1.逻辑回归常常被当作二分类问题。
2.线性神经元输出xw,经过sigmoid函数映射为(0,1)中,sigmoid = 1/1+e^-x
3.伯努利原理:抛硬币正:p = θ ,y=1 抛硬币反:p=1-θ,y=0
p(y|θ)=θ^y(1-θ)^1-y
4.极大似然:抛硬币每次独立,若硬币不均匀,则每次抛出的概率会偏向高概率的结果,那么将每次抛出的结果概率累起来,则会趋向最大。
p(y) = py1py2py3...pyn = θ^y1(1-θ)^1-y1....θ^yn(1-θ)^1-yn
我们有,p(yi) = sigm(x,w)= 1/1+e^-xiwi = θ
最大似然可以写作:p(y|X,w) = sigm(xiw)^yi *(1-sigm(xiw))^1-yi
我们两边取对数,取负数,将最大化变最小化:
-logp(y|X,w)=- (yi logsigm(xiw)* 1-yi log(1-sigm(xiw)) )
令L(w)= -logp(y|X,w) ,L(w)就是我们需要的损失函数,只需求出最小的损失函数,梯度下降后就可以得到 =w ,pytorch中 有这样的交叉熵损失函数(用极大似然得到的函数),
nn.CrossEntropyLoss()
import matplotlib.pyplot as plt
import torch
from torch import nn,optim
# 500*2的大小
cluster = torch.ones(500, 2)
# 4,-4为期望值,2为标准差生成一堆数据
data0 = torch.normal(4 * cluster, 2)
data1 = torch.normal(-4 * cluster, 2)
label0 = torch.zeros(500)
label1 = torch.ones(500)
x = torch.cat((data0, data1), ).type(torch.FloatTensor)
y = torch.cat((label0, label1), ).type(torch.LongTensor)
plt.scatter(x.numpy()[:, 0], x.numpy()[:, 1], c=y.numpy(), s=10, lw=0, cmap='RdYlGn')
plt.show()
class Net(nn.Module):
def __init__(self):
super(Net, self).__init__()
# 输入包括两个特征,x1,x2横轴纵轴
self.linear = nn.Linear(2, 2)
def forward(self, x):
x = self.linear(x)
x = torch.sigmoid(x)
return x
CUDA = torch.cuda.is_available()
if CUDA:
net = Net().cuda()
inputs = x.cuda()
target = y.cuda()
else:
net = Net()
inputs = x
target = y
criterion = torch.nn.CrossEntropyLoss()
optimizer = optim.SGD(net.parameters(), lr=2e-2)
def draw(output):
if CUDA:
output = output.cpu()
# 清空画布
plt.cla()
output = torch.max((output), 1)[1]
pred_y = output.data.numpy().squeeze()
target_y = y.numpy()
plt.scatter(x.numpy()[:, 0], x.numpy()[:, 1], c=pred_y, s=10, lw=0, cmap='RdYlGn')
accuracy = sum(pred_y == target_y) / 1000.0
plt.text(1.5, -4, 'Accuracy=%s' % (accuracy), fontdict={'size': 20, 'color': 'red'})
plt.pause(0.1)
def Train(model,criterion,optimizer,epochs):
for epochs in range(epochs):
output = model(inputs)
loss = criterion(output,target)
optimizer.zero_grad()
loss.backward()
optimizer.step() #权重更新
if epochs % 40 == 0:
draw(output)
return model,loss
Train(net,criterion,optimizer,1000)