def main():
""" Main function. """
# pylint: disable=too-many-locals
# pylint: disable=maybe-no-member
np.random.seed(0)
reporter = BasicReporter()
Z_tr, X_tr, Y_tr, ZX_tr, Z_te, X_te, Y_te, ZX_te = get_data()
sgp_options = load_options(gp_instances.all_simple_gp_args,
'GP',
reporter=reporter)
mfgp_options = load_options(mf_gp.all_mf_gp_args,
'MFGP',
reporter=reporter)
mfgp_options.mean_func_type = 'median'
# Fit the GPs.
sgp_fitter = gp_instances.SimpleGPFitter(ZX_tr,
Y_tr,
sgp_options,
reporter=reporter)
sgp, opt_s = sgp_fitter.fit_gp()
mfgp_fitter = mf_gp.MFGPFitter(Z_tr,
X_tr,
Y_tr,
mfgp_options,
reporter=reporter)
mfgp, opt_mf = mfgp_fitter.fit_gp()
opt_s = (np.array(opt_s).round(4))
opt_mf = (np.array(opt_mf).round(4))
s_bounds = sgp_fitter.hp_bounds.round(3)
mf_bounds = mfgp_fitter.hp_bounds.round(3)
# Print out some fitting statistics
_print_str_results(reporter, 'Opt-pts', str(opt_s), str(opt_mf))
_print_str_results(reporter, 'Opt-bounds', str(s_bounds), str(mf_bounds))
# The marginal likelihoods
sgp_lml = sgp.compute_log_marginal_likelihood()
mfgp_lml = mfgp.compute_log_marginal_likelihood()
_print_float_results(reporter, 'Log_Marg_Like', sgp_lml, mfgp_lml)
# Train errors
s_pred, _ = sgp.eval(ZX_tr)
mf_pred, _ = mfgp.eval_at_fidel(Z_tr, X_tr)
sgp_tr_err = compute_average_sq_prediction_error(Y_tr, s_pred)
mfgp_tr_err = compute_average_sq_prediction_error(Y_tr, mf_pred)
_print_float_results(reporter, 'Train Error', sgp_tr_err, mfgp_tr_err)
# Test errors
s_pred, _ = sgp.eval(ZX_te)
mf_pred, _ = mfgp.eval_at_fidel(Z_te, X_te)
sgp_te_err = compute_average_sq_prediction_error(Y_te, s_pred)
mfgp_te_err = compute_average_sq_prediction_error(Y_te, mf_pred)
_print_float_results(reporter, 'Test Error', sgp_te_err, mfgp_te_err)
def _get_ga_optimiser_args(cls,
obj,
num_evals,
mutation_op,
init_pool,
init_pool_vals=None,
expects_inputs_to_be_iterable=True):
""" Returns arguments for the optimiser. """
if expects_inputs_to_be_iterable:
def _obj_wrap(_obj, _x):
""" A wrapper for the optimiser for GA. """
ret = _obj([_x])
return ret[0]
def _get_obj_wrap(_obj):
""" Returns an optimiser for GA. """
return lambda x: _obj_wrap(_obj, x)
obj = _get_obj_wrap(obj)
if init_pool_vals is None:
init_pool_vals = [obj(nn) for nn in init_pool]
reporter = get_reporter('silent')
options = load_options(ga_opt_args, reporter=reporter)
options.pre_eval_points = init_pool
options.pre_eval_vals = init_pool_vals
options.pre_eval_true_vals = init_pool_vals
options.num_mutations_per_epoch = int(
np.clip(3 * np.sqrt(num_evals), 5, 100))
# print 'GA opt args: ', num_evals, options.num_mutations_per_epoch
return obj, options, reporter, mutation_op
def test_initial_sampling(self):
""" Test for initialisation sampling. """
self.report('Testing sample initialisation.')
prob = _get_gpb_problem()
acquisitions = ['mf_gp_ucb', 'gp_ucb', 'gp_ei']
options = option_handler.load_options(
mf_gp_bandit.all_mf_gp_bandit_args, reporter=prob.reporter)
for acq in acquisitions:
options.acq = acq
options.gpb_init_capital = prob.mfof.opt_fidel_cost * 23.2
mfgpb = mf_gp_bandit.MFGPBandit(prob.mfof, options, prob.reporter)
mfgpb.optimise_initialise()
hf_idxs = are_opt_fidel_queries(mfgpb.history.query_fidels,
prob.mfof.opt_fidel)
hf_vals = mfgpb.history.query_vals[hf_idxs]
num_hf_queries = len(hf_vals)
self.report((
'Initialised %s with %d queries (%d at opt_fidel). Init capital = '
+ '%0.4f (%0.4f used) ') %
(acq, len(mfgpb.history.query_vals), num_hf_queries,
options.gpb_init_capital, mfgpb.spent_capital),
'test_result')
assert mfgpb.spent_capital <= 1.1 * options.gpb_init_capital
assert mfgpb.history.curr_opt_vals[-1] == mfgpb.gpb_opt_val
assert is_non_decreasing_sequence(mfgpb.history.curr_opt_vals)
assert num_hf_queries == 0 or hf_vals.max() == mfgpb.gpb_opt_val
assert mfgpb.num_opt_fidel_queries == num_hf_queries
assert mfgpb.history.query_at_opt_fidel.sum() == num_hf_queries
def optimize_chemist(func_caller,
worker_manager,
budget,
mode=None,
init_pool=None,
acq='hei',
options=None,
reporter='default'):
""" Chemist optimization from a function caller. """
if options is None:
reporter = get_reporter(reporter)
options = load_options(all_chemist_args, reporter=reporter)
# TODO: what is this option?
if acq is not None:
options.acq = acq
if mode is not None:
options.mode = mode
# Initial queries
if not hasattr(options,
'pre_eval_points') or options.pre_eval_points is None:
if init_pool is None:
init_pool = get_initial_pool()
options.get_initial_points = lambda n: init_pool[:n]
return (Chemist(func_caller,
worker_manager,
options=options,
reporter=reporter)).optimise(budget)
def __init__(self,
func_caller,
worker_manager,
mutation_op,
crossover_op=None,
options=None,
reporter=None):
""" Constructor.
mutation_op: A function which takes in a list of objects and modifies them.
crossover_op: A function which takes in two objects and performs a cross-over
operation.
So far we have not implemented cross-over but included here in case we want to
include it in the future.
For other arguments, see BlackboxOptimiser
"""
if options is None:
reporter = get_reporter(reporter)
options = load_options(ga_opt_args, reporter=reporter)
super(GAOptimiser, self).__init__(func_caller,
worker_manager,
model=None,
options=options,
reporter=reporter)
self.mutation_op = mutation_op
self.crossover_op = crossover_op
self.to_eval_points = []
def mf_random_optimiser_from_func_caller(func_caller,
worker_manager,
max_capital,
mode,
options=None,
reporter='default',
*args,
**kwargs):
""" Creates a MF EuclideanRandomOptimiser Object and optimises the function. """
reporter = get_reporter(reporter)
if isinstance(func_caller.domain, domains.EuclideanDomain) and \
isinstance(func_caller.fidel_space, domains.EuclideanDomain):
optimiser_constructor = MFEuclideanRandomOptimiser
dflt_list_of_options = mf_euclidean_random_optimiser_args
else:
raise ValueError(
('MF Random optimiser not implemented for (domain, fidel_space) ' +
'of types (%s, %s).') %
(type(func_caller.domain), type(func_caller.fidel_space)))
# Load options
if options is None:
options = load_options(dflt_list_of_options)
options.mode = mode
# Create optimiser
optimiser = optimiser_constructor(func_caller,
worker_manager,
options=options,
reporter=reporter,
*args,
**kwargs)
# optimise and return
return optimiser.optimise(max_capital)
def nasbot(func_caller,
worker_manager,
budget,
tp_comp=None,
mode=None,
init_pool=None,
acq='hei',
options=None,
reporter='default'):
""" NASBOT optimisation from a function caller. """
nn_type = func_caller.domain.nn_type
if options is None:
reporter = get_reporter(reporter)
options = load_options(all_nasbot_args, reporter=reporter)
if acq is not None:
options.acq = acq
if mode is not None:
options.mode = mode
if tp_comp is None:
tp_comp = get_default_otmann_distance(nn_type, 1.0)
# Initial queries
if not hasattr(options,
'pre_eval_points') or options.pre_eval_points is None:
if init_pool is None:
init_pool = get_initial_pool(nn_type)
options.get_initial_points = lambda n: init_pool[:n]
return (NASBOT(func_caller,
worker_manager,
tp_comp,
options=options,
reporter=reporter)).optimise(budget)
def main():
"""
Maximizes a function given a config file containing the hyperparameters and the
corresponding domain bounds.
"""
# Loading Options
euc_gpb_args = get_all_gp_bandit_args_from_gp_args(euclidean_gp_args)
options = load_options(euc_gpb_args + dragonfly_args, cmd_line=True)
if options.config is None:
raise ValueError('Config file is required.')
# Parsing config file
expt_dir = os.path.dirname(
os.path.abspath(os.path.realpath(options.config)))
if not os.path.exists(expt_dir):
raise ValueError("Experiment directory does not exist.")
param_spec = config_parser(options.config)
exp_info = param_spec['exp_info']
obj = imp.load_source(exp_info['name'],
os.path.join(expt_dir, exp_info['name'] + '.py'))
options.capital_type = 'return_value'
if options.max_capital == 0.0:
if options.budget == 0.0:
raise ValueError('Specify the budget in budget or max_capital.')
options.max_capital = options.budget
domain = create_domain(param_spec['domain'])
opt_val, opt_pt, history = maximise_function(
obj.main,
domain=domain,
options=options,
max_capital=options.max_capital)
print('Optimum Value in %d evals: %0.4f' % (options.max_capital, opt_val))
print('Optimum Point: %s' % (opt_pt))
def optimise_with_method_on_func_caller(method,
func_caller,
worker_manager,
max_capital,
options=None,
reporter='default'):
""" A wrapper that can be used to call all optimisation method. """
if options is None:
from gp_bandit import all_gp_bandit_args
reporter = get_reporter(reporter)
options = load_options(all_gp_bandit_args, reporter=reporter)
# The mode should be specified in the first three letters of the method argument
options.mode = method[0:3]
if method in ['synRAND', 'asyRAND']:
return random_optimiser_from_func_caller(func_caller,
worker_manager,
max_capital,
options=options,
reporter=reporter)
elif method in [
'asyTS', 'synTS', 'asyHTS', 'asyUCB', 'asyBUCB', 'synBUCB',
'synUCBPE', 'asyEI'
]:
from gp_bandit import gpb_from_func_caller
options.acq = method[3:]
return gpb_from_func_caller(func_caller,
worker_manager,
max_capital,
options=options,
reporter='default')
else:
raise ValueError('Unknown method %s!' % (method))
def __init__(self, constraint_checker=None, options=None, reporter=None):
""" Constructor. """
self.reporter = get_reporter(reporter)
self.constraint_checker = constraint_checker
if options is None:
options = load_options(nn_modifier_args)
self.options = options
def main():
""" Main function. """
options = load_options(all_mf_gp_bandit_args)
prob, options = get_problem_parameters(options)
# Set other variables
all_methods = NONFINITE_METHODS + FINITE_METHODS
method_options = {key: options for key in all_methods}
noisy_str = 'noiseless' if not NOISY else 'noisy%0.3f' % (
prob.mfof.noise_var)
save_file_prefix = '%s-%s-p%d-d%d' % (prob.experiment_name, noisy_str,
prob.mfof.fidel_dim,
prob.mfof.domain_dim)
reporter = get_reporter('default')
experimenter = MFOptExperimenter(prob.experiment_name,
prob.mfof,
prob.max_capital,
all_methods,
prob.num_experiments,
SAVE_RESULTS_DIR,
save_file_prefix=save_file_prefix,
method_options=method_options,
method_reporter=reporter,
reporter=reporter)
experimenter.run_experiments()
def __init__(self, options, reporter='default'):
""" Constructor. """
super(GPFitter, self).__init__()
self.reporter = get_reporter(reporter)
if isinstance(options, list):
options = load_options(options, 'GP', reporter=self.reporter)
self.options = options
self._set_up()
def _get_euc_gpb_arguments(cls, num_dims):
""" Returns options for GP Bandits on a Euclidean space. """
reporter = reporters.get_reporter('silent')
gpb_args = gp_bandit.get_all_gp_bandit_args_from_gp_args(all_simple_gp_args)
options = load_options(gpb_args, reporter=reporter)
options.get_initial_points = lambda n: np.random.random((n, num_dims))
options.num_init_evals = 20
return options
def gpb_from_func_caller(func_caller,
worker_manager,
max_capital,
is_mf,
mode=None,
acq=None,
mf_strategy=None,
domain_add_max_group_size=-1,
options=None,
reporter='default'):
""" GP Bandit optimisation from a utils.function_caller.FunctionCaller instance.
domain_add_max_group_size indicates whether we should use an additive model or not.
If its negative, then use a non-additive model. If its positive, then use an additive
model with maximum group size as given. If zero, then use the default in options.
"""
# pylint: disable=too-many-branches
reporter = get_reporter(reporter)
if is_mf:
if isinstance(func_caller.fidel_space, domains.EuclideanDomain) and \
isinstance(func_caller.domain, domains.EuclideanDomain):
optimiser_constructor = EuclideanGPBandit
dflt_list_of_options = get_all_mf_gp_bandit_args_from_gp_args(
euclidean_mf_gp_args)
else:
if isinstance(func_caller.domain, domains.EuclideanDomain):
optimiser_constructor = EuclideanGPBandit
dflt_list_of_options = get_all_gp_bandit_args_from_gp_args(
euclidean_gp_args)
else:
raise ValueError('Random optimiser not implemented for domain of type %s.'%( \
type(func_caller.domain)))
# Load options
if options is None:
reporter = get_reporter(reporter)
options = load_options(dflt_list_of_options, reporter=reporter)
if acq is not None:
options.acq = acq
if mode is not None:
options.mode = mode
if mf_strategy is not None:
options.mf_strategy = mf_strategy
# Additive model
if domain_add_max_group_size >= 0:
if is_mf:
options.domain_use_additive_gp = True
if domain_add_max_group_size > 0:
options.domain_add_max_group_size = domain_add_max_group_size
else:
options.use_additive_gp = True
if domain_add_max_group_size > 0:
options.add_max_group_size = domain_add_max_group_size
# create optimiser and return
optimiser = optimiser_constructor(func_caller,
worker_manager,
is_mf=is_mf,
options=options,
reporter=reporter)
return optimiser.optimise(max_capital)
def maximise_function(func,
max_capital,
domain=None,
domain_bounds=None,
config=None,
options=None,
hp_tune_criterion='post_sampling',
hp_tune_method='slice',
init_capital=None,
init_capital_frac=None,
num_init_evals=20):
"""
Maximizes a function given a function and domain bounds of the hyperparameters
and returns optimal value and optimal point.
"""
reporter = get_reporter('default')
if options is None:
euc_gpb_args = get_all_gp_bandit_args_from_gp_args(euclidean_gp_args)
options = load_options(euc_gpb_args)
options.gpb_hp_tune_criterion = hp_tune_criterion
options.gpb_post_hp_tune_method = hp_tune_method
options.init_capital = init_capital
options.init_capital_frac = init_capital_frac
options.num_init_evals = num_init_evals
# Check for Domains
if domain is None:
if config is not None:
param_spec = config_parser(options.config)
domain = create_domain(param_spec['domain'])
elif domain_bounds is not None:
domain = domains.EuclideanDomain(domain_bounds)
else:
raise ValueError(
'Domain or path to config file or domain_bounds have to be given.'
)
# Create worker manager and function caller
worker_manager = SyntheticWorkerManager(num_workers=1,
time_distro='caller_eval_cost')
if isinstance(domain, domains.EuclideanDomain):
func_caller = EuclideanFunctionCaller(func, domain, vectorised=False)
else:
func_caller = FunctionCaller(func, domain)
# Create GPBandit opbject and run optimiser
gpb = EuclideanGPBandit(func_caller,
worker_manager,
reporter=reporter,
options=options)
opt_val, opt_pt, history = gpb.optimise(max_capital)
opt_pt = func_caller.get_raw_domain_coords(opt_pt)
history.curr_opt_points = [
func_caller.get_raw_domain_coords(pt) for pt in history.curr_opt_points
]
history.query_points = [
func_caller.get_raw_domain_coords(pt) for pt in history.query_points
]
return opt_val, opt_pt, history
def __init__(self, X, Y, options=None, reporter=None, *args, **kwargs):
""" Constructor. """
self.X = X
self.Y = Y
self.reporter = get_reporter(reporter)
self.num_data = len(X)
if options is None:
options = load_options(mol_gp_args, 'GPFitter', reporter=reporter)
super(MolGPFitter, self).__init__(options, *args, **kwargs)
def __init__(self, X, Y, options=None, reporter=None):
""" Constructor. """
self.dim = len(X[0])
reporter = get_reporter(reporter)
if options is None:
options = load_options(euclidean_gp_args,
'EuclideanGP',
reporter=reporter)
super(EuclideanGPFitter, self).__init__(X, Y, options, reporter)
def prep_optimiser_args(obj_func, optimiser_args):
""" Returns the options and reporter. """
reporter = get_reporter('default')
options = load_options(optimiser_args, reporter=reporter)
options.pre_eval_points = get_initial_pool()
options.pre_eval_vals = [obj_func(mol) for mol in options.pre_eval_points]
options.pre_eval_true_vals = options.pre_eval_vals
options_clone = deepcopy(options)
return options, options_clone, reporter
def _get_optimiser_args(self, optimiser_args=chemist.all_chemist_args):
""" Returns the options and reporter. """
reporter = get_reporter('default')
options = load_options(optimiser_args, reporter=reporter)
options.pre_eval_points = self.ga_init_pool
options.pre_eval_vals = self.ga_init_vals
options.pre_eval_true_vals = self.ga_init_vals
options_clone = deepcopy(options)
return options, options_clone, reporter
def fit_molgp_with_dataset(dataset, kernel_type):
""" Fits an MolGP to this dataset. """
options = load_options(mol_gp.mol_gp_args, '')
options.kernel_type = kernel_type
gp_fitter = mol_gp.MolGPFitter(dataset[0],
dataset[1],
options=options,
reporter=None)
_, fitted_gp, _ = gp_fitter.fit_gp()
return fitted_gp
def fit_se_gp_with_dataset(dataset):
""" A wrapper to fit a gp using the dataset. """
options = load_options(euclidean_gp_args)
options.kernel_type = 'se'
ret_fit_gp = (EuclideanGPFitter(dataset[0],
dataset[1],
options=options,
reporter=None)).fit_gp()
assert ret_fit_gp[0] == 'fitted_gp'
return ret_fit_gp[1]
def __init__(self, ZZ, XX, YY, options=None, reporter=None):
""" Constructor. options should either be a Namespace, a list or None. """
reporter = get_reporter(reporter)
if options is None:
options = load_options(euclidean_mf_gp_args, 'MF-GP', reporter)
self.fidel_dim = len(ZZ[0])
self.domain_dim = len(XX[0])
self.input_dim = self.fidel_dim + self.domain_dim
super(EuclideanMFGPFitter, self).__init__(ZZ, XX, YY, options,
reporter)
def __init__(self, mf_opt_func, options=None, reporter=None):
""" Constructor. """
self.reporter = get_reporter(reporter)
if options is None:
options = load_options(all_mf_gp_bandit_args, reporter=reporter)
self.options = options
# Set up mfgp and mfof attributes
self.mfof = mf_opt_func # mfof refers to an MFOptFunction object.
self.mfgp = None
# Other set up
self._set_up()
def fit_matern_gp_with_dataset(dataset, nu=-1.0):
""" A wrapper to fit a GP with a matern kernel using the dataset. """
options = load_options(euclidean_gp_args)
options.kernel_type = 'matern'
options.matern_nu = nu
ret_fit_gp = (EuclideanGPFitter(dataset[0],
dataset[1],
options=options,
reporter=None)).fit_gp()
assert ret_fit_gp[0] == 'fitted_gp'
return ret_fit_gp[1]
def __init__(self, ZZ, XX, YY, options=None, reporter=None):
""" Constructor. """
reporter = get_reporter(reporter)
if isinstance(options, list):
options = load_options(options, 'MFGP', reporter=self.reporter)
self.ZZ = ZZ
self.XX = XX
self.YY = YY
self.num_tr_data = len(self.YY)
ZX = get_ZX_from_ZZ_XX(ZZ, XX)
super(MFGPFitter, self).__init__(ZX, YY, options, reporter)
def __init__(self,
func_caller,
worker_manager,
options=None,
reporter=None):
""" Constructor. """
self.reporter = get_reporter(reporter)
if options is None:
options = load_options(all_gp_bandit_args, reporter=reporter)
super(GPBandit, self).__init__(func_caller, worker_manager, options,
self.reporter)
def get_options_and_reporter(method, init_points, init_vals):
""" Returns the options and reporter. """
reporter = get_reporter('default')
if method in ['GA', 'randGA']:
options = load_options(ga_optimiser.ga_opt_args, reporter=reporter)
else:
raise ValueError('Unknown method %s.' % (method))
options.pre_eval_points = init_points
options.pre_eval_vals = init_vals
options.pre_eval_true_vals = init_vals
return options, reporter
def __init__(self,
func_caller,
worker_manager,
options=None,
reporter=None):
""" Constructor. """
self.reporter = get_reporter(reporter)
if options is None:
options = load_options(blackbox_opt_args, reporter=reporter)
super(RandomOptimiser, self).__init__(func_caller, worker_manager,
options, self.reporter)
def __init__(self,
func_caller,
worker_manager,
options=None,
reporter=None):
if options is None:
reporter = get_reporter(reporter)
options = load_options(all_chemist_args, reporter=reporter)
super(Chemist, self).__init__(func_caller,
worker_manager,
options=options,
reporter=reporter)
def fit_nngp_with_dataset(dataset, kernel_type, dist_type):
""" Fits an NNGP to this dataset. """
options = load_options(nn_gp.nn_gp_args, '')
options.kernel_type = kernel_type
options.dist_type = dist_type
gp_fitter = nn_gp.NNGPFitter(dataset[0],
dataset[1],
dataset[-1],
options=options,
reporter=None)
_, fitted_gp, _ = gp_fitter.fit_gp()
return fitted_gp