RNNLayer.h

Go to the documentation of this file.

// @(#)root/tmva/tmva/dnn/rnn:$Id$
// Author: Saurav Shekhar 19/07/17
 
/**********************************************************************************
 * Project: TMVA - a Root-integrated toolkit for multivariate data analysis       *
 * Package: TMVA                                                                  *
 * Class : BasicRNNLayer                                                          *
 *                                                                                *
 * Description:                                                                   *
 *       NeuralNetwork                                                            *
 *                                                                                *
 * Authors (alphabetical):                                                        *
 *       Saurav Shekhar    <sauravshekhar01@gmail.com> - ETH Zurich, Switzerland  *
 *                                                                                *
 * Copyright (c) 2005-2015:                                                       *
 * All rights reserved.                                                           *
 *       CERN, Switzerland                                                        *
 *                                                                                *
 * For the licensing terms see $ROOTSYS/LICENSE.                                  *
 * For the list of contributors see $ROOTSYS/README/CREDITS.                      *
 **********************************************************************************/
 
//#pragma once
 
//////////////////////////////////////////////////////////////////////
// <Description> //
//////////////////////////////////////////////////////////////////////
 
#ifndef TMVA_DNN_RNN_LAYER
#define TMVA_DNN_RNN_LAYER
 
#include <cmath>
#include <iostream>
#include <vector>
#include <string>
 
#include "TMatrix.h"
#include "TMVA/DNN/Functions.h"
 
namespace TMVA
{
namespace DNN
{
 
namespace RNN {
 
//______________________________________________________________________________
//
// Basic RNN Layer
//______________________________________________________________________________
 
/** \class BasicRNNLayer
      Generic implementation
*/
template<typename Architecture_t>
      class TBasicRNNLayer : public VGeneralLayer<Architecture_t>
{
 
public:
 
   using Tensor_t = typename Architecture_t::Tensor_t;
   using Matrix_t = typename Architecture_t::Matrix_t;
   using Scalar_t = typename Architecture_t::Scalar_t;
 
   using LayerDescriptor_t = typename Architecture_t::RecurrentDescriptor_t;
   using WeightsDescriptor_t = typename Architecture_t::FilterDescriptor_t;
   using TensorDescriptor_t = typename Architecture_t::TensorDescriptor_t;
   using HelperDescriptor_t = typename Architecture_t::DropoutDescriptor_t;
 
   using RNNWorkspace_t = typename Architecture_t::RNNWorkspace_t;
   using RNNDescriptors_t = typename Architecture_t::RNNDescriptors_t;
 
private:
 
   size_t fTimeSteps;              ///< Timesteps for RNN
   size_t fStateSize;              ///< Hidden state size of RNN
   bool   fRememberState;          ///< Remember state in next pass
   bool   fReturnSequence = false; ///< Return in output full sequence or just last element in time
 
   DNN::EActivationFunction fF;  ///< Activation function of the hidden state
 
   Matrix_t fState;                ///< Hidden State
   Matrix_t &fWeightsInput;         ///< Input weights, fWeights[0]
   Matrix_t &fWeightsState;         ///< Prev state weights, fWeights[1]
   Matrix_t &fBiases;               ///< Biases
 
   Tensor_t fDerivatives; ///< First fDerivatives of the activations
   Matrix_t &fWeightInputGradients; ///< Gradients w.r.t. the input weights
   Matrix_t &fWeightStateGradients; ///< Gradients w.r.t. the recurring weights
   Matrix_t &fBiasGradients;        ///< Gradients w.r.t. the bias values
 
   Tensor_t fWeightsTensor;
   Tensor_t fWeightGradientsTensor;
 
   typename Architecture_t::ActivationDescriptor_t fActivationDesc;
 
   TDescriptors *fDescriptors = nullptr; ///< Keeps all the RNN descriptors
   TWorkspace   *fWorkspace = nullptr;   // workspace needed for GPU computation (CudNN)
 
   Matrix_t fCell; ///< Empty matrix for RNN
 
   // tensors used internally for the forward and backward pass
   Tensor_t fX;  ///<  cached input tensor as T x B x I
   Tensor_t fY;  ///<  cached output tensor as T x B x S
   Tensor_t fDx; ///< cached   gradient on the input (output of backward)   as T x B x I
   Tensor_t fDy; ///< cached  activation gradient (input of backward)   as T x B x S
 
 
public:
 
   /** Constructor */
   TBasicRNNLayer(size_t batchSize, size_t stateSize, size_t inputSize,
                  size_t timeSteps, bool rememberState = false, bool returnSequence = false,
                  DNN::EActivationFunction f = DNN::EActivationFunction::kTanh,
                  bool training = true, DNN::EInitialization fA = DNN::EInitialization::kZero);
 
   /** Copy Constructor */
   TBasicRNNLayer(const TBasicRNNLayer &);
 
   /*! Destructor. */
   virtual ~TBasicRNNLayer();
 
   /*! Initialize the weights according to the given initialization
    **  method. */
   void Initialize() override;
 
   /*! Initialize the state
    **  method. */
   void InitState(DNN::EInitialization m = DNN::EInitialization::kZero);
 
   /*! Compute and return the next state with given input
   *  matrix */
   void Forward(Tensor_t &input, bool isTraining = true) override;
 
   /*! Forward for a single cell (time unit) */
   void CellForward(const Matrix_t &input, Matrix_t & dF);
 
   /*! Backpropagates the error. Must only be called directly at the corresponding
    *  call to Forward(...). */
   void Backward(Tensor_t &gradients_backward,
                 const Tensor_t &activations_backward) override;
 
   /* Updates weights and biases, given the learning rate */
   void Update(const Scalar_t learningRate);
 
   /*! Backward for a single time unit
    * a the corresponding call to Forward(...). */
   inline Matrix_t & CellBackward(Matrix_t & state_gradients_backward,
                              const Matrix_t & precStateActivations,
                              const Matrix_t & input, Matrix_t & input_gradient, Matrix_t &dF);
 
   /** Prints the info about the layer */
   void Print() const override;
 
   /*! Writes the information and the weights about the layer in an XML node. */
   void AddWeightsXMLTo(void *parent) override;
 
   /*! Read the information and the weights about the layer from XML node. */
   void ReadWeightsFromXML(void *parent) override;
 
   void InitTensors();
   // void InitializeDescriptors();
   // void ReleaseDescriptors();
   // void InitializeWorkspace();
   // void FreeWorkspace();
 
   /** Getters */
   size_t GetTimeSteps() const { return fTimeSteps; }
   size_t GetStateSize() const { return fStateSize; }
   size_t GetInputSize() const { return this->GetInputWidth(); }
   inline bool DoesRememberState()  const {return fRememberState;}
   inline bool DoesReturnSequence() const { return fReturnSequence; }
   inline DNN::EActivationFunction GetActivationFunction()  const {return fF;}
   Matrix_t        & GetState()            {return fState;}  // RNN Hidden state
   const Matrix_t & GetState()       const  {return fState;}
   Matrix_t &GetCell() { return fCell; } // this returns an empty matrixfor RNN
   const Matrix_t &GetCell() const { return fCell; }
 
   Matrix_t        & GetWeightsInput()        {return fWeightsInput;}
   const Matrix_t & GetWeightsInput()   const {return fWeightsInput;}
   Matrix_t        & GetWeightsState()        {return fWeightsState;}
   const Matrix_t & GetWeightsState()   const {return fWeightsState;}
   Tensor_t       & GetDerivatives()        {return fDerivatives;}
   const Tensor_t & GetDerivatives()   const {return fDerivatives;}
   // Matrix_t &GetDerivativesAt(size_t i) { return fDerivatives[i]; }
   // const Matrix_t &GetDerivativesAt(size_t i) const { return fDerivatives[i]; }
 
   Matrix_t        & GetBiasesState()              {return fBiases;}
   const Matrix_t & GetBiasesState()         const {return fBiases;}
   Matrix_t        & GetBiasStateGradients()            {return fBiasGradients;}
   const Matrix_t & GetBiasStateGradients() const {return fBiasGradients;}
   Matrix_t        & GetWeightInputGradients()         {return fWeightInputGradients;}
   const Matrix_t & GetWeightInputGradients()    const {return fWeightInputGradients;}
   Matrix_t        & GetWeightStateGradients()         {return fWeightStateGradients;}
   const Matrix_t & GetWeightStateGradients()    const {return fWeightStateGradients;}
 
   Tensor_t &GetWeightsTensor()  { return fWeightsTensor;  }
   const Tensor_t &GetWeightsTensor() const { return fWeightsTensor; }
   Tensor_t &GetWeightGradientsTensor() { return fWeightGradientsTensor; }
   const Tensor_t &GetWeightGradientsTensor() const { return fWeightGradientsTensor; }
 
   Tensor_t &GetX() { return fX; }
   Tensor_t &GetY() { return fY; }
   Tensor_t &GetDX() { return fDx; }
   Tensor_t &GetDY() { return fDy; }
};
 
//______________________________________________________________________________
//
// BasicRNNLayer Implementation
//______________________________________________________________________________
template <typename Architecture_t>
TBasicRNNLayer<Architecture_t>::TBasicRNNLayer(size_t batchSize, size_t stateSize, size_t inputSize, size_t timeSteps,
                                               bool rememberState, bool returnSequence, DNN::EActivationFunction f, bool /*training*/,
                                               DNN::EInitialization fA)
   // TODO inputDepth and outputDepth changed to batchSize??
   : VGeneralLayer<Architecture_t>(batchSize, 1, timeSteps, inputSize, 1, (returnSequence) ? timeSteps : 1 ,
                                   stateSize, 2, {stateSize, stateSize}, {inputSize, stateSize}, 1, {stateSize}, {1},
                                   batchSize, (returnSequence) ? timeSteps : 1, stateSize, fA),
     fTimeSteps(timeSteps), fStateSize(stateSize), fRememberState(rememberState), fReturnSequence(returnSequence), fF(f), fState(batchSize, stateSize),
     fWeightsInput(this->GetWeightsAt(0)), fWeightsState(this->GetWeightsAt(1)),
     fBiases(this->GetBiasesAt(0)), fDerivatives(timeSteps, batchSize, stateSize), // create tensor time x bs x S
     fWeightInputGradients(this->GetWeightGradientsAt(0)), fWeightStateGradients(this->GetWeightGradientsAt(1)),
     fBiasGradients(this->GetBiasGradientsAt(0)), fWeightsTensor({0}), fWeightGradientsTensor({0})
{
   InitTensors();
}
 
//______________________________________________________________________________
template <typename Architecture_t>
TBasicRNNLayer<Architecture_t>::TBasicRNNLayer(const TBasicRNNLayer &layer)
   : VGeneralLayer<Architecture_t>(layer), fTimeSteps(layer.fTimeSteps), fStateSize(layer.fStateSize),
     fRememberState(layer.fRememberState), fReturnSequence(layer.fReturnSequence), fF(layer.GetActivationFunction()),
     fState(layer.GetBatchSize(), layer.GetStateSize()),
     fWeightsInput(this->GetWeightsAt(0)), fWeightsState(this->GetWeightsAt(1)), fBiases(this->GetBiasesAt(0)),
     fDerivatives(layer.GetDerivatives().GetShape()), fWeightInputGradients(this->GetWeightGradientsAt(0)),
     fWeightStateGradients(this->GetWeightGradientsAt(1)), fBiasGradients(this->GetBiasGradientsAt(0)),
     fWeightsTensor({0}), fWeightGradientsTensor({0})
{
 
   Architecture_t::Copy(fDerivatives, layer.GetDerivatives() );
 
   // Gradient matrices not copied
   Architecture_t::Copy(fState, layer.GetState());
   InitTensors();
}
 
template <typename Architecture_t>
TBasicRNNLayer<Architecture_t>::~TBasicRNNLayer()
{
   if (fDescriptors) {
      Architecture_t::ReleaseRNNDescriptors(fDescriptors);
      delete fDescriptors;
   }
 
   if (fWorkspace) {
      Architecture_t::FreeRNNWorkspace(fWorkspace);
      delete fWorkspace;
   }
}
 
//______________________________________________________________________________
template<typename Architecture_t>
void TBasicRNNLayer<Architecture_t>::Initialize()
{
   // auto m = this->GetInitialization();
   // DNN::initialize<Architecture_t>(fWeightsInput, m);
   // DNN::initialize<Architecture_t>(fWeightsState, m);
   // DNN::initialize<Architecture_t>(fBiases,  DNN::EInitialization::kZero);
 
   VGeneralLayer<Architecture_t>::Initialize();
 
   Architecture_t::InitializeRNNDescriptors(fDescriptors, this);
   Architecture_t::InitializeRNNWorkspace(fWorkspace, fDescriptors, this);
}
 
//______________________________________________________________________________
template <typename Architecture_t>
void TBasicRNNLayer<Architecture_t>::InitTensors()
{
   // fix output tensor for Cudnn must be a tensor of B x T x S of right layout
   Architecture_t::InitializeRNNTensors(this);
}
//______________________________________________________________________________
template <typename Architecture_t>
auto TBasicRNNLayer<Architecture_t>::InitState(DNN::EInitialization /*m*/) -> void
{
   DNN::initialize<Architecture_t>(this->GetState(),  DNN::EInitialization::kZero);
 
   Architecture_t::InitializeActivationDescriptor(fActivationDesc,this->GetActivationFunction());
}
 
//______________________________________________________________________________
template<typename Architecture_t>
auto TBasicRNNLayer<Architecture_t>::Print() const
-> void
{
   std::cout << " RECURRENT Layer: \t ";
   std::cout << " (NInput = " << this->GetInputSize();  // input size
   std::cout << ", NState = " << this->GetStateSize();  // hidden state size
   std::cout << ", NTime  = " << this->GetTimeSteps() << " )";  // time size
   std::cout << "\tOutput = ( " << this->GetOutput().GetFirstSize() << " , " << this->GetOutput().GetHSize() << " , " << this->GetOutput().GetWSize() << " )\n";
}
 
template <typename Architecture_t>
auto debugMatrix(const typename Architecture_t::Matrix_t &A, const std::string name = "matrix")
-> void
{
  std::cout << name << "\n";
  for (size_t i = 0; i < A.GetNrows(); ++i) {
    for (size_t j = 0; j < A.GetNcols(); ++j) {
        std::cout << A(i, j) << " ";
    }
    std::cout << "\n";
  }
  std::cout << "********\n";
}
 
 
//______________________________________________________________________________
template <typename Architecture_t>
void TBasicRNNLayer<Architecture_t>::Forward(Tensor_t &input, bool isTraining ) // B x T x D
{
 
   //printf("doing RNNLayer forward\n");
   // for Cudnn
   if (Architecture_t::IsCudnn()) {
 
      Tensor_t &x = this->fX;
      Tensor_t &y = this->fY;
 
      Architecture_t::Rearrange(x, input);
 
      // why passing the first weight, better to pass all weight tensor (including bias)
      // LM 05/24
      //const auto &weights = this->GetWeightsAt(0);
      const auto & weights = this->GetWeightsTensor();
 
      // Tensor_t cx({1}); // not used for normal RNN
      // Tensor_t cy({1}); // not used for normal RNN
 
      // hx is fState - tensor are of right shape
      auto &hx = this->GetState();
      auto &cx = this->GetCell();
      // use same for hy and cy
      auto &hy = this->GetState();
      auto &cy = this->GetCell();
 
      auto & rnnDesc = static_cast<RNNDescriptors_t &>(*fDescriptors);
      auto & rnnWork = static_cast<RNNWorkspace_t &>(*fWorkspace);
 
      //printf("doing RNNLayer forward  - calling cudnn forwsrd\n");
 
      Architecture_t::RNNForward(x, hx, cx, weights, y, hy, cy, rnnDesc, rnnWork, isTraining);
 
      if (fReturnSequence) {
         Architecture_t::Rearrange(this->GetOutput(), y);    // swap B and T from y to Output
      }
      else {
         // tmp is a reference to y (full cudnn output)
         Tensor_t tmp = (y.At(y.GetShape()[0] - 1)).Reshape({y.GetShape()[1], 1, y.GetShape()[2]});
         Architecture_t::Copy(this->GetOutput(), tmp);
      }
      return;
   }
 
   // FORWARD for CPU architecture
   // D : input size
   // H : state size
   // T : time size
   // B : batch size
 
   Tensor_t arrInput (fTimeSteps, this->GetBatchSize(), this->GetInputWidth() );
   //for (size_t t = 0; t < fTimeSteps; ++t) arrInput.emplace_back(this->GetBatchSize(), this->GetInputWidth()); // T x B x D
   Architecture_t::Rearrange(arrInput, input);
   Tensor_t arrOutput ( fTimeSteps, this->GetBatchSize(), fStateSize);
   //for (size_t t = 0; t < fTimeSteps;++t) arrOutput.emplace_back(this->GetBatchSize(), fStateSize); // T x B x H
 
   if (!this->fRememberState) InitState(DNN::EInitialization::kZero);
 
   for (size_t t = 0; t < fTimeSteps; ++t) {
      Matrix_t arrInput_m = arrInput.At(t).GetMatrix();
      Matrix_t df_m = fDerivatives.At(t).GetMatrix();
      CellForward(arrInput_m, df_m );
      Matrix_t arrOutput_m = arrOutput.At(t).GetMatrix();
      Architecture_t::Copy(arrOutput_m, fState);
   }
 
   if (fReturnSequence)
      Architecture_t::Rearrange(this->GetOutput(), arrOutput);  // B x T x D
   else {
      // get T[end[]]
 
      Tensor_t tmp = arrOutput.At(fTimeSteps - 1); // take last time step
      // shape of tmp is  for CPU (column wise) B x D ,   need to reshape to  make a B x D x 1
      //  and transpose it to 1 x D x B  (this is how output is expected in columnmajor format)
      tmp = tmp.Reshape({tmp.GetShape()[0], tmp.GetShape()[1], 1});
      assert(tmp.GetSize() == this->GetOutput().GetSize());
      assert(tmp.GetShape()[0] == this->GetOutput().GetShape()[2]); // B is last dim in output and first in tmp
      Architecture_t::Rearrange(this->GetOutput(), tmp);
      // keep array output
      fY = arrOutput;
   }
}
 
//______________________________________________________________________________
template <typename Architecture_t>
auto inline TBasicRNNLayer<Architecture_t>::CellForward(const Matrix_t &input, Matrix_t &dF)
-> void
{
   // State = act(W_input . input + W_state . state + bias)
   const DNN::EActivationFunction fAF = this->GetActivationFunction();
   Matrix_t tmpState(fState.GetNrows(), fState.GetNcols());
   Architecture_t::MultiplyTranspose(tmpState, fState, fWeightsState);
   Architecture_t::MultiplyTranspose(fState, input, fWeightsInput);
   Architecture_t::ScaleAdd(fState, tmpState);
   Architecture_t::AddRowWise(fState, fBiases);
   Tensor_t inputActivFunc(dF);
   Tensor_t tState(fState);
 
   // DNN::evaluateDerivative<Architecture_t>(dFt, fAF, fState);
   // DNN::evaluate<Architecture_t>(tState, fAF);
 
   Architecture_t::Copy(inputActivFunc, tState);
   Architecture_t::ActivationFunctionForward(tState, fAF, fActivationDesc);
 
}
 
//____________________________________________________________________________
template <typename Architecture_t>
auto inline TBasicRNNLayer<Architecture_t>::Backward(Tensor_t &gradients_backward,         // B x T x D
                                                     const Tensor_t &activations_backward) -> void  // B x T x D
                                                   //   std::vector<Matrix_t> & /*inp1*/, std::vector<Matrix_t> &
                                                   //   /*inp2*/) -> void
{
   //BACKWARD for CUDNN
   if (Architecture_t::IsCudnn() ) {
 
      Tensor_t &x = this->fX;
      Tensor_t &y = this->fY;
      Tensor_t &dx = this->fDx;
      Tensor_t &dy = this->fDy;
 
      // input size is stride[1] of input tensor that is B x T x inputSize
      assert(activations_backward.GetStrides()[1] == this->GetInputSize() );
 
      Architecture_t::Rearrange(x, activations_backward);
 
      if (!fReturnSequence) {
 
         //Architecture_t::InitializeZero(dy);
         Architecture_t::InitializeZero(dy);
 
         //Tensor_t tmp1 = y.At(y.GetShape()[0] - 1).Reshape({y.GetShape()[1], 1, y.GetShape()[2]});
         Tensor_t tmp2 = dy.At(dy.GetShape()[0] - 1).Reshape({dy.GetShape()[1], 1, dy.GetShape()[2]});
 
         //Architecture_t::Copy(tmp1, this->GetOutput());
         Architecture_t::Copy(tmp2, this->GetActivationGradients());
      }
      else {
         Architecture_t::Rearrange(y, this->GetOutput());
         Architecture_t::Rearrange(dy, this->GetActivationGradients());
      }
 
 
 
      // for cudnn Matrix_t and Tensor_t are same type
      //const auto &weights = this->GetWeightsTensor();
      auto &weights = this->GetWeightsTensor();
      auto &weightGradients = this->GetWeightGradientsTensor();
      // note that cudnnRNNBackwardWeights accumulate the weight gradients.
      // We need then to initialize the tensor to zero every time
      Architecture_t::InitializeZero(weightGradients);
 
 
      // hx is fState
      auto &hx = this->GetState();
      auto &cx = this->GetCell();
      // use same for hy and cy
      auto &dhy = hx;
      auto &dcy = cx;
      auto &dhx = hx;
      auto &dcx = cx;
 
 
      auto & rnnDesc = static_cast<RNNDescriptors_t &>(*fDescriptors);
      auto & rnnWork = static_cast<RNNWorkspace_t &>(*fWorkspace);
 
      Architecture_t::RNNBackward(x, hx, cx, y, dy, dhy, dcy, weights, dx, dhx, dcx, weightGradients, rnnDesc, rnnWork);
 
      if (gradients_backward.GetSize() != 0)
         Architecture_t::Rearrange(gradients_backward, dx);
 
      return;
   }
 
   // BACKWARD FOR CPU
   // activations backward is input
   // gradients_backward is activationGradients of layer before it, which is input layer
   // currently gradient_backward is for input(x) and not for state
   // TODO use this to change initial state??
 
 
  bool dummy = false;
  if (gradients_backward.GetSize() == 0) {
     dummy = true;
  }
  Tensor_t arr_gradients_backward ( fTimeSteps, this->GetBatchSize(), this->GetInputSize());
  //for (size_t t = 0; t < fTimeSteps; ++t) arr_gradients_backward.emplace_back(this->GetBatchSize(), this->GetInputSize()); // T x B x D
 
  if (!dummy) {
      // TODO gradients_backward will be written back on the matrix
     //Architecture_t::Rearrange(arr_gradients_backward, gradients_backward);
  }
  Tensor_t arr_activations_backward ( fTimeSteps, this->GetBatchSize(), this->GetInputSize());
  //for (size_t t = 0; t < fTimeSteps; ++t) arr_activations_backward.emplace_back(this->GetBatchSize(), this->GetInputSize());  // T x B x D
  Architecture_t::Rearrange(arr_activations_backward, activations_backward);
 
   Matrix_t state_gradients_backward(this->GetBatchSize(), fStateSize);  // B x H
   DNN::initialize<Architecture_t>(state_gradients_backward,  DNN::EInitialization::kZero);
 
   Matrix_t initState(this->GetBatchSize(), fStateSize);  // B x H
   DNN::initialize<Architecture_t>(initState,   DNN::EInitialization::kZero);
 
   Tensor_t arr_output (  fTimeSteps, this->GetBatchSize(), fStateSize);
   Tensor_t arr_actgradients(fTimeSteps, this->GetBatchSize(), fStateSize);
 
   if (fReturnSequence) {
      Architecture_t::Rearrange(arr_output, this->GetOutput());
      Architecture_t::Rearrange(arr_actgradients, this->GetActivationGradients());
   } else {
      //
      arr_output = fY;
 
      Architecture_t::InitializeZero(arr_actgradients);
      // need to reshape to pad a time dimension = 1 (note here is columnmajor tensors)
      Tensor_t tmp_grad = arr_actgradients.At(fTimeSteps - 1).Reshape({this->GetBatchSize(), fStateSize, 1});
      assert(tmp_grad.GetSize() == this->GetActivationGradients().GetSize());
      assert(tmp_grad.GetShape()[0] ==
             this->GetActivationGradients().GetShape()[2]); // B in tmp is [0] and [2] in input act. gradients
 
      Architecture_t::Rearrange(tmp_grad, this->GetActivationGradients());
   }
 
   // reinitialize weights and biases gradients to 0
   fWeightInputGradients.Zero();
   fWeightStateGradients.Zero();
   fBiasGradients.Zero();
 
   for (size_t t = fTimeSteps; t > 0; t--) {
      //const Matrix_t & currStateActivations = arr_output[t - 1];
      Matrix_t actgrad_m = arr_actgradients.At(t - 1).GetMatrix();
      Architecture_t::ScaleAdd(state_gradients_backward, actgrad_m);
 
      Matrix_t actbw_m = arr_activations_backward.At(t - 1).GetMatrix();
      Matrix_t gradbw_m = arr_gradients_backward.At(t - 1).GetMatrix();
 
      // compute derivatives of activations
      Tensor_t  df = fDerivatives.At(t-1);
      Tensor_t dy =  Tensor_t(state_gradients_backward);
      //Tensor_t dy =  arr_actgradients.At(t - 1);
      Tensor_t y = arr_output.At(t-1);
      Architecture_t::ActivationFunctionBackward(df, y,
                                                 dy, df, //do in place (should work)
                                              this->GetActivationFunction(), fActivationDesc);
 
      Matrix_t df_m = df.GetMatrix();
 
      // Architecture_t::PrintTensor(df, "dy before");
      if (t > 1) {
         Matrix_t precStateActivations = arr_output.At(t - 2).GetMatrix();
         CellBackward(state_gradients_backward, precStateActivations, actbw_m, gradbw_m, df_m);
 
      } else {
         const Matrix_t & precStateActivations = initState;
         CellBackward(state_gradients_backward, precStateActivations, actbw_m, gradbw_m, df_m);
 
      }
   }
   if (!dummy) {
      Architecture_t::Rearrange(gradients_backward, arr_gradients_backward );
   }
}
 
//______________________________________________________________________________
template <typename Architecture_t>
auto inline TBasicRNNLayer<Architecture_t>::CellBackward(Matrix_t & state_gradients_backward,
                                                     const Matrix_t & precStateActivations,
                                                     const Matrix_t & input, Matrix_t & input_gradient, Matrix_t &dF)
-> Matrix_t &
{
   return Architecture_t::RecurrentLayerBackward(state_gradients_backward, fWeightInputGradients, fWeightStateGradients,
                                                 fBiasGradients, dF, precStateActivations, fWeightsInput,
                                                 fWeightsState, input, input_gradient);
}
 
//______________________________________________________________________________
template <typename Architecture_t>
void TBasicRNNLayer<Architecture_t>::AddWeightsXMLTo(void *parent)
{
   auto layerxml = gTools().xmlengine().NewChild(parent, nullptr, "RNNLayer");
 
   // write All other info like stateSize, inputSize, timeSteps,rememberState
   gTools().xmlengine().NewAttr(layerxml, nullptr, "StateSize", gTools().StringFromInt(this->GetStateSize()));
   gTools().xmlengine().NewAttr(layerxml, nullptr, "InputSize", gTools().StringFromInt(this->GetInputSize()));
   gTools().xmlengine().NewAttr(layerxml, nullptr, "TimeSteps", gTools().StringFromInt(this->GetTimeSteps()));
   gTools().xmlengine().NewAttr(layerxml, nullptr, "RememberState", gTools().StringFromInt(this->DoesRememberState()));
   gTools().xmlengine().NewAttr(layerxml, nullptr, "ReturnSequence", gTools().StringFromInt(this->DoesReturnSequence()));
 
   // write weights and bias matrices
   this->WriteMatrixToXML(layerxml, "InputWeights", this -> GetWeightsAt(0));
   this->WriteMatrixToXML(layerxml, "StateWeights", this -> GetWeightsAt(1));
   this->WriteMatrixToXML(layerxml, "Biases",  this -> GetBiasesAt(0));
 
 
}
 
//______________________________________________________________________________
template <typename Architecture_t>
void TBasicRNNLayer<Architecture_t>::ReadWeightsFromXML(void *parent)
{
   // Read weights and biases
   this->ReadMatrixXML(parent,"InputWeights", this -> GetWeightsAt(0));
   this->ReadMatrixXML(parent,"StateWeights", this -> GetWeightsAt(1));
   this->ReadMatrixXML(parent,"Biases", this -> GetBiasesAt(0));
 
}
 
} // namespace RNN
} // namespace DNN
} // namespace TMVA
 
#endif