def __init__(self, config_space, objective_func, R,
num_iter=10, n_workers=1, eta=3, es_gap=9, rho=0.7,
random_state=1, method_id="Default"):
BaseFacade.__init__(self, objective_func, n_workers=n_workers, need_lc=True, method_name=method_id)
self.seed = random_state
self.config_space = config_space
self.config_space.seed(self.seed)
self.R = R
self.num_iter = num_iter
self.eta = eta
self.logeta = lambda x: log(x) / log(self.eta)
self.s_max = int(self.logeta(R))
self.inner_iteration_n = (self.s_max + 1) * (self.s_max + 1)
types, bounds = get_types(config_space)
self.num_config = len(bounds)
self.surrogate = RandomForestWithInstances(types=types, bounds=bounds)
self.acquisition_func = EI(model=self.surrogate)
# TODO: add SMAC's optimization algorithm.
self.acq_optimizer = RandomSampling(self.acquisition_func, config_space,
n_samples=max(500, 50 * self.num_config))
self.incumbent_configs = []
self.incumbent_obj = []
self.lcnet_model = LC_ES()
self.early_stop_gap = es_gap
self.es_rho = rho
self.lc_training_x = None
self.lc_training_y = None
def __init__(self,
config_space: ConfigurationSpace,
objective_func,
R,
num_iter=10000,
eta=3,
p=0.3,
n_workers=1,
random_state=1,
method_id='Default'):
BaseFacade.__init__(self,
objective_func,
n_workers=n_workers,
method_name=method_id)
self.config_space = config_space
self.seed = random_state
self.config_space.seed(self.seed)
self.p = p
self.R = R
self.eta = eta
self.logeta = lambda x: log(x) / log(self.eta)
self.s_max = int(self.logeta(self.R))
self.B = (self.s_max + 1) * self.R
self.num_iter = num_iter
types, bounds = get_types(config_space)
self.num_config = len(bounds)
self.surrogate = RandomForestWithInstances(types=types, bounds=bounds)
self.acquisition_func = EI(model=self.surrogate)
self.acq_optimizer = RandomSampling(self.acquisition_func,
config_space,
n_samples=max(
500, 50 * self.num_config))
self.incumbent_configs = []
self.incumbent_obj = []
def update_weight(self):
max_r = self.iterate_r[-1]
incumbent_configs = self.target_x[max_r]
test_x = convert_configurations_to_array(incumbent_configs)
test_y = np.array(self.target_y[max_r], dtype=np.float64)
r_list = self.weighted_surrogate.surrogate_r
K = len(r_list)
if len(test_y) >= 3:
# Get previous weights
if self.weight_method in [
'rank_loss_softmax', 'rank_loss_single', 'rank_loss_p_norm'
]:
preserving_order_p = list()
preserving_order_nums = list()
for i, r in enumerate(r_list):
fold_num = 5
if i != K - 1:
mean, var = self.weighted_surrogate.surrogate_container[
r].predict(test_x)
tmp_y = np.reshape(mean, -1)
preorder_num, pair_num = MFSE.calculate_preserving_order_num(
tmp_y, test_y)
preserving_order_p.append(preorder_num / pair_num)
preserving_order_nums.append(preorder_num)
else:
if len(test_y) < 2 * fold_num:
preserving_order_p.append(0)
else:
# 5-fold cross validation.
kfold = KFold(n_splits=fold_num)
cv_pred = np.array([0] * len(test_y))
for train_idx, valid_idx in kfold.split(test_x):
train_configs, train_y = test_x[
train_idx], test_y[train_idx]
valid_configs, valid_y = test_x[
valid_idx], test_y[valid_idx]
types, bounds = get_types(self.config_space)
_surrogate = RandomForestWithInstances(
types=types, bounds=bounds)
_surrogate.train(train_configs, train_y)
pred, _ = _surrogate.predict(valid_configs)
cv_pred[valid_idx] = pred.reshape(-1)
preorder_num, pair_num = MFSE.calculate_preserving_order_num(
cv_pred, test_y)
preserving_order_p.append(preorder_num / pair_num)
preserving_order_nums.append(preorder_num)
if self.weight_method == 'rank_loss_softmax':
order_weight = np.array(np.sqrt(preserving_order_nums))
trans_order_weight = order_weight - np.max(order_weight)
# Softmax mapping.
new_weights = np.exp(trans_order_weight) / sum(
np.exp(trans_order_weight))
elif self.weight_method == 'rank_loss_p_norm':
trans_order_weight = np.array(preserving_order_p)
power_sum = np.sum(
np.power(trans_order_weight, self.power_num))
new_weights = np.power(trans_order_weight,
self.power_num) / power_sum
else:
_idx = np.argmax(np.array(preserving_order_nums))
new_weights = [0.] * K
new_weights[_idx] = 1.
elif self.weight_method == 'rank_loss_prob':
# For basic surrogate i=1:K-1.
mean_list, var_list = list(), list()
for i, r in enumerate(r_list[:-1]):
mean, var = self.weighted_surrogate.surrogate_container[
r].predict(test_x)
mean_list.append(np.reshape(mean, -1))
var_list.append(np.reshape(var, -1))
sample_num = 100
min_probability_array = [0] * K
for _ in range(sample_num):
order_preseving_nums = list()
# For basic surrogate i=1:K-1.
for idx in range(K - 1):
sampled_y = np.random.normal(mean_list[idx],
var_list[idx])
_num, _ = MFSE.calculate_preserving_order_num(
sampled_y, test_y)
order_preseving_nums.append(_num)
fold_num = 5
# For basic surrogate i=K. cv
if len(test_y) < 2 * fold_num:
order_preseving_nums.append(0)
else:
# 5-fold cross validation.
kfold = KFold(n_splits=fold_num)
cv_pred = np.array([0] * len(test_y))
for train_idx, valid_idx in kfold.split(test_x):
train_configs, train_y = test_x[train_idx], test_y[
train_idx]
valid_configs, valid_y = test_x[valid_idx], test_y[
valid_idx]
types, bounds = get_types(self.config_space)
_surrogate = RandomForestWithInstances(
types=types, bounds=bounds)
_surrogate.train(train_configs, train_y)
_pred, _var = _surrogate.predict(valid_configs)
sampled_pred = np.random.normal(
_pred.reshape(-1), _var.reshape(-1))
cv_pred[valid_idx] = sampled_pred
_num, _ = MFSE.calculate_preserving_order_num(
cv_pred, test_y)
order_preseving_nums.append(_num)
max_id = np.argmax(order_preseving_nums)
min_probability_array[max_id] += 1
new_weights = np.array(min_probability_array) / sample_num
elif self.weight_method == 'opt_based':
mean_list, var_list = list(), list()
for i, r in enumerate(r_list):
if i != K - 1:
mean, var = self.weighted_surrogate.surrogate_container[
r].predict(test_x)
tmp_y = np.reshape(mean, -1)
tmp_var = np.reshape(var, -1)
mean_list.append(tmp_y)
var_list.append(tmp_var)
else:
if len(test_y) < 8:
mean_list.append(np.array([0] * len(test_y)))
var_list.append(np.array([0] * len(test_y)))
else:
# 5-fold cross validation.
kfold = KFold(n_splits=5)
cv_pred = np.array([0] * len(test_y))
cv_var = np.array([0] * len(test_y))
for train_idx, valid_idx in kfold.split(test_x):
train_configs, train_y = test_x[
train_idx], test_y[train_idx]
valid_configs, valid_y = test_x[
valid_idx], test_y[valid_idx]
types, bounds = get_types(self.config_space)
_surrogate = RandomForestWithInstances(
types=types, bounds=bounds)
_surrogate.train(train_configs, train_y)
pred, var = _surrogate.predict(valid_configs)
cv_pred[valid_idx] = pred.reshape(-1)
cv_var[valid_idx] = var.reshape(-1)
mean_list.append(cv_pred)
var_list.append(cv_var)
means = np.array(mean_list)
vars = np.array(var_list) + 1e-8
def min_func(x):
x = np.reshape(np.array(x), (1, len(x)))
ensemble_vars = 1 / (x @ (1 / vars))
ensemble_means = x @ (means / vars) * ensemble_vars
ensemble_means = np.reshape(ensemble_means, -1)
self.logger.info("Loss:" + str(x))
return MFSE.calculate_ranking_loss(ensemble_means, test_y)
constraints = [{
'type': 'eq',
'fun': lambda x: np.sum(x) - 1
}, {
'type': 'ineq',
'fun': lambda x: x - 0
}, {
'type': 'ineq',
'fun': lambda x: 1 - x
}]
res = minimize(min_func,
np.array([1e-8] * K),
constraints=constraints)
new_weights = res.x
else:
raise ValueError('Invalid weight method: %s!' %
self.weight_method)
else:
old_weights = list()
for i, r in enumerate(r_list):
_weight = self.weighted_surrogate.surrogate_weight[r]
old_weights.append(_weight)
new_weights = old_weights.copy()
self.logger.info(
'[%s] %d-th Updating weights: %s' %
(self.weight_method, self.weight_changed_cnt, str(new_weights)))
# Assign the weight to each basic surrogate.
for i, r in enumerate(r_list):
self.weighted_surrogate.surrogate_weight[r] = new_weights[i]
self.weight_changed_cnt += 1
# Save the weight data.
self.hist_weights.append(new_weights)
np.save(
'data/%s_weights_%s.npy' % (self.method_name, self.method_name),
np.asarray(self.hist_weights))
class SMAC_ES(BaseFacade):
def __init__(self, config_space, objective_func, R,
num_iter=10, n_workers=1, eta=3, es_gap=9, rho=0.7,
random_state=1, method_id="Default"):
BaseFacade.__init__(self, objective_func, n_workers=n_workers, need_lc=True, method_name=method_id)
self.seed = random_state
self.config_space = config_space
self.config_space.seed(self.seed)
self.R = R
self.num_iter = num_iter
self.eta = eta
self.logeta = lambda x: log(x) / log(self.eta)
self.s_max = int(self.logeta(R))
self.inner_iteration_n = (self.s_max + 1) * (self.s_max + 1)
types, bounds = get_types(config_space)
self.num_config = len(bounds)
self.surrogate = RandomForestWithInstances(types=types, bounds=bounds)
self.acquisition_func = EI(model=self.surrogate)
# TODO: add SMAC's optimization algorithm.
self.acq_optimizer = RandomSampling(self.acquisition_func, config_space,
n_samples=max(500, 50 * self.num_config))
self.incumbent_configs = []
self.incumbent_obj = []
self.lcnet_model = LC_ES()
self.early_stop_gap = es_gap
self.es_rho = rho
self.lc_training_x = None
self.lc_training_y = None
def iterate(self):
for _ in range(self.inner_iteration_n):
T = self.choose_next(self.num_workers)
extra_info = None
lc_info = dict()
lc_conf_mapping = dict()
# assume no same configuration in the same batch: T
for item in T:
conf_id = get_configuration_id(item.get_dictionary())
sha = hashlib.sha1(conf_id.encode('utf8'))
conf_id = sha.hexdigest()
lc_conf_mapping[conf_id] = item
total_iter_num = self.R // self.early_stop_gap
for iter_num in range(1, 1 + total_iter_num):
self.logger.info('start iteration gap %d' % iter_num)
ret_val, early_stops = self.run_in_parallel(T, self.early_stop_gap, extra_info)
val_losses = [item['loss'] for item in ret_val]
ref_list = [item['ref_id'] for item in ret_val]
for item in ret_val:
conf_id = item['ref_id']
if not self.restart_needed:
if conf_id not in lc_info:
lc_info[conf_id] = []
lc_info[conf_id].extend(item['lc_info'])
else:
lc_info[conf_id] = item['lc_info']
if iter_num == total_iter_num:
self.incumbent_configs.extend(T)
self.incumbent_obj.extend(val_losses)
T = [config for i, config in enumerate(T) if not early_stops[i]]
extra_info = [ref for i, ref in enumerate(ref_list) if not early_stops[i]]
if len(T) == 0:
break
if len(self.incumbent_obj) >= 2 * self.num_config and iter_num != total_iter_num:
# learning curve based early stop strategy.
ref_list = extra_info
early_stops = self.stop_early(T)
T = [config for i, config in enumerate(T) if not early_stops[i]]
extra_info = [ref for i, ref in enumerate(ref_list) if not early_stops[i]]
if len(T) == 0:
break
# keep learning curve data
for item, config in lc_conf_mapping.items():
lc_data = lc_info[item]
if len(lc_data) > 0:
n_epochs = len(lc_data)
# self.logger.info('insert one learning curve data into dataset.')
t_idx = np.arange(1, n_epochs + 1) / n_epochs
conf_data = convert_configurations_to_array([config])
x = np.repeat(conf_data, t_idx.shape[0], axis=0)
x = np.concatenate((x, t_idx[:, None]), axis=1)
y = np.array(lc_data)
if self.lc_training_x is None:
self.lc_training_x, self.lc_training_y = x, y
else:
self.lc_training_x = np.concatenate((self.lc_training_x, x), 0)
self.lc_training_y = np.concatenate((self.lc_training_y, y), 0)
self.logger.info('training data shape: %s' % str(self.lc_training_x.shape))
if len(self.incumbent_obj) >= 2 * self.num_config and len(self.incumbent_obj) % self.num_config == 0:
self.lcnet_model.train(self.lc_training_x, self.lc_training_y)
self.add_stage_history(self.stage_id, self.global_incumbent)
self.stage_id += 1
self.remove_immediate_model()
@BaseFacade.process_manage
def run(self):
try:
for iter in range(self.num_iter):
self.logger.info('-' * 50)
self.logger.info("SMAC with ES algorithm: %d/%d iteration starts" % (iter, self.num_iter))
start_time = time.time()
self.iterate()
time_elapsed = (time.time() - start_time) / 60
self.logger.info("iteration took %.2f min." % time_elapsed)
self.save_intemediate_statistics()
except Exception as e:
print(e)
self.logger.error(str(e))
# clear the immediate result.
self.remove_immediate_model()
def stop_early(self, T):
configs_data = convert_configurations_to_array(T)
x_test = None
for i in range(configs_data.shape[0]):
x = np.concatenate((configs_data[i, None, :], np.array([[1.0]])), axis=1)
if x_test is None:
x_test = x
else:
x_test = np.concatenate((x_test, x), 0)
m, v = self.lcnet_model.predict(x_test)
best_accuracy = 1 - self.global_incumbent
s = np.sqrt(v)
less_p = norm.cdf((best_accuracy - m) / s)
self.logger.info('early stop prob: %s' % str(less_p))
early_stop_flag = (less_p >= self.es_rho)
self.logger.info('early stop vector: %s' % str(early_stop_flag))
for i, flag in enumerate(early_stop_flag):
if flag:
self.incumbent_configs.append(T[i])
self.incumbent_obj.append(m[i])
return early_stop_flag
def choose_next(self, num_config):
if len(self.incumbent_obj) < 2 * self.num_config:
return sample_configurations(self.config_space, num_config)
# print('choose next starts!')
self.logger.info('train feature is: %s' % str(self.incumbent_configs[-5:]))
self.logger.info('train target is: %s' % str(self.incumbent_obj))
self.surrogate.train(convert_configurations_to_array(self.incumbent_configs),
np.array(self.incumbent_obj, dtype=np.float64))
conf_cnt = 0
total_cnt = 0
next_configs = []
while conf_cnt < num_config and total_cnt < 5 * num_config:
incumbent = dict()
best_index = np.argmin(self.incumbent_obj)
incumbent['obj'] = self.incumbent_obj[best_index]
incumbent['config'] = self.incumbent_configs[best_index]
self.acquisition_func.update(model=self.surrogate, eta=incumbent)
rand_config = self.acq_optimizer.maximize(batch_size=1)[0]
if rand_config not in next_configs:
next_configs.append(rand_config)
conf_cnt += 1
total_cnt += 1
if conf_cnt < num_config:
next_configs = expand_configurations(next_configs, self.config_space, num_config)
return next_configs
def get_incumbent(self, num_inc=1):
assert (len(self.incumbent_obj) == len(self.incumbent_configs))
indices = np.argsort(self.incumbent_obj)
configs = [self.incumbent_configs[i] for i in indices[0:num_inc]]
targets = [self.incumbent_obj[i] for i in indices[0: num_inc]]
return configs, targets
class BOHB(BaseFacade):
""" The implementation of BOHB.
The paper can be found in https://arxiv.org/abs/1807.01774 .
"""
def __init__(self,
config_space: ConfigurationSpace,
objective_func,
R,
num_iter=10000,
eta=3,
p=0.3,
n_workers=1,
random_state=1,
method_id='Default'):
BaseFacade.__init__(self,
objective_func,
n_workers=n_workers,
method_name=method_id)
self.config_space = config_space
self.seed = random_state
self.config_space.seed(self.seed)
self.p = p
self.R = R
self.eta = eta
self.logeta = lambda x: log(x) / log(self.eta)
self.s_max = int(self.logeta(self.R))
self.B = (self.s_max + 1) * self.R
self.num_iter = num_iter
types, bounds = get_types(config_space)
self.num_config = len(bounds)
self.surrogate = RandomForestWithInstances(types=types, bounds=bounds)
self.acquisition_func = EI(model=self.surrogate)
self.acq_optimizer = RandomSampling(self.acquisition_func,
config_space,
n_samples=max(
500, 50 * self.num_config))
self.incumbent_configs = []
self.incumbent_obj = []
def iterate(self, skip_last=0):
for s in reversed(range(self.s_max + 1)):
# Set initial number of configurations
n = int(ceil(self.B / self.R / (s + 1) * self.eta**s))
# Set initial number of iterations per config
r = self.R * self.eta**(-s)
# Sample n configurations according to BOHB strategy.
T = self.choose_next(n)
extra_info = None
last_run_num = None
for i in range((s + 1) - int(skip_last)): # changed from s + 1
# Run each of the n configs for <iterations>
# and keep best (n_configs / eta) configurations.
n_configs = n * self.eta**(-i)
n_iterations = r * self.eta**(i)
n_iter = n_iterations
if last_run_num is not None and not self.restart_needed:
n_iter -= last_run_num
last_run_num = n_iterations
self.logger.info(
"BOHB: %d configurations x %d iterations each" %
(int(n_configs), int(n_iterations)))
ret_val, early_stops = self.run_in_parallel(
T, n_iter, extra_info)
val_losses = [item['loss'] for item in ret_val]
ref_list = [item['ref_id'] for item in ret_val]
if int(n_iterations) == self.R:
self.incumbent_configs.extend(T)
self.incumbent_obj.extend(val_losses)
# Select a number of best configurations for the next loop.
# Filter out early stops, if any.
indices = np.argsort(val_losses)
if len(T) == sum(early_stops):
break
if len(T) >= self.eta:
T = [T[i] for i in indices if not early_stops[i]]
extra_info = [
ref_list[i] for i in indices if not early_stops[i]
]
reduced_num = int(n_configs / self.eta)
T = T[0:reduced_num]
extra_info = extra_info[0:reduced_num]
else:
T = [T[indices[0]]]
extra_info = [ref_list[indices[0]]]
incumbent_loss = val_losses[indices[0]]
self.add_stage_history(
self.stage_id, min(self.global_incumbent, incumbent_loss))
self.stage_id += 1
self.remove_immediate_model()
@BaseFacade.process_manage
def run(self):
try:
for iter in range(self.num_iter):
self.logger.info('-' * 50)
self.logger.info("BOHB algorithm: %d/%d iteration starts" %
(iter, self.num_iter))
start_time = time.time()
self.iterate()
time_elapsed = (time.time() - start_time) / 60
self.logger.info("Iteration took %.2f min." % time_elapsed)
self.save_intemediate_statistics()
except Exception as e:
print(e)
self.logger.error(str(e))
# Clean the immediate result.
self.remove_immediate_model()
def choose_next(self, num_config):
if len(self.incumbent_obj) < 3:
return sample_configurations(self.config_space, num_config)
self.logger.info('Train feature is: %s' %
str(self.incumbent_configs[:5]))
self.logger.info('Train target is: %s' % str(self.incumbent_obj))
self.surrogate.train(
convert_configurations_to_array(self.incumbent_configs),
np.array(self.incumbent_obj, dtype=np.float64))
config_cnt = 0
total_sample_cnt = 0
config_candidates = []
while config_cnt < num_config and total_sample_cnt < 3 * num_config:
if random.random() < self.p:
rand_config = self.config_space.sample_configuration(1)
else:
# print('use surrogate to produce candidate.')
incumbent = dict()
best_index = np.argmin(self.incumbent_obj)
incumbent['obj'] = self.incumbent_obj[best_index]
incumbent['config'] = self.incumbent_configs[best_index]
self.acquisition_func.update(model=self.surrogate,
eta=incumbent)
rand_config = self.acq_optimizer.maximize(batch_size=1)[0]
if rand_config not in config_candidates:
config_candidates.append(rand_config)
config_cnt += 1
total_sample_cnt += 1
if config_cnt < num_config:
config_candidates = expand_configurations(config_candidates,
self.config_space,
num_config)
return config_candidates
def get_incumbent(self, num_inc=1):
assert (len(self.incumbent_obj) == len(self.incumbent_configs))
indices = np.argsort(self.incumbent_obj)
configs = [self.incumbent_configs[i] for i in indices[0:num_inc]]
targets = [self.incumbent_obj[i] for i in indices[0:num_inc]]
return configs, targets