0. 案例背景
我们将建立一个逻辑回归模型来预测一个学生是否被大学录取。假设你是一个大学系的管理员,你想根据两次考试的结果来决定每个申请人的录取机会。你有以前的申请人的历史数据,你可以用它作为逻辑回归的训练集。对于每一个培训例子,你有两个考试的申请人的分数和录取决定。为了做到这一点,我们将建立一个分类模型,根据考试成绩估计入学概率。
1. 导入pythony库
#导入机器学习三大件:Numpy, Pandas, Matplotlib
import numpy as np
import pandas as pd
import matplotlib
import matplotlib.pyplot as plt
%matplotlib inline
for i in [np, pd, matplotlib]:
print(i.__version__)
输出:
1.17.4
0.25.3
3.1.2
2. 导入数据集
import os
path = 'data' + os.sep + 'LogiReg_data.txt'
pdData = pd.read_csv(path, header=None, names=['Exam1', 'Exam2', 'Admitted'])
print(pdData.head(8))
print("Data shape:",pdData.shape)
输出:
Exam1 Exam2 Admitted
0 34.623660 78.024693 0
1 30.286711 43.894998 0
2 35.847409 72.902198 0
3 60.182599 86.308552 1
4 79.032736 75.344376 1
5 45.083277 56.316372 0
6 61.106665 96.511426 1
7 75.024746 46.554014 1
Data shape: (100, 3)
3. 数据可视化
positive = pdData[pdData['Admitted'] == 1] # 返回Admitted列值为1的样本
negative = pdData[pdData['Admitted'] == 0] # 返回Admitted列值为1的样本
fig, ax = plt.subplots(figsize=(10,5))
ax.scatter(positive['Exam1'], positive['Exam2'], s=30, c='b', marker='o', label='Admitted')
ax.scatter(negative['Exam1'], negative['Exam2'], s=30, c='r', marker='x', label='Not Admitted')
ax.legend()
ax.set_xlabel('Exam1 Score')
ax.set_ylabel('Exam2 Score')
输出:
从上图可以看出:数据具有一定的可区分性。红色点(负样本)总体位于左下方,而蓝色点(正样本)总体位于右上方。
4. 数据处理
在模型构建中为了方便书写和计算,将偏置项用W0表示,因此在数据所有样本的第一列插入1,即b=W0*1
try:
pdData.insert(0, 'Ones', 1) # 写到try...except结构中以防第二次执行时报错
except:
pass
orig_data = pdData.as_matrix() # 将Pandas的DataFrame转换成矩阵形成
print(type(orig_data))
print(orig_data.shape)
cols = orig_data.shape[1]
X = orig_data[:,0:cols-1]
y = orig_data[:,cols-1:cols]
print("X shape:",X.shape)
print("y shape:",y.shape)
#X = np.matrix(X.values)
#y = np.matrix(data.iloc[:,3:4].values) #np.array(y.values)
W = np.zeros([3, 1])
print("W shape:",W.shape)
输出:
<class 'numpy.ndarray'>
(100, 4)
X shape: (100, 3)
y shape: (100, 1)
W shape: (3, 1)
根据X, y, W的shape以及矩阵乘法相关知识可得:XW=y
5. 用Python手动实现逻辑回归
目标:建立逻辑回归模型,即求解出参数𝑊W。由于数据共有3列:Exam1,Exam2,Admitted,其中Admitted作为标签使用。因此𝑊可以表示为3*1的向量(这里偏置项用𝑊0代替,方便书写和矩阵计算)。
设定阈值,根据阈值判断录取结果
需要完成的模块
sigmoid
: 映射到概率的函数model
: 返回预测结果值cost
: 根据参数计算损失gradient
: 计算每个参数的梯度方向descent
: 进行参数更新accuracy
: 计算精度
5.1 sigmoid 函数
def sigmoid(z):
return 1 / (1 + np.exp(-z))
nums = np.arange(-10, 10, step=0.01)
fig, ax = plt.subplots(figsize=(10,5))
ax.plot(nums, sigmoid(nums), 'r')
plt.show()
输出:
可以看到Sigmoid函数的值域为(0,1),可以当作概率来看待。
5.2 model函数
def model(X, W):
return sigmoid(np.dot(X, W))
5.3 损失函数
根据逻辑回归目标函数:
公式推导见:逻辑回归(Logistic Regression)算法原理推导
def Loss(X, y, W):
left = np.multiply(y, np.log(model(X, W)))
right = np.multiply(1 - y, np.log(1 - model(X, W)))
return -np.sum(left + right) / (len(X))
print(Loss(X, y, W))
输出:
0.6931471805599453
5.4 梯度函数
公式推导见:逻辑回归(Logistic Regression)算法原理推导
def Gradient(X, y, W):
grad = np.zeros(W.shape)
error = (model(X, W)- y)
grad = np.dot(X.T, error)/len(X)
return grad
grad = Gradient(X, y, W)
print(grad)
输出:
[[ -0.1 ]
[-12.00921659]
[-11.26284221]]
输出即为初始的梯度。
5.5 梯度下降(Gradient descent)
比较3中不同梯度下降方法
STOP_ITER = 0
STOP_COST = 1
STOP_GRAD = 2
def stopCriterion(type, value, threshold):
#设定三种不同的停止策略
if type == STOP_ITER: #
return value > threshold
elif type == STOP_COST:
return abs(value[-1]-value[-2]) < threshold
elif type == STOP_GRAD:
return np.linalg.norm(value) < threshold
def shuffleData(data):
np.random.shuffle(data)
cols = data.shape[1]
X = data[:, 0:cols-1]
y = data[:, cols-1:]
return X, y
import time
def descent(data, W, batchSize, stopType, thresh, alpha):
#梯度下降求解
init_time = time.time()
i = 0 # 迭代次数
k = 0 # batch
X, y = shuffleData(data)
grad = np.zeros(W.shape) # 计算的梯度
costs = [Loss(X, y, W)] # 损失值
while True:
grad = Gradient(X[k:k+batchSize], y[k:k+batchSize], W)
k += batchSize #取batch数量个数据
if k >= n:
k = 0
X, y = shuffleData(data) #重新洗牌
W = W - alpha*grad # 参数更新
costs.append(Loss(X, y, W)) # 计算新的损失
i += 1
if stopType == STOP_ITER:
value = i
elif stopType == STOP_COST:
value = costs
elif stopType == STOP_GRAD:
value = grad
if stopCriterion(stopType, value, thresh):
break
return W, i-1, costs, grad, time.time() - init_time
def runExpe(data, W, batchSize, stopType, thresh, alpha):
#import pdb; pdb.set_trace();
W, iter, costs, grad, dur = descent(data, W, batchSize, stopType, thresh, alpha)
name = "Original" if (data[:,1]>2).sum() > 1 else "Scaled"
name += " data - learning rate: {} - ".format(alpha)
if batchSize==n: strDescType = "Gradient"
elif batchSize==1: strDescType = "Stochastic"
else: strDescType = "Mini-batch ({})".format(batchSize)
name += strDescType + " descent - Stop: "
if stopType == STOP_ITER: strStop = "{} iterations".format(thresh)
elif stopType == STOP_COST: strStop = "costs change < {}".format(thresh)
else: strStop = "gradient norm < {}".format(thresh)
name += strStop
print ("***{}\nW: {} - \nIter: {} - Last cost: {:03.2f} - Duration: {:03.2f}s".format(
name, W, iter, costs[-1], dur))
fig, ax = plt.subplots(figsize=(12,4))
ax.plot(np.arange(len(costs)), costs, 'r')
ax.set_xlabel('Iterations')
ax.set_ylabel('Cost')
ax.set_title(name.upper() + ' - Error vs. Iteration')
return W
5.6 研究不同的停止策略对结果的影响
5.6.1 设定迭代次数
n=100 #选择的梯度下降方法是基于所有样本的,因为总共就有100个样本
runExpe(orig_data, W, n, STOP_ITER, thresh=5000, alpha=0.000001)
输出:
***Original data - learning rate: 1e-06 - Gradient descent - Stop: 5000 iterations
W: [[-0.00027127]
[ 0.00705232]
[ 0.00376711]] -
Iter: 5000 - Last cost: 0.63 - Duration: 0.72s
Out[297]:
array([[-0.00027127],
[ 0.00705232],
[ 0.00376711]])
5.6.2 根据损失值停止
设定阈值 1E-6, 差不多需要110 000次迭代
n = 100
runExpe(orig_data, W, n, STOP_COST, thresh=0.000001, alpha=0.001)
输出:
***Original data - learning rate: 0.001 - Gradient descent - Stop: costs change < 1e-06
W: [[-5.13364014]
[ 0.04771429]
[ 0.04072397]] -
Iter: 109901 - Last cost: 0.38 - Duration: 15.96s
Out[268]:
array([[-5.13364014],
[ 0.04771429],
[ 0.04072397]])
5.6.3 根据梯度变化停止
设定阈值 0.05,差不多需要40 000次迭代
n = 100
runExpe(orig_data, W, n, STOP_GRAD, thresh=0.05, alpha=0.001)
输出:
***Original data - learning rate: 0.001 - Gradient descent - Stop: gradient norm < 0.05
Theta: [[1.14934134]
[2.79258729]
[2.56635486]] - Iter: 40045 - Last cost: 0.49 - Duration: 6.46s
Out[174]:
array([[-2.37033409],
[ 0.02721692],
[ 0.01899456]])
5.7 对比不同的梯度下降方法
注意:下面几个随机梯度下降的例子每次运行的结果可能不一样,因为是每次选取的随机样本不一样。
5.7.1 随机梯度下降(Stochastic descent)
n = 1
runExpe(orig_data, W, n, STOP_ITER, thresh=5000, alpha=0.001)
输出:
***Original data - learning rate: 0.001 - Gradient descent - Stop: 5000 iterations
W: [[-0.36353926]
[ 0.04708261]
[-0.06730932]] -
Iter: 5000 - Last cost: 1.42 - Duration: 0.71s
Out[301]:
array([[-0.36353926],
[ 0.04708261],
[-0.06730932]])
由上图可以看出,损失值振荡幅度很大。尝试调小学习率。
runExpe(orig_data, W, 1, STOP_ITER, thresh=15000, alpha=0.00002)
输出:
***Original data - learning rate: 2e-05 - Stochastic descent - Stop: 15000 iterations
Theta: [[1.14934134]
[2.79258729]
[2.56635486]] - Iter: 15000 - Last cost: 0.63 - Duration: 0.73s
Out[177]:
array([[-0.02101212],
[ 0.01132223],
[ 0.00104678]])
可以看出振荡幅度稍微有些缓和,可自行尝试更小的学习率,观察损失值振荡是否还会减弱。
5.7.2 批量随机梯度下降(Mini-batch descent)
可尝试16,32,64等不同的批大小:
批大小为16时:
runExpe(orig_data, W, 16, STOP_ITER, thresh=15000, alpha=0.001)
输出:
***Original data - learning rate: 0.001 - Mini-batch (16) descent - Stop: 15000 iterations
W: [[-1.01560025]
[ 0.01830399]
[ 0.01316626]] -
Iter: 15000 - Last cost: 0.58 - Duration: 2.10s
Out[304]:
array([[-1.01560025],
[ 0.01830399],
[ 0.01316626]])
批大小为64时:
runExpe(orig_data, W, 64, STOP_ITER, thresh=15000, alpha=0.001)
输出:
***Original data - learning rate: 0.001 - Mini-batch (64) descent - Stop: 15000 iterations
W: [[-0.99214625]
[ 0.01786299]
[ 0.00733927]] -
Iter: 15000 - Last cost: 0.56 - Duration: 2.10s
Out[306]:
array([[-0.99214625],
[ 0.01786299],
[ 0.00733927]])
对比批大小16和64的结果图:很明显批大小为64时损失值振荡幅度更小。(注意:两张图的纵坐标范围大小不一样)
5.8 研究数据处理对结果的影响
在机器学习建模前,通常需要对数据进行预处理。常见的处理方法是标准化(或称归一化),即将数据按其属性(按列进行)减去其均值,然后除以其方差,得到均值为0,方差为1的数据集。
from sklearn import preprocessing as pp
scaled_data = orig_data.copy()
scaled_data[:, 1:3] = pp.scale(orig_data[:, 1:3])
n = 100
runExpe(scaled_data, W, n, STOP_ITER, thresh=5000, alpha=0.001)
输出:
***Scaled data - learning rate: 0.001 - Gradient descent - Stop: 5000 iterations
W: [[0.3080807 ]
[0.86494967]
[0.77367651]] -
Iter: 5000 - Last cost: 0.38 - Duration: 0.74s
Out[316]:
array([[0.3080807 ],
[0.86494967],
[0.77367651]])
5.6中利用原始数据建模预测的结果,最好的结果只能达到达到0.61。而对数据进行预处理后,已经达到了0.38好结果。通过调整其它参数,将会得到后面所示更好的结果。
n= 100
runExpe(scaled_data, W, n, STOP_GRAD, thresh=0.02, alpha=0.001)
输出:
***Scaled data - learning rate: 0.001 - Gradient descent - Stop: gradient norm < 0.02
W: [[1.0707921 ]
[2.63030842]
[2.41079787]] -
Iter: 59422 - Last cost: 0.22 - Duration: 9.05s
Out[319]:
array([[1.0707921 ],
[2.63030842],
[2.41079787]])
n = 100
runExpe(scaled_data, W, 1, STOP_GRAD, thresh=0.002/5, alpha=0.001)
输出:
***Scaled data - learning rate: 0.001 - Stochastic descent - Stop: gradient norm < 0.0004
W: [[1.14745059]
[2.79268354]
[2.56859444]] -
Iter: 72681 - Last cost: 0.22 - Duration: 3.32s
Out[320]:
array([[1.14745059],
[2.79268354],
[2.56859444]])
n = 100
runExpe(scaled_data, W, 16, STOP_GRAD, thresh=0.002*2, alpha=0.001)
输出:
***Scaled data - learning rate: 0.001 - Mini-batch (16) descent - Stop: gradient norm < 0.004
W: [[1.04991433]
[2.57978973]
[2.36638254]] -
Iter: 56083 - Last cost: 0.22 - Duration: 3.44s
Out[321]:
array([[1.04991433],
[2.57978973],
[2.36638254]])
通过尝试不同的参数发现,最好结果停留在损失值为0.22的结果上。
5.9 精确度计算函数
#设定阈值
def predict(X, W):
return [1 if x >= 0.5 else 0 for x in model(X, W)]
n = 100
W = runExpe(scaled_data, W, 64, STOP_GRAD, thresh=0.002*2, alpha=0.001) # 将W保存下来
scaled_X = scaled_data[:, :3]
y = scaled_data[:, 3]
predictions = predict(scaled_X, W)
correct = [1 if ((a == 1 and b == 1) or (a == 0 and b == 0)) else 0 for (a, b) in zip(predictions, y)]
accuracy = (sum(map(int, correct)) % len(correct))
print ('accuracy = {0}%'.format(accuracy))
输出:
***Scaled data - learning rate: 0.001 - Mini-batch (64) descent - Stop: gradient norm < 0.004
W: [[1.11200715]
[2.7166569 ]
[2.49287266]] -
Iter: 29877 - Last cost: 0.22 - Duration: 2.98s
accuracy = 89%
可以发现,模型最终的精确度达到89%。