opencv/sources/modules/dnn/src/layers/recurrent_layers.cpp

/*M///////////////////////////////////////////////////////////////////////////////////////

//

//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.

//

//  By downloading, copying, installing or using the software you agree to this license.

//  If you do not agree to this license, do not download, install,

//  copy or use the software.

//

//

//                           License Agreement

//                For Open Source Computer Vision Library

//

// Copyright (C) 2013, OpenCV Foundation, all rights reserved.

// Copyright (C) 2017, Intel Corporation, all rights reserved.

// Third party copyrights are property of their respective owners.

//

// Redistribution and use in source and binary forms, with or without modification,

// are permitted provided that the following conditions are met:

//

//   * Redistribution's of source code must retain the above copyright notice,

//     this list of conditions and the following disclaimer.

//

//   * Redistribution's in binary form must reproduce the above copyright notice,

//     this list of conditions and the following disclaimer in the documentation

//     and/or other materials provided with the distribution.

//

//   * The name of the copyright holders may not be used to endorse or promote products

//     derived from this software without specific prior written permission.

//

// This software is provided by the copyright holders and contributors "as is" and

// any express or implied warranties, including, but not limited to, the implied

// warranties of merchantability and fitness for a particular purpose are disclaimed.

// In no event shall the Intel Corporation or contributors be liable for any direct,

// indirect, incidental, special, exemplary, or consequential damages

// (including, but not limited to, procurement of substitute goods or services;

// loss of use, data, or profits; or business interruption) however caused

// and on any theory of liability, whether in contract, strict liability,

// or tort (including negligence or otherwise) arising in any way out of

// the use of this software, even if advised of the possibility of such damage.

//

//M*/

#include "../precomp.hpp"
#include <iostream>
#include <cmath>
#include <opencv2/dnn/shape_utils.hpp>

#ifdef HAVE_CUDA
#include "../cuda4dnn/primitives/recurrent_cells.hpp"
using namespace cv::dnn::cuda4dnn;
#endif

#include "layers_common.hpp"

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);
}

//TODO: make utils method
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, "Function supports only floating point types");
}

static void sigmoid(const Mat &src, Mat &dst)

{
    cv::exp(-src, dst);
    cv::pow(1 + dst, -1, dst);
}

typedef void (*ActivationFunction)(const Mat &src, Mat &dst);
static ActivationFunction get_activation_function(const String& activation) {
    // most used activations for PyTorch and TF : Tanh, Sigmoid
    // if you need to support more optional activations use std::map instead
    if (activation == "Tanh")
    {
        return tanh;
    }
    else if (activation == "Sigmoid")
    {
        return sigmoid;
    }
    else
    {
        CV_Error(Error::StsNotImplemented,
                 cv::format("Activation function [%s] for layer LSTM  is not supported", activation.c_str()));
    }
}

class LSTMLayerImpl CV_FINAL : public LSTMLayer
{
    int numTimeStamps, numSamples, numHidden;
    bool allocated;

    MatShape outTailShape;  //shape of single output sample
    MatShape outTsShape;    //shape of N output samples

    enum layout_t : int {
        SEQ_BATCH_HID = 0,
        BATCH_SEQ_HID = 1
    };

    bool useTimestampDim;
    bool produceCellOutput;
    float forgetBias, cellClip;
    bool useCellClip, usePeephole;
    bool reverse;   // If true, go in negative direction along the time axis
    bool bidirectional;  // If true, produces both forward and reversed directions along time axis
    layout_t layout;  // If layout == BATCH_SEQ_HID, uses batch_size x seq_length x num_hidden for input and output
                      // else uses seq_length x batch_size x num_hidden

    ActivationFunction f_activation;
    ActivationFunction g_activation;
    ActivationFunction h_activation;
    bool isDefaultActivations{true};

#if CV_TRY_AVX
    bool useAVX;
#endif
#if CV_TRY_AVX2
    bool useAVX2;
#endif

    // CUDA needs input blobs to be rearranged in a specific way, but some transformations
    // in ONNXImporter are destructive, so we keep a copy.
    std::vector<Mat> originalBlobs;

public:

    LSTMLayerImpl(const LayerParams& params)
        : numTimeStamps(0), numSamples(0)
#if CV_TRY_AVX
          , useAVX(checkHardwareSupport(CPU_AVX))
#endif
#if CV_TRY_AVX2
          , useAVX2(checkHardwareSupport(CPU_AVX2))
#endif
    {
        setParamsFrom(params);

        if (params.get<bool>("is_onnx", false))
        {
            // collect copies of onnx blobs
            originalBlobs.insert(originalBlobs.begin(), blobs.begin(), blobs.begin() + 3);
            blobs.erase(blobs.begin(), blobs.begin() + 3);
        }

        bidirectional = params.get<bool>("bidirectional", false);
        if (!blobs.empty())
        {
            CV_Assert(blobs.size() >= 3);

            blobs[2] = blobs[2].reshape(1, 1);

            const Mat& Wh = blobs[0];
            const Mat& Wx = blobs[1];
            const Mat& bias = blobs[2];
            const Mat& hInternal = blobs[3];
            const Mat& cInternal = blobs[4];
            CV_CheckEQ(Wh.dims, 2, "");
            CV_CheckEQ(Wx.dims, 2, "");
            CV_CheckEQ(Wh.rows, Wx.rows, "");
            CV_CheckEQ(Wh.rows, (1 + static_cast<int>(bidirectional))*4*Wh.cols, "");
            CV_CheckEQ(Wh.rows, (int)bias.total(), "");
            // Only perform these checks if hInternal and cInternal are not empty matrices
            // e.g. inputs are not given by a user
            if(!hInternal.empty()){
                CV_CheckEQ(hInternal.cols, Wh.cols, "");
            }
            if(!cInternal.empty()){
                CV_CheckEQ(cInternal.cols, Wh.cols, "");
            }
            if (!hInternal.empty() && !cInternal.empty()){ //otherwise check in forward
                CV_CheckEQ(hInternal.rows, cInternal.rows, "");
            }
            CV_Assert(Wh.type() == Wx.type() && Wx.type() == bias.type());

            // Peephole weights.
            if (blobs.size() > 5)
            {
                CV_Assert(blobs.size() == 8);
                const int N = Wh.cols;
                for (int i = 5; i < 8; ++i)
                {
                    CV_Assert(blobs[i].rows == N && blobs[i].cols == N);
                    CV_Assert(blobs[i].type() == bias.type());
                }
            }
        }
        layout = (layout_t) params.get<int>("layout", SEQ_BATCH_HID);
        useTimestampDim = params.get<bool>("use_timestamp_dim", true);
        produceCellOutput = params.get<bool>("produce_cell_output", false);
        forgetBias = params.get<float>("forget_bias", 0.0f);
        cellClip = params.get<float>("cell_clip", 0.0f);
        useCellClip = params.get<bool>("use_cell_clip", false);
        usePeephole = params.get<bool>("use_peephole", false);
        reverse = params.get<bool>("reverse", false);
        numHidden = params.get<int>("hidden_size", 1);
        CV_Assert(!reverse || !bidirectional);

        // read activations
        DictValue activations = params.get<DictValue>("activations", DictValue(String()));
        if (activations.size() == 1) // if activations wasn't specified use default
        {
            f_activation = sigmoid;
            g_activation = tanh;
            h_activation = tanh;
            isDefaultActivations = true;
        } else {
            CV_Assert(activations.size() == 3);
            f_activation = get_activation_function(activations.getStringValue(0));
            g_activation = get_activation_function(activations.getStringValue(1));
            h_activation = get_activation_function(activations.getStringValue(2));
            isDefaultActivations = activations.getStringValue(0) == "Sigmoid"
                                   && activations.getStringValue(1) == "Tanh"
                                   && activations.getStringValue(2) == "Tanh";
        }

        allocated = false;
        outTailShape.clear();
    }

    void setUseTimstampsDim(bool use) CV_OVERRIDE

    {
        CV_Assert(!allocated);
        useTimestampDim = use;
    }

    void setProduceCellOutput(bool produce) CV_OVERRIDE

    {
        CV_Assert(!allocated);
        produceCellOutput = produce;
    }

    void setOutShape(const MatShape &outTailShape_) CV_OVERRIDE

    {
        CV_Assert(!allocated || total(outTailShape) == total(outTailShape_));
        outTailShape = outTailShape_;
    }

    void setWeights(const Mat &Wh, const Mat &Wx, const Mat &bias) CV_OVERRIDE

    {
        CV_Assert(Wh.dims == 2 && Wx.dims == 2);
        CV_Assert(Wh.rows == Wx.rows);
        CV_Assert(Wh.rows == 4*Wh.cols);
        CV_Assert(Wh.rows == (int)bias.total());
        CV_Assert(Wh.type() == Wx.type() && Wx.type() == bias.type());

        blobs.resize(3);
        blobs[0] = Mat(Wh.clone());
        blobs[1] = Mat(Wx.clone());
        blobs[2] = Mat(bias.clone()).reshape(1, 1);
    }

    bool supportBackend(int backendId) CV_OVERRIDE

    {
        return backendId == DNN_BACKEND_OPENCV
               || (backendId == DNN_BACKEND_CUDA && isDefaultActivations && !reverse && !usePeephole);
    }

    bool getMemoryShapes(const std::vector<MatShape> &inputs,

                         const int requiredOutputs,

                         std::vector<MatShape> &outputs,

                         std::vector<MatShape> &internals) const CV_OVERRIDE

    {
        CV_Assert((!usePeephole && blobs.size() == 5) || (usePeephole && blobs.size() == 8));
        CV_Assert((inputs.size() == 1 || inputs.size() == 3));
        const MatShape& inp0 = inputs[0];

        const Mat &Wh = blobs[0], &Wx = blobs[1];
        int _numOut = Wh.size[1];
        int _numInp = Wx.size[1];
        MatShape outTailShape_(outTailShape), outResShape;

        if (!outTailShape_.empty())
            CV_Assert(total(outTailShape_) == _numOut);
        else
            outTailShape_.assign(1, _numOut);

        int _numSamples;
        if (useTimestampDim)
        {
            CV_Assert(inp0.size() >= 2 && total(inp0, 2) == _numInp);
            if (layout == SEQ_BATCH_HID) {
                _numSamples = inp0[1];
                outResShape.push_back(inp0[0]);
            } else {
                _numSamples = inp0[0];
                outResShape.push_back(inp0[1]);
            }
        }
        else
        {
            CV_Assert(inp0.size() >= 2 && total(inp0, 1) == _numInp);
            _numSamples = inp0[0];
        }

        outResShape.push_back(_numSamples);
        outResShape.insert(outResShape.end(), outTailShape_.begin(), outTailShape_.end());
        outResShape.back() *= (1 + static_cast<int>(bidirectional));

        outputs.assign(1, outResShape);
        if (produceCellOutput)
        {
            // the producer is ONNX, so CellState is different
            if (!originalBlobs.empty())
            {
                int shp[] = {(1 + static_cast<int>(bidirectional)), _numSamples, numHidden};
                MatShape newShape(shp, shp + sizeof(shp)/sizeof(shp[0]));
                outputs.push_back(newShape);
            }
            else
            {
                outputs.push_back(outResShape);
            }
        }

        internals.assign(1, shape(_numSamples, _numOut)); // hInternal
        internals.push_back(shape(_numSamples, _numOut)); // cInternal
        internals.push_back(shape(_numSamples, 1)); // dummyOnes
        internals.push_back(shape(_numSamples, 4*_numOut)); // gates

        return false;
    }

    void finalize(InputArrayOfArrays inputs_arr, OutputArrayOfArrays) CV_OVERRIDE

    {
        std::vector<Mat> input;
        inputs_arr.getMatVector(input);

        CV_Assert((!usePeephole && blobs.size() == 5) || (usePeephole && blobs.size() == 8));
        CV_Assert((input.size() == 1 || input.size() == 3));
        const Mat& inp0 = input[0];

        Mat &Wh = blobs[0], &Wx = blobs[1];
        int numOut = Wh.size[1];
        int numInp = Wx.size[1];

        if (!outTailShape.empty())
            CV_Assert(total(outTailShape) == numOut);
        else
            outTailShape.assign(1, numOut);

        if (useTimestampDim)
        {
            CV_Assert(inp0.dims >= 2 && (int)inp0.total(2) == numInp);
            if (layout == SEQ_BATCH_HID){
                numTimeStamps = inp0.size[0];
                numSamples = inp0.size[1];
            }else{
                numTimeStamps = inp0.size[1];
                numSamples = inp0.size[0];
            }
        }
        else
        {
            CV_Assert(inp0.dims >= 2 && (int)inp0.total(1) == numInp);
            numTimeStamps = 1;
            numSamples = inp0.size[0];
        }

        outTsShape.clear();
        outTsShape.push_back(numSamples);
        outTsShape.insert(outTsShape.end(), outTailShape.begin(), outTailShape.end());
        outTsShape.back() *= (1 + static_cast<int>(bidirectional));

        allocated = true;
    }

    void forward(InputArrayOfArrays inputs_arr, OutputArrayOfArrays outputs_arr, OutputArrayOfArrays internals_arr) CV_OVERRIDE

    {
        CV_TRACE_FUNCTION();
        CV_TRACE_ARG_VALUE(name, "name", name.c_str());

        if (inputs_arr.depth() == CV_16F)
        {
            forward_fallback(inputs_arr, outputs_arr, internals_arr);
            return;
        }

        std::vector<Mat> input, output, internals;
        inputs_arr.getMatVector(input);
        outputs_arr.getMatVector(output);
        internals_arr.getMatVector(internals);

        if (layout == BATCH_SEQ_HID){
            //swap axis 0 and 1 input x
            cv::Mat tmp;
            // Since python input is 4 dimentional and C++ input 3 dimentinal
            // we need to proccess each differently
            if (input[0].dims == 4){
                // here !!!
                CV_Assert(input[0].size[3] == 1);
                cv::transposeND(input[0], {1, 0, 2, 3}, tmp); //back to seq_len, batch_size, hidden_size format
            }else{
                cv::transposeND(input[0], {1, 0, 2}, tmp); //back to seq_len, batch_size, hidden_size format
            }
            input[0] = tmp;
        }

        Mat cOut = produceCellOutput ? output[0].clone() : Mat();
        const bool needYcTransform = !originalBlobs.empty(); // if the producer is onnx
        const int numDirs = 1 + static_cast<int>(bidirectional);
        for (int i = 0; i < numDirs; ++i)
        {
            Mat Wh = blobs[0];
            Mat Wx = blobs[1];
            Mat bias = blobs[2];

            Mat h_0, c_0;
            // Handle h_0 and c_0 based on input size
            h_0 = (input.size() >= 2) ? input[1].reshape(1, input[1].size[0] * input[1].size[1]) : blobs[3];
            c_0 = (input.size() == 3) ? input[2].reshape(1, input[2].size[0] * input[2].size[1]) : blobs[4];

            // Perform checks if input size is 2 or 3
            if (input.size() >= 2) {
                CV_CheckEQ(h_0.cols, Wh.cols, "");
                CV_CheckEQ(h_0.cols, c_0.cols, "");
                CV_CheckEQ(h_0.rows, c_0.rows, "");
            }


            Mat pI, pF, pO;

            Wh = Wh.rowRange(i * Wh.rows / numDirs, (i + 1) * Wh.rows / numDirs);
            Wx = Wx.rowRange(i * Wx.rows / numDirs, (i + 1) * Wx.rows / numDirs);
            bias = bias.colRange(i * bias.cols / numDirs, (i + 1) * bias.cols / numDirs);
            h_0 = h_0.rowRange(i * h_0.rows / numDirs, (i + 1) * h_0.rows / numDirs);
            c_0 = c_0.rowRange(i * c_0.rows / numDirs, (i + 1) * c_0.rows / numDirs);

            if (usePeephole)
            {
                pI = blobs[5];
                pF = blobs[6];
                pO = blobs[7];

                pI = pI.rowRange(i * pI.rows / numDirs, (i + 1) * pI.rows / numDirs);
                pI = pI.colRange(i * pI.cols / numDirs, (i + 1) * pI.cols / numDirs);

                pF = pF.rowRange(i * pF.rows / numDirs, (i + 1) * pF.rows / numDirs);
                pF = pF.colRange(i * pF.cols / numDirs, (i + 1) * pF.cols / numDirs);

                pO = pO.rowRange(i * pO.rows / numDirs, (i + 1) * pO.rows / numDirs);
                pO = pO.colRange(i * pO.cols / numDirs, (i + 1) * pO.cols / numDirs);
            }

            int numOut = Wh.size[1];
            Mat hInternal = internals[0], cInternal = internals[1],
                    dummyOnes = internals[2], gates = internals[3];
            h_0.copyTo(hInternal);
            c_0.copyTo(cInternal);
            dummyOnes.setTo(1.);

            int numSamplesTotal = numTimeStamps*numSamples;
            Mat xTs = input[0].reshape(1, numSamplesTotal);

            Mat hOutTs = output[0].reshape(1, numSamplesTotal);
            hOutTs = hOutTs.colRange(i * hOutTs.cols / numDirs, (i + 1) * hOutTs.cols / numDirs);
            Mat cOutTs;
            if (produceCellOutput)
            {
                cOutTs = cOut.reshape(1, numSamplesTotal);
                cOutTs = cOutTs.colRange(i * cOutTs.cols / numDirs, (i + 1) * cOutTs.cols / numDirs);
            }

#if CV_TRY_AVX2 || CV_TRY_AVX
            bool canUseAvx = gates.isContinuous() && bias.isContinuous()
                && Wx.depth() == CV_32F && gates.depth() == CV_32F
                && bias.depth() == CV_32F && Wx.cols >= 8;
            bool canUseAvx_hInternal = hInternal.isContinuous() && gates.isContinuous() && bias.isContinuous()
                && Wh.depth() == CV_32F && hInternal.depth() == CV_32F && gates.depth() == CV_32F
                && Wh.cols >= 8;
#endif

            int tsStart, tsEnd, tsInc;
            if (reverse || i == 1) {
                tsStart = numTimeStamps - 1;
                tsEnd = -1;
                tsInc = -1;
            }
            else {
                tsStart = 0;
                tsEnd = numTimeStamps;
                tsInc = 1;
            }
            for (int ts = tsStart; ts != tsEnd; ts += tsInc)
            {
                Range curRowRange(ts*numSamples, (ts + 1)*numSamples);
                Mat xCurr = xTs.rowRange(curRowRange);

#if CV_TRY_AVX2
                if (useAVX2 && canUseAvx && xCurr.isContinuous())
                {
                    for (int n = 0; n < xCurr.rows; n++) {
                        opt_AVX2::fastGEMM1T(
                            xCurr.ptr<float>(n),
                            Wx.ptr<float>(),
                            Wx.step1(),
                            bias.ptr<float>(),
                            gates.ptr<float>(n),
                            Wx.rows,
                            Wx.cols
                        );
                    }
                }
                else
#endif
#if CV_TRY_AVX
                if (useAVX && canUseAvx && xCurr.isContinuous())
                {
                    for (int n = 0; n < xCurr.rows; n++) {
                        opt_AVX::fastGEMM1T(
                            xCurr.ptr<float>(n),
                            Wx.ptr<float>(),
                            Wx.step1(),
                            bias.ptr<float>(),
                            gates.ptr<float>(n),
                            Wx.rows,
                            Wx.cols
                        );
                    }
                }
                else
#endif
                {
                    gemm(xCurr, Wx, 1, gates, 0, gates, GEMM_2_T);      // Wx * x_t
                    gemm(dummyOnes, bias, 1, gates, 1, gates);          //+b
                }

#if CV_TRY_AVX2
                if (useAVX2 && canUseAvx_hInternal)
                {
                    for (int n = 0; n < hInternal.rows; n++) {
                        opt_AVX2::fastGEMM1T(
                            hInternal.ptr<float>(n),
                            Wh.ptr<float>(),
                            Wh.step1(),
                            gates.ptr<float>(n),
                            gates.ptr<float>(n),
                            Wh.rows,
                            Wh.cols
                        );
                    }
                }
                else
#endif
#if CV_TRY_AVX
                if (useAVX && canUseAvx_hInternal)
                {
                    for (int n = 0; n < hInternal.rows; n++) {
                        opt_AVX::fastGEMM1T(
                            hInternal.ptr<float>(n),
                            Wh.ptr<float>(),
                            Wh.step1(),
                            gates.ptr<float>(n),
                            gates.ptr<float>(n),
                            Wh.rows,
                            Wh.cols
                        );
                    }
                }
                else
#endif
                {
                    gemm(hInternal, Wh, 1, gates, 1, gates, GEMM_2_T);  //+Wh * h_{t-1}
                }

                Mat gateI = gates.colRange(0*numOut, 1*numOut);
                Mat gateF = gates.colRange(1*numOut, 2*numOut);
                Mat gateO = gates.colRange(2*numOut, 3*numOut);
                Mat gateG = gates.colRange(3*numOut, 4*numOut);

                if (forgetBias)
                    add(gateF, forgetBias, gateF);

                if (usePeephole)
                {
                    Mat gatesIF = gates.colRange(0, 2*numOut);
                    gemm(cInternal, pI, 1, gateI, 1, gateI);
                    gemm(cInternal, pF, 1, gateF, 1, gateF);
                    f_activation(gatesIF, gatesIF);
                }
                else
                {
                    Mat gatesIFO = gates.colRange(0, 3*numOut);
                    f_activation(gatesIFO, gatesIFO);
                }

                g_activation(gateG, gateG);

                //compute c_t
                multiply(gateF, cInternal, gateF);  // f_t (*) c_{t-1}
                multiply(gateI, gateG, gateI);      // i_t (*) g_t
                add(gateF, gateI, cInternal);       // c_t = f_t (*) c_{t-1} + i_t (*) g_t

                if (useCellClip)
                {
                    min(cInternal, cellClip, cInternal);
                    max(cInternal, -cellClip, cInternal);
                }
                if (usePeephole)
                {
                    gemm(cInternal, pO, 1, gateO, 1, gateO);
                    f_activation(gateO, gateO);
                }

                //compute h_t
                h_activation(cInternal, hInternal);
                multiply(gateO, hInternal, hInternal);

                //save results in output blobs
                hInternal.copyTo(hOutTs.rowRange(curRowRange));
                if (produceCellOutput)
                    cInternal.copyTo(cOutTs.rowRange(curRowRange));
            }
        }
        // transpose to match batch first output
        if (layout == BATCH_SEQ_HID){
            cv::Mat tmp;
            cv::transposeND(output[0], {1, 0, 2}, tmp);
            output[0] = tmp;
        }
        if (needYcTransform && produceCellOutput)
        {
            fixCellState(cOut, numDirs);
        }
        if (produceCellOutput)
        {
            cOut.copyTo(output[1]);
        }
    }

    void fixCellState(Mat& cOut, int numDirs)

    {
        // seq, batch, dirs, hidden
        int shp[] = {0, numSamples, numDirs, numHidden};
        cOut = cOut.reshape(1, sizeof(shp)/sizeof(shp[0]), shp);

        // permute to {0, 2, 1, 3};
        cv::Mat newCellState;
        // transpose to match batch first output
        if (layout == BATCH_SEQ_HID){
            cv::transposeND(cOut, {2, 0, 1, 3}, newCellState);
        }
        else{
            cv::transposeND(cOut, {0, 2, 1, 3}, newCellState);
        }
        cOut = newCellState;

        if (numDirs == 1)
        {
            // Slice: Yh = Y[-1, :, :, :]
            Range ranges[] = {cv::Range(cOut.size[0] - 1, cOut.size[0]), cv::Range::all(), cv::Range::all(), cv::Range::all()};
            cOut = cOut(ranges);
            // Reshape: 1x1xBxH -> 1xBxH
            int shp[] = {1, numSamples, numHidden};
            cOut = cOut.reshape(1, sizeof(shp)/sizeof(shp[0]), shp);
        }
        else
        {
            // Slice: SxDxBxH -> last sequence, first direction
            Range ranges1[] = {cv::Range(cOut.size[0] - 1, cOut.size[0]), cv::Range(0, 1), cv::Range::all(), cv::Range::all()};
            Mat part1 = cOut(ranges1);

            // Slice: SxDxBxH -> first sequence, last direction
            Range ranges2[] = {cv::Range(0, 1), cv::Range(cOut.size[1] - 1, cOut.size[1]), cv::Range::all(), cv::Range::all()};
            Mat part2 = cOut(ranges2);

            int shp[] = {1, part1.size[2] * part1.size[3]};
            part1 = part1.reshape(1, sizeof(shp)/sizeof(shp[0]), shp);
            part2 = part2.reshape(1, sizeof(shp)/sizeof(shp[0]), shp);

            vconcat(part1, part2, cOut);

            // Reshape: 1x2xBxH -> 2xBxH
            int finalShape[] = {2, numSamples, numHidden};
            cOut = cOut.reshape(1, sizeof(finalShape)/sizeof(finalShape[0]), finalShape);
        }
    }

#ifdef HAVE_CUDA
    Ptr<BackendNode> initCUDA(void *context_, const std::vector<Ptr<BackendWrapper>> &inputs,

                              const std::vector<Ptr<BackendWrapper>> &outputs) override

    {
        const int numDirs = 1 + static_cast<int>(bidirectional);
        auto toIFCO = [numDirs] (Mat& in) {
            int first = in.size[0];
            int rest = in.total() / first / 4;
            // every weight blob contains weights for Input, Output, Forget and Cell gates
            Mat m = in.reshape(1, {first, 4, rest});
            Mat outputGate = m.col(1);
            Mat forgetGate = m.col(2);
            Mat cellGate = m.col(3);
            // IOFC -> IFOC
            std::swap_ranges(outputGate.begin<float>(), outputGate.end<float>(), forgetGate.begin<float>());
            std::swap(outputGate, forgetGate);
            // IFOC -> IFCO
            std::swap_ranges(outputGate.begin<float>(), outputGate.end<float>(), cellGate.begin<float>());
            in = in.reshape(1, numDirs);
        };

        Mat& b = originalBlobs[2];
        // B is a concatenation of biases for Wh and Wx
        b = b.reshape(1, originalBlobs[2].size[0]*2);

        for (auto& m : originalBlobs)
        {
            toIFCO(m);
        }

        b = b.reshape(1, static_cast<int>(b.total()));

        Mat ordered_weights;
        // Wx_f, Wh_f, [Wx_b, Wh_b,] b
        for (int i = 0; i < numDirs; ++i)
        {
            for (size_t j = 0; j < 2; ++j) // Wx, Wh
            {
                Mat oneDirection = originalBlobs[j].row(i);
                ordered_weights.push_back(oneDirection.reshape(1, static_cast<int>(oneDirection.total())));
            }
        }
        ordered_weights.push_back(b);

        // Pass hidden states as is
        Mat h0 = blobs[3];
        Mat c0 = blobs[4];

        CV_Assert(!inputs.empty());
        auto input_wrapper = inputs[0].dynamicCast<CUDABackendWrapper>();
        auto input_shape = input_wrapper->getShape();

        RNNConfiguration config
        {
            input_shape[0],      // seqLength;
            1,                   // numLayers;
            numHidden,           // hiddenSize;
            input_shape[2],      // inputSize;
            input_shape[1],      // miniBatch;
            bidirectional
        };


        auto *context = reinterpret_cast<cuda4dnn::csl::CSLContext *>(context_);
        return make_cuda_node<cuda4dnn::LSTMOp>(preferableTarget, std::move(context->stream),
                                                std::move(context->cudnn_handle),
                                                ordered_weights, h0, c0,
                                                config);
    }
#endif
};

Ptr<LSTMLayer> LSTMLayer::create(const LayerParams& params)

{
    return Ptr<LSTMLayer>(new LSTMLayerImpl(params));
}

int LSTMLayer::inputNameToIndex(String inputName)

{
    if (toLowerCase(inputName) == "x")
        return 0;
    return -1;
}

int LSTMLayer::outputNameToIndex(const String& outputName)

{
    if (toLowerCase(outputName) == "h")
        return 0;
    else if (toLowerCase(outputName) == "c")
        return 1;
    return -1;
}


class RNNLayerImpl : public RNNLayer
{
    int numX, numH, numO;
    int numSamples, numTimestamps, numSamplesTotal;
    int dtype;
    Mat Whh, Wxh, bh;
    Mat Who, bo;
    bool produceH;

public:

    RNNLayerImpl(const LayerParams& params)
        : numX(0), numH(0), numO(0), numSamples(0), numTimestamps(0), numSamplesTotal(0), dtype(0)
    {
        setParamsFrom(params);
        type = "RNN";
        produceH = false;
    }

    void setProduceHiddenOutput(bool produce = false) CV_OVERRIDE

    {
        produceH = produce;
    }

    void setWeights(const Mat &W_xh, const Mat &b_h, const Mat &W_hh, const Mat &W_ho, const Mat &b_o) CV_OVERRIDE

    {
        CV_Assert(W_hh.dims == 2 && W_xh.dims == 2);
        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]);

        blobs.resize(5);
        blobs[0] = Mat(W_xh.clone());
        blobs[1] = Mat(b_h.clone());
        blobs[2] = Mat(W_hh.clone());
        blobs[3] = Mat(W_ho.clone());
        blobs[4] = Mat(b_o.clone());
    }

    bool getMemoryShapes(const std::vector<MatShape> &inputs,

                         const int requiredOutputs,

                         std::vector<MatShape> &outputs,

                         std::vector<MatShape> &internals) const CV_OVERRIDE

    {
        CV_Assert(inputs.size() >= 1 && inputs.size() <= 2);

        Mat Who_ = blobs[3];
        Mat Wxh_ = blobs[0];

        int numTimestamps_ = inputs[0][0];
        int numSamples_ = inputs[0][1];

        int numO_ = Who_.rows;
        int numH_ = Wxh_.rows;

        outputs.clear();
        int dims[] = {numTimestamps_, numSamples_, numO_};
        outputs.push_back(shape(dims, 3));
        dims[2] = numH_;
        if (produceH)
            outputs.push_back(shape(dims, 3));

        internals.assign(2, shape(numSamples_, numH_));
        internals.push_back(shape(numSamples_, 1));

        return false;
    }

    void finalize(InputArrayOfArrays inputs_arr, OutputArrayOfArrays) CV_OVERRIDE

    {
        std::vector<Mat> input, outputs;
        inputs_arr.getMatVector(input);

        CV_Assert(input.size() >= 1 && input.size() <= 2);

        Wxh = blobs[0];
        bh  = blobs[1];
        Whh = blobs[2];
        Who = blobs[3];
        bo  = blobs[4];

        numH = Wxh.rows;
        numX = Wxh.cols;
        numO = Who.rows;

        const Mat& inp0 = input[0];

        CV_Assert(inp0.dims >= 2);
        CV_Assert(inp0.total(2) == numX);
        dtype = CV_32F;
        CV_Assert(inp0.type() == dtype);
        numTimestamps = inp0.size[0];
        numSamples = inp0.size[1];
        numSamplesTotal = numTimestamps * numSamples;

        bh = bh.reshape(1, 1); //is 1 x numH Mat
        bo = bo.reshape(1, 1); //is 1 x numO Mat
    }

    void reshapeOutput(std::vector<Mat> &output)

    {
        output.resize(produceH ? 2 : 1);
        int sz0[] = { numTimestamps, numSamples, numO };
        output[0].create(3, sz0, dtype);
        if (produceH)
        {
            int sz1[] = { numTimestamps, numSamples, numH };
            output[1].create(3, sz1, dtype);
        }
    }

    void forward(InputArrayOfArrays inputs_arr, OutputArrayOfArrays outputs_arr, OutputArrayOfArrays internals_arr) CV_OVERRIDE

    {
        CV_TRACE_FUNCTION();
        CV_TRACE_ARG_VALUE(name, "name", name.c_str());

        if (inputs_arr.depth() == CV_16F)
        {
            forward_fallback(inputs_arr, outputs_arr, internals_arr);
            return;
        }

        std::vector<Mat> input, output, internals;
        inputs_arr.getMatVector(input);
        outputs_arr.getMatVector(output);
        internals_arr.getMatVector(internals);

        Mat xTs = input[0].reshape(1, numSamplesTotal);
        Mat oTs = output[0].reshape(1, numSamplesTotal);
        Mat hTs = produceH ? output[1].reshape(1, numSamplesTotal) : Mat();
        Mat hCurr = internals[0];
        Mat hPrev = internals[1];
        Mat dummyBiasOnes = internals[2];

        hPrev.setTo(0.);
        dummyBiasOnes.setTo(1.);

        for (int ts = 0; ts < numTimestamps; ts++)
        {
            Range curRowRange = Range(ts * numSamples, (ts + 1) * numSamples);
            Mat xCurr = xTs.rowRange(curRowRange);

            gemm(hPrev, Whh, 1, hCurr, 0, hCurr, GEMM_2_T); // W_{hh} * h_{prev}
            gemm(xCurr, Wxh, 1, hCurr, 1, hCurr, GEMM_2_T); //+W_{xh} * x_{curr}
            gemm(dummyBiasOnes, bh, 1, hCurr, 1, hCurr);    //+bh
            tanh(hCurr, hPrev);

            Mat oCurr = oTs.rowRange(curRowRange);
            gemm(hPrev, Who, 1, oCurr, 0, oCurr, GEMM_2_T); // W_{ho} * h_{prev}
            gemm(dummyBiasOnes, bo, 1, oCurr, 1, oCurr);    //+b_o
            tanh(oCurr, oCurr);

            if (produceH)
                hPrev.copyTo(hTs.rowRange(curRowRange));
        }
    }
};

CV_EXPORTS_W Ptr<RNNLayer> RNNLayer::create(const LayerParams& params)

{
    return Ptr<RNNLayer>(new RNNLayerImpl(params));
}

class GRULayerImpl CV_FINAL : public GRULayer
{
    int numTimeStamps, numSamples;
    bool allocated;

    MatShape outTailShape;  //shape of single output sample
    MatShape outTsShape;    //shape of N output samples
    bool bidirectional;     // If true, produces both forward and reversed directions along time axis

public:

    GRULayerImpl(const LayerParams& params) : numTimeStamps(0), numSamples(0)
    {
        setParamsFrom(params);

        bidirectional = params.get<bool>("bidirectional", false);
        if (!blobs.empty())
        {
            CV_Assert(blobs.size() >= 3);

            blobs[2] = blobs[2].reshape(1, 1);

            const Mat& Wh = blobs[0];
            const Mat& Wx = blobs[1];
            const Mat& bias = blobs[2];
            const Mat& hInternal = blobs[3];
            CV_CheckEQ(Wh.dims, 2, "");
            CV_CheckEQ(Wx.dims, 2, "");
            CV_CheckEQ(Wh.rows, Wx.rows, "");
            CV_CheckEQ(Wh.rows, (1 + static_cast<int>(bidirectional)) * 3 * Wh.cols, "");
            CV_CheckEQ(Wh.rows * 2, (int)bias.total(), "");
            CV_CheckEQ(hInternal.cols, Wh.cols, "");
            CV_CheckTypeEQ(Wh.type(), Wx.type(), "");
            CV_CheckTypeEQ(Wx.type(), bias.type(), "");
        }

        allocated = false;
        outTailShape.clear();
    }

    bool getMemoryShapes(const std::vector<MatShape> &inputs,

                         const int requiredOutputs,

                         std::vector<MatShape> &outputs,

                         std::vector<MatShape> &internals) const CV_OVERRIDE

    {
        CV_Assert(inputs.size() == 1);
        const MatShape& inp0 = inputs[0];

        const Mat &Wh = blobs[0], &Wx = blobs[1];
        int _numOut = Wh.size[1];
        int _numInp = Wx.size[1];
        MatShape outTailShape_(outTailShape), outResShape;

        if (!outTailShape_.empty())
            CV_Assert(total(outTailShape_) == _numOut);
        else
            outTailShape_.assign(1, _numOut);

        int _numSamples;
        CV_Assert(inp0.size() >= 2 && total(inp0, 2) == _numInp);
        _numSamples = inp0[1];
        outResShape.push_back(inp0[0]);

        outResShape.push_back(_numSamples);
        outResShape.insert(outResShape.end(), outTailShape_.begin(), outTailShape_.end());
        outResShape.back() *= (1 + static_cast<int>(bidirectional));

        outputs.assign(1, outResShape);

        internals.assign(1, shape(_numSamples, _numOut));     // hInternal
        internals.push_back(shape(_numSamples, 1));           // dummyOnes
        internals.push_back(shape(_numSamples, 2 * _numOut)); // gates
        internals.push_back(shape(_numSamples, 2 * _numOut)); // gates_b
        internals.push_back(shape(_numSamples, 1 * _numOut)); // h_linear
        internals.push_back(shape(_numSamples, _numOut));     // ones

        return false;
    }

    void finalize(InputArrayOfArrays inputs_arr, OutputArrayOfArrays) CV_OVERRIDE

    {
        std::vector<Mat> input;
        inputs_arr.getMatVector(input);

        CV_Assert(input.size() == 1);
        const Mat& inp0 = input[0];

        Mat &Wh = blobs[0], &Wx = blobs[1];
        int numOut = Wh.size[1];
        int numInp = Wx.size[1];

        if (!outTailShape.empty())
            CV_Assert(total(outTailShape) == numOut);
        else
            outTailShape.assign(1, numOut);

        CV_Assert(inp0.dims >= 2 && (int)inp0.total(2) == numInp);
        numTimeStamps = inp0.size[0];
        numSamples = inp0.size[1];

        outTsShape.clear();
        outTsShape.push_back(numSamples);
        outTsShape.insert(outTsShape.end(), outTailShape.begin(), outTailShape.end());
        outTsShape.back() *= (1 + static_cast<int>(bidirectional));

        allocated = true;
    }

    void forward(InputArrayOfArrays inputs_arr, OutputArrayOfArrays outputs_arr, OutputArrayOfArrays internals_arr) CV_OVERRIDE

    {
        CV_TRACE_FUNCTION();
        CV_TRACE_ARG_VALUE(name, "name", name.c_str());

        if (inputs_arr.depth() == CV_16F)
        {
            forward_fallback(inputs_arr, outputs_arr, internals_arr);
            return;
        }

        std::vector<Mat> input, output, internals;
        inputs_arr.getMatVector(input);
        outputs_arr.getMatVector(output);
        internals_arr.getMatVector(internals);

        const int numDirs = 1 + static_cast<int>(bidirectional);
        for (int i = 0; i < numDirs; ++i)
        {
            const Mat &Wh = blobs[0].rowRange(i * blobs[0].rows / numDirs, (i + 1) * blobs[0].rows / numDirs);
            const Mat &Wx = blobs[1].rowRange(i * blobs[1].rows / numDirs, (i + 1) * blobs[1].rows / numDirs);
            const Mat &bias = blobs[2].colRange(i * blobs[2].cols / numDirs, (i + 1) * blobs[2].cols / numDirs);
            const Mat &h_0 = blobs[3].rowRange(i * blobs[3].rows / numDirs, (i + 1) * blobs[3].rows / numDirs);

            const Mat &bx = bias.colRange(0, bias.cols / 2);
            const Mat &bh = bias.colRange(bias.cols / 2, bias.cols);

            Mat hInternal = internals[0], dummyOnes = internals[1], gates = internals[2],
                b_rz = internals[3], n_t = internals[4], ones = internals[5];
            h_0.copyTo(hInternal);
            dummyOnes.setTo(1.);
            ones.setTo(1.);

            int numOut = Wh.size[1];
            const Mat& wx_rz = Wx.rowRange(0, 2 * numOut);
            const Mat& wh_rz = Wh.rowRange(0, 2 * numOut);
            b_rz = bx.colRange(0, 2 * numOut) + bh.colRange(0, 2 * numOut);
            const Mat& wx_n = Wx.rowRange(2 * numOut, 3 * numOut);
            const Mat& wh_n = Wh.rowRange(2 * numOut, 3 * numOut);
            const Mat& b_in = bx.colRange(2 * numOut, 3 * numOut);
            const Mat& b_hn = bh.colRange(2 * numOut, 3 * numOut);

            int numSamplesTotal = numTimeStamps * numSamples;
            Mat xTs = input[0].reshape(1, numSamplesTotal);

            Mat hOutTs = output[0].reshape(1, numSamplesTotal);
            hOutTs = hOutTs.colRange(i * hOutTs.cols / numDirs, (i + 1) * hOutTs.cols / numDirs);
            Mat cOutTs = Mat();

            int tsStart, tsEnd, tsInc;
            if (i == 1) {
                tsStart = numTimeStamps - 1;
                tsEnd = -1;
                tsInc = -1;
            }
            else {
                tsStart = 0;
                tsEnd = numTimeStamps;
                tsInc = 1;
            }
            for (int ts = tsStart; ts != tsEnd; ts += tsInc)
            {
                Range curRowRange(ts * numSamples, (ts + 1) * numSamples);
                Mat xCurr = xTs.rowRange(curRowRange);

                // calculate r_t = sigmoid(x * Wx_r + h_(t-1) * Wh_r + b_r)
                // calculate z_t = sigmoid(x * Wx_z + h_(t-1) * Wh_z + b_z)
                gemm(xCurr, wx_rz, 1, gates, 0, gates, GEMM_2_T);      // x * Wx_rz
                gemm(hInternal, wh_rz, 1, gates, 1, gates, GEMM_2_T);  // + h_(t-1) * Wh_rz
                gemm(dummyOnes, b_rz, 1, gates, 1, gates);             // + b_rz
                sigmoid(gates, gates);                                 // sigmoid()

                Mat z = gates.colRange(0, gates.cols / 2);
                Mat r = gates.colRange(gates.cols / 2, gates.cols);

                // calculate n_t = tanh(r (*) (h_(t-1) * Wh_n + b_hn) + x * Wx_n + b_in)
                gemm(hInternal, wh_n, 1, n_t, 0, n_t, GEMM_2_T);       // h_(t-1) * Wh_n
                gemm(dummyOnes, b_hn, 1, n_t, 1, n_t);                 // + b_hn
                multiply(r, n_t, n_t);                                 // r (*) (h_(t-1) * Wh_n + b_hn)

                gemm(xCurr, wx_n, 1, n_t, 1, n_t, GEMM_2_T);          // + x * Wx_n
                gemm(dummyOnes, b_in, 1, n_t, 1, n_t);                // + b_in
                tanh(n_t, n_t);                                       // tanh()

                //compute next h_t = z (*) h_(t-1) + (1 - z) (*) n_t
                multiply(z, hInternal, hInternal);                    // z (*) h_{t-1}
                subtract(ones, z, z);                                 // 1 - z
                multiply(z, n_t, z);                                  // (1 - z) * n
                add(z, hInternal, hInternal);                         // z (*) h_(t-1) + (1 - z) (*) n_t

                //save results in output blobs
                hInternal.copyTo(hOutTs.rowRange(curRowRange));
            }
        }
    }
};

Ptr<GRULayer> GRULayer::create(const LayerParams &params) {
    return Ptr<GRULayer>(new GRULayerImpl(params));
}

}
}