Statistics
| Branch: | Revision:

ffmpeg / libavutil / pca.c @ 4869f47e

History | View | Annotate | Download (6.22 KB)

1
/*
2
 * Principal component analysis
3
 * Copyright (c) 2004 Michael Niedermayer <michaelni@gmx.at>
4
 *
5
 * This library is free software; you can redistribute it and/or
6
 * modify it under the terms of the GNU Lesser General Public
7
 * License as published by the Free Software Foundation; either
8
 * version 2 of the License, or (at your option) any later version.
9
 *
10
 * This library is distributed in the hope that it will be useful,
11
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
12
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
13
 * Lesser General Public License for more details.
14
 *
15
 * You should have received a copy of the GNU Lesser General Public
16
 * License along with this library; if not, write to the Free Software
17
 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
18
 *
19
 */
20

    
21
/**
22
 * @file pca.c
23
 * Principal component analysis
24
 */
25

    
26
#include "common.h"
27
#include "pca.h"
28

    
29
typedef struct PCA{
30
    int count;
31
    int n;
32
    double *covariance;
33
    double *mean;
34
}PCA;
35

    
36
PCA *ff_pca_init(int n){
37
    PCA *pca;
38
    if(n<=0)
39
        return NULL;
40

    
41
    pca= av_mallocz(sizeof(PCA));
42
    pca->n= n;
43
    pca->count=0;
44
    pca->covariance= av_mallocz(sizeof(double)*n*n);
45
    pca->mean= av_mallocz(sizeof(double)*n);
46

    
47
    return pca;
48
}
49

    
50
void ff_pca_free(PCA *pca){
51
    av_freep(&pca->covariance);
52
    av_freep(&pca->mean);
53
    av_free(pca);
54
}
55

    
56
void ff_pca_add(PCA *pca, double *v){
57
    int i, j;
58
    const int n= pca->n;
59

    
60
    for(i=0; i<n; i++){
61
        pca->mean[i] += v[i];
62
        for(j=i; j<n; j++)
63
            pca->covariance[j + i*n] += v[i]*v[j];
64
    }
65
    pca->count++;
66
}
67

    
68
int ff_pca(PCA *pca, double *eigenvector, double *eigenvalue){
69
    int i, j, k, pass;
70
    const int n= pca->n;
71
    double z[n];
72

    
73
    memset(eigenvector, 0, sizeof(double)*n*n);
74

    
75
    for(j=0; j<n; j++){
76
        pca->mean[j] /= pca->count;
77
        eigenvector[j + j*n] = 1.0;
78
        for(i=0; i<=j; i++){
79
            pca->covariance[j + i*n] /= pca->count;
80
            pca->covariance[j + i*n] -= pca->mean[i] * pca->mean[j];
81
            pca->covariance[i + j*n] = pca->covariance[j + i*n];
82
        }
83
        eigenvalue[j]= pca->covariance[j + j*n];
84
        z[j]= 0;
85
    }
86

    
87
    for(pass=0; pass < 50; pass++){
88
        double sum=0;
89

    
90
        for(i=0; i<n; i++)
91
            for(j=i+1; j<n; j++)
92
                sum += fabs(pca->covariance[j + i*n]);
93

    
94
        if(sum == 0){
95
            for(i=0; i<n; i++){
96
                double maxvalue= -1;
97
                for(j=i; j<n; j++){
98
                    if(eigenvalue[j] > maxvalue){
99
                        maxvalue= eigenvalue[j];
100
                        k= j;
101
                    }
102
                }
103
                eigenvalue[k]= eigenvalue[i];
104
                eigenvalue[i]= maxvalue;
105
                for(j=0; j<n; j++){
106
                    double tmp= eigenvector[k + j*n];
107
                    eigenvector[k + j*n]= eigenvector[i + j*n];
108
                    eigenvector[i + j*n]= tmp;
109
                }
110
            }
111
            return pass;
112
        }
113

    
114
        for(i=0; i<n; i++){
115
            for(j=i+1; j<n; j++){
116
                double covar= pca->covariance[j + i*n];
117
                double t,c,s,tau,theta, h;
118

    
119
                if(pass < 3 && fabs(covar) < sum / (5*n*n)) //FIXME why pass < 3
120
                    continue;
121
                if(fabs(covar) == 0.0) //FIXME shouldnt be needed
122
                    continue;
123
                if(pass >=3 && fabs((eigenvalue[j]+z[j])/covar) > (1LL<<32) && fabs((eigenvalue[i]+z[i])/covar) > (1LL<<32)){
124
                    pca->covariance[j + i*n]=0.0;
125
                    continue;
126
                }
127

    
128
                h= (eigenvalue[j]+z[j]) - (eigenvalue[i]+z[i]);
129
                theta=0.5*h/covar;
130
                t=1.0/(fabs(theta)+sqrt(1.0+theta*theta));
131
                if(theta < 0.0) t = -t;
132

    
133
                c=1.0/sqrt(1+t*t);
134
                s=t*c;
135
                tau=s/(1.0+c);
136
                z[i] -= t*covar;
137
                z[j] += t*covar;
138

    
139
#define ROTATE(a,i,j,k,l) {\
140
    double g=a[j + i*n];\
141
    double h=a[l + k*n];\
142
    a[j + i*n]=g-s*(h+g*tau);\
143
    a[l + k*n]=h+s*(g-h*tau); }
144
                for(k=0; k<n; k++) {
145
                    if(k!=i && k!=j){
146
                        ROTATE(pca->covariance,FFMIN(k,i),FFMAX(k,i),FFMIN(k,j),FFMAX(k,j))
147
                    }
148
                    ROTATE(eigenvector,k,i,k,j)
149
                }
150
                pca->covariance[j + i*n]=0.0;
151
            }
152
        }
153
        for (i=0; i<n; i++) {
154
            eigenvalue[i] += z[i];
155
            z[i]=0.0;
156
        }
157
    }
158

    
159
    return -1;
160
}
161

    
162
#ifdef TEST
163

    
164
#undef printf
165
#undef random
166
#include <stdio.h>
167
#include <stdlib.h>
168

    
169
int main(){
170
    PCA *pca;
171
    int i, j, k;
172
#define LEN 8
173
    double eigenvector[LEN*LEN];
174
    double eigenvalue[LEN];
175

    
176
    pca= ff_pca_init(LEN);
177

    
178
    for(i=0; i<9000000; i++){
179
        double v[2*LEN+100];
180
        double sum=0;
181
        int pos= random()%LEN;
182
        int v2= (random()%101) - 50;
183
        v[0]= (random()%101) - 50;
184
        for(j=1; j<8; j++){
185
            if(j<=pos) v[j]= v[0];
186
            else       v[j]= v2;
187
            sum += v[j];
188
        }
189
/*        for(j=0; j<LEN; j++){
190
            v[j] -= v[pos];
191
        }*/
192
//        sum += random()%10;
193
/*        for(j=0; j<LEN; j++){
194
            v[j] -= sum/LEN;
195
        }*/
196
//        lbt1(v+100,v+100,LEN);
197
        ff_pca_add(pca, v);
198
    }
199

    
200

    
201
    ff_pca(pca, eigenvector, eigenvalue);
202
    for(i=0; i<LEN; i++){
203
        pca->count= 1;
204
        pca->mean[i]= 0;
205

    
206
//        (0.5^|x|)^2 = 0.5^2|x| = 0.25^|x|
207

    
208

    
209
//        pca.covariance[i + i*LEN]= pow(0.5, fabs
210
        for(j=i; j<LEN; j++){
211
            printf("%f ", pca->covariance[i + j*LEN]);
212
        }
213
        printf("\n");
214
    }
215

    
216
#if 1
217
    for(i=0; i<LEN; i++){
218
        double v[LEN];
219
        double error=0;
220
        memset(v, 0, sizeof(v));
221
        for(j=0; j<LEN; j++){
222
            for(k=0; k<LEN; k++){
223
                v[j] += pca->covariance[FFMIN(k,j) + FFMAX(k,j)*LEN] * eigenvector[i + k*LEN];
224
            }
225
            v[j] /= eigenvalue[i];
226
            error += fabs(v[j] - eigenvector[i + j*LEN]);
227
        }
228
        printf("%f ", error);
229
    }
230
    printf("\n");
231
#endif
232
    for(i=0; i<LEN; i++){
233
        for(j=0; j<LEN; j++){
234
            printf("%9.6f ", eigenvector[i + j*LEN]);
235
        }
236
        printf("  %9.1f %f\n", eigenvalue[i], eigenvalue[i]/eigenvalue[0]);
237
    }
238

    
239
    return 0;
240
}
241
#endif