python-tensorflow实现非线性回归（两个案例）

'''
tensorflow 实现非线性回归
案例一：y = ax^2+bx+c
案例二：y = a*exp(-(x-u)**2/(x*sigma**2))+b

这个和神经网络目前来看还没有关系，好像完全就是利用模型，然后去进行反向传播，最后去计算的最优参数
'''

import tensorflow as tf
import numpy as np
import matplotlib.pyplot as plt

'''案例一：y = ax^2+bx+c'''
def demo1():
    #构造数据
    x = np.arange(-10,10,0.1)
    y = 3*x**2+2*x+1

    plt.plot(x,y,'*')
    # plt.show()

    #构造模型
    xx = tf.placeholder(dtype=tf.float32,shape=[x.shape[0]],name='xx')
    a = tf.Variable(initial_value=0.0,name='a')
    b = tf.Variable(initial_value=0.0,name='b')
    c = tf.Variable(initial_value=0.0,name='c')
    yy = tf.placeholder(dtype=tf.float32,shape=[y.shape[0]],name='yy')
    predict = a*xx**2+b*xx+c

    loss = tf.sqrt(tf.reduce_mean((yy-predict)**2))

    Optimizer = tf.train.GradientDescentOptimizer(learning_rate=0.0005).minimize(loss)

    with tf.Session() as sess:
        tf.global_variables_initializer().run()

        for i in range(7000):
            Optimizer.run(feed_dict={xx:x,yy:y})
            print("loss:%f"%(loss.eval(feed_dict={xx:x,yy:y})))
        print('a:%f,b:%f,c:%f'%(a.eval(),b.eval(),c.eval()))
        plt.plot(x,a.eval()*x**2+b.eval()*x+c.eval())
        plt.show()

迭代了7000次后，达到了较小的损失

loss:0.500241
loss:0.500210
loss:0.500238
a:3.012125,b:2.000031,c:0.941524

输出结果

第二个案例

'''案例二：y = a*exp(-(x-u)**2/(x*sigma**2))+b'''
def demo2():
    #构造数据
    x = np.arange(-10,10,0.1)
    a = 2
    u = 1.0
    s = 1.0
    b = 2
    y = a*np.exp(-(x-u)**2*s)+b

    plt.plot(x,y,'*')
    # plt.show()

    #构造模型
    xx = tf.placeholder(dtype=tf.float32,shape=[x.shape[0]],name='xx')
    aa = tf.Variable(initial_value=3.5,name='aa')
    bb = tf.Variable(initial_value=2.2,name='bb')
    uu = tf.Variable(initial_value=1.5,name='uu')
    ss = tf.Variable(initial_value=2.0,name='sigma')
    yy = tf.placeholder(dtype=tf.float32,shape=[y.shape[0]],name='xx')
    predict = aa*tf.exp(-(xx-uu)**2*ss)+bb

    loss = tf.sqrt(tf.reduce_mean((yy-predict)**2))

    Optimizer = tf.train.AdamOptimizer(learning_rate=0.0005).minimize(loss)

    with tf.Session() as sess:
        tf.global_variables_initializer().run()

        for i in range(5000):
            Optimizer.run(feed_dict={xx:x,yy:y})
            print("loss:%f"%(loss.eval(feed_dict={xx:x,yy:y})))
        print('aa:%f,uu:%f,ss:%f,bb:%f'%(aa.eval(),uu.eval(),ss.eval(),bb.eval()))
        plt.plot(x,aa.eval()*np.exp(-(x-uu.eval())**2*ss.eval())+bb.eval())
        plt.show()
# demo2()

loss:0.000044
loss:0.000041
loss:0.000031
aa:2.000023,uu:1.000000,ss:0.999975,bb:2.000025

结果图