샘플 데이터셋 생성

import torch
import numpy as np
import matplotlib.pyplot as plt

x = torch.tensor(np.arange(10))
y = 2*x + 1
y

tensor([ 1,  3,  5,  7,  9, 11, 13, 15, 17, 19])

plt.title('y = 2x + b', fontsize=15)
plt.scatter(x, y)
plt.show()

# random w, b 생성
w = torch.randn(1, requires_grad=True)
b = torch.randn(1, requires_grad=True)
w, b

(tensor([0.8925], requires_grad=True), tensor([1.2460], requires_grad=True))

# Hypothesis Function 정의
y_hat = w*x + b

# Mean Squared Error(MSE) 오차 정의
loss = ((y_hat - y)**2).mean()

# BackPropagation (Gradient 계산)
loss.backward()

# 결과 출력
print(f'w gradient: {w.grad.item():.2f}, b gradient: {b.grad.item():.2f}')

w gradient: -60.91, b gradient: -9.48

w, b 의 직접 계산한 Gradient와 비교

w_grad = (2*(y_hat - y)*x).mean().item()
b_grad = (2*(y_hat - y)).mean().item()

print(f'w gradient: {w_grad:.2f}, b gradient: {b_grad:.2f}')

w gradient: -60.91, b gradient: -9.48

y_hat = w*x + b
print(y_hat.requires_grad)

with torch.no_grad():
    y_hat = w*x + b
    
print(y_hat.requires_grad)

True
False