All Downloads are FREE. Search and download functionalities are using the official Maven repository.

org.machinelearning.svm.libsvm.svm Maven / Gradle / Ivy

There is a newer version: 2024.11.2
Show newest version





package org.machinelearning.svm.libsvm;
import java.io.*;
import java.util.*;

//
// Kernel Cache
//
// l is the number of total data items
// size is the cache size limit in bytes
//
class Cache {
	private final int l;
	private long size;
	private final class head_t
	{
		head_t prev, next;	// a cicular list
		float[] data;
		int len;		// data[0,len) is cached in this entry
	}
	private final head_t[] head;
	private head_t lru_head;

	Cache(int l_, long size_)
	{
		l = l_;
		size = size_;
		head = new head_t[l];
		for(int i=0;i= len if nothing needs to be filled)
	// java: simulate pointer using single-element array
	int get_data(int index, float[][] data, int len)
	{
		head_t h = head[index];
		if(h.len > 0) lru_delete(h);
		int more = len - h.len;

		if(more > 0)
		{
			// free old space
			while(size < more)
			{
				head_t old = lru_head.next;
				lru_delete(old);
				size += old.len;
				old.data = null;
				old.len = 0;
			}

			// allocate new space
			float[] new_data = new float[len];
			if(h.data != null) System.arraycopy(h.data,0,new_data,0,h.len);
			h.data = new_data;
			size -= more;
			do {int tmp=h.len; h.len=len; len=tmp;} while(false);
		}

		lru_insert(h);
		data[0] = h.data;
		return len;
	}

	void swap_index(int i, int j)
	{
		if(i==j) return;

		if(head[i].len > 0) lru_delete(head[i]);
		if(head[j].len > 0) lru_delete(head[j]);
		do {float[] tmp=head[i].data; head[i].data=head[j].data; head[j].data=tmp;} while(false);
		do {int tmp=head[i].len; head[i].len=head[j].len; head[j].len=tmp;} while(false);
		if(head[i].len > 0) lru_insert(head[i]);
		if(head[j].len > 0) lru_insert(head[j]);

		if(i>j) do {int tmp=i; i=j; j=tmp;} while(false);
		for(head_t h = lru_head.next; h!=lru_head; h=h.next)
		{
			if(h.len > i)
			{
				if(h.len > j)
					do {float tmp=h.data[i]; h.data[i]=h.data[j]; h.data[j]=tmp;} while(false);
				else
				{
					// give up
					lru_delete(h);
					size += h.len;
					h.data = null;
					h.len = 0;
				}
			}
		}
	}
}

//
// Kernel evaluation
//
// the static method k_function is for doing single kernel evaluation
// the constructor of Kernel prepares to calculate the l*l kernel matrix
// the member function get_Q is for getting one column from the Q Matrix
//
abstract class QMatrix {
	abstract float[] get_Q(int column, int len);
	abstract double[] get_QD();
	abstract void swap_index(int i, int j);
};

abstract class Kernel extends QMatrix {
	private svm_node[][] x;
	private final double[] x_square;

	// svm_parameter
	private final int kernel_type;
	private final int degree;
	private final double gamma;
	private final double coef0;

	abstract float[] get_Q(int column, int len);
	abstract double[] get_QD();

	void swap_index(int i, int j)
	{
		do {svm_node[] tmp=x[i]; x[i]=x[j]; x[j]=tmp;} while(false);
		if(x_square != null) do {double tmp=x_square[i]; x_square[i]=x_square[j]; x_square[j]=tmp;} while(false);
	}

	private static double powi(double base, int times)
	{
		double tmp = base, ret = 1.0;

		for(int t=times; t>0; t/=2)
		{
			if(t%2==1) ret*=tmp;
			tmp = tmp * tmp;
		}
		return ret;
	}

	double kernel_function(int i, int j)
	{
		switch(kernel_type)
		{
			case svm_parameter.LINEAR:
				return dot(x[i],x[j]);
			case svm_parameter.POLY:
				return powi(gamma*dot(x[i],x[j])+coef0,degree);
			case svm_parameter.RBF:
				return Math.exp(-gamma*(x_square[i]+x_square[j]-2*dot(x[i],x[j])));
			case svm_parameter.SIGMOID:
				return Math.tanh(gamma*dot(x[i],x[j])+coef0);
			case svm_parameter.PRECOMPUTED:
				return x[i][(int)(x[j][0].value)].value;
			default:
				return 0;	// java
		}
	}

	Kernel(int l, svm_node[][] x_, svm_parameter param)
	{
		this.kernel_type = param.kernel_type;
		this.degree = param.degree;
		this.gamma = param.gamma;
		this.coef0 = param.coef0;

		x = (svm_node[][])x_.clone();

		if(kernel_type == svm_parameter.RBF)
		{
			x_square = new double[l];
			for(int i=0;i y[j].index)
					++j;
				else
					++i;
			}
		}
		return sum;
	}

	static double k_function(svm_node[] x, svm_node[] y,
					svm_parameter param)
	{
		switch(param.kernel_type)
		{
			case svm_parameter.LINEAR:
				return dot(x,y);
			case svm_parameter.POLY:
				return powi(param.gamma*dot(x,y)+param.coef0,param.degree);
			case svm_parameter.RBF:
			{
				double sum = 0;
				int xlen = x.length;
				int ylen = y.length;
				int i = 0;
				int j = 0;
				while(i < xlen && j < ylen)
				{
					if(x[i].index == y[j].index)
					{
						double d = x[i++].value - y[j++].value;
						sum += d*d;
					}
					else if(x[i].index > y[j].index)
					{
						sum += y[j].value * y[j].value;
						++j;
					}
					else
					{
						sum += x[i].value * x[i].value;
						++i;
					}
				}

				while(i < xlen)
				{
					sum += x[i].value * x[i].value;
					++i;
				}

				while(j < ylen)
				{
					sum += y[j].value * y[j].value;
					++j;
				}

				return Math.exp(-param.gamma*sum);
			}
			case svm_parameter.SIGMOID:
				return Math.tanh(param.gamma*dot(x,y)+param.coef0);
			case svm_parameter.PRECOMPUTED:
				return	x[(int)(y[0].value)].value;
			default:
				return 0;	// java
		}
	}
}

// An SMO algorithm in Fan et al., JMLR 6(2005), p. 1889--1918
// Solves:
//
//	min 0.5(\alpha^T Q \alpha) + p^T \alpha
//
//		y^T \alpha = \delta
//		y_i = +1 or -1
//		0 <= alpha_i <= Cp for y_i = 1
//		0 <= alpha_i <= Cn for y_i = -1
//
// Given:
//
//	Q, p, y, Cp, Cn, and an initial feasible point \alpha
//	l is the size of vectors and matrices
//	eps is the stopping tolerance
//
// solution will be put in \alpha, objective value will be put in obj
//
class Solver {
	int active_size;
	byte[] y;
	double[] G;		// gradient of objective function
	static final byte LOWER_BOUND = 0;
	static final byte UPPER_BOUND = 1;
	static final byte FREE = 2;
	byte[] alpha_status;	// LOWER_BOUND, UPPER_BOUND, FREE
	double[] alpha;
	QMatrix Q;
	double[] QD;
	double eps;
	double Cp,Cn;
	double[] p;
	int[] active_set;
	double[] G_bar;		// gradient, if we treat free variables as 0
	int l;
	boolean unshrink;	// XXX

	static final double INF = Double.POSITIVE_INFINITY;

	double get_C(int i)
	{
		return (y[i] > 0)? Cp : Cn;
	}
	void update_alpha_status(int i)
	{
		if(alpha[i] >= get_C(i))
			alpha_status[i] = UPPER_BOUND;
		else if(alpha[i] <= 0)
			alpha_status[i] = LOWER_BOUND;
		else alpha_status[i] = FREE;
	}
	boolean is_upper_bound(int i) { return alpha_status[i] == UPPER_BOUND; }
	boolean is_lower_bound(int i) { return alpha_status[i] == LOWER_BOUND; }
	boolean is_free(int i) {  return alpha_status[i] == FREE; }

	// java: information about solution except alpha,
	// because we cannot return multiple values otherwise...
	static class SolutionInfo {
		double obj;
		double rho;
		double upper_bound_p;
		double upper_bound_n;
		double r;	// for Solver_NU
	}

	void swap_index(int i, int j)
	{
		Q.swap_index(i,j);
		do {byte tmp=y[i]; y[i]=y[j]; y[j]=tmp;} while(false);
		do {double tmp=G[i]; G[i]=G[j]; G[j]=tmp;} while(false);
		do {byte tmp=alpha_status[i]; alpha_status[i]=alpha_status[j]; alpha_status[j]=tmp;} while(false);
		do {double tmp=alpha[i]; alpha[i]=alpha[j]; alpha[j]=tmp;} while(false);
		do {double tmp=p[i]; p[i]=p[j]; p[j]=tmp;} while(false);
		do {int tmp=active_set[i]; active_set[i]=active_set[j]; active_set[j]=tmp;} while(false);
		do {double tmp=G_bar[i]; G_bar[i]=G_bar[j]; G_bar[j]=tmp;} while(false);
	}

	void reconstruct_gradient()
	{
		// reconstruct inactive elements of G from G_bar and free variables

		if(active_size == l) return;

		int i,j;
		int nr_free = 0;

		for(j=active_size;j 2*active_size*(l-active_size))
		{
			for(i=active_size;iInteger.MAX_VALUE/100 ? Integer.MAX_VALUE : 100*l);
		int counter = Math.min(l,1000)+1;
		int[] working_set = new int[2];

		while(iter < max_iter)
		{
			// show progress and do shrinking

			if(--counter == 0)
			{
				counter = Math.min(l,1000);
				if(shrinking!=0) do_shrinking();
				svm.info(".");
			}

			if(select_working_set(working_set)!=0)
			{
				// reconstruct the whole gradient
				reconstruct_gradient();
				// reset active set size and check
				active_size = l;
				svm.info("*");
				if(select_working_set(working_set)!=0)
					break;
				else
					counter = 1;	// do shrinking next iteration
			}

			int i = working_set[0];
			int j = working_set[1];

			++iter;

			// update alpha[i] and alpha[j], handle bounds carefully

			float[] Q_i = Q.get_Q(i,active_size);
			float[] Q_j = Q.get_Q(j,active_size);

			double C_i = get_C(i);
			double C_j = get_C(j);

			double old_alpha_i = alpha[i];
			double old_alpha_j = alpha[j];

			if(y[i]!=y[j])
			{
				double quad_coef = QD[i]+QD[j]+2*Q_i[j];
				if (quad_coef <= 0)
					quad_coef = 1e-12;
				double delta = (-G[i]-G[j])/quad_coef;
				double diff = alpha[i] - alpha[j];
				alpha[i] += delta;
				alpha[j] += delta;

				if(diff > 0)
				{
					if(alpha[j] < 0)
					{
						alpha[j] = 0;
						alpha[i] = diff;
					}
				}
				else
				{
					if(alpha[i] < 0)
					{
						alpha[i] = 0;
						alpha[j] = -diff;
					}
				}
				if(diff > C_i - C_j)
				{
					if(alpha[i] > C_i)
					{
						alpha[i] = C_i;
						alpha[j] = C_i - diff;
					}
				}
				else
				{
					if(alpha[j] > C_j)
					{
						alpha[j] = C_j;
						alpha[i] = C_j + diff;
					}
				}
			}
			else
			{
				double quad_coef = QD[i]+QD[j]-2*Q_i[j];
				if (quad_coef <= 0)
					quad_coef = 1e-12;
				double delta = (G[i]-G[j])/quad_coef;
				double sum = alpha[i] + alpha[j];
				alpha[i] -= delta;
				alpha[j] += delta;

				if(sum > C_i)
				{
					if(alpha[i] > C_i)
					{
						alpha[i] = C_i;
						alpha[j] = sum - C_i;
					}
				}
				else
				{
					if(alpha[j] < 0)
					{
						alpha[j] = 0;
						alpha[i] = sum;
					}
				}
				if(sum > C_j)
				{
					if(alpha[j] > C_j)
					{
						alpha[j] = C_j;
						alpha[i] = sum - C_j;
					}
				}
				else
				{
					if(alpha[i] < 0)
					{
						alpha[i] = 0;
						alpha[j] = sum;
					}
				}
			}

			// update G

			double delta_alpha_i = alpha[i] - old_alpha_i;
			double delta_alpha_j = alpha[j] - old_alpha_j;

			for(int k=0;k= max_iter)
		{
			if(active_size < l)
			{
				// reconstruct the whole gradient to calculate objective value
				reconstruct_gradient();
				active_size = l;
				svm.info("*");
			}
			System.err.print("\nWARNING: reaching max number of iterations\n");
			throw new RuntimeException("WARNING: reaching max number of iterations");
		}

		// calculate rho

		si.rho = calculate_rho();

		// calculate objective value
		{
			double v = 0;
			int i;
			for(i=0;i= Gmax)
					{
						Gmax = -G[t];
						Gmax_idx = t;
					}
			}
			else
			{
				if(!is_lower_bound(t))
					if(G[t] >= Gmax)
					{
						Gmax = G[t];
						Gmax_idx = t;
					}
			}

		int i = Gmax_idx;
		float[] Q_i = null;
		if(i != -1) // null Q_i not accessed: Gmax=-INF if i=-1
			Q_i = Q.get_Q(i,active_size);

		for(int j=0;j= Gmax2)
						Gmax2 = G[j];
					if (grad_diff > 0)
					{
						double obj_diff;
						double quad_coef = QD[i]+QD[j]-2.0*y[i]*Q_i[j];
						if (quad_coef > 0)
							obj_diff = -(grad_diff*grad_diff)/quad_coef;
						else
							obj_diff = -(grad_diff*grad_diff)/1e-12;

						if (obj_diff <= obj_diff_min)
						{
							Gmin_idx=j;
							obj_diff_min = obj_diff;
						}
					}
				}
			}
			else
			{
				if (!is_upper_bound(j))
				{
					double grad_diff= Gmax-G[j];
					if (-G[j] >= Gmax2)
						Gmax2 = -G[j];
					if (grad_diff > 0)
					{
						double obj_diff;
						double quad_coef = QD[i]+QD[j]+2.0*y[i]*Q_i[j];
						if (quad_coef > 0)
							obj_diff = -(grad_diff*grad_diff)/quad_coef;
						else
							obj_diff = -(grad_diff*grad_diff)/1e-12;

						if (obj_diff <= obj_diff_min)
						{
							Gmin_idx=j;
							obj_diff_min = obj_diff;
						}
					}
				}
			}
		}

		if(Gmax+Gmax2 < eps || Gmin_idx == -1)
			return 1;

		working_set[0] = Gmax_idx;
		working_set[1] = Gmin_idx;
		return 0;
	}

	private boolean be_shrunk(int i, double Gmax1, double Gmax2)
	{
		if(is_upper_bound(i))
		{
			if(y[i]==+1)
				return(-G[i] > Gmax1);
			else
				return(-G[i] > Gmax2);
		}
		else if(is_lower_bound(i))
		{
			if(y[i]==+1)
				return(G[i] > Gmax2);
			else
				return(G[i] > Gmax1);
		}
		else
			return(false);
	}

	void do_shrinking()
	{
		int i;
		double Gmax1 = -INF;		// max { -y_i * grad(f)_i | i in I_up(\alpha) }
		double Gmax2 = -INF;		// max { y_i * grad(f)_i | i in I_low(\alpha) }

		// find maximal violating pair first
		for(i=0;i= Gmax1)
						Gmax1 = -G[i];
				}
				if(!is_lower_bound(i))
				{
					if(G[i] >= Gmax2)
						Gmax2 = G[i];
				}
			}
			else
			{
				if(!is_upper_bound(i))
				{
					if(-G[i] >= Gmax2)
						Gmax2 = -G[i];
				}
				if(!is_lower_bound(i))
				{
					if(G[i] >= Gmax1)
						Gmax1 = G[i];
				}
			}
		}

		if(unshrink == false && Gmax1 + Gmax2 <= eps*10)
		{
			unshrink = true;
			reconstruct_gradient();
			active_size = l;
		}

		for(i=0;i i)
				{
					if (!be_shrunk(active_size, Gmax1, Gmax2))
					{
						swap_index(i,active_size);
						break;
					}
					active_size--;
				}
			}
	}

	double calculate_rho()
	{
		double r;
		int nr_free = 0;
		double ub = INF, lb = -INF, sum_free = 0;
		for(int i=0;i 0)
					ub = Math.min(ub,yG);
				else
					lb = Math.max(lb,yG);
			}
			else if(is_upper_bound(i))
			{
				if(y[i] < 0)
					ub = Math.min(ub,yG);
				else
					lb = Math.max(lb,yG);
			}
			else
			{
				++nr_free;
				sum_free += yG;
			}
		}

		if(nr_free>0)
			r = sum_free/nr_free;
		else
			r = (ub+lb)/2;

		return r;
	}

}

//
// Solver for nu-svm classification and regression
//
// additional constraint: e^T \alpha = constant
//
final class Solver_NU extends Solver
{
	private SolutionInfo si;

	void Solve(int l, QMatrix Q, double[] p, byte[] y,
		   double[] alpha, double Cp, double Cn, double eps,
		   SolutionInfo si, int shrinking)
	{
		this.si = si;
		super.Solve(l,Q,p,y,alpha,Cp,Cn,eps,si,shrinking);
	}

	// return 1 if already optimal, return 0 otherwise
	int select_working_set(int[] working_set)
	{
		// return i,j such that y_i = y_j and
		// i: maximizes -y_i * grad(f)_i, i in I_up(\alpha)
		// j: minimizes the decrease of obj value
		//    (if quadratic coefficeint <= 0, replace it with tau)
		//    -y_j*grad(f)_j < -y_i*grad(f)_i, j in I_low(\alpha)

		double Gmaxp = -INF;
		double Gmaxp2 = -INF;
		int Gmaxp_idx = -1;

		double Gmaxn = -INF;
		double Gmaxn2 = -INF;
		int Gmaxn_idx = -1;

		int Gmin_idx = -1;
		double obj_diff_min = INF;

		for(int t=0;t= Gmaxp)
					{
						Gmaxp = -G[t];
						Gmaxp_idx = t;
					}
			}
			else
			{
				if(!is_lower_bound(t))
					if(G[t] >= Gmaxn)
					{
						Gmaxn = G[t];
						Gmaxn_idx = t;
					}
			}

		int ip = Gmaxp_idx;
		int in = Gmaxn_idx;
		float[] Q_ip = null;
		float[] Q_in = null;
		if(ip != -1) // null Q_ip not accessed: Gmaxp=-INF if ip=-1
			Q_ip = Q.get_Q(ip,active_size);
		if(in != -1)
			Q_in = Q.get_Q(in,active_size);

		for(int j=0;j= Gmaxp2)
						Gmaxp2 = G[j];
					if (grad_diff > 0)
					{
						double obj_diff;
						double quad_coef = QD[ip]+QD[j]-2*Q_ip[j];
						if (quad_coef > 0)
							obj_diff = -(grad_diff*grad_diff)/quad_coef;
						else
							obj_diff = -(grad_diff*grad_diff)/1e-12;

						if (obj_diff <= obj_diff_min)
						{
							Gmin_idx=j;
							obj_diff_min = obj_diff;
						}
					}
				}
			}
			else
			{
				if (!is_upper_bound(j))
				{
					double grad_diff=Gmaxn-G[j];
					if (-G[j] >= Gmaxn2)
						Gmaxn2 = -G[j];
					if (grad_diff > 0)
					{
						double obj_diff;
						double quad_coef = QD[in]+QD[j]-2*Q_in[j];
						if (quad_coef > 0)
							obj_diff = -(grad_diff*grad_diff)/quad_coef;
						else
							obj_diff = -(grad_diff*grad_diff)/1e-12;

						if (obj_diff <= obj_diff_min)
						{
							Gmin_idx=j;
							obj_diff_min = obj_diff;
						}
					}
				}
			}
		}

		if(Math.max(Gmaxp+Gmaxp2,Gmaxn+Gmaxn2) < eps || Gmin_idx == -1)
			return 1;

		if(y[Gmin_idx] == +1)
			working_set[0] = Gmaxp_idx;
		else
			working_set[0] = Gmaxn_idx;
		working_set[1] = Gmin_idx;

		return 0;
	}

	private boolean be_shrunk(int i, double Gmax1, double Gmax2, double Gmax3, double Gmax4)
	{
		if(is_upper_bound(i))
		{
			if(y[i]==+1)
				return(-G[i] > Gmax1);
			else
				return(-G[i] > Gmax4);
		}
		else if(is_lower_bound(i))
		{
			if(y[i]==+1)
				return(G[i] > Gmax2);
			else
				return(G[i] > Gmax3);
		}
		else
			return(false);
	}

	void do_shrinking()
	{
		double Gmax1 = -INF;	// max { -y_i * grad(f)_i | y_i = +1, i in I_up(\alpha) }
		double Gmax2 = -INF;	// max { y_i * grad(f)_i | y_i = +1, i in I_low(\alpha) }
		double Gmax3 = -INF;	// max { -y_i * grad(f)_i | y_i = -1, i in I_up(\alpha) }
		double Gmax4 = -INF;	// max { y_i * grad(f)_i | y_i = -1, i in I_low(\alpha) }

		// find maximal violating pair first
		int i;
		for(i=0;i Gmax1) Gmax1 = -G[i];
				}
				else	if(-G[i] > Gmax4) Gmax4 = -G[i];
			}
			if(!is_lower_bound(i))
			{
				if(y[i]==+1)
				{
					if(G[i] > Gmax2) Gmax2 = G[i];
				}
				else	if(G[i] > Gmax3) Gmax3 = G[i];
			}
		}

		if(unshrink == false && Math.max(Gmax1+Gmax2,Gmax3+Gmax4) <= eps*10)
		{
			unshrink = true;
			reconstruct_gradient();
			active_size = l;
		}

		for(i=0;i i)
				{
					if (!be_shrunk(active_size, Gmax1, Gmax2, Gmax3, Gmax4))
					{
						swap_index(i,active_size);
						break;
					}
					active_size--;
				}
			}
	}

	double calculate_rho()
	{
		int nr_free1 = 0,nr_free2 = 0;
		double ub1 = INF, ub2 = INF;
		double lb1 = -INF, lb2 = -INF;
		double sum_free1 = 0, sum_free2 = 0;

		for(int i=0;i 0)
			r1 = sum_free1/nr_free1;
		else
			r1 = (ub1+lb1)/2;

		if(nr_free2 > 0)
			r2 = sum_free2/nr_free2;
		else
			r2 = (ub2+lb2)/2;

		si.r = (r1+r2)/2;
		return (r1-r2)/2;
	}
}

//
// Q matrices for various formulations
//
class SVC_Q extends Kernel
{
	private final byte[] y;
	private final Cache cache;
	private final double[] QD;

	SVC_Q(svm_problem prob, svm_parameter param, byte[] y_)
	{
		super(prob.l, prob.x, param);
		y = (byte[])y_.clone();
		cache = new Cache(prob.l,(long)(param.cache_size*(1<<20)));
		QD = new double[prob.l];
		for(int i=0;i 0) y[i] = +1; else y[i] = -1;
		}

		Solver s = new Solver();
		s.Solve(l, new SVC_Q(prob,param,y), minus_ones, y,
			alpha, Cp, Cn, param.eps, si, param.shrinking);

		double sum_alpha=0;
		for(i=0;i0)
				y[i] = +1;
			else
				y[i] = -1;

		double sum_pos = nu*l/2;
		double sum_neg = nu*l/2;

		for(i=0;i 0)
			{
				++nSV;
				if(prob.y[i] > 0)
				{
					if(Math.abs(alpha[i]) >= si.upper_bound_p)
					++nBSV;
				}
				else
				{
					if(Math.abs(alpha[i]) >= si.upper_bound_n)
						++nBSV;
				}
			}
		}

		svm.info("nSV = "+nSV+", nBSV = "+nBSV+"\n");

		decision_function f = new decision_function();
		f.alpha = alpha;
		f.rho = si.rho;
		return f;
	}

	// Platt's binary SVM Probablistic Output: an improvement from Lin et al.
	private static void sigmoid_train(int l, double[] dec_values, double[] labels,
				  double[] probAB)
	{
		double A, B;
		double prior1=0, prior0 = 0;
		int i;

		for (i=0;i 0) prior1+=1;
			else prior0+=1;

		int max_iter=100;	// Maximal number of iterations
		double min_step=1e-10;	// Minimal step taken in line search
		double sigma=1e-12;	// For numerically strict PD of Hessian
		double eps=1e-5;
		double hiTarget=(prior1+1.0)/(prior1+2.0);
		double loTarget=1/(prior0+2.0);
		double[] t= new double[l];
		double fApB,p,q,h11,h22,h21,g1,g2,det,dA,dB,gd,stepsize;
		double newA,newB,newf,d1,d2;
		int iter;

		// Initial Point and Initial Fun Value
		A=0.0; B=Math.log((prior0+1.0)/(prior1+1.0));
		double fval = 0.0;

		for (i=0;i0) t[i]=hiTarget;
			else t[i]=loTarget;
			fApB = dec_values[i]*A+B;
			if (fApB>=0)
				fval += t[i]*fApB + Math.log(1+Math.exp(-fApB));
			else
				fval += (t[i] - 1)*fApB +Math.log(1+Math.exp(fApB));
		}
		for (iter=0;iter= 0)
				{
					p=Math.exp(-fApB)/(1.0+Math.exp(-fApB));
					q=1.0/(1.0+Math.exp(-fApB));
				}
				else
				{
					p=1.0/(1.0+Math.exp(fApB));
					q=Math.exp(fApB)/(1.0+Math.exp(fApB));
				}
				d2=p*q;
				h11+=dec_values[i]*dec_values[i]*d2;
				h22+=d2;
				h21+=dec_values[i]*d2;
				d1=t[i]-p;
				g1+=dec_values[i]*d1;
				g2+=d1;
			}

			// Stopping Criteria
			if (Math.abs(g1)= min_step)
			{
				newA = A + stepsize * dA;
				newB = B + stepsize * dB;

				// New function value
				newf = 0.0;
				for (i=0;i= 0)
						newf += t[i]*fApB + Math.log(1+Math.exp(-fApB));
					else
						newf += (t[i] - 1)*fApB +Math.log(1+Math.exp(fApB));
				}
				// Check sufficient decrease
				if (newf=max_iter)
			svm.info("Reaching maximal iterations in two-class probability estimates\n");
		probAB[0]=A;probAB[1]=B;
	}

	private static double sigmoid_predict(double decision_value, double A, double B)
	{
		double fApB = decision_value*A+B;
		if (fApB >= 0)
			return Math.exp(-fApB)/(1.0+Math.exp(-fApB));
		else
			return 1.0/(1+Math.exp(fApB)) ;
	}

	// Method 2 from the multiclass_prob paper by Wu, Lin, and Weng
	private static void multiclass_probability(int k, double[][] r, double[] p)
	{
		int t,j;
		int iter = 0, max_iter=Math.max(100,k);
		double[][] Q=new double[k][k];
		double[] Qp=new double[k];
		double pQp, eps=0.005/k;

		for (t=0;tmax_error)
					max_error=error;
			}
			if (max_error=max_iter)
			svm.info("Exceeds max_iter in multiclass_prob\n");
	}

	// Cross-validation decision values for probability estimates
	private static void svm_binary_svc_probability(svm_problem prob, svm_parameter param, double Cp, double Cn, double[] probAB)
	{
		int i;
		int nr_fold = 5;
		int[] perm = new int[prob.l];
		double[] dec_values = new double[prob.l];

		// random shuffle
		for(i=0;i0)
					p_count++;
				else
					n_count++;

			if(p_count==0 && n_count==0)
				for(j=begin;j 0 && n_count == 0)
				for(j=begin;j 0)
				for(j=begin;j 5*std)
				count=count+1;
			else
				mae+=Math.abs(ymv[i]);
		mae /= (prob.l-count);
		svm.info("Prob. model for test data: target value = predicted value + z,\nz: Laplace distribution e^(-|z|/sigma)/(2sigma),sigma="+mae+"\n");
		return mae;
	}

	// label: label name, start: begin of each class, count: #data of classes, perm: indices to the original data
	// perm, length l, must be allocated before calling this subroutine
	private static void svm_group_classes(svm_problem prob, int[] nr_class_ret, int[][] label_ret, int[][] start_ret, int[][] count_ret, int[] perm)
	{
		int l = prob.l;
		int max_nr_class = 16;
		int nr_class = 0;
		int[] label = new int[max_nr_class];
		int[] count = new int[max_nr_class];
		int[] data_label = new int[l];
		int i;

		for(i=0;i 0) ++nSV;
			model.l = nSV;
			model.SV = new svm_node[nSV][];
			model.sv_coef[0] = new double[nSV];
			model.sv_indices = new int[nSV];
			int j = 0;
			for(i=0;i 0)
				{
					model.SV[j] = prob.x[i];
					model.sv_coef[0][j] = f.alpha[i];
					model.sv_indices[j] = i+1;
					++j;
				}
		}
		else
		{
			// classification
			int l = prob.l;
			int[] tmp_nr_class = new int[1];
			int[][] tmp_label = new int[1][];
			int[][] tmp_start = new int[1][];
			int[][] tmp_count = new int[1][];
			int[] perm = new int[l];

			// group training data of the same class
			svm_group_classes(prob,tmp_nr_class,tmp_label,tmp_start,tmp_count,perm);
			int nr_class = tmp_nr_class[0];
			int[] label = tmp_label[0];
			int[] start = tmp_start[0];
			int[] count = tmp_count[0];

			if(nr_class == 1)
				svm.info("WARNING: training data in only one class. See README for details.\n");

			svm_node[][] x = new svm_node[l][];
			int i;
			for(i=0;i 0)
							nonzero[si+k] = true;
					for(k=0;k 0)
							nonzero[sj+k] = true;
					++p;
				}

			// build output

			model.nr_class = nr_class;

			model.label = new int[nr_class];
			for(i=0;i some folds may have zero elements
		if((param.svm_type == svm_parameter.C_SVC ||
		    param.svm_type == svm_parameter.NU_SVC) && nr_fold < l)
		{
			int[] tmp_nr_class = new int[1];
			int[][] tmp_label = new int[1][];
			int[][] tmp_start = new int[1][];
			int[][] tmp_count = new int[1][];

			svm_group_classes(prob,tmp_nr_class,tmp_label,tmp_start,tmp_count,perm);

			int nr_class = tmp_nr_class[0];
			int[] start = tmp_start[0];
			int[] count = tmp_count[0];

			// random shuffle and then data grouped by fold using the array perm
			int[] fold_count = new int[nr_fold];
			int c;
			int[] index = new int[l];
			for(i=0;i0)?1:-1;
			else
				return sum;
		}
		else
		{
			int nr_class = model.nr_class;
			int l = model.l;

			double[] kvalue = new double[l];
			for(i=0;i 0)
						++vote[i];
					else
						++vote[j];
					p++;
				}

			int vote_max_idx = 0;
			for(i=1;i vote[vote_max_idx])
					vote_max_idx = i;

			return model.label[vote_max_idx];
		}
	}

	public static double svm_predict(svm_model model, svm_node[] x)
	{
		int nr_class = model.nr_class;
		double[] dec_values;
		if(model.param.svm_type == svm_parameter.ONE_CLASS ||
				model.param.svm_type == svm_parameter.EPSILON_SVR ||
				model.param.svm_type == svm_parameter.NU_SVR)
			dec_values = new double[1];
		else
			dec_values = new double[nr_class*(nr_class-1)/2];
		double pred_result = svm_predict_values(model, x, dec_values);
		return pred_result;
	}

	public static double svm_predict_probability(svm_model model, svm_node[] x, double[] prob_estimates)
	{
		if ((model.param.svm_type == svm_parameter.C_SVC || model.param.svm_type == svm_parameter.NU_SVC) &&
		    model.probA!=null && model.probB!=null)
		{
			int i;
			int nr_class = model.nr_class;
			double[] dec_values = new double[nr_class*(nr_class-1)/2];
			svm_predict_values(model, x, dec_values);

			double min_prob=1e-7;
			double[][] pairwise_prob=new double[nr_class][nr_class];

			int k=0;
			for(i=0;i prob_estimates[prob_max_idx])
					prob_max_idx = i;
			return model.label[prob_max_idx];
		}
		else
			return svm_predict(model, x);
	}

	static final String svm_type_table[] =
	{
		"c_svc","nu_svc","one_class","epsilon_svr","nu_svr",
	};

	static final String kernel_type_table[]=
	{
		"linear","polynomial","rbf","sigmoid","precomputed"
	};

	public static void svm_save_model(String model_file_name, svm_model model) throws IOException
	{
		DataOutputStream fp = new DataOutputStream(new BufferedOutputStream(new FileOutputStream(model_file_name)));

		svm_parameter param = model.param;

		fp.writeBytes("svm_type "+svm_type_table[param.svm_type]+"\n");
		fp.writeBytes("kernel_type "+kernel_type_table[param.kernel_type]+"\n");

		if(param.kernel_type == svm_parameter.POLY)
			fp.writeBytes("degree "+param.degree+"\n");

		if(param.kernel_type == svm_parameter.POLY ||
		   param.kernel_type == svm_parameter.RBF ||
		   param.kernel_type == svm_parameter.SIGMOID)
			fp.writeBytes("gamma "+param.gamma+"\n");

		if(param.kernel_type == svm_parameter.POLY ||
		   param.kernel_type == svm_parameter.SIGMOID)
			fp.writeBytes("coef0 "+param.coef0+"\n");

		int nr_class = model.nr_class;
		int l = model.l;
		fp.writeBytes("nr_class "+nr_class+"\n");
		fp.writeBytes("total_sv "+l+"\n");

		{
			fp.writeBytes("rho");
			for(int i=0;i 1)
				return "nu <= 0 or nu > 1";

		if(svm_type == svm_parameter.EPSILON_SVR)
			if(param.p < 0)
				return "p < 0";

		if(param.shrinking != 0 &&
		   param.shrinking != 1)
			return "shrinking != 0 and shrinking != 1";

		if(param.probability != 0 &&
		   param.probability != 1)
			return "probability != 0 and probability != 1";

		if(param.probability == 1 &&
		   svm_type == svm_parameter.ONE_CLASS)
			return "one-class SVM probability output not supported yet";

		// check whether nu-svc is feasible

		if(svm_type == svm_parameter.NU_SVC)
		{
			int l = prob.l;
			int max_nr_class = 16;
			int nr_class = 0;
			int[] label = new int[max_nr_class];
			int[] count = new int[max_nr_class];

			int i;
			for(i=0;i Math.min(n1,n2))
						return "specified nu is infeasible";
				}
			}
		}

		return null;
	}

	public static int svm_check_probability_model(svm_model model)
	{
		if (((model.param.svm_type == svm_parameter.C_SVC || model.param.svm_type == svm_parameter.NU_SVC) &&
		model.probA!=null && model.probB!=null) ||
		((model.param.svm_type == svm_parameter.EPSILON_SVR || model.param.svm_type == svm_parameter.NU_SVR) &&
		 model.probA!=null))
			return 1;
		else
			return 0;
	}

	public static void svm_set_print_string_function(svm_print_interface print_func)
	{
		if (print_func == null)
			svm_print_string = svm_print_stdout;
		else
			svm_print_string = print_func;
	}
}




© 2015 - 2024 Weber Informatics LLC | Privacy Policy