Fixed a bunch of bugs in multitask logistic regression

shogun-toolbox · Jun 26, 2012 · 3ee917a · 3ee917a
1 parent a937587
commit 3ee917a
Show file tree

Hide file tree

Showing 2 changed files with 41 additions and 17 deletions.
diff --git a/src/shogun/lib/slep/slep_mt_lr.cpp b/src/shogun/lib/slep/slep_mt_lr.cpp
@@ -28,7 +28,8 @@ slep_result_t slep_mt_lr(
 	double lambda, lambda_max, beta;
 	double funcp = 0.0, func = 0.0;
 
-	int n_tasks = options.n_nodes;
+	int n_tasks = options.n_tasks;
+	//SG_SPRINT("N tasks = %d \n", n_tasks);
 
 	int iter = 1;
 	bool done = false;
@@ -64,9 +65,9 @@ slep_result_t slep_mt_lr(
 			for (i=task_ind_start; i<task_ind_end; i++)
 			{
 				if (y[i]>0)
-					b = m1[t]/(m1[t]+m2[t]);
+					b = double(m1[t])/(m1[t]+m2[t]);
 				else
-					b = -m2[t]/(m1[t]+m2[t]);
+					b = -double(m2[t])/(m1[t]+m2[t]);
 
 				features->add_to_dense_vec(b,i,ATb+t*n_feats,n_feats);
 			}
@@ -76,21 +77,23 @@ slep_result_t slep_mt_lr(
 		double q_bar = 0.0;
 		if (options.q==1)
 			q_bar = CMath::ALMOST_INFTY;
-		else if (options.q>1e-6)
+		if (options.q>1e-6)
 			q_bar = 1;
-		else
-			q_bar = options.q/(options.q-1);
+
+		q_bar = options.q/(options.q-1);
 		lambda_max = 0.0;
 
-		for (t=0; t<n_tasks; t++)
+		for (i=0; i<n_feats; i++)
 		{
+			double sum = 0.0;
+			for (t=0; t<n_tasks; t++)
+				sum += CMath::pow(fabs(ATb[t*n_feats+i]),q_bar);
 			lambda_max = 
-				CMath::max(lambda_max, 
-						SGVector<float64_t>::qnorm(ATb+t*n_feats, n_feats, q_bar));
+				CMath::max(lambda_max, CMath::pow(sum,1.0/q_bar)); 
 		}
 
 		lambda_max /= n_vecs;
-
+		
 		lambda = z*lambda_max;
 	}
 	else 
@@ -105,8 +108,7 @@ slep_result_t slep_mt_lr(
 				w(i,j) = options.initial_w[j*n_feats+i];
 	}
 	SGVector<double> c(n_tasks);
-	for (t=0; t<n_tasks; t++)
-		c[t] = CMath::log(m1[t]/m2[t]);
+	c.zero();
 
 	double* s = SG_CALLOC(double, n_feats*n_tasks);
 	double* sc = SG_CALLOC(double, n_tasks);
@@ -121,6 +123,7 @@ slep_result_t slep_mt_lr(
 		for (i=task_ind_start; i<task_ind_end; i++)
 			Aw[i] = features->dense_dot(i,w.matrix+t*n_feats,n_feats);
 	}
+
 	double* Av = SG_MALLOC(double, n_vecs);
 	double* As = SG_MALLOC(double, n_vecs);
 
@@ -149,9 +152,10 @@ slep_result_t slep_mt_lr(
 
 	double fun_x = 0.0;
 
-	while (!done && iter < options.max_iter) 
+	while (!done && iter <= options.max_iter) 
 	{
 		beta = (alphap-1.0)/alpha;
+		//SG_SPRINT("beta = %f \n", beta);
 
 		for (i=0; i<n_feats*n_tasks; i++)
 			s[i] = w[i] + beta*wwp[i];
@@ -161,7 +165,11 @@ slep_result_t slep_mt_lr(
 		for (i=0; i<n_vecs; i++)
 			As[i] = Aw[i] + beta*(Aw[i]-Awp[i]);
 
+		//SG_SPRINT("As = %f\n",SGVector<float64_t>::dot(As,As,n_vecs));
+
 		double fun_s = 0.0;
+		for (i=0; i<n_tasks*n_feats; i++)
+			g[i] = 0.0;
 		for (t=0; t<n_tasks; t++)
 		{
 			int task_ind_start = options.ind[t];
@@ -183,6 +191,10 @@ slep_result_t slep_mt_lr(
 			}
 		}
 		fun_s /= n_vecs;
+
+		//SG_SPRINT("fun_s = %f\n", fun_s);
+
+		//SG_SPRINT("g = %f\n", SGVector<float64_t>::dot(g,g,n_feats*n_tasks));
 
 		for (i=0; i<n_feats*n_tasks; i++)
 			wp[i] = w[i];
@@ -193,17 +205,23 @@ slep_result_t slep_mt_lr(
 		for (i=0; i<n_vecs; i++)
 			Awp[i] = Aw[i];
 
-		while (true)
+		int inner_iter = 1;
+		while (inner_iter < 100)
 		{
 			// v = s - g / L
 			for (i=0; i<n_feats*n_tasks; i++)
 				v[i] = s[i] - g[i]*(1.0/L);
 
 			for (t=0; t<n_tasks; t++)
 				c[t] = sc[t] - gc[t]*(1.0/L);
-
+			
 			eppMatrix(w.matrix, v, n_feats, n_tasks, lambda/L, options.q);
 
+			//SG_SPRINT("params [%d,%d,%f,%f]\n", n_feats, n_tasks, lambda/L, options.q);
+
+			//w.display_matrix();
+			//SG_SPRINT("w = %f \n", SGVector<float64_t>::dot(w.matrix,w.matrix,n_feats*n_tasks));
+
 			// v = x - s
 			for (i=0; i<n_feats*n_tasks; i++)
 				v[i] = w[i] - s[i];
@@ -224,6 +242,9 @@ slep_result_t slep_mt_lr(
 			}
 			fun_x /= n_vecs;
 
+			//SG_SPRINT("Aw = %f\n", SGVector<float64_t>::dot(Aw,Aw,n_vecs));
+			//c.display_vector();
+
 			double r_sum = SGVector<float64_t>::dot(v,v,n_feats*n_tasks);
 			double l_sum = fun_x - fun_s - SGVector<float64_t>::dot(v,g,n_feats*n_tasks);
 
@@ -234,6 +255,8 @@ slep_result_t slep_mt_lr(
 			}
 			r_sum /= 2.0;
 
+			//SG_SPRINT("sums = [%f, %f, %f]\n", r_sum, l_sum, fun_x);
+
 			if (r_sum <= 1e-20)
 			{
 				gradient_break = true;
@@ -244,6 +267,7 @@ slep_result_t slep_mt_lr(
 				break;
 			else 
 				L = CMath::max(2*L, l_sum/r_sum);
+			inner_iter++;
 		}
 
 		alphap = alpha;

diff --git a/src/shogun/transfer/multitask/MultitaskLogisticRegression.cpp b/src/shogun/transfer/multitask/MultitaskLogisticRegression.cpp
@@ -83,7 +83,7 @@ bool CMultitaskLogisticRegression::train_machine(CFeatures* data)
 	ASSERT(features);
 	ASSERT(m_labels);
 
-	SGVector<float64_t> y = ((CRegressionLabels*)m_labels)->get_labels();
+	SGVector<float64_t> y = ((CBinaryLabels*)m_labels)->get_labels();
 
 	slep_options options = slep_options::default_options();
 	options.q = m_q;
@@ -100,7 +100,7 @@ bool CMultitaskLogisticRegression::train_machine(CFeatures* data)
 			CTaskGroup* task_group = (CTaskGroup*)m_task_relation;
 			SGVector<index_t> ind = task_group->get_SLEP_ind();
 			options.ind = ind.vector;
-			options.n_nodes = ind.vlen-1;
+			options.n_tasks = ind.vlen-1;
 
 			slep_result_t result = slep_mt_lr(features, y.vector, m_z, options);
 			m_tasks_w = result.w;