libsvm代码阅读：关于Solver类分析（二）

update：2014-2-27 LinJM @HQU  『 libsvm专栏地址：http://blog.csdn.net/column/details/libsvm.html 』

如果你看完了上篇博文的伪代码，那么我们就可以开始谈谈它的源代码了。

下面先贴出它的类定义，一些成员函数的具体实现先忽略。

// 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 {
public:
	Solver() {};
	virtual ~Solver() {};//用虚析构函数的原因是：保证根据实际运行适当的析构函数

	struct SolutionInfo {
		double obj;
		double rho;
		double upper_bound_p;
		double upper_bound_n;
		double r;	// for Solver_NU
	};

	void Solve(int l, const QMatrix& Q, const double *p_, const schar *y_,
		   double *alpha_, double Cp, double Cn, double eps,
		   SolutionInfo* si, int shrinking);
protected:
	int active_size;//计算时实际参加运算的样本数目，经过shrink处理后，该数目小于全部样本数
	schar *y;	    //样本所属类别，该值只能取-1或+1。
	double *G;		// gradient of objective function = （Q alpha + p）
	enum { LOWER_BOUND, UPPER_BOUND, FREE };
	char *alpha_status;	// LOWER_BOUND, UPPER_BOUND, FREE 
	double *alpha;		//
	const QMatrix *Q;	
	const double *QD;
	double eps;		//误差限
	double Cp,Cn;
	double *p;
	int *active_set;
	double *G_bar;		// gradient, if we treat free variables as 0
	int l;
	bool unshrink;	// XXX
	//返回对应于样本的C。设置不同的Cp和Cn是为了处理数据的不平衡
	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;
	}
	bool is_upper_bound(int i) { return alpha_status[i] == UPPER_BOUND; }
	bool is_lower_bound(int i) { return alpha_status[i] == LOWER_BOUND; }
	bool is_free(int i) { return alpha_status[i] == FREE; }
	void swap_index(int i, int j);//交换样本i和j的内容，包括申请的内存的地址
	void reconstruct_gradient();  //重新计算梯度。
	virtual int select_working_set(int &i, int &j);//选择工作集
	virtual double calculate_rho();
	virtual void do_shrinking();//对样本集做缩减。
private:
	bool be_shrunk(int i, double Gmax1, double Gmax2);	
};

下面我们来看看SMO如何选择工作集(working set B)，选择的约束如下：

	// return i,j such that
	// 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)

论文中的公式如下：

int Solver::select_working_set(int &out_i, int &out_j)
{
	// return i,j such that
	// 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)
//select i	
	double Gmax = -INF;
	double Gmax2 = -INF;
	int Gmax_idx = -1;
	int Gmin_idx = -1;
	double obj_diff_min = INF;

	for(int t=0;t<active_size;t++)
		if(y[t]==+1)	//若类别为1
		{
			if(!is_upper_bound(t))//若alpha<C
				if(-G[t] >= Gmax)
				{
					Gmax = -G[t];// -y[t]*G[t]=-1*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;
	const Qfloat *Q_i = NULL;
	if(i != -1) // NULL Q_i not accessed: Gmax=-INF if i=-1
		Q_i = Q->get_Q(i,active_size);
//select j
	for(int j=0;j<active_size;j++)
	{
		if(y[j]==+1)
		{
			if (!is_lower_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)/TAU;

					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)/TAU;

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

	if(Gmax+Gmax2 < eps)
		return 1;

	out_i = Gmax_idx;
	out_j = Gmin_idx;
	return 0;
}

配合上面几个公式看，这段代码还是很清晰了。

下面来看看它的构造函数，这个构造函数是solver类的核心。这个算法也结合上一篇博文的algorithm2来看。其中要注意的是get_Q是获取核函数。

void Solver::Solve(int l, const QMatrix& Q, const double *p_, const schar *y_,
		   double *alpha_, double Cp, double Cn, double eps,
		   SolutionInfo* si, int shrinking)
{
	this->l = l;
	this->Q = &Q;
	QD=Q.get_QD();//这个是获取核函数（如果分类的话在SVC_Q中定义）

	clone(p, p_,l);
	clone(y, y_,l);
	clone(alpha,alpha_,l);

	this->Cp = Cp;
	this->Cn = Cn;
	this->eps = eps;
	unshrink = false;

	// initialize alpha_status
	{
		alpha_status = new char[l];
		for(int i=0;i<l;i++)
			update_alpha_status(i);
	}

	// initialize active set (for shrinking)
	{
		active_set = new int[l];
		for(int i=0;i<l;i++)
			active_set[i] = i;
		active_size = l;
	}

	// initialize gradient
	{
		G = new double[l];
		G_bar = new double[l];
		int i;
		for(i=0;i<l;i++)
		{
			G[i] = p[i];
			G_bar[i] = 0;
		}
		for(i=0;i<l;i++)
			if(!is_lower_bound(i))
			{
				const Qfloat *Q_i = Q.get_Q(i,l);
				double alpha_i = alpha[i];
				int j;
				for(j=0;j<l;j++)
					G[j] += alpha_i*Q_i[j];
				if(is_upper_bound(i))
					for(j=0;j<l;j++)
						G_bar[j] += get_C(i) * Q_i[j]; //这里见文献LIBSVM: A Library for SVM公式（33）
			}
	}

	// optimization step

	int iter = 0;
	int max_iter = max(10000000, l>INT_MAX/100 ? INT_MAX : 100*l);
	int counter = min(l,1000)+1;
	
	while(iter < max_iter)
	{
		// show progress and do shrinking

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

		int i,j;
		if(select_working_set(i,j)!=0)
		{
			// reconstruct the whole gradient
			reconstruct_gradient();
			// reset active set size and check
			active_size = l;
			info("*");
			if(select_working_set(i,j)!=0)
				break;
			else
				counter = 1;	// do shrinking next iteration
		}
		
		++iter;

		// update alpha[i] and alpha[j], handle bounds carefully
		
		const Qfloat *Q_i = Q.get_Q(i,active_size);
		const Qfloat *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 = TAU;
			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 = TAU;
			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<active_size;k++)
		{
			G[k] += Q_i[k]*delta_alpha_i + Q_j[k]*delta_alpha_j;
		}

		// update alpha_status and G_bar

		{
			bool ui = is_upper_bound(i);
			bool uj = is_upper_bound(j);
			update_alpha_status(i);
			update_alpha_status(j);
			int k;
			if(ui != is_upper_bound(i))
			{
				Q_i = Q.get_Q(i,l);
				if(ui)
					for(k=0;k<l;k++)
						G_bar[k] -= C_i * Q_i[k];
				else
					for(k=0;k<l;k++)
						G_bar[k] += C_i * Q_i[k];
			}

			if(uj != is_upper_bound(j))
			{
				Q_j = Q.get_Q(j,l);
				if(uj)
					for(k=0;k<l;k++)
						G_bar[k] -= C_j * Q_j[k];
				else
					for(k=0;k<l;k++)
						G_bar[k] += C_j * Q_j[k];
			}
		}
	}

	if(iter >= max_iter)
	{
		if(active_size < l)
		{
			// reconstruct the whole gradient to calculate objective value
			reconstruct_gradient();
			active_size = l;
			info("*");
		}
		fprintf(stderr,"\nWARNING: reaching max number of iterations\n");
	}

	// calculate rho

	si->rho = calculate_rho();

	// calculate objective value
	{
		double v = 0;
		int i;
		for(i=0;i<l;i++)
			v += alpha[i] * (G[i] + p[i]);

		si->obj = v/2;
	}

	// put back the solution
	{
		for(int i=0;i<l;i++)
			alpha_[active_set[i]] = alpha[i];
	}

	// juggle everything back
	/*{
		for(int i=0;i<l;i++)
			while(active_set[i] != i)
				swap_index(i,active_set[i]);
				// or Q.swap_index(i,active_set[i]);
	}*/

	si->upper_bound_p = Cp;
	si->upper_bound_n = Cn;

	info("\noptimization finished, #iter = %d\n",iter);

	delete[] p;
	delete[] y;
	delete[] alpha;
	delete[] alpha_status;
	delete[] active_set;
	delete[] G;
	delete[] G_bar;
}

本文地址：http://blog.csdn.net/linj_m/article/details/19705911

微博：林建民-机器视觉