线性回归 更简洁的实现方式

Cleaner Code of Linear Regression

更简洁的线性回归实现

import numpy as np
import torch
from torch.utils import data
from d2l import torch as d2l
from torch import nn

# 生成数据集
true_w = torch.tensor([2,-3.4])
true_b = 4.2
featrues , labels = d2l.synthetic_data(true_w , true_b , 1000) # 这里生成的是真实的y我们后面会预测他和期待值之间的差距

# 读取数据集
def load_data(data_arrays , batch_size , is_train=True): # @save
    # is_train is for torch.shuffle()
    '''构造一个pytorch数据迭代器'''
    dataset = data.TensorDataset(*data_arrays)
    return data.DataLoader(dataset , batch_size , shuffle = is_train)

# test
batch_size = 10
data_iter = load_data((featrues , labels) , batch_size)
next(iter(data_iter))

[tensor([[-0.8651, -0.6773],
         [-0.5669, -0.3115],
         [-0.4197, -0.8198],
         [-0.1276, -0.3684],
         [-0.8580, -0.8185],
         [-0.4955,  0.1050],
         [ 0.2148,  0.1534],
         [-2.0009,  0.6106],
         [-1.3546,  1.1515],
         [-0.2777, -0.7998]]),
 tensor([[ 4.7699],
         [ 4.1208],
         [ 6.1584],
         [ 5.1972],
         [ 5.2708],
         [ 2.8533],
         [ 4.1034],
         [-1.8997],
         [-2.4302],
         [ 6.3858]])]

# 定义模型
net = nn.Sequential(nn.Linear(2 , 1))
print(net[0].weight.data)
print(net[0].bias.data)# 访问隐藏层和偏置层

tensor([[-0.0941,  0.1806]])
tensor([0.2378])

net[0].weight.data.normal_(0 , 0.01)
net[0].bias.data.fill_(0) #通过data.normal_和fill_重写参数值
print(net[0].weight.data)
print(net[0].bias.data)# 访问隐藏层和偏置层

tensor([[ 0.0048, -0.0003]])
tensor([0.])

loss = nn.MSELoss()
trainer = torch.optim.SGD(net.parameters() , lr = 0.03)# 定义优化算法

num_eopch = 3
for epoch in range(num_eopch) :
    for X , y in data_iter :
        l = loss(net(X) , y)
        trainer.zero_grad()
        l.backward()
        trainer.step()
    l = loss(net(featrues) , labels)
    print(f'epoch{epoch + 1} , loss {l :f}')

epoch1 , loss 0.000185
epoch2 , loss 0.000097
epoch3 , loss 0.000096

w = net[0].weight.data
print(f'w 的估计误差 {true_w - w.reshape(true_w.shape)}')

w 的估计误差 tensor([ 0.0002, -0.0007])

b = net[0].bias.data
print(f'b 的估计误差 {true_b - b}')

b 的估计误差 tensor([0.0002])