def predict(self, graphData, rfPath, stopProbs=[0.5], damping=0.05):
print "load random forest"
rf = vigra.learning.RandomForest(rfPath,'rf')
mg = graphs.mergeGraph(graphData.rag)
df = graphs.NeuroDynamicFeatures(graphData.rag, mg)
# assign features
df.assignEdgeCues(graphData.eX)
df.assignNodeCues(graphData.nX)
df.assignEdgeSizes(graphData.eSize)
df.assignNodeSizes(graphData.nSize)
#NOT assign labels
#df.assignLabels(eY)
# register callbacks
df.registerCallbacks()
stopProbsArray = numpy.array(stopProbs,dtype=numpy.float32)
labelsArray = df.predict(rf=rf, stopProbs=stopProbsArray, damping=float(damping))
rLabels = []
for i,sp in enumerate(stopProbs):
print "sp",sp
rLabels.append(labelsArray[i,:])
return rLabels
def predict(self, graphData, rfPath, stopProbs=[0.5], damping=0.05):
print "load random forest"
rf = vigra.learning.RandomForest(rfPath, 'rf')
mg = graphs.mergeGraph(graphData.rag)
df = graphs.NeuroDynamicFeatures(graphData.rag, mg)
# assign features
df.assignEdgeCues(graphData.eX)
df.assignNodeCues(graphData.nX)
df.assignEdgeSizes(graphData.eSize)
df.assignNodeSizes(graphData.nSize)
#NOT assign labels
#df.assignLabels(eY)
# register callbacks
df.registerCallbacks()
stopProbsArray = numpy.array(stopProbs, dtype=numpy.float32)
labelsArray = df.predict(rf=rf,
stopProbs=stopProbsArray,
damping=float(damping))
rLabels = []
for i, sp in enumerate(stopProbs):
print "sp", sp
rLabels.append(labelsArray[i, :])
return rLabels
def getNewRf(self, graphData, eY, rfPath, rfPathNew):
print "load random forest"
rf = vigra.learning.RandomForest(rfPath,'rf')
X = vigra.impex.readHDF5(rfPath, 'X')
Y = vigra.impex.readHDF5(rfPath, 'Y')
mg = graphs.mergeGraph(graphData.rag)
df = graphs.NeuroDynamicFeatures(graphData.rag, mg)
# assign features
df.assignEdgeCues(graphData.eX)
df.assignNodeCues(graphData.nX)
df.assignEdgeSizes(graphData.eSize)
df.assignNodeSizes(graphData.nSize)
# assign labels
df.assignLabels(eY)
# register callbacks
df.registerCallbacks()
ret = df.getNewFeatureByClustering(rf=rf,noiseMagnitude=self.noise)
if len(ret) == 1 :
return "done"
nN, nX, nY = ret
print "collected",nN,"new training instances"
nX = nX[0:nN,:]
nY = nY[0:nN,:]
X = numpy.concatenate([X,nX], axis=0)
Y = numpy.concatenate([Y,nY], axis=0)
print "train random forest"
rf = self.getFreshRf(treeCount=1000)
oob = rf.learnRF(X, Y)
print oob
print "save random forest"
rf.writeHDF5(rfPathNew, 'rf')
print "save out of bag"
oobA = numpy.array([oob],dtype=numpy.float32)
vigra.impex.writeHDF5(oobA, rfPathNew, 'oob')
print "save features and labels "
vigra.impex.writeHDF5(X, rfPathNew, 'X')
vigra.impex.writeHDF5(Y, rfPathNew, 'Y')
#sys.exit(0)
return "not_done"
def cluster(wardness):
edgeLengthsNew = numpy.concatenate(
[eLen, numpy.zeros(nAdditionalEdges)]).astype('float32')
edgeIndicatorNew = numpy.concatenate(
[edgeIndicator, numpy.zeros(nAdditionalEdges)]).astype('float32')
nodeSizes = nodeSizes_.copy()
nodeLabels = vgraph.graphMap(originalGraph, 'node', dtype='uint32')
nodeFeatures = vgraph.graphMap(liftedGraph, 'node', addChannelDim=True)
nodeFeatures[:] = 0
outWeight = vgraph.graphMap(liftedGraph,
item='edge',
dtype=numpy.float32)
mg = vgraph.mergeGraph(liftedGraph)
clusterOp = vgraph.minEdgeWeightNodeDist(mg,
edgeWeights=edgeIndicatorNew,
edgeLengths=edgeLengthsNew,
nodeFeatures=nodeFeatures,
nodeSizes=nodeSizes,
nodeLabels=nodeLabels,
beta=0.5,
metric='l1',
wardness=wardness,
outWeight=outWeight)
clusterOp.setLiftedEdges(whereLifted)
hc = vgraph.hierarchicalClustering(clusterOp,
nodeNumStopCond=1,
buildMergeTreeEncoding=False)
hc.cluster()
assert mg.edgeNum == 0, str(mg.edgeNum)
# FIXME I am disabling these checks for now, but will need to investigate this further
# They can fail because with defects and seg mask we can get unconnected pieces in the graph
# if we have completely defected slcies, we get a non-connected merge graph
# TODO I don't know if this is a problem, if it is, we could first remove them
# and then add dummy values later
#if not with_defects:
# assert mg.nodeNum == 1, str(mg.nodeNum)
#else:
# # TODO test hypothesis
# assert mg.nodeNum == len(ds.defect_slice_list) + 1, "%i, %i" % (mg.nodeNum, len(ds.defect_slice_list) + 1)
tweight = edgeIndicatorNew.copy()
hc.ucmTransform(tweight)
whereInLifted = liftedGraph.findEdges(extraUV)
assert whereInLifted.min() >= 0
feat = tweight[whereInLifted]
assert feat.shape[0] == extraUV.shape[0]
return feat[:, None]
def getNewRf(self, graphData, eY, rfPath, rfPathNew):
print "load random forest"
rf = vigra.learning.RandomForest(rfPath, 'rf')
X = vigra.impex.readHDF5(rfPath, 'X')
Y = vigra.impex.readHDF5(rfPath, 'Y')
mg = graphs.mergeGraph(graphData.rag)
df = graphs.NeuroDynamicFeatures(graphData.rag, mg)
# assign features
df.assignEdgeCues(graphData.eX)
df.assignNodeCues(graphData.nX)
df.assignEdgeSizes(graphData.eSize)
df.assignNodeSizes(graphData.nSize)
# assign labels
df.assignLabels(eY)
# register callbacks
df.registerCallbacks()
ret = df.getNewFeatureByClustering(rf=rf, noiseMagnitude=self.noise)
if len(ret) == 1:
return "done"
nN, nX, nY = ret
print "collected", nN, "new training instances"
nX = nX[0:nN, :]
nY = nY[0:nN, :]
X = numpy.concatenate([X, nX], axis=0)
Y = numpy.concatenate([Y, nY], axis=0)
print "train random forest"
rf = self.getFreshRf(treeCount=1000)
oob = rf.learnRF(X, Y)
print oob
print "save random forest"
rf.writeHDF5(rfPathNew, 'rf')
print "save out of bag"
oobA = numpy.array([oob], dtype=numpy.float32)
vigra.impex.writeHDF5(oobA, rfPathNew, 'oob')
print "save features and labels "
vigra.impex.writeHDF5(X, rfPathNew, 'X')
vigra.impex.writeHDF5(Y, rfPathNew, 'Y')
#sys.exit(0)
return "not_done"
def initialTraining(self, graphData, eY, rfPath):
# do the inital training
mg = graphs.mergeGraph(graphData.rag)
df = graphs.NeuroDynamicFeatures(graphData.rag, mg)
# assign features
df.assignEdgeCues(graphData.eX)
df.assignNodeCues(graphData.nX)
df.assignEdgeSizes(graphData.eSize)
df.assignNodeSizes(graphData.nSize)
# assign labels
df.assignLabels(eY)
# register callbacks
df.registerCallbacks()
# the training set from level 0
print "compute inital training set"
features, labels = df.computeInitalTrainingSet()
print "features/labels.shape", features.shape, labels.shape
print "train random forest"
rf = self.getFreshRf(treeCount=100)
oob = rf.learnRF(features, labels)
print "OOB", oob
print "save random forest"
rf.writeHDF5(rfPath, 'rf')
print "save out of bag"
oobA = numpy.array([oob],dtype=numpy.float32)
vigra.impex.writeHDF5(oobA, rfPath, 'oob')
print "save features and labels "
vigra.impex.writeHDF5(features, rfPath, 'X')
vigra.impex.writeHDF5(labels, rfPath, 'Y')
def initialTraining(self, graphData, eY, rfPath):
# do the inital training
mg = graphs.mergeGraph(graphData.rag)
df = graphs.NeuroDynamicFeatures(graphData.rag, mg)
# assign features
df.assignEdgeCues(graphData.eX)
df.assignNodeCues(graphData.nX)
df.assignEdgeSizes(graphData.eSize)
df.assignNodeSizes(graphData.nSize)
# assign labels
df.assignLabels(eY)
# register callbacks
df.registerCallbacks()
# the training set from level 0
print "compute inital training set"
features, labels = df.computeInitalTrainingSet()
print "features/labels.shape", features.shape, labels.shape
print "train random forest"
rf = self.getFreshRf(treeCount=100)
oob = rf.learnRF(features, labels)
print "OOB", oob
print "save random forest"
rf.writeHDF5(rfPath, 'rf')
print "save out of bag"
oobA = numpy.array([oob], dtype=numpy.float32)
vigra.impex.writeHDF5(oobA, rfPath, 'oob')
print "save features and labels "
vigra.impex.writeHDF5(features, rfPath, 'X')
vigra.impex.writeHDF5(labels, rfPath, 'Y')
print "load sizes" edgeSize = vigra.impex.readHDF5(dopt['ragL1EdgeSize'], 'data') nodeSize = vigra.impex.readHDF5(dopt['ragL1NodeSize'], 'data') print "load label file" edgeLabels = vigra.impex.readHDF5(dopt['ragL1EdgeGt'], 'data') print "load rag" rag = graphs.loadGridRagHDF5(dopt['ragL1'], 'data') gridGraph = rag.baseGraph print "get merge graph" mg = graphs.mergeGraph(rag) df = graphs.NeuroDynamicFeatures(rag, mg) # assign features df.assignEdgeCues(edgeCues) df.assignNodeCues(nodeCues) df.assignEdgeSizes(edgeSize) df.assignNodeSizes(nodeSize) # assign labels df.assignLabels(edgeLabels) # register callbacks df.registerCallbacks()
# load image and convert to LAB
img = vigra.impex.readImage(filepath)
# get super-pixels with slic on LAB image
imgLab = vigra.colors.transform_RGB2Lab(img)
labels, nseg = vigra.analysis.slicSuperpixels(imgLab, slicWeight, superpixelDiameter)
labels = vigra.analysis.labelImage(labels)
gridGraph = graphs.gridGraph(img.shape[0:2])
rag = graphs.regionAdjacencyGraph(gridGraph, labels)
# get the merge graph
mg = graphs.mergeGraph(rag)
# do n runs where we erase k edges in each run
n = 3
k = 200
for r in range(n):
erased = 0
while erased < k:
# get a random edge
randEdgeId = numpy.random.randint(rag.edgeNum)
print("random edge:", randEdgeId)
# edge could be gone
