Usage of tf.variable() function in tensorflow

reference

tensorflow And tf.Variable() function

sketch

tf.Variable() The function is used to create variables (Variable), A variable is a special tensor (), It can be any shape and type of tensor .
The syntax format is as follows :

tf.Variable.init(initial_value, trainable=True, 
collections=None, validate_shape=True, name=None))

Be careful : When starting the diagram ( Before the operation ), All variables must be clearly defined . Variables are often used to store and update parameters , In the process of calculating the chart, its value will be saved until the end of the program , This formally distinguishes general tensors . General Tensorflow In the process of operation, the tensor just flows through the calculation diagram , Will not be preserved ; Related operations involving variables must pass session Session control .

Particular attention ： stay tensorflow The definition and initialization of variables in are separate .

Initialize variable ： Used to tf.initialize_all_variables() function , But in 2017 year 3 month 2 After number , This function is no longer used , use tf.global_variables_initializer() replace tf.initialize_all_variables()

tf.global_variables_initializer() Used to initialize all variables ;w.initializer Used to initialize a single variable .

Besides : You can also use tf.get_variable () Function to define variables . In essence ,tf.Variable() Is the real definition variable , and tf.get_variable() Is to get variables , The main design purpose is to share variables , But if you can't get it, it's equivalent to newly defined variables .
Besides , Use tf.Variable when , If a naming conflict is detected , The system will take care of itself . Use tf.get_variable() when , The system will not handle conflicts , It's a mistake
Besides, besides , Use tf.Variable() when , If duplicate names are detected , The system will automatically handle , The processing method is generally to add "_n". However ,tf.get_variable() If a duplicate name conflict is found, an error will be reported directly .

example

Use tf.Variable() Create variables

import tensorflow._api.v2.compat.v1 as tf
tf.disable_v2_behavior()

w = tf.Variable([6,8,6])	# Create variables w
print(w)	# Look at the shape, It's not worth it .
with tf.Session() as sess:
    sess.run(w.initializer)		# Initialize variable 
    print(sess.run(w))	# Look at the value of the variable 

Output :

''' <tf.Variable 'Variable:0' shape=(3,) dtype=int32_ref> [6 8 6] '''

tf.get_variable() And tf.Variable()

import tensorflow._api.v2.compat.v1 as tf
tf.disable_v2_behavior()

import tensorflow as tf

V1 = tf.Variable([1,2,3],name="V1")
V2 = tf.Variable([2,3,4],name="V1")
V3 = tf.Variable([3,4,5],name="V1_1")

print("V1:",V1.name)
print("V2:",V2.name)
print("V3:",V3.name)

Output :

''' V1: V1:0 V2: V1_1:0 V3: V1_1_1:0 '''

Summary

• Generally, parameters are used initial_value And name That's enough
• tf.get_variable() And tf.Variable() Are used to define variables