-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathpolynomial.py
172 lines (152 loc) · 6.31 KB
/
polynomial.py
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
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
import sys
import json
from itertools import product
from sklearn.linear_model import LogisticRegression
from sklearn.metrics import classification_report, confusion_matrix
from sklearn.model_selection import StratifiedShuffleSplit
import random as rn
import numpy as np
import matplotlib.pyplot as plt
rn.seed(1)
np.random.seed(1)
lmin = 1.0e-3
lmax = 1-1.0e-3
#
# f(x,y) = c0 * 1 + c1 * x + c2 * y + c3 * x^2 + c4 * x * y + c5 * y^2
#
# 2.25 -3.0*x -18.0*y + 35.0 * x * x - 64.0 * x * y + 50.0 * y * y
# prp ./polynomial.py 1000 2.25 -3.0 -18.0 35.0 -64.0 50.0
args = sys.argv
true_w = np.zeros(6)
nPoints = int(args[1])
true_w[0] = float(args[2])
true_w[1] = float(args[3])
true_w[2] = float(args[4])
true_w[3] = float(args[5])
true_w[4] = float(args[6])
true_w[5] = float(args[7])
if (true_w[0] == 0):
print ('For easy scaling & comparison, we use a non-zero intercept here... exiting')
sys.exit(0)
def getF(w, data):
return np.dot(data, w)
def generateData():
ones = np.ones(nPoints*15)
x = (lmax - lmin) * np.random.random_sample(nPoints*15) + lmin
y = (lmax - lmin) * np.random.random_sample(nPoints*15) + lmin
allData = np.stack((ones, x, y, x*x, x*y, y*y),axis=-1)
fvals = getF(true_w, allData)
data, labels = [], []
zeroCount, oneCount = 0, 0
for i,f in enumerate(fvals):
if ( (zeroCount < nPoints) and (f > 1.0e-1) ):
labels.append(0.0)
data.append(allData[i])
zeroCount = zeroCount + 1
if ( (oneCount < nPoints) and (f < -1.0e-1) ):
labels.append(1.0)
data.append(allData[i])
oneCount = oneCount + 1
if (zeroCount == nPoints) and (oneCount == nPoints):
break;
data=np.array([np.array(xi) for xi in data])
labels = np.array(labels)
return data, labels
data, labels = generateData()
zeroPoints = np.where(labels == 0.0)
data0 = data[zeroPoints[0]]
onePoints = np.where(labels == 1.0)
data1 = data[onePoints[0]]
print ('n_zero, n_one:', len(data0), len(data1))
sss = StratifiedShuffleSplit(n_splits=1, test_size=0.2, random_state=1).split(data, labels)
train_indices, test_indices = next(sss)
# True boundary
true_w_scaled = true_w/true_w[0]
# SK. Just the linear data
lr = LogisticRegression(max_iter=10000, verbose=0, tol=1.0e-8, solver='lbfgs' )
lr.fit(data[train_indices][:,1:3], labels[train_indices])
linear_predicted_labels = lr.predict(data[test_indices][:,1:3])
w = [lr.intercept_[0], lr.coef_[0][0], lr.coef_[0][1]]
linear_w_scaled = w/w[0]
# SK. Augmented with non linear data
lr2 = LogisticRegression(max_iter=10000, verbose=0, tol=1.0e-8, solver='lbfgs' )
lr2.fit(data[train_indices][:,1:6], labels[train_indices])
nonlinear_predicted_labels = lr2.predict(data[test_indices][:,1:6])
w = [lr2.intercept_[0], lr2.coef_[0][0], lr2.coef_[0][1], lr2.coef_[0][2], lr2.coef_[0][3], lr2.coef_[0][4]]
nonlinear_w_scaled = w/w[0]
from scipy import optimize
data_train = data[train_indices]
labels_train = labels[train_indices]
def getLLforMin(w):
f = getF(w, data_train)
LL = np.dot(labels_train.T, f) - np.sum (np.log(1.0 + np.exp(f)))
return -LL
def getdLLbydwforMin(w):
f = getF(w, data_train)
dLLbydW = np.dot(data_train.T, labels_train - 1.0/(1+np.exp(-f)))
return -dLLbydW
w0 = true_w * 1.0e-9
print ('\nsolve df/dw = 0 \n')
root = optimize.fsolve(getdLLbydwforMin, w0)
solve_dLLbydw_0_w_scaled = root / root[0]
print ('True Weights:',true_w_scaled)
print ('Linear scikit-learn Weights:',linear_w_scaled)
print ('Augmented/Nonlinear scikit-learn Weights:',nonlinear_w_scaled)
print ('Solving dLL/dw = 0:',solve_dLLbydw_0_w_scaled)
#
# Plot decisoon boundaries
#
def solveQuad (w, z):
a = w[5]
b = w[4] * z + w[2]
c = w[0] + w[1] * z + w[3] * z * z
disc = b*b - 4 * a * c
y1, y2 = -99, -99
if (disc > 0):
y1 = (-b + (b*b - 4 * a * c)**0.5 ) / (2*a)
y2 = (-b - (b*b - 4 * a * c)**0.5 ) / (2*a)
return y1, y2
def getBoundary (w):
low = (lmax-lmin)/100
randX = np.random.uniform(low, high=low*2)
xvals = np.linspace(lmin+randX, lmax-randX, 100)
boundary_data = []
for i, x in enumerate(xvals):
y_1, y_2 = solveQuad (w, x)
if (y_1 != -99):
boundary_data.append ([x, y_1])
boundary_data.append ([x, y_2])
boundary_data=np.array([np.array(xi) for xi in boundary_data])
return boundary_data
xvals = np.linspace(lmin, lmax, 100)
linear_boundary_data = np.stack((xvals, ( - linear_w_scaled[0] - linear_w_scaled[1] * xvals ) / linear_w_scaled[2]),axis=-1)
nonlinear_boundary_data = getBoundary (nonlinear_w_scaled)
true_boundary_data = getBoundary (true_w_scaled)
fsolve_boundary_data = getBoundary (solve_dLLbydw_0_w_scaled)
fig = plt.figure(figsize=(8,8),dpi=720)
filename = 'predicted-boundaries'
plt.xlim(-0.05, 1.05)
plt.ylim(-0.05, 1.05)
#plt.plot(true_boundary_data[:, 0], true_boundary_data[:, 1], color='g', linestyle='', markersize=3,marker='*')
plt.plot(nonlinear_boundary_data[:, 0], nonlinear_boundary_data[:, 1], color='b', linestyle='', markersize=3,marker='<')
plt.plot(fsolve_boundary_data[:, 0], fsolve_boundary_data[:, 1], color='g', linestyle='', markersize=3,marker='>')
plt.plot(linear_boundary_data[:, 0], linear_boundary_data[:, 1], color='r')
plt.plot(data0[:, 1], data0[:, 2], color='k', linestyle='', markersize=1,marker='.')
plt.plot(data1[:, 1], data1[:, 2], color='c', linestyle='', markersize=1,marker='+')
fig.savefig(filename + '.png', format='png', dpi=720)
plt.close()
# fsolve predictions
fsolve_predicted_fvals = getF(solve_dLLbydw_0_w_scaled, data[test_indices])
fsolve_predicted_labels = np.zeros(fsolve_predicted_fvals.size)
onePoints = np.where(fsolve_predicted_fvals < 0.0)
fsolve_predicted_labels[onePoints] = 1.0
target_names = ['zero', 'one']
print ('LR linear:')
print (confusion_matrix(labels[test_indices], linear_predicted_labels))
print (classification_report(labels[test_indices], linear_predicted_labels, digits=4, target_names=target_names))
print ('LR nonlinear:')
print (confusion_matrix(labels[test_indices], nonlinear_predicted_labels))
print (classification_report(labels[test_indices], nonlinear_predicted_labels, digits=4, target_names=target_names))
print ('fsolve dLLbydw=0:')
print (confusion_matrix(labels[test_indices], fsolve_predicted_labels))
print (classification_report(labels[test_indices], fsolve_predicted_labels, digits=4, target_names=target_names))