本算法算是本人在学习机器学习过程中的入门算法之一。
该算法实现了纯手工打造,每一行代码都包含原创作者的心血,是算法中的劳斯莱斯,向该算法的原创作者致敬!
参考地址:点击打开
计算较为繁琐,需要用到sigmoid函数和梯度下降算法,算法的步骤主要如下:
- 二项分布概率公式
- 最大似然估计法
- 求导数
- 梯度下降算法和优化
上图的主要过程是对交叉熵里的w求导,手写的拍照,可能有些杂乱,而且略有错误,下面再次手动推导一下。
交叉熵是概率统计里一种描述预测差异的方法,主要思路是单分类任务的结果服从二项分布,而且其出现一次的概率为:
P ( x = k ) = ( n k ) P k ( 1 − P ) n − k P(x = k) = { n \choose k} P_{}^{k} (1- P) _{}^{n-k} P(x=k)=(kn)Pk(1−P)n−k
其总的概率分布为:
∑
i
=
1
n
P
(
x
=
i
)
=
∑
i
=
i
n
(
n
i
)
P
i
(
1
−
P
)
n
−
i
=
1
\sum _{i = 1}^{n}P(x = i) = \sum_{i=i}^{n} { n \choose i} P_{}^{i} (1- P) _{}^{n-i} = 1
i=1∑nP(x=i)=i=i∑n(in)Pi(1−P)n−i=1
因为二分类任务的输出一般只有两种(0和1),用y 和1-y来表示,因此产生的交叉熵为:
−
l
o
g
p
i
y
(
1
−
p
i
)
1
−
y
-logp_{i}^{y}(1-p_i)^{1-y}
−logpiy(1−pi)1−y,对此式取对数并化简:
−
l
o
g
(
p
i
y
(
1
−
p
i
)
1
−
y
)
=
−
[
y
l
o
g
p
i
+
(
1
−
y
)
l
o
g
(
1
−
p
i
)
]
=
−
[
y
l
o
g
p
i
+
l
o
g
(
1
−
p
i
)
−
y
l
o
g
(
1
−
p
i
)
]
=
−
[
y
l
o
g
p
i
1
−
p
i
+
l
o
g
(
1
−
p
i
)
]
-log (p_{i}^{y}(1-p_i)^{1-y}) = -[ylogp_i + (1-y)log(1-p_i)]= \\ -[ylogp_i + log(1-p_i) - ylog(1-p_i) ]= \\ -[ylog\frac{p_i}{1-p_i} + log(1-p_i)]
−log(piy(1−pi)1−y)=−[ylogpi+(1−y)log(1−pi)]=−[ylogpi+log(1−pi)−ylog(1−pi)]=−[ylog1−pipi+log(1−pi)]
对w求偏导:
∂
f
∂
w
=
−
∂
l
o
g
(
p
i
y
(
1
−
p
i
)
1
−
y
)
∂
w
=
−
∂
[
y
l
o
g
p
i
1
−
p
i
+
l
o
g
(
1
−
p
i
)
]
∂
w
=
−
∂
[
y
l
o
g
1
1
+
e
−
w
x
1
−
1
1
+
e
−
w
x
+
l
o
g
(
1
−
1
1
+
e
−
w
x
)
]
∂
w
=
−
∂
[
y
l
o
g
1
1
+
e
−
w
x
e
−
w
x
1
+
e
−
w
x
+
l
o
g
(
e
−
w
x
1
+
e
−
w
x
)
]
∂
w
=
−
∂
[
y
l
o
g
1
e
−
w
x
+
l
o
g
(
e
−
w
x
1
+
e
−
w
x
)
]
∂
w
=
−
∂
[
y
w
x
+
l
o
g
(
e
−
w
x
1
+
e
−
w
x
)
]
∂
w
=
−
y
x
−
∂
[
l
o
g
(
e
−
w
x
1
+
e
−
w
x
)
]
∂
w
=
−
y
x
−
1
+
e
−
w
x
e
−
w
x
−
e
−
w
x
(
−
x
)
e
−
w
x
+
(
1
+
e
−
w
x
)
e
−
w
x
(
−
x
)
[
1
+
e
−
w
x
]
2
=
−
y
x
−
1
+
e
−
w
x
e
−
w
x
e
−
2
w
x
(
x
)
−
e
−
w
x
x
−
x
e
−
2
w
x
[
1
+
e
−
w
x
]
2
=
−
y
x
−
−
x
1
+
e
−
w
x
=
(
y
^
−
y
)
x
\frac {\partial f}{\partial w} = -\frac {\partial log (p_{i}^{y}(1-p_i)^{1-y})}{\partial w} =\\ -\frac {\partial \biggl [ylog\frac{p_i}{1-p_i} + log(1-p_i) \biggr]}{\partial w} =\\ -\frac {\partial \biggl [ylog\frac{\frac{1}{ 1+e^{-wx}}}{1 - \frac {1}{1 + e^{-wx}}}+ log(1-\frac {1}{1 + e^{-wx}}) \biggr]}{\partial w} = \\ -\frac {\partial \biggl [ylog\frac{\frac{1}{ 1+e^{-wx}}}{ \frac {e^{-wx}}{1 + e^{-wx}}}+ log(\frac {e^{-wx}}{1 + e^{-wx}}) \biggr]}{\partial w} =\\ -\frac {\partial \biggl [ylog\frac{1}{e^{-wx}}+ log(\frac {e^{-wx}}{1 + e^{-wx}}) \biggr]}{\partial w} = \\ -\frac {\partial \biggl [ywx+ log(\frac {e^{-wx}}{1 + e^{-wx}}) \biggr]}{\partial w} =\\ -yx - \frac {\partial \biggl [ log(\frac {e^{-wx}}{1 + e^{-wx}}) \biggr]}{\partial w} = \\ -yx- \frac{1+ e^{-wx}}{e^{-wx}} \frac {-e^{-wx} (-x) e^{-wx} + (1+e^{-wx} )e^{-wx}(-x)}{[1+e^{-wx}]^2}=\\ -yx - \frac{1+ e^{-wx}}{e^{-wx}} \frac {e^{-2wx} (x) -e^{-wx}x- x e^{-2wx} }{[1+e^{-wx}]^2}=\\ -yx - \frac{-x}{1+e^{-wx}} = (\hat y - y)x
∂w∂f=−∂w∂log(piy(1−pi)1−y)=−∂w∂[ylog1−pipi+log(1−pi)]=−∂w∂[ylog1−1+e−wx11+e−wx1+log(1−1+e−wx1)]=−∂w∂[ylog1+e−wxe−wx1+e−wx1+log(1+e−wxe−wx)]=−∂w∂[yloge−wx1+log(1+e−wxe−wx)]=−∂w∂[ywx+log(1+e−wxe−wx)]=−yx−∂w∂[log(1+e−wxe−wx)]=−yx−e−wx1+e−wx[1+e−wx]2−e−wx(−x)e−wx+(1+e−wx)e−wx(−x)=−yx−e−wx1+e−wx[1+e−wx]2e−2wx(x)−e−wxx−xe−2wx=−yx−1+e−wx−x=(y^−y)x
上图中,y和 y ^ \hat y y^的顺序能不能交换呢?测试是不可以的,为什么呢?此处还有存疑,有时间再好好想想。
进一步的阐述说明:
- sigmoid函数没有偏执bias参数。
- 上面导数推导图中的连乘号不准确,实际是表达了预测结果与实际结果两个矩阵向量的点积运算。
- 算法测度损失使用交叉熵。从熵的性质可以知道,交叉熵是熵的最大值,使用交叉熵的主要原因是便于计算吗?
- 二分类问题中,y 和 y ^ \hat y y^的概率和是1,次数之和是n,而不是y 的值为1, y ^ \hat y y^的值为0;如果是这样的话,就没法求导了。
- 划线模块draw_desion_line的原理是怎样的呢?在模块load_dataset中对x增加一个维度的原因是什么呢?此问题是二分类问题画图的核心问题之一 ,初学者一般都会提出这样一个问题,这个问题的实质请参考如下网址的解释说明:https://blog.csdn.net/xq151750111/article/details/121871627,以及:https://scipython.com/blog/plotting-the-decision-boundary-of-a-logistic-regression-model/;https://stats.stackexchange.com/questions/202668/calculation-of-decision-boundaries-with-perceptron。回到第二个问题,测试集也可以将第一列设置为0,不影响最后的画图。
代码:
import numpy as np
import matplotlib.pyplot as plt
def sigmoid(z):
return 1.0/(1+np.exp(-z))
# datas NxD
# labs Nx1
# w Dx1
def weight_update(datas,labs,w,alpha=0.01):
z = np.dot(datas,w) # Nx1
h = sigmoid(z) # Nx1
Error = -( labs-h ) # Nx1
w = w - alpha*np.dot(datas.T,Error)
return w
# 随机梯度下降
def train_LR_batch(datas,labs,batchsize=80,n_epoch=2,alpha=0.005):
print("epoch:%d,alpha:%.8f batchsize:%d"%(n_epoch,alpha,batchsize))
N,D = np.shape(datas)
print("N:%f D:%f"% (N,D))
# weight 初始化
w = np.ones([D,1]) # Dx1
N_batch = N//batchsize #取整数
#print("n:%d d:%d batchsize:%d"%(N,D,batchsize))
for i in range(n_epoch):
# 数据打乱
rand_index = np.random.permutation(N).tolist()
# 每个batch 更新一下weight
for j in range(N_batch):
#print("i:%d j:%d N_batch:%d\r\n"%(i,j,N_batch))
# alpha = 4.0/(i+j+1) +0.01
index = rand_index[j*batchsize:(j+1)*batchsize]
batch_datas = datas[index]
batch_labs = labs[index]
w=weight_update(batch_datas,batch_labs,w,alpha)
error = test_accuracy(datas,labs,w)
#print("epoch %d error %.2f%%"%(i,error*100))
return w
def train_LR(datas,labs,n_epoch=2,alpha=0.005):
N,D = np.shape(datas)
print("N:%f D:%f"% (N,D))
w = np.ones([D,1]) # Dx1
# 进行n_epoch轮迭代
for i in range(n_epoch):
w = weight_update(datas,labs,w,alpha)
error_rate=test_accuracy(datas,labs,w)
print("epoch %d error %.3f%%"%(i,error_rate*100))
return w
def test_accuracy(datas,labs,w):
N,D = np.shape(datas)
z = np.dot(datas,w) # Nx1
h = sigmoid(z) # Nx1
lab_det = (h>0.5).astype(np.float)
error_rate=np.sum(np.abs(labs-lab_det))/N
return error_rate
def draw_desion_line(datas,labs,w,name="0.jpg"):
dic_colors={0:(.8,0,0),1:(0,.8,0)}
#print(labs)
# 画数据点
for i in range(2):
#print(i)
index = np.where(labs==i)[0]
sub_datas = datas[index]
#print(sub_datas)
plt.scatter(sub_datas[:,1],sub_datas[:,2],s=16.,color=dic_colors[i])
# 画判决线
min_x = np.min(datas[:,1])
max_x = np.max(datas[:,1])
w = w[:,0]
x = np.arange(min_x,max_x,0.01)
#print(x)
y = -(x*w[1]+w[0])/w[2]
#y = -(x*w[2]+w[1])/w[0]
plt.plot(x,y)
#print(w)
plt.savefig(name)
def load_dataset(file):
with open(file,"r",encoding="utf-8") as f:
lines = f.read().splitlines()
# 取 lab 维度为 N x 1
labs = [line.split("\t")[-1] for line in lines]
labs = np.array(labs).astype(np.float32)
labs= np.expand_dims(labs,axis=-1) # Nx1
#rint(labs)
# 取数据 增加 一维全是1的特征
datas = [line.split("\t")[:-1] for line in lines]
datas = np.array(datas).astype(np.float32)
N,D = np.shape(datas)
# 增加一个维度
datas = np.c_[np.ones([N,1]),datas]
#print(datas)
return datas,labs
if __name__ == "__main__":
''' 实验1 基础测试数据'''
# 加载数据
file = "testset.txt"
datas,labs = load_dataset(file)
#weights = train_LR(datas,labs,alpha=0.001,n_epoch=800)
weights = train_LR_batch(datas,labs,batchsize=80,alpha=0.001,n_epoch=800)
print(weights)
draw_desion_line(datas,labs,weights,name="test_1.jpg")
训练需要的数据集testset.txt文件:
-0.017612 14.053064 0
-1.395634 4.662541 1
-0.752157 6.538620 0
-1.322371 7.152853 0
0.423363 11.054677 0
0.406704 7.067335 1
0.667394 12.741452 0
-2.460150 6.866805 1
0.569411 9.548755 0
-0.026632 10.427743 0
0.850433 6.920334 1
1.347183 13.175500 0
1.176813 3.167020 1
-1.781871 9.097953 0
-0.566606 5.749003 1
0.931635 1.589505 1
-0.024205 6.151823 1
-0.036453 2.690988 1
-0.196949 0.444165 1
1.014459 5.754399 1
1.985298 3.230619 1
-1.693453 -0.557540 1
-0.576525 11.778922 0
-0.346811 -1.678730 1
-2.124484 2.672471 1
1.217916 9.597015 0
-0.733928 9.098687 0
-3.642001 -1.618087 1
0.315985 3.523953 1
1.416614 9.619232 0
-0.386323 3.989286 1
0.556921 8.294984 1
1.224863 11.587360 0
-1.347803 -2.406051 1
1.196604 4.951851 1
0.275221 9.543647 0
0.470575 9.332488 0
-1.889567 9.542662 0
-1.527893 12.150579 0
-1.185247 11.309318 0
-0.445678 3.297303 1
1.042222 6.105155 1
-0.618787 10.320986 0
1.152083 0.548467 1
0.828534 2.676045 1
-1.237728 10.549033 0
-0.683565 -2.166125 1
0.229456 5.921938 1
-0.959885 11.555336 0
0.492911 10.993324 0
0.184992 8.721488 0
-0.355715 10.325976 0
-0.397822 8.058397 0
0.824839 13.730343 0
1.507278 5.027866 1
0.099671 6.835839 1
-0.344008 10.717485 0
1.785928 7.718645 1
-0.918801 11.560217 0
-0.364009 4.747300 1
-0.841722 4.119083 1
0.490426 1.960539 1
-0.007194 9.075792 0
0.356107 12.447863 0
0.342578 12.281162 0
-0.810823 -1.466018 1
2.530777 6.476801 1
1.296683 11.607559 0
0.475487 12.040035 0
-0.783277 11.009725 0
0.074798 11.023650 0
-1.337472 0.468339 1
-0.102781 13.763651 0
-0.147324 2.874846 1
0.518389 9.887035 0
1.015399 7.571882 0
-1.658086 -0.027255 1
1.319944 2.171228 1
2.056216 5.019981 1
-0.851633 4.375691 1
-1.510047 6.061992 0
-1.076637 -3.181888 1
1.821096 10.283990 0
3.010150 8.401766 1
-1.099458 1.688274 1
-0.834872 -1.733869 1
-0.846637 3.849075 1
1.400102 12.628781 0
1.752842 5.468166 1
0.078557 0.059736 1
0.089392 -0.715300 1
1.825662 12.693808 0
0.197445 9.744638 0
0.126117 0.922311 1
-0.679797 1.220530 1
0.677983 2.556666 1
0.761349 10.693862 0
-2.168791 0.143632 1
1.388610 9.341997 0
0.317029 14.739025 0
anaconda jupyter notebook中的运行结果截图: