def test_key_to_coordinate(self):
f = opengm.SparseFunction([2, 3, 4], 0)
c = numpy.ones(3, dtype=numpy.uint64)
for key, cTrue in enumerate(opengm.shapeWalker(f.shape)):
f.keyToCoordinate(key, c)
for ct, cOwn in zip(cTrue, c):
assert ct == cOwn
def test_dense_assignment(self):
f = opengm.SparseFunction()
fDense = numpy.zeros([3, 4])
fDense[0, 1] = 1
fDense[0, 2] = 2
f.assignDense(fDense, 0)
assert f.dimension == 2
assert f.shape[0] == 3
assert f.shape[1] == 4
assert f[[0, 0]] == 0
assert f[[0, 1]] == 1
assert f[[0, 2]] == 2
for c in opengm.shapeWalker(f.shape):
assert f[c] == fDense[c[0], c[1]]
assert len(f.container) == 2
def genericSolverCheck(solverClass, params, gms, semiRings,checkPartial=False,checkMarginals=False,testPythonVisitor=True,testLpInterface=False):
class PyCallback(object):
def __init__(self):
self.inBegin=False
self.inEnd=False
self.inVisit=False
def begin(self,inference):
self.inBegin=True
def end(self,inference):
self.inEnd=True
def visit(self,inference):
self.inVisit=True
for operator, accumulator in semiRings:
for gmGen in gms:
gm = gmGen[operator]
for param in params:
# start inference
solver = solverClass(gm=gm, accumulator=accumulator, parameter=param)
solver2 = solverClass(gm=gm, accumulator=accumulator, parameter=param)
if (testLpInterface==True):
c=0
for vi in range(gm.numberOfVariables):
nl = gm.numberOfLabels(vi)
for l in range(nl):
assert c==solver.lpNodeVariableIndex(vi,l)
c+=1
cv=c
for fi in range(gm.numberOfFactors):
if gm[fi].numberOfVariables>1:
s = gm[fi].size
for l in range(nl):
assert solver.lpFactorVariableIndex(fi,s)>0 or cv==0
sw = opengm.shapeWalker(gm[fi].shape)
for c in sw:
assert solver.lpFactorVariableIndex(fi,c)>0 or cv==0
solver2.addConstraint(lpVariableIndices=[0,1],coefficients=[1,1],lowerBound=0.0,upperBound=1.0)
solver2.addConstraints(lpVariableIndices=[ [0,1],[0,2] ,[1,2]],coefficients=[ [1,1],[2,2],[1,2]],lowerBounds=[0,0,0],upperBounds=[1,1,1])
solver.infer()
arg = solver.arg() # no used?
value = solver.value()
bound = solver.bound()
if testPythonVisitor==True:
solver = solverClass(gm=gm, accumulator=accumulator, parameter=param)
callback=PyCallback()
pvisitor=solver.pythonVisitor(callback,1)
solver.infer(pvisitor)
assert callback.inBegin == True
assert callback.inEnd == True
if checkPartial :
pOptimal = solver.partialOptimality()
assert len(pOptimal)==gm.numberOfVariables
#assert len(numpy.where(pOptimal==True)[0]) == gm.numberOfVariables
if checkMarginals :
visRange=numpy.arange(gm.numberOfVariables)
marginal = solver.marginals(visRange)
assert marginal.shape[0]==gm.numberOfVariables
assert marginal.shape[1]==gm.numberOfLabels(0)
fis1 = gm.factorSubset(order=1).factorIndices
fis2 = gm.factorSubset(order=2).factorIndices
assert len(fis1)!=0
assert len(fis2)!=0
factorMarginal1 = solver.factorMarginals(fis1)
assert factorMarginal1.ndim==2
assert factorMarginal1.shape[0]==len(fis1)
assert factorMarginal1.shape[1]==gm.numberOfLabels(0)
factorMarginal2 = solver.factorMarginals(fis2)
assert factorMarginal2.ndim==3
assert factorMarginal2.shape[0]==len(fis2)
assert factorMarginal2.shape[1]==gm.numberOfLabels(0)
assert factorMarginal2.shape[2]==gm.numberOfLabels(0)
def genericSolverCheck(solverClass, params, gms, semiRings,checkPartial=False,checkMarginals=False,testPythonVisitor=True,testLpInterface=False):
class PyCallback(object):
def __init__(self):
self.inBegin=False
self.inEnd=False
self.inVisit=False
def begin(self,inference):
self.inBegin=True
def end(self,inference):
self.inEnd=True
def visit(self,inference):
self.inVisit=True
for operator, accumulator in semiRings:
for gmGen in gms:
gm = gmGen[operator]
for param in params:
# start inference
solver = solverClass(gm=gm, accumulator=accumulator, parameter=param)
solver2 = solverClass(gm=gm, accumulator=accumulator, parameter=param)
if (testLpInterface==True):
c=0
for vi in range(gm.numberOfVariables):
nl = gm.numberOfLabels(vi)
for l in range(nl):
assert c==solver.lpNodeVariableIndex(vi,l)
c+=1
cv=c
for fi in range(gm.numberOfFactors):
if gm[fi].numberOfVariables>1:
s = gm[fi].size
for l in range(nl):
assert solver.lpFactorVariableIndex(fi,s)>0 or cv==0
sw = opengm.shapeWalker(gm[fi].shape)
for c in sw:
assert solver.lpFactorVariableIndex(fi,c)>0 or cv==0
solver2.addConstraint(lpVariableIndices=[0,1],coefficients=[1,1],lowerBound=0.0,upperBound=1.0)
solver2.addConstraints(lpVariableIndices=[ [0,1],[0,2] ,[1,2]],coefficients=[ [1,1],[2,2],[1,2]],lowerBounds=[0,0,0],upperBounds=[1,1,1])
solver.infer()
arg = solver.arg() # no used?
value = solver.value()
bound = solver.bound()
if testPythonVisitor==True:
solver = solverClass(gm=gm, accumulator=accumulator, parameter=param)
callback=PyCallback()
pvisitor=solver.pythonVisitor(callback,1)
solver.infer(pvisitor)
assert callback.inBegin == True
assert callback.inEnd == True
if checkPartial :
pOptimal = solver.partialOptimality()
assert len(pOptimal)==gm.numberOfVariables
#assert len(numpy.where(pOptimal==True)[0]) == gm.numberOfVariables
if checkMarginals :
visRange=numpy.arange(gm.numberOfVariables)
marginal = solver.marginals(visRange)
assert marginal.shape[0]==gm.numberOfVariables
assert marginal.shape[1]==gm.numberOfLabels(0)
fis1 = gm.factorSubset(order=1).factorIndices
fis2 = gm.factorSubset(order=2).factorIndices
assert len(fis1)!=0
assert len(fis2)!=0
factorMarginal1 = solver.factorMarginals(fis1)
assert factorMarginal1.ndim==2
assert factorMarginal1.shape[0]==len(fis1)
assert factorMarginal1.shape[1]==gm.numberOfLabels(0)
factorMarginal2 = solver.factorMarginals(fis2)
assert factorMarginal2.ndim==3
assert factorMarginal2.shape[0]==len(fis2)
assert factorMarginal2.shape[1]==gm.numberOfLabels(0)
assert factorMarginal2.shape[2]==gm.numberOfLabels(0)
import opengm
import numpy
# some graphical model
# -with 3 variables with 4 labels.
# -with 2 random 2-order functions
# -connected to 2 factors
gm = opengm.gm([4] * 3)
f = numpy.random.rand(4, 4).astype(numpy.float32)
gm.addFactor(gm.addFunction(f), [0, 1])
f = numpy.random.rand(4, 4).astype(numpy.float32)
gm.addFactor(gm.addFunction(f), [1, 2])
# iterate over all factors of the graphical model
for factor in gm.factors():
# iterate over all labelings with a "shape walker"
for coord in opengm.shapeWalker(f.shape):
print " f[", coord, "]=", factor[coord]
def myFunc( labels):
p=1.0
s=0
for l in labels:
p*=l
s+=l
return p+s
pf=opengm.PythonFunction(myFunc,[2,2])
print pf.shape
print pf[0,0]
print pf[1,1]
for c in opengm.shapeWalker(pf.shape):
print c," ",pf[c]
unaries=numpy.random.rand(5 , 5,2)
potts=opengm.PottsFunction([2,2],0.0,0.1)
gm=opengm.grid2d2Order(unaries=unaries,regularizer=potts)
viewFunction=opengm.modelViewFunction(gm[0])
for c in opengm.shapeWalker(viewFunction.shape):
print c," ",viewFunction[c]
viewFunction=opengm.modelViewFunction(gm[25])
print viewFunction.shape