def obtain_latin_hypercube(self):
hps = self.hp_utils.config_space._hyperparameters
hp_idx = {"numerical": [], "categorical": []}
choices = []
for var_name, hp in hps.items():
idx = self.hp_utils.config_space._hyperparameter_idx[var_name]
dist = utils.distribution_type(self.hp_utils.config_space,
var_name)
if dist is str or dist is bool:
hp_idx["categorical"].append(idx)
choices.append(hp.choices)
else:
hp_idx["numerical"].append(idx)
n_num_dim = len(hp_idx["numerical"])
n_dim = n_num_dim + len(hp_idx["categorical"])
rnd_grid = np.array([
self.rng.permutation(list(range(1, self.n_points + 1)))
for _ in range(n_num_dim)
])
hp_confs = [[
self.value_in_grid(rnd_grid, hp_idx, i, k, choices)
for i in range(n_dim)
] for k in range(self.n_points)]
hp_confs = self.hp_utils.revert_hp_confs(hp_confs)
return hp_confs
def _initial_sampler(self):
"""
random sampling for an initialization
Returns
-------
hyperparameter configurations: list
"""
hps = self.hp_utils.config_space._hyperparameters
sample = [None for _ in range(len(hps))]
for var_name, hp in hps.items():
idx = self.hp_utils.config_space._hyperparameter_idx[var_name]
dist = utils.distribution_type(self.hp_utils.config_space, var_name)
if dist is str or dist is bool:
# categorical
choices = hp.choices
rnd = self.rng.randint(len(choices))
sample[idx] = choices[rnd]
else:
# numerical
rnd = self.rng.uniform()
sample[idx] = self.hp_utils.revert_hp(rnd, var_name)
return sample
def sample(self):
hps_conf, _ = self.hp_utils.load_hps_conf(convert=True,
do_sort=True,
index_from_conf=False)
hp_conf = []
for idx, hps in enumerate(hps_conf):
n_lower = self.gamma_func(len(hps))
lower_hps, upper_hps = hps[:n_lower], hps[n_lower:]
var_name = self.hp_utils.config_space._idx_to_hyperparameter[idx]
var_type = utils.distribution_type(self.hp_utils.config_space,
var_name)
if var_type in [float, int]:
pe_lower, pe_upper = self._construct_numerical_parzen_estimator(
var_name, var_type, lower_hps, upper_hps)
hp_value = self._compare_candidates(pe_lower, pe_upper)
else:
pe_lower, pe_upper, choices = self._construct_categorical_parzen_estimator(
var_name, lower_hps, upper_hps)
hp_value = self._compare_candidates(pe_lower, pe_upper,
choices)
hp_conf.append(hp_value)
return self.hp_utils.revert_hp_conf(hp_conf)
def sample(self):
hps_conf, _ = self.hp_utils.load_hps_conf(convert=True,
do_sort=True,
index_from_conf=False)
hp_confs = [] # ndarray (D, n_ei_candidates)
choices_list = []
n_evals = len(hps_conf[0])
n_lower = self.gamma_func(n_evals)
basis_loglikelihoods_lower = np.zeros(
(len(hps_conf), n_lower + 1, self.n_ei_candidates))
basis_loglikelihoods_upper = np.zeros(
(len(hps_conf), n_evals - n_lower + 1, self.n_ei_candidates))
for idx, hps in enumerate(hps_conf):
lower_hps, upper_hps = hps[:n_lower], hps[n_lower:]
var_name = self.hp_utils.config_space._idx_to_hyperparameter[idx]
var_type = utils.distribution_type(self.hp_utils.config_space,
var_name)
if var_type in [float, int]:
pe_lower, pe_upper = self._construct_numerical_parzen_estimator(
var_name, var_type, lower_hps, upper_hps)
choices_list.append(None)
else:
pe_lower, pe_upper, choices = self._construct_categorical_parzen_estimator(
var_name, lower_hps, upper_hps)
choices_list.append(choices)
samples_lower = pe_lower.sample_from_density_estimator(
self.rng, self.n_ei_candidates)
hp_confs.append(samples_lower)
basis_loglikelihoods_lower[idx] += pe_lower.basis_loglikelihood(
samples_lower)
basis_loglikelihoods_upper[idx] += pe_upper.basis_loglikelihood(
samples_lower)
hp_conf = self._compare_configurations(basis_loglikelihoods_lower,
basis_loglikelihoods_upper,
hp_confs, choices_list)
return self.hp_utils.revert_hp_conf(hp_conf)
def sample(self):
hps_conf, _ = self.hp_utils.load_hps_conf(convert=True,
do_sort=True,
index_from_conf=False)
hp_conf = []
for idx, hps in enumerate(hps_conf):
n_lower = self.gamma_func(len(hps))
lower_vals, upper_vals = hps[:n_lower], hps[n_lower:]
var_name = self.hp_utils.config_space._idx_to_hyperparameter[idx]
var_type = utils.distribution_type(self.hp_utils.config_space,
var_name)
if var_type in [float, int]:
hp_value = self._sample_numerical(var_name, var_type,
lower_vals, upper_vals)
else:
hp_value = self._sample_categorical(var_name, lower_vals,
upper_vals)
hp_conf.append(hp_value)
return self.hp_utils.revert_hp_conf(hp_conf)