Fix perceptron BPG

Author: DubiousCactus

Algorithm.cpp

@@ -329,38 +329,40 @@ void Algorithm::train_perceptrons_BPG(Eigen::MatrixXd &weights) {
  }
 
  std::vector<int> misclassified_elements; //Row indexes of the criterion function
+ misclassified_elements.push_back(1); //Add a value to start the loop
  Eigen::MatrixXd criterion_function(input_data->getNbClasses(), input_data->getNbTrainingElements());
  criterion_function.setZero();
 
- c = 200;
+ c = 200; //Safety counter: stop if still misclassified elements anyway
  /* Iterate while there are misclassified elements and counter not equal to zero */
  while (!misclassified_elements.empty() && c--) {
-/* Empty the misclassified elements */
-misclassified_elements.clear();
-
 /* Update the criterion function */
 for (int i = 0; i < input_data->getNbTrainingElements(); i++)
  criterion_function.col(i) = outputVectors.col(i).cwiseProduct(weights.transpose() * augmented_data.col(i));
 
+int n = 0;
 for (int i = 0; i < input_data->getNbClasses(); i++) {
+ /* Empty the misclassified elements */
+ misclassified_elements.clear();
+
  /* Find all misclassified elements */
- for (int j = 0; j < criterion_function.rows(); j++)
-for (int k = 0; j < criterion_function.cols(); j++)
- if (criterion_function(j, k) < 0) //Misclassified !
-misclassified_elements.push_back(k); //Add the row index to the list
+ for (int k = 0; k < criterion_function.cols(); k++)
+if (criterion_function(n, k) < 0) //Misclassified !
+ misclassified_elements.push_back(k); //Add the col index to the list
 
- /* Calculate the gradiant and update the weights */
- Eigen::VectorXd gradiant;
- gradiant.setZero();
+ /* Calculate the gradient and update the weights */
+ Eigen::VectorXd gradient(weights.rows());
+ gradient.setZero();
 
- int k = 0;
  for (auto const &misclassified_element : misclassified_elements)
-gradiant += misclassified_element * outputVectors.row(k++);
+gradient += augmented_data.col(misclassified_element) * outputVectors(n, misclassified_element);
 
- gradiant *= LEARNING_RATE;
+ gradient *= LEARNING_RATE;
 
  /* Update the weights */
- weights.col(i) += gradiant;
+ weights.col(n) += gradient;
+
+ n++; //Next class index (don't use i because of the offset of ORL
 }
  }
 }