NeuroBook/Include/realization/neuronbase.mqh
super.admin 951d788cd9 convert
2025-05-30 16:12:30 +02:00

1045 lines
83 KiB
MQL5

//+------------------------------------------------------------------+
//| NeuronBase.mqh |
//| Copyright 2021, MetaQuotes Ltd. |
//| https://www.mql5.com |
//+------------------------------------------------------------------+
#property copyright "Copyright 2021, MetaQuotes Ltd."
#property link "https://www.mql5.com"
//+------------------------------------------------------------------+
//| Connect libraries |
//+------------------------------------------------------------------+
#include <Math\Stat\Normal.mqh>
#include "buffer.mqh"
#include "layerdescription.mqh"
#include "activation.mqh"
//+------------------------------------------------------------------+
//| Class CNeuronBase |
//| Purpose: Base class for a fully connected neural layer |
//+------------------------------------------------------------------+
class CNeuronBase : public CObject
{
protected:
bool m_bTrain;
CMyOpenCL* m_cOpenCL;
CActivation* m_cActivation;
ENUM_OPTIMIZATION m_eOptimization;
CBufferType* m_cOutputs;
CBufferType* m_cWeights;
CBufferType* m_cDeltaWeights;
CBufferType* m_cGradients;
CBufferType* m_cMomenum[2];
//---
virtual bool SGDUpdate(int batch_size, TYPE learningRate,
VECTOR &Lambda);
virtual bool MomentumUpdate(int batch_size, TYPE learningRate,
VECTOR &Beta, VECTOR &Lambda);
virtual bool AdaGradUpdate(int batch_size, TYPE learningRate,
VECTOR &Lambda);
virtual bool RMSPropUpdate(int batch_size, TYPE learningRate,
VECTOR &Beta, VECTOR &Lambda);
virtual bool AdaDeltaUpdate(int batch_size,
VECTOR &Beta, VECTOR &Lambda);
virtual bool AdamUpdate(int batch_size, TYPE learningRate,
VECTOR &Beta, VECTOR &Lambda);
virtual bool SetActivation(ENUM_ACTIVATION_FUNCTION function, VECTOR &params);
public:
CNeuronBase(void);
~CNeuronBase(void);
//---
virtual bool Init(const CLayerDescription *description);
virtual bool SetOpenCL(CMyOpenCL *opencl);
virtual bool FeedForward(CNeuronBase *prevLayer);
virtual bool CalcOutputGradient(CBufferType *target, ENUM_LOSS_FUNCTION loss);
virtual bool CalcHiddenGradient(CNeuronBase *prevLayer);
virtual bool CalcDeltaWeights(CNeuronBase *prevLayer, bool read);
virtual bool UpdateWeights(int batch_size, TYPE learningRate, VECTOR &Beta, VECTOR &Lambda);
//---
virtual void TrainMode(bool flag) { m_bTrain = flag; }
virtual bool TrainMode(void) const { return m_bTrain; }
//---
virtual CBufferType *GetOutputs(void) const { return(m_cOutputs); }
virtual CBufferType *GetGradients(void) const { return(m_cGradients); }
virtual CBufferType *GetWeights(void) const { return(m_cWeights); }
virtual CBufferType *GetDeltaWeights(void) const { return(m_cDeltaWeights);}
virtual bool SetOutputs(CBufferType* buffer, bool delete_prevoius = true);
//--- methods for working with files
virtual bool Save(const int file_handle);
virtual bool Load(const int file_handle);
//--- method of identifying the object
virtual int Type(void) const { return(defNeuronBase); }
virtual ulong Rows(void) const { return(m_cOutputs.Rows()); }
virtual ulong Cols(void) const { return(m_cOutputs.Cols()); }
virtual ulong Total(void) const { return(m_cOutputs.Total()); }
};
//+------------------------------------------------------------------+
//| Class constructor |
//+------------------------------------------------------------------+
CNeuronBase::CNeuronBase(void) : m_eOptimization(Adam)
{
m_cOpenCL = NULL;
m_cActivation = new CActivationSwish();
m_cOutputs = new CBufferType();
m_cWeights = new CBufferType();
m_cDeltaWeights = new CBufferType();
m_cGradients = new CBufferType();
m_cMomenum[0] = new CBufferType();
m_cMomenum[1] = new CBufferType();
}
//+------------------------------------------------------------------+
//| Class destructor |
//+------------------------------------------------------------------+
CNeuronBase::~CNeuronBase(void)
{
if(m_cActivation)
delete m_cActivation;
if(m_cOutputs)
delete m_cOutputs;
if(m_cWeights)
delete m_cWeights;
if(m_cDeltaWeights)
delete m_cDeltaWeights;
if(m_cGradients)
delete m_cGradients;
if(m_cMomenum[0])
delete m_cMomenum[0];
if(m_cMomenum[1])
delete m_cMomenum[1];
}
//+------------------------------------------------------------------+
//| Method for passing a pointer to the context handling object |
//| OpenCL |
//+------------------------------------------------------------------+
bool CNeuronBase::SetOpenCL(CMyOpenCL *opencl)
{
if(!opencl)
{
if(m_cOutputs)
m_cOutputs.BufferFree();
if(m_cGradients)
m_cGradients.BufferFree();
if(m_cWeights)
m_cWeights.BufferFree();
if(m_cDeltaWeights)
m_cDeltaWeights.BufferFree();
for(int i = 0; i < 2; i++)
{
if(m_cMomenum[i])
m_cMomenum[i].BufferFree();
}
if(m_cActivation)
m_cActivation.SetOpenCL(m_cOpenCL, Rows(), Cols());
m_cOpenCL = opencl;
return true;
}
if(m_cOpenCL)
delete m_cOpenCL;
m_cOpenCL = opencl;
if(m_cOutputs)
m_cOutputs.BufferCreate(opencl);
if(m_cGradients)
m_cGradients.BufferCreate(opencl);
if(m_cWeights)
m_cWeights.BufferCreate(opencl);
if(m_cDeltaWeights)
m_cDeltaWeights.BufferCreate(opencl);
for(int i = 0; i < 2; i++)
{
if(m_cMomenum[i])
m_cMomenum[i].BufferCreate(opencl);
}
if(m_cActivation)
m_cActivation.SetOpenCL(m_cOpenCL, Rows(), Cols());
//---
return(!!m_cOpenCL);
}
//+------------------------------------------------------------------+
//| Class initialization method |
//+------------------------------------------------------------------+
bool CNeuronBase::Init(const CLayerDescription *desc)
{
//--- source data control block
if(!desc || desc.type != Type() || desc.count <= 0)
return false;
//--- create a results buffer
if(!m_cOutputs)
if(!(m_cOutputs = new CBufferType()))
return false;
if(!m_cOutputs.BufferInit(1, desc.count, 0))
return false;
//--- create a buffer of error gradients
if(!m_cGradients)
if(!(m_cGradients = new CBufferType()))
return false;
if(!m_cGradients.BufferInit(1, desc.count, 0))
return false;
//--- delete unused objects for the source data layer
if(desc.window <= 0)
{
if(m_cActivation)
delete m_cActivation;
if(m_cWeights)
delete m_cWeights;
if(m_cDeltaWeights)
delete m_cDeltaWeights;
if(m_cMomenum[0])
delete m_cMomenum[0];
if(m_cMomenum[1])
delete m_cMomenum[1];
if(m_cOpenCL)
if(!m_cOutputs.BufferCreate(m_cOpenCL))
return false;
m_eOptimization = desc.optimization;
return true;
}
//--- initialize the activation function object
VECTOR ar_temp = desc.activation_params;
if(!SetActivation(desc.activation, ar_temp))
return false;
//--- initialize the weight matrix object
if(!m_cWeights)
if(!(m_cWeights = new CBufferType()))
return false;
if(!m_cWeights.BufferInit(desc.count, desc.window + 1, 0))
return false;
double weights[];
TYPE sigma = (TYPE)(desc.activation == AF_LRELU ?
2.0 / (MathPow(1 + desc.activation_params[0], 2) * desc.window) :
1.0 / desc.window);
if(!MathRandomNormal(0, MathSqrt(sigma), (uint)m_cWeights.Total(), weights))
return false;
for(uint i = 0; i < m_cWeights.Total(); i++)
if(!m_cWeights.m_mMatrix.Flat(i, (TYPE)weights[i]))
return false;
//--- initialize the object of gradient accumulation at the weight matrix level
if(!m_cDeltaWeights)
if(!(m_cDeltaWeights = new CBufferType()))
return false;
if(!m_cDeltaWeights.BufferInit(desc.count, desc.window + 1, 0))
return false;
//--- initialize the momentum objects
switch(desc.optimization)
{
case None:
case SGD:
for(int i = 0; i < 2; i++)
if(m_cMomenum[i])
delete m_cMomenum[i];
break;
case MOMENTUM:
case AdaGrad:
case RMSProp:
if(!m_cMomenum[0])
if(!(m_cMomenum[0] = new CBufferType()))
return false;
if(!m_cMomenum[0].BufferInit(desc.count, desc.window + 1, 0))
return false;
if(m_cMomenum[1])
delete m_cMomenum[1];
break;
case AdaDelta:
case Adam:
for(int i = 0; i < 2; i++)
{
if(!m_cMomenum[i])
if(!(m_cMomenum[i] = new CBufferType()))
return(false);
if(!m_cMomenum[i].BufferInit(desc.count, desc.window + 1, 0))
return false;
}
break;
default:
return false;
break;
}
//--- save the parameter optimization object
m_eOptimization = desc.optimization;
return true;
}
//+------------------------------------------------------------------+
//| Feed-forward method |
//+------------------------------------------------------------------+
bool CNeuronBase::FeedForward(CNeuronBase * prevLayer)
{
//--- control block
if(!prevLayer || !m_cOutputs || !m_cWeights || !prevLayer.GetOutputs() || !m_cActivation)
return false;
CBufferType *input_data = prevLayer.GetOutputs();
//--- branching of the algorithm depending on the device used for performing operations
if(!m_cOpenCL)
{
if(m_cWeights.Cols() != (input_data.Total() + 1))
return false;
//---
MATRIX m = input_data.m_mMatrix;
if(!m.Reshape(1, input_data.Total() + 1))
return false;
m[0, m.Cols() - 1] = 1;
m_cOutputs.m_mMatrix = m.MatMul(m_cWeights.m_mMatrix.Transpose());
}
else // OpenCL block
{
//--- check data buffers
if(input_data.GetIndex() < 0)
return false;
if(m_cWeights.GetIndex() < 0)
return false;
if(m_cOutputs.GetIndex() < 0)
return false;
//--- pass arguments to the kernel
if(!m_cOpenCL.SetArgumentBuffer(def_k_PerceptronFeedForward, def_pff_inputs, input_data.GetIndex()))
return false;
if(!m_cOpenCL.SetArgumentBuffer(def_k_PerceptronFeedForward, def_pff_weights, m_cWeights.GetIndex()))
return false;
if(!m_cOpenCL.SetArgumentBuffer(def_k_PerceptronFeedForward, def_pff_outputs, m_cOutputs.GetIndex()))
return false;
if(!m_cOpenCL.SetArgument(def_k_PerceptronFeedForward, def_pff_inputs_total, (int)input_data.Total()))
return false;
//--- place kernel to the execution queue
uint off_set[] = {0};
uint NDRange[] = {(uint)m_cOutputs.Total()};
if(!m_cOpenCL.Execute(def_k_PerceptronFeedForward, 1, off_set, NDRange))
return false;
}
//---
return m_cActivation.Activation(m_cOutputs);
}
//+------------------------------------------------------------------+
//| Method for calculating error gradient for the results layer |
//+------------------------------------------------------------------+
bool CNeuronBase::CalcOutputGradient(CBufferType * target, ENUM_LOSS_FUNCTION loss)
{
//--- control block
if(!target || !m_cOutputs || !m_cGradients ||
target.Total() < m_cOutputs.Total() ||
m_cGradients.Total() < m_cOutputs.Total())
return false;
//--- branching of the algorithm depending on the device used for performing operations
if(!m_cOpenCL)
{
switch(loss)
{
case LOSS_MAE:
m_cGradients.m_mMatrix = target.m_mMatrix - m_cOutputs.m_mMatrix;
break;
case LOSS_MSE:
m_cGradients.m_mMatrix = (target.m_mMatrix - m_cOutputs.m_mMatrix) * 2;
break;
case LOSS_CCE:
m_cGradients.m_mMatrix = target.m_mMatrix / (m_cOutputs.m_mMatrix + (TYPE)FLT_MIN) *
log(m_cOutputs.m_mMatrix) * (-1);
break;
case LOSS_BCE:
m_cGradients.m_mMatrix = (target.m_mMatrix - m_cOutputs.m_mMatrix) /
(MathPow(m_cOutputs.m_mMatrix, 2) - m_cOutputs.m_mMatrix + (TYPE)FLT_MIN);
break;
default:
m_cGradients.m_mMatrix = target.m_mMatrix - m_cOutputs.m_mMatrix;
break;
}
}
else // OpenCL block
{
//--- check data buffers
if(target.GetIndex() < 0)
return false;
if(m_cOutputs.GetIndex() < 0)
return false;
if(m_cGradients.GetIndex() < 0)
return false;
//--- pass arguments to the kernel
if(!m_cOpenCL.SetArgumentBuffer(def_k_CalcOutputGradient, def_outgr_target, target.GetIndex()))
return false;
if(!m_cOpenCL.SetArgumentBuffer(def_k_CalcOutputGradient, def_outgr_outputs, m_cOutputs.GetIndex()))
return false;
if(!m_cOpenCL.SetArgumentBuffer(def_k_CalcOutputGradient, def_outgr_gradients, m_cGradients.GetIndex()))
return false;
if(!m_cOpenCL.SetArgument(def_k_CalcOutputGradient, def_outgr_loss_function, (int)loss))
return false;
//--- place kernel to the execution queue
uint NDRange[] = { (uint)m_cOutputs.Total() };
uint off_set[] = {0};
if(!m_cOpenCL.Execute(def_k_CalcOutputGradient, 1, off_set, NDRange))
return false;
}
//---
return true;
}
//+------------------------------------------------------------------+
//| Method for propagating gradient through hidden layer |
//+------------------------------------------------------------------+
bool CNeuronBase::CalcHiddenGradient(CNeuronBase *prevLayer)
{
//--- adjust the input gradient to the derivative of the activation function
if(!m_cActivation.Derivative(m_cGradients))
return false;
//--- check buffers of the previous layer
if(!prevLayer)
return false;
CBufferType *input_data = prevLayer.GetOutputs();
CBufferType *input_gradient = prevLayer.GetGradients();
if(!input_data || !input_gradient || input_data.Total() != input_gradient.Total())
return false;
//--- check size correspondence between the source data buffer and the weight matrix
if(!m_cWeights || m_cWeights.Cols() != (input_data.Total() + 1))
return false;
//--- branching of the algorithm depending on the device used for performing operations
if(!m_cOpenCL)
{
MATRIX grad = m_cGradients.m_mMatrix.MatMul(m_cWeights.m_mMatrix);
grad.Reshape(input_data.Rows(), input_data.Cols());
input_gradient.m_mMatrix = grad;
}
else // OpenCL block
{
//--- check data buffers
if(m_cWeights.GetIndex() < 0)
return false;
if(input_gradient.GetIndex() < 0)
return false;
if(m_cGradients.GetIndex() < 0)
return false;
//--- pass arguments to the kernel
if(!m_cOpenCL.SetArgumentBuffer(def_k_CalcHiddenGradient, def_hidgr_gradient_inputs, input_gradient.GetIndex()))
return false;
if(!m_cOpenCL.SetArgumentBuffer(def_k_CalcHiddenGradient, def_hidgr_weights, m_cWeights.GetIndex()))
return false;
if(!m_cOpenCL.SetArgumentBuffer(def_k_CalcHiddenGradient, def_hidgr_gradients, m_cGradients.GetIndex()))
return false;
if(!m_cOpenCL.SetArgument(def_k_CalcHiddenGradient, def_hidgr_outputs_total, (int)m_cGradients.Total()))
return false;
//--- place kernel to the execution queue
uint NDRange[] = {(uint)input_data.Total()};
uint off_set[] = {0};
if(!m_cOpenCL.Execute(def_k_CalcHiddenGradient, 1, off_set, NDRange))
return false;
}
//---
return true;
}
//+------------------------------------------------------------------+
//| Method for propagating the error gradient to the weight matrix |
//+------------------------------------------------------------------+
bool CNeuronBase::CalcDeltaWeights(CNeuronBase * prevLayer, bool read)
{
//--- control block
if(!prevLayer || !m_cDeltaWeights || !m_cGradients)
return false;
CBufferType *Inputs = prevLayer.GetOutputs();
if(!Inputs)
return false;
//--- branching of the algorithm depending on the device used for performing operations
if(!m_cOpenCL)
{
MATRIX m = Inputs.m_mMatrix;
m.Resize(1, Inputs.Total() + 1);
m[0, Inputs.Total()] = 1;
m = m_cGradients.m_mMatrix.Transpose().MatMul(m);
m_cDeltaWeights.m_mMatrix += m;
}
else // OpenCL block
{
//--- check data buffers
if(m_cGradients.GetIndex() < 0)
return false;
if(m_cDeltaWeights.GetIndex() < 0)
return false;
if(Inputs.GetIndex() < 0)
return false;
//--- pass arguments to the kernel
if(!m_cOpenCL.SetArgumentBuffer(def_k_CalcDeltaWeights, def_delt_delta_weights, m_cDeltaWeights.GetIndex()))
return false;
if(!m_cOpenCL.SetArgumentBuffer(def_k_CalcDeltaWeights, def_delt_inputs, Inputs.GetIndex()))
return false;
if(!m_cOpenCL.SetArgumentBuffer(def_k_CalcDeltaWeights, def_delt_gradients, m_cGradients.GetIndex()))
return false;
//--- place kernel to the execution queue
uint NDRange[] = {(uint)m_cGradients.Total(), (uint)Inputs.Total()};
uint off_set[] = {0, 0};
if(!m_cOpenCL.Execute(def_k_CalcDeltaWeights, 2, off_set, NDRange))
return false;
if(read && !m_cDeltaWeights.BufferRead())
return false;
}
//---
return true;
}
//+------------------------------------------------------------------+
//| Weight matrix update method |
//+------------------------------------------------------------------+
bool CNeuronBase::UpdateWeights(int batch_size, TYPE learningRate, VECTOR &Beta, VECTOR &Lambda)
{
//--- control block
if(!m_cDeltaWeights || !m_cWeights || m_cWeights.Total() < m_cDeltaWeights.Total() || batch_size <= 0)
return false;
//--- gradient branching depending on the activation function used
bool result = false;
switch(m_eOptimization)
{
case None:
result = true;
break;
case SGD:
result = SGDUpdate(batch_size, learningRate, Lambda);
break;
case MOMENTUM:
result = MomentumUpdate(batch_size, learningRate, Beta, Lambda);
break;
case AdaGrad:
result = AdaGradUpdate(batch_size, learningRate, Lambda);
break;
case RMSProp:
result = RMSPropUpdate(batch_size, learningRate, Beta, Lambda);
break;
case AdaDelta:
result = AdaDeltaUpdate(batch_size, Beta, Lambda);
break;
case Adam:
result = AdamUpdate(batch_size, learningRate, Beta, Lambda);
break;
}
//---
return result;
}
//+------------------------------------------------------------------+
//| Update the weight matrix using stochastic gradient descent |
//+------------------------------------------------------------------+
bool CNeuronBase::SGDUpdate(int batch_size, TYPE learningRate, VECTOR &Lambda)
{
//--- branching of the algorithm depending on the device used for performing operations
if(!m_cOpenCL)
{
TYPE lr = learningRate / ((TYPE)batch_size);
m_cWeights.m_mMatrix -= m_cWeights.m_mMatrix * Lambda[1] + Lambda[0];
m_cWeights.m_mMatrix += m_cDeltaWeights.m_mMatrix * lr;
m_cDeltaWeights.m_mMatrix.Fill(0);
}
else
{
//--- check data buffers
if(m_cWeights.GetIndex() < 0)
return false;
if(m_cDeltaWeights.GetIndex() < 0)
return false;
//--- pass arguments to the kernel
if(!m_cOpenCL.SetArgumentBuffer(def_k_SGDUpdate, def_sgd_delta_weights, m_cDeltaWeights.GetIndex()))
return false;
if(!m_cOpenCL.SetArgumentBuffer(def_k_SGDUpdate, def_sgd_weights, m_cWeights.GetIndex()))
return false;
if(!m_cOpenCL.SetArgument(def_k_SGDUpdate, def_sgd_total, m_cWeights.Total()))
return false;
if(!m_cOpenCL.SetArgument(def_k_SGDUpdate, def_sgd_batch_size, batch_size))
return false;
if(!m_cOpenCL.SetArgument(def_k_SGDUpdate, def_sgd_learningRate, learningRate))
return false;
if(!m_cOpenCL.SetArgument(def_k_SGDUpdate, def_sgd_Lambda1, Lambda[0]))
return false;
if(!m_cOpenCL.SetArgument(def_k_SGDUpdate, def_sgd_Lambda2, Lambda[1]))
return false;
//--- place kernel to the execution queue
int NDRange[] = { (int)((m_cWeights.Total() + 3) / 4) };
int off_set[] = {0};
if(!m_cOpenCL.Execute(def_k_SGDUpdate, 1, off_set, NDRange))
return false;
}
return true;
}
//+------------------------------------------------------------------+
//| Update the weight matrix using the momentum method |
//+------------------------------------------------------------------+
bool CNeuronBase::MomentumUpdate(int batch_size, TYPE learningRate, VECTOR &Beta, VECTOR &Lambda)
{
if(Beta[0] == 0)
return SGDUpdate(batch_size, learningRate, Lambda);
//--- control block
if(!m_cMomenum[0])
return false;
if(m_cMomenum[0].Total() < m_cWeights.Total())
return false;
//--- branching of the algorithm depending on the device used for performing operations
if(!m_cOpenCL)
{
TYPE lr = learningRate / ((TYPE)batch_size);
m_cWeights.m_mMatrix -= m_cWeights.m_mMatrix * Lambda[1] + Lambda[0];
m_cMomenum[0].m_mMatrix = m_cDeltaWeights.m_mMatrix * lr + m_cMomenum[0].m_mMatrix * Beta[0] ;
m_cWeights.m_mMatrix += m_cMomenum[0].m_mMatrix;
m_cDeltaWeights.m_mMatrix.Fill(0);
}
else // OpenCL block
{
//--- check data buffers
if(m_cWeights.GetIndex() < 0)
return false;
if(m_cDeltaWeights.GetIndex() < 0)
return false;
if(m_cMomenum[0].GetIndex() < 0)
return false;
//--- pass arguments to the kernel
if(!m_cOpenCL.SetArgumentBuffer(def_k_MomentumUpdate, def_moment_delta_weights, m_cDeltaWeights.GetIndex()))
return false;
if(!m_cOpenCL.SetArgumentBuffer(def_k_MomentumUpdate, def_moment_weights, m_cWeights.GetIndex()))
return false;
if(!m_cOpenCL.SetArgumentBuffer(def_k_MomentumUpdate, def_moment_momentum, m_cMomenum[0].GetIndex()))
return false;
if(!m_cOpenCL.SetArgument(def_k_MomentumUpdate, def_moment_total, m_cWeights.Total()))
return false;
if(!m_cOpenCL.SetArgument(def_k_MomentumUpdate, def_moment_batch_size, batch_size))
return false;
if(!m_cOpenCL.SetArgument(def_k_MomentumUpdate, def_moment_learningRate, learningRate))
return false;
if(!m_cOpenCL.SetArgument(def_k_MomentumUpdate, def_moment_Lambda1, Lambda[0]))
return false;
if(!m_cOpenCL.SetArgument(def_k_MomentumUpdate, def_moment_Lambda2, Lambda[1]))
return false;
if(!m_cOpenCL.SetArgument(def_k_MomentumUpdate, def_moment_beta, Beta[0]))
return false;
//--- place kernel to the execution queue
int NDRange[] = { (int)((m_cWeights.Total() + 3) / 4) };
int off_set[] = {0};
if(!m_cOpenCL.Execute(def_k_MomentumUpdate, 1, off_set, NDRange))
return false;
}
return true;
}
//+------------------------------------------------------------------+
//| Updating the weight matrix using the AdaGrad method |
//+------------------------------------------------------------------+
bool CNeuronBase::AdaGradUpdate(int batch_size, TYPE learningRate, VECTOR &Lambda)
{
//--- control block
if(!m_cMomenum[0])
return false;
if(m_cMomenum[0].Total() < m_cWeights.Total())
return false;
//--- branching of the algorithm depending on the device used for performing operations
if(!m_cOpenCL)
{
m_cWeights.m_mMatrix -= m_cWeights.m_mMatrix * Lambda[1] + Lambda[0];
MATRIX delta = m_cDeltaWeights.m_mMatrix / ((TYPE)batch_size);
MATRIX G = m_cMomenum[0].m_mMatrix = m_cMomenum[0].m_mMatrix + delta.Power(2);
G = MathPow(MathSqrt(G) + 1e-32, -1);
G = G * learningRate;
m_cWeights.m_mMatrix += G * delta;
m_cDeltaWeights.m_mMatrix.Fill(0);
}
else // OpenCL block
{
//--- check data buffers
if(m_cWeights.GetIndex() < 0)
return false;
if(m_cDeltaWeights.GetIndex() < 0)
return false;
if(m_cMomenum[0].GetIndex() < 0)
return false;
//--- pass arguments to the kernel
if(!m_cOpenCL.SetArgumentBuffer(def_k_AdaGradUpdate, def_adagrad_delta_weights, m_cDeltaWeights.GetIndex()))
return false;
if(!m_cOpenCL.SetArgumentBuffer(def_k_AdaGradUpdate, def_adagrad_weights, m_cWeights.GetIndex()))
return false;
if(!m_cOpenCL.SetArgumentBuffer(def_k_AdaGradUpdate, def_adagrad_momentum, m_cMomenum[0].GetIndex()))
return false;
if(!m_cOpenCL.SetArgument(def_k_AdaGradUpdate, def_adagrad_total, m_cWeights.Total()))
return false;
if(!m_cOpenCL.SetArgument(def_k_AdaGradUpdate, def_adagrad_batch_size, batch_size))
return false;
if(!m_cOpenCL.SetArgument(def_k_AdaGradUpdate, def_adagrad_learningRate, learningRate))
return false;
if(!m_cOpenCL.SetArgument(def_k_AdaGradUpdate, def_adagrad_Lambda1, Lambda[0]))
return false;
if(!m_cOpenCL.SetArgument(def_k_AdaGradUpdate, def_adagrad_Lambda2, Lambda[1]))
return false;
//--- place kernel to the execution queue
int NDRange[] = { (int)((m_cWeights.Total() + 3) / 4) };
int off_set[] = {0};
if(!m_cOpenCL.Execute(def_k_AdaGradUpdate, 1, off_set, NDRange))
return false;
}
return true;
}
//+------------------------------------------------------------------+
//| Updating the weight matrix using the RMSProp method |
//+------------------------------------------------------------------+
bool CNeuronBase::RMSPropUpdate(int batch_size, TYPE learningRate, VECTOR &Beta, VECTOR &Lambda)
{
//--- control block
if(!m_cMomenum[0])
return false;
if(m_cMomenum[0].Total() < m_cWeights.Total())
return false;
//--- branching of the algorithm depending on the device used for performing operations
if(!m_cOpenCL)
{
TYPE lr = learningRate;
m_cWeights.m_mMatrix -= m_cWeights.m_mMatrix * Lambda[1] + Lambda[0];
MATRIX delta = m_cDeltaWeights.m_mMatrix / ((TYPE)batch_size);
MATRIX G = m_cMomenum[0].m_mMatrix = m_cMomenum[0].m_mMatrix * Beta[0] + delta.Power(2) * (1 - Beta[0]);
G = MathPow(MathSqrt(G) + 1e-32, -1);
G = G * learningRate;
m_cWeights.m_mMatrix += G * delta;
m_cDeltaWeights.m_mMatrix.Fill(0);
}
else // OpenCL block
{
//--- check data buffers
if(m_cWeights.GetIndex() < 0)
return false;
if(m_cDeltaWeights.GetIndex() < 0)
return false;
if(m_cMomenum[0].GetIndex() < 0)
return false;
//--- pass arguments to the kernel
if(!m_cOpenCL.SetArgumentBuffer(def_k_RMSPropUpdate, def_rms_delta_weights, m_cDeltaWeights.GetIndex()))
return false;
if(!m_cOpenCL.SetArgumentBuffer(def_k_RMSPropUpdate, def_rms_weights, m_cWeights.GetIndex()))
return false;
if(!m_cOpenCL.SetArgumentBuffer(def_k_RMSPropUpdate, def_rms_momentum, m_cMomenum[0].GetIndex()))
return false;
if(!m_cOpenCL.SetArgument(def_k_RMSPropUpdate, def_rms_total, m_cWeights.Total()))
return false;
if(!m_cOpenCL.SetArgument(def_k_RMSPropUpdate, def_rms_batch_size, batch_size))
return false;
if(!m_cOpenCL.SetArgument(def_k_RMSPropUpdate, def_rms_learningRate, learningRate))
return false;
if(!m_cOpenCL.SetArgument(def_k_RMSPropUpdate, def_rms_Lambda1, Lambda[0]))
return false;
if(!m_cOpenCL.SetArgument(def_k_RMSPropUpdate, def_rms_Lambda2, Lambda[1]))
return false;
if(!m_cOpenCL.SetArgument(def_k_RMSPropUpdate, def_rms_beta, Beta[0]))
return false;
//--- place kernel to the execution queue
int NDRange[] = { (int)((m_cWeights.Total() + 3) / 4) };
int off_set[] = {0};
if(!m_cOpenCL.Execute(def_k_RMSPropUpdate, 1, off_set, NDRange))
return false;
}
//---
return true;
}
//+------------------------------------------------------------------+
//| Updating the weight matrix using the AdaDelta method |
//+------------------------------------------------------------------+
bool CNeuronBase::AdaDeltaUpdate(int batch_size, VECTOR &Beta, VECTOR &Lambda)
{
//--- control block
for(int i = 0; i < 2; i++)
{
if(!m_cMomenum[i])
return false;
if(m_cMomenum[i].Total() < m_cWeights.Total())
return false;
}
//--- branching of the algorithm depending on the device used for performing operations
if(!m_cOpenCL)
{
MATRIX delta = m_cDeltaWeights.m_mMatrix / ((TYPE)batch_size);
MATRIX W = m_cMomenum[0].m_mMatrix = m_cMomenum[0].m_mMatrix * Beta[0] + m_cWeights.m_mMatrix.Power(2) * (1 - Beta[0]);
m_cMomenum[1].m_mMatrix = m_cMomenum[1].m_mMatrix * Beta[1] + delta.Power(2) * (1 - Beta[1]);
m_cWeights.m_mMatrix -= m_cWeights.m_mMatrix * Lambda[1] + Lambda[0];
W = MathSqrt(W) / (MathSqrt(m_cMomenum[1].m_mMatrix) + 1e-32);
m_cWeights.m_mMatrix += W * delta;
m_cDeltaWeights.m_mMatrix.Fill(0);
}
else // OpenCL block
{
//--- create data buffers
if(m_cWeights.GetIndex() < 0)
return false;
if(m_cDeltaWeights.GetIndex() < 0)
return false;
if(m_cMomenum[0].GetIndex() < 0)
return false;
if(m_cMomenum[1].GetIndex() < 0)
return false;
//--- pass arguments to the kernel
if(!m_cOpenCL.SetArgumentBuffer(def_k_AdaDeltaUpdate, def_adadelt_delta_weights, m_cDeltaWeights.GetIndex()))
return false;
if(!m_cOpenCL.SetArgumentBuffer(def_k_AdaDeltaUpdate, def_adadelt_weights, m_cWeights.GetIndex()))
return false;
if(!m_cOpenCL.SetArgumentBuffer(def_k_AdaDeltaUpdate, def_adadelt_momentumW, m_cMomenum[0].GetIndex()))
return false;
if(!m_cOpenCL.SetArgumentBuffer(def_k_AdaDeltaUpdate, def_adadelt_momentumG, m_cMomenum[1].GetIndex()))
return false;
if(!m_cOpenCL.SetArgument(def_k_AdaDeltaUpdate, def_adadelt_total, m_cWeights.Total()))
return false;
if(!m_cOpenCL.SetArgument(def_k_AdaDeltaUpdate, def_adadelt_batch_size, batch_size))
return false;
if(!m_cOpenCL.SetArgument(def_k_AdaDeltaUpdate, def_adadelt_Lambda1, Lambda[0]))
return false;
if(!m_cOpenCL.SetArgument(def_k_AdaDeltaUpdate, def_adadelt_Lambda2, Lambda[1]))
return false;
if(!m_cOpenCL.SetArgument(def_k_AdaDeltaUpdate, def_adadelt_beta1, Beta[0]))
return false;
if(!m_cOpenCL.SetArgument(def_k_AdaDeltaUpdate, def_adadelt_beta2, Beta[1]))
return false;
//--- place kernel to the execution queue
int NDRange[] = { (int)((m_cWeights.Total() + 3) / 4) };
int off_set[] = {0};
if(!m_cOpenCL.Execute(def_k_AdaDeltaUpdate, 1, off_set, NDRange))
return false;
}
//---
return true;
}
//+------------------------------------------------------------------+
//| Updating the weight matrix using the Adam method |
//+------------------------------------------------------------------+
bool CNeuronBase::AdamUpdate(int batch_size, TYPE learningRate, VECTOR &Beta, VECTOR &Lambda)
{
//--- control block
for(int i = 0; i < 2; i++)
{
if(!m_cMomenum[i])
return false;
if(m_cMomenum[i].Total() != m_cWeights.Total())
return false;
}
//--- branching of the algorithm depending on the device used for performing operations
if(!m_cOpenCL)
{
MATRIX delta = m_cDeltaWeights.m_mMatrix / ((TYPE)batch_size);
m_cMomenum[0].m_mMatrix = m_cMomenum[0].m_mMatrix * Beta[0] + delta * (1 - Beta[0]);
m_cMomenum[1].m_mMatrix = m_cMomenum[1].m_mMatrix * Beta[1] + MathPow(delta, 2) * (1 - Beta[1]);
MATRIX M = m_cMomenum[0].m_mMatrix / (1 - Beta[0]);
MATRIX V = m_cMomenum[1].m_mMatrix / (1 - Beta[1]);
m_cWeights.m_mMatrix -= m_cWeights.m_mMatrix * Lambda[1] + Lambda[0];
m_cWeights.m_mMatrix += M * learningRate / MathSqrt(V);
m_cDeltaWeights.m_mMatrix.Fill(0);
}
else // OpenCL block
{
//--- check data buffers
if(m_cWeights.GetIndex() < 0)
return false;
if(m_cDeltaWeights.GetIndex() < 0)
return false;
if(m_cMomenum[0].GetIndex() < 0)
return false;
if(m_cMomenum[1].GetIndex() < 0)
return false;
//--- pass arguments to the kernel
if(!m_cOpenCL.SetArgumentBuffer(def_k_AdamUpdate, def_adam_delta_weights, m_cDeltaWeights.GetIndex()))
return false;
if(!m_cOpenCL.SetArgumentBuffer(def_k_AdamUpdate, def_adam_weights, m_cWeights.GetIndex()))
return false;
if(!m_cOpenCL.SetArgumentBuffer(def_k_AdamUpdate, def_adam_momentumM, m_cMomenum[0].GetIndex()))
return false;
if(!m_cOpenCL.SetArgumentBuffer(def_k_AdamUpdate, def_adam_momentumV, m_cMomenum[1].GetIndex()))
return false;
if(!m_cOpenCL.SetArgument(def_k_AdamUpdate, def_adam_total, (int)m_cWeights.Total()))
return false;
if(!m_cOpenCL.SetArgument(def_k_AdamUpdate, def_adam_batch_size, batch_size))
return false;
if(!m_cOpenCL.SetArgument(def_k_AdamUpdate, def_adam_Lambda1, Lambda[0]))
return false;
if(!m_cOpenCL.SetArgument(def_k_AdamUpdate, def_adam_Lambda2, Lambda[1]))
return false;
if(!m_cOpenCL.SetArgument(def_k_AdamUpdate, def_adam_beta1, Beta[0]))
return false;
if(!m_cOpenCL.SetArgument(def_k_AdamUpdate, def_adam_beta2, Beta[1]))
return false;
if(!m_cOpenCL.SetArgument(def_k_AdamUpdate, def_adam_learningRate, learningRate))
return false;
//--- place kernel to the execution queue
int NDRange[] = { (int)((m_cWeights.Total() + 3) / 4) };
int off_set[] = {0};
if(!m_cOpenCL.Execute(def_k_AdamUpdate, 1, off_set, NDRange))
return false;
}
//---
return true;
}
//+------------------------------------------------------------------+
//| Method for saving class elements to a file |
//+------------------------------------------------------------------+
bool CNeuronBase::Save(const int file_handle)
{
//--- control block
if(file_handle == INVALID_HANDLE)
return false;
//--- write data to the results buffer
if(!m_cOutputs)
return false;
if(FileWriteInteger(file_handle, Type()) <= 0 ||
FileWriteInteger(file_handle, (uint)m_cOutputs.Total()) <= 0)
return false;
//--- check and write the source data layer flag
if(!m_cActivation || !m_cWeights)
{
if(FileWriteInteger(file_handle, 1) <= 0)
return false;
return true;
}
if(FileWriteInteger(file_handle, 0) <= 0)
return false;
int momentums = 0;
switch(m_eOptimization)
{
case SGD:
momentums = 0;
break;
case MOMENTUM:
case AdaGrad:
case RMSProp:
momentums = 1;
break;
case AdaDelta:
case Adam:
momentums = 2;
break;
default:
return false;
break;
}
for(int i = 0; i < momentums; i++)
if(!m_cMomenum[i])
return false;
//--- save the matrix of weighs, moments and activation function
if(FileWriteInteger(file_handle, (int)m_eOptimization) <= 0 ||
FileWriteInteger(file_handle, momentums) <= 0)
return false;
if(!m_cWeights.Save(file_handle) || !m_cActivation.Save(file_handle))
return false;
for(int i = 0; i < momentums; i++)
if(!m_cMomenum[i].Save(file_handle))
return false;
//---
return true;
}
//+------------------------------------------------------------------+
//| Method for restoring class state from data in the file |
//+------------------------------------------------------------------+
bool CNeuronBase::Load(const int file_handle)
{
//--- control block
if(file_handle == INVALID_HANDLE)
return false;
//--- load the results buffer
if(!m_cOutputs)
if(!(m_cOutputs = new CBufferType()))
return false;
int outputs = FileReadInteger(file_handle);
if(!m_cOutputs.BufferInit(1, outputs, 0))
return false;
//--- create an error gradient buffer
if(!m_cGradients)
if(!(m_cGradients = new CBufferType()))
return false;
if(!m_cGradients.BufferInit(1, outputs, 0))
return false;
//--- check the source data layer flag
int input_layer = FileReadInteger(file_handle);
if(input_layer == 1)
{
if(m_cActivation)
delete m_cActivation;
if(m_cWeights)
delete m_cWeights;
if(m_cDeltaWeights)
delete m_cDeltaWeights;
if(m_cMomenum[0])
delete m_cMomenum[0];
if(m_cMomenum[1])
delete m_cMomenum[1];
if(m_cOpenCL)
if(!m_cOutputs.BufferCreate(m_cOpenCL))
return false;
m_eOptimization = None;
return true;
}
//---
m_eOptimization = (ENUM_OPTIMIZATION)FileReadInteger(file_handle);
int momentums = FileReadInteger(file_handle);
//--- create objects before loading data
if(!m_cWeights)
if(!(m_cWeights = new CBufferType()))
return false;
//--- load data from the file
if(!m_cWeights.Load(file_handle))
return false;
//--- activation function
if(FileReadInteger(file_handle) != defActivation)
return false;
ENUM_ACTIVATION_FUNCTION activation = (ENUM_ACTIVATION_FUNCTION)FileReadInteger(file_handle);
if(!SetActivation(activation, VECTOR::Zeros(2)))
return false;
if(!m_cActivation.Load(file_handle))
return false;
//---
for(int i = 0; i < momentums; i++)
{
if(!m_cMomenum[i])
if(!(m_cMomenum[i] = new CBufferType()))
return false;
if(!m_cMomenum[i].Load(file_handle))
return false;
}
//--- initialize the remaining buffers
if(!m_cDeltaWeights)
if(!(m_cDeltaWeights = new CBufferType()))
return false;
if(!m_cDeltaWeights.BufferInit(m_cWeights.m_mMatrix.Rows(), m_cWeights.m_mMatrix.Cols(), 0))
return false;
//--- pass pointer to an OpenCL object to objects
SetOpenCL(m_cOpenCL);
//---
return true;
}
//+------------------------------------------------------------------+
//| |
//+------------------------------------------------------------------+
bool CNeuronBase::SetOutputs(CBufferType * buffer, bool delete_prevoius = true)
{
if(!buffer)
return false;
//---
if(delete_prevoius)
if(!!m_cOutputs)
delete m_cOutputs;
m_cOutputs = buffer;
//---
return true;
}
//+------------------------------------------------------------------+
//| |
//+------------------------------------------------------------------+
bool CNeuronBase::SetActivation(ENUM_ACTIVATION_FUNCTION function, VECTOR &params)
{
if(m_cActivation)
delete m_cActivation;
switch(function)
{
case AF_LINEAR:
if(!(m_cActivation = new CActivationLine()))
return false;
break;
case AF_SIGMOID:
if(!(m_cActivation = new CActivationSigmoid()))
return false;
break;
case AF_LRELU:
if(!(m_cActivation = new CActivationLReLU()))
return false;
break;
case AF_TANH:
if(!(m_cActivation = new CActivationTANH()))
return false;
break;
case AF_SOFTMAX:
if(!(m_cActivation = new CActivationSoftMAX()))
return false;
break;
case AF_SWISH:
if(!(m_cActivation = new CActivationSwish()))
return false;
break;
default:
if(!(m_cActivation = new CActivation()))
return false;
break;
}
if(!m_cActivation.Init(params))
return false;
m_cActivation.SetOpenCL(m_cOpenCL, m_cOutputs.Rows(), m_cOutputs.Cols());
return true;
}
//+------------------------------------------------------------------+