From 09a17dfa249b264f7c1d0b89c82d9f2a6cfbf745 Mon Sep 17 00:00:00 2001
From: sosokker <sirin_29@hotmail.co.th>
Date: Wed, 23 Aug 2023 23:49:44 +0700
Subject: [PATCH] Update Layers/ Use as parent class/ Add Dense Layers

---
 src/Activation/Activation.hpp | 16 +++++++++
 src/Layers/DenseLayer.cpp     | 26 +++++++++++++++
 src/Layers/DenseLayer.hpp     | 17 ++++++++++
 src/Layers/Layers.hpp         | 63 ++++++++++-------------------------
 4 files changed, 76 insertions(+), 46 deletions(-)
 create mode 100644 src/Activation/Activation.hpp
 create mode 100644 src/Layers/DenseLayer.cpp
 create mode 100644 src/Layers/DenseLayer.hpp

diff --git a/src/Activation/Activation.hpp b/src/Activation/Activation.hpp
new file mode 100644
index 0000000..e51f6f7
--- /dev/null
+++ b/src/Activation/Activation.hpp
@@ -0,0 +1,16 @@
+#include <functional>
+#include <Eigen/Dense>
+#include <Layers/Layers.hpp>
+
+class Activation : public Layer {
+public:
+    Activation(std::function<Eigen::VectorXd(const Eigen::VectorXd&)> activation,
+               std::function<Eigen::VectorXd(const Eigen::VectorXd&)> activation_prime);
+
+    void forward(const Eigen::VectorXd& input) override;
+    void backward(const Eigen::VectorXd& output_gradient, double learning_rate) override;
+
+private:
+    std::function<Eigen::VectorXd(const Eigen::VectorXd&)> activation;
+    std::function<Eigen::VectorXd(const Eigen::VectorXd&)> activation_prime;
+};
diff --git a/src/Layers/DenseLayer.cpp b/src/Layers/DenseLayer.cpp
new file mode 100644
index 0000000..14ec230
--- /dev/null
+++ b/src/Layers/DenseLayer.cpp
@@ -0,0 +1,26 @@
+#include "DenseLayer.hpp"
+
+DenseLayer::DenseLayer(int input_size, int output_size) {
+    weights = Eigen::MatrixXd::Random(output_size, input_size);
+    bias = Eigen::VectorXd::Random(output_size);
+}
+
+void DenseLayer::forward(const Eigen::VectorXd& input) {
+    this->input = input;
+    output = weights * input + bias;
+}
+
+void DenseLayer::backward(const Eigen::VectorXd& output_gradient, double learning_rate) {
+    Eigen::MatrixXd weights_gradient = output_gradient * input.transpose();
+    input_gradient = weights.transpose() * output_gradient;
+    weights -= learning_rate * weights_gradient;
+    bias -= learning_rate * output_gradient;
+}
+
+const Eigen::VectorXd& DenseLayer::getOutput() const {
+    return output;
+}
+
+const Eigen::VectorXd& DenseLayer::getInputGradient() const {
+    return input_gradient;
+}
diff --git a/src/Layers/DenseLayer.hpp b/src/Layers/DenseLayer.hpp
new file mode 100644
index 0000000..b2d30e4
--- /dev/null
+++ b/src/Layers/DenseLayer.hpp
@@ -0,0 +1,17 @@
+#include <Eigen/Dense>
+#include "Layers.hpp"
+
+class DenseLayer : public Layer {
+public:
+    DenseLayer(int input_size, int output_size);
+
+    void forward(const Eigen::VectorXd& input) override;
+    void backward(const Eigen::VectorXd& output_gradient, double learning_rate) override;
+
+    const Eigen::VectorXd& getOutput() const;
+    const Eigen::VectorXd& getInputGradient() const;
+
+private:
+    Eigen::MatrixXd weights;
+    Eigen::VectorXd bias;
+};
diff --git a/src/Layers/Layers.hpp b/src/Layers/Layers.hpp
index 9de4a44..a5f1239 100644
--- a/src/Layers/Layers.hpp
+++ b/src/Layers/Layers.hpp
@@ -1,51 +1,22 @@
-#pragma once
-#include <vector>
-#include <utility>
+#ifndef LAYERS_HPP
+#define LAYERS_HPP
+
+#include <Eigen/Dense>
 
 class Layer {
 public:
-    virtual void forward(const std::vector<double>& inputs, bool training) = 0;
-    virtual void backward(std::vector<double>& dvalues) = 0;
-    virtual std::pair<std::vector<std::vector<double>>, std::vector<std::vector<double>>> get_parameters() const = 0;
-    virtual void set_parameters(const std::vector<std::vector<double>>& weights, const std::vector<std::vector<double>>& biases) = 0;
+    Layer();
+
+    virtual void forward(const Eigen::VectorXd& input_data);
+    virtual void backward(const Eigen::VectorXd& output_gradient, double learning_rate);
+
+    const Eigen::VectorXd& getOutput() const;
+    const Eigen::VectorXd& getInputGradient() const;
+
+protected:
+    Eigen::VectorXd input;
+    Eigen::VectorXd output;
+    Eigen::VectorXd input_gradient;
 };
 
-class Layer_Dense : public Layer {
-private:
-    double weight_regularizer_l1, weight_regularizer_l2,
-        bias_regularizer_l1, bias_regularizer_l2;
-    std::vector<std::vector<double>> weights, biases;
-    std::vector<double> inputs;
-
-public:
-    Layer_Dense(int n_inputs, int n_neurons,
-                double weight_regularizer_l1,
-                double weight_regularizer_l2,
-                double bias_regularizer_l1,
-                double bias_regularizer_l2);
-
-    void forward(const std::vector<double>& inputs, bool training) override;
-    void backward(std::vector<double>& dvalues) override;
-    std::pair<std::vector<std::vector<double>>, std::vector<std::vector<double>>> get_parameters() const override;
-    void set_parameters(const std::vector<std::vector<double>>& weights, const std::vector<std::vector<double>>& biases) override;
-};
-
-class Layer_Dropout : public Layer {
-private:
-    double rate;
-
-public:
-    Layer_Dropout(double rate);
-
-    void forward(const std::vector<double>& inputs, bool training) override;
-    void backward(std::vector<double>& dvalues) override;
-};
-
-
-class Layer_Input {
-private:
-    std::vector<double> output;
-
-public:
-    void forward(const std::vector<double>& inputs, bool training);
-};
\ No newline at end of file
+#endif // LAYERS_HPP