
with tf.Session(graph = g1) as sess:
    tf.global_variables_initializer().run()
    with tf.variable_scope("", reuse=True):
        print(sess.run(tf.get_variable("v")))
with tf.Session(graph = g2) as sess:
    tf.global_variables_initializer().run()
    with tf.variable_scope("", reuse=True):
        print(sess.run(tf.get_variable("v")))

2.TensorFlow Calculation model - tensor
The tensor is tensorflow Form of management data , All the data are expressed in the form of tensor , Functionally speaking , Tensor can be simply understood as multidimensional data , Zero order tensors represent scalars , It's a number , The first tensor is a vector , That is, a one-dimensional array , The first n The order tensor can be understood as a n For data .

import tensorflow as tf
a = tf.constant([1.0, 2.0], name="a")
b = tf.constant([2.0, 3.0], name="b")
result = a + b
print result

sess = tf.InteractiveSession ()
print(result.eval())
sess.close()

The output of the code is Tensor(“add:0”, shape=(2,), dtype=float32)
No result of the addition is output , Just a reference to the result , The structure of the instant tensor . From the structure of the tensor, we can see , The first attribute of the tensor is not only the unique identifier of the tensor , The technical process of tensor is also given , With node:src_output Form gives ,add Is the node name ,0 Indicates the number of output nodes . The second attribute is the dimension of the tensor ,shape(2,) The representation is a one-dimensional array , The length is 2. The third attribute is the type .
Note here that the tensor itself does not store specific numbers , Only through the conversation can we get the specific value .

3.TensorFlow Calculation model - conversation
Session owned and managed tensorflow All resources at program run time , So you need to close the session to help the system reclaim resources after the calculation , Otherwise, there will be a memory leak .tensorflow In general use python Context manager to use sessions , Just put all the calculations in with Internal , When the context manager exits, it will automatically release all resources .

with tf.Session() as sess:
    print(sess.run(result))