def __init__(self, evaluator, evaluable, **args):
assert not isinstance(evaluable, TopologyEvolvable)
if isinstance(evaluable, Module):
evaluable = MaskedModule(evaluable)
else:
evaluable = MaskedParameters(evaluable)
Learner.__init__(self, evaluator, evaluable, **args)
def __init__(self, evaluator, evaluable, **args):
Learner.__init__(self, evaluator, evaluable, **args)
if isinstance(evaluable, ParameterContainer):
# in this case we have to wrap the evaluator
self.wrappingEvaluable = evaluable.copy()
self.wrappingEvaluable.name = 'opt-by-'+self.__class__.__name__
def wrappedEvaluator(x):
self.wrappingEvaluable._setParameters(x)
return evaluator(self.wrappingEvaluable)
self.evaluator = wrappedEvaluator
self.x0 = evaluable.params.copy()
self.bestEvaluable = self.x0.copy()
else:
self.x0 = evaluable.copy()
if self.minimize:
# then we need to change the sign of the evaluations
tmp = self.evaluator
self.evaluator = lambda x: -tmp(x)
if self.desiredEvaluation != None:
self.desiredEvaluation *= -1
# the first guess at the solution (it must be an array)
assert type(self.x0) == ndarray
self.xdim = size(self.x0)
if self.initialSearchRange != None:
for i in range(self.xdim):
mi, ma = self.initialSearchRange[i]
self.x0[i] = rand()*(ma-mi)+mi
self.noisyEvaluator = evaluator.noisy
def __init__(self, evaluator, evaluable, **args):
Learner.__init__(self, evaluator, evaluable, **args)
if isinstance(evaluable, ParameterContainer):
# in this case we have to wrap the evaluator
self.wrappingEvaluable = evaluable.copy()
self.wrappingEvaluable.name = 'opt-by-'+self.__class__.__name__
def wrappedEvaluator(x):
self.wrappingEvaluable._setParameters(x)
return evaluator(self.wrappingEvaluable)
self.evaluator = wrappedEvaluator
self.x0 = evaluable.params.copy()
self.bestEvaluable = self.x0.copy()
else:
self.x0 = evaluable.copy()
if self.minimize:
# then we need to change the sign of the evaluations
tmp = self.evaluator
self.evaluator = lambda x: -tmp(x)
if self.desiredEvaluation != None:
self.desiredEvaluation *= -1
# the first guess at the solution (it must be an array)
assert type(self.x0) == ndarray
self.xdim = size(self.x0)
if self.initialSearchRange != None:
for i in range(self.xdim):
mi, ma = self.initialSearchRange[i]
self.x0[i] = rand()*(ma-mi)+mi
self.noisyEvaluator = evaluator.noisy
def learn(self, maxSteps = None):
""" Some BlackBoxOptimizers can only be called one time, and currently
do not support iteratively adding more steps. """
if not self.online:
if self.maxEvaluations != None:
if maxSteps != None:
maxSteps = min(maxSteps, self.maxEvaluations - self.steps)
else:
maxSteps = self.maxEvaluations-self.steps
self._batchLearn(maxSteps)
else:
Learner.learn(self, maxSteps)
if self.wrappingEvaluable != None and isinstance(self.bestEvaluable, ndarray):
xopt = self.bestEvaluable
self.wrappingEvaluable._setParameters(xopt)
self.bestEvaluable = self.wrappingEvaluable
if self.minimize:
self.bestEvaluation *= -1
return self.bestEvaluable, self.bestEvaluation
def learn(self, maxSteps = None):
""" Some BlackBoxOptimizers can only be called one time, and currently
do not support iteratively adding more steps. """
if not self.online:
if self.maxEvaluations != None:
if maxSteps != None:
maxSteps = min(maxSteps, self.maxEvaluations - self.steps)
else:
maxSteps = self.maxEvaluations-self.steps
self._batchLearn(maxSteps)
else:
Learner.learn(self, maxSteps)
if self.wrappingEvaluable != None and isinstance(self.bestEvaluable, ndarray):
xopt = self.bestEvaluable
self.wrappingEvaluable._setParameters(xopt)
self.bestEvaluable = self.wrappingEvaluable
if self.minimize:
self.bestEvaluation *= -1
return self.bestEvaluable, self.bestEvaluation