import torch.nn
from torch.optim import Adam
import torch
x = torch.tensor([[2., 2.], [1., -2.], [1., 4.]])
y = torch.tensor([3., 1., 3.])
w = torch.ones(2, requires_grad=True)
criterion = torch.nn.MSELoss()
optimizer = Adam([w])
opt_step = 0
predict = torch.mv(x, w)
loss = criterion(predict, y)
while loss > 1/9 + 0.000001:
opt_step = opt_step + 1
optimizer.zero_grad()
loss.backward()
optimizer.step()
predict = torch.mv(x, w)
loss = criterion(predict, y)
print(f'The optimal step is {opt_step}, the final loss is {loss}.')
print(w)
第一段代码,使用优化器进行优化。未涉及“特征值”,“标签”等概念。
from torch.optim import Adam
import torch
x = torch.tensor([[2., 2.], [1., -2.], [1., 4.]])
y = torch.tensor([3., 1., 3.]).reshape(-1, 1)
fc = torch.nn.Linear(2, 1, bias=False)
criterion = torch.nn.MSELoss()
optimizer = Adam(fc.parameters())
predict = fc(x)
loss = criterion(predict, y)
opt_step = 0
while loss > 1/9 + 0.00000001:
opt_step = opt_step + 1
optimizer.zero_grad()
loss.backward()
optimizer.step()
predict = fc(x)
loss = criterion(predict, y)
weights = next(fc.parameters())
print(weights)
这是第二段代码,采用了一个简单的神经元,通过优化神经元的权重得到拟合值。
二者的y有区别,一个加了reshape(-1, 1),另一个没有。
原因:在神经网络的语义里,一组特征值对应一个标签。所以要加上reshape(-1, 1),让特征值和标签一一对应。
reshape(-1, 1)的含义参考这篇文章