void Assignment::StochasticBackPropagateTaskGPU(unsigned int numEpochs) {
InitCLContext();
InitCLResources();
std::cout << std::endl;
std::cout << "You have selected the back propagation task on the GPU " <<
"using stochastic gradient descent." << std::endl;
std::cout << "The neuronal network will now train " << this->trainingData->numberOfSamples <<
" samples for " << numEpochs << " epochs with a learning rate of " <<
this->learningRate << " using a local group size of " << this->localGroupSize <<
" and stop the time." << std::endl;
CTimer timer;
timer.Start();
for (unsigned int j = 0; j < numEpochs; j++) {
std::cout << "Starting with epoch " << j << std::endl;
double crossEntropy = 0.0f;
zeroCrossEntropyGPU();
for (unsigned int i = 0; i < this->trainingData->numberOfSamples; i++) {
zeroDeltaBuffersGPU();
feedForwardGPU(i, 1);
calculateCrossEntropyGPU(i, 1);
gradientDescentGPU(i, 1);
updateWeightsGPU();
if ( i % 100 == 0) {
crossEntropy += readCrossEntropyGPU();
zeroCrossEntropyGPU();
}
}
crossEntropy += readCrossEntropyGPU();
std::cout << "Done." << std::endl;
std::cout << "Accumulated crossEntropy error for this epoch: " << crossEntropy << std::endl;
}
timer.Stop();
copyWeightBuffersFromDevice();
std::cout << "Done with back propagation (stochastic)." << std::endl;
std::cout << "The task took " << timer.GetElapsedMilliseconds() <<
" milliseconds to complete." << std::endl;
ReleaseClResources();
ReleaseCLContext();
}
void Assignment::batchBackPropagateTaskGPU(unsigned int numEpochs, unsigned int batchSize) {
InitCLContext();
InitCLResources();
std::cout << std::endl;
std::cout << "You have selected the back propagation task on the GPU " <<
"using batch gradient descent with a batch size of " << batchSize << "." << std::endl;
std::cout << "The gpu will use local group size of " << this->localGroupSize <<
" and try to evaluate " << this->parallelBackpropagationSize << " inputs at once." << std::endl;
std::cout << "The neuronal network will now train " << this->trainingData->numberOfSamples <<
" samples for " << numEpochs << " epochs with a learning rate of " << this->learningRate <<
" and stop the time." << std::endl;
CTimer timer;
timer.Start();
int fullBatches = this->trainingData->numberOfSamples / batchSize;
int lastBatchSize = this->trainingData->numberOfSamples % batchSize;
//iterate over epochs
for (unsigned int i = 0; i < numEpochs; i++) {
std::cout << "Starting with epoch " << i << std::endl;
double crossEntropy = 0.0;
//iterate over batches
for (int j = 0; j < fullBatches; j++) {
crossEntropy += calculateBackPropForBatch(j * batchSize, batchSize);
}
//do the last small batch
if (lastBatchSize > 0) {
crossEntropy += calculateBackPropForBatch(fullBatches * batchSize, lastBatchSize);
}
std::cout << "Done." << std::endl;
std::cout << "Accumulated crossEntropy error for this epoch: " << crossEntropy << std::endl;
}
timer.Stop();
copyWeightBuffersFromDevice();
std::cout << "Done with back propagation (batch)." << std::endl;
std::cout << "The task took " << timer.GetElapsedMilliseconds() <<
" milliseconds to complete." << std::endl;
ReleaseClResources();
ReleaseCLContext();
}
void Assignment::feedForwardTaskGPU() {
InitCLContext();
InitCLResources();
std::cout << std::endl;
std::cout << "You have selected the feed forward task on the GPU." << std::endl;
std::cout << "Will now feed forward " << this->parallelBackpropagationSize <<
" samples in parallel using a local group size of " << this->localGroupSize <<
", stop the time and accumulate the crossEntropy error." << std::endl;
CTimer timer;
double crossEntropy = 0.0;
timer.Start();
int fullBatches = this->trainingData->numberOfSamples / this->parallelBackpropagationSize;
int lastBatchSize = this->trainingData->numberOfSamples % this->parallelBackpropagationSize;
for (int i = 0; i < fullBatches; i++) {
zeroCrossEntropyGPU();
feedForwardGPU(i * this->parallelBackpropagationSize, this->parallelBackpropagationSize);
//printFeedForwardResultGPU(this->parallelBackpropagationSize);
calculateCrossEntropyGPU(i * this->parallelBackpropagationSize, this->parallelBackpropagationSize);
crossEntropy += readCrossEntropyGPU();
}
if (lastBatchSize > 0) {
feedForwardGPU(fullBatches * this->parallelBackpropagationSize, lastBatchSize);
calculateCrossEntropyGPU(fullBatches * this->parallelBackpropagationSize, lastBatchSize);
}
crossEntropy += readCrossEntropyGPU();
timer.Stop();
std::cout << "done." << std::endl;
std::cout << "The crossEntropy error is " << crossEntropy << "." << std::endl;
std::cout << "The task took " << timer.GetElapsedMilliseconds() <<
" milliseconds to complete." << std::endl;
ReleaseClResources();
ReleaseCLContext();
}
double CConvolutionSeparableTask::ConvolutionChannelCPU(unsigned int Channel)
{
CTimer timer;
timer.Start();
//horizontal pass
for(int y = 0; y < (int)m_Height; y++)
for(int x = 0; x < (int)m_Width; x++)
{
float value = 0;
//apply horizontal kernel
for(int k = -m_KernelRadius; k <= m_KernelRadius; k++)
{
int sx = x + k;
if(sx >= 0 && sx < (int)m_Width)
value += m_hSourceChannels[Channel][y * m_Pitch + sx] * m_hKernelHorizontal[m_KernelRadius - k];
}
m_hCPUWorkingBuffer[y * m_Pitch + x] = value;
//m_hCPUResultChannels[Channel][y * m_Pitch + x] = value;
}
//vertical pass
for(int x = 0; x < (int)m_Width; x++)
for(int y = 0; y < (int)m_Height; y++)
{
float value = 0;
//apply horizontal kernel
for(int k = -m_KernelRadius; k <= m_KernelRadius; k++)
{
int sy = y + k;
if(sy >= 0 && sy < (int)m_Height)
value += m_hCPUWorkingBuffer[sy * m_Pitch + x] * m_hKernelVertical[m_KernelRadius - k];
}
m_hCPUResultChannels[Channel][y * m_Pitch + x] = value;
}
timer.Stop();
return timer.GetElapsedMilliseconds();
}
