ComputationalGraphInvolution.h

#pragma once
 
#if _MSC_VER >= 1900 && defined(_M_X64)
 
#include "ComputationalGraph.h"
#include "Tensor.h"
#include "BackendConv2D.h"
#include "BackendFold.h"
#include "BackendAdaptiveAvgPool2D.h"
#include "BackendBatchNorm2D.h"
#include "BackendReLU.h"
 
#include <vector>
 
namespace FLImaging
{
    namespace AI
    {
        template <typename T>
        class FL_EXPORT CComputationalGraphInvolution : public CComputationalGraph<T>
        {
        private:
            CComputationalGraphInvolution();
 
        protected:
            CComputationalGraphInvolution(const CComputationalGraphInvolution<T>& cg);
 
        public:
            // Operand1
            // 4th dim : Batch size
            // 3rd dim : Channels
            // 2nd dim : Height
            // 1st dim : Width
            // 
            // Operand2 (Channel Reduce Kernel)
            // 4th dim : Channels / Reduction ratio
            // 3rd dim : Channels
            // 2nd dim : 1
            // 1st dim : 1
            // 
            // Operand3 (Channel to spatial Kernel)
            // 4th dim : KernelX * KernelY
            // 3rd dim : Channels / Reduction ratio
            // 2nd dim : 1
            // 1st dim : 1
 
            CComputationalGraphInvolution(const CComputationalBase<T>& cbOperand1, const CTensor<T>& tsrKernel1, const CTensor<T>& tsrKernel2, int64_t i64KernelX = 3, int64_t i64KernelY = 3, int64_t i64StrideX = 1, int64_t i64StrideY = 1, int64_t i64GroupCount = 1, T tEpsilon = (T)FL_EPSILON_FLOAT, T tMomentum = (T).1, bool bAffine = true);
            virtual ~CComputationalGraphInvolution();
 
            virtual CTensor<T>& Forward() override;
            virtual CTensor<T>* Backward() override;
            virtual CComputationalBase<T>* Clone() const override;
 
            virtual const std::vector<int64_t>& GetEstimatedShape(bool bRecursive = true) const override;
            virtual const CResult PrintNodeParamInfo() const override;
 
            virtual const CResult GetBinaryData(Base::CFLData& fldBinary, bool bSuperClass = false, int32_t i32Version = -1, bool bDumpMode = false) const override;
            virtual const CResult GetBinaryData(Base::CFLData* pFldBinary, bool bSuperClass = false, int32_t i32Version = -1, bool bDumpMode = false) const override;
 
            virtual const CResult SetBinaryData(const Base::CFLData& fldBinary, int64_t* pI64Offset = nullptr) override;
            virtual const CResult SetBinaryData(const Base::CFLData* pFldBinary, int64_t* pI64Offset = nullptr) override;
 
            virtual int64_t GetRequiredTemporaryMemory(bool bTraining = false, bool bRecursively = true, int64_t i64BatchSize = 1, int64_t i64MemoryIndex = 0) const override;
            virtual int64_t GetRequiredDedicatedMemory(bool bTraining = false, bool bRecursively = true, int64_t i64BatchSize = 1) const override;
 
            DeclareGetClassType();
            SupportToDuplicateObjectWithoutCreateNewObject(CComputationalGraphInvolution, *this);
 
        protected:
            const CResult CheckShape();
            const CResult GetCommonValues(std::vector<int64_t>& vctPooledShape, std::vector<int64_t>& vctReduceOutputShape, std::vector<int64_t>& vctSpatialOutputShape, bool& bPoolSkip) const;
            const CResult InvolutionMainForward(const CTensor<T>* pTsrOperand, const CTensor<T>* pTsrKernel, CTensor<T>* pTsrResult, const std::vector<int64_t> vctResultShape);
            const CResult InvolutionMainBackwardImage(const CTensor<T>* pTsrDy, const CTensor<T>* pTsrW, CTensor<T>* pTsrDx, const std::vector<int64_t> vctXShape, bool bAddGradient, CTensor<T>* pTsrDerivativeTemp);
            const CResult InvolutionMainBackwardKernel(const CTensor<T>* pTsrDy, const CTensor<T>* pTsrX, CTensor<T>* pTsrDw, std::vector<int64_t> vctWShape);
 
            int64_t m_i64KernelX;
            int64_t m_i64KernelY;
            int64_t m_i64StrideX;
            int64_t m_i64StrideY;
            int64_t m_i64GroupCount;
            T m_tEpsilon;
            T m_tMomentum;
            bool m_bAffine;
 
            CTensor<T> m_tsrPoolVal;
            CTensor<T> m_tsrKernelBuffer;
            CTensor<T> m_tsrReduceVal;
            CTensor<T> m_tsrBNVarienceBuffer;
            CTensor<T> m_tsrBNMeanBuffer;
            CTensor<T> m_tsrBNSaveWeightGradient;
            CTensor<T> m_tsrBNSaveBiasGradient;
            CTensor<T> m_tsrReLUVal;
            CTensor<T> m_tsrMulKernel;
 
            CBackendAdaptiveAvgPool2D<T> m_backendAdaptiveAvgPool2D;
            CBackendConv2D<T> m_backendConv2DReduce;
            CBackendConv2D<T> m_backendConv2DSpatial;
            CBackendBatchNorm2D<T> m_backendBatchNorm2D;
            CBackendReLU<T> m_backendReLU;
 
        public:
            DeclareGetSignletonObject(CComputationalGraphInvolution);
        };
 
        #define CCGFInvolution(...) (*(new CComputationalGraphInvolution<float>(__VA_ARGS__)))
        #define CCGDInvolution(...) (*(new CComputationalGraphInvolution<double>(__VA_ARGS__)))
 
        #define CCGTInvolution(T, ...) (*(new CComputationalGraphInvolution<T>(__VA_ARGS__)))
    }
}
 
#endif