Predecting age from weight and height, using neural networks

There are three parts to the question you asked.

1st part (what’s wrong with the code) Ans: There is nothing wrong as such. Just proper lr and initilization

2nd part(where to use sigmoid). Ans: Nowhere in regression.

3rd part(correct alpha): Ans: start with 1 run the model and observe loss. Make alpha=1e-1 and so on. In the entire process observe loss. Whichever alpha yields small loss, choose that.

Code for 1st part: (Make following chnages)

b=np.zeros((1,))
alpha=1e-5

Yeah, that’s it 🙂 You may keep a list of loss for each epoch loss. I kept it for your code and got following graph. (Note I used only 50 epochs. you may keep higher values)

Full Code

import numpy as np
import matplotlib.pyplot as plt

# weight,height,age

data=[[50,160,15],
      [45,156,13],
      [55,162,17],
      [56,163,18],
      [58,163.4,20],
      [12,85,2],
      [14.2,95,3],
      [15.4,100,4],
      [32,138.4,10],
      [25.8,128.3,8]]

# missing age data
mys=[22.4,121.1]

w1=np.random.rand()
w2=np.random.rand()
b=0
alpha=0.00001

loss = []
for i in range(50):
  ri=np.random.randint(len(data))
  point=data[ri]

  pred=(point[0]*w1)+(point[1]*w2)+b
  target=point[2]
  cost=(pred-target)**2


  dcost_pred=2*(pred-target)
  dpred_w1=point[1]
  dpred_w2=point[0]
  dpred_db=1

  dcost_w1=dcost_pred*dpred_w1
  dcost_w2=dcost_pred*dpred_w2
  dcost_b=dcost_pred*dpred_db

  w1=(w1-alpha*dcost_w1)
  w2=w2-alpha*dcost_w2
  b=b-alpha*dcost_b
  loss.append(cost)

plt.plot(loss)