Merge pull request #55 from akleeman/ransac_testing

Add additional tests of ransac with groups.

Author: akleeman

albatross/evaluate.h

@@ -56,7 +56,7 @@ negative_log_likelihood(const Eigen::VectorXd &deviation,
   const double rank = static_cast<double>(diag.size());
   const double mahalanobis = deviation.dot(ldlt.solve(deviation));
   const double log_det = log_sum(diag);
-  return -0.5 * (log_det + mahalanobis + rank * log(2 * M_PI));
+  return 0.5 * (log_det + mahalanobis + rank * log(2 * M_PI));
 }
 
 /*

albatross/models/ransac.h

@@ -101,10 +101,9 @@ ransac(const typename RansacFunctions<FitType>::Fitter &fitter,
         inliers.push_back(test_ind);
       }
     }
-
     // If there is enough agreement, consider this random set of inliers
     // as a candidate model.
-    if (inliers.size() > min_inliers) {
+    if (inliers.size() + random_sample_size >= min_inliers) {
       const auto inlier_inds = concatenate_subset_of_groups(inliers, indexer);
       ref_inds.insert(ref_inds.end(), inlier_inds.begin(), inlier_inds.end());
       std::sort(ref_inds.begin(), ref_inds.end());

tests/test_evaluate.cc

@@ -44,10 +44,17 @@ TEST(test_evaluate, test_negative_log_likelihood) {
   cov << 1., 0.9, 0.8, 0.9, 1., 0.9, 0.8, 0.9, 1.;
 
   const auto nll = albatross::negative_log_likelihood(x, cov);
-  EXPECT_NEAR(nll, -6.0946974293510134, 1e-6);
+  EXPECT_NEAR(nll, 6.0946974293510134, 1e-6);
 
   const auto ldlt_nll = albatross::negative_log_likelihood(x, cov.ldlt());
   EXPECT_NEAR(nll, ldlt_nll, 1e-6);
+
+  const DiagonalMatrixXd diagonal_matrix = cov.diagonal().asDiagonal();
+  const Eigen::MatrixXd dense_diagonal = diagonal_matrix.toDenseMatrix();
+  const auto diag_nll = albatross::negative_log_likelihood(x, diagonal_matrix);
+  const auto dense_diag_nll =
+      albatross::negative_log_likelihood(x, dense_diagonal);
+  EXPECT_NEAR(diag_nll, dense_diag_nll, 1e-6);
 }
 
 TEST_F(LinearRegressionTest, test_leave_one_out) {
@@ -67,27 +74,6 @@ TEST_F(LinearRegressionTest, test_leave_one_out) {
   EXPECT_LT(in_sample_rmse, out_of_sample_rmse);
 }
 
-// Group values by interval, but return keys that once sorted won't be
-// in order
-std::string group_by_interval(const double &x) {
-  if (x <= 3) {
-    return "2";
-  } else if (x <= 6) {
-    return "3";
-  } else {
-    return "1";
-  }
-}
-
-bool is_monotonic_increasing(const Eigen::VectorXd &x) {
-  for (Eigen::Index i = 0; i < x.size() - 1; i++) {
-    if (x[i + 1] - x[i] <= 0.) {
-      return false;
-    }
-  }
-  return true;
-}
-
 TEST_F(LinearRegressionTest, test_cross_validated_predict) {
   const auto folds = leave_one_group_out<double>(dataset_, group_by_interval);
 

tests/test_outlier.cc

@@ -40,6 +40,30 @@ TEST(test_outlier, test_ransac) {
             modified.features.end());
 }
 
+// Group values by interval, but return keys that once sorted won't be
+// in order
+std::string group_by_modulo(const double &x) {
+  const int x_int = static_cast<int>(x);
+  return std::to_string(x_int % 4);
+}
+
+TEST(test_outlier, test_ransac_groups) {
+  auto dataset = make_toy_linear_data();
+  const auto model_ptr = toy_gaussian_process();
+
+  EvaluationMetric<JointDistribution> nll =
+      albatross::evaluation_metrics::negative_log_likelihood;
+
+  dataset.targets.mean[5] = -300.;
+
+  const auto fold_indexer =
+      leave_one_group_out_indexer<double>(dataset, group_by_modulo);
+  const auto modified =
+      ransac(dataset, fold_indexer, model_ptr.get(), nll, 0., 1, 1, 20);
+
+  EXPECT_LE(modified.features.size(), dataset.features.size());
+}
+
 TEST(test_outlier, test_ransac_gp) {
   auto dataset = make_toy_linear_data();
 

tests/test_utils.h

@@ -282,6 +282,27 @@ inline auto random_spherical_points(std::size_t n = 10, double radius = 1.,
   return points;
 }
 
+// Group values by interval, but return keys that once sorted won't be
+// in order
+inline std::string group_by_interval(const double &x) {
+  if (x <= 3) {
+    return "2";
+  } else if (x <= 6) {
+    return "3";
+  } else {
+    return "1";
+  }
+}
+
+inline bool is_monotonic_increasing(const Eigen::VectorXd &x) {
+  for (Eigen::Index i = 0; i < x.size() - 1; i++) {
+    if (x[i + 1] - x[i] <= 0.) {
+      return false;
+    }
+  }
+  return true;
+}
+
 } // namespace albatross
 
 #endif