Updating API of LSTM layer, fixed LSTM bug in tanh() implementation

LSTM bug is caused by different behaviour of std::tanh() and hand-
crafted
tanh() via std::exp().

modules/dnn/CMakeLists.txt

@@ -24,7 +24,7 @@ ocv_module_include_directories(include ${PROTOBUF_INCLUDE_DIR})
 OCV_OPTION(${the_module}_WITH_BLAS "Use external BLAS library to speedup processing" OFF)
 include(cmake/OpenCVFindCBLAS.cmake)
 
-ocv_glob_module_sources(${PROTOBUF_SRCS} ${PROTOBUF_HDRS} ${CBLAS_H_PATH})
+ocv_glob_module_sources(${PROTOBUF_SRCS} ${PROTOBUF_HDRS} ${CBLAS_H_PROXY_PATH})
 ocv_create_module(${PROTOBUF_LIBRARIES})
 ocv_add_samples()
 ocv_add_accuracy_tests()
@@ -37,6 +37,7 @@ if(${the_module}_WITH_BLAS AND HAVE_BLAS)
     add_definitions(-DHAVE_CBLAS=1)
     ocv_module_include_directories(${${the_module}_BLAS_INCLUDE_DIR})
     ocv_add_dependencies(${the_module} ${${the_module}_BLAS_LIBRARIES})
+    target_link_libraries(${the_module} ${${the_module}_BLAS_LIBRARIES})
 
     if(${the_module}_BLAS_BINARIES)
         ocv_install_target(${the_module} EXPORT ${the_module}_BLAS_BINARIES

modules/dnn/cmake/OpenCVFindCBLAS.cmake

@@ -1,6 +1,6 @@
 macro(_find_file_in_dirs VAR NAME DIRS)
     find_path(${VAR} ${NAME} ${DIRS} NO_DEFAULT_PATH)
-    set(${VAR} ${${VAR}})
+    set(${VAR} ${${VAR}}/${NAME})
     unset(${VAR} CACHE)
 endmacro()
 
@@ -16,7 +16,7 @@ if(${the_module}_WITH_BLAS)
     endif()
     if(NOT HAVE_BLAS)
         include(cmake/OpenCVFindMKL.cmake)
-        if(MKL_FOUND)
+        if(MKL_FOUND AND FALSE)
             set(BLAS_INCLUDE_DIR    ${MKL_INCLUDE_DIRS})
             set(BLAS_LIBRARIES      ${MKL_LIBRARIES}   )
             set(BLAS_CBLAS_H        "mkl_cblas.h"      )
@@ -52,8 +52,9 @@ if(${the_module}_WITH_BLAS)
         if(NOT CBLAS_H_PATH)
             message(WARNING "CBLAS header '${${_bp}_CBLAS_H}' not found into '${${_bp}_INCLUDE_DIR}'")
         endif()
-        add_custom_command(OUTPUT ${CMAKE_CURRENT_BINARY_DIR}/cblas.h #TARGET ${the_module} PRE_BUILD
+
-                           COMMAND ${CMAKE_COMMAND} ARGS -E echo "\#include \"${CBLAS_H_PATH}\"" > ${CMAKE_CURRENT_BINARY_DIR}/cblas.h
+        set(CBLAS_H_PROXY_PATH ${CMAKE_CURRENT_BINARY_DIR}/opencv_cblas.hpp)
-                           COMMENT "Adding proxy cblas.h header")
+        set(_include_str "\#include \"${CBLAS_H_PATH}\"")
+        file(WRITE ${CBLAS_H_PROXY_PATH} ${_include_str})
     endif()
 endif()

modules/dnn/include/opencv2/dnn/all_layers.hpp

@@ -101,8 +101,8 @@ namespace dnn
         @f$W_{x?} \in R^{N_c \times N_x}@f$, @f$W_h? \in R^{N_c \times N_h}@f$, @f$b_? \in R^{N_c}@f$.
 
         For simplicity and performance purposes we use @f$ W_x = [W_{xi}; W_{xf}; W_{xo}, W_{xg}] @f$
-        (i.e. @f$W_x@f$ is vertical contacentaion of @f$ W_{x?} @f$), @f$ W_x \in R^{4N_c x N_x} @f$.
+        (i.e. @f$W_x@f$ is vertical contacentaion of @f$ W_{x?} @f$), @f$ W_x \in R^{4N_c \times N_x} @f$.
-        The same for @f$ W_h = [W_{hi}; W_{hf}; W_{ho}, W_{hg}], W_h \in R^{4N_c x N_h} @f$
+        The same for @f$ W_h = [W_{hi}; W_{hf}; W_{ho}, W_{hg}], W_h \in R^{4N_c \times N_h} @f$
         and for @f$ b = [b_i; b_f, b_o, b_g]@f$, @f$b \in R^{4N_c} @f$.
 
         @param Wh is matrix defining how previous output is transformed to internal gates (i.e. according to abovemtioned notation is @f$ W_h @f$)
@@ -111,16 +111,44 @@ namespace dnn
         */
         virtual void setWeights(const Blob &Wh, const Blob &Wx, const Blob &b) = 0;
 
-        /** In common case it uses three inputs (@f$x_t@f$, @f$h_{t-1}@f$ and @f$c_{t-1}@f$) to compute compute two outputs (@f$h_t@f$ and @f$c_t@f$).
+        /** @brief Set @f$ h_{t-1} @f$ value that will be used in next forward() calls.
-
+          * @details By-default @f$ h_{t-1} @f$ is inited by zeros and updated after each forward() call.
-        @param input could contain three inputs: @f$x_t@f$, @f$h_{t-1}@f$ and @f$c_{t-1}@f$.
+          */
-        @param output contains computed outputs: @f$h_t@f$ and @f$c_t@f$.
+        virtual void setH(const Blob &H) = 0;
-
+        /** @brief Returns current @f$ h_{t-1} @f$ value (deep copy). */
-        The first input @f$x_t@f$ is required.
+        virtual Blob getH() const = 0;
-        The second and third inputs are optional: if they weren't set than layer will use internal @f$h_{t-1}@f$ and @f$c_{t-1}@f$ from previous calls,
+
-        but at the first call they will be filled by zeros.
+        /** @brief Set @f$ c_{t-1} @f$ value that will be used in next forward() calls.
-        Size of the last dimension of @f$x_t@f$ must be @f$N_x@f$, (@f$N_h@f$ for @f$h_{t-1}@f$ and @f$N_c@f$ for @f$c_{t-1}@f$).
+          * @details By-default @f$ c_{t-1} @f$ is inited by zeros and updated after each forward() call.
-        Sizes of remainder dimensions could be any, but thay must be consistent among @f$x_t@f$, @f$h_{t-1}@f$ and @f$c_{t-1}@f$.
+          */
+        virtual void setC(const Blob &C) = 0;
+        /** @brief Returns current @f$ c_{t-1} @f$ value (deep copy). */
+        virtual Blob getC() const = 0;
+
+        /** @brief Specifies either interpet first dimension of input blob as timestamp dimenion either as sample.
+          *
+          * If flag is set to true then shape of input blob will be interpeted as [`T`, `N`, `[data dims]`] where `T` specifies number of timpestamps, `N` is number of independent streams.
+          * In this case each forward() call will iterate through `T` timestamps and update layer's state `T` times.
+          *
+          * If flag is set to false then shape of input blob will be interpeted as [`N`, `[data dims]`].
+          * In this case each forward() call will make one iteration and produce one timestamp with shape [`N`, `[out dims]`].
+          */
+        virtual void setUseTimstampsDim(bool use = true) = 0;
+
+        /** @brief If this flag is set to true then layer will produce @f$ c_t @f$ as second output.
+         * @details Shape of the second output is the same as first output.
+         */
+        virtual void setProduceCellOutput(bool produce = false) = 0;
+
+        /** In common case it use single input with @f$x_t@f$ values to compute output(s) @f$h_t@f$ (and @f$c_t@f$).
+         * @param input should contain packed values @f$x_t@f$
+         * @param output contains computed outputs: @f$h_t@f$ (and @f$c_t@f$ if setProduceCellOutput() flag was set to true).
+         *
+         * If setUseTimstampsDim() is set to true then @p input[0] should has at least two dimensions with the following shape: [`T`, `N`, `[data dims]`],
+         * where `T` specifies number of timpestamps, `N` is number of independent streams (i.e. x_{t_0 + t}^{stream} is @p input[0][t, stream, ...]).
+         *
+         * If setUseTimstampsDim() is set to fase then @p input[0] should contain single timestamp, its shape should has form [`N`, `[data dims]`] with at least one dimension.
+         * (i.e. x_{t}^{stream} = @p input[0][stream, ...]).
         */
         void forward(std::vector<Blob*> &input, std::vector<Blob> &output);
     };

modules/dnn/include/opencv2/dnn/blob.hpp

@@ -107,6 +107,18 @@ namespace dnn
 
         bool operator== (const BlobShape &r) const;
 
+        /** @brief Contacenates two shapes */
+        BlobShape operator+ (const BlobShape &r) const;
+
+        /** @brief Returns shape of passed Mat. */
+        static BlobShape like(const Mat &m);
+        /** @brief Returns shape of passed Mat. */
+        static BlobShape like(const UMat &m);
+
+#ifdef CV_CXX_MOVE_SEMANTICS
+        //TBD
+#endif
+
     private:
         cv::AutoBuffer<int,4> sz;
     };
@@ -228,6 +240,11 @@ namespace dnn
          */
         Blob &reshape(const BlobShape &shape);
 
+        /** @brief Changes shape of the blob without copying the data.
+         * @returns shallow copy of original blob with new shape.
+         */
+        Blob reshaped(const BlobShape &newShape) const;
+
         /** @brief Returns type of the blob. */
         int type() const;
 

modules/dnn/include/opencv2/dnn/blob.inl.hpp

@@ -185,6 +185,16 @@ inline bool BlobShape::operator==(const BlobShape &r) const
     return this->equal(r);
 }
 
+inline BlobShape BlobShape::like(const Mat &m)
+{
+    return BlobShape(m.dims, (const int*)m.size);
+}
+
+inline BlobShape BlobShape::like(const UMat &m)
+{
+    return BlobShape(m.dims, (const int*)m.size);
+}
+
 CV_EXPORTS std::ostream &operator<< (std::ostream &stream, const BlobShape &shape);
 
 /////////////////////////////////////////////////////////////////////
@@ -277,6 +287,17 @@ inline BlobShape Blob::shape() const
     return BlobShape(dims(), sizes());
 }
 
+
+inline BlobShape BlobShape::operator+(const BlobShape &r) const
+{
+    BlobShape newShape(this->dims() + r.dims(), (int*)NULL);
+    for (int i = 0; i < this->dims(); i++)
+        newShape[i] = (*this)[i];
+    for (int i = 0; i < r.dims(); i++)
+        newShape[this->dims() + i] = r[i];
+    return newShape;
+}
+
 inline bool Blob::equalShape(const Blob &other) const
 {
     if (this->dims() != other.dims())
@@ -366,6 +387,13 @@ inline Blob &Blob::reshape(const BlobShape &newShape)
     return *this;
 }
 
+inline Blob Blob::reshaped(const BlobShape &newShape) const
+{
+    Blob res(*this); //also, res.shareFrom(*this) could be used
+    res.reshape(newShape);
+    return res;
+}
+
 }
 }
 

modules/dnn/src/layers/op_blas.cpp

@@ -1,9 +1,11 @@
 #include "op_blas.hpp"
 
 #if HAVE_CBLAS
-#include "cblas.h"
+#include "opencv_cblas.hpp"
 #endif
 
+#include <iostream>
+
 namespace cv
 {
 namespace dnn
@@ -14,18 +16,19 @@ void gemm(InputArray A, InputArray B, double alpha, InputOutputArray C, double b
     cv::gemm(A, B, alpha, C, beta, C, flags);
 }
 
-inline void SwapRowCols(const Mat &A, int &rows, int &cols, bool transA)
+inline void SwapRowCols(const Mat &A, int &rows, int &cols, bool isTrans)
 {
-    rows = (transA) ? A.cols : A.rows;
+    CV_DbgAssert(A.dims == 2);
-    cols = (transA) ? A.rows : A.cols;
+    rows = (isTrans) ? A.cols : A.rows;
+    cols = (isTrans) ? A.rows : A.cols;
 }
 
 void gemmCPU(const Mat &A, const Mat &B, double alpha, Mat &C, double beta, int flags /*= 0*/)
 {
     #if HAVE_CBLAS
-    int transA = flags & GEMM_1_T;
+    bool transA = static_cast<bool>(flags & GEMM_1_T);
-    int transB = flags & GEMM_2_T;
+    bool transB = static_cast<bool>(flags & GEMM_2_T);
-    int transC = flags & GEMM_3_T;
+    bool transC = static_cast<bool>(flags & GEMM_3_T);
 
     int Arows, Acols, Brows, Bcols, Crows, Ccols;
     SwapRowCols(A, Arows, Acols, transA);
@@ -34,9 +37,9 @@ void gemmCPU(const Mat &A, const Mat &B, double alpha, Mat &C, double beta, int
 
     CV_DbgAssert(!(flags & GEMM_3_T));
     CV_Assert(Acols == Brows && Arows == Crows && Bcols == Ccols);
-    CV_DbgAssert(A.isContinuous() && B.isContinuous() && C.isContinuous());
+    CV_Assert(A.isContinuous() && B.isContinuous() && C.isContinuous());
-    CV_DbgAssert(A.type() == CV_32F || A.type() == CV_64F);
+    CV_Assert(A.type() == CV_32F || A.type() == CV_64F);
-    CV_DbgAssert(A.type() == B.type() && B.type() == C.type());
+    CV_Assert(A.type() == B.type() && B.type() == C.type());
 
     if (C.type() == CV_32F)
     {

modules/dnn/src/layers/recurrent_layers.cpp

@@ -43,35 +43,118 @@
 #include "recurrent_layers.hpp"
 #include "op_blas.hpp"
 #include <iostream>
+#include <cmath>
 
 namespace cv
 {
 namespace dnn
 {
 
+template<typename Dtype>
+static void tanh(const Mat &src, Mat &dst)
+{
+    MatConstIterator_<Dtype> itSrc = src.begin<Dtype>();
+    MatIterator_<Dtype> itDst = dst.begin<Dtype>();
+
+    for (; itSrc != src.end<Dtype>(); itSrc++, itDst++)
+        *itDst = std::tanh(*itSrc);
+}
+
+static void tanh(const Mat &src, Mat &dst)
+{
+    dst.create(src.dims, (const int*)src.size, src.type());
+
+    if (src.type() == CV_32F)
+        tanh<float>(src, dst);
+    else if (src.type() == CV_64F)
+        tanh<double>(src, dst);
+    else
+        CV_Error(Error::StsUnsupportedFormat, "Functions supports only floating point types");
+}
+
+static void sigmoid(const Mat &src, Mat &dst)
+{
+    cv::exp(-src, dst);
+    cv::pow(1 + dst, -1, dst);
+}
+
 class LSTMLayerImpl : public LSTMLayer
 {
+    int numOut, numTimeStamps, numSamples, numInp;
+    Mat hInternal, cInternal;
+    Mat gates, dummyOnes;
+    int dtype;
+    bool allocated;
+
+    bool useTimestampDim;
+    bool produceCellOutput;
+
 public:
 
     LSTMLayerImpl()
     {
         type = "LSTM";
+        useTimestampDim = true;
+        produceCellOutput = false;
+        allocated = false;
     }
 
-    int nH, nX, nC, numSamples;
+    void setUseTimstampsDim(bool use)
-    Mat prevH, prevC;
+    {
-    Mat gates, dummyOnes;
+        CV_Assert(!allocated);
+        useTimestampDim = use;
+    }
+
+    void setProduceCellOutput(bool produce)
+    {
+        CV_Assert(!allocated);
+        produceCellOutput = produce;
+    }
+
+    void setC(const Blob &C)
+    {
+        CV_Assert(!allocated || C.total() == cInternal.total());
+        C.matRefConst().copyTo(cInternal);
+    }
+
+    void setH(const Blob &H)
+    {
+        CV_Assert(!allocated || H.total() == hInternal.total());
+        H.matRefConst().copyTo(hInternal);
+    }
+
+    Blob getC() const
+    {
+        CV_Assert(!cInternal.empty());
+
+        //TODO: add convinient Mat -> Blob constructor
+        Blob res;
+        res.fill(BlobShape::like(cInternal), cInternal.type(), cInternal.data);
+        return res;
+    }
+
+    Blob getH() const
+    {
+        CV_Assert(!hInternal.empty());
+
+        Blob res;
+        res.fill(BlobShape::like(hInternal), hInternal.type(), hInternal.data);
+        return res;
+    }
 
     void setWeights(const Blob &Wh, const Blob &Wx, const Blob &bias)
     {
         CV_Assert(Wh.dims() == 2 && Wx.dims() == 2);
-        CV_Assert(Wh.size(0) == Wx.size(0) && Wh.size(0) % 4 == 0);
+        CV_Assert(Wh.size(0) == Wx.size(0));
+        CV_Assert(Wh.size(0) == 4*Wh.size(1));
         CV_Assert(Wh.size(0) == (int)bias.total());
+        CV_Assert(Wh.type() == Wx.type() && Wx.type() == bias.type());
 
         blobs.resize(3);
         blobs[0] = Wh;
         blobs[1] = Wx;
         blobs[2] = bias;
+        blobs[2].reshape(BlobShape(1, (int)bias.total()));
     }
 
     void allocate(const std::vector<Blob*> &input, std::vector<Blob> &output)
@@ -79,103 +162,83 @@ public:
         CV_Assert(blobs.size() == 3);
         Blob &Wh = blobs[0], &Wx = blobs[1];
 
-        nH = Wh.size(1);
+        numOut = Wh.size(1);
-        nX = Wx.size(1);
+        numInp = Wx.size(1);
-        nC = Wh.size(0) / 4;
 
-        CV_Assert(input.size() >= 1 && input.size() <= 3);
+        CV_Assert(input.size() == 1);
-        CV_Assert(input[0]->size(-1) == nX);
+        CV_Assert(input[0]->dims() > 2 && (int)input[0]->total(2) == numInp);
 
-        BlobShape inpShape = input[0]->shape();
+        numTimeStamps = input[0]->size(0);
-        numSamples = input[0]->total(0, input[0]->dims()-1);
+        numSamples = input[0]->size(1);
+        dtype = input[0]->type();
 
-        BlobShape hShape = inpShape;
+        CV_Assert(dtype == CV_32F || dtype == CV_64F);
-        hShape[-1] = nH;
+        CV_Assert(Wh.type() == dtype);
-        BlobShape cShape = inpShape;
-        cShape[-1] = nC;
 
+        BlobShape outShape(numTimeStamps, numSamples, numOut);
         output.resize(2);
-        output[0].create(hShape, input[0]->type());
+        output[0].create(outShape, dtype);
-        output[1].create(cShape, input[0]->type());
+        output[1].create(outShape, dtype);
 
-        if (input.size() < 2)
+        hInternal.create(numSamples, numOut, dtype);
-        {
+        hInternal.setTo(0);
-            prevH.create(numSamples, nH, input[0]->type());
-            prevH.setTo(0);
-        }
-        else
-            CV_Assert(input[1]->shape() == hShape);
 
-        if (input.size() < 3)
+        cInternal.create(numSamples, numOut, dtype);
-        {
+        cInternal.setTo(0);
-            prevC.create(numSamples, nC, input[0]->type());
-            prevC.setTo(0);
-        }
-        else
-            CV_Assert(input[2]->shape() == cShape);
 
-        gates.create(numSamples, 4*nC, input[0]->type());
+        gates.create(numSamples, 4*numOut, dtype);
-        dummyOnes.create(numSamples, 1, input[0]->type());
-        dummyOnes.setTo(1);
-    }
 
-    Mat ep, em;
+        dummyOnes.create(numSamples, 1, dtype);
-    void tanh(Mat &x, Mat &d)
+        dummyOnes.setTo(1);
-    {
-        //TODO: two exp() is bad idea
-        cv::exp(-x, em);
-        cv::exp( x, ep);
-        cv::divide(ep - em, ep + em, d);
-    }
 
-    void sigmoid(Mat &x)
+        allocated = true;
-    {
-        cv::exp(-x, x);
-        cv::pow(1 + x, -1, x);
     }
 
     void forward(std::vector<Blob*> &input, std::vector<Blob> &output)
     {
-        CV_DbgAssert(blobs.size() == 3);
+        const Mat &Wh = blobs[0].matRefConst();
-        const Mat &Wh = blobs[0].matRefConst(), &Wx = blobs[1].matRefConst();
+        const Mat &Wx = blobs[1].matRefConst();
-        Mat bias = blobs[2].matRefConst().reshape(1, 1);
+        const Mat &bias = blobs[2].matRefConst();
-        CV_DbgAssert(Wh.type() == CV_32F && Wx.type() == CV_32F && bias.type() == CV_32F);
+
-
+        int numSamplesTotal = numTimeStamps*numSamples;
-        int szx[] = { numSamples, nX };
+        Mat xTs = input[0]->reshaped(BlobShape(numSamplesTotal, numInp)).matRefConst();
-        int szc[] = { numSamples, nC };
+
-        Mat xCurr = input[0]->matRefConst().reshape(1, 2, szx);
+        BlobShape outMatShape(numSamplesTotal, numOut);
-        Mat hPrev = (input.size() >= 2) ? input[1]->matRefConst().reshape(1, 2, szc) : prevH;
+        Mat hOutTs = output[0].reshaped(outMatShape).matRef();
-        Mat cPrev = (input.size() >= 3) ? input[2]->matRefConst().reshape(1, 2, szc) : prevC;
+        Mat cOutTs = (produceCellOutput) ? output[1].reshaped(outMatShape).matRef() : Mat();
-        CV_Assert(xCurr.type() == CV_32F && hPrev.type() == CV_32F && cPrev.type() == CV_32F);
+
-
+        for (int ts = 0; ts < numTimeStamps; ts++)
-        Mat hCurr = output[0].matRef().reshape(1, 2, szc);
+        {
-        Mat cCurr = output[1].matRef().reshape(1, 2, szc);
+            Range curRowRange(ts*numSamples, (ts + 1)*numSamples);
-        CV_Assert(hCurr.type() == CV_32F && cCurr.type() == CV_32F);
+            Mat xCurr = xTs.rowRange(curRowRange);
-
+
-        gemmCPU(xCurr, Wx, 1, gates, 0, GEMM_2_T); // Wx * x_t
+            gemmCPU(xCurr, Wx, 1, gates, 0, GEMM_2_T);      // Wx * x_t
-        gemmCPU(hPrev, Wh, 1, gates, 1, GEMM_2_T); //+Wh * h_{t-1}
+            gemmCPU(hInternal, Wh, 1, gates, 1, GEMM_2_T);  //+Wh * h_{t-1}
-        gemmCPU(dummyOnes, bias, 1, gates, 1);     //+b
+            gemmCPU(dummyOnes, bias, 1, gates, 1);          //+b
-
+
-        Mat gatesDiv = gates.reshape(1, 4*numSamples);
+            Mat getesIFO = gates.colRange(0, 3*numOut);
-        Mat getesIFO = gatesDiv(Range(0, 3*numSamples), Range::all());
+            Mat gateI = gates.colRange(0*numOut, 1*numOut);
-        Mat gateI = gatesDiv(Range(0*numSamples, 1*numSamples), Range::all());
+            Mat gateF = gates.colRange(1*numOut, 2*numOut);
-        Mat gateF = gatesDiv(Range(1*numSamples, 2*numSamples), Range::all());
+            Mat gateO = gates.colRange(2*numOut, 3*numOut);
-        Mat gateO = gatesDiv(Range(2*numSamples, 3*numSamples), Range::all());
+            Mat gateG = gates.colRange(3*numOut, 4*numOut);
-        Mat gateG = gatesDiv(Range(3*numSamples, 4*numSamples), Range::all());
+
-
+            sigmoid(getesIFO, getesIFO);
-        sigmoid(getesIFO);
+            tanh(gateG, gateG);
-        tanh(gateG, gateG);
+
-
+            //compute c_t
-        cv::add(gateF.mul(cPrev), gateI.mul(gateG), cCurr);
+            cv::multiply(gateF, cInternal, gateF);  // f_t (*) c_{t-1}
-
+            cv::multiply(gateI, gateG, gateI);      // i_t (*) g_t
-        tanh(cCurr, hCurr);
+            cv::add(gateF, gateI, cInternal);       // c_t = f_t (*) c_{t-1} + i_t (*) g_t
-        cv::multiply(gateO, hCurr, hCurr);
+
-
+            //compute h_t
-        //save answers for next iteration
+            tanh(cInternal, hInternal);
-        if (input.size() <= 2)
+            cv::multiply(gateO, hInternal, hInternal);
-            hCurr.copyTo(hPrev);
+
-        if (input.size() <= 3)
+            //save results in output blobs
-            cCurr.copyTo(cPrev);
+            hInternal.copyTo(hOutTs.rowRange(curRowRange));
+            if (produceCellOutput)
+                cInternal.copyTo(cOutTs.rowRange(curRowRange));
+        }
     }
 };
 
@@ -193,6 +256,7 @@ void LSTMLayer::forward(std::vector<Blob*>&, std::vector<Blob>&)
 class RNNLayerImpl : public RNNLayer
 {
     int nX, nH, nO, nSamples;
+    int dtype;
     Mat Whh, Wxh, bh;
     Mat Who, bo;
     Mat hPrevInternal, dummyBiasOnes;
@@ -210,7 +274,6 @@ public:
         CV_Assert(W_hh.size(0) == W_xh.size(0) && W_hh.size(0) == W_hh.size(1) && (int)b_h.total() == W_xh.size(0));
         CV_Assert(W_ho.size(0) == (int)b_o.total());
         CV_Assert(W_ho.size(1) == W_hh.size(1));
-        //TODO: Check type
 
         blobs.resize(5);
         blobs[0] = W_hh;
@@ -262,16 +325,6 @@ public:
         bo = bo.reshape(1, 1); //is 1 x nO mat
     }
 
-    //in-place tanh function
-    static void tanh(Mat &x) // 2 / (1 + e^(-2x)) - 1
-    {
-        x.convertTo(x, x.type(), -2);   // -2x
-        cv::exp(x, x);                  // e^(-2x)
-        x.convertTo(x, x.type(), 1, 1); // 1 + e^(-2x)
-        cv::pow(x, -1, x);              // 1 / (1 + e^(-2x))
-        x.convertTo(x, x.type(), 2, -1);// 2 / (1 + e^(-2x)) - 1
-    }
-
     void forward(std::vector<Blob*> &input, std::vector<Blob> &output)
     {
         Mat xCurr = input[0]->matRefConst();
@@ -291,12 +344,12 @@ public:
 
         gemmCPU(hPrev, Whh, 1, hCurr, 0, GEMM_2_T); // W_{hh} * h_{prev}
         gemmCPU(xCurr, Wxh, 1, hCurr, 1, GEMM_2_T); //+W_{xh} * x_{curr}
-        gemmCPU(dummyBiasOnes, bh, 1, hCurr, 1);      //+bh
+        gemmCPU(dummyBiasOnes, bh, 1, hCurr, 1);    //+bh
-        tanh(hCurr);
+        tanh(hCurr, hCurr);
 
         gemmCPU(hPrev, Who, 1, oCurr, 0, GEMM_2_T); // W_{ho} * h_{prev}
-        gemmCPU(dummyBiasOnes, bo, 1, oCurr, 1);      //+b_o
+        gemmCPU(dummyBiasOnes, bo, 1, oCurr, 1);    //+b_o
-        tanh(oCurr);
+        tanh(oCurr, oCurr);
 
         if (input.size() < 2) //save h_{prev}
             hCurr.copyTo(hPrevInternal);
@@ -314,4 +367,4 @@ CV_EXPORTS_W Ptr<RNNLayer> RNNLayer::create()
 }
 
 }
-}
+}

modules/dnn/test/test_layers.cpp

@@ -177,6 +177,23 @@ TEST(Layer_Test_Reshape_Split_Slice, Accuracy)
     normAssert(input, output);
 }
 
+enum RunLayerMode
+{
+    ALLOC_ONLY          = 1,
+    FORWARD_ONLY        = 2,
+    ALLOC_AND_FORWARD   = 3
+};
+
+void runLayer(Ptr<Layer> layer, std::vector<Blob> &inpBlobs, std::vector<Blob> &outBlobs, int mode=ALLOC_AND_FORWARD)
+{
+    std::vector<Blob*> inpPtrs(inpBlobs.size());
+    for (size_t i = 0; i < inpBlobs.size(); i++)
+        inpPtrs[i] = &inpBlobs[i];
+
+    if (mode & ALLOC_ONLY) layer->allocate(inpPtrs, outBlobs);
+    if (mode & FORWARD_ONLY) layer->forward(inpPtrs, outBlobs);
+}
+
 class Layer_LSTM_Test : public ::testing::Test
 {
 public:
@@ -233,6 +250,28 @@ TEST_F(Layer_LSTM_Test, BasicTest_2)
     EXPECT_EQ(outputs[1].shape(), BlobShape(1, 2, 3, Nc));
 }
 
+TEST(Layer_LSTM_Test_Accuracy_Reference_with_, CaffeRecurrent)
+{
+
+    Ptr<LSTMLayer> layer = LSTMLayer::create();
+
+    Blob Wx = blobFromNPY(_tf("lstm.prototxt.w_0.npy"));
+    Blob Wh = blobFromNPY(_tf("lstm.prototxt.w_2.npy"));
+    Blob b  = blobFromNPY(_tf("lstm.prototxt.w_1.npy"));
+    layer->setWeights(Wh, Wx, b);
+
+    Blob inp = blobFromNPY(_tf("blob.npy"));
+    std::vector<Blob> inputs(1, inp), outputs;
+    runLayer(layer, inputs, outputs, ALLOC_ONLY | FORWARD_ONLY);
+
+    Blob &h_t_gathered = outputs[0];
+    Blob h_t_reference = blobFromNPY(_tf("lstm.prototxt.h_1.npy"));
+
+    //h_t_gathered.reshape(h_t_reference.shape());
+
+    normAssert(h_t_reference, h_t_gathered);
+}
+
 
 class Layer_RNN_Test : public ::testing::Test
 {