def init_random_generator(cls, rdm_object=None):
""" Return a RandomGenerator. Can deal with multiple input: Seed/
RandomGenerator/None
"""
if (rdm_object is None):
return RandomGenerator()
elif (ut.is_int(rdm_object)):
return RandomGenerator(seed=rdm_object)
elif (isinstance(rdm_object, RandomGenerator)):
return rdm_object
else:
raise NotImplementedError()
def _init_params_BO(self, **args_optim):
""" Provides different ways to initialize bo depending of the input type:
<None> / <int>: returns an integer (nb of points to be eval)
e.g. None >> 2 * nb_params
<str>: random_init based on a string. returns a <P x N np.array>
P is the population size N number of parameters
e.g. '40_lhs' >> 40 points drawn by latin hypercube sampling
'50_uniform' >> 50 points drawn uniformly
range of each params is infered from bounds
<P * N array>: passs it though
"""
init_obj = args_optim['init_obj']
nb_params = args_optim['nb_params']
bounds = args_optim['bounds_params']
if (init_obj is None):
init_args = nb_params *2
elif ut.is_int(init_obj):
init_args = init_obj
elif ut.is_string(init_obj):
bits = init_obj.split("_",1)
nb_points_init = int(bits[0])
if(bits[1] == 'lhs'):
size_pop = [nb_points_init, nb_params]
limits = np.array(bounds).T
init_args = self.rdm_gen.init_population_lhs(size_pop, limits)
else:
distrib_one_param = [bits[1]+'_'+ bounds[i][0] + '_' + bounds[i][1] for i in range(nb_params)]
init_matrix = [self.rdm_gen.gen_rdmfunc_from_string(d, dim = nb_points_init) for d in distrib_one_param]
init_args = np.array(init_matrix).T
else:
init_args = init_obj
return init_args
def gen_rdmfunc_from_string(self, method_rdm, dim=1):
"""return a function that when called return random variables according
to the distribution described by the string, with specific dim
convention: dim[-1] correspond to the dim of the RV while dim[:-1] to
the size of the population
TODO: works only for 1D dim
"""
if ut.is_list(method_rdm):
# return a lst of function
# should it be a function returning a list??
func = [
self.gen_rdmfunc_from_string(meth, dim) for meth in method_rdm
]
else:
args = ut.splitString(method_rdm)
if (ut.is_int(dim)):
dimRV = dim
else:
dimRV = dim[-1]
if (args[0] in ['uniform', 'normal']):
if (len(args) == 1):
first_args = np.repeat(0, dimRV)
second_args = np.repeat(1, dimRV)
elif (len(args) == 3):
first_args = np.repeat(float(args[1]), dimRV)
second_args = np.repeat(float(args[2]), dimRV)
elif (len(args) == (1 + 2 * dimRV)):
first_args = np.array(
[float(args[1 + 2 * d]) for d in range(dimRV)])
second_args = np.array(
[float(args[2 + 2 * d]) for d in range(dimRV)])
else:
raise NotImplementedError()
if (dim == 1):
# such that function return a value instead of an array
# may change
first_args, second_args = first_args[0], second_args[0]
dim = None
if args[0] == 'normal':
def func():
return self.normal(first_args, second_args, size=dim)
else:
def func():
return self.uniform(first_args, second_args, size=dim)
elif (args[0] == 'determ'):
if (len(args) == 1):
constant = np.repeat(0, dim)
elif (len(args) == 2):
constant = np.repeat(float(args[1]), dim)
elif (len(args) == (1 + dimRV)):
constant = np.array(args[1:])
if (ut.is_list(dim)):
dim_pop = np.product(dim[:-1])
constant = np.tile(constant, dim_pop).reshape(dim)
else:
raise NotImplementedError()
def func():
return constant
else:
raise NotImplementedError()
return func
def _run_BO2(self, options, **args_call):
""" Bayesian optimization using GPYOpt
def __init__(self, f, domain = None, constraints = None, cost_withGradients = None,
model_type = 'GP', X = None, Y = None, initial_design_numdata = 5,
initial_design_type='random', acquisition_type ='EI', normalize_Y = True,
exact_feval = False, acquisition_optimizer_type = 'lbfgs', model_update_interval=1,
evaluator_type = 'sequential',batch_size = 1, num_cores = 1, verbosity=False,
verbosity_model = False, maximize=False, de_duplication=False, **kwargs)
def run_optimization(self, maxiter = 0, max_time = np.inf, eps = 1e-8,
context = None, verbosity=False, save_models_parameters= True,
report_file = None, evaluations_file = None, models_file=None):
parameters
----------
options.constraints: <str> or None
encode the type of constraints which can be used. This string is converted
to be passed to the BayesianOptimizer constructor with the following structure
[{'name':'name1', 'constraint':obj_constraint1}, .., {'name':'nameN', 'constraint':obj_constraintN}]
where obj_constraintX can be a function or a string which can be exec to generate a function
a constraint is of the form c(x) <= 0 (i.e. a parameter is accepted if c(x) <= 0)
e.g. 'step_a' difference between two consecutive parameters should be less than a
e.g. 'step_a_a0_aN' same as above plus the first (last) parameter is also compared to a0 (aN)
e.g. 'smooth_s' smoothness calculated should be <=s
"""
#Init BO
model = options['model']
nb_params = options['nb_params']
name_params = [str(i) for i in range(nb_params)]
options['name_params'] = name_params
bounds_bo = [{'name': name_params[i], 'type': 'continuous',
'domain': options['bounds_params'][i]} for i in range(nb_params)]
constraints = options.get('constraints')
cst = self._build_constraints(constraints, **options)
args_BO = {'acquisition_type':options['acq'],
'acquisition_optimizer_type':options['acq_opt_type'],
'num_cores':options['num_cores'],
'domain': bounds_bo,
'optim_num_anchor':options['optim_num_anchor'],
'optim_num_samples':options['optim_num_samples'],
'acquisition_jitter':options['acquisition_jitter'],
'acquisition_weight':options['acquisition_weight'],
'batch_size':options['batch_size'],
'evaluator_type':options['batch_method'],
'num_inducing':options['num_inducing'],
'model_type':options['model_type'],
'ARD':options['ARD'],
'acquisition_weight_lindec':options['acquisition_weight_lindec'],
'constraints':cst}
# definition of the cost function
def cost(params):
return model(np.squeeze(params), **args_call)
#V0.1 NON DEFAULT LIKELIHOOD
if(args_BO['model_type'] == 'GP_CUSTOM_LIK'):
logger.info('Use of GP_CUSTOM_LIK: enforce **normalize_Y** as False ')
logger.info('Use of GP_CUSTOM_LIK: enforce **inf_method** as Laplace')
args_BO['normalize_Y'] = False
args_BO['inf_method'] = 'Laplace'
args_BO['likelihood'] = options.get('likelihood', 'Binomial_10')
#V0.1 transfer learning
to_transfer = options['to_transfer']
if((to_transfer is not None) and (args_BO['model_type'] == 'GP_STACKED')):
args_BO_transfer = copy.copy(args_BO)
args_BO_transfer['model_type'] = 'GP'
args_BO_transfer['X'] = to_transfer['X']
args_BO_transfer['Y'] = to_transfer['Y']
regressor_to_transfer = GPyOpt.methods.BayesianOptimization(cost, **args_BO)
regressor_to_transfer.model._create_model(args_BO_transfer['X'], args_BO_transfer['Y'])
args_BO['prev'] = regressor_to_transfer.model
init = options['init_params']
if(init is None):
args_BO['initial_design_numdata'] = int(3 * nb_params)
args_BO['initial_design_type'] = options['initial_design_type']
elif(ut.is_int(init)):
args_BO['initial_design_numdata'] = init
args_BO['initial_design_type'] = options['initial_design_type']
else:
args_BO['X'] = init
logger.info('Init of the GP: acquisition of {} points'.format(init.shape[0]))
args_BO['Y'] = np.array([cost(x) for x in init])
ker = options['kernel'] # matern52 /matern32
if(ker is not None):
ard = options['ARD']
if(ker == 'matern52'):
args_BO['ker'] = eval("GPy.kern.Matern52(self.input_dim = nb_params, ARD={})".format(ard))
elif(ker == 'matern32'):
args_BO['ker'] = eval("GPy.kern.Matern32(self.input_dim = nb_params, ARD={})".format(ard))
else:
logger.warning("{} not a valid kernel".format(ker))
bo = GPyOpt.methods.BayesianOptimization(cost, **args_BO)
#Initialization phase
# Exploration-Exploitation phase
max_time = options['max_time']
bo.run_optimization(max_iter = options['maxiter'], max_time = max_time)
time_left = max_time - bo.cum_time
max_reached = (time_left < 0)
# Exploitation phase
# should it be forced eben if max_time is reached ? No so far
exploit_steps = options['exploit_steps']
if(exploit_steps > 0):
if(not(max_reached)):
bo.acquisition_type = 'LCB'
bo.acquisition_weight = 0.000001
bo.kwargs['acquisition_weight'] = 0.000001
bo.acquisition = bo._acquisition_chooser()
#maybe to add new logic when Batch/Sparse
if(bo.batch_size > 1):
bo.batch_size = 1
bo.evaluator = bo._evaluator_chooser()
logger.info('Exploitation (i.e. ucb with k=0) for {}'.format(exploit_steps))
logger.info('Still {} s left'.format(time_left))
bo.run_optimization(exploit_steps, max_time = time_left)
time_left -= bo.cum_time
max_reached = (time_left < 0)
if((_still_potentially_better(bo)) and (exploit_steps <= 50) and not(max_reached)):
logger.info('2nd round of exploitation for {}'.format(exploit_steps))
logger.info('Still {} s left'.format(time_left))
bo.run_optimization(exploit_steps, max_time = time_left)
# generate results
optim_params = bo.x_opt
optim_value = bo.fx_opt
optim_params_exp = _get_best_exp_from_BO(bo)
optim_value_cputed = model(optim_params)
nfev = len(bo.X)
resultTest = {'x': optim_params, 'x_exp':optim_params_exp, 'fun': optim_value,
'fun_ver': optim_value_cputed, 'nfev':nfev, 'nit':nfev, 'sucess':True}
resultTest['gp_kernel_optim_names'] = bo.model.model.parameter_names()
resultTest['gp_kernel_optim_vals'] = bo.model.model.param_array
resultTest['nb_processes'] = bo.num_cores
resultTest['nb_cpus'] = self.mp.n_cpus
resultTest['X_evol'] = bo.X
resultTest['Y_evol'] = bo.Y
resultTest['Y_best'] = bo.Y_best
resultTest['still_potentially_better'] = _still_potentially_better(bo)
resultTest['maxtime'] = max_reached
# Close pool of processors used (if needed)
self.mp.close_mp()
return resultTest