void MultiScaleTransforms::PrepareTransform(double* kernel, double scale)
{
ao::uvector<double> shape;
size_t kernelSize;
MakeShapeFunction(scale, shape, kernelSize);
memset(kernel, 0, sizeof(double) * _width * _height);
FFTConvolver::PrepareSmallKernel(kernel, _width, _height, shape.data(), kernelSize);
}
void MultiScaleTransforms::Transform(const ao::uvector<double*>& images, double* scratch, double scale)
{
ao::uvector<double> shape;
size_t kernelSize;
MakeShapeFunction(scale, shape, kernelSize);
memset(scratch, 0, sizeof(double) * _width * _height);
FFTConvolver::PrepareSmallKernel(scratch, _width, _height, shape.data(), kernelSize);
for(double*const* imageIter = images.begin(); imageIter!=images.end(); ++imageIter)
FFTConvolver::ConvolveSameSize(_fftwManager, *imageIter, scratch, _width, _height);
}
void FastMultiScaleClean<ImageSetType>::executeMajorIterationForScale(double currentScale, double nextScale, bool& reachedStopGain, bool& canCleanFurther)
{
ImageBufferAllocator& allocator = *_dataImageOriginal->Allocator();
// Scale down the image so that the "scale size" is 10 pixels
// 10 pixels is so that the peak finding is still reasonably accurate
_rescaledWidth = size_t(ceil(double(_originalWidth)*(10.0/currentScale))),
_rescaledHeight = size_t(ceil(double(_originalHeight)*(10.0/currentScale)));
double rescaleFactor = double(_rescaledWidth) / _originalWidth;
if(_rescaledWidth > _originalWidth || _rescaledHeight > _originalHeight)
{
_rescaledWidth = _originalWidth;
_rescaledHeight = _originalHeight;
rescaleFactor = 1.0;
}
ImageSetType
largeScaleImage(_rescaledWidth*_rescaledHeight, *_dataImageOriginal),
nextScaleImage(_rescaledWidth*_rescaledHeight, *_dataImageOriginal),
currentScaleModel(_rescaledWidth*_rescaledHeight, *_modelImage);
std::vector<double*> scaledPsfs(_originalPsfs->size());
_dataImageLargeScale = &largeScaleImage;
_dataImageNextScale = &nextScaleImage;
_scaledPsfs = &scaledPsfs;
canCleanFurther = true;
size_t cpuCount = this->_threadCount;
double thresholdBias = pow(this->_multiscaleThresholdBias, log2(currentScale));
std::cout << "Threshold bias: " << thresholdBias << '\n';
double oldSubtractionGain = this->_subtractionGain;
this->_subtractionGain *= sqrt(thresholdBias);
// Fill the large and next scale images with the rescaled images
FFTResampler imageResampler(_originalWidth, _originalHeight, _rescaledWidth, _rescaledHeight, cpuCount, false);
imageResampler.Start();
for(size_t i=0; i!=_dataImageOriginal->ImageCount(); ++i)
imageResampler.AddTask(_dataImageOriginal->GetImage(i), largeScaleImage.GetImage(i));
for(size_t i=0; i!=_originalPsfs->size(); ++i)
{
scaledPsfs[i] = allocator.Allocate(_rescaledWidth*_rescaledHeight);
imageResampler.AddTask((*_originalPsfs)[i], scaledPsfs[i]);
}
imageResampler.Finish();
for(size_t i=0; i!=_dataImageOriginal->ImageCount(); ++i)
memcpy(nextScaleImage.GetImage(i), largeScaleImage.GetImage(i), sizeof(double)*_rescaledWidth*_rescaledHeight);
for(size_t i=0; i!=currentScaleModel.ImageCount(); ++i)
memset(currentScaleModel.GetImage(i), 0, sizeof(double)*_rescaledWidth*_rescaledHeight);
// Convolve the large and next scale images and the psfs
double* kernelImage = allocator.Allocate(_rescaledWidth * _rescaledHeight);
ao::uvector<double> shape;
size_t kernelSize;
MakeShapeFunction(currentScale * rescaleFactor, shape, kernelSize);
memset(kernelImage, 0, sizeof(double) * _rescaledWidth * _rescaledHeight);
FFTConvolver::PrepareSmallKernel(kernelImage, _rescaledWidth, _rescaledHeight, shape.data(), kernelSize);
for(size_t i=0; i!=largeScaleImage.ImageCount(); ++i)
FFTConvolver::ConvolveSameSize(largeScaleImage.GetImage(i), kernelImage, _rescaledWidth, _rescaledHeight);
for(size_t i=0; i!=scaledPsfs.size(); ++i) {
FFTConvolver::ConvolveSameSize(scaledPsfs[i], kernelImage, _rescaledWidth, _rescaledHeight);
FFTConvolver::ConvolveSameSize(scaledPsfs[i], kernelImage, _rescaledWidth, _rescaledHeight);
}
MakeShapeFunction(nextScale * rescaleFactor, shape, kernelSize);
memset(kernelImage, 0, sizeof(double) * _rescaledWidth * _rescaledHeight);
FFTConvolver::PrepareSmallKernel(kernelImage, _rescaledWidth, _rescaledHeight, shape.data(), kernelSize);
for(size_t i=0; i!=nextScaleImage.ImageCount(); ++i)
FFTConvolver::ConvolveSameSize(nextScaleImage.GetImage(i), kernelImage, _rescaledWidth, _rescaledHeight);
//FitsWriter writer;
//writer.SetImageDimensions(_rescaledWidth, _rescaledHeight);
//writer.Write("multiscale-kernel.fits", kernelImage);
//writer.Write("multiscale-current-scale.fits", largeScaleImage.GetImage(0));
//writer.Write("multiscale-next-scale.fits", nextScaleImage.GetImage(0));
//writer.Write("multiscale-psf.fits", scaledPsfs[0]);
//writer.SetImageDimensions(kernelSize, kernelSize);
//writer.Write("multiscale-shape.fits", shape.data());
size_t componentX=0, componentY=0;
findPeak(componentX, componentY);
std::cout << "Initial peak: " << peakDescription(largeScaleImage, componentX, componentY, rescaleFactor) << '\n';
size_t peakIndex = componentX + componentY*_rescaledWidth;
double peakNormalized = _dataImageLargeScale->JoinedValueNormalized(peakIndex) * rescaleFactor * rescaleFactor;
double firstThreshold = this->_threshold, stopGainThreshold = fabs(peakNormalized*(1.0-this->_stopGain)/thresholdBias);
std::cout << "Scale-adjusted threshold: " << firstThreshold*thresholdBias << ", major iteration stops at " << stopGainThreshold*thresholdBias << '\n';
if(stopGainThreshold > firstThreshold)
{
firstThreshold = stopGainThreshold;
std::cout << "Next major iteration for this scale at: " << stopGainThreshold << '\n';
}
else if(this->_stopGain != 1.0) {
std::cout << "Major iteration threshold reached global threshold of " << this->_threshold << " for this scale.\n";
}
std::vector<ao::lane<CleanTask>*> taskLanes(cpuCount);
std::vector<ao::lane<CleanResult>*> resultLanes(cpuCount);
boost::thread_group threadGroup;
for(size_t i=0; i!=cpuCount; ++i)
{
taskLanes[i] = new ao::lane<CleanTask>(1);
resultLanes[i] = new ao::lane<CleanResult>(1);
CleanThreadData cleanThreadData;
cleanThreadData.startY = (_rescaledHeight*i)/cpuCount;
cleanThreadData.endY = _rescaledHeight*(i+1)/cpuCount;
cleanThreadData.parent = this;
threadGroup.add_thread(new boost::thread(&FastMultiScaleClean<ImageSetType>::cleanThreadFunc, this, &*taskLanes[i], &*resultLanes[i], cleanThreadData));
}
while(fabs(peakNormalized) > firstThreshold*thresholdBias && this->_iterationNumber < this->_maxIter && !(largeScaleImage.IsComponentNegative(peakIndex) && this->_stopOnNegativeComponent))
{
if(this->_iterationNumber <= 10 ||
(this->_iterationNumber <= 100 && this->_iterationNumber % 10 == 0) ||
(this->_iterationNumber <= 1000 && this->_iterationNumber % 100 == 0) ||
this->_iterationNumber % 1000 == 0)
std::cout << "Iteration " << this->_iterationNumber << ": " << peakDescription(largeScaleImage, componentX, componentY, rescaleFactor) << '\n';
CleanTask task;
task.cleanCompX = componentX;
task.cleanCompY = componentY;
task.peak = largeScaleImage.Get(peakIndex);
for(size_t i=0; i!=cpuCount; ++i)
taskLanes[i]->write(task);
currentScaleModel.AddComponent(largeScaleImage, peakIndex, this->_subtractionGain * rescaleFactor * rescaleFactor);
double peakUnnormalized = 0.0;
for(size_t i=0; i!=cpuCount; ++i)
{
CleanResult result;
resultLanes[i]->read(result);
if(result.peakLevelUnnormalized >= peakUnnormalized && std::isfinite(result.peakLevelUnnormalized))
{
peakUnnormalized = result.peakLevelUnnormalized;
componentX = result.nextPeakX;
componentY = result.nextPeakY;
}
}
if(peakUnnormalized == 0.0)
{
std::cout << "No more components found at current scale: continuing to next scale.\n";
canCleanFurther = false;
break;
}
peakIndex = componentX + componentY*_rescaledWidth;
peakNormalized = largeScaleImage.JoinedValueNormalized(peakIndex) * rescaleFactor * rescaleFactor;
++this->_iterationNumber;
}
for(size_t i=0; i!=cpuCount; ++i)
taskLanes[i]->write_end();
threadGroup.join_all();
for(size_t i=0; i!=cpuCount; ++i)
{
delete taskLanes[i];
delete resultLanes[i];
}
std::cout << "Stopped on peak " << peakNormalized << '\n';
reachedStopGain = fabs(peakNormalized) <= stopGainThreshold*thresholdBias;
for(size_t i=0; i!=scaledPsfs.size(); ++i)
allocator.Free(scaledPsfs[i]);
MakeShapeFunction(currentScale * rescaleFactor, shape, kernelSize);
memset(kernelImage, 0, sizeof(double) * _rescaledWidth * _rescaledHeight);
FFTConvolver::PrepareSmallKernel(kernelImage, _rescaledWidth, _rescaledHeight, shape.data(), kernelSize);
double* convolvedModel = allocator.Allocate(_originalWidth * _originalHeight);
double* preparedPsf = allocator.Allocate(_originalWidth * _originalHeight);
FFTResampler modelResampler(_rescaledWidth, _rescaledHeight, _originalWidth, _originalHeight, cpuCount, false);
for(size_t i=0; i!=currentScaleModel.ImageCount(); ++i)
{
FFTConvolver::ConvolveSameSize(currentScaleModel.GetImage(i), kernelImage, _rescaledWidth, _rescaledHeight);
//writer.SetImageDimensions(_rescaledWidth, _rescaledHeight);
//writer.Write("multiscale-model-small.fits", currentScaleModel.GetImage(i));
modelResampler.RunSingle(currentScaleModel.GetImage(i), convolvedModel);
double *modelPtr = convolvedModel;
double *dest = _modelImage->GetImage(i), *destEnd = dest + _originalWidth*_originalHeight;
while(dest != destEnd) {
*dest += *modelPtr;
++modelPtr;
++dest;
}
FFTConvolver::PrepareKernel(preparedPsf, (*_originalPsfs)[_modelImage->PSFIndex(i)], _originalWidth, _originalHeight);
FFTConvolver::ConvolveSameSize(convolvedModel, preparedPsf, _originalWidth, _originalHeight);
dest = _dataImageOriginal->GetImage(i);
destEnd = dest + _originalWidth*_originalHeight;
modelPtr = convolvedModel;
while(dest != destEnd) {
*dest -= *modelPtr;
++modelPtr;
++dest;
}
}
allocator.Free(preparedPsf);
allocator.Free(convolvedModel);
allocator.Free(kernelImage);
this->_subtractionGain = oldSubtractionGain;
}
