BackendBatchNorm2D.h

#pragma once
 
#if _MSC_VER >= 1900 && defined(_M_X64)
 
#include "BackendBase.h"
 
namespace FLImaging
{
    namespace AI
    {
        #ifdef CUDNN_MODE
        template<typename T>
        class CCuda_ComputationalGraphBatchNorm2D_Cudnn;
        #endif
 
        template <typename T>
        class FL_EXPORT CTensor;
 
        template <typename T>
        class FL_EXPORT CBackendBase;
 
        template <typename T>
        class FL_EXPORT CBackendBatchNorm2D : public CBackendBase<T>
        {
 
        public:
            CBackendBatchNorm2D();
            CBackendBatchNorm2D(const CBackendBatchNorm2D<T>& bbn);
            virtual ~CBackendBatchNorm2D();
 
            virtual const CResult SetEpsilon(T tEpsilon);
            virtual const CResult SetMomentum(T tMomentum);
            virtual const CResult UseAffine(bool bAffine);
            virtual T GetEpsilon();
            virtual T GetMomentum();
            virtual bool IsAffineUsed();
 
            virtual const CResult Forward(CTensor<T>* pTsrX, CTensor<T>* pTsrWeight, CTensor<T>* pTsrBias, CTensor<T>* pTsrY, CTensor<T>* pTsrVarBuffer, CTensor<T>* pTsrMeanBuffer, CTensor<T>* pTsrInferVar, CTensor<T>* pTsrInferMean, bool bTrainingModeEnabled);
            virtual const CResult Derivative(CTensor<T>* pTsrX, CTensor<T>* pTsrWeight, CTensor<T>* pTsrBias, 
                                             CTensor<T>* pTsrVarBuffer, CTensor<T>* pTsrMeanBuffer, CTensor<T>* pTsrInferVar, CTensor<T>* pTsrInferMean, 
                                             CTensor<T>* pTsrDy, CTensor<T>* pTsrDx, CTensor<T>* pTsrDw, CTensor<T>* pTsrDb, bool bTrainingModeEnabled, 
                                             bool bAddGradient, bool bWeightAddGradient, bool bBiasAddGradient, CTensor<T>* pTsrAddGradientTemp = nullptr, CTensor<T>* pTsrWeightAddGradientTemp = nullptr, CTensor<T>* pTsrBiasAddGradientTemp = nullptr);
 
        protected:
            virtual const CResult BatchNorm2D_CUDA(CTensor<T>* pTsrX, CTensor<T>* pTsrWeight, CTensor<T>* pTsrBias, CTensor<T>* pTsrY, CTensor<T>* pTsrVarBuffer, CTensor<T>* pTsrMeanBuffer, CTensor<T>* pTsrInferVar, CTensor<T>* pTsrInferMean, bool bTrainingModeEnabled);
            virtual const CResult BatchNorm2D_CPU(CTensor<T>* pTsrX, CTensor<T>* pTsrWeight, CTensor<T>* pTsrBias, CTensor<T>* pTsrY, CTensor<T>* pTsrVarBuffer, CTensor<T>* pTsrMeanBuffer, CTensor<T>* pTsrInferVar, CTensor<T>* pTsrInferMean, bool bTrainingModeEnabled);
 
            virtual const CResult Derivative_CUDA(CTensor<T>* pTsrX, CTensor<T>* pTsrWeight, CTensor<T>* pTsrBias,
                                                  CTensor<T>* pTsrVarBuffer, CTensor<T>* pTsrMeanBuffer, CTensor<T>* pTsrInferVar, CTensor<T>* pTsrInferMean,
                                                  CTensor<T>* pTsrDy, CTensor<T>* pTsrDx, CTensor<T>* pTsrDw, CTensor<T>* pTsrDb, bool bTrainingModeEnabled, bool bAddGradient, bool bWeightAddGradient, bool bBiasAddGradient);
            virtual const CResult Derivative_CPU(CTensor<T>* pTsrX, CTensor<T>* pTsrWeight, CTensor<T>* pTsrBias,
                                                 CTensor<T>* pTsrVarBuffer, CTensor<T>* pTsrMeanBuffer, CTensor<T>* pTsrInferVar, CTensor<T>* pTsrInferMean,
                                                 CTensor<T>* pTsrDy, CTensor<T>* pTsrDx, CTensor<T>* pTsrDw, CTensor<T>* pTsrDb, bool bTrainingModeEnabled);
 
 
            DeclareGetClassType();
            SupportToDuplicateObjectWithoutCreateNewObject(CBackendBatchNorm2D<T>, *this);
 
        protected:
            T m_tEpsilon;
            T m_tMomentum;
            bool m_bAffine;
 
            #ifdef CUDNN_MODE
            CCuda_ComputationalGraphBatchNorm2D_Cudnn<T>* m_pCudnn;
            #endif
        };
    }
}
 
#endif