815 lines
56 KiB
MQL5
815 lines
56 KiB
MQL5
//+------------------------------------------------------------------+
|
|
//| NeuronLSTM.mqh |
|
|
//| Copyright 2021, MetaQuotes Ltd. |
|
|
//| https://www.mql5.com |
|
|
//+------------------------------------------------------------------+
|
|
#property copyright "Copyright 2021, MetaQuotes Ltd."
|
|
#property link "https://www.mql5.com"
|
|
//+------------------------------------------------------------------+
|
|
//| Connect libraries |
|
|
//+------------------------------------------------------------------+
|
|
#include "neuronbase.mqh"
|
|
#include <Arrays\ArrayObj.mqh>
|
|
//+------------------------------------------------------------------+
|
|
//| Class CNeuronLSTM |
|
|
//| Purpose: Class for implementing a recurrent LSTM block |
|
|
//+------------------------------------------------------------------+
|
|
class CNeuronLSTM : public CNeuronBase
|
|
{
|
|
protected:
|
|
CNeuronBase* m_cForgetGate;
|
|
CNeuronBase* m_cInputGate;
|
|
CNeuronBase* m_cNewContent;
|
|
CNeuronBase* m_cOutputGate;
|
|
CArrayObj* m_cMemorys;
|
|
CArrayObj* m_cHiddenStates;
|
|
CArrayObj* m_cInputs;
|
|
CArrayObj* m_cForgetGateOuts;
|
|
CArrayObj* m_cInputGateOuts;
|
|
CArrayObj* m_cNewContentOuts;
|
|
CArrayObj* m_cOutputGateOuts;
|
|
CBufferType* m_cInputGradient;
|
|
int m_iDepth;
|
|
|
|
void ClearBuffer(CArrayObj *buffer);
|
|
bool InsertBuffer(CArrayObj *&array, CBufferType *element, bool create_new = true);
|
|
CBufferType* CreateBuffer(CArrayObj *&array);
|
|
|
|
public:
|
|
CNeuronLSTM(void);
|
|
~CNeuronLSTM(void);
|
|
//---
|
|
virtual bool Init(const CLayerDescription *desc) override;
|
|
virtual bool SetOpenCL(CMyOpenCL *opencl) override;
|
|
virtual bool FeedForward(CNeuronBase *prevLayer) override;
|
|
virtual bool CalcHiddenGradient(CNeuronBase *prevLayer) override;
|
|
virtual bool CalcDeltaWeights(CNeuronBase *prevLayer, bool read) override
|
|
{ return (!m_cOpenCL ? true : m_cDeltaWeights.BufferRead()); }
|
|
virtual bool UpdateWeights(int batch_size, TYPE learningRate,
|
|
VECTOR &Beta, VECTOR &Lambda) override;
|
|
//---
|
|
virtual int GetDepth(void) const { return m_iDepth; }
|
|
//--- file handling methods
|
|
virtual bool Save(const int file_handle) override;
|
|
virtual bool Load(const int file_handle) override;
|
|
//--- object identification method
|
|
virtual int Type(void) override const { return(defNeuronLSTM); }
|
|
};
|
|
//+------------------------------------------------------------------+
|
|
//| Class constructor |
|
|
//+------------------------------------------------------------------+
|
|
CNeuronLSTM::CNeuronLSTM(void) : m_iDepth(2)
|
|
{
|
|
m_cForgetGate = new CNeuronBase();
|
|
m_cInputGate = new CNeuronBase();
|
|
m_cNewContent = new CNeuronBase();
|
|
m_cOutputGate = new CNeuronBase();
|
|
m_cMemorys = new CArrayObj();
|
|
m_cHiddenStates = new CArrayObj();
|
|
m_cInputs = new CArrayObj();
|
|
m_cForgetGateOuts = new CArrayObj();
|
|
m_cInputGateOuts = new CArrayObj();
|
|
m_cNewContentOuts = new CArrayObj();
|
|
m_cOutputGateOuts = new CArrayObj();
|
|
m_cInputGradient = new CBufferType();
|
|
}
|
|
//+------------------------------------------------------------------+
|
|
//| Class destructor |
|
|
//+------------------------------------------------------------------+
|
|
CNeuronLSTM::~CNeuronLSTM(void)
|
|
{
|
|
if(m_cForgetGate)
|
|
delete m_cForgetGate;
|
|
if(m_cInputGate)
|
|
delete m_cInputGate;
|
|
if(m_cNewContent)
|
|
delete m_cNewContent;
|
|
if(m_cOutputGate)
|
|
delete m_cOutputGate;
|
|
if(m_cMemorys)
|
|
delete m_cMemorys;
|
|
if(m_cHiddenStates)
|
|
delete m_cHiddenStates;
|
|
if(m_cInputs)
|
|
delete m_cInputs;
|
|
if(m_cForgetGateOuts)
|
|
delete m_cForgetGateOuts;
|
|
if(m_cInputGateOuts)
|
|
delete m_cInputGateOuts;
|
|
if(m_cNewContentOuts)
|
|
delete m_cNewContentOuts;
|
|
if(m_cOutputGateOuts)
|
|
delete m_cOutputGateOuts;
|
|
if(m_cInputGradient)
|
|
delete m_cInputGradient;
|
|
}
|
|
//+------------------------------------------------------------------+
|
|
//| Class initialization method |
|
|
//+------------------------------------------------------------------+
|
|
bool CNeuronLSTM::Init(const CLayerDescription *desc)
|
|
{
|
|
//--- control block
|
|
if(!desc || desc.type != Type() || desc.count <= 0 || desc.window == 0)
|
|
return false;
|
|
//--- create a description for the internal neural layers
|
|
CLayerDescription *temp = new CLayerDescription();
|
|
if(!temp)
|
|
return false;
|
|
temp.type = defNeuronBase;
|
|
temp.window = desc.window + desc.count;
|
|
temp.count = desc.count;
|
|
temp.activation = AF_SIGMOID;
|
|
temp.activation_params[0] = 1;
|
|
temp.activation_params[1] = 0;
|
|
temp.optimization = desc.optimization;
|
|
//--- call the initialization method of the parent class
|
|
CLayerDescription *temp2 = new CLayerDescription();
|
|
if(!temp2 || !temp2.Copy(desc))
|
|
return false;
|
|
temp2.window = 0;
|
|
if(!CNeuronBase::Init(temp2))
|
|
return false;
|
|
delete temp2;
|
|
if(!InsertBuffer(m_cHiddenStates, m_cOutputs, false))
|
|
return false;
|
|
m_iDepth = (int)fmax(desc.window_out, 2);
|
|
//--- initialize ForgetGate
|
|
if(!m_cForgetGate)
|
|
{
|
|
if(!(m_cForgetGate = new CNeuronBase()))
|
|
return false;
|
|
}
|
|
if(!m_cForgetGate.Init(temp))
|
|
return false;
|
|
if(!InsertBuffer(m_cForgetGateOuts, m_cForgetGate.GetOutputs(), false))
|
|
return false;
|
|
//--- initialize InputGate
|
|
if(!m_cInputGate)
|
|
{
|
|
if(!(m_cInputGate = new CNeuronBase()))
|
|
return false;
|
|
}
|
|
if(!m_cInputGate.Init(temp))
|
|
return false;
|
|
if(!InsertBuffer(m_cInputGateOuts, m_cInputGate.GetOutputs(), false))
|
|
return false;
|
|
//--- initialize OutputGate
|
|
if(!m_cOutputGate)
|
|
{
|
|
if(!(m_cOutputGate = new CNeuronBase()))
|
|
return false;
|
|
}
|
|
if(!m_cOutputGate.Init(temp))
|
|
return false;
|
|
if(!InsertBuffer(m_cOutputGateOuts, m_cOutputGate.GetOutputs(), false))
|
|
return false;
|
|
//--- initialize NewContent
|
|
if(!m_cNewContent)
|
|
{
|
|
if(!(m_cNewContent = new CNeuronBase()))
|
|
return false;
|
|
}
|
|
temp.activation = AF_TANH;
|
|
if(!m_cNewContent.Init(temp))
|
|
return false;
|
|
if(!InsertBuffer(m_cNewContentOuts, m_cNewContent.GetOutputs(), false))
|
|
return false;
|
|
//--- initialize InputGradient buffer
|
|
if(!m_cInputGradient)
|
|
{
|
|
if(!(m_cInputGradient = new CBufferType()))
|
|
return false;
|
|
}
|
|
if(!m_cInputGradient.BufferInit(1, temp.window, 0))
|
|
return false;
|
|
delete temp;
|
|
//--- initialize Memory
|
|
CBufferType *buffer = CreateBuffer(m_cMemorys);
|
|
if(!buffer)
|
|
return false;
|
|
if(!InsertBuffer(m_cMemorys, buffer, false))
|
|
{
|
|
delete buffer;
|
|
return false;
|
|
}
|
|
//--- initialize HiddenStates
|
|
if(!(buffer = CreateBuffer(m_cHiddenStates)))
|
|
return false;
|
|
if(!InsertBuffer(m_cHiddenStates, buffer, false))
|
|
{
|
|
delete buffer;
|
|
return false;
|
|
}
|
|
//---
|
|
SetOpenCL(m_cOpenCL);
|
|
//---
|
|
return true;
|
|
}
|
|
//+------------------------------------------------------------------+
|
|
//| Method for passing a pointer to the OpenCL object to all |
|
|
//| internal objects |
|
|
//+------------------------------------------------------------------+
|
|
bool CNeuronLSTM::SetOpenCL(CMyOpenCL *opencl)
|
|
{
|
|
//--- call of the method of the parent class
|
|
CNeuronBase::SetOpenCL(opencl);
|
|
//--- call the relevant method for all internal layers
|
|
m_cForgetGate.SetOpenCL(m_cOpenCL);
|
|
m_cInputGate.SetOpenCL(m_cOpenCL);
|
|
m_cOutputGate.SetOpenCL(m_cOpenCL);
|
|
m_cNewContent.SetOpenCL(m_cOpenCL);
|
|
m_cInputGradient.BufferCreate(m_cOpenCL);
|
|
for(int i = 0; i < m_cMemorys.Total(); i++)
|
|
{
|
|
CBufferType *temp = m_cMemorys.At(i);
|
|
temp.BufferCreate(m_cOpenCL);
|
|
}
|
|
for(int i = 0; i < m_cHiddenStates.Total(); i++)
|
|
{
|
|
CBufferType *temp = m_cHiddenStates.At(i);
|
|
temp.BufferCreate(m_cOpenCL);
|
|
}
|
|
//---
|
|
return(!!m_cOpenCL);
|
|
}
|
|
//+------------------------------------------------------------------+
|
|
//| Method for deleting unnecessary data from the stack |
|
|
//+------------------------------------------------------------------+
|
|
void CNeuronLSTM::ClearBuffer(CArrayObj *buffer)
|
|
{
|
|
if(!buffer)
|
|
return;
|
|
int total = buffer.Total();
|
|
if(total > m_iDepth + 1)
|
|
buffer.DeleteRange(m_iDepth + 1, total);
|
|
}
|
|
//+------------------------------------------------------------------+
|
|
//| Feed-forward method |
|
|
//+------------------------------------------------------------------+
|
|
bool CNeuronLSTM::FeedForward(CNeuronBase *prevLayer)
|
|
{
|
|
//--- check the relevance of all objects
|
|
if(!prevLayer || !prevLayer.GetOutputs() || !m_cOutputs ||
|
|
!m_cForgetGate || !m_cInputGate || !m_cOutputGate ||
|
|
!m_cNewContent)
|
|
return false;
|
|
//--- prepare blanks for new buffers
|
|
if(!m_cForgetGate.SetOutputs(CreateBuffer(m_cForgetGateOuts), false))
|
|
return false;
|
|
if(!m_cInputGate.SetOutputs(CreateBuffer(m_cInputGateOuts), false))
|
|
return false;
|
|
if(!m_cOutputGate.SetOutputs(CreateBuffer(m_cOutputGateOuts), false))
|
|
return false;
|
|
if(!m_cNewContent.SetOutputs(CreateBuffer(m_cNewContentOuts), false))
|
|
return false;
|
|
CBufferType *memory = CreateBuffer(m_cMemorys);
|
|
if(!memory)
|
|
return false;
|
|
CBufferType *hidden = CreateBuffer(m_cHiddenStates);
|
|
if(!hidden)
|
|
{
|
|
delete memory;
|
|
return false;
|
|
}
|
|
//--- only to check the gradient
|
|
//memory.m_mMatrix.Fill(0);
|
|
//hidden.m_mMatrix.Fill(0);
|
|
//if(!!m_cOpenCL)
|
|
// {
|
|
// memory.BufferWrite();
|
|
// hidden.BufferWrite();
|
|
// }
|
|
//--- create the buffer of source data
|
|
if(!m_cInputs)
|
|
{
|
|
m_cInputs = new CArrayObj();
|
|
if(!m_cInputs)
|
|
{
|
|
delete memory;
|
|
delete hidden;
|
|
return false;
|
|
}
|
|
}
|
|
CNeuronBase *inputs = new CNeuronBase();
|
|
if(!inputs)
|
|
{
|
|
delete memory;
|
|
delete hidden;
|
|
return false;
|
|
}
|
|
CLayerDescription *desc = new CLayerDescription();
|
|
if(!desc)
|
|
{
|
|
delete inputs;
|
|
delete memory;
|
|
delete hidden;
|
|
return false;
|
|
}
|
|
desc.type = defNeuronBase;
|
|
desc.count = (int)(prevLayer.GetOutputs().Total() + m_cOutputs.Total());
|
|
desc.window = 0;
|
|
if(!inputs.Init(desc))
|
|
{
|
|
delete inputs;
|
|
delete memory;
|
|
delete hidden;
|
|
delete desc;
|
|
return false;
|
|
}
|
|
delete desc;
|
|
inputs.SetOpenCL(m_cOpenCL);
|
|
CBufferType *inputs_buffer = inputs.GetOutputs();
|
|
if(!inputs_buffer)
|
|
{
|
|
delete inputs;
|
|
delete memory;
|
|
delete hidden;
|
|
return false;
|
|
}
|
|
if(!inputs_buffer.Concatenate(prevLayer.GetOutputs(), hidden, prevLayer.Total(), hidden.Total()))
|
|
{
|
|
delete inputs;
|
|
delete memory;
|
|
delete hidden;
|
|
return false;
|
|
}
|
|
//--- feed-forward pass through internal neural layers
|
|
if(!m_cForgetGate.FeedForward(inputs))
|
|
{
|
|
delete inputs;
|
|
delete memory;
|
|
delete hidden;
|
|
return false;
|
|
}
|
|
if(!m_cInputGate.FeedForward(inputs))
|
|
{
|
|
delete inputs;
|
|
delete memory;
|
|
delete hidden;
|
|
return false;
|
|
}
|
|
if(!m_cOutputGate.FeedForward(inputs))
|
|
{
|
|
delete inputs;
|
|
delete memory;
|
|
delete hidden;
|
|
return false;
|
|
}
|
|
if(!m_cNewContent.FeedForward(inputs))
|
|
{
|
|
delete inputs;
|
|
delete memory;
|
|
delete hidden;
|
|
return false;
|
|
}
|
|
//--- branching of the algorithm across computing devices
|
|
CBufferType *fg = m_cForgetGate.GetOutputs();
|
|
CBufferType *ig = m_cInputGate.GetOutputs();
|
|
CBufferType *og = m_cOutputGate.GetOutputs();
|
|
CBufferType *nc = m_cNewContent.GetOutputs();
|
|
if(!m_cOpenCL)
|
|
{
|
|
memory.m_mMatrix *= fg.m_mMatrix;
|
|
memory.m_mMatrix += ig.m_mMatrix * nc.m_mMatrix;
|
|
hidden.m_mMatrix = MathTanh(memory.m_mMatrix) * og.m_mMatrix;
|
|
}
|
|
else
|
|
{
|
|
//--- check buffers
|
|
if(fg.GetIndex() < 0 || ig.GetIndex() < 0 || og.GetIndex() < 0 ||
|
|
nc.GetIndex() < 0 || memory.GetIndex() < 0 || hidden.GetIndex() < 0)
|
|
return false;
|
|
//--- pass parameters to the kernel
|
|
if(!m_cOpenCL.SetArgumentBuffer(def_k_LSTMFeedForward, def_lstmff_forgetgate, fg.GetIndex()))
|
|
return false;
|
|
if(!m_cOpenCL.SetArgumentBuffer(def_k_LSTMFeedForward, def_lstmff_inputgate, ig.GetIndex()))
|
|
return false;
|
|
if(!m_cOpenCL.SetArgumentBuffer(def_k_LSTMFeedForward, def_lstmff_newcontent, nc.GetIndex()))
|
|
return false;
|
|
if(!m_cOpenCL.SetArgumentBuffer(def_k_LSTMFeedForward, def_lstmff_outputgate, og.GetIndex()))
|
|
return false;
|
|
if(!m_cOpenCL.SetArgumentBuffer(def_k_LSTMFeedForward, def_lstmff_memory, memory.GetIndex()))
|
|
return false;
|
|
if(!m_cOpenCL.SetArgumentBuffer(def_k_LSTMFeedForward, def_lstmff_hiddenstate, hidden.GetIndex()))
|
|
return false;
|
|
if(!m_cOpenCL.SetArgument(def_k_LSTMFeedForward, def_lstmff_outputs_total, (int)m_cOutputs.Total()))
|
|
return false;
|
|
//--- launch the kernel
|
|
int NDRange[] = {(int)(m_cOutputs.Total() + 3) / 4};
|
|
int off_set[] = {0};
|
|
if(!m_cOpenCL.Execute(def_k_LSTMFeedForward, 1, off_set, NDRange))
|
|
return false;
|
|
}
|
|
//--- copy the hidden state to the results buffer of the neural layer
|
|
m_cOutputs = hidden;
|
|
//--- save the current state
|
|
if(!m_cInputs.Insert(inputs, 0))
|
|
{
|
|
delete inputs;
|
|
delete memory;
|
|
delete hidden;
|
|
return false;
|
|
}
|
|
ClearBuffer(m_cInputs);
|
|
if(!InsertBuffer(m_cForgetGateOuts, m_cForgetGate.GetOutputs(), false))
|
|
{
|
|
delete memory;
|
|
delete hidden;
|
|
return false;
|
|
}
|
|
if(!InsertBuffer(m_cInputGateOuts, m_cInputGate.GetOutputs(), false))
|
|
{
|
|
delete memory;
|
|
delete hidden;
|
|
return false;
|
|
}
|
|
if(!InsertBuffer(m_cOutputGateOuts, m_cOutputGate.GetOutputs(), false))
|
|
{
|
|
delete memory;
|
|
delete hidden;
|
|
return false;
|
|
}
|
|
if(!InsertBuffer(m_cNewContentOuts, m_cNewContent.GetOutputs(), false))
|
|
{
|
|
delete memory;
|
|
delete hidden;
|
|
return false;
|
|
}
|
|
if(!InsertBuffer(m_cMemorys, memory, false))
|
|
{
|
|
delete hidden;
|
|
return false;
|
|
}
|
|
if(!InsertBuffer(m_cHiddenStates, hidden, false))
|
|
return false;
|
|
//---
|
|
return true;
|
|
}
|
|
//+------------------------------------------------------------------+
|
|
//| Method for adding data to the stack |
|
|
//+------------------------------------------------------------------+
|
|
bool CNeuronLSTM::InsertBuffer(CArrayObj *&array, CBufferType *element, bool create_new = true)
|
|
{
|
|
//--- control block
|
|
if(!element)
|
|
return false;
|
|
if(!array)
|
|
{
|
|
array = new CArrayObj();
|
|
if(!array)
|
|
return false;
|
|
}
|
|
//---
|
|
if(create_new)
|
|
{
|
|
CBufferType *buffer = new CBufferType();
|
|
if(!buffer)
|
|
return false;
|
|
buffer.m_mMatrix = element.m_mMatrix;
|
|
if(!array.Insert(buffer, 0))
|
|
{
|
|
delete buffer;
|
|
return false;
|
|
}
|
|
}
|
|
else
|
|
{
|
|
if(!array.Insert(element, 0))
|
|
{
|
|
delete element;
|
|
return false;
|
|
}
|
|
}
|
|
//--- remove unnecessary history from the buffer
|
|
ClearBuffer(array);
|
|
//---
|
|
return true;
|
|
}
|
|
//+------------------------------------------------------------------+
|
|
//| Method for creating a new data buffer |
|
|
//+------------------------------------------------------------------+
|
|
CBufferType *CNeuronLSTM::CreateBuffer(CArrayObj *&array)
|
|
{
|
|
if(!array)
|
|
{
|
|
array = new CArrayObj();
|
|
if(!array)
|
|
return NULL;
|
|
}
|
|
CBufferType *buffer = new CBufferType();
|
|
if(!buffer)
|
|
return NULL;
|
|
if(array.Total() <= 0)
|
|
{
|
|
if(!buffer.BufferInit(m_cOutputs.Rows(), m_cOutputs.Cols(), 0))
|
|
{
|
|
delete buffer;
|
|
return NULL;
|
|
}
|
|
}
|
|
else
|
|
{
|
|
CBufferType *temp = array.At(0);
|
|
if(!temp)
|
|
{
|
|
delete buffer;
|
|
return NULL;
|
|
}
|
|
buffer.m_mMatrix = temp.m_mMatrix;
|
|
}
|
|
//---
|
|
if(m_cOpenCL)
|
|
{
|
|
if(!buffer.BufferCreate(m_cOpenCL))
|
|
delete buffer;
|
|
}
|
|
//---
|
|
return buffer;
|
|
}
|
|
//+------------------------------------------------------------------+
|
|
//| Method for propagating error gradient through hidden layer |
|
|
//+------------------------------------------------------------------+
|
|
bool CNeuronLSTM::CalcHiddenGradient(CNeuronBase *prevLayer)
|
|
{
|
|
//--- check the relevance of all objects
|
|
if(!prevLayer || !prevLayer.GetGradients() ||
|
|
!m_cGradients || !m_cForgetGate || !m_cForgetGateOuts ||
|
|
!m_cInputGate || !m_cInputGateOuts || !m_cOutputGate ||
|
|
!m_cOutputGateOuts || !m_cNewContent || !m_cNewContentOuts)
|
|
return false;
|
|
//--- check the presence of feed-forward data
|
|
int total = (int)fmin(m_cMemorys.Total(), m_cHiddenStates.Total()) - 1;
|
|
if(total <= 0)
|
|
return false;
|
|
//--- make pointers to buffers of gradients and results of internal layers
|
|
CBufferType *fg_grad = m_cForgetGate.GetGradients();
|
|
if(!fg_grad)
|
|
return false;
|
|
CBufferType *fg_out = m_cForgetGate.GetOutputs();
|
|
if(!fg_out)
|
|
return false;
|
|
CBufferType *ig_grad = m_cInputGate.GetGradients();
|
|
if(!ig_grad)
|
|
return false;
|
|
CBufferType *ig_out = m_cInputGate.GetOutputs();
|
|
if(!ig_out)
|
|
return false;
|
|
CBufferType *og_grad = m_cOutputGate.GetGradients();
|
|
if(!og_grad)
|
|
return false;
|
|
CBufferType *og_out = m_cOutputGate.GetOutputs();
|
|
if(!og_out)
|
|
return false;
|
|
CBufferType *nc_grad = m_cNewContent.GetGradients();
|
|
if(!nc_grad)
|
|
return false;
|
|
CBufferType *nc_out = m_cNewContent.GetOutputs();
|
|
if(!nc_out)
|
|
return false;
|
|
//---
|
|
ulong out_total = m_cOutputs.Total();
|
|
//--- loop through the accumulated history
|
|
for(int i = 0; i < total; i++)
|
|
{
|
|
//--- get pointers to buffers from the stack
|
|
CBufferType *fg = m_cForgetGateOuts.At(i);
|
|
if(!fg)
|
|
return false;
|
|
CBufferType *ig = m_cInputGateOuts.At(i);
|
|
if(!ig)
|
|
return false;
|
|
CBufferType *og = m_cOutputGateOuts.At(i);
|
|
if(!og)
|
|
return false;
|
|
CBufferType *nc = m_cNewContentOuts.At(i);
|
|
if(!nc)
|
|
return false;
|
|
CBufferType *memory = m_cMemorys.At(i + 1);
|
|
if(!memory)
|
|
return false;
|
|
CBufferType *hidden = m_cHiddenStates.At(i);
|
|
if(!hidden)
|
|
return false;
|
|
CNeuronBase *inputs = m_cInputs.At(i);
|
|
if(!inputs)
|
|
return false;
|
|
//--- branching of the algorithm across computing devices
|
|
if(!m_cOpenCL)
|
|
{
|
|
//--- calculate the gradient at the output of each internal layer
|
|
MATRIX m = hidden.m_mMatrix / (og.m_mMatrix + 1e-8);
|
|
//--- OutputGate gradient
|
|
MATRIX grad = m_cGradients.m_mMatrix;
|
|
og_grad.m_mMatrix = grad * m;
|
|
//--- memory gradient
|
|
grad *= og.m_mMatrix;
|
|
//--- adjust the gradient to the derivative
|
|
grad *= MathPow(m, 2) * (-1) + 1;
|
|
//--- InputGate gradient
|
|
ig_grad.m_mMatrix = grad * nc.m_mMatrix;
|
|
//--- NewContent gradient
|
|
nc_grad.m_mMatrix = grad * ig.m_mMatrix;
|
|
//--- ForgetGates gradient
|
|
fg_grad.m_mMatrix = grad * memory.m_mMatrix;
|
|
}
|
|
else
|
|
{
|
|
//--- check buffers
|
|
if(hidden.GetIndex() < 0)
|
|
return false;
|
|
if(m_cGradients.GetIndex() < 0)
|
|
return false;
|
|
if(ig.GetIndex() < 0)
|
|
return false;
|
|
if(og.GetIndex() < 0)
|
|
return false;
|
|
if(nc.GetIndex() < 0)
|
|
return false;
|
|
if(memory.GetIndex() < 0)
|
|
return false;
|
|
if(fg_grad.GetIndex() < 0)
|
|
return false;
|
|
if(ig_grad.GetIndex() < 0)
|
|
return false;
|
|
if(og_grad.GetIndex() < 0)
|
|
return false;
|
|
if(nc_grad.GetIndex() < 0)
|
|
return false;
|
|
//--- pass parameters to the kernel
|
|
if(!m_cOpenCL.SetArgumentBuffer(def_k_LSTMHiddenGradients, def_lstmhgr_fg_gradients, fg_grad.GetIndex()))
|
|
return false;
|
|
if(!m_cOpenCL.SetArgumentBuffer(def_k_LSTMHiddenGradients, def_lstmhgr_gradients, m_cGradients.GetIndex()))
|
|
return false;
|
|
if(!m_cOpenCL.SetArgumentBuffer(def_k_LSTMHiddenGradients, def_lstmhgr_ig_gradients, ig_grad.GetIndex()))
|
|
return false;
|
|
if(!m_cOpenCL.SetArgumentBuffer(def_k_LSTMHiddenGradients, def_lstmhgr_inputgate, ig.GetIndex()))
|
|
return false;
|
|
if(!m_cOpenCL.SetArgumentBuffer(def_k_LSTMHiddenGradients, def_lstmhgr_memory, memory.GetIndex()))
|
|
return false;
|
|
if(!m_cOpenCL.SetArgumentBuffer(def_k_LSTMHiddenGradients, def_lstmhgr_nc_gradients, nc_grad.GetIndex()))
|
|
return false;
|
|
if(!m_cOpenCL.SetArgumentBuffer(def_k_LSTMHiddenGradients, def_lstmhgr_newcontent, nc.GetIndex()))
|
|
return false;
|
|
if(!m_cOpenCL.SetArgumentBuffer(def_k_LSTMHiddenGradients, def_lstmhgr_og_gradients, og_grad.GetIndex()))
|
|
return false;
|
|
if(!m_cOpenCL.SetArgumentBuffer(def_k_LSTMHiddenGradients, def_lstmhgr_outputgate, og.GetIndex()))
|
|
return false;
|
|
if(!m_cOpenCL.SetArgumentBuffer(def_k_LSTMHiddenGradients, def_lstmhgr_outputs, hidden.GetIndex()))
|
|
return false;
|
|
if(!m_cOpenCL.SetArgument(def_k_LSTMHiddenGradients, def_lstmhgr_outputs_total, (int)m_cOutputs.Total()))
|
|
return false;
|
|
//--- launch the kernel
|
|
int NDRange[] = { (int)(m_cOutputs.Total() + 3) / 4 };
|
|
int off_set[] = {0};
|
|
if(!m_cOpenCL.Execute(def_k_LSTMHiddenGradients, 1, off_set, NDRange))
|
|
return false;
|
|
}
|
|
//--- copy the corresponding historical data to the buffers of internal layers
|
|
if(!m_cForgetGate.SetOutputs(fg, false))
|
|
return false;
|
|
if(!m_cInputGate.SetOutputs(ig, false))
|
|
return false;
|
|
if(!m_cOutputGate.SetOutputs(og, false))
|
|
return false;
|
|
if(!m_cNewContent.SetOutputs(nc, false))
|
|
return false;
|
|
//--- propagate gradient through the inner layers
|
|
if(!m_cForgetGate.CalcHiddenGradient(inputs))
|
|
return false;
|
|
if(!m_cInputGradient)
|
|
{
|
|
m_cInputGradient = new CBufferType();
|
|
if(!m_cInputGradient)
|
|
return false;
|
|
m_cInputGradient.m_mMatrix = inputs.GetGradients().m_mMatrix;
|
|
m_cInputGradient.BufferCreate(m_cOpenCL);
|
|
}
|
|
else
|
|
{
|
|
m_cInputGradient.Scaling(0);
|
|
if(!m_cInputGradient.SumArray(inputs.GetGradients()))
|
|
return false;
|
|
}
|
|
if(!m_cInputGate.CalcHiddenGradient(inputs))
|
|
return false;
|
|
if(!m_cInputGradient.SumArray(inputs.GetGradients()))
|
|
return false;
|
|
if(!m_cOutputGate.CalcHiddenGradient(inputs))
|
|
return false;
|
|
if(!m_cInputGradient.SumArray(inputs.GetGradients()))
|
|
return false;
|
|
if(!m_cNewContent.CalcHiddenGradient(inputs))
|
|
return false;
|
|
if(!inputs.GetGradients().SumArray(m_cInputGradient))
|
|
return false;
|
|
//--- project gradient onto weight matrices of internal layers
|
|
if(!m_cForgetGate.CalcDeltaWeights(inputs, false))
|
|
return false;
|
|
if(!m_cInputGate.CalcDeltaWeights(inputs, false))
|
|
return false;
|
|
if(!m_cOutputGate.CalcDeltaWeights(inputs, false))
|
|
return false;
|
|
if(!m_cNewContent.CalcDeltaWeights(inputs, false))
|
|
return false;
|
|
//--- if gradient for the current state is calculated, pass it to the previous layer
|
|
//--- and write the hidden state gradient to the gradient buffer for a new iteration
|
|
if(!inputs.GetGradients().Split((i == 0 ? prevLayer.GetGradients() : inputs.GetGradients()), m_cGradients, (uint)prevLayer.Total()))
|
|
return false;
|
|
}
|
|
//---
|
|
return true;
|
|
}
|
|
//+------------------------------------------------------------------+
|
|
//| Method for updating weight matrices |
|
|
//+------------------------------------------------------------------+
|
|
bool CNeuronLSTM::UpdateWeights(int batch_size, TYPE learningRate, VECTOR &Beta, VECTOR &Lambda)
|
|
{
|
|
//--- check the state of objects
|
|
if(!m_cForgetGate || !m_cInputGate || !m_cOutputGate || !m_cNewContent || m_iDepth <= 0)
|
|
return false;
|
|
int batch = batch_size * m_iDepth;
|
|
//--- update weight matrices of internal layers
|
|
if(!m_cForgetGate.UpdateWeights(batch, learningRate, Beta, Lambda))
|
|
return false;
|
|
if(!m_cInputGate.UpdateWeights(batch, learningRate, Beta, Lambda))
|
|
return false;
|
|
if(!m_cOutputGate.UpdateWeights(batch, learningRate, Beta, Lambda))
|
|
return false;
|
|
if(!m_cNewContent.UpdateWeights(batch, learningRate, Beta, Lambda))
|
|
return false;
|
|
//---
|
|
return true;
|
|
}
|
|
//+------------------------------------------------------------------+
|
|
//| Method for saving class elements to a file |
|
|
//+------------------------------------------------------------------+
|
|
bool CNeuronLSTM::Save(const int file_handle)
|
|
{
|
|
//--- call of the method of the parent class
|
|
if(!CNeuronBase::Save(file_handle))
|
|
return false;
|
|
//--- save constants
|
|
if(FileWriteInteger(file_handle, m_iDepth) <= 0)
|
|
return false;
|
|
//--- call the relevant method for all internal layers
|
|
if(!m_cForgetGate.Save(file_handle))
|
|
return false;
|
|
if(!m_cInputGate.Save(file_handle))
|
|
return false;
|
|
if(!m_cOutputGate.Save(file_handle))
|
|
return false;
|
|
if(!m_cNewContent.Save(file_handle))
|
|
return false;
|
|
//---
|
|
return true;
|
|
}
|
|
//+------------------------------------------------------------------+
|
|
//| Method for restoring the class from a file |
|
|
//+------------------------------------------------------------------+
|
|
bool CNeuronLSTM::Load(const int file_handle)
|
|
{
|
|
//--- call of the method of the parent class
|
|
if(!CNeuronBase::Load(file_handle))
|
|
return false;
|
|
//--- read constants
|
|
m_iDepth = FileReadInteger(file_handle);
|
|
//--- call the relevant method for all internal layers
|
|
if(FileReadInteger(file_handle) != defNeuronBase || !m_cForgetGate.Load(file_handle))
|
|
return false;
|
|
if(FileReadInteger(file_handle) != defNeuronBase || !m_cInputGate.Load(file_handle))
|
|
return false;
|
|
if(FileReadInteger(file_handle) != defNeuronBase || !m_cOutputGate.Load(file_handle))
|
|
return false;
|
|
if(FileReadInteger(file_handle) != defNeuronBase || !m_cNewContent.Load(file_handle))
|
|
return false;
|
|
//--- initialize Memory
|
|
if(m_cMemorys.Total() > 0)
|
|
m_cMemorys.Clear();
|
|
CBufferType *buffer = CreateBuffer(m_cMemorys);
|
|
if(!buffer)
|
|
return false;
|
|
if(!m_cMemorys.Add(buffer))
|
|
return false;
|
|
//--- initialize HiddenStates
|
|
if(m_cHiddenStates.Total() > 0)
|
|
m_cHiddenStates.Clear();
|
|
buffer = CreateBuffer(m_cHiddenStates);
|
|
if(!buffer)
|
|
return false;
|
|
if(!m_cHiddenStates.Add(buffer))
|
|
return false;
|
|
//--- clear the remaining stacks
|
|
if(!m_cInputs)
|
|
m_cInputs.Clear();
|
|
if(!m_cForgetGateOuts)
|
|
m_cForgetGateOuts.Clear();
|
|
if(!m_cInputGateOuts)
|
|
m_cInputGateOuts.Clear();
|
|
if(!m_cNewContentOuts)
|
|
m_cNewContentOuts.Clear();
|
|
if(!m_cOutputGateOuts)
|
|
m_cOutputGateOuts.Clear();
|
|
//---
|
|
return true;
|
|
}
|
|
//+------------------------------------------------------------------+
|