def __network_mode(self, workers):
printf("Starting optimization with network workers")
if not workers:
raise RuntimeError("You have to specify a list of remote"
" worker addresses")
# Connect to workers with ZeroMQ.
self.cc = NetworkClientComm(addresses=parse_worker_addresses(workers))
def __network_mode(self, workers):
printf("Starting optimization with network workers")
if not workers:
raise RuntimeError("You have to specify a list of remote"
" worker addresses")
# Connect to workers with ZeroMQ.
self.cc = NetworkClientComm(addresses=parse_worker_addresses(workers))
def __local_mode(self, workers, xargs, evaluator_path):
if evaluator_path is None:
raise ValueError("You have to specify a path to the evaluator"
" while running in local mode.")
printf("Starting optimization with local workers (%d)" % workers)
# Prepare the ZeroMQ communication layer.
self.cc = LocalClientComm()
# Arguments to be passed to evaluator processes.
evaluator_args = xargs + ["-local-mode", "-local-pull-address",
self.cc.push_addr, "-local-publish-address", self.cc.sub_addr]
# Launch worker processes.
spawn_workers(workers, evaluator_path, evaluator_args)
def __local_mode(self, workers, xargs, evaluator_path):
if evaluator_path is None:
raise ValueError("You have to specify a path to the evaluator"
" while running in local mode.")
printf("Starting optimization with local workers (%d)" % workers)
# Prepare the ZeroMQ communication layer.
self.cc = LocalClientComm()
# Arguments to be passed to evaluator processes.
evaluator_args = xargs + [
"-local-mode", "-local-pull-address", self.cc.push_addr,
"-local-publish-address", self.cc.sub_addr
]
# Launch worker processes.
spawn_workers(workers, evaluator_path, evaluator_args)
class HybridOptimizer(object):
"The hybrid (two-step) optimization engine."
def __init__(self,
evaluator_path=None,
local=True,
local_workers=None,
remote_workers=None,
stop_flag=None,
unknown_args=None,
**kwargs):
self.cc = None # Client communicator.
self.stop_flag = stop_flag
self.extra_args = kwargs
# Initialize relation with workers.
if local:
self.__local_mode(local_workers, unknown_args, evaluator_path)
else:
self.__network_mode(remote_workers)
def __del__(self):
# Close communicator.
if self.cc:
self.cc.close()
def __local_mode(self, workers, xargs, evaluator_path):
if evaluator_path is None:
raise ValueError("You have to specify a path to the evaluator"
" while running in local mode.")
printf("Starting optimization with local workers (%d)" % workers)
# Prepare the ZeroMQ communication layer.
self.cc = LocalClientComm()
# Arguments to be passed to evaluator processes.
evaluator_args = xargs + [
"-local-mode", "-local-pull-address", self.cc.push_addr,
"-local-publish-address", self.cc.sub_addr
]
# Launch worker processes.
spawn_workers(workers, evaluator_path, evaluator_args)
def __network_mode(self, workers):
printf("Starting optimization with network workers")
if not workers:
raise RuntimeError("You have to specify a list of remote"
" worker addresses")
# Connect to workers with ZeroMQ.
self.cc = NetworkClientComm(addresses=parse_worker_addresses(workers))
def __fetch_constraints(self):
u"Fetch constraints from any worker."
_sent = False
while not _sent:
_sent = self.cc.get_options(0, wait=False)
if check_stop_flag(self.stop_flag):
sys.exit(-1)
resp = self.cc.resp_options(0, wait=True)
self.no_vars = resp["num_params"]
self.mins = resp["min_constr"]
self.maxes = resp["max_constr"]
def prepare_optimization(self):
# Wait until (presumably) all workers are awake and ready
# to avoid unfair distribution of tasks.
time.sleep(1)
printf("Waiting for initial connection")
self.__fetch_constraints()
printf("Received initial data from workers")
def run(self):
self.prepare_optimization()
args = dict(no_vars=self.no_vars,
mins=self.mins,
maxes=self.maxes,
comm=self.cc,
stop_flag=self.stop_flag)
ga_results = self.run_genetic(args, **self.extra_args)
self.display_info(ga_results)
lm_results = self.run_levmar(args, ga_results, **self.extra_args)
self.display_info(lm_results)
self.save_output(lm_results)
def run_genetic(self, args, **extra):
u"Prepare and run genetic step."
# Translate all additional arguments that are understood in this step.
ga_args = dict(args)
if extra["ga_seed"]:
ga_args["seed"] = extra["ga_seed"]
if extra["ga_iter"]:
ga_args["generations"] = extra["ga_iter"]
if extra["ga_size"]:
ga_args["size"] = extra["ga_size"]
if extra["ga_allele"] is not None:
ga_args["allele"] = extra["ga_allele"]
# Run genetic algorithm optimization step...
ga_opt = GeneticOptimization(**ga_args)
return ga_opt.run() # Return optimized parameters vector.
def run_levmar(self, args, start_vector, **extra):
u"Prepare and run levmar step."
lm_args = dict(args)
lm_args["p0"] = start_vector
if extra["lm_iter"]:
lm_args["max_iter"] = extra["lm_iter"]
if extra["lm_central"] is not None:
lm_args["central"] = extra["lm_central"]
lm_opt = LevmarOptimization(**lm_args)
with warnings.catch_warnings():
warnings.filterwarnings("ignore")
return lm_opt.run()
def display_info(self, parameters):
u"Asks evaluator for parameters description and displays it."
self.cc.get_stats(parameters, "2%d" % id(parameters))
response = self.cc.resp_stats("2%d" % id(parameters), wait=True)
printf("Statistics for given parameters:")
printf(response["stats"])
def save_output(self, parameters):
u"Save optimization results"
self.cc.save_output(parameters, 1)
response = self.cc.resp_save(1, wait=True)
if response["status"] == "" and response["files"]:
print "Saved files: %s" % ', '.join(response["files"])
class HybridOptimizer(object):
"The hybrid (two-step) optimization engine."
def __init__(self, evaluator_path=None, local=True, local_workers=None,
remote_workers=None, stop_flag=None, unknown_args=None,
**kwargs):
self.cc = None # Client communicator.
self.stop_flag = stop_flag
self.extra_args = kwargs
# Initialize relation with workers.
if local:
self.__local_mode(local_workers, unknown_args, evaluator_path)
else:
self.__network_mode(remote_workers)
def __del__(self):
# Close communicator.
if self.cc:
self.cc.close()
def __local_mode(self, workers, xargs, evaluator_path):
if evaluator_path is None:
raise ValueError("You have to specify a path to the evaluator"
" while running in local mode.")
printf("Starting optimization with local workers (%d)" % workers)
# Prepare the ZeroMQ communication layer.
self.cc = LocalClientComm()
# Arguments to be passed to evaluator processes.
evaluator_args = xargs + ["-local-mode", "-local-pull-address",
self.cc.push_addr, "-local-publish-address", self.cc.sub_addr]
# Launch worker processes.
spawn_workers(workers, evaluator_path, evaluator_args)
def __network_mode(self, workers):
printf("Starting optimization with network workers")
if not workers:
raise RuntimeError("You have to specify a list of remote"
" worker addresses")
# Connect to workers with ZeroMQ.
self.cc = NetworkClientComm(addresses=parse_worker_addresses(workers))
def __fetch_constraints(self):
u"Fetch constraints from any worker."
_sent = False
while not _sent:
_sent = self.cc.get_options(0, wait=False)
if check_stop_flag(self.stop_flag):
sys.exit(-1)
resp = self.cc.resp_options(0, wait=True)
self.no_vars = resp["num_params"]
self.mins = resp["min_constr"]
self.maxes = resp["max_constr"]
def prepare_optimization(self):
# Wait until (presumably) all workers are awake and ready
# to avoid unfair distribution of tasks.
time.sleep(1)
printf("Waiting for initial connection")
self.__fetch_constraints()
printf("Received initial data from workers")
def run(self):
self.prepare_optimization()
args = dict(no_vars=self.no_vars, mins=self.mins, maxes=self.maxes,
comm=self.cc, stop_flag=self.stop_flag)
ga_results = self.run_genetic(args, **self.extra_args)
self.display_info(ga_results)
lm_results = self.run_levmar(args, ga_results, **self.extra_args)
self.display_info(lm_results)
self.save_output(lm_results)
def run_genetic(self, args, **extra):
u"Prepare and run genetic step."
# Translate all additional arguments that are understood in this step.
ga_args = dict(args)
if extra["ga_seed"]:
ga_args["seed"] = extra["ga_seed"]
if extra["ga_iter"]:
ga_args["generations"] = extra["ga_iter"]
if extra["ga_size"]:
ga_args["size"] = extra["ga_size"]
if extra["ga_allele"] is not None:
ga_args["allele"] = extra["ga_allele"]
# Run genetic algorithm optimization step...
ga_opt = GeneticOptimization(**ga_args)
return ga_opt.run() # Return optimized parameters vector.
def run_levmar(self, args, start_vector, **extra):
u"Prepare and run levmar step."
lm_args = dict(args)
lm_args["p0"] = start_vector
if extra["lm_iter"]:
lm_args["max_iter"] = extra["lm_iter"]
if extra["lm_central"] is not None:
lm_args["central"] = extra["lm_central"]
lm_opt = LevmarOptimization(**lm_args)
with warnings.catch_warnings():
warnings.filterwarnings("ignore")
return lm_opt.run()
def display_info(self, parameters):
u"Asks evaluator for parameters description and displays it."
self.cc.get_stats(parameters, "2%d" % id(parameters))
response = self.cc.resp_stats("2%d" % id(parameters), wait=True)
printf("Statistics for given parameters:")
printf(response["stats"])
def save_output(self, parameters):
u"Save optimization results"
self.cc.save_output(parameters, 1)
response = self.cc.resp_save(1, wait=True)
if response["status"] == "" and response["files"]:
print "Saved files: %s" % ', '.join(response["files"])
