state update in each iteration for perceptron (#4320)

* state update in each iteration for perceptron
* test for custom init

src/shogun/classifier/Perceptron.cpp

@@ -35,7 +35,9 @@ void CPerceptron::init()
 	max_iter = 1000;
 	learn_rate = 0.1;
 	m_initialize_hyperplane = true;
-	SG_ADD(&max_iter, "initialize_hyperplane", "Whether to initialize hyperplane.", MS_AVAILABLE);
+	SG_ADD(
+	    &m_initialize_hyperplane, "initialize_hyperplane",
+	    "Whether to initialize hyperplane.", MS_AVAILABLE);
 	SG_ADD(&max_iter, "max_iter", "Maximum number of iterations.", MS_AVAILABLE);
 	SG_ADD(&learn_rate, "learn_rate", "Learning rate.", MS_AVAILABLE);
 }
@@ -65,15 +67,20 @@ bool CPerceptron::train_machine(CFeatures* data)
 	ASSERT(num_vec==train_labels.vlen)
 	SGVector<float64_t> output(num_vec);
 
-	SGVector<float64_t> w = get_w();
+	SGVector<float64_t> w;
 	if (m_initialize_hyperplane)
 	{
 		w = SGVector<float64_t>(num_feat);
+		set_w(w);
+
 		//start with uniform w, bias=0
+		w.set_const(1.0 / num_feat);
 		bias=0;
-		linalg::add_scalar(w, 1.0 / num_feat);
 	}
-
+	else
+	{
+		w = get_w();
+	}
 	//loop till we either get everything classified right or reach max_iter
 	while (!converged && iter < max_iter)
 	{
@@ -105,8 +112,6 @@ bool CPerceptron::train_machine(CFeatures* data)
 	else
 		SG_WARNING("Perceptron algorithm did not converge after %d iterations.\n", max_iter)
 
-	set_w(w);
-
 	return converged;
 }
 

tests/unit/classifier/Perceptron_unittest.cc

@@ -59,3 +59,25 @@ TEST(Perceptron, train)
 	auto acc = some<CAccuracyMeasure>();
 	EXPECT_EQ(acc->evaluate(results, test_labels), 1.0);
 }
+
+TEST(Perceptron, custom_hyperplane_initialization)
+{
+	auto env = linear_test_env->getBinaryLabelData();
+	auto features = wrap(env->get_features_train());
+	auto labels = wrap(env->get_labels_train());
+	auto test_features = wrap(env->get_features_test());
+	auto test_labels = wrap(env->get_labels_test());
+
+	auto perceptron = some<CPerceptron>(features, labels);
+	perceptron->train();
+
+	auto weights = perceptron->get_w();
+
+	auto perceptron_initialized = some<CPerceptron>(features, labels);
+	perceptron_initialized->set_initialize_hyperplane(false);
+	perceptron_initialized->set_w(weights);
+	perceptron_initialized->set_max_iter(1);
+
+	perceptron_initialized->train();
+	EXPECT_TRUE(perceptron_initialized->get_w().equals(weights));
+}