class ProperSurrogateModel(SurrogateModel):
def __init__(self, fitness, configuration, controller):
super(ProperSurrogateModel, self).__init__(fitness, configuration,
controller)
if configuration.classifier == 'SupportVectorMachine':
self.classifier = SupportVectorMachineClassifier()
else:
logging.error('Classifier type {} not found'.format(
configuration.classifier))
if configuration.regressor == 'GaussianProcess':
self.regressor = GaussianProcessRegressor(controller)
else:
logging.error('Regressor type {} not found'.format(
configuration.regressor))
def train(self, pop):
dimensions = self.fitness.dimensions
return self.classifier.train(pop) and self.regressor.train(
pop, self.configuration, dimensions)
def add_training_instance(self, part, code, fitness):
self.classifier.add_training_instance(part, code)
self.regressor.add_training_instance(part, fitness)
def __init__(self, fitness, configuration, controller):
super(ProperSurrogateModel, self).__init__(fitness, configuration,
controller)
if configuration.classifier == 'SupportVectorMachine':
self.classifier = SupportVectorMachineClassifier()
else:
logging.error('Classifier type {} not found'.format(
configuration.classifier))
if configuration.regressor == 'GaussianProcess':
self.regressor = GaussianProcessRegressor(controller)
else:
logging.error('Regressor type {} not found'.format(
configuration.regressor))
def __init__(self, configuration, controller, fitness):
super(ProperSurrogateModel, self).__init__(configuration,
controller,
fitness)
if configuration.classifier == 'SupportVectorMachine':
self.classifier = SupportVectorMachineClassifier()
else:
logging.error('Classifier type ' + str(configuration.classifier) + ' not found')
self.regressor = self.regressor_constructor()
try:
if self.configuration.sample_on == "ei":
self.max_uncertainty = self.max_ei
elif self.configuration.sample_on == "s":
self.max_uncertainty = self.max_s2
except:
if self.max_uncertainty:
pass
else:
logging.debug("Sampling scheme wasnt specified, using Expected Improvment")
self.max_uncertainty = self.max_ei
class ProperSurrogateModel(SurrogateModel):
def __init__(self, configuration, controller, fitness):
super(ProperSurrogateModel, self).__init__(configuration,
controller,
fitness)
if configuration.classifier == 'SupportVectorMachine':
self.classifier = SupportVectorMachineClassifier()
else:
logging.error('Classifier type ' + str(configuration.classifier) + ' not found')
self.regressor = self.regressor_constructor()
try:
if self.configuration.sample_on == "ei":
self.max_uncertainty = self.max_ei
elif self.configuration.sample_on == "s":
self.max_uncertainty = self.max_s2
except:
if self.max_uncertainty:
pass
else:
logging.debug("Sampling scheme wasnt specified, using Expected Improvment")
self.max_uncertainty = self.max_ei
def get_regressor(self):
return self.regressor
def get_classifier(self):
return self.classifier
def get_copy(self):
model_copy = ProperSurrogateModel(self.configuration, self.controller)
model_copy.set_state_dictionary(self.get_state_dictionary())
return model_copy
def predict(self, particles):
try:
#logging.debug("Using tranformation function for the regressor")
trans_particles = particles
except:
trans_particles = [self.fitness.transformation_function(part) for part in particles]
MU, S2, EI, P = self.regressor.predict(trans_particles)
return self.classifier.predict(particles), MU, S2, EI, P
def train(self, hypercube=None):
self.was_trained = True
if self.classifier.train() and self.regressor.train():
logging.info("Trained Surrogate Model")
else:
logging.info("Couldnt Train Surrogate Model")
return False
def regressor_constructor(self):
controller = self.controller
configuration = self.configuration
if self.configuration.regressor == 'GaussianProcess':
return GaussianProcessRegressor(controller, configuration)
elif self.configuration.regressor == 'GaussianProcess2':
return GaussianProcessRegressor2(controller, configuration)
elif self.configuration.regressor == 'GaussianProcess3':
return GaussianProcessRegressor3(controller, configuration)
elif self.configuration.regressor == 'KMeansGaussianProcessRegressor':
return KMeansGaussianProcessRegressor(controller, configuration)
elif self.configuration.regressor == 'DPGMMGaussianProcessRegressor':
return DPGMMGaussianProcessRegressor(controller, configuration)
elif self.configuration.regressor == 'R':
return GaussianProcessRegressorRpy(controller, configuration)
else:
raise Exception('Regressor type ' + str(configuration.regressor) + ' not found')
def add_training_instance(self, part, code, fitness, addReturn):
self.classifier.add_training_instance(part, code)
if addReturn[0] == 0: ## only update regressor if the fitness function produced a result
try:
trans_part = self.fitness.transformation_function(part)
#logging.debug("Using tranformation function for the regressor")
except:
trans_part = part
self.regressor.add_training_instance(trans_part, fitness)
def contains_training_instance(self, part):
try:
trans_part = self.fitness.transformation_function(part)
#logging.debug("Using tranformation function for the regressor")
except:
trans_part = part
return self.regressor.contains_training_instance(trans_part) or self.classifier.contains_training_instance(part)
def get_training_instance(self, part):
code = self.classifier.get_training_instance(part)
fitness = None
if self.regressor.contains_training_instance(part):
fitness = self.regressor.get_training_instance(part)
return code, fitness
def model_failed(self, part):
return False
def max_ei(self, designSpace, hypercube=None, npts=10):
D = len(designSpace)
n_bins = npts*ones(D)
grid = False
if grid:
if hypercube:
result = mgrid[[slice(h_min, h_max, npts*1.0j) for h_max, h_min , n in zip(hypercube[0],hypercube[1], n_bins)]]
z = result.reshape(D,-1).T
else:
bounds = [(d["min"],d["max"]) for d in designSpace]
result = mgrid[[slice(row[0], row[1], npts*1.0j) for row, n in zip(bounds, n_bins)]]
z = result.reshape(D,-1).T
'''
x,y,v = mgrid[designSpace[0]["min"]:designSpace[0]["max"]:(int(designSpace[0]["max"]-designSpace[0]["min"])+1)*1.0j,designSpace[1]["min"]:designSpace[1]["max"]:(int(designSpace[1]["max"]-designSpace[1]["min"])+1)*1.0j , designSpace[2]["min"]:designSpace[2]["max"]:(int(designSpace[2]["max"]-designSpace[2]["min"])+1)*1.0j]
x=reshape(x,-1)
y=reshape(y,-1)
v=reshape(v,-1)
z = array([[a,b,c] for (a,b,c) in zip(x,y,v)])
'''
try:
zClass, MU, S2, EI, P = self.predict(z)
filteredEI=[]
filteredZ=[]
for i,ei in enumerate(EI):
if zClass[i]==0:
filteredEI.append(ei)
filteredZ.append(z[i])
EI = array(filteredEI)
return filteredZ[argmax(EI)]
except Exception,e:
logging.error("Finding max S2 failed: " + str(e))
return None
else: ## more memory efficient yet slower