逻辑斯蒂回归
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
| import torch
import torch.nn.functional as F
x_data= torch.Tensor([[1.0],[2.0],[3.0]])
y_data= torch.Tensor([[0],[0],[1]])
class Logistic_model(torch.nn.Module):
def __init__(self):
super(Logistic_model,self).__init__()
self.linear = torch.nn.Linear(1,1)
def forward(self,x):
y_prev = F.sigmoid(self.linear(x))
return y_prev
model = Logistic_model()
criterion = torch.nn.BCELoss(reduction= 'sum')
optimizer = torch.optim.SGD(model.parameters(),lr=0.01)
for epoch in range (1,1000):
y_prev = model(x_data)
loss = criterion(y_prev,y_data)
print(epoch,loss.item())
optimizer.zero_grad()
loss.backward()
optimizer.step()
print('w = ', model.linear.weight.item())
print('b = ', model.linear.bias.item())
x_test = torch.Tensor([[4.0]])
y_test = model(x_test)
print('y_pred = ', y_test.data)
|
与上节课实现思路基本一致，区别点如下
- 使用了Logistic Regression将具体的数值映射到0，1表示成功与失败
- 换用BCELoss来表示损失函数，通过比较交叉熵的数值来判断两个分布之间的接近程度，而不是判断具体数值之间的接近程度
本例展示了这套PyTorch编程架构的弹性，可以适用于较多的网状神经网络
