tensorflow实现线性回归总结

1、知识点

"""
模拟一个y = 0.7x+0.8的案例

报警：
    1、initialize_all_variables (from tensorflow.python.ops.variables) is deprecated and will be removed after 2017-03-02
        解决方法：由于使用了tf.initialize_all_variables() 初始化变量，该方法已过时，使用tf.global_variables_initializer()就不会了
        
tensorboard查看数据：
    1、收集变量信息
         tf.summary.scalar()
         tf.summary.histogram()
         merge = tf.summary.merge_all()
    2、创建事件机制
         fileWriter = tf.summary.FileWriter(logdir='',graph=sess.graph)
    3、在sess中运行并合并merge
        summary = sess.run(merge)
    4、在循环训练中将变量添加到事件中
        fileWriter.add_summary(summary,i) #i为训练次数
        
保存并加载训练模型：
    1、创建保存模型saver对象
       saver = tf.train.Saver()
    2、保存模型
        saver.save(sess,'./ckpt/model')
    3、利用保存的模型加载模型，变量初始值从保存模型读取
        if os.path.exists('./ckpt/checkpoint'):
            saver.restore(sess,'./ckpt/model')
            
创建变量域：
    with tf.variable_scope("data"):
"""

2、代码

 

# coding = utf-8

import tensorflow as tf
import  os
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'

def myLinear():
    """
    自实现线性回归
    :return:
    """
    with tf.variable_scope("data"):
        #1、准备数据
        x = tf.random_normal((100,1),mean=0.5,stddev=1,name='x')
        y_true = tf.matmul(x,[[0.7]])+0.8 #矩阵相乘至少为2维

    with tf.variable_scope("model"):
        #2、初始化权重和偏置
        weight = tf.Variable(tf.random_normal((1,1)),name='w')
        bias = tf.Variable(0.0,name='b')
        y_predict = tf.matmul(x,weight)+bias

    with tf.variable_scope("loss"):
        #3、计算损失值
        loss = tf.reduce_mean(tf.square(y_true-y_predict))

    with tf.variable_scope("train"):
        #4、梯度下降优化loss
        train_op = tf.train.GradientDescentOptimizer(learning_rate=0.01).minimize(loss)

    #初始化变量
    init_op = tf.global_variables_initializer()

    ############收集变量信息存到tensorboard查看###############
    #收集变量
    tf.summary.scalar('losses',loss)#1维
    tf.summary.histogram('weight',weight) #高维
    tf.summary.histogram('bias', bias)  # 高维
    merged = tf.summary.merge_all() #将变量合并
    #########################################################

    #####################保存并加载模型###############
    saver = tf.train.Saver()
    #################################################
    #5、循环训练
    with tf.Session() as sess:
        sess.run(init_op) #运行是初始化变量
        if os.path.exists('./ckpt/checkpoint'):
            saver.restore(sess,'./ckpt/model')

        #建立事件机制
        fileWriter = tf.summary.FileWriter(logdir='./tmp',graph=sess.graph)
        print("初始化权重为：%f,偏置为：%f" %(weight.eval(),bias.eval()))
        for i in range(501):
            summary = sess.run(merged)  # 运行并合并
            fileWriter.add_summary(summary,i)
            sess.run(train_op)
            if i%10==0 :
                print("第%d次训练权重为：%f,偏置为：%f" % (i,weight.eval(), bias.eval()))
        saver.save(sess,'./ckpt/model')
    return None


if __name__ == '__main__':
    myLinear()

 3、代码

import tensorflow as tf
import csv
import numpy as np
import matplotlib.pyplot as plt
# 设置学习率
learning_rate = 0.01
# 设置训练次数
train_steps = 1000
with open('D:/Machine Learning/Data_wrangling/鲍鱼数据集.csv') as file:
    reader = csv.reader(file)
    a, b = [], []
    for item in reader:
        b.append(item[8])
        del(item[8])
        a.append(item)
    file.close()
x_data = np.array(a)
y_data = np.array(b)
for i in range(len(x_data)):
    y_data[i] = float(y_data[i])
    for j in range(len(x_data[i])):
        x_data[i][j] = float(x_data[i][j])
# 定义各影响因子的权重
weights = tf.Variable(np.ones([8,1]),dtype = tf.float32)
x_data_ = tf.placeholder(tf.float32, [None, 8])
y_data_ = tf.placeholder(tf.float32, [None, 1])
bias = tf.Variable(1.0, dtype = tf.float32)#定义偏差值
# 构建模型为：y_model = w1X1 + w2X2 + w3X3 + w4X4 + w5X5 + w6X6 + w7X7 + w8X8 + bias
y_model = tf.add(tf.matmul(x_data_ , weights), bias)
# 定义损失函数
loss = tf.reduce_mean(tf.pow((y_model - y_data_), 2))
#训练目标为损失值最小，学习率为0.01
train_op = tf.train.GradientDescentOptimizer(0.01).minimize(loss)
with tf.Session() as sess:
    sess.run(tf.global_variables_initializer())
    print("Start training!")
    lo = []
    sample = np.arange(train_steps)
    for i in range(train_steps):
        for (x,y) in zip(x_data, y_data):
            z1 = x.reshape(1,8)
            z2 = y.reshape(1,1)
            sess.run(train_op, feed_dict = {x_data_ : z1, y_data_ : z2})
        l = sess.run(loss, feed_dict = {x_data_ : z1, y_data_ : z2}) 
        lo.append(l)
    print(weights.eval(sess))
    print(bias.eval(sess))
    # 绘制训练损失变化图
    plt.plot(sample, lo, marker="*", linewidth=1, linestyle="--", color="red")
    plt.title("The variation of the loss")
    plt.xlabel("Sampling Point")
    plt.ylabel("Loss")
    plt.grid(True)
    plt.show()