08 线性回归从零开始实现

temp0723

 

 

 

 

In [64]:

import torch
from torch import Tensor
import random as r
import matplotlib.pyplot as plt
from tianfy.pytools import batch_data
import os


os.environ['KMP_DUPLICATE_LIB_OK'] = 'True'

In [70]:

def get_data(w, b, num_samples):
    num_features = w.shape[0]
    X = torch.normal(0, 1, (num_samples, num_features))
    y = X @ w + b
    y += torch.normal(0, 0.01, y.shape)
    return X, y.reshape(-1, 1)

In [81]:

w = Tensor([[3.5, 4]]).T
b = 2.5
X, y = get_data(w, b, 1000)
n_samples, n_features = X.shape
w = torch.zeros((n_features + 1, 1), requires_grad=True)
X = torch.cat((torch.ones((n_samples, 1)), X), dim=1)

def model(X, w):
    return X @ w


def loss(y_hat, y):
    return (y_hat - y) ** 2 / 2


def sgd(w, lr, batch_size):
    with torch.no_grad():
        w -= lr * w.grad / batch_size
        w.grad.zero_()


lr = 0.03
epochs = 5
batch_size = 10
for epoch in range(epochs):
    for X_, y_ in batch_data(X, y, batch_size):
        l = loss(model(X_, w), y_)
        l.sum().backward()
        sgd(w, lr, batch_size)
    with torch.no_grad():
        train_l = loss(model(X, w), y)
        print(f"epoch{epoch+1}, loss {train_l.mean():2f}, params {(w).view(-1).tolist()}")

 

epoch1, loss 0.040367, params [2.391767740249634, 3.330096483230591, 3.79638671875]
epoch2, loss 0.000144, params [2.4951794147491455, 3.491722822189331, 3.9897348880767822]
epoch3, loss 0.000050, params [2.4995713233947754, 3.499561071395874, 3.999636650085449]
epoch4, loss 0.000050, params [2.499753475189209, 3.4999406337738037, 4.000146389007568]
epoch5, loss 0.000050, params [2.499760627746582, 3.4999589920043945, 4.0001726150512695]