Merge pull request #1327 from borglab/fix/gbt-determinant

gtsam/linear/GaussianBayesTree.cpp

@@ -31,18 +31,37 @@ namespace gtsam {
   template class BayesTreeCliqueBase<GaussianBayesTreeClique, GaussianFactorGraph>;
   template class BayesTree<GaussianBayesTreeClique>;
 
-  /* ************************************************************************* */
-  namespace internal
-  {
-    /* ************************************************************************* */
-  double logDeterminant(const GaussianBayesTreeClique::shared_ptr& clique,
-                        double& parentSum) {
-    parentSum += clique->conditional()
-                     ->R()
-                     .diagonal()
-                     .unaryExpr([](double x) { return log(x); })
-                     .sum();
-    return 0;
+  /* ************************************************************************ */
+  namespace internal {
+
+  /**
+   * @brief Struct to help with traversing the Bayes Tree
+   * for log-determinant computation.
+   * Records the data which is passed to the child nodes in pre-order visit.
+   */
+  struct LogDeterminantData {
+    // Use pointer so we can get the full result after tree traversal
+    double* logDet;
+    LogDeterminantData(double* logDet)
+        : logDet(logDet) {}
+  };
+  /* ************************************************************************ */
+  LogDeterminantData& logDeterminant(
+      const GaussianBayesTreeClique::shared_ptr& clique,
+      LogDeterminantData& parentSum) {
+    auto cg = clique->conditional();
+    double logDet;
+    if (cg->get_model()) {
+      Vector diag = cg->R().diagonal();
+      cg->get_model()->whitenInPlace(diag);
+      logDet = diag.unaryExpr([](double x) { return log(x); }).sum();
+    } else {
+      logDet =
+          cg->R().diagonal().unaryExpr([](double x) { return log(x); }).sum();
+    }
+    // Add the current clique's log-determinant to the overall sum
+    (*parentSum.logDet) += logDet;
+    return parentSum;
   }
   }  // namespace internal
 
@@ -87,7 +106,14 @@ namespace gtsam {
       return 0.0;
     } else {
       double sum = 0.0;
-      treeTraversal::DepthFirstForest(*this, sum, internal::logDeterminant);
+      // Store the log-determinant in this struct.
+      internal::LogDeterminantData rootData(&sum);
+      // No need to do anything for post-operation.
+      treeTraversal::no_op visitorPost;
+      // Limits OpenMP threads if we're mixing TBB and OpenMP
+      TbbOpenMPMixedScope threadLimiter;
+      // Traverse the GaussianBayesTree depth first and call logDeterminant on each node.
+      treeTraversal::DepthFirstForestParallel(*this, rootData, internal::logDeterminant, visitorPost);
       return sum;
     }
   }
@@ -106,7 +132,3 @@ namespace gtsam {
 
 
 } // \namespace gtsam
-
-
-
-

gtsam/linear/tests/testGaussianBayesTree.cpp

@@ -15,18 +15,18 @@
  * @author Kai Ni
  */
 
-#include <iostream>
 #include <CppUnitLite/TestHarness.h>
+#include <gtsam/base/debug.h>
+#include <gtsam/base/numericalDerivative.h>
+#include <gtsam/inference/Symbol.h>
+#include <gtsam/linear/GaussianBayesTree.h>
+#include <gtsam/linear/GaussianConditional.h>
+#include <gtsam/linear/GaussianJunctionTree.h>
 
 #include <boost/assign/list_of.hpp>
 #include <boost/assign/std/list.hpp>  // for operator +=
 #include <boost/assign/std/set.hpp>   // for operator +=
-
-#include <gtsam/base/debug.h>
-#include <gtsam/base/numericalDerivative.h>
-#include <gtsam/linear/GaussianJunctionTree.h>
-#include <gtsam/linear/GaussianBayesTree.h>
-#include <gtsam/linear/GaussianConditional.h>
+#include <iostream>
 
 using namespace boost::assign;
 using namespace std::placeholders;
@@ -321,6 +321,35 @@ TEST(GaussianBayesTree, determinant_and_smallestEigenvalue) {
   EXPECT_DOUBLES_EQUAL(expectedDeterminant,actualDeterminant,expectedDeterminant*1e-6);// relative tolerance
 }
 
+/* ************************************************************************* */
+/// Test to expose bug in GaussianBayesTree::logDeterminant.
+TEST(GaussianBayesTree, LogDeterminant) {
+  using symbol_shorthand::L;
+  using symbol_shorthand::X;
+
+  // Create a factor graph that will result in
+  // a bayes tree with at least 2 nodes.
+  GaussianFactorGraph fg;
+  Key x1 = X(1), x2 = X(2), l1 = L(1);
+  SharedDiagonal unit2 = noiseModel::Unit::Create(2);
+  fg += JacobianFactor(x1, 10 * I_2x2, -1.0 * Vector2::Ones(), unit2);
+  fg += JacobianFactor(x2, 10 * I_2x2, x1, -10 * I_2x2, Vector2(2.0, -1.0),
+                       unit2);
+  fg += JacobianFactor(l1, 5 * I_2x2, x1, -5 * I_2x2, Vector2(0.0, 1.0), unit2);
+  fg +=
+      JacobianFactor(x2, -5 * I_2x2, l1, 5 * I_2x2, Vector2(-1.0, 1.5), unit2);
+  fg += JacobianFactor(x3, 10 * I_2x2, x2, -10 * I_2x2, Vector2(2.0, -1.0),
+                       unit2);
+  fg += JacobianFactor(x3, 10 * I_2x2, -1.0 * Vector2::Ones(), unit2);
+
+  // create corresponding Bayes net and Bayes tree:
+  boost::shared_ptr<gtsam::GaussianBayesNet> bn = fg.eliminateSequential();
+  boost::shared_ptr<gtsam::GaussianBayesTree> bt = fg.eliminateMultifrontal();
+
+  // Test logDeterminant
+  EXPECT_DOUBLES_EQUAL(bn->logDeterminant(), bt->logDeterminant(), 1e-9);
+}
+
 /* ************************************************************************* */
 int main() { TestResult tr; return TestRegistry::runAllTests(tr);}
 /* ************************************************************************* */