Done with activation softmax backwards

src/neural_network/layer/layer_dense.cpp

@@ -43,6 +43,9 @@
 #include <math/add.hpp>
 #include <random/uniform.hpp>
 
+#include <math/transpose.hpp>
+#include <math/sum.hpp>
+
 //---------------------------------------------------------------------------------------------------------------------------
 // PRIVATE CONSTANTS
 //---------------------------------------------------------------------------------------------------------------------------
@@ -70,7 +73,8 @@ namespace panic{
 layer_dense::layer_dense() {
 	weights.resize(0,0);
 	biases.resize(0);
-	outputs.resize(0,0);
+	dweights.resize(0,0);
+	dbiases.resize(0);
 }
 
 //--------------------------------------------------------------------------------------------------------------------------
@@ -86,9 +90,7 @@ layer_dense::layer_dense(panic::types::uint_t input_size, panic::types::uint_t n
 
 	biases.resize(neurons);
 	biases.fill(0);
-	//panic::random::uniform(biases);
-
-	outputs.resize(0,0);
+	
 }
 
 
@@ -99,7 +101,9 @@ layer_dense::layer_dense(panic::types::uint_t input_size, panic::types::uint_t n
 //		Calculated the forward pass:
 //		outputs = inputs * weights + biases
 //--------------------------------------------------------------------------------------------------------------------------
-bool layer_dense::forward(const panic::tensor::real_matrix& inputs){
+bool layer_dense::forward(const panic::tensor::real_matrix& input_data){
+
+	inputs = input_data;
 
 	if (inputs.cols() != weights.rows()){
 		return false;
@@ -128,9 +132,16 @@ bool layer_dense::forward(const panic::tensor::real_matrix& inputs){
 //		Calculated the backward pass:
 //		??
 //--------------------------------------------------------------------------------------------------------------------------
-bool layer_dense::backward(const panic::tensor::real_matrix& dinputs){
+bool layer_dense::backward(const panic::tensor::real_matrix& dvalues){
 
 
+	// Gradients on parameters
+	dweights = panic::math::matmul(panic::math::transpose(inputs), dvalues);
+	dbiases = panic::math::sum_colwise(dvalues);
+
+	// Gradients on values
+	dinputs = panic::math::matmul(dvalues, panic::math::transpose(weights));
+
 	return true;
 }