#pragma once

#include "core/omp_config.h"

#include "utils/vector.h"
#include "utils/matrix.h"
#include "utils/matcast.h"

#include "numerics/clip.h"
#include "numerics/log.h"
#include "numerics/sub.h"

#include "Loss.h"


namespace neural_networks{

	template <typename T>
	struct Loss_MeanSquaredError : Loss<T>  {

		utils::Matrix<T> dinputs;
		utils::Matrix<T> y_true;

		utils::Vector<T> sample_losses;

	
		utils::Vector<T> forward(const utils::Matrix<T>& y_pred, const utils::Matrix<T>& y_true) override{
			
			
			// Calculate loss
			sample_losses = numerics::mean_rowwise(numerics::pow(numerics::sub(y_true, y_pred), T{2}));


			// Return losses
			return sample_losses;
		}
		
		void backward(const utils::Matrix<T>& dvalues, const utils::Matrix<T>& y_true) override{


			// Number of samples
			const T samples = static_cast<T> (dvalues.rows());
			// Number of outputs in every sample
			const T outputs = static_cast<T> (dvalues.cols());


			// Gradient values
			dinputs = numerics::mul(numerics::div(numerics::sub(y_true, dvalues), outputs), T{-2});
			// Normalise gradient
			dinputs = numerics::div(dinputs, samples);


		}
	};



} // end namespace neural_networks