def __init__(self,
expFileName,
sysPath,
messengerHost,
messengerPort,
messengerUser,
messengerPassword,
sync,
aggregator,
learnerFactory,
dataSourceFactory,
stoppingCriterion,
initHandler=InitializationHandler()):
self._uniqueId = str(os.getpid())
self.expFileName = expFileName
self.sysPath = sysPath
self.messengerHost = messengerHost
self.messengerPort = messengerPort
self.messengerUser = messengerUser
self.messengerPassword = messengerPassword
self.sync = sync
self.aggregator = aggregator
self.learnerFactory = learnerFactory
self.dataSourceFactory = dataSourceFactory
self.dataScheduler = IntervalDataScheduler()
self.initHandler = initHandler
self.stoppingCriterion = stoppingCriterion
def __init__(self,
executionMode,
messengerHost,
messengerPort,
numberOfNodes,
sync,
aggregator,
learnerFactory,
dataSourceFactory,
stoppingCriterion,
initHandler=InitializationHandler(),
dataScheduler=IntervalDataScheduler,
minStartNodes=0,
minStopNodes=0,
sleepTime=5):
self.executionMode = executionMode
if executionMode == 'cpu':
self.devices = None
self.modelsPer = None
else:
self.devices = []
if os.environ.get('CUDA_VISIBLE_DEVICES') is None:
gpuIds = range(
str(subprocess.check_output(["nvidia-smi",
"-L"])).count('UUID'))
else:
gpuIds = os.environ.get('CUDA_VISIBLE_DEVICES').split(',')
for taskid in gpuIds:
self.devices.append('cuda:' + str(taskid))
self.modelsPer = math.ceil(numberOfNodes * 1.0 / len(self.devices))
print(self.modelsPer, "models per gpu on", ','.join(self.devices))
self.messengerHost = messengerHost
self.messengerPort = messengerPort
self.numberOfNodes = numberOfNodes
self.sync = sync
self.aggregator = aggregator
self.learnerFactory = learnerFactory
self.dataSourceFactory = dataSourceFactory
self.stoppingCriterion = stoppingCriterion
self.initHandler = initHandler
self.dataScheduler = dataScheduler
self._uniqueId = str(os.getpid())
self.sleepTime = sleepTime
self.minStartNodes = minStartNodes
self.minStopNodes = minStopNodes
aggregator = Average()
stoppingCriterion = MaxAmountExamples(2800)
dsFactory = FileDataSourceFactory(
filename="../../../../data/textualMNIST/mnist_train.txt",
decoder=MNISTDecoder(),
numberOfNodes=numberOfNodes,
indices='roundRobin',
shuffle=False,
cache=False)
learnerFactory = PytorchLearnerFactory(network=DropoutNet(),
updateRule=updateRule,
learningRate=learningRate,
learningParams=learningParams,
lossFunction=lossFunction,
batchSize=batchSize,
syncPeriod=syncPeriod)
initHandler = InitializationHandler()
exp = Experiment(executionMode=executionMode,
messengerHost=messengerHost,
messengerPort=messengerPort,
numberOfNodes=numberOfNodes,
sync=sync,
aggregator=aggregator,
learnerFactory=learnerFactory,
dataSourceFactory=dsFactory,
stoppingCriterion=stoppingCriterion,
initHandler=initHandler)
exp.run("MNISTtorchCNNwithDropOut")