def __init__(self, **kwargs):
"""
CosmoHammer sampler implementation
"""
self.pool = MpiPool(self._getMapFunction())
self.rank = self.pool.rank
super(MpiCosmoHammerSampler, self).__init__(pool=self.pool, **kwargs)
def __init__(self, func, low, high, particleCount=25,req=1e-5, threads=1, InPos=None):
self.threads = threads
self.pool = MpiPool(self._getMapFunction())
super(MpiParticleSwarmOptimizer, self).__init__(func, low, high,
req=req,
particleCount=particleCount,
pool=self.pool,
InPos=InPos)
def __init__(self, **kwargs):
"""
CosmoHammer sampler implementation
"""
self.pool = MpiPool(self._getMapFunction())
self.rank = self.pool.rank
super(MpiCosmoHammerSampler, self).__init__(pool=self.pool, **kwargs)
def __init__(self, func, low, high, particleCount=25, threads=1):
self.threads = threads
pool = MpiPool(self._getMapFunction())
super(MpiParticleSwarmOptimizer,
self).__init__(func,
low,
high,
particleCount=particleCount,
pool=pool)
class MpiParticleSwarmOptimizer(ParticleSwarmOptimizer):
"""
PSO with support for MPI to distribute the workload over multiple nodes
"""
def __init__(self, func, low, high, particleCount=25,req=1e-5, threads=1, InPos=None):
self.threads = threads
self.pool = MpiPool(self._getMapFunction())
super(MpiParticleSwarmOptimizer, self).__init__(func, low, high,
req=req,
particleCount=particleCount,
pool=self.pool,
InPos=InPos)
def _getMapFunction(self):
if self.threads > 1:
pool = multiprocessing.Pool(self.threads)
return pool.map
else:
return map
def _converged(self, p, m, n):
if(self.isMaster()):
converged = super(MpiParticleSwarmOptimizer, self)._converged( p, m, n)
else:
converged = False
converged = mpiBCast(converged)
return converged
def _get_fitness(self,swarm):
mapFunction = self.pool.map
mpiSwarm = mpiBCast(swarm)
pos = numpy.array([part.position for part in mpiSwarm])
results = mapFunction(self.func, pos)
lnprob = numpy.array([l[0] for l in results])
for i, particle in enumerate(swarm):
particle.fitness = lnprob[i]
particle.position = pos[i]
def isMaster(self):
return self.pool.isMaster()
class MpiCosmoHammerSampler(CosmoHammerSampler):
"""
A sampler implementation extending the regular sampler in order to allow for distributing
the computation with MPI.
:param kwargs:
key word arguments passed to the CosmoHammerSampler
"""
def __init__(self, **kwargs):
"""
CosmoHammer sampler implementation
"""
self.pool = MpiPool(self._getMapFunction())
self.rank = self.pool.rank
super(MpiCosmoHammerSampler, self).__init__(pool=self.pool, **kwargs)
def _getMapFunction(self):
"""
Returns the build in map function
"""
return map
def createSampleFileUtil(self):
"""
Returns a new instance of a File Util
"""
return SampleFileUtil(self.filePrefix, self.isMaster(), reuseBurnin=self.reuseBurnin)
def sampleBurnin(self, p0):
"""
Starts the sampling process. The master node (mpi rank = 0) persists the result to the disk
"""
p0 = mpiBCast(p0)
self.log("MPI Process rank "+ str(self.rank)+" starts sampling")
return super(MpiCosmoHammerSampler, self).sampleBurnin(p0);
def sample(self, burninPos, burninProb, burninRstate, datas):
"""
Starts the sampling process. The master node (mpi rank = 0) persists the result to the disk
"""
burninPos = mpiBCast(burninPos)
burninProb = mpiBCast(burninProb)
burninRstate = mpiBCast(burninRstate)
self.log("MPI Process rank "+ str(self.rank)+" starts sampling")
super(MpiCosmoHammerSampler, self).sample(burninPos, burninProb, burninRstate, datas);
def loadBurnin(self):
"""
loads the burn in form the file system
"""
if(self.isMaster()):
pos, prob, rstate = super(MpiCosmoHammerSampler, self).loadBurnin()
else:
pos, prob, rstate = []
pos = mpiBCast(pos)
prob = mpiBCast(prob)
rstate = mpiBCast(rstate)
self.log("loading done")
return pos, prob, rstate
def createInitPos(self):
"""
Factory method to create initial positions
"""
#bcast the positions to ensure that all mpi nodes start at the same position
return mpiBCast(super(MpiCosmoHammerSampler, self).createInitPos())
def isMaster(self):
"""
Returns true if the rank is 0
"""
return self.pool.isMaster()
def gather(self,value):
"""
Returns the value. Can be overridden with MPI
"""
#if(self.isMaster()):
return mpiMean(value)
start_time = time.time()
#path2load = '/mnt/lnec/sibirrer/input.txt'
path2load = str(sys.argv[1])
f = open(path2load, 'rb')
[lensDES, walkerRatio, n_burn, n_run, mean_start, sigma_start, lowerLimit, upperLimit, path2dump] = dill.load(f)
f.close()
end_time = time.time()
#print end_time - start_time, 'time used for initialisation'
# run the computation
from easylens.Fitting.mcmc import MCMC_sampler
sampler = MCMC_sampler(lensDES, fix_center=False)
samples = sampler.mcmc_CH(walkerRatio, n_run, n_burn, mean_start, sigma_start, lowerLimit, upperLimit, threadCount=1, init_pos=None, mpi_monch=True)
# save the output
pool = MpiPool(None)
if pool.isMaster():
f = open(path2dump, 'wb')
pickle.dump(samples, f)
f.close()
end_time = time.time()
print(end_time - start_time, 'total time needed for computation')
print('Result saved in:', path2dump)
print('============ CONGRATULATION, YOUR JOB WAS SUCCESSFUL ================ ')
__author__ = 'sibirrer'
#this file is ment to be a shell script to be run with Monch cluster
# set up the scene
from cosmoHammer.util.MpiUtil import MpiPool
import time
import sys
import pickle
import os
from lenstronomy.Workflow.fitting_sequence import FittingSequence
pool = MpiPool(None)
start_time = time.time()
job_name = str(sys.argv[1])
if pool.isMaster():
print("job %s loaded" % job_name)
# hoffman2 specifics
dir_path_cluster = '/u/flashscratch/s/sibirrer/'
path2load = os.path.join(dir_path_cluster, job_name) + ".txt"
path2dump = os.path.join(dir_path_cluster, job_name) + "_out.txt"
f = open(path2load, 'rb')
input = pickle.load(f)
f.close()
[
fitting_kwargs_list, multi_band_list, kwargs_model, kwargs_constraints,
kwargs_likelihood, kwargs_params, init_samples
class MpiCosmoHammerSampler(CosmoHammerSampler):
"""
A sampler implementation extending the regular sampler in order to allow for distributing
the computation with MPI.
:param kwargs:
key word arguments passed to the CosmoHammerSampler
"""
def __init__(self, **kwargs):
"""
CosmoHammer sampler implementation
"""
self.pool = MpiPool(self._getMapFunction())
self.rank = self.pool.rank
super(MpiCosmoHammerSampler, self).__init__(pool=self.pool, **kwargs)
def _getMapFunction(self):
"""
Returns the build in map function
"""
return map
def createSampleFileUtil(self):
"""
Returns a new instance of a File Util
"""
return SampleFileUtil(self.filePrefix,
self.isMaster(),
reuseBurnin=self.reuseBurnin)
def sampleBurnin(self, p0):
"""
Starts the sampling process. The master node (mpi rank = 0) persists the result to the disk
"""
p0 = mpiBCast(p0)
self.log("MPI Process rank " + str(self.rank) + " starts sampling")
return super(MpiCosmoHammerSampler, self).sampleBurnin(p0)
def sample(self, burninPos, burninProb, burninRstate, datas):
"""
Starts the sampling process. The master node (mpi rank = 0) persists the result to the disk
"""
burninPos = mpiBCast(burninPos)
burninProb = mpiBCast(burninProb)
burninRstate = mpiBCast(burninRstate)
self.log("MPI Process rank " + str(self.rank) + " starts sampling")
super(MpiCosmoHammerSampler, self).sample(burninPos, burninProb,
burninRstate, datas)
def loadBurnin(self):
"""
loads the burn in form the file system
"""
if (self.isMaster()):
pos, prob, rstate = super(MpiCosmoHammerSampler, self).loadBurnin()
else:
pos, prob, rstate = []
pos = mpiBCast(pos)
prob = mpiBCast(prob)
rstate = mpiBCast(rstate)
self.log("loading done")
return pos, prob, rstate
def createInitPos(self):
"""
Factory method to create initial positions
"""
#bcast the positions to ensure that all mpi nodes start at the same position
return mpiBCast(super(MpiCosmoHammerSampler, self).createInitPos())
def isMaster(self):
"""
Returns true if the rank is 0
"""
return self.pool.isMaster()
