BackendLayerNorm.h

#pragma once
 
#if _MSC_VER >= 1900 && defined(_M_X64)
 
#include "BackendBase.h"
 
namespace FLImaging
{
    namespace AI
    {
        template <typename T>
        class FL_EXPORT CTensor;
 
        template <typename T>
        class FL_EXPORT CBackendBase;
 
        template <typename T>
        class FL_EXPORT CBackendLayerNorm : public CBackendBase<T>
        {
 
        public:
            CBackendLayerNorm();
            CBackendLayerNorm(const CBackendLayerNorm<T>& bbn);
            virtual ~CBackendLayerNorm();
 
            virtual const CResult SetEpsilon(T tEpsilon);
            virtual const CResult UseAffine(bool bAffine);
            virtual const CResult UseBias(bool bBias);
            virtual const CResult SetNormalizeShape(const std::vector<int64_t>& vtNormalizeShape);
            virtual T GetEpsilon();
            virtual bool IsAffineUsed();
            virtual bool IsBiasUsed();
 
            virtual const CResult Forward(CTensor<T>* pTsrX, CTensor<T>* pTsrWeight, CTensor<T>* pTsrBias, CTensor<T>* pTsrY, CTensor<T>* pTsrVarBuffer, CTensor<T>* pTsrMeanBuffer,  bool bTrainingModeEnabled);
            virtual const CResult Derivative(CTensor<T>* pTsrX, CTensor<T>* pTsrWeight, CTensor<T>* pTsrBias, CTensor<T>* pTsrVarBuffer, CTensor<T>* pTsrMeanBuffer, 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 LayerNorm_CUDA(CTensor<T>* pTsrX, CTensor<T>* pTsrWeight, CTensor<T>* pTsrBias, CTensor<T>* pTsrY, CTensor<T>* pTsrVarBuffer, CTensor<T>* pTsrMeanBuffer, bool bTrainingModeEnabled);
            virtual const CResult LayerNorm_CPU(CTensor<T>* pTsrX, CTensor<T>* pTsrWeight, CTensor<T>* pTsrBias, CTensor<T>* pTsrY, CTensor<T>* pTsrVarBuffer, CTensor<T>* pTsrMeanBuffer, bool bTrainingModeEnabled);
            virtual const CResult Derivative_CUDA(CTensor<T>* pTsrX, CTensor<T>* pTsrWeight, CTensor<T>* pTsrBias,CTensor<T>* pTsrVarBuffer, CTensor<T>* pTsrMeanBuffer,   CTensor<T>* pTsrDy, CTensor<T>* pTsrDx, CTensor<T>* pTsrDw, CTensor<T>* pTsrDb, bool bTrainingModeEnabled, bool bAddGradient);
            virtual const CResult Derivative_CPU(CTensor<T>* pTsrX, CTensor<T>* pTsrWeight, CTensor<T>* pTsrBias, CTensor<T>* pTsrVarBuffer, CTensor<T>* pTsrMeanBuffer, CTensor<T>* pTsrDy, CTensor<T>* pTsrDx, CTensor<T>* pTsrDw, CTensor<T>* pTsrDb, bool bTrainingModeEnabled);
 
            DeclareGetClassType();
            SupportToDuplicateObjectWithoutCreateNewObject(CBackendLayerNorm<T>, *this);
 
        protected:
            T m_tEpsilon;
            bool m_bAffine;
            bool m_bBias;
            std::vector<int64_t>& m_vtNormalizeShape;
        };
    }
}
 
#endif