Realize linear regression with tensorflow (including problems and solutions in the process)

The experiment purpose ：

Through this experiment, let students understand TensorFlow The basic idea of constructing the whole neural network training model , master TensorFlow The process and method of realizing linear regression , And learn to use tensorboard graph To view and check the neural network model designed by yourself in a graphical way .
Experimental instruments, equipment and materials ：
Installed with Python Computer running environment .

Experimental content

One ．TensorFlow Environmental installation

1. stay anaconda prompt Window for installation （anaconda3） It's best to create a new virtual environment , Then operate and install in the environment （ The default environment is base）

conda create -n tf2 python=3.6.5   

This is a new one named tf2, also python The version is 3.6.5 An environment of （python The version number should match your own version number ）.
Switch to an environment ：conda activate Environment name .
2. Enter the just created tf2 Environmental Science ：

conda activate tf2

3. install TensorFlow2.4.0

pip install tensorflow-cpu==2.4.0

Or mirror installation pip install tensorflow-cpu==2.4.0 -i https://pypi.tuna.tsinghua.edu.cn/simple
It may be a little long here , Make sure that the network cannot be disconnected . It will install other packages , These packages are also often used in machine learning . If the scarlet letter appears , Install it again . Until it appears successfull installed …, Indicates that it has been successfully installed tensorflow package . The picture below is 2.5.0 Schematic diagram of successful installation of version ：

4. Then if you use pycharm Tool software , No need to install TensorFlow, Just configure the environment

Two ． utilize TensorFlow Carry out linear regression experiment

Given a batch by y = 3x + 2 Generated data sets (x, y), Build a linear regression model h(x) = wx + b, Predict w = 3 and b = 2.

The experimental requirements ：

1 Generate fitted data sets

The dataset contains only one eigenvector , Note that the error term must satisfy the Gaussian distribution . Used numpy and matplotlib library .numpy yes Python An open source numerical Scientific Computing Library , Can be used to store and process large matrices .matplotlib yes Python Drawing library of , It can be connected with numpy Use it together , Provides an effective MatLab Open source alternatives
The code is as follows ：

# First, import. 3 Databases :
import numpy as np
import tensorflow as tf
import matplotlib.pyplot as plt
# Randomly generate data points 100 individual , The random probability conforms to Gaussian distribution （ Normal distribution ）
num_points = 100
vectors_set = []
for i in range(num_points):
 x1 = np.random.normal(0.,0.55)
 y1 = x1 * 0.1 + 0.3 + np.random.normal(0.0,0.03)
 vectors_set.append([x1,y1])
# Define eigenvectors x
x_data = [v[0] for v in vectors_set]
# Define the label vector y
y_data = [v[1] for v in vectors_set]
# Press [x_data,y_data] stay X-Y In the coordinate system, it is displayed in dot mode , call plt Establish the coordinate system and print out the values 
plt.scatter(x_data,y_data,c='b')
plt.show()

The resulting data distribution is as follows ：

2 Construct a linear regression model Graph

# utilize TensorFlow Randomly generated w and b, For graphic display, you need , Define the names of myw and myb
w = tf.Variable(tf.random_uniform([1],-1.,1.),name='myw')
b = tf.Variable(tf.zeros([1]),name='myb')
# Based on randomly generated w and b, Combined with the above randomly generated feature vector x_data, Estimated value after calculation 
y = w * x_data + b
# Estimated value y And actual value y_data The mean square deviation between is taken as the loss 
loss = tf.reduce_mean(tf.square(y-y_data,name='mysquare'), name='myloss')
# The gradient descent method is used to optimize the parameters 
optimizer = tf.train.GradientDescentOptimizer(0.5)
train = optimizer.minimize(loss,name='mytrain')

3 stay Session Run the built Graph

#global_variables_initializer initialization Variable Equivariant 
sess = tf.Session()
init = tf.global_variables_initializer()
sess.run(init)
print("w=", sess.run(w),"b=",sess.run(b),sess.run(loss))
# iteration 20 Time train
for step in range(20):
 sess.run(train)
 print("w=", sess.run(w),"b=",sess.run(b),sess.run(loss))
# Write to disk , Provide tensorboard Show in browser with 
writer = tf.summary.FileWriter("./mytmp",sess.graph)

Print w and b, Change of loss value , You can see the loss value from 0.24 drop to 0.0008.

4 Draw the fitting curve

plt.scatter(x_data,y_data,c='b')
plt.plot(x_data,sess.run(w)*x_data+sess.run(b))
plt.show()

Pictured ：

5 tensorboard Show the application of neural network graph

In the above program design, there is code ：writer = tf.summary.FileWriter("./mytmp",sess.graph)
After running the modified code, the node information of the whole neural network can be written to ./mytmp Under the table of contents （ This directory is in the same directory as the previously established program ）. stay cmd Pass through “cd Catalog ” Switch to the directory , Input “dir” The command displays the log files just run in this directory , Last input tensorboard --logdir=D:\PyCharm2018.3.7\workpase\eg001\mytmp, The display message appears after the enter operation , In the last line of the message “TensorBoard 2.0.2 at http://localhost:6006/ (Press CTRL+C to quit)”, open chrome browser , Type in the browser address bar http://localhost:6006, It will show the graphic display of the neural network just programmed . As shown in the figure below ：

Experiment records, problems encountered in the experiment and solutions ：

1. Program running error

tensorflow Show no random_uniform modular
terms of settlement ：tf2.0 I changed my name in , use tf.random.uniform Instead of

2. Program running error

TensorFlow2.0 The version running program reports an error
terms of settlement ：

import tensorflow as tf

Change it to

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

tf.disable_v2_behavior()： This function can be called at the beginning of the program （ In creating a tensor 、 Before graphics or other structures , And before initializing the device ）. It will TensorFlow 1.x and 2.x Switch between all different global behaviors to predetermined 1.x Behavior , It's shielding 2.x edition .

3. Format error

To look at first tensorboard Installation position , I used Anaconda3, Can be in Anaconda3 Search under tensorflow, Find the bottom Scripts Look under the file for tensorboard.exe file , If it exists, there is no configuration tensorboard environment variable .

After this is done, there is a problem ,

There can be no spaces in the path , Just delete it

4. Page error reporting

Get into http://localhost:6006 Will report a mistake

Change the user name of the host to localhost （Win10 System host user name modification （ After modification, you need to restart the computer ）, Or an error ,（ This method is not only useless , It will also lead to the subsequent lack of access to the Internet , Readers are not advised to use , Because the author has used , Here are only records for reference ）

stay pycharm Try again in ：（ Turn off the cmd Command line ！！！ According to observation , If you want to open another new calculation chart , Be sure to put the original cmd The command line window closes , Repeat the above steps , Otherwise, the previous calculation diagram will still be opened .）
stay pycharm The menu bar of , choice View–Tool Windows–Terminal
And then execute ：

 tensorboard --logdir=mytmp

After entering the page , Successful implementation ！！！