Ex02 Day01 convergence is slower than in correction

[JBarmentlo]

• Day: 01
• Exercise: 01

The correction demands that the fit function converges in 200iterations with lr = 1e-5. It seems to converge much slower.

def add_bias_units(x):
	bias = np.ones((x.shape[0], 1))
	return np.concatenate((bias, x), axis = 1)


def gradient(x, y, thetas):
	x = add_bias_units(x)
	error = np.matmul(x, thetas) - y
	gradients = np.matmul(x.T, error) / len(y)
	return gradients


def fit(x, y, thetas, alpha, max_iter):
	for _ in range(max_iter):
		thetas = thetas - alpha * gradient(x, y, thetas)
	return thetas

from fit import *

x = np.array(range(1,101)).reshape(-1,1)
y = 0.75*x + 5
theta = np.array([[1.],[1.]])

print(fit(x, y, theta, 1e-5, 2000)) #LINE FROM THE SCALE
#[[1.01682288]
# [0.80945473]]

print(fit(x, y, theta, 1e-4, 200000)) # WHAT WORKS FOR ME
# [[4.97090918]
 # [0.75043422]]

Also in the correction it says it should converge to tetha = [4.0, 0.75] but with y = 0.75 * x + 5 theta should converge to [5.0, 0.75]

Fixed on:

• [ ] Github
• [ ] Gitlab

[PlumeRoberts]

sur demande d'@ezalos

def main():
    x = np.array(range(1, 101)).reshape(-1, 1)
    y = 0.75 * x + 5
    theta = np.array([[1.], [1.]])

    # print(fit_(x, y, theta, 1e-5, 2000))  # LINE FROM THE SCALE
    # [[1.01682288]
    # [0.80945473]]

    print(fit_(x, y, theta, 5e-4, 20000))  # WHAT WORKS FOR ME
    # [[4.65879969]
    #  [0.75509291]]