pytorch-lin-reg.py

import torch
from torch.autograd import Variable

dtype = torch.FloatTensor  # Datatype of our Tensors

# Equation: y = wX + b
y = Variable(torch.rand(10).type(dtype), requires_grad=False)
X = Variable(torch.rand(10).type(dtype), requires_grad=False)
w = Variable(torch.randn(10).type(dtype), requires_grad=True)
b = Variable(torch.randn(10).type(dtype), requires_grad=True)

learning_rate = 1e-2  # Learning rate

for iter in range(500):
    pred_y = X.mul(w).add(b)  # y = wX + b
    loss = (pred_y - y).pow(2).sum()  # Sum of squared errors
    print(iter, loss.data[0])  # Iteration and loss
    loss.backward()  # Compute gradients
    # Gradient descent
    w.data -= learning_rate * w.grad.data
    b.data -= learning_rate * b.grad.data
    # Manually the zero the gradient buffers
    w.grad.data.zero_()
    b.grad.data.zero_()