Merge pull request #36 from Az-r-ow/bce

BCE

src/NeuralNet/Network.cpp

@@ -35,6 +35,9 @@ void Network::setLoss(LOSS loss) {
       this->cmpLoss = MCE::cmpLoss;
       this->cmpLossGrad = MCE::cmpLossGrad;
       break;
+    case LOSS::BCE:
+      this->cmpLoss = BCE::cmpLoss;
+      this->cmpLossGrad = BCE::cmpLossGrad;
     default:
       assert(false && "Loss not defined");
       break;

src/NeuralNet/losses/BCE.hpp

@@ -0,0 +1,23 @@
+#pragma once
+
+#include "Loss.hpp"
+
+namespace NeuralNet {
+/**
+ * Binary Cross-Entropy
+ */
+class BCE : public Loss {
+ public:
+  static double cmpLoss(const Eigen::MatrixXd &o, const Eigen::MatrixXd &y) {
+    Eigen::MatrixXd loss =
+        -(y.array() * o.array().log() + (1.0 - y.array()).log());
+    return loss.sum();
+  }
+
+  static Eigen::MatrixXd cmpLossGrad(const Eigen::MatrixXd &yHat,
+                                     const Eigen::MatrixXd &y) {
+    return (yHat.array() - y.array()) / (yHat.array() * (1.0 - y.array()));
+  }
+};
+
+}  // namespace NeuralNet

src/NeuralNet/losses/Loss.hpp

@@ -18,12 +18,12 @@ class Loss {
   /**
    * @brief This function computes the loss gradient w.r.t the outputs
    *
-   * @param o The outputs from the output layer
+   * @param yHat The outputs from the output layer
    * @param y The labels (expected vals)
    *
    * @return The current iteration's gradient
    */
-  static Eigen::MatrixXd cmpLossGrad(const Eigen::MatrixXd &o,
+  static Eigen::MatrixXd cmpLossGrad(const Eigen::MatrixXd &yHat,
                                      const Eigen::MatrixXd &y);
 };
 }  // namespace NeuralNet

src/NeuralNet/losses/losses.hpp

@@ -5,5 +5,6 @@
 
 #pragma once
 
-#include "MCE.hpp"
+#include "BCE.hpp"  // Binary Cross-Entropy
+#include "MCE.hpp"  // Multiclass Cross-Entropy
 #include "Quadratic.hpp"

src/NeuralNet/utils/Enums.hpp

@@ -17,6 +17,7 @@ enum class WEIGHT_INIT {
 
 enum class LOSS {
   MCE,  // Multi-class Cross Entropy
-  QUADRATIC
+  QUADRATIC,
+  BCE  // Binary Cross-Entropy
 };
 }  // namespace NeuralNet

src/bindings/NeuralNetPy.cpp

@@ -67,7 +67,8 @@ PYBIND11_MODULE(NeuralNetPy, m) {
 
   py::enum_<LOSS>(m, "LOSS")
       .value("QUADRATIC", LOSS::QUADRATIC)
-      .value("MCE", LOSS::MCE);
+      .value("MCE", LOSS::MCE)
+      .value("BCE", LOSS::BCE);
 
   py::module optimizers_m = m.def_submodule("optimizers", R"pbdoc(
       Optimizers

tests/test-callbacks.cpp

@@ -2,6 +2,7 @@
 #include <callbacks/Callback.hpp>
 #include <callbacks/EarlyStopping.hpp>
 #include <catch2/catch_test_macros.hpp>
+#include <catch2/matchers/catch_matchers.hpp>
 #include <utils/Variants.hpp>
 #include <vector>
 
@@ -11,7 +12,7 @@ TEST_CASE(
     "EarlyStopping callback throws exception when the metric is not found",
     "[callback]") {
   std::shared_ptr<Callback> earlyStopping =
-      std::make_shared<EarlyStopping>("LOSS", 0.1);
+      std::make_shared<EarlyStopping>("NOT_A_METRIC", 0.1);
 
   Network network;
 

tests/test-network.cpp

@@ -214,7 +214,7 @@ SCENARIO("The network updates the weights and biases as pre-calculated") {
 
   Network checkpoint;
 
-  Model::load_from_file("checkpoint-0.bin", checkpoint);
+  Model::load_from_file("N9NeuralNet7NetworkE-checkpoint-0.bin", checkpoint);
 
   REQUIRE(checkpoint.getNumLayers() == network.getNumLayers());