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
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
|
#include "ranksvmtn.h"
#include<iostream>
#include<list>
#include"../tools/matrixIO.h"
using namespace std;
using namespace Eigen;
// Main terminating criteria
const int maxiter = 60; // max iteration count
const double prec=1e-10; // precision
// conjugate gradient
const double cg_prec=1e-10; // precision
const int cg_maxiter = 10;
const int ls_maxiter = 10;
// line search
const double line_prec=1e-10; // precision
const double line_turb=1e-15; // purturbation
int cal_Hs(RidList &D,const vector<int> &rank,const VectorXd &corr,const VectorXd &alpha,const VectorXd s,VectorXd &Hs)
{
Hs = VectorXd::Zero(s.rows());
VectorXd Ds(D.getSize());
for (int i=0;i<D.getSize();++i)
Ds(i) = D.getVec(i).dot(s);
VectorXd gamma(D.getSize());
for (int i=0;i<D.getSize();)
{
double g=0;
for (int j = D.getqSize()-1;j>=0;--j)
if (corr[rank[i+j]]>0)
gamma[rank[i+j]]=g;
else
g+=Ds[rank[i+j]];
g=0;
i+=D.getqSize();
for (int j = D.getqSize();j>0;--j)
if (corr[rank[i-j]]<0)
gamma[rank[i-j]]=g;
else
g+=Ds[rank[i-j]];
}
VectorXd tmp = alpha.cwiseProduct(Ds)-gamma;
VectorXd res = 0*s;
for (int i=0;i<D.getSize();++i)
res = res + D.getVec(i)*tmp[i];
Hs = s + C*res;
return 0;
}
int cg_solve(RidList &D,const vector<int> &rank,const VectorXd &corr,const VectorXd &alph,const VectorXd &b, VectorXd &x)
{
double alpha,beta,r_1,r_2;
int iter=0;
VectorXd q;
VectorXd Hs;
cal_Hs(D,rank,corr,alph,x,Hs);
VectorXd res = b - Hs;
VectorXd p = res;
while (1)
{
// Non preconditioned version
r_1 = res.dot(res);
if (iter)
LOG(INFO) << "CG iter "<<iter<<", r:"<<r_1;
if (r_1<cg_prec) // Terminate condition
break;
if (iter >= cg_maxiter)
{
LOG(INFO) << "CG forced termination by maxiter";
break;
}
if (iter){
beta = r_1 / r_2;
p = res + p * beta;
}
cal_Hs(D,rank,corr,alph,p,q);
alpha = r_1/p.dot(q);
x=x+p*alpha;
res=res-q*alpha;
++iter;
r_2=r_1;
}
return 0;
}
void ranksort(int l,int r,vector<int> &rank,VectorXd &ref)
{
int i=l,j=r,k;
double mid=ref(rank[(l+r)>>1]);
while (i<=j)
{
while (ref[rank[i]]<mid) ++i;
while (ref[rank[j]]>mid) --j;
if (i<=j)
{
k=rank[i];
rank[i]=rank[j];
rank[j]=k;
++i;
--j;
}
}
if (j>l)
ranksort(l,j,rank,ref);
if (i<r)
ranksort(i,r,rank,ref);
}
int cal_alpha_beta(const VectorXd &dw,const VectorXd &corr,RidList &D,vector<int> &rank,VectorXd &yt,VectorXd &alpha,VectorXd &beta)
{
long n = dw.rows();
yt = dw - corr;
alpha=VectorXd::Zero(n);
beta=VectorXd::Zero(n);
for (int i=0;i<n;++i) rank[i]=i;
for (int i=0;i<D.getSize();i+=D.getqSize())
{
int ed=i+D.getqSize()-1;
ranksort(i,ed,rank,yt);
double a=0,b=0;
for (int j=i;j<=ed;++j)
if (corr[rank[j]]<0)
{
alpha[rank[j]]=a;
beta[rank[j]]=b;
}
else
{
a+=1;
b+=yt[rank[j]];
}
a=b=0;
for (int j=ed;j>=i;--j)
if (corr[rank[j]]>0)
{
alpha[rank[j]]=a;
beta[rank[j]]=b;
}
else
{
a+=1;
b+=yt[rank[j]];
}
}
}
// line search using newton method
int line_search(const VectorXd &w,RidList &D,const VectorXd &corr,const VectorXd &step,double &t)
{
VectorXd Dd(D.getSize());
for (int i=0;i<D.getSize();++i)
Dd(i) = D.getVec(i).dot(step);
VectorXd alpha,beta,yt;
VectorXd grad;
VectorXd Hs;
int n=D.getSize();
vector<int> rank(D.getSize());
int iter = 0;
double g,h;
t = 0;
while (1)
{
grad=w+t*step;
for (int i=0;i<D.getSize();++i)
Dd(i) = D.getVec(i).dot(grad);
cal_alpha_beta(Dd,corr,D,rank,yt,alpha,beta);
VectorXd tmp = alpha.cwiseProduct(yt)-beta;
VectorXd res = 0*grad;
for (int i=0;i<n;++i)
res = res + D.getVec(i)*tmp[i];
grad = grad + C*res;
g = grad.dot(step);
cal_Hs(D,rank,corr,alpha,step,Hs);
h = Hs.dot(step);
g=g+line_turb;
h = h+line_turb;
t=t-g/h;
if (g*g/h<line_prec)
break;
++iter;
LOG(INFO) << "line search iter "<<iter<<", prec:"<<g*g/h;
if (iter >= ls_maxiter)
{
LOG(INFO) << "line search forced termination by maxiter";
break;
}
}
return 0;
}
int train_orig(int fsize, RidList &Data,const VectorXd &corr,VectorXd &weight){
int iter = 0;
long n=Data.getSize();
LOG(INFO) << "training with feature size:" << fsize << " Data size:" << n << " Query size:" << Data.getuSize();
VectorXd grad(fsize);
VectorXd step(fsize);
vector<int> rank(n);
double obj,t;
VectorXd dw(n);
VectorXd yt;
VectorXd alpha,beta;
while (true)
{
iter+=1;
if (iter> maxiter)
{
LOG(INFO)<< "Maxiter reached";
break;
}
for (int i=0;i<n;++i)
dw(i) = Data.getVec(i).dot(weight);
cal_alpha_beta(dw,corr,Data,rank,yt,alpha,beta);
// Generate support vector matrix sv & gradient
obj = (weight.dot(weight) + C*(alpha.dot(yt.cwiseProduct(yt))-beta.dot(yt)))/2;
VectorXd tmp = alpha.cwiseProduct(yt)-beta;
VectorXd res = 0*weight;
for (int i=0;i<n;++i)
res = res + Data.getVec(i)*tmp[i];
grad = weight + C*res;
step = grad*0;
// Solve
cg_solve(Data,rank,corr,alpha,grad,step);
// do line search
line_search(weight,Data,corr,step,t);
weight=weight+step*t;
// When dec is small enough
LOG(INFO)<<"Iter: "<<iter<<" Obj: " <<obj << " Newton decr:"<<step.dot(grad)/2 << " linesearch: "<< -t ;
if (step.dot(grad) < prec * obj)
break;
}
return 0;
}
int RSVMTN::train(RidList &D){
VectorXd corr(D.getSize());
vector<int> A1,A2;
int i,j;
LOG(INFO)<<"Processing input";
for (i=0;i<D.getSize();++i)
corr(i)=D.getL(i)>0?0.5:-0.5;
train_orig(fsize,D,corr,model.weight);
return 0;
};
int RSVMTN::predict(DataList &D, vector<double> &res){
res.clear();
for (int i=0;i<D.getSize();++i)
res.push_back(((D.getData()[i])->feature).dot(model.weight));
return 0;
};
|
