def Matching(cf):
if cf.compute.useCUDA and cf.compute.gpuID is not None:
ca.SetCUDADevice(cf.compute.gpuID)
# prepare output directory
common.Mkdir_p(os.path.dirname(cf.io.outputPrefix))
# Output loaded config
if cf.io.outputPrefix is not None:
cfstr = Config.ConfigToYAML(MatchingConfigSpec, cf)
with open(cf.io.outputPrefix + "parsedconfig.yaml", "w") as f:
f.write(cfstr)
mType = ca.MEM_DEVICE if cf.compute.useCUDA else ca.MEM_HOST
I0 = common.LoadITKImage(cf.study.I0, mType)
I1 = common.LoadITKImage(cf.study.I1, mType)
#ca.DivC_I(I0,255.0)
#ca.DivC_I(I1,255.0)
grid = I0.grid()
ca.ThreadMemoryManager.init(grid, mType, 1)
#common.DebugHere()
# TODO: need to work on these
t = [x*1./cf.optim.nTimeSteps for x in range(cf.optim.nTimeSteps+1)]
checkpointinds = range(1,len(t))
checkpointstates = [(ca.Field3D(grid,mType),ca.Field3D(grid,mType)) for idx in checkpointinds]
p = MatchingVariables(I0,I1, cf.vectormomentum.sigma, t,checkpointinds, checkpointstates, cf.vectormomentum.diffOpParams[0], cf.vectormomentum.diffOpParams[1], cf.vectormomentum.diffOpParams[2], cf.optim.Niter, cf.optim.stepSize, cf.optim.maxPert, cf.optim.nTimeSteps, integMethod = cf.optim.integMethod, optMethod=cf.optim.method, nInv=cf.optim.NIterForInverse,plotEvery=cf.io.plotEvery, plotSlice = cf.io.plotSlice, quiverEvery = cf.io.quiverEvery, outputPrefix = cf.io.outputPrefix)
RunMatching(p)
# write output
if cf.io.outputPrefix is not None:
# reset all variables by shooting once, may have been overwritten
CAvmCommon.IntegrateGeodesic(p.m0,p.t,p.diffOp,\
p.m, p.g, p.ginv,\
p.scratchV1, p.scratchV2,p. scratchV3,\
p.checkpointstates, p.checkpointinds,\
Ninv=p.nInv, integMethod = p.integMethod)
common.SaveITKField(p.m0, cf.io.outputPrefix+"m0.mhd")
common.SaveITKField(p.ginv, cf.io.outputPrefix+"phiinv.mhd")
common.SaveITKField(p.g, cf.io.outputPrefix+"phi.mhd")
def BuildAtlas(cf):
"""Worker for running Atlas construction on a subset of individuals.
Runs Atlas on this subset sequentially. The variations retuned are
summed up to get update for all individuals
"""
localRank = Compute.GetMPIInfo()['local_rank']
rank = Compute.GetMPIInfo()['rank']
# prepare output directory
common.Mkdir_p(os.path.dirname(cf.io.outputPrefix))
# just one reporter process on each node
isReporter = rank == 0
cf.study.numSubjects = len(cf.study.subjectImages)
if isReporter:
# Output loaded config
if cf.io.outputPrefix is not None:
cfstr = Config.ConfigToYAML(AtlasConfigSpec, cf)
with open(cf.io.outputPrefix + "parsedconfig.yaml", "w") as f:
f.write(cfstr)
#common.DebugHere()
# if MPI check if processes are greater than number of subjects. it is okay if there are more subjects than processes
if cf.compute.useMPI and (cf.study.numSubjects < cf.compute.numProcesses):
raise Exception("Please don't use more processes " +
"than total number of individuals")
# subdivide data, create subsets for this thread to work on
nodeSubjectIds = cf.study.subjectIds[rank::cf.compute.numProcesses]
nodeImages = cf.study.subjectImages[rank::cf.compute.numProcesses]
nodeWeights = cf.study.subjectWeights[rank::cf.compute.numProcesses]
numLocalSubjects = len(nodeImages)
print 'rank:', rank, ', localRank:', localRank, ', nodeImages:', nodeImages, ', nodeWeights:', nodeWeights
# mem type is determined by whether or not we're using CUDA
mType = ca.MEM_DEVICE if cf.compute.useCUDA else ca.MEM_HOST
# load data in memory
# load intercepts
J_array = [
common.LoadITKImage(f, mType) if isinstance(f, str) else f
for f in nodeImages
]
# get imGrid from data
imGrid = J_array[0].grid()
# atlas image
atlas = ca.Image3D(imGrid, mType)
# allocate memory to store only the initial momenta for each individual in this thread
m_array = [ca.Field3D(imGrid, mType) for i in range(numLocalSubjects)]
# allocate only one copy of scratch memory to be reused for each local individual in this thread in loop
p = WarpVariables(imGrid,
mType,
cf.vectormomentum.diffOpParams[0],
cf.vectormomentum.diffOpParams[1],
cf.vectormomentum.diffOpParams[2],
cf.optim.NIterForInverse,
cf.vectormomentum.sigma,
cf.optim.stepSize,
integMethod=cf.optim.integMethod)
# memory to accumulate numerators and denominators for atlas from
# local individuals which will be summed across MPI threads
sumSplatI = ca.Image3D(imGrid, mType)
sumJac = ca.Image3D(imGrid, mType)
# start up the memory manager for scratch variables
ca.ThreadMemoryManager.init(imGrid, mType, 0)
# need some host memory in np array format for MPI reductions
if cf.compute.useMPI:
mpiImageBuff = None if mType == ca.MEM_HOST else ca.Image3D(
imGrid, ca.MEM_HOST)
t = [
x * 1. / (cf.optim.nTimeSteps) for x in range(cf.optim.nTimeSteps + 1)
]
cpinds = range(1, len(t))
msmtinds = [
len(t) - 2
] # since t=0 is not in cpinds, thats just identity deformation so not checkpointed
cpstates = [(ca.Field3D(imGrid, mType), ca.Field3D(imGrid, mType))
for idx in cpinds]
gradAtMsmts = [ca.Image3D(imGrid, mType) for idx in msmtinds]
EnergyHistory = []
# TODO: better initializations
# initialize atlas image with zeros.
ca.SetMem(atlas, 0.0)
# initialize momenta with zeros
for m0_individual in m_array:
ca.SetMem(m0_individual, 0.0)
'''
# initial template image
ca.SetMem(groupState.I0, 0.0)
tmp = ca.ManagedImage3D(imGrid, mType)
for tdisc in tdiscGroup:
if tdisc.J is not None:
ca.Copy(tmp, tdisc.J)
groupState.I0 += tmp
del tmp
if cf.compute.useMPI:
Compute.Reduce(groupState.I0, mpiImageBuff)
# divide by total num subjects
groupState.I0 /= cf.study.numSubjects
'''
# preprocessinput
# assign atlas reference to p.I0. This reference will not change.
p.I0 = atlas
# run the loop
for it in range(cf.optim.Niter):
# run one iteration of warp for each individual and update
# their own initial momenta and also accumulate SplatI and Jac
ca.SetMem(sumSplatI, 0.0)
ca.SetMem(sumJac, 0.0)
TotalVEnergy = np.array([0.0])
TotalIEnergy = np.array([0.0])
for itsub in range(numLocalSubjects):
# initializations for this subject, this only assigns
# reference to image variables
p.m0 = m_array[itsub]
Imsmts = [J_array[itsub]]
# run warp iteration
VEnergy, IEnergy = RunWarpIteration(nodeSubjectIds[itsub], cf, p,
t, Imsmts, cpinds, cpstates,
msmtinds, gradAtMsmts, it)
# gather relevant results
ca.Add_I(sumSplatI, p.sumSplatI)
ca.Add_I(sumJac, p.sumJac)
TotalVEnergy[0] += VEnergy
TotalIEnergy[0] += IEnergy
# if there are multiple nodes we'll need to sum across processes now
if cf.compute.useMPI:
# do an MPI sum
Compute.Reduce(sumSplatI, mpiImageBuff)
Compute.Reduce(sumJac, mpiImageBuff)
# also sum up energies of other nodes
mpi4py.MPI.COMM_WORLD.Allreduce(mpi4py.MPI.IN_PLACE,
TotalVEnergy,
op=mpi4py.MPI.SUM)
mpi4py.MPI.COMM_WORLD.Allreduce(mpi4py.MPI.IN_PLACE,
TotalIEnergy,
op=mpi4py.MPI.SUM)
EnergyHistory.append([TotalVEnergy[0], TotalIEnergy[0]])
# now divide to get the new atlas image
ca.Div(atlas, sumSplatI, sumJac)
# keep track of energy in this iteration
if isReporter and cf.io.plotEvery > 0 and ((
(it + 1) % cf.io.plotEvery == 0) or (it == cf.optim.Niter - 1)):
# plots
AtlasPlots(cf, p, atlas, m_array, EnergyHistory)
if isReporter:
# print out energy
(VEnergy, IEnergy) = EnergyHistory[-1]
print "Iter", it, "of", cf.optim.Niter, ":", VEnergy + IEnergy, '(Total) = ', VEnergy, '(Vector) + ', IEnergy, '(Image)'
# write output images and fields
AtlasWriteOutput(cf, atlas, m_array, nodeSubjectIds, isReporter)
def MatchingImageMomenta(cf):
"""Runs matching for image momenta pair."""
if cf.compute.useCUDA and cf.compute.gpuID is not None:
ca.SetCUDADevice(cf.compute.gpuID)
common.DebugHere()
# prepare output directory
common.Mkdir_p(os.path.dirname(cf.io.outputPrefix))
# Output loaded config
if cf.io.outputPrefix is not None:
cfstr = Config.ConfigToYAML(MatchingImageMomentaConfigSpec, cf)
with open(cf.io.outputPrefix + "parsedconfig.yaml", "w") as f:
f.write(cfstr)
# mem type is determined by whether or not we're using CUDA
mType = ca.MEM_DEVICE if cf.compute.useCUDA else ca.MEM_HOST
# load data in memory
I0 = common.LoadITKImage(cf.study.I, mType)
m0 = common.LoadITKField(cf.study.m, mType)
J1 = common.LoadITKImage(cf.study.J, mType)
n1 = common.LoadITKField(cf.study.n, mType)
# get imGrid from data
imGrid = I0.grid()
# create time array with checkpointing info for this geodesic to be estimated
(s, scratchInd,
rCpinds) = CAvmHGM.HGMSetUpTimeArray(cf.optim.nTimeSteps, [1.0], 0.001)
tDiscGeodesic = CAvmHGMCommon.HGMSetupTimeDiscretizationResidual(
s, rCpinds, imGrid, mType)
# create the state variable for geodesic that is going to hold all info
p0 = ca.Field3D(imGrid, mType)
geodesicState = CAvmHGMCommon.HGMResidualState(
I0,
p0,
imGrid,
mType,
cf.vectormomentum.diffOpParams[0],
cf.vectormomentum.diffOpParams[1],
cf.vectormomentum.diffOpParams[2],
s,
cf.optim.NIterForInverse,
1.0,
cf.vectormomentum.sigmaM,
cf.vectormomentum.sigmaI,
cf.optim.stepSize,
integMethod=cf.optim.integMethod)
# initialize with zero
ca.SetMem(geodesicState.p0, 0.0)
# start up the memory manager for scratch variables
ca.ThreadMemoryManager.init(imGrid, mType, 0)
EnergyHistory = []
# run the loop
for it in range(cf.optim.Niter):
# shoot the geodesic forward
CAvmHGMCommon.HGMIntegrateGeodesic(geodesicState.p0, geodesicState.s,
geodesicState.diffOp,
geodesicState.p, geodesicState.rho,
geodesicState.rhoinv, tDiscGeodesic,
geodesicState.Ninv,
geodesicState.integMethod)
# integrate the geodesic backward
CAvmHGMCommon.HGMIntegrateAdjointsResidual(geodesicState,
tDiscGeodesic, m0, J1, n1)
# TODO: verify it should just be log map/simple image matching when sigmaM=\infty
# gradient descent step for geodesic.p0
CAvmHGMCommon.HGMTakeGradientStepResidual(geodesicState)
# compute and print energy
(VEnergy, IEnergy,
MEnergy) = MatchingImageMomentaComputeEnergy(geodesicState, m0, J1,
n1)
EnergyHistory.append(
[VEnergy + IEnergy + MEnergy, VEnergy, IEnergy, MEnergy])
print "Iter", it, "of", cf.optim.Niter, ":", VEnergy + IEnergy + MEnergy, '(Total) = ', VEnergy, '(Vector) + ', IEnergy, '(Image Match) + ', MEnergy, '(Momenta Match)'
# plots
if cf.io.plotEvery > 0 and (((it + 1) % cf.io.plotEvery == 0) or
(it == cf.optim.Niter - 1)):
MatchingImageMomentaPlots(cf,
geodesicState,
tDiscGeodesic,
EnergyHistory,
m0,
J1,
n1,
writeOutput=True)
# write output
MatchingImageMomentaWriteOuput(cf, geodesicState)
def GeodesicShooting(cf):
# prepare output directory
common.Mkdir_p(os.path.dirname(cf.io.outputPrefix))
# Output loaded config
if cf.io.outputPrefix is not None:
cfstr = Config.ConfigToYAML(GeodesicShootingConfigSpec, cf)
with open(cf.io.outputPrefix + "parsedconfig.yaml", "w") as f:
f.write(cfstr)
mType = ca.MEM_DEVICE if cf.useCUDA else ca.MEM_HOST
#common.DebugHere()
I0 = common.LoadITKImage(cf.study.I0, mType)
m0 = common.LoadITKField(cf.study.m0, mType)
grid = I0.grid()
ca.ThreadMemoryManager.init(grid, mType, 1)
# set up diffOp
if mType == ca.MEM_HOST:
diffOp = ca.FluidKernelFFTCPU()
else:
diffOp = ca.FluidKernelFFTGPU()
diffOp.setAlpha(cf.diffOpParams[0])
diffOp.setBeta(cf.diffOpParams[1])
diffOp.setGamma(cf.diffOpParams[2])
diffOp.setGrid(grid)
g = ca.Field3D(grid, mType)
ginv = ca.Field3D(grid, mType)
mt = ca.Field3D(grid, mType)
It = ca.Image3D(grid, mType)
t = [
x * 1. / cf.integration.nTimeSteps
for x in range(cf.integration.nTimeSteps + 1)
]
checkpointinds = range(1, len(t))
checkpointstates = [(ca.Field3D(grid, mType), ca.Field3D(grid, mType))
for idx in checkpointinds]
scratchV1 = ca.Field3D(grid, mType)
scratchV2 = ca.Field3D(grid, mType)
scratchV3 = ca.Field3D(grid, mType)
# scale momenta to shoot
cf.study.scaleMomenta = float(cf.study.scaleMomenta)
if abs(cf.study.scaleMomenta) > 0.000000:
ca.MulC_I(m0, float(cf.study.scaleMomenta))
CAvmCommon.IntegrateGeodesic(m0,t,diffOp, mt, g, ginv,\
scratchV1,scratchV2,scratchV3,\
keepstates=checkpointstates,keepinds=checkpointinds,
Ninv=cf.integration.NIterForInverse, integMethod = cf.integration.integMethod)
else:
ca.Copy(It, I0)
ca.Copy(mt, m0)
ca.SetToIdentity(ginv)
ca.SetToIdentity(g)
# write output
if cf.io.outputPrefix is not None:
# scale back shotmomenta before writing
if abs(cf.study.scaleMomenta) > 0.000000:
ca.ApplyH(It, I0, ginv)
ca.CoAd(mt, ginv, m0)
ca.DivC_I(mt, float(cf.study.scaleMomenta))
common.SaveITKImage(It, cf.io.outputPrefix + "I1.mhd")
common.SaveITKField(mt, cf.io.outputPrefix + "m1.mhd")
common.SaveITKField(ginv, cf.io.outputPrefix + "phiinv.mhd")
common.SaveITKField(g, cf.io.outputPrefix + "phi.mhd")
GeodesicShootingPlots(g, ginv, I0, It, cf)
if cf.io.saveFrames:
SaveFrames(checkpointstates, checkpointinds, I0, It, m0, mt, cf)
def BuildHGM(cf):
"""Worker for running Hierarchical Geodesic Model (HGM)
n for group geodesic estimation on a subset of individuals.
Runs HGM on this subset sequentially. The variations retuned
are summed up to get update for all individuals"""
size = Compute.GetMPIInfo()['size']
rank = Compute.GetMPIInfo()['rank']
name = Compute.GetMPIInfo()['name']
localRank = Compute.GetMPIInfo()['local_rank']
nodename = socket.gethostname()
# prepare output directory
common.Mkdir_p(os.path.dirname(cf.io.outputPrefix))
# just one reporter process on each node
isReporter = rank == 0
cf.study.numSubjects = len(cf.study.subjectIntercepts)
if isReporter:
# Output loaded config
if cf.io.outputPrefix is not None:
cfstr = Config.ConfigToYAML(HGMConfigSpec, cf)
with open(cf.io.outputPrefix + "parsedconfig.yaml", "w") as f:
f.write(cfstr)
#common.DebugHere()
# if MPI check if processes are greater than number of subjects. it is okay if there are more subjects than processes
if cf.compute.useMPI and (cf.study.numSubjects < cf.compute.numProcesses):
raise Exception("Please don't use more processes " +
"than total number of individuals")
# subdivide data, create subsets for this thread to work on
nodeSubjectIds = cf.study.subjectIds[rank::cf.compute.numProcesses]
nodeIntercepts = cf.study.subjectIntercepts[rank::cf.compute.numProcesses]
nodeSlopes = cf.study.subjectSlopes[rank::cf.compute.numProcesses]
nodeBaselineTimes = cf.study.subjectBaselineTimes[rank::cf.compute.
numProcesses]
sys.stdout.write(
"This is process %d of %d with name: %s on machinename: %s and local rank: %d.\nnodeIntercepts: %s\n nodeSlopes: %s\n nodeBaselineTimes: %s\n"
% (rank, size, name, nodename, localRank, nodeIntercepts, nodeSlopes,
nodeBaselineTimes))
# mem type is determined by whether or not we're using CUDA
mType = ca.MEM_DEVICE if cf.compute.useCUDA else ca.MEM_HOST
# load data in memory
# load intercepts
J = [
common.LoadITKImage(f, mType) if isinstance(f, str) else f
for f in nodeIntercepts
]
# load slopes
n = [
common.LoadITKField(f, mType) if isinstance(f, str) else f
for f in nodeSlopes
]
# get imGrid from data
imGrid = J[0].grid()
# create time array with checkpointing info for group geodesic
(t, Jind, gCpinds) = HGMSetUpTimeArray(cf.optim.nTimeStepsGroup,
nodeBaselineTimes, 0.0000001)
tdiscGroup = CAvmHGMCommon.HGMSetupTimeDiscretizationGroup(
t, J, n, Jind, gCpinds, mType, nodeSubjectIds)
# create time array with checkpointing info for residual geodesic
(s, scratchInd, rCpinds) = HGMSetUpTimeArray(cf.optim.nTimeStepsResidual,
[1.0], 0.0000001)
tdiscResidual = CAvmHGMCommon.HGMSetupTimeDiscretizationResidual(
s, rCpinds, imGrid, mType)
# create group state and residual state
groupState = CAvmHGMCommon.HGMGroupState(
imGrid,
mType,
cf.vectormomentum.diffOpParamsGroup[0],
cf.vectormomentum.diffOpParamsGroup[1],
cf.vectormomentum.diffOpParamsGroup[2],
t,
cf.optim.NIterForInverse,
cf.vectormomentum.varIntercept,
cf.vectormomentum.varSlope,
cf.vectormomentum.varInterceptReg,
cf.optim.stepSizeGroup,
integMethod=cf.optim.integMethodGroup)
#ca.Copy(groupState.I0, common.LoadITKImage('/usr/sci/projects/ADNI/nikhil/software/vectormomentumtest/TestData/FlowerData/Longitudinal/GroupGeodesic/I0.mhd', mType))
# note that residual state is treated a scratch variable in this algorithm and reused for computing residual geodesics of multiple individual
residualState = CAvmHGMCommon.HGMResidualState(
None,
None,
imGrid,
mType,
cf.vectormomentum.diffOpParamsResidual[0],
cf.vectormomentum.diffOpParamsResidual[1],
cf.vectormomentum.diffOpParamsResidual[2],
s,
cf.optim.NIterForInverse,
cf.vectormomentum.varIntercept,
cf.vectormomentum.varSlope,
cf.vectormomentum.varInterceptReg,
cf.optim.stepSizeResidual,
integMethod=cf.optim.integMethodResidual)
# start up the memory manager for scratch variables
ca.ThreadMemoryManager.init(imGrid, mType, 0)
# need some host memory in np array format for MPI reductions
if cf.compute.useMPI:
mpiImageBuff = None if mType == ca.MEM_HOST else ca.Image3D(
imGrid, ca.MEM_HOST)
mpiFieldBuff = None if mType == ca.MEM_HOST else ca.Field3D(
imGrid, ca.MEM_HOST)
for i in range(len(groupState.t) - 1, -1, -1):
if tdiscGroup[i].J is not None:
indx_last_individual = i
break
'''
# initial template image
ca.SetMem(groupState.I0, 0.0)
tmp = ca.ManagedImage3D(imGrid, mType)
for tdisc in tdiscGroup:
if tdisc.J is not None:
ca.Copy(tmp, tdisc.J)
groupState.I0 += tmp
del tmp
if cf.compute.useMPI:
Compute.Reduce(groupState.I0, mpiImageBuff)
# divide by total num subjects
groupState.I0 /= cf.study.numSubjects
'''
# run the loop
for it in range(cf.optim.Niter):
# compute HGM variation for group
HGMGroupVariation(groupState, tdiscGroup, residualState, tdiscResidual,
cf.io.outputPrefix, rank, it)
common.CheckCUDAError("Error after HGM iteration")
# compute gradient for momenta (m is used as scratch)
# if there are multiple nodes we'll need to sum across processes now
if cf.compute.useMPI:
# do an MPI sum
Compute.Reduce(groupState.sumSplatI, mpiImageBuff)
Compute.Reduce(groupState.sumJac, mpiImageBuff)
Compute.Reduce(groupState.madj, mpiFieldBuff)
# also sum up energies of other nodes
# intercept
Eintercept = np.array([groupState.EnergyHistory[-1][1]])
mpi4py.MPI.COMM_WORLD.Allreduce(mpi4py.MPI.IN_PLACE,
Eintercept,
op=mpi4py.MPI.SUM)
groupState.EnergyHistory[-1][1] = Eintercept[0]
Eslope = np.array([groupState.EnergyHistory[-1][2]])
mpi4py.MPI.COMM_WORLD.Allreduce(mpi4py.MPI.IN_PLACE,
Eslope,
op=mpi4py.MPI.SUM)
groupState.EnergyHistory[-1][2] = Eslope[0]
ca.Copy(groupState.m, groupState.m0)
groupState.diffOp.applyInverseOperator(groupState.m)
ca.Sub_I(groupState.m, groupState.madj)
#groupState.diffOp.applyOperator(groupState.m)
# now take gradient step in momenta for group
if cf.optim.method == 'FIXEDGD':
# take fixed stepsize gradient step
ca.Add_MulC_I(groupState.m0, groupState.m, -cf.optim.stepSizeGroup)
else:
raise Exception("Unknown optimization scheme: " + cf.optim.method)
# end if
# now divide to get the new base image for group
ca.Div(groupState.I0, groupState.sumSplatI, groupState.sumJac)
# keep track of energy in this iteration
if isReporter and cf.io.plotEvery > 0 and ((
(it + 1) % cf.io.plotEvery == 0) or (it == cf.optim.Niter - 1)):
HGMPlots(cf,
groupState,
tdiscGroup,
residualState,
tdiscResidual,
indx_last_individual,
writeOutput=True)
if isReporter:
(VEnergy, IEnergy, SEnergy) = groupState.EnergyHistory[-1]
print datetime.datetime.now().time(
), " Iter", it, "of", cf.optim.Niter, ":", VEnergy + IEnergy + SEnergy, '(Total) = ', VEnergy, '(Vector) + ', IEnergy, '(Intercept) + ', SEnergy, '(Slope)'
# write output images and fields
HGMWriteOutput(cf, groupState, tdiscGroup, isReporter)
def BuildGeoReg(cf):
"""Worker for running geodesic estimation on a subset of individuals
"""
#common.DebugHere()
localRank = Compute.GetMPIInfo()['local_rank']
rank = Compute.GetMPIInfo()['rank']
# prepare output directory
common.Mkdir_p(os.path.dirname(cf.io.outputPrefix))
# just one reporter process on each node
isReporter = rank == 0
# load filenames and times for all subjects
(subjectsIds, subjectsImagePaths,
subjectsTimes) = GeoRegLoadSubjectsDetails(cf.study.subjectFile)
cf.study.numSubjects = len(subjectsIds)
if isReporter:
# Output loaded config
if cf.io.outputPrefix is not None:
cfstr = Config.ConfigToYAML(GeoRegConfigSpec, cf)
with open(cf.io.outputPrefix + "parsedconfig.yaml", "w") as f:
f.write(cfstr)
# if MPI check if processes are greater than number of subjects. it is okay if there are more subjects than processes
if cf.compute.useMPI and (len(subjectsIds) < cf.compute.numProcesses):
raise Exception("Please don't use more processes " +
"than total number of individuals")
nodeSubjectsIds = subjectsIds[rank::cf.compute.numProcesses]
nodeSubjectsImagePaths = subjectsImagePaths[rank::cf.compute.numProcesses]
nodeSubjectsTimes = subjectsTimes[rank::cf.compute.numProcesses]
numLocalSubjects = len(nodeSubjectsImagePaths)
if cf.study.initializationsFile is not None:
(subjectsInitialImages,
subjectsInitialMomenta) = GeoRegLoadSubjectsInitializations(
cf.study.initializationsFile)
nodeSubjectsInitialImages = subjectsInitialImages[rank::cf.compute.
numProcesses]
nodeSubjectsInitialMomenta = subjectsInitialMomenta[rank::cf.compute.
numProcesses]
print 'rank:', rank, ', localRank:', localRank, ', numberSubjects/TotalSubjects:', len(
nodeSubjectsImagePaths
), '/', cf.study.numSubjects, ', nodeSubjectsImagePaths:', nodeSubjectsImagePaths, ', nodeSubjectsTimes:', nodeSubjectsTimes
# mem type is determined by whether or not we're using CUDA
mType = ca.MEM_DEVICE if cf.compute.useCUDA else ca.MEM_HOST
# setting gpuid should be handled in gpu
# if using GPU set device based on local rank
#if cf.compute.useCUDA:
# ca.SetCUDADevice(localRank)
# get image size information
dummyImToGetGridInfo = common.LoadITKImage(nodeSubjectsImagePaths[0][0],
mType)
imGrid = dummyImToGetGridInfo.grid()
if cf.study.setUnitSpacing:
imGrid.setSpacing(ca.Vec3Df(1.0, 1.0, 1.0))
if cf.study.setZeroOrigin:
imGrid.setOrigin(ca.Vec3Df(0, 0, 0))
#del dummyImToGetGridInfo;
# start up the memory manager for scratch variables
ca.ThreadMemoryManager.init(imGrid, mType, 0)
# allocate memory
p = GeoRegVariables(imGrid,
mType,
cf.vectormomentum.diffOpParams[0],
cf.vectormomentum.diffOpParams[1],
cf.vectormomentum.diffOpParams[2],
cf.optim.NIterForInverse,
cf.vectormomentum.sigma,
cf.optim.stepSize,
integMethod=cf.optim.integMethod)
# for each individual run geodesic regression for each subject
for i in range(numLocalSubjects):
# initializations for this subject
if cf.study.initializationsFile is not None:
# assuming the initializations are already preprocessed, in terms of intensities, origin and voxel scalings.
p.I0 = common.LoadITKImage(nodeSubjectsInitialImages[i], mType)
p.m0 = common.LoadITKField(nodeSubjectsInitialMomenta[i], mType)
else:
ca.SetMem(p.m0, 0.0)
ca.SetMem(p.I0, 0.0)
# allocate memory specific to this subject in steps a, b and c
# a. create time array with checkpointing info for regression geodesic, allocate checkpoint memory
(t, msmtinds, cpinds) = GeoRegSetUpTimeArray(cf.optim.nTimeSteps,
nodeSubjectsTimes[i],
0.001)
cpstates = [(ca.Field3D(imGrid, mType), ca.Field3D(imGrid, mType))
for idx in cpinds]
# b. allocate gradAtMeasurements of the length of msmtindex for storing residuals
gradAtMsmts = [ca.Image3D(imGrid, mType) for idx in msmtinds]
# c. load timepoint images for this subject
Imsmts = [
common.LoadITKImage(f, mType) if isinstance(f, str) else f
for f in nodeSubjectsImagePaths[i]
]
# reset stepsize if adaptive stepsize changed it inside
p.stepSize = cf.optim.stepSize
# preprocessimages
GeoRegPreprocessInput(nodeSubjectsIds[i], cf, p, t, Imsmts, cpinds,
cpstates, msmtinds, gradAtMsmts)
# run regression for this subject
# REMEMBER
# msmtinds index into cpinds
# gradAtMsmts is parallel to msmtinds
# cpinds index into t
EnergyHistory = RunGeoReg(nodeSubjectsIds[i], cf, p, t, Imsmts, cpinds,
cpstates, msmtinds, gradAtMsmts)
# write output images and fields for this subject
# TODO: BEWARE There are hardcoded numbers inside preprocessing code specific for ADNI/OASIS brain data.
GeoRegWriteOuput(nodeSubjectsIds[i], cf, p, t, Imsmts, cpinds,
cpstates, msmtinds, gradAtMsmts, EnergyHistory)
# clean up memory specific to this subject
del t, Imsmts, cpinds, cpstates, msmtinds, gradAtMsmts
