TF2实战-回归问题

TF2实战-回归问题

一、回归算法模型

找到参数w’与b’

//linear regression 模型
(wX+b-Y)^2
loss=sum(w*xi+b-yi)^2

minimize loss：
w’*x+b’-> y

//梯度下降
x’ = x-0.005*(dy/dx) //x沿着函数值下降的方向移动
w’ = w-lr*(dy/dw)

//loss下降的方向就是y’->y的方向
loss=sum(wxi+b-yi)^2
w’ = w-lr(dloss/dw) // dL/dw = 2sum(wxi+b-yi)xi
b’ = b-lr*(dloss/db) // dL/db = 2sum(wxi+b-yi)
w’*x+b’->y

二、利用numpy实战

import numpy as np
import csv

def calculate_loss(points,w0,b0):
    loss = 0
    for i in range(len(points)):
        xi = points[i,0]
        yi = points[i,1]
        loss += (w0*xi+b0-yi)**2
    return loss/float(len(points))

def calculate_tidu(points,w1,b1,lr):
    dw = 0
    db = 0
    N = len(points)
    for i in range(len(points)):
        xi = points[i, 0]
        yi = points[i, 1]
        dw += 2/N*(w1*xi+b1-yi)*xi
        db += 2/N*(w1*xi+b1-yi)
    w_new = w1-lr*dw
    b_new = b1-lr*db
    return w_new,b_new

def run():
    points = np.genfromtxt('data.csv',delimiter=',')
    lr = 0.0001
    w_orignal = 0
    b_orignal = 0
    loss0 = calculate_loss(points, w_orignal,b_orignal)
    print('the orignal is:w = {0},b = {1}, loss = {2}'.format(w_orignal,b_orignal,loss0))
    w_new = w_orignal
    b_new = b_orignal
    for i in range(100000):
        w_new,b_new = calculate_tidu(points, w_new, b_new, lr)
        loss_new = calculate_loss(points, w_new, b_new)
        print('the running is:w = {0},b = {1}, loss = {2}'.format(w_new, b_new, loss_new))

if __name__ =='__main__':
    run()

三、手写数字体的识别实验

基础理论：