Some performance tuning

2008-02-18 19:05:46 +00:00
parent 65bf7b1c36
commit 4de9e7109a
4 changed files with 27 additions and 3 deletions
@@ -429,6 +429,12 @@ class DenseLpr {

    ////////// Private Member Variables //////////

+    /** @brief The number of finite difference prunes. */
+    int num_finite_difference_prunes_;
+
+    /** @brief The number of far-field prunes. */
+    int num_far_field_prunes_;
+
    /** @brief The local polynomial order. */
    int lpr_order_;

@@ -757,11 +763,18 @@ class DenseLpr {
 		      Vector *query_magnitude_weight_diagrams,
 		      Vector *query_influence_values) {
 
+      // Clear prune statistics.
+      num_finite_difference_prunes_ = num_far_field_prunes_ = 0;
+
      // This is the basic N-body based computation.
      BasicCompute_(queries, query_regression_estimates,
 		    query_confidence_bands, query_magnitude_weight_diagrams,
 		    query_influence_values);

+      printf("Number of finite difference prunes: %d\n",
+	     num_finite_difference_prunes_);
+      printf("Number of far-field prunes: %d\n", num_far_field_prunes_);
+
      // If the reference dataset is being used for training, then
      // compute variance and degrees of freedom.
      if(query_influence_values != NULL) {
@@ -473,6 +473,9 @@ void DenseLpr<TKernel, TPruneRule>::DualtreeLprCanonical_
 	      &(qnode->stat().postponed_weight_diagram_numerator_e_));
    qnode->stat().postponed_weight_diagram_numerator_used_error_ += 
      delta_weight_diagram_numerator_used_error;
+
+    // Keep track of the number of finite difference prunes.
+    num_finite_difference_prunes_++;
    return;
  }

@@ -480,7 +483,8 @@ void DenseLpr<TKernel, TPruneRule>::DualtreeLprCanonical_
  // moments if the maximum distance between the two nodes is within
  // the bandwidth! This if-statement does not apply to the Gaussian
  // kernel, so I need to fix in the future!
-  if(kernel_aux_.kernel_.bandwidth_sq() >= dsqd_range.hi) {
+  if(kernel_aux_.kernel_.bandwidth_sq() >= dsqd_range.hi &&
+     rnode->count() > 32) {

    for(index_t q = qnode->begin(); q < qnode->end(); q++) {
      for(index_t j = 0; j < row_length_; j++) {
@@ -512,6 +516,9 @@ void DenseLpr<TKernel, TPruneRule>::DualtreeLprCanonical_

    la::AddTo(weight_diagram_numerator_dl, 
 	      &(qnode->stat().postponed_weight_diagram_numerator_l_));
+
+    // Keep track of the far-field prunes.
+    num_far_field_prunes_++;
    return;
  }

@@ -55,18 +55,22 @@ int main(int argc, char *argv[]) {
  DatasetScaler::ScaleDataByMinMax(queries, references, false);

  // Do fast algorithm.
+  printf("Running the fast algorithm...\n");
  Vector fast_lpr_results;
  DenseLpr<EpanKernelAux, RelativePruneLpr> fast_lpr;
  fast_lpr.Init(references, reference_targets, local_linear_module);
  fast_lpr.PrintDebug();
  fast_lpr.get_regression_estimates(&fast_lpr_results);
+  printf("Finished running the fast algorithm...\n");

  // Do naive algorithm.
+  printf("Running the naive algorithm...\n");
  Vector naive_lpr_results;
  NaiveLpr<EpanKernel> naive_lpr;
  naive_lpr.Init(references, reference_targets, local_linear_module);
  naive_lpr.PrintDebug();
  naive_lpr.get_regression_estimates(&naive_lpr_results);
+  printf("Finished running the naive algorithm...\n");

  printf("Maximum relative difference: %g\n",
 	 MatrixUtil::MaxRelativeDifference(naive_lpr_results,
@@ -122,12 +122,12 @@ class RelativePruneLpr {
      denominator_used_error = kernel_error *
 	(rnode->stat().sum_data_outer_products_error_norm_);
      denominator_n_pruned = rnode->stat().sum_data_outer_products_alloc_norm_;
-      
+
      // The total norm error for each query point for approximating
      // the B^T W(q)^2 B matrix.
      weight_diagram_numerator_used_error = squared_kernel_error *
 	(rnode->stat().sum_data_outer_products_error_norm_);
-      
+
      // Check pruning condition.
      return (numerator_used_error <= numerator_allowed_err &&
 	      denominator_used_error <= denominator_allowed_err &&