Backwards Pass

Activation Softmax with CategoricalCrossentrophy is almost done from page. 234

include/modules/neural_networks/activation_functions/Activation_Softmax.h

@@ -11,30 +11,47 @@
 #include "./numerics/matdiv.h"
 
 
+
 namespace neural_networks{
 
 	template <typename T>
 	struct Activation_Softmax{
 
-			utils::Matrix<T> exp_values;
+			//utils::Matrix<T> exp_values;
-			utils::Matrix<T> probabilities;
+			//utils::Matrix<T> probabilities;
 			utils::Matrix<T> outputs;
 
+			utils::Matrix<T> dinputs;
+
 	        void forward(const utils::Matrix<T>& inputs){
 	
 	        	// Get unnormalized probabilities
-			    exp_values = numerics::matexp(numerics::matsubtract(inputs, numerics::matmax(inputs, "rows"), "col"));
+			    utils::Matrix<T> exp_values = numerics::matexp(numerics::matsubtract(inputs, numerics::matmax(inputs, "rows"), "col"));
 
 			    // Normalize them for each sample
-			    probabilities = numerics::matdiv(exp_values, numerics::matsum(exp_values, "col"), "col");
+			    utils::Matrix<T> probabilities = numerics::matdiv(exp_values, numerics::matsum(exp_values, "col"), "col");
-
-
 			    outputs = probabilities;
-
 	        }
 
+	        void backward(const utils::Matrix<T>& dvalues){
+
+	        	const uint64_t rows = dvalues.rows();
+	        	const uint64_t cols = dvalues.cols();
+
+	        	if ((dinputs.rows() != rows) || dinputs.cols() != cols){
+	        		dinputs.resize(rows, cols);
+	        	}
+
+	        	for (uint64_t i = 0; i < rows; ++i){
+	        		T dot = T{0};
+	        		for (uint64_t j = 0; j < cols; ++j){
+	        			dot += outputs(i,j) * dvalues(i,j);
+	        		}
+	        		for (uint64_t j = 0; j < cols; ++j){
+	        			dinputs(i,j) = outputs(i,j) * (dvalues(i,j) - dot);
+	        		}
+	        	}
+	        }
 	};
 
-
-
 } // end namespace neural_networks

include/modules/neural_networks/activation_functions/Activation_Softmax_Loss_CategoricalCrossentropy.h

@@ -0,0 +1,68 @@
+#pragma once
+
+#include "./core/omp_config.h"
+
+#include "./utils/vector.h"
+#include "./utils/matrix.h"
+
+#include "./numerics/matmax.h"
+#include "./numerics/matsubtract.h"
+#include "./numerics/matexp.h"
+#include "./numerics/matdiv.h"
+
+#include "./modules/neural_networks/activation_functions/Activation_Softmax.h"
+#include "./modules/neural_networks/loss/Loss_CategoricalCrossentrophy.h"
+
+
+namespace neural_networks{
+
+	template <typename Td, typename Ti>
+	struct Activation_Softmax_Loss_CategoricalCrossentropy{
+
+			neural_networks::Activation_Softmax<Td> activation;
+			neural_networks::Loss_CategoricalCrossentrophy<Td, Ti> loss;
+
+			//utils::Matrix<T> exp_values;
+			//utils::Matrix<T> probabilities;
+			utils::Matrix<Td> outputs;
+			utils::Matrix<Td> dinputs;
+
+	        utils::Vector<Td> forward(const utils::Matrix<Td>& inputs, const utils::Matrix<Ti>& y_true){
+				
+				// Output layer's activation function
+				activation.forward(inputs);
+	        	// Set the output
+			    outputs = activation.outputs;
+			    // Calculate and return loss value
+			    Td data_loss = loss.calculate(inputs, y_true);
+			    return data_loss;
+	        }
+
+	        void backward(const utils::Matrix<Td>& dvalues, const utils::Matrix<Ti>& y_true){
+
+	        	// Number of samples
+	        	const uint64_t samples = y_true.rows();
+
+	        	// If the labels are one-hot encoded,
+	        	// turn them into discrete values
+
+	        	const uint64_t rows = dvalues.rows();
+	        	const uint64_t cols = dvalues.cols();
+
+	        	if ((dinputs.rows() != rows) || dinputs.cols() != cols){
+	        		dinputs.resize(rows, cols);
+	        	}
+
+	        	for (uint64_t i = 0; i < rows; ++i){
+	        		Td dot = Td{0};
+	        		for (uint64_t j = 0; j < cols; ++j){
+	        			dot += outputs(i,j) * dvalues(i,j);
+	        		}
+	        		for (uint64_t j = 0; j < cols; ++j){
+	        			dinputs(i,j) = outputs(i,j) * (dvalues(i,j) - dot);
+	        		}
+	        	}
+	        }
+	};
+
+} // end namespace neural_networks

include/modules/neural_networks/loss/Loss.h

@@ -13,9 +13,11 @@ namespace neural_networks{
 	struct Loss{
 
 			utils::Vector<Td> sample_losses;
+			utils::Matrix<Td> dinputs;
 			Td data_loss;
 
 			virtual utils::Vector<Td> forward(const utils::Matrix<Td>& output, const utils::Matrix<Ti>& y) = 0;
+			virtual void backward(const utils::Matrix<Td>& dvalues, const utils::Matrix<Ti>& y) = 0;
 
 			Td calculate(const utils::Matrix<Td>& output, const utils::Matrix<Ti>& y){
 				

include/modules/neural_networks/loss/Loss_CategoricalCrossentrophy.h

@@ -17,6 +17,9 @@ namespace neural_networks{
 	template <typename Td, typename Ti>
 	struct Loss_CategoricalCrossentrophy : Loss<Td, Ti>  {
 
+		utils::Matrix<Td> dinputs;
+
+	
 		utils::Vector<Td> forward(const utils::Matrix<Td>& y_pred, const utils::Matrix<Ti>& y_true) override{
 			
 			utils::Vector<Td> correct_confidences(y_true.rows(), Td{0});
@@ -48,6 +51,32 @@ namespace neural_networks{
 
 			return negative_log_likelihoods;
 		}
+		
+		void backward(const utils::Matrix<Td>& dvalues, const utils::Matrix<Ti>& y_true) override{
+
+			// Number of samples
+			const Td samples = static_cast<Td> (y_true.rows());
+			// Number of labels in every sample
+			// We'll use the first samle to count them
+			const Ti labels = dvalues.cols();
+
+			utils::Matrix<Ti> y_temp;
+
+			if (y_true.cols() == 1){
+			 	
+			 	y_temp = utils::eye(labels, y_true.get_col(0));
+			 }else{
+			 	y_temp = y_true;
+			 }
+
+
+			 // Calculate the gradient
+			 numerics::inplace_matscalar(y_temp,Ti{-1});
+			 dinputs = numerics::matdiv(utils::matcast<Td, Ti>(y_temp), dvalues);
+			 numerics::inplace_matdiv(dinputs, samples);
+
+
+		}
 	};
 
 

include/modules/neural_networks/neural_networks.h

@@ -9,6 +9,7 @@
 
 #include "activation_functions/Activation_ReLU.h"
 #include "activation_functions/Activation_Softmax.h"
+#include "activation_functions/Activation_Softmax_Loss_CategoricalCrossentropy.h"
 
 #include "loss/Loss.h" // Base
 #include "loss/Loss_CategoricalCrossentrophy.h"

include/numerics/matdiv.h

@@ -8,7 +8,6 @@
 
 namespace numerics{
 
-// ---------------- Serial baseline ----------------
 	template <typename T>
 	utils::Matrix<T> matdiv(const utils::Matrix<T>& A, const utils::Vector<T>& b, std::string method){
         
@@ -33,6 +32,60 @@ namespace numerics{
 	}
 
 
+
+
+    template <typename T>
+    void inplace_matdiv(utils::Matrix<T>& A, const utils::Matrix<T>& B){
+        
+        const uint64_t rows = A.rows();
+        const uint64_t cols = A.cols();
+
+        if ((rows != B.rows()) || (cols != B.cols())){
+            throw std::runtime_error("inplace_matdiv: rows and cols are not the same'");
+        }
+
+        for (uint64_t i = 0; i < rows; ++i){
+            for (uint64_t j = 0; j < cols; ++j){
+                A(i,j) /= B(i,j);
+            }
+        }
+    }
+
+
+    template <typename T>
+    utils::Matrix<T> matdiv(const utils::Matrix<T>& A, const utils::Matrix<T>& B){
+        
+        const uint64_t rows = A.rows();
+        const uint64_t cols = A.cols();
+
+        if ((rows != B.rows()) || (cols != B.cols())){
+            throw std::runtime_error("matdiv: choose div by: 'row' or 'col'");
+        }
+
+        utils::Matrix<T> C = A;
+
+        inplace_matdiv(C, B);
+
+        return C;
+    }
+
+    template <typename T>
+    void inplace_matdiv(utils::Matrix<T>& A, const T b){
+        
+        const uint64_t rows = A.rows();
+        const uint64_t cols = A.cols();
+
+
+        for (uint64_t i = 0; i < rows; ++i){
+            for (uint64_t j = 0; j < cols; ++j){
+                A(i,j) /= b;
+            }
+        }
+    }
+
+
+
+
 } // namespace numerics
 
 #endif // _matdiv_n_

include/utils/generators.h

@@ -2,3 +2,4 @@
 #pragma once
 
 #include "./utils/generators/linspace.h"
+#include "./utils/generators/eye.h"

include/utils/generators/eye.h

@@ -0,0 +1,24 @@
+#pragma once
+
+#include "utils/vector.h"
+#include "utils/matrix.h"
+
+
+namespace utils{
+
+	template <typename T>
+	utils::Matrix<T> eye(const T a, const utils::Vector<T>& b){
+
+		const uint64_t N = b.size();
+		utils::Matrix<T> C(N, a, T{0});
+
+		for (uint64_t i = 0; i < N; ++i){
+			C(i, b[i]) = T{1};
+		}
+		return C;
+	}
+
+
+
+
+} // end namespace utils

obj/main.d

@@ -2,7 +2,8 @@ obj/main.o: src/main.cpp include/./core/omp_config.h \
  include/./utils/utils.h include/./utils/vector.h \
  include/./utils/random.h include/./utils/matrix.h \
  include/./utils/generators.h include/./utils/generators/linspace.h \
- include/utils/vector.h include/./utils/matcast.h \
+ include/utils/vector.h include/./utils/generators/eye.h \
+ include/utils/matrix.h include/./utils/matcast.h \
  include/./numerics/numerics.h include/./numerics/max.h \
  include/./numerics/exp.h include/./numerics/log.h \
  include/./numerics/vecclip.h include/./numerics/vecexp.h \
@@ -30,16 +31,16 @@ obj/main.o: src/main.cpp include/./core/omp_config.h \
  include/./decomp/decomp.h include/./modules/mesh/mesh.h \
  include/modules/mesh/mesh1d.h include/modules/fluids/fluids.h \
  include/modules/fluids/diffusion1d.h include/core/global_config.h \
- include/utils/matrix.h \
  include/./modules/neural_networks/neural_networks.h \
  include/./modules/neural_networks/datasets/spiral.h \
  include/./modules/neural_networks/datasets/vertical.h \
  include/./modules/neural_networks/layers/Dense_Layer.h \
  include/./modules/neural_networks/activation_functions/Activation_ReLU.h \
  include/./modules/neural_networks/activation_functions/Activation_Softmax.h \
- include/./modules/neural_networks/loss/Loss.h \
+ include/./modules/neural_networks/activation_functions/Activation_Softmax_Loss_CategoricalCrossentropy.h \
- include/./numerics/vecmean.h \
+ include/./modules/neural_networks/loss/Loss_CategoricalCrossentrophy.h \
- include/./modules/neural_networks/loss/Loss_CategoricalCrossentrophy.h
+ include/./modules/neural_networks/loss/./Loss.h \
+ include/./numerics/vecmean.h
 include/./core/omp_config.h:
 include/./utils/utils.h:
 include/./utils/vector.h:
@@ -48,6 +49,8 @@ include/./utils/matrix.h:
 include/./utils/generators.h:
 include/./utils/generators/linspace.h:
 include/utils/vector.h:
+include/./utils/generators/eye.h:
+include/utils/matrix.h:
 include/./utils/matcast.h:
 include/./numerics/numerics.h:
 include/./numerics/max.h:
@@ -95,13 +98,13 @@ include/modules/mesh/mesh1d.h:
 include/modules/fluids/fluids.h:
 include/modules/fluids/diffusion1d.h:
 include/core/global_config.h:
-include/utils/matrix.h:
 include/./modules/neural_networks/neural_networks.h:
 include/./modules/neural_networks/datasets/spiral.h:
 include/./modules/neural_networks/datasets/vertical.h:
 include/./modules/neural_networks/layers/Dense_Layer.h:
 include/./modules/neural_networks/activation_functions/Activation_ReLU.h:
 include/./modules/neural_networks/activation_functions/Activation_Softmax.h:
-include/./modules/neural_networks/loss/Loss.h:
+include/./modules/neural_networks/activation_functions/Activation_Softmax_Loss_CategoricalCrossentropy.h:
-include/./numerics/vecmean.h:
 include/./modules/neural_networks/loss/Loss_CategoricalCrossentrophy.h:
+include/./modules/neural_networks/loss/./Loss.h:
+include/./numerics/vecmean.h:

src/main.cpp

@@ -24,20 +24,20 @@ int main(int argc, char const *argv[])
 {   
 
     utils::Mf X(10,2, 0);
-    utils::Matrix<uint64_t> y(10,1, 0);
+    utils::Matrix<int64_t> y(10,1, 0);
-    utils::Vector<uint64_t> class_targets;
+    utils::Vector<int64_t> class_targets;
     float loss;
     float accuracy;
 
     //neural_networks::create_spital_data<float, uint64_t>(10000, 3, X, y);
-    neural_networks::create_vertical_data<float, uint64_t>(100, 3, X, y);
+    neural_networks::create_vertical_data<float, int64_t>(100, 3, X, y);
 
     neural_networks::Dense_Layer<float> dense1(2, 3);
     neural_networks::Activation_ReLU<float> activation1;
     neural_networks::Dense_Layer<float> dense2(3, 3);
     neural_networks::Activation_Softmax<float> activation2;
     
-    neural_networks::Loss_CategoricalCrossentrophy<float, uint64_t> loss_funtion;
+    neural_networks::Loss_CategoricalCrossentrophy<float, int64_t> loss_funtion;
 
     float lowest_loss = 9999999;
     utils::Mf best_dense_1_weights = dense1.weights;
@@ -47,7 +47,7 @@ int main(int argc, char const *argv[])
 
     utils::Vf vectRND;
 
-    utils::Vector<uint64_t> predections;
+    utils::Vector<int64_t> predections;
 
 
 
@@ -73,10 +73,10 @@ int main(int argc, char const *argv[])
             // it takes the output of the second dense layer here and returns loss
             loss = loss_funtion.calculate(activation2.outputs, y);
 
-            predections = numerics::matargmax_row<uint64_t, float>(activation2.outputs);
+            predections = numerics::matargmax_row<int64_t, float>(activation2.outputs);
 
             if (y.cols() < 1){
-                class_targets = numerics::matargmax_row<uint64_t, uint64_t>(y);
+                class_targets = numerics::matargmax_row<int64_t, int64_t>(y);
             }else{
                 class_targets = y.get_col(0);
             }