Created
May 26, 2013 03:51
-
-
Save spaghetti-source/5651656 to your computer and use it in GitHub Desktop.
Support Vector Machine (gradient descent with log-sum-exp approximation)
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
// | |
// Support Vector Machine | |
// (gradient descent with log-sum-exp approximation) | |
// | |
// original: | |
// G(w) := min_{+,-} { <w, x_+> - <w,x_-> } --> maximize | |
// approx: | |
// G'(w) := -log sum_{+,-} exp (-<w, x_+> - <w,x_->) --> maximize | |
// <=> | |
// L(w) := sum_{+,-} exp -<w, x_+ - x_-> --> minimize | |
// | |
// algoritihm: | |
// w = w - dL(w) | |
// where | |
// dL(w) = sum_{+,-} (x_+ - x_-) exp -<w, x_+ - x_-> | |
// | |
#include <iostream> | |
#include <cstdio> | |
#include <cstdlib> | |
#include <cmath> | |
#include <map> | |
#include <vector> | |
#include <cstring> | |
#include <functional> | |
#include <algorithm> | |
using namespace std; | |
#define ALL(c) c.begin(), c.end() | |
#define FOR(i,c) for(typeof(c.begin())i=c.begin();i!=c.end();++i) | |
#define REP(i,n) for(int i=0;i<n;++i) | |
#define fst first | |
#define snd second | |
double randn() { | |
double r = 0; | |
for (int i = 0; i < 12; ++i) | |
r += 1.0 * rand() / RAND_MAX; | |
return r - 6; | |
} | |
int DIM = 100; | |
int SIZE = 200; | |
typedef vector<double> Vector; | |
int main() { | |
// dataset | |
Vector v(DIM); | |
REP(i, DIM) v[i] = randn(); | |
vector< Vector > xp, xn; | |
REP(i, SIZE) { | |
Vector x(DIM); | |
REP(i, DIM) x[i] = randn(); | |
double dot = 0; | |
REP(i, DIM) dot += x[i] * v[i]; | |
if (dot > 0) xp.push_back( x ); | |
else if (dot <-0) xn.push_back( x ); | |
} | |
// learning | |
Vector w(DIM); | |
REP(i, DIM) w[i] = randn(); | |
REP(epoch, 100) { | |
Vector dw(DIM); | |
REP(p, xp.size()) REP(q, xn.size()) { | |
double dot = 0; | |
REP(i, DIM) dot += (xp[p][i] - xn[q][i]) * w[i]; | |
REP(i, DIM) dw[i] -= (xp[p][i] - xn[q][i]) * exp(-dot); | |
} | |
double norm = 0; | |
double eps = 0.001/SIZE; | |
REP(i, DIM) { | |
w[i] -= eps * dw[i]; | |
norm = norm + w[i] * w[i]; | |
} | |
norm = sqrt(norm); | |
REP(i, DIM) { | |
w[i] /= norm; | |
} | |
// verify | |
int miss = 0; | |
REP(p, xp.size()) { | |
double dot = 0; | |
REP(i, DIM) dot += xp[p][i] * w[i]; | |
if (dot < 0) ++miss; | |
} | |
REP(p, xn.size()) { | |
double dot = 0; | |
REP(i, DIM) dot += xn[p][i] * w[i]; | |
if (dot > 0) ++miss; | |
} | |
printf(" miss: %5d\n", miss); | |
} | |
} |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment