Deep learning (tentsorflow2. version) three good student performance problems (1)

 

🥰 Blog's front page ：knighthood2001

Comments are welcome on thumb up ️

️ love python, Looking forward to progress and growth with you ！！️

Catalog

The introduction of "three good" students' achievements

Build a neural network to solve the problem of "three good" students' grades

Code of constructed neural network

Code explanation

---

The introduction of "three good" students' achievements

         Let's look at such a problem :  A school is going to select three good students , We know , Three good students “ Three good " It means having good moral character 、 Study well 、 Good sports : But to be selected , Now we all need quantitative , in other words The school will score according to moral education 、 Intellectual education and physical education 3 To calculate a total score ,  Then determine who can be selected as the "three good students" according to the total score . hypothesis The school's rule for calculating the total score is : Moral education share 60% , Intellectual education accounts for 30%, Sports account for 10% . If this rule is expressed by a formula, it is like this

Total score = Moral education score *0.6 + Intellectual education points * 0.3 + Physical education scores *0.1

You can see , The formula for calculating the total score of three good students is actually 3 Each term score is multiplied by a weight (weight) value , Then add and sum .

---

The above is the background for us to solve the problem . that , We need to Problem solved That's true : Parents with two children , Know your child's 3 Item score and total score , however The school did not tell parents the rules for calculating the total score . Parents guess that the way to calculate the total score is to put 3 The term scores are multiplied by different weights and then added to obtain , The only thing I don't know is how much these weights are . Now parents want to use artificial intelligence neural network method to roughly calculate this 3 How many weights are there respectively . Let's assume that the first parent's child A Of Moral education is divided into 90、 Intellectual education is divided into 80、  Physical education is divided into 70、 The total score is 85, And separately w1、w2、w3 To represent moral education 、 The weight multiplied by the scores of intellectual education and physical education , You can get this formula :

90 * w1 + 80 * w2 + 70 * w3 = 85

Another child B Of Moral education is divided into 98、 Intellectual education is divided into 95、 Physical education is divided into 87、 The total score is 96, We can get this formula :

98 * w1 + 95 * w2 + 87 * w3 = 96

From the mathematical method of solving equations , There is a total of... In these two formulas 3 An unknown number , In theory, as long as there is 3 An inequivalent formula , You can work out the answer . But we only have data for two students , There are only two formulas , It is impossible to solve this problem by solving the equation . So at this point , You can use Neural network method To try to solve this problem .

---

Build a neural network to solve the problem of "three good" students' grades

Theoretical knowledge

① Neural network model diagrams generally contain 1 Input layers 、1 One or more hidden layers , as well as 1 Output layers .

② Generally speaking ,  Input layer It describes the form of input data ; We use squares to represent a number of each input data  ( Or a field ), It is called an input node ; The input node is usually x Named after the , If there are multiple values , Then use x1,x2,...,xn To represent the .

③ Hidden layer It is the most important part to describe the structure of the neural network model we designed ; There may be more than one hidden layer ; There will be... In every layer 1 One or more neurons , We use circles to represent , It's called a neuron node or a hidden node , Sometimes referred to simply as nodes ; Each node receives the data from the previous layer and outputs the data to the next layer after certain operations , Conform to the characteristics of neurons , These operations on neuron nodes are called computational operations or operations （operation, abbreviation op）

④ Output layer It is generally the last layer of the neural network model , Will contain 1 One or more output nodes represented by diamonds , The output node represents the final result of the whole neural network : The nodes of the output layer are generally used to y Named after the , But not necessarily .

⑤ We are in the neural network model diagram , It is generally agreed in the... Of each node Lower right ( Sometimes it's on the lower left because of the crowd ) The name of the tag node , At the node upper left Mark the calculations made by this node , for example ,x1、x2、x3、n1、n2、n3、y Are all node names ,“*w1”、 “*w2”、 “*w3” These represent node operations .

Now let's go back to the model itself , It's a standard Feedforward neural networks , That is, a neural network in which signals are always transmitted forward . The input layer has 3 Nodes x1、x2、x3, They represent the moral education points mentioned above 、 Intellectual education and physical education . Because the problem is relatively simple , Hidden layer we only designed one layer , Among them is 3 Nodes n1、n2、n3, Respectively for 3 Input scores for processing , The way to deal with it is to multiply by 3 A weight w1、w2、w3. The output layer has only one node y,  Because we only require a total score ; node y The operation is to n1、n2、n3 this 3 The value output from the nodes Add to sum .

Code of constructed neural network

import tensorflow as tf
# placeholder and eager execution Are not compatible 
tf.compat.v1.disable_eager_execution()

#  Define three input nodes 
x1 = tf.compat.v1.placeholder(dtype=tf.float32)
x2 = tf.compat.v1.placeholder(dtype=tf.float32)
x3 = tf.compat.v1.placeholder(dtype=tf.float32)

#  Define weights （ Variable parameters ）
w1 = tf.Variable(0.1, dtype=tf.float32)
w2 = tf.Variable(0.1, dtype=tf.float32)
w3 = tf.Variable(0.1, dtype=tf.float32)

#  Hidden layer 
n1 = x1 * w1
n2 = x2 * w2
n3 = x3 * w3

#  Output layer 
y = n1 + n2 + n3

#  conversation , An object that manages the operation of a neural network 
sess = tf.compat.v1.Session()

init = tf.compat.v1.global_variables_initializer()

#  stay sess Run the initialization function in the session 
sess.run(init)

#  Perform a neural network calculation 
result = sess.run([x1, x2, x3, w1, w2, w3, y], feed_dict={x1: 90, x2: 80, x3: 70})
print(result)
# [array(90., dtype=float32), array(80., dtype=float32), array(70., dtype=float32), 0.1, 0.1, 0.1, 24.0]

---

Code explanation

tf.compat.v1.disable_eager_execution()

Need to add this , as a result of placeholder and eager execution Are not compatible

#  Define three input nodes 
x1 = tf.compat.v1.placeholder(dtype=tf.float32)
x2 = tf.compat.v1.placeholder(dtype=tf.float32)
x3 = tf.compat.v1.placeholder(dtype=tf.float32)

Definition 3 Input nodes ,placeholder（ Place holder ）, So called placeholders , That is, when writing the program, you are not sure what number to enter , And only when the program is running will it enter , When programming, just define this node , First “ Take a seat ”.

dtype yes “data type” Abbreviation , Indicates the type of value represented by the placeholder ,tf.float32 yes tensorflow Medium 32 Bit floating point , The box 32 Bit binary to represent a decimal , commonly 32 Bit floating-point numbers can meet the needs of calculation .

#  Define weights （ Variable parameters ）
w1 = tf.Variable(0.1, dtype=tf.float32)
w2 = tf.Variable(0.1, dtype=tf.float32)
w3 = tf.Variable(0.1, dtype=tf.float32)

This is used to define the weight of each score , In the neural network , Neuron parameters such as weights that change frequently during training ,tensorflow Call them variables .

Definition w1、w2、w3 Except that the function uses tf.Variable Out of function , Other and definition placeholders x1、x2、x3 It's similar , There is another difference except that dtype Parameter to specify the value type , also Another initial value parameter was passed in , This parameter is not in the form of a named parameter , This is because tf.Variable The function specifies that the first parameter is used to specify that the initial value of the variable parameter can be seen , We put w1、w2、w3 The initial values of are set to 0.1.

#  Hidden layer 
n1 = x1 * w1
n2 = x2 * w2
n3 = x3 * w3

#  Output layer 
y = n1 + n2 + n3

Here we define the hidden layer and the output layer .

---

This completes the definition of neural network model , Next, let's look at how to input data into this neural network and get the result of the operation .

#  conversation , An object that manages the operation of a neural network 
sess = tf.compat.v1.Session()

init = tf.compat.v1.global_variables_initializer()

#  stay sess Run the initialization function in the session 
sess.run(init)

         So let's define a sess Variable , It contains a session （session） object , We can A conversation is simply understood as an object that manages the operation of a neural network , With the conversation object , Our neural network can be formally operated .

         The first step of session object management neural network , Generally, all variable parameters should be initialized , That is to give all variable parameters their own initial values ,

         First let the variables init be equal to global_variables_initializer  The return value of this function , It returns an object dedicated to initializing variable parameters . Then call the session object sess Member function of run(), close init Variables as parameters , We can initialize all the variable parameters in the neural network model we defined before .run(init) Is in the sess Run the initialization function in the session . What initial values are assigned to each variable parameter , It was defined by us just now w1、w2、w3 Is determined by the first parameter of . We set it to 0.1.

#  Perform a neural network calculation 
result = sess.run([x1, x2, x3, w1, w2, w3, y], feed_dict={x1: 90, x2: 80, x3: 70})
print(result)
# [array(90., dtype=float32), array(80., dtype=float32), array(70., dtype=float32), 0.1, 0.1, 0.1, 24.0]
result = sess.run([x1, x2, x3, w1, w2, w3, y], feed_dict={x1: 98, x2: 95, x3: 87})
print(result)
# [array(98., dtype=float32), array(95., dtype=float32), array(87., dtype=float32), 0.1, 0.1, 0.1, 28.0]

         Here we make a real calculation ,sess.run Functional The first parameter Is an array , Represents which result items we need to view ; Another parameter feed_dict, Represents the data we want to input .

give the result as follows

# [array(90., dtype=float32), array(80., dtype=float32), array(70., dtype=float32), 0.1, 0.1, 0.1, 24.0]

# [array(98., dtype=float32), array(95., dtype=float32), array(87., dtype=float32), 0.1, 0.1, 0.1, 28.0]

After checking ：

90*0.1+80*0.1+70*0.1=24

98*0.1+95*0.1+87*0.1=28

correct , It shows that the result of the neural network is correct .

Follow up the training of neural network , Coming soon ......