Easiest understanding about Gradient descent

gradient_descent.py

#%% making dummy data ----------------------------------------
from random import *
import math
import matplotlib.pyplot as plt
import numpy as np

def get_biased_point_randomly():
    x = randint(1, 100) 
    y = x + randint(1,20) * pow(-1, x)
    y = math.floor(y)
    return [x,y]

points = []
for i in range(100):
    points.append(get_biased_point_randomly())

points = np.asarray(points)
points



#%% plotting dummy data ----------------------------------------
def draw_points():
    plt.scatter(points.T[0], points.T[1])
draw_points()

#%%
b = randint(-100,100)
b = 0 # If you want learn for 'b', escape this line.
w = randint(-2,2)

def draw_regLine():
    linear_line = [
        [0,b],
        [100,100*w]
    ]
    linear_line = np.asarray(linear_line)
    plt.plot(linear_line.T[0], linear_line.T[1])

draw_points(); draw_regLine()

#%% calculate loss ----------------------------------------
    # ex : estimated x
    # rx : real x
    
def calLoss(w):
    loss_arr = []
    for point in points:
        rx = point[1]
        ex = w * point[0]
        loss_arr.append(ex-rx)
    return math.floor(
            pow(sum(loss_arr)/ len(points), 2) 
        )

print('loss is : ', calLoss(w))


#%% manualy w updating ----------------------------------------
w = w+.1
draw_points(); draw_regLine()
plt.title(calLoss(w))



#%% auto gradient descent ----------------------------------------
    # gradient is (delta loss) over (delta w)

delta = .1; learning_rate = .0001
def get_gradient():
    return ( (calLoss(w+delta) - calLoss(w-delta)) / (2*delta) )
print(get_gradient())


#%% auto gradient descent ----------------------------------------
    # if you repeat this Block Cell, w will be adjust automatically
w = w - (learning_rate*get_gradient())
draw_points(); draw_regLine()
plt.title("Loss:"+str(calLoss(w)) + " " + "gradient:"+str(get_gradient()))

#%% draw loss=f(w) graph ----------------------------------------
    # this is not countinuously
    # this is computational function

loss_coordinates = []
i = -10
while i < 10:
    loss_coordinates.append([i, calLoss(i)])
    i = i + .1
loss_coordinates

plt.plot(loss_coordinates)

[오류 수정]
print(gradient) 대신 print(get_gradient())로 변경

[섹션 추가]
#%% draw loss=f(w) graph 섹션 추가됨

더이상 수정하지 않습니다.
ipynb 로 만든 다른 파일 참조 : https://gist.github.com/fredriccliver/0e5c79a2c0c277de332aa85bf686312a