pytorch线性回归

import torch
import numpy as np

函数关系是 y = -5*x +3

x_train = torch.unsqueeze(torch.linspace(-10,10,100),dim=1)
y_train = -5*x_train + 3 + torch.randn(x_train.size())
print("x_train[:10]:{}".format(x_train[:10]))
print("y_train[:10]:{}".format(y_train[:10]))

x_train[:10]:tensor([[-10.0000],
        [ -9.7980],
        [ -9.5960],
        [ -9.3939],
        [ -9.1919],
        [ -8.9899],
        [ -8.7879],
        [ -8.5859],
        [ -8.3838],
        [ -8.1818]])
y_train[:10]:tensor([[51.5397],
        [52.7127],
        [52.1205],
        [50.8858],
        [47.1663],
        [47.2642],
        [47.2564],
        [45.4092],
        [46.2278],
        [43.5305]])

创建模型

class myLinearReg(torch.nn.Module):
    def __init__(self):
        super(myLinearReg,self).__init__()
        self.fc = torch.nn.Linear(1,1)#(1,1)表示输入参数个数是1，输出参数个数是1
        
    def forward(self,x):
        return self.fc(x)
model = myLinearReg()

定义损失函数优化器

loss = torch.nn.MSELoss()
#定义优化器
optimizer = torch.optim.SGD(model.parameters(),lr = 0.0001)

训练

epochs = 1000
loss_list = []
epoch_list = []
for epoch in range(epochs):
    #前向传播过程
    y_hat = model(x_train)
    loss_ = loss(y_hat,y_train)
    #反向传播
    optimizer.zero_grad()
    loss_.backward()
    optimizer.step()
    
    if (epoch+1)%20 == 0:
        print("Epoch:{}/{},loss:{:.5f}".format(epoch+1,epochs,loss_.detach().numpy()))
        loss_list.append(loss_.detach().numpy())
        epoch_list.append(epoch+1)
    

    Epoch:20/1000,loss:910.51935
    Epoch:40/1000,loss:695.19537
    ............................
    Epoch:1000/1000,loss:6.75332

可视化

import matplotlib.pyplot as plt
%matplotlib
plt.scatter(x_train.numpy(),y_train.numpy(),label="original data",c = 'g',linewidth = 1)

#把模型转为预测模式
model.eval()
y_pred = model(x_train)
plt.plot(x_train.numpy(),y_pred.detach().numpy(),c='r',label='fit',linewidth=1)
plt.show()

plot.png

plt.scatter(epoch_list,loss_list,label="loss",linewidth=1)
plt.show()

loss.png

list(model.named_parameters())

[('fc.weight',
  Parameter containing:
  tensor([[-4.9977]], requires_grad=True)),
 ('fc.bias',
  Parameter containing:
  tensor([0.5304], requires_grad=True))]