protected void displayWeights(
final double osemspeedup,
final ArrayList<RandomAccessibleInterval<FloatType>> weights,
final RandomAccessibleInterval<FloatType> overlapImg,
final ImgFactory<FloatType> imgFactory) {
final DisplayImage d = new DisplayImage();
d.exportImage(overlapImg, bb, "Number of views per pixel");
final Img<FloatType> w = imgFactory.create(overlapImg, new FloatType());
final Img<FloatType> wosem = imgFactory.create(overlapImg, new FloatType());
// split up into many parts for multithreading
final Vector<ImagePortion> portions =
FusionHelper.divideIntoPortions(
Views.iterable(weights.get(0)).size(), Threads.numThreads() * 2);
// set up executor service
final ExecutorService taskExecutor = Executors.newFixedThreadPool(Threads.numThreads());
final ArrayList<Callable<String>> tasks = new ArrayList<Callable<String>>();
for (final ImagePortion portion : portions) {
tasks.add(
new Callable<String>() {
@Override
public String call() throws Exception {
final ArrayList<Cursor<FloatType>> cursors = new ArrayList<Cursor<FloatType>>();
final Cursor<FloatType> sum = w.cursor();
final Cursor<FloatType> sumOsem = wosem.cursor();
for (final RandomAccessibleInterval<FloatType> imgW : weights) {
final Cursor<FloatType> c = Views.iterable(imgW).cursor();
c.jumpFwd(portion.getStartPosition());
cursors.add(c);
}
sum.jumpFwd(portion.getStartPosition());
sumOsem.jumpFwd(portion.getStartPosition());
for (long j = 0; j < portion.getLoopSize(); ++j) {
double sumW = 0;
double sumOsemW = 0;
for (final Cursor<FloatType> c : cursors) {
final float w = c.next().get();
sumW += w;
sumOsemW += Math.min(1, w * osemspeedup);
}
sum.next().set((float) sumW);
sumOsem.next().set((float) sumOsemW);
}
return "done.";
}
});
}
// run threads
try {
// invokeAll() returns when all tasks are complete
taskExecutor.invokeAll(tasks);
} catch (final Exception e) {
IOFunctions.println("Failed to compute weight normalization for deconvolution: " + e);
e.printStackTrace();
return;
}
taskExecutor.shutdown();
d.exportImage(w, bb, "Sum of weights per pixel");
d.exportImage(wosem, bb, "OSEM=" + osemspeedup + ", sum of weights per pixel");
}
/**
* Fuses one stack, i.e. all angles/illuminations for one timepoint and channel
*
* @param timepoint
* @param channel
* @return
*/
public boolean fuseStacksAndGetPSFs(
final TimePoint timepoint,
final Channel channel,
final ImgFactory<FloatType> imgFactory,
final int osemIndex,
double osemspeedup,
WeightType weightType,
final HashMap<Channel, ChannelPSF> extractPSFLabels,
final long[] psfSize,
final HashMap<Channel, ArrayList<Pair<Pair<Angle, Illumination>, String>>> psfFiles,
final boolean transformLoadedPSFs) {
// TODO: get rid of this hack
if (files != null) {
weightType = WeightType.LOAD_WEIGHTS;
IOFunctions.println("WARNING: LOADING WEIGHTS FROM IMAGES, files.length()=" + files.length);
}
// get all views that are fused for this timepoint & channel
this.viewDescriptions =
FusionHelper.assembleInputData(spimData, timepoint, channel, viewIdsToProcess);
if (this.viewDescriptions.size() == 0) return false;
this.imgs = new HashMap<ViewId, RandomAccessibleInterval<FloatType>>();
this.weights = new HashMap<ViewId, RandomAccessibleInterval<FloatType>>();
final Img<FloatType> overlapImg;
if (weightType == WeightType.WEIGHTS_ONLY)
overlapImg = imgFactory.create(bb.getDimensions(), new FloatType());
else overlapImg = null;
final boolean extractPSFs =
(extractPSFLabels != null) && (extractPSFLabels.get(channel).getLabel() != null);
final boolean loadPSFs = (psfFiles != null);
if (extractPSFs) ePSF = new ExtractPSF<FloatType>();
else if (loadPSFs) ePSF = loadPSFs(channel, viewDescriptions, psfFiles, transformLoadedPSFs);
else {
ePSF = assignOtherChannel(channel, extractPSFLabels);
}
if (ePSF == null) return false;
// remember the extracted or loaded PSFs
extractPSFLabels.get(channel).setExtractPSFInstance(ePSF);
// we will need to run some batches until all is fused
for (int i = 0; i < viewDescriptions.size(); ++i) {
final ViewDescription vd = viewDescriptions.get(i);
IOFunctions.println(
"Transforming view "
+ i
+ " of "
+ (viewDescriptions.size() - 1)
+ " (viewsetup="
+ vd.getViewSetupId()
+ ", tp="
+ vd.getTimePointId()
+ ")");
IOFunctions.println(
"("
+ new Date(System.currentTimeMillis())
+ "): Reserving memory for transformed & weight image.");
// creating the output
RandomAccessibleInterval<FloatType> transformedImg; // might be null if WEIGHTS_ONLY
final RandomAccessibleInterval<FloatType>
weightImg; // never null (except LOAD_WEIGHTS which is not implemented yet)
if (weightType == WeightType.WEIGHTS_ONLY) transformedImg = overlapImg;
else transformedImg = imgFactory.create(bb.getDimensions(), new FloatType());
IOFunctions.println(
"("
+ new Date(System.currentTimeMillis())
+ "): Transformed image factory: "
+ imgFactory.getClass().getSimpleName());
// loading the input if necessary
final RandomAccessibleInterval<FloatType> img;
if (weightType == WeightType.WEIGHTS_ONLY && !extractPSFs) {
img = null;
} else {
IOFunctions.println("(" + new Date(System.currentTimeMillis()) + "): Loading image.");
img = ProcessFusion.getImage(new FloatType(), spimData, vd, true);
if (Img.class.isInstance(img))
IOFunctions.println(
"("
+ new Date(System.currentTimeMillis())
+ "): Input image factory: "
+ ((Img<FloatType>) img).factory().getClass().getSimpleName());
}
// initializing weights
IOFunctions.println(
"("
+ new Date(System.currentTimeMillis())
+ "): Initializing transformation & weights: "
+ weightType.name());
spimData.getViewRegistrations().getViewRegistration(vd).updateModel();
final AffineTransform3D transform =
spimData.getViewRegistrations().getViewRegistration(vd).getModel();
final long[] offset = new long[] {bb.min(0), bb.min(1), bb.min(2)};
if (weightType == WeightType.PRECOMPUTED_WEIGHTS || weightType == WeightType.WEIGHTS_ONLY)
weightImg = imgFactory.create(bb.getDimensions(), new FloatType());
else if (weightType == WeightType.NO_WEIGHTS)
weightImg =
Views.interval(
new ConstantRandomAccessible<FloatType>(
new FloatType(1), transformedImg.numDimensions()),
transformedImg);
else if (weightType == WeightType.VIRTUAL_WEIGHTS) {
final Blending blending = getBlending(img, blendingBorder, blendingRange, vd);
weightImg =
new TransformedRealRandomAccessibleInterval<FloatType>(
blending, new FloatType(), transformedImg, transform, offset);
} else // if ( processType == ProcessType.LOAD_WEIGHTS )
{
IOFunctions.println("WARNING: LOADING WEIGHTS FROM: '" + new File(files[i]) + "'");
ImagePlus imp = StackImgLoaderIJ.open(new File(files[i]));
weightImg = imgFactory.create(bb.getDimensions(), new FloatType());
StackImgLoaderIJ.imagePlus2ImgLib2Img(imp, (Img<FloatType>) weightImg, false);
imp.close();
if (debugImport) {
imp = ImageJFunctions.show(weightImg);
imp.setTitle("ViewSetup " + vd.getViewSetupId() + " Timepoint " + vd.getTimePointId());
}
}
// split up into many parts for multithreading
final Vector<ImagePortion> portions =
FusionHelper.divideIntoPortions(
Views.iterable(transformedImg).size(), Threads.numThreads() * 4);
// set up executor service
final ExecutorService taskExecutor = Executors.newFixedThreadPool(Threads.numThreads());
final ArrayList<Callable<String>> tasks = new ArrayList<Callable<String>>();
IOFunctions.println(
"("
+ new Date(System.currentTimeMillis())
+ "): Transforming image & computing weights.");
for (final ImagePortion portion : portions) {
if (weightType == WeightType.WEIGHTS_ONLY) {
final Interval imgInterval =
new FinalInterval(
ViewSetupUtils.getSizeOrLoad(
vd.getViewSetup(),
vd.getTimePoint(),
spimData.getSequenceDescription().getImgLoader()));
final Blending blending = getBlending(imgInterval, blendingBorder, blendingRange, vd);
tasks.add(
new TransformWeights(
portion, imgInterval, blending, transform, overlapImg, weightImg, offset));
} else if (weightType == WeightType.PRECOMPUTED_WEIGHTS) {
final Blending blending = getBlending(img, blendingBorder, blendingRange, vd);
tasks.add(
new TransformInputAndWeights(
portion, img, blending, transform, transformedImg, weightImg, offset));
} else if (weightType == WeightType.NO_WEIGHTS
|| weightType == WeightType.VIRTUAL_WEIGHTS
|| weightType == WeightType.LOAD_WEIGHTS) {
tasks.add(new TransformInput(portion, img, transform, transformedImg, offset));
} else {
throw new RuntimeException(weightType.name() + " not implemented yet.");
}
}
try {
// invokeAll() returns when all tasks are complete
taskExecutor.invokeAll(tasks);
} catch (final InterruptedException e) {
IOFunctions.println("Failed to compute fusion: " + e);
e.printStackTrace();
return false;
}
taskExecutor.shutdown();
// extract PSFs if wanted
if (extractPSFs) {
final ArrayList<double[]> llist =
getLocationsOfCorrespondingBeads(
timepoint, vd, extractPSFLabels.get(channel).getLabel());
IOFunctions.println(
"("
+ new Date(System.currentTimeMillis())
+ "): Extracting PSF for viewsetup "
+ vd.getViewSetupId()
+ " using label '"
+ extractPSFLabels.get(channel).getLabel()
+ "'"
+ " ("
+ llist.size()
+ " corresponding detections available)");
ePSF.extractNextImg(img, vd, transform, llist, psfSize);
}
if (weightType != WeightType.WEIGHTS_ONLY) imgs.put(vd, transformedImg);
weights.put(vd, weightImg);
}
// normalize the weights
final ArrayList<RandomAccessibleInterval<FloatType>> weightsSorted =
new ArrayList<RandomAccessibleInterval<FloatType>>();
for (final ViewDescription vd : viewDescriptions) weightsSorted.add(weights.get(vd));
IOFunctions.println(
"("
+ new Date(System.currentTimeMillis())
+ "): Computing weight normalization for deconvolution.");
final WeightNormalizer wn;
if (weightType == WeightType.WEIGHTS_ONLY
|| weightType == WeightType.PRECOMPUTED_WEIGHTS
|| weightType == WeightType.LOAD_WEIGHTS) wn = new WeightNormalizer(weightsSorted);
else if (weightType == WeightType.VIRTUAL_WEIGHTS)
wn = new WeightNormalizer(weightsSorted, imgFactory);
else // if ( processType == ProcessType.NO_WEIGHTS )
wn = null;
if (wn != null && !wn.process()) return false;
// put the potentially modified weights back
for (int i = 0; i < viewDescriptions.size(); ++i)
weights.put(viewDescriptions.get(i), weightsSorted.get(i));
this.minOverlappingViews = wn.getMinOverlappingViews();
this.avgOverlappingViews = wn.getAvgOverlappingViews();
IOFunctions.println(
"("
+ new Date(System.currentTimeMillis())
+ "): Minimal number of overlapping views: "
+ getMinOverlappingViews()
+ ", using "
+ (this.minOverlappingViews = Math.max(1, this.minOverlappingViews)));
IOFunctions.println(
"("
+ new Date(System.currentTimeMillis())
+ "): Average number of overlapping views: "
+ getAvgOverlappingViews()
+ ", using "
+ (this.avgOverlappingViews = Math.max(1, this.avgOverlappingViews)));
if (osemIndex == 1) osemspeedup = getMinOverlappingViews();
else if (osemIndex == 2) osemspeedup = getAvgOverlappingViews();
IOFunctions.println(
"("
+ new Date(System.currentTimeMillis())
+ "): Adjusting for OSEM speedup = "
+ osemspeedup);
if (weightType == WeightType.WEIGHTS_ONLY)
displayWeights(osemspeedup, weightsSorted, overlapImg, imgFactory);
else adjustForOSEM(weights, weightType, osemspeedup);
IOFunctions.println(
"("
+ new Date(System.currentTimeMillis())
+ "): Finished precomputations for deconvolution.");
return true;
}