-
Notifications
You must be signed in to change notification settings - Fork 2
/
svm_test.py
179 lines (152 loc) · 7.71 KB
/
svm_test.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
172
173
174
175
176
177
178
179
from numpy import *
from plotBoundary import *
# import your SVM training code
import svm
import numpy
from utils import *
from svm import *
from plotBoundary import *
class SVMTrain(Train):
def _generateTitle(self):
return ", error=%s, %s formulation with $C=%s$"
@counted
def _train(self, X, Y):
## save for use in _getPredictor below
self.tX = X
self.tY = Y
## first, cvxopt output suppression
if not self.printInfo: cvxopt.solvers.options['show_progress'] = False
phi = makePhi(X,self.M)
n,m = phi.shape
primal = self.params['primal']
C = self.params['C']
if primal:
self.result = svm.primal(phi,Y,C)
w = numpy.array(self.result['x'][:m])
b = self.result['x'][m]
else:
self.kernel = self.params['kernel']
self.result = svm.dual(phi,Y,C,self.kernel)
self.alphaD = numpy.array(self.result['x'])
w,b,self.dualS,self.dualM = svm.dualWeights(phi, Y, self.kernel, self.alphaD, C)
self.sv = self.numSupport(self.alphaD)
self.slack = numpy.array(self.result['z'])[:-n]
return w,b
@counted
def _getPredictor(self):
self.kernel = self.params['kernel']
if not self.params['primal']:
return makeKernelPredictor(self.w, self.b, self.M, self.alphaD, self.tX, self.tY, self.kernel, self.params['C'], self.params)
#else: return super(SVMTrain, self)._getPredictor()
else:
return super(SVMTrain, self)._getPredictor()
def dummy():
## a very simple test
xDummy = numpy.array([[0,0],[1,1],[1,2]])
assert xDummy.shape == (3,2)
## quadratic basis function
M=2
phiDummy = makePhi(xDummy,M)
assert phiDummy.shape == (3,6)
n,m = phiDummy.shape
yDummy = numpy.array([[-1], [1], [1]])
assert yDummy.shape == (3,1)
# Carry out training, primal and/or dual
C = 10e10
C = 0.25
p = svm.primal(phiDummy,yDummy,C)
w = numpy.array(p['x'][:m])
assert w.shape == (6,1)
b = p['x'][m]
assert isinstance(b, (int,float))
dummyPredictor = makePredictor(w,b,M,'svm')
plotDecisionBoundary(xDummy, yDummy, dummyPredictor, [-1, 0, 1], title = 'SVM Validate')
def cSweep():
problemClass = "svm"
varName = "C"
## try an exponential sweep up to around 250
#cVals = numpy.array([0] + [10**i for i in numpy.arange(-4,2.5,0.2)])
## -2 to 3
cVals = numpy.array([0] + [10**i for i in numpy.arange(-2,3.1,1)])
# ## first try the primal with linear basis functions, linear kernel
# resultsLin = numpy.array([SVMTrain({'primal':True, 'C':C}, basisfunc='lin', plot=False, printInfo=False)() for c in cVals])
# trainingErrLin = resultsLin[:,0]
# validationErrLin = resultsLin[:,1]
# plotTVError(lambdaVals, trainingErrLin, validationErrorLin, problemClass=problemClass, varName=varName, linQuad='linear', extra='')
# # now try the primal with quadratic basis function, linear kernel
# resultsQuad = numpy.array([SVMTrain({'primal':True, 'C':C}, basisfunc='quad', plot=False, printInfo=False)() for c in cVals])
# trainingErrQuad = resultsQuad[:,0]
# validationErrQuad = resultsQuad[:,1]
# plotTVError(lambdaVals, trainingErrQuad, validationErrorQuad, problemClass=problemClass, varName=varName, linQuad='quadratic', extra='')
def plotTVErrorWrapper(cVals, primal, kernelName, kernel, linQuad, problemClass, varName):
beta = 100
gaussianKernel = Kernel(makeGaussian(beta))
print "primal=%s, linQuad=%s, kernelName=%s, beta=%s" %(primal, linQuad, kernelName, beta)
results = numpy.array([SVMTrain({'primal':(True if primal=='primal' else False),'C':C, 'kernel':gaussianKernel,'kernelName':'Gaussian','beta':beta}, dataSetName='ls',problemClass='svm', basisfunc='lin', printInfo=True, plot=False, meshSize=145.)() for C in cVals])
print results
## a test
#results = numpy.array([(0,C) for C in cVals])
trainingErr = results[:,0]
validationErr = results[:,1]
gm = results[:,2]
sv = results[:,3]
plotTVError(cVals, trainingErr, validationErr, gm=gm, sv=sv, problemClass=problemClass, varName=varName, linQuad=('quadratic' if linQuad=='quad' else 'linear'), extra=' and a %s kernel, %s form' %(kernelName, primal))
return numpy.array([results, trainingErr, validationErr])
kernelName = "Gaussian"
linQuad = "lin"
p = 'dual'
kernel = gaussianKernel
plotTVErrorWrapper(cVals, p, kernelName, kernel, linQuad, problemClass, varName)
# kernels = [('linear',linearKernel), ('quadratic',squaredKernel), ('Gaussian',gaussianKernel)]
# lq = ['lin','quad']
# p = ['primal','dual']
# ## plot and return all 3*2*2 = 12 possibilities
# result = numpy.array([plotTVErrorWrapper(cVals, primal, kernelName, kernel, linQuad, problemClass, varName) for kernelName, kernel in kernels for linQuad in lq for primal in p])
#return cVals, result
if __name__=='__main__':
#print SVMTrain({'primal':False,'C':0.1, 'kernel':gaussianKernel}, problemClass='svm', basisfunc='quad', printInfo=True, plot=True)()
#gaussianKernel = Kernel(lambda a,b: exp(-0.3*((a-b).T.dot(a-b))))
#a=SVMTrain({'primal':False,'C':10, 'kernel':squaredKernel}, problemClass='svm', basisfunc='lin', printInfo=True, plot=True)
#e = a()
#print e
#cSweep()
## a simple check
dummyX = numpy.array([[1,1],
[2,2],
[3,3]])
dummyY = numpy.array([[-1],
[1],
[-1]])
#dummyX = numpy.array([[1,1],[0.9,1.02],[1.02,0.93],[1,1],[0.9,1.02],[1.02,0.93],[1,1],[0.9,1.02],[1.02,0.93],
# [2.02,2.2],[2.3,2.4],[2,2],[2.02,2.2],[2.3,2.4],[2,2],[2.02,2.2],[2.3,2.4],[2,2],
# [1.4, 1.4],
# [1.6,1.6]
# ])
#dummyY = numpy.array([[-1],[-1],[-1],[-1],[-1],[-1],[-1],[-1],[-1],
# [1],[1],[1],[1],[1],[1],[1],[-1],[1],
# [-1],
# [1]])
beta = 1
#b=SVMTrain({'primal':True,'C':1, 'kernel':gaussianKernel,'kernelName':'Gaussian','beta':beta}, problemClass='svm', basisfunc='lin', printInfo=True, plot=True)
#print b._computeError(dummyX, dummyY)
b=SVMTrain({'primal':False,'C':1, 'kernel':linearKernel,'kernelName':'linear','beta':beta}, dataSetName='nonlin',problemClass='svm', basisfunc='lin', printInfo=True, plot=True, meshSize=145.)
print b()
## try some more nonlinear data
## worked well with beta=10
#s=SVMTrain({'primal':False,'C':.1, 'kernel':gaussianKernel}, problemClass='svm', basisfunc='lin', dataSetName='nls',printInfo=True, plot=True)
#print s()
# now try with quadratic basis function
problemName = 'nls'
knn = [(linearKernel,'linear'), (squaredKernel,'second-order polynomial'), (gaussianKernel, 'Gaussian')]
kernel,kernelName = knn[0]
def plotTV(pn, k, kn):
results = numpy.array([SVMTrain({'C':c,'primal':False,'kernel':k,'kernelName':kn,'beta':beta}, problemClass='svm', basisfunc='lin', plot=False, printInfo=True, dataSetName=pn, meshSize=125.)() for c in [10**i for i in [-2,-1,0,1,2]]])
trainingErr = results[:,0]
validationErr = results[:,1]
gm = results[:,2]
sv = results[:,3]
plotTVError([10**i for i in [-2,-1,0,1,2]], trainingErr, validationErr, gm=gm, sv=sv, problemClass="svm", varName="C", linQuad='lin', extra=' for ' + pn.upper() + " data (dual with " + kn + " kernel)")
return results
#results = [plotTV(pn, k, kn) for pn in ['ls','nls','nonlin'] for k,kn in knn]
vals = [(pn, k, kn) for pn in ['ls','nls','nonlin'] for k,kn in knn]
pl.show()