def setZ(self, z):
def handleClic(curve, edge):
print curve, "edge", edge
curve.setPen(pg.mkPen({'color': (0, 0, 1), 'width': 6}))
pen = pg.mkPen({'color': "FF0", 'width': 6})
vb = self.viewBox
def vbAdd(item):
if item.zValue() < vb.zValue():
item.setZValue(vb.zValue() + 1)
#item.setParentItem(vb.childGroup)
vb.addedItems.append(item)
dataSlice = self.dataDict['raw'][:, :, z]
labelSlice = self.labels[:, :, z]
self.imgItem.setImage(dataSlice)
self.imgItem.update()
#self.viewBox.update()
#for curve in self.curves:
# self.viewBox.removeItem(curve)
self.curves = []
slicesEdges = vigra.graphs.SliceEdges(self.rag)
with vigra.Timer("findSlicesEdges"):
slicesEdges.findSlicesEdges(labelSlice)
with vigra.Timer("render them"):
visibleEdges = slicesEdges.visibleEdges()
for edge in visibleEdges:
edge = long(edge)
lineIds = slicesEdges.lineIds(edge)
totalLine = []
for lid in lineIds:
line = slicesEdges.line(long(lid)).astype('float32') / 2.0
totalLine.append(line)
totalLine.append([[float('nan'), float('nan')]])
totalLine = numpy.concatenate(totalLine, axis=0)
lx = totalLine[:, 0]
ly = totalLine[:, 1]
#with vigra.Timer("get curve"):
curve = pg.PlotCurveItem(clickable=True, parent=vb.childGroup)
curve.setPen(pen)
curve.setData(lx, ly, connect="finite")
curve.sigClicked.connect(partial(handleClic, edge=edge))
self.curves.append(curve)
#with vigra.Timer("add curve"):
#vbAdd(curve)
#self.viewBox.update()
self.allCurves.setCurves(self.curves)
self.viewBox.addItem(self.allCurves)
with vigra.Timer("update auto range"):
vb.updateAutoRange()
def _update_rendering(self):
if not self.render:
return
op = self.topLevelOperatorView
if not self._renderMgr.ready:
shape = op.InputData.meta.shape[1:4]
self._renderMgr.setup(op.InputData.meta.shape[1:4])
# remove nonexistent objects
self._shownObjects3D = dict((k, v) for k, v in self._shownObjects3D.iteritems()
if k in op.MST.value.object_lut.keys())
lut = numpy.zeros(op.MST.value.nodeNum+1, dtype=numpy.int32)
for name, label in self._shownObjects3D.iteritems():
objectSupervoxels = op.MST.value.objects[name]
lut[objectSupervoxels] = label
if self._showSegmentationIn3D:
# Add segmentation as label, which is green
lut[:] = numpy.where( op.MST.value.getSuperVoxelSeg() == 2, self._segmentation_3d_label, lut )
import vigra
with vigra.Timer("remapping"):
self._renderMgr.volume = lut[op.MST.value.supervoxelUint32] # (Advanced indexing)
self._update_colors()
self._renderMgr.update()
def setZ(self, z):
vb = self.viewBox
dataSlice = self.dataDict['raw'][:, :, z]
labelSlice = self.labels[:, :, z]
self.imgItem.setImage(dataSlice)
self.imgItem.update()
slicesEdges = vigra.graphs.SliceEdges(self.rag)
with vigra.Timer("find slices"):
slicesEdges.findSlicesEdges(labelSlice)
if self.allCurves is None:
self.allCurves = AllCurves()
#self.viewBox.addItem(self.allCurves)
self.curves = self.allCurves.curves
ci = 0
#with vigra.Timer("build curves"):
visibleEdges = slicesEdges.visibleEdges()
for edge in visibleEdges:
edge = long(edge)
lineIds = slicesEdges.lineIds(edge)
totalLine = []
for lid in lineIds:
line = slicesEdges.line(long(lid)).astype('float32') / 2.0
totalLine.append(line)
totalLine.append([[float('nan'), float('nan')]])
totalLine = numpy.concatenate(totalLine, axis=0)
lx = totalLine[:, 0]
ly = totalLine[:, 1]
try:
curve = self.curves[ci]
except:
curve = BvPlotCurveItem(clickable=False, parent=self.imgItem)
self.curves.append(curve)
curve.viewer = self
#self.viewBox.addItem(curve,ignoreBounds=True)
ci += 1
leftTop = numpy.nanmin(lx), numpy.nanmin(ly)
rightBottom = numpy.nanmax(lx), numpy.nanmax(ly)
curve.setToolTip("id%d" % edge)
curve.bRect.setCoords(leftTop[0], leftTop[1], rightBottom[0],
rightBottom[1])
curve.edge = edge
curve.setPen(self.getPen(edge))
curve.setData(lx, ly, connect="finite")
curve.setVisible(True)
self.nCurves = ci
print self.nCurves
for cii in range(ci, len(self.curves)):
self.curves[cii].setVisible(False)
def largeSeedWatershed(raw, pmap, seeds, membraneWidth=7.0, visu=False):
blockShape = (100, ) * 3
cOpts = vbw.convOpts
pmap = numpy.require(pmap, dtype='float32')
with vigra.Timer("add noise"):
mx = pmap.max()
sshape = pmap.squeeze().shape
noise = numpy.random.rand(*sshape) * (0.05 * mx)
noise = vigra.taggedView(noise.astype('float32'), 'xyz')
opts = vbw.convOpts(blockShape=blockShape, sigma=4.0)
noise = vbw.gaussianSmooth(noise, options=opts)
pmap += noise
with vigra.Timer("smoothed tie breaker"):
# compute a smoothed map as tie breaker
opts = vbw.convOpts(blockShape=blockShape, sigma=membraneWidth / 2.0)
gaussianSmoothedPmap = vbw.gaussianSmooth(pmap, options=opts)
addEps = 0.3
growingMap = gaussianSmoothedPmap
growingMap *= addEps
#
opts = vbw.convOpts(blockShape=blockShape, sigma=membraneWidth / 7.50)
notSoMuch = vbw.gaussianSmooth(pmap, options=opts)
# get the actual growing mapz
growingMap += notSoMuch
growingMap /= 1.0 + addEps
with vigra.Timer("watershedsNew"):
# do the actual watershed
growingMap = vigra.taggedView(growingMap, 'xyz')
seeds = numpy.require(seeds, dtype='uint32')
seeds = vigra.taggedView(seeds, 'xyz')
seg, nSeg = vigra.analysis.watershedsNew(image=growingMap, seeds=seeds)
if visu:
grayData = [(raw, "raw"), (pmap, "pmap"), (growingMap, "growingMap")]
segData = [(seeds, "seeds"), (seg, "seg")]
skneuro.addHocViewer(grayData, segData)
return seg, nSeg
def scrolled(d):
if d > 0:
d = 5
else:
d = -5
if self.ndim == 3:
newSlice = min(self.nSlices - 1, self.currentSlice - d)
newSlice = max(0, newSlice)
self.currentSlice = newSlice
with vigra.Timer("scroll"):
self.setZ(self.currentSlice)
def computeRag(oversegFile, dset):
oversegFileH5 = h5py.File(oversegFile)
oversegDset = oversegFileH5[dset]
try:
maxLabel = oversegDset.attrs['maxLabel']
#print("load max label")
except:
print("compute max label")
maxLabel = h5Max(oversegDset)
oversegDset.attrs['maxLabel'] = maxLabel
oversegFileH5.close()
print("max label", maxLabel)
nLabels = maxLabel + 1
cs = nifty.hdf5.CacheSettings()
cs.hashTabelSize = 977
cs.nBytes = 990000000
cs.rddc = 0.5
h5File = nifty.hdf5.openFile(oversegFile, cs)
labelsArray = nifty.hdf5.Hdf5ArrayUInt32(h5File, dset)
ragFile = os.path.join(workDir, 'rag.h5')
if not os.path.isfile(ragFile):
with vigra.Timer("rag"):
gridRag = nifty.graph.rag.gridRagHdf5(labelsArray,
nLabels,
blockShape=[150, 150, 150],
numberOfThreads=20)
print("serialize")
serialization = gridRag.serialize()
print("save serialization")
ragH5File = h5py.File(ragFile)
ragH5File['data'] = serialization
ragH5File.close()
else:
ragH5File = h5py.File(ragFile, 'r')
serialization = ragH5File['data']
gridRag = nifty.graph.rag.gridRagHdf5(labelsArray,
nLabels,
serialization=serialization)
ragH5File.close()
return gridRag
if True:
#pmapH5 = h5py.File(pmapPath,'r')
#pmapDset = pmapH5['data']
print("load pmap in ram (since we have enough")
pmapArray = pmapDset[:, :, :, :]
pmapArray = pmapArray[:, :, :, 0]
heightMapH5 = h5py.File(heightMapFile, 'w')
heightMapDset = heightMapH5.create_dataset('data',
shape=pmapDset.shape[0:3],
chunks=(100, 100, 100),
dtype='float32')
with vigra.Timer("st"):
params = {
"axisResolution": [2.0, 2.0, 2.0],
"featureBlockShape": [400, 400, 400],
"invertPmap": False,
#"roiBegin": [0,0,0],
#"roiEnd": [1000,1000,1000],
#"nWorkers":1,
}
print(pmapDset.shape)
membraneOverseg3D(pmapArray, heightMapDset, **params)
heightMapH5.close()
pmapH5.close()
if True:
options.blockShape = (64, ) * 3
print "hessianEv"
ev = bw.hessianOfGaussianFirstEigenvalue(volume, options)
print ev.shape, ev.dtype
options.stdDev = (4.5, ) * 3
print "smooth"
ev = bw.gaussianSmooth(ev, options)
vigra.impex.writeHDF5(ev, "growmap.h5", 'data')
else:
ev = vigra.impex.readHDF5("growmap.h5", 'data')
with vigra.Timer("watershedsNew"):
labels, nseg = vigra.analysis.unionFindWatershed3D(ev, (100, 100, 100))
print "gridGraph"
gridGraph = graphs.gridGraph(labels.shape)
rag = graphs.regionAdjacencyGraph(gridGraph, labels)
rag.writeHDF5("rag.h5", 'data')
else:
rag = vigra.graphs.loadGridRagHDF5("rag.h5", 'data')
labels = rag.labels
app = QtGui.QApplication([])
class DownCtrl(QtGui.QWidget):
def __init__(self, *args, **kwargs):
def oversegPipeline(pmapPath,
outFolder,
outName,
pixelResNm=2.0,
reduceBy=[5, 10, 20]):
scale = pixelResNm / 2.0
if not os.path.exists(outFolder):
os.makedirs(outFolder)
class DummyWithStatement:
def __enter__(self):
pass
def __exit__(self, type, value, traceback):
pass
def makeBall(r):
size = 2 * r + 1
mask = numpy.zeros([size] * 3)
for x0 in range(-1 * r, r + 1):
for x1 in range(-1 * r, r + 1):
for x2 in range(-1 * r, r + 1):
if math.sqrt(x0**2 + x1**2 + x2**2) <= r:
mask[x0 + r, x1 + r, x2 + r] = 1
return mask, (r, r, r)
def membraneOverseg3D(pmapDset, heightMapDset, **kwargs):
axisResolution = kwargs.get("axisResolution", ['4nm'] * 3)
featureBlockShape = kwargs.get("featureBlockShape", ['100'] * 3)
shape = pmapDset.shape[0:3]
roiBegin = kwargs.get("roiBegin", [0] * 3)
roiEnd = kwargs.get("roiEnd", shape)
nWorkers = kwargs.get("nWorkers", cpu_count())
invertPmap = kwargs.get("invertPmap", False)
blocking = nifty.tools.blocking(roiBegin=roiBegin,
roiEnd=roiEnd,
blockShape=featureBlockShape)
margin = [45, 45, 45]
def pmapToHeightMap(pmap):
r = int(min(3.0 * scale, 1.0) + 0.5)
footprint, origin = makeBall(r=r)
blurredSmall = fastfilters.gaussianSmoothing(pmap, 1.0 * scale)
blurredLarge = fastfilters.gaussianSmoothing(pmap, 6.0 * scale)
blurredSuperLarge = fastfilters.gaussianSmoothing(
pmap, 10.0 * scale)
combined = pmap + blurredSuperLarge * 0.3 + 0.15 * blurredLarge + 0.1 * blurredSmall
r = int(min(5.0 * scale, 1.0) + 0.5)
footprint, origin = makeBall(r=r)
combined = scipy.ndimage.percentile_filter(
input=combined,
#size=(20,20,20),
footprint=footprint,
#origin=origin,
mode='reflect',
percentile=50.0)
if False:
nifty.viewer.view3D(pmap, show=False, title='pm', cmap='gray')
nifty.viewer.view3D(medianImg,
show=False,
title='medianImg',
cmap='gray')
nifty.viewer.view3D(combined,
show=False,
title='combined',
cmap='gray')
pylab.show()
return combined
numberOfBlocks = blocking.numberOfBlocks
lock = threading.Lock()
noLock = DummyWithStatement()
done = [0]
for blockIndex in range(numberOfBlocks):
blockWithHalo = blocking.getBlockWithHalo(blockIndex, margin)
block = blocking.getBlock(blockIndex)
outerBlock = blockWithHalo.outerBlock
innerBlock = blockWithHalo.innerBlock
innerBlockLocal = blockWithHalo.innerBlockLocal
#print("B ",block.begin, block.end)
#print("O ",outerBlock.begin, outerBlock.end)
#print("I ",innerBlock.begin, innerBlock.end)
#print("IL",innerBlockLocal.begin, innerBlockLocal.end)
with nifty.tools.progressBar(size=numberOfBlocks) as bar:
def f(blockIndex):
blockWithHalo = blocking.getBlockWithHalo(
blockIndex, margin, margin)
#print "fo"
outerBlock = blockWithHalo.outerBlock
outerSlicing = nifty.tools.getSlicing(outerBlock.begin,
outerBlock.end)
b, e = outerBlock.begin, outerBlock.end
#print bi
#
maybeLock = [lock, noLock][isinstance(pmapDset, numpy.ndarray)]
with maybeLock:
if invertPmap:
outerPmap = 1.0 - pmapDset[b[0]:e[0], b[1]:e[1],
b[2]:e[2]]
else:
outerPmap = pmapDset[b[0]:e[0], b[1]:e[1], b[2]:e[2]]
heightMap = pmapToHeightMap(outerPmap)
#
innerBlockLocal = blockWithHalo.innerBlockLocal
b, e = innerBlockLocal.begin, innerBlockLocal.end
innerHeightMap = heightMap[b[0]:e[0], b[1]:e[1], b[2]:e[2]]
b, e = blockWithHalo.innerBlock.begin, blockWithHalo.innerBlock.end
if isinstance(heightMapDset, numpy.ndarray):
#print("NOT LOCKED")
heightMapDset[b[0]:e[0], b[1]:e[1],
b[2]:e[2]] = innerHeightMap
with lock:
done[0] += 1
bar.update(done[0])
else:
with lock:
#print("locked",b,e)
heightMapDset[b[0]:e[0], b[1]:e[1],
b[2]:e[2]] = innerHeightMap
done[0] += 1
bar.update(done[0])
nifty.tools.parallelForEach(range(blocking.numberOfBlocks),
f=f,
nWorkers=nWorkers)
def makeSmallerSegNifty(oseg, volume_feat, reduceBySetttings,
wardnessSettings, baseFilename):
import nifty
import nifty.graph
import nifty.graph.rag
import nifty.graph.agglo
nrag = nifty.graph.rag
nagglo = nifty.graph.agglo
print("overseg in c order starting at zero")
oseg = numpy.require(oseg, dtype='uint32', requirements='C')
oseg -= 1
print("make rag")
rag = nifty.graph.rag.gridRag(oseg)
print("volfeatshape")
vFeat = numpy.require(volume_feat, dtype='float32', requirements='C')
print("accumulate means and counts")
eFeatures, nFeatures = nrag.accumulateMeanAndLength(
rag, vFeat, [100, 100, 100], -1)
eMeans = eFeatures[:, 0]
eSizes = eFeatures[:, 1]
nSizes = nFeatures[:, 1]
print("get clusterPolicy")
for wardness in wardnessSettings:
for reduceBy in reduceBySetttings:
print("wardness", wardness, "reduceBy", reduceBy)
numberOfNodesStop = int(
float(rag.numberOfNodes) / float(reduceBy) + 0.5)
numberOfNodesStop = max(1, numberOfNodesStop)
numberOfNodesStop = min(rag.numberOfNodes, numberOfNodesStop)
clusterPolicy = nagglo.edgeWeightedClusterPolicy(
graph=rag,
edgeIndicators=eMeans,
edgeSizes=eSizes,
nodeSizes=nSizes,
numberOfNodesStop=numberOfNodesStop,
sizeRegularizer=float(wardness))
print("do clustering")
agglomerativeClustering = nagglo.agglomerativeClustering(
clusterPolicy)
agglomerativeClustering.run()
seg = agglomerativeClustering.result() #out=[1,2,3,4])
print("make seg dense")
dseg = nifty.tools.makeDense(seg)
print(dseg.dtype, type(dseg))
print("project to pixels")
pixelData = nrag.projectScalarNodeDataToPixels(
rag, dseg.astype('uint32'))
print("done")
outFilename = baseFilename + str(wardness) + "_" + str(
reduceBy) + ".h5"
agglosegH5 = h5py.File(outFilename, 'w')
agglosegDset = agglosegH5.create_dataset('data',
shape=oseg.shape[0:3],
chunks=(100, 100,
100),
dtype='uint32',
compression="gzip")
agglosegDset[:, :, :] = pixelData
agglosegH5.close()
if True:
pmapH5 = h5py.File(pmapPath, 'r')
pmapDset = pmapH5['data']
shape = list(pmapDset.shape[0:3])
subset = None
sshape = (subset, ) * 3
heightMapFile = os.path.join(outFolder, "%s_heightMap.h5" % outName)
heightMapH5 = h5py.File(heightMapFile, 'w')
heightMapDset = heightMapH5.create_dataset('data',
shape=shape,
dtype='float32',
chunks=(64, ) * 3)
print("bra", heightMapFile)
#sys.exit(1)
print("load pmap in ram (since we have enough")
if subset is None:
pmapArray = pmapDset[:, :, :]
else:
pmapArray = pmapDset[0:subset, 0:subset, 0:subset]
pmapH5.close()
if shape[0] >= 2000:
featureBlockShape = [350, 350, 350]
elif shape[0] < 2000 and shape[0] >= 1000:
featureBlockShape = [250, 250, 250]
elif shape[0] < 1000 and shape[0] >= 500:
featureBlockShape = [100, 100, 100]
with vigra.Timer("st"):
params = {
"axisResolution": [2.0, 2.0, 2.0],
"featureBlockShape": featureBlockShape,
"invertPmap": False,
#"roiBegin": [0,0,0],
#"roiEnd": sshape
#"nWorkers":1,
}
membraneOverseg3D(pmapArray, heightMapDset, **params)
heightMapH5.close()
if True:
with vigra.Timer("read hmap"):
heightMapFile = os.path.join(outFolder,
"%s_heightMap.h5" % outName)
heightMapH5 = h5py.File(heightMapFile, 'r')
heightMapDset = heightMapH5['data']
heightMap = heightMapDset[:, :, :]
shape = list(heightMap.shape)
heightMapH5.close()
with vigra.Timer("do overseg"):
overseg, nseg = vigra.analysis.unionFindWatershed3D(
heightMap.T, blockShape=(100, 100, 100))
overseg = overseg.T
with vigra.Timer("write res"):
oversegFile = os.path.join(outFolder,
"%s_ufd_overseg.h5" % outName)
oversegH5 = h5py.File(oversegFile, 'w')
oversegDset = oversegH5.create_dataset('data',
shape=shape,
data=overseg,
chunks=(64, ) * 3,
compression='gzip')
oversegDset.attrs['nseg'] = nseg
oversegH5.close()
#if False:
# import nifty.hdf5
# with vigra.Timer("ws"):
# nifty.hdf5.unionFindWatershed(heightMapFile,'data', oversegFile,'data',[100,100,100])
if True:
heightMapFile = os.path.join(outFolder, "%s_heightMap.h5" % outName)
heightMapH5 = h5py.File(heightMapFile, 'r')
heightMapDset = heightMapH5['data']
oversegFile = os.path.join(outFolder, "%s_ufd_overseg.h5" % outName)
oversegH5 = h5py.File(oversegFile, 'r')
oversegDset = oversegH5['data']
print("read hmap")
heightMap = heightMapDset[:, :, :]
print("read oseg")
overseg = oversegDset[:, :, :]
heightMapH5.close()
oversegH5.close()
print("make smaller")
agglosegBaseFile = os.path.join(outFolder, "%s_agglo_seg" % outName)
makeSmallerSegNifty(overseg, heightMap, reduceBy, [0.3],
agglosegBaseFile)
def singleBlockPipeline(rawFile, pmapFile, oversegFile, ilpFile, outFolder, outName, sigmas=(2.0,4.0,8.0), filtOnRaw=True, filtOnPmap=True, blockShape=(254,)*3,
minSegSize=None ,beta=0.5, stages=[1,1,1]):
if not os.path.exists(outFolder):
os.makedirs(outFolder)
# save stuff
def saveRes(data,name):
f5 = h5py.File(os.path.join(outFolder,'%s_%s.h5'%(outName,name)),'w')
f5['data'] = data
f5.close()
# load stuff
def loadRes(name):
f5 = h5py.File(os.path.join(outFolder,'%s_%s.h5'%(outName,name)),'r')
data = f5['data'][:]
f5.close()
return data
rawH5File = h5py.File(rawFile,'r')
osegH5File = h5py.File(oversegFile,'r')
pmapH5File = h5py.File(pmapFile,'r')
# dsets
rawDset = rawH5File['data']
osegDset = osegH5File['data']
pmapDset = pmapH5File['data']
print("raw",rawDset.shape)
print("oseg",osegDset.shape)
print("pmap",pmapDset.shape)
shape = rawDset.shape
#load labels
uvIdsTrain, labelsTrain = getIlpLabels(ilpFile)
blocking = nifty.tools.blocking(roiBegin=[0]*3, roiEnd=shape,
blockShape=blockShape)
nBlocks = blocking.numberOfBlocks
lock = threading.Lock()
lock2 = threading.Lock()
blockRes = [None]*nBlocks
if len(sigmas) > 0:
maxSigma = max(sigmas)
if stages[0] == 1:
print("STAGE 0")
filtList = []
filtListPmap = []
bs = 100
hs = [s//2 for s in shape]
raw = rawDset[hs[0] : hs[0] + bs, hs[1] : hs[1] + bs, hs[2] : hs[2] + bs].astype('float32')
pm = pmapDset[hs[0] : hs[0] + bs, hs[1] : hs[1] + bs, hs[2] : hs[2] + bs].astype('float32')
for sigma in sigmas:
if filtOnRaw:
filtList.append(GaussianGradientMagnitude(raw=raw, sigma=sigma))
filtList.append(LaplacianOfGaussian(raw=raw, sigma=sigma))
filtList.append(HessianOfGaussianEv(raw=raw, sigma=sigma))
filtList.append(GaussianSmoothing(raw=raw, sigma=sigma))
if filtOnPmap:
filtListPmap.append(GaussianGradientMagnitude(raw=pm, sigma=sigma))
filtListPmap.append(LaplacianOfGaussian(raw=pm, sigma=sigma))
filtListPmap.append(HessianOfGaussianEv(raw=pm, sigma=sigma))
filtListPmap.append(GaussianSmoothing(raw=pm, sigma=sigma))
def f(blockIndex):
# labels halo
blockWithBorder = blocking.getBlockWithHalo(blockIndex, [0,0,0],[1,1,1])
outerBlock = blockWithBorder.outerBlock
bAcc, eAcc = outerBlock.begin, outerBlock.end
# filters
filterRes= []
if len(sigmas) > 0 and (filtOnRaw or filtOnPmap):
halo = [int(round(3.5*maxSigma + 3.0))] * 3
else:
halo = [0]*3
blockWithBorder = blocking.addHalo(outerBlock, halo)
outerBlock = blockWithBorder.outerBlock
innerBlockLocal = blockWithBorder.innerBlockLocal
bFilt, eFilt = outerBlock.begin, outerBlock.end
bFiltCore, eFiltCore = innerBlockLocal.begin, innerBlockLocal.end
with lock:
rawFilt = rawDset[bFilt[0]:eFilt[0], bFilt[1]:eFilt[1], bFilt[2]:eFilt[2]].astype('float32')
pmFilt = pmapDset[bFilt[0]:eFilt[0], bFilt[1]:eFilt[1], bFilt[2]:eFilt[2]].astype('float32')
oseg = osegDset[bAcc[0]:eAcc[0], bAcc[1]:eAcc[1], bAcc[2]:eAcc[2]]
raw = rawFilt[bFiltCore[0]:eFiltCore[0], bFiltCore[1]:eFiltCore[1], bFiltCore[2]:eFiltCore[2]]/255.0
pmap = pmFilt[bFiltCore[0]:eFiltCore[0], bFiltCore[1]:eFiltCore[1], bFiltCore[2]:eFiltCore[2]]/255.0
filterRes.append(raw[...,None])
filterRes.append(pmap[...,None])
for filt in filtList:
res = filt(rawFilt)
res = res[bFiltCore[0]:eFiltCore[0], bFiltCore[1]:eFiltCore[1], bFiltCore[2]:eFiltCore[2],:]
filterRes.append(res)
for filt in filtListPmap:
res = filt(pmFilt)
res = res[bFiltCore[0]:eFiltCore[0], bFiltCore[1]:eFiltCore[1], bFiltCore[2]:eFiltCore[2],:]
filterRes.append(res)
#for vd in filterRes:
# print(vd.shape,vd.dtype)
#
voxelData = numpy.concatenate(filterRes, axis=3)
voxelData = numpy.require(voxelData,requirements=['C_CONTIGUOUS'],dtype='float32')
#print(voxelData.strides)
#last axis is continous
assert voxelData.strides[3] == 4
blockData = nseg.BlockData(blocking=blocking, blockIndex=blockIndex,
numberOfChannels=voxelData.shape[3],
numberOfBins=20)
blockData.accumulate(oseg, voxelData)
blockRes[blockIndex] = blockData
nifty.tools.parallelForEach(iterable=range(nBlocks), f=f,
nWorkers=-1, showBar=True,
size=nBlocks, name="ComputeFeatures")
def mergePairwise(toMerge):
while(len(toMerge) != 1):
newList = []
l = len(toMerge)
pairs = (l+1)//2
print(l)
for p in range(pairs):
p0 = p*2
p1 = p*2 + 1
if(p1<l):
b0 = toMerge[p0]
b1 = toMerge[p1]
b0.merge(b1)
newList.append(b0)
else:
newList.append(toMerge[p0])
toMerge = newList
return toMerge[0]
mergedRes = mergePairwise(blockRes)
uvIds = mergedRes.uvIds()
allFeat = mergedRes.extractFeatures()
nodeCounts = mergedRes.nodeCounts()
edgeCounts = mergedRes.edgeCounts()
# mean pmap
nodeMeansP = mergedRes.nodeMeans(1)
edgeMeansP = mergedRes.edgeMeans(1)
saveRes(nodeCounts,'nodeCounts')
saveRes(edgeCounts,'edgeCounts')
saveRes(nodeMeansP,'nodeMeansP')
saveRes(edgeMeansP,'edgeMeansP')
saveRes(uvIds,'uvIds')
saveRes(allFeat,'allFeat')
if stages[1] == 1:
from sklearn.ensemble import RandomForestClassifier
#load stuff
uvIds = loadRes('uvIds')
allFeat = loadRes('allFeat')
nodeCounts = loadRes('nodeCounts')
edgeCounts = loadRes('edgeCounts')
nodeMeansP = loadRes('nodeMeansP')
edgeMeansP = loadRes('edgeMeansP')
# get the graph
numberOfNodes = uvIds.max() + 1
g = nifty.graph.UndirectedGraph(numberOfNodes)
g.insertEdges(uvIds)
print("graph",g)
if False:
print("ucm feat")
print("edgeMeansP",edgeMeansP)
ucmFeat = nagglo.ucmFeatures(graph=g, edgeIndicators=edgeMeansP,
edgeSizes=edgeCounts, nodeSizes=nodeCounts,
sizeRegularizers=numpy.arange(0.025,1,0.1))
allFeat = numpy.concatenate([allFeat, ucmFeat],axis=1)
sys.exit()
print("rest")
# mapping
uvToIndex = dict()
for i,uv in enumerate(uvIds):
uv = long(uv[0]),long(uv[1])
uvToIndex[uv] = i
featTrain = []
lTrain = []
for i,(uv,l) in enumerate(zip(uvIdsTrain,labelsTrain)):
uv = long(uv[0]),long(uv[1])
if uv not in uvToIndex:
print(g,uv)
assert False
else:
i = uvToIndex[uv]
featTrain.append(allFeat[i,:][None,:])
lTrain.append(l)
featTrain = numpy.concatenate(featTrain,axis=0).astype('float32')
labelsTrain = numpy.array(lTrain,dtype='uint32')#[:,None]
print("training set size",featTrain.shape,labelsTrain.shape)
# train the rf
#rf = vigra.learning.RandomForest(treeCount = 10000,
# #min_split_node_size=20,
# #sample_classes_individually=True
#)
#oob = rf.learnRF(featTrain, labelsTrain)
rf = RandomForestClassifier(n_estimators=1000, n_jobs=40)
rf.fit(featTrain, labelsTrain)
p1 = rf.predict_proba(allFeat)[:,1]
#print("oob",oob)
#with vigra.Timer("predict"):
# p1 = rf.predictProbabilities(allFeat)[:,1]
# get weights
eps = 0.00001
beta = float(beta)
p1 = numpy.clip(p1, eps, 1.0 - eps)
p0 = 1.0 - p1
w = numpy.log(p0/p1) + numpy.log((1.0-beta)/beta)
w *= edgeCounts
obj = nifty.graph.multicut.multicutObjective(g, w)
ilpFactory = obj.multicutIlpFactory()
fusionMove = obj.fusionMoveSettings(mcFactory=ilpFactory)
factory = obj.fusionMoveBasedFactory(fusionMove=fusionMove,
stopIfNoImprovement=200)
solver = ilpFactory.create(obj)
visitor = obj.multicutVerboseVisitor()
with vigra.Timer("optimzie"):
ret = solver.optimize(visitor=visitor)
ret = nifty.tools.makeDense(ret)
resName = '%s_%s.h5'%(outName,"resultSeg")
resultH5File = h5py.File(os.path.join(outFolder,resName),'w')
resultSegDset = resultH5File.create_dataset('data', shape=shape, chunks=(64,)*3, compression='gzip', dtype='uint32')
blocking = nifty.tools.blocking(roiBegin=[0]*3, roiEnd=shape,
blockShape=(256,)*3)
nBlocks = blocking.numberOfBlocks
def f(blockIndex):
# labels halo
block = blocking.getBlock(blockIndex)
b, e = block.begin, block.end
# load files
with lock:
oseg = osegDset[b[0]:e[0], b[1]:e[1], b[2]:e[2]]
relabeld = numpy.take(ret, oseg)
with lock2:
resultSegDset[b[0]:e[0], b[1]:e[1], b[2]:e[2]] = relabeld
#print("relabeld shape",relabeld.shape)
nifty.tools.parallelForEach(iterable=range(nBlocks), f=f,
nWorkers=-1, showBar=True,
size=nBlocks, name="Braaa")
resultH5File.close()
if stages[2] == 1:
if minSegSize is not None:
with vigra.Timer("load the segIn"):
resName = '%s_%s.h5'%(outName,"resultSeg")
resultH5File = h5py.File(os.path.join(outFolder,resName),'r')
segIn = resultH5File['data'][:,:,:]#[0:200,0:200,0:200]
with vigra.Timer("load the pmap"):
# load the pmap
pmap = pmapDset[:,:,:,]#[0:200,0:200,0:200]
with vigra.Timer("make rag"):
rag = nifty.graph.rag.gridRag(segIn)
print(rag)
with vigra.Timer("acc"):
eFeatures, nFeatures = nifty.graph.rag.accumulateMeanAndLength(rag, pmap, [100,100,100],-1)
eMeans = eFeatures[:,0]
eSizes = eFeatures[:,1]
nSizes = nFeatures[:,1]
with vigra.Timer("cluster"):
res = nagglo.sizeLimitClustering(graph=rag, nodeSizes=nSizes, minimumNodeSize=int(minSegSize),
edgeIndicators=eMeans,edgeSizes=eSizes,
sizeRegularizer=0.001, gamma=0.999,
makeDenseLabels=True)
#res = numpy.arange(rag.nodeIdUpperBound + 1)
resName = '%s_%s.h5'%(outName,"outSegSmallRemoved")
resultH5File = h5py.File(os.path.join(outFolder,resName),'w')
resultSegDset = resultH5File.create_dataset('data', shape=shape, chunks=(64,)*3, compression='gzip', dtype='uint32')
with vigra.Timer("makeDense"):
def f(blockIndex):
# labels halo
block = blocking.getBlock(blockIndex)
b, e = block.begin, block.end
oseg = segIn[b[0]:e[0], b[1]:e[1], b[2]:e[2]]
relabeld = numpy.take(res, oseg)
with lock:
resultSegDset[b[0]:e[0], b[1]:e[1], b[2]:e[2]] = relabeld
#print("relabeld shape",relabeld.shape)
nifty.tools.parallelForEach(iterable=range(nBlocks), f=f,
nWorkers=-1, showBar=True,
size=nBlocks, name="write results")
rawH5File.close()
pmapH5File.close()
osegH5File.close()
def getLargeSeeds(raw, pmap, membraneWidth, threshold, rank, visu=False):
with vigra.Timer("ballRankOrderFilter"):
# make mebrane wider with ballRankOrderFilter
r = int(membraneWidth * 0.35 + 0.5)
r = max(r, 1)
widerPmap1 = denoise.ballRankOrderFilter(pmap, radius=r, rank=rank)
widerPmap = denoise.ballRankOrderFilter(widerPmap1,
radius=r,
rank=rank)
widerPmap1 = None
with vigra.Timer("normalize"):
# renormalize
widerPmap -= widerPmap.min()
widerPmap /= widerPmap.max()
with vigra.Timer("binarize"):
# binarize
binaryPmap = numpy.zeros(widerPmap.shape, dtype='uint8')
binaryPmap[pmap > threshold] = 1
#widerPmap = None # save mem
if visu == False:
widerPmap = None
with vigra.Timer("multiBinaryDilation"):
# morphology
# 1) make membrane wider by r
r = int(membraneWidth * 1.2 + 0.5)
r = max(r, 2)
mBinaryPmapA = vigra.filters.multiBinaryDilation(binaryPmap, r)
#binaryPmap = None # save mem
if visu == False:
binaryPmap = None
#with vigra.Timer("multiBinaryErosion"):
# # morphology
# # 1) make membrane smaller by r
# r = int(membraneWidth*0.1 + 0.5)
# r = max(r, 1)
# mBinaryPmapB = vigra.filters.multiBinaryErosion(mBinaryPmapA,r)
# if visu == False:
# mBinaryPmapA = None
with vigra.Timer("labelVolumeWithBackground"):
# get seeds
invertedBinaryPmap = 1 - mBinaryPmapA
if visu == False:
mBinaryPmapB = None
invertedBinaryPmap = numpy.require(invertedBinaryPmap, dtype='uint32')
ccImg = vigra.analysis.labelVolumeWithBackground(invertedBinaryPmap)
if visu == False:
invertedBinaryPmap = None
if visu:
grayData = [
(raw, "raw"),
(pmap, "pmap"),
(widerPmap, "widerPmap"),
(binaryPmap, "binaryPmap"),
(mBinaryPmapA, "mBinaryPmapA"),
#(mBinaryPmapB,"mBinaryPmapB"),
]
segData = [
(ccImg, "seeds"),
#(seg, "seg")
]
skneuro.addHocViewer(grayData, segData)
return ccImg
for fi in mInputs:
name = fi.name()
voxelData = fi().squeeze()
if voxelData.ndim == 3:
voxelData = voxelData[:, :, :, None]
for c in range(voxelData.shape[3]):
vd = voxelData[:, :, :, c]
print "DO ACCUMULATOION :", name
#raise RuntimeError("check for existence here")
with vigra.Timer("accumulate features"):
eFeatures, nFeatures = learning.accumulateFeatures(rag=rag,
volume=vd)
#with vigra.Timer("save edge features"):
vigra.impex.writeHDF5(eFeatures,
featureDir + name + '_c_%d.h5' % c,
'edgeFeatures')
vigra.impex.writeHDF5(nFeatures,
featureDir + name + '_c_%d.h5' % c,
'nodeFeatures')
del eFeatures
del nFeatures
del vd
labPath = ('segMaskOnly.h5', 'data') # labeled image path
# load volume
labels = vigra.impex.readHDF5(*labPath).astype(np.uint32) #[:,:,0:20]
volume = vigra.impex.readHDF5(*imPath) #[:,:,0:20]
volume = volume.astype('float32')
gridGraph = graphs.gridGraph(labels.shape)
rag = graphs.regionAdjacencyGraph(gridGraph, labels)
featureExtractor = graphs.gridRagFeatureExtractor(rag, labels)
accSimpel = featureExtractor.accumulatedFeaturesSimple
with vigra.Timer("acc raw"):
miMa = float(volume.min()), float(volume.max())
accFeatRaw = featureExtractor.accumulatedFeatures(volume, miMa[0], miMa[1])
with vigra.Timer("acc grad"):
grad = vigra.filters.gaussianGradientMagnitude(volume, 1.0)
accFeatGrad = accSimpel(grad)
with vigra.Timer("acc hessian"):
grad = vigra.filters.hessianOfGaussianEigenvalues(volume, 1.0)[:, :, 0]
accFeatHess = accSimpel(grad)
gui = vigra.graphs.TinyEdgeLabelGui(rag=rag,
img=volume,
edgeLabels=accFeatHess[:, 0],
labelMode=False)
gui.startGui()
import vigra
from vigra import graphs
filepath = '12003.jpg'
img = vigra.impex.readImage(filepath).astype('float32')
imgM = img.copy()
sigmaS = 1.0
sigmaB = 5.0
with vigra.Timer("ransk"):
for c in range(3):
print c
imgC = img[:, :, c].squeeze()
imgM[:, :,
c] = vigra.histogram.gaussianRankOrder(imgC,
sigmas=(sigmaS, sigmaS,
sigmaB),
ranks=(0.5, ),
bins=255).squeeze()
vigra.imshow(vigra.taggedView(imgM, 'xyc'))
vigra.show()
import vigra
from vigra import graphs
from vigra import numpy
from vigra import Timer
from vigra import blockwise as bw
numpy.random.seed(42)
# input
shape = (500, 500, 500)
data = numpy.random.rand(*shape).astype('float32')
print "make options object"
options = bw.BlockwiseConvolutionOptions3D()
print type(options)
sigma = 1.0
options.stdDev = (sigma, ) * 3
options.blockShape = (128, ) * 3
print "stddev", options.stdDev
print "call blockwise filter"
with vigra.Timer("AllThread"):
res = bw.gaussianSmooth(data, options)
with vigra.Timer("1thread"):
resRef = vigra.gaussianSmoothing(data, sigma)
print numpy.sum(numpy.abs(res - resRef))
import skneuro
import vigra
from volumina.api import Viewer
from PyQt4.QtGui import QApplication
import skneuro.denoising as dn
p = '/home/tbeier/Desktop/blocks/data_sub_3.h5'
data = vigra.impex.readHDF5(p, 'data')[0:200, :,
0:200].astype('uint8').squeeze()
app = QApplication(sys.argv)
v = Viewer()
v.addGrayscaleLayer(data, name="raw")
with vigra.Timer("get ranks 8*2"):
a = dn.ballRankOrder(data,
radius=12,
takeNth=2,
ranks=(0.01, 0.1, 0.5, 0.9, 0.99),
useHistogram=True,
minVal=0.0,
maxVal=255.0,
nBins=256)
v.addGrayscaleLayer(a, name="0.5 8* 2")
v.setWindowTitle("data")
v.showMaximized()
app.exec_()
shape = [s / 2 for s in shape]
timg = vigra.sampling.resize(img, shape)
# get orientation
tensor = vigra.filters.structureTensor(timg, 1.0, 2.0)
eigenrep = vigra.filters.tensorEigenRepresentation2D(tensor)
# print get oriented repulsive edges
edgePmap = eigenrep[:, :, 0].copy()
edgePmap -= edgePmap.min()
edgePmap /= edgePmap.max()
graph = agraph.gridGraph(shape)
model = agraph.liftedMcModel(graph)
with vigra.Timer("add long range edges"):
agraph.addLongRangeEdges(model, edgePmap, 0.5, 2, 5)
settings = agraph.settingsParallelLiftedMc(model)
solver = agraph.parallelLiftedMc(model, settings)
out = solver.run()
nodeLabels = model.edgeLabelsToNodeLabels(out)
nodeLabels = nodeLabels.reshape(shape)
vigra.imshow(nodeLabels)
vigra.show()
import vigra
from vigra import numpy
import h5py
import skneuro
from skneuro import oversegmentation as oseg
with vigra.Timer("load pmap"):
path = "/home/tbeier/Desktop/blocks/data_sub_3.h5"
f = h5py.File(path, 'r')
raw = f['data'][0:200, 0:200, 0:200]
f.close()
def normalVol(shape, center, scale):
size = numpy.prod(shape)
a = numpy.random.normal(center, scale, size).reshape(shape)
a = vigra.taggedView(a, 'xyz')
return a
def augmentGaussian(data, lAdd, gAdd, gMult):
"""
lAdd : sigma of local additive gaussian noise
gAdd : sigma of global additive gaussian noise
gMult : sigma of global multiplicative guasian noise
"""
data = vigra.taggedView(data, 'xyz')
shape = data.shape
# local and global additive and multiplicative
# gaussian noise
