def mini_smac(learn_delta):
sample_num_m = s_mid
sample_num_l = s_min
if not learn_delta:
sample_num_m = s_min
start_time = time.time()
config_space = create_configspace()
types, bounds = get_types(config_space)
num_hp = len(bounds)
surrogate = RandomForestWithInstances(types=types, bounds=bounds)
acquisition_func = EI(model=surrogate)
acq_optimizer = RandomSampling(acquisition_func, config_space, n_samples=max(500, 50 * num_hp))
X = []
y = []
y_delta = []
c = []
inc_y = 1.
# Initial design.
for _ in range(num_init):
init_configs = sample_configurations(config_space, num_init)
for config in init_configs:
perf_t, _ = objective_function((config.get_dictionary(), sample_num_m))
X.append(config)
y.append(perf_t)
if perf_t < inc_y:
inc_y = perf_t
c.append([time.time()-start_time, inc_y])
if learn_delta:
perf_l, _ = objective_function((config.get_dictionary(), sample_num_l))
y_delta.append(perf_t - perf_l)
else:
y_delta.append(perf_t)
# BO iterations.
for _ in range(num_iter - num_init):
# Update the surrogate model.
surrogate.train(convert_configurations_to_array(X), np.array(y, dtype=np.float64))
# Use EI acq to choose next config.
incumbent = dict()
best_index = np.argmin(y)
incumbent['obj'] = y[best_index]
incumbent['config'] = X[best_index]
acquisition_func.update(model=surrogate, eta=incumbent)
next_config = acq_optimizer.maximize(batch_size=1)[0]
perf_t, _ = objective_function((next_config.get_dictionary(), sample_num_m))
X.append(next_config)
y.append(perf_t)
if perf_t < inc_y:
inc_y = perf_t
c.append([time.time() - start_time, inc_y])
if learn_delta:
perf_l, _ = objective_function((config.get_dictionary(), sample_num_l))
y_delta.append(perf_t - perf_l)
else:
y_delta.append(perf_t)
return [convert_configurations_to_array(X), np.array(y_delta, dtype=np.float64)]
def choose_next(self, num_config):
if len(self.target_y[self.iterate_r[-1]]) == 0:
return sample_configurations(self.config_space, num_config)
conf_cnt = 0
next_configs = []
total_cnt = 0
incumbent = dict()
max_r = self.iterate_r[-1]
best_index = np.argmin(self.target_y[max_r])
incumbent['config'] = self.target_x[max_r][best_index]
approximate_obj = self.weighted_surrogate.predict(
convert_configurations_to_array([incumbent['config']]))[0]
incumbent['obj'] = approximate_obj
self.weighted_acquisition_func.update(model=self.weighted_surrogate,
eta=incumbent)
while conf_cnt < num_config and total_cnt < 2 * num_config:
rand_config = self.weighted_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 update_weight_vector(self):
rho = self.rho
max_r = self.iterate_r[-1]
incumbent_configs = self.target_x[max_r]
test_x = convert_configurations_to_array(incumbent_configs)
test_y = minmax_normalization(self.target_y[max_r])
r_list = self.weighted_surrogate.surrogate_r
cur_confidence = self.weighted_surrogate.surrogate_weight
curr_list = [cur_confidence[r] for r in r_list[:-1]]
# calculate correlation coefficient.
corrcoef_list = []
for i, r in enumerate(r_list):
mean, _ = self.weighted_surrogate.surrogate_container[r].predict(
test_x)
tmp_y = np.reshape(mean, -1)
# corrcoef = np.corrcoef(np.vstack((test_y, tmp_y)))[0][1]/2 + 0.5
corrcoef = max(0, np.corrcoef(np.vstack((test_y, tmp_y)))[0][1])
corrcoef_list.append(corrcoef)
corrcoef_list = np.array(corrcoef_list)
self.logger.info('scale method %d, before normalization: %s' %
(self.scale_method, str(corrcoef_list)))
if self.scale_method <= 4:
corrcoef_list = corrcoef_list**self.scale_method / sum(
corrcoef_list**self.scale_method)
elif self.scale_method == 6:
corrcoef_list = corrcoef_list**2 / sum(corrcoef_list**2)
elif self.scale_method == 7:
ref = corrcoef_list**2
follow_list = (corrcoef_list != max(corrcoef_list)) * ref
corrcoef_list = follow_list / sum(follow_list) * 0.5 + (
corrcoef_list == max(corrcoef_list)) * 0.5
self.logger.info('after normalization: %s' % str(corrcoef_list))
if sum(np.isnan(corrcoef_list)) == len(corrcoef_list):
corrcoef_list = [item for item in curr_list]
self.logger.info('escape nan, current update', corrcoef_list)
assert len(cur_confidence) == len(corrcoef_list)
self.logger.info('cur rho %.3f; conf vector/update vector: %s/%s' %
(self.rho, str(cur_confidence), str(corrcoef_list)))
updated_weights = list()
for i, r in enumerate(r_list):
if self.scale_method == 7:
self.weighted_surrogate.surrogate_weight[r] = corrcoef_list[i]
else:
self.weighted_surrogate.surrogate_weight[
r] = corrcoef_list[i] * (1 - rho) + rho * cur_confidence[r]
updated_weights.append(self.weighted_surrogate.surrogate_weight[r])
print('update surrogate weight:', r,
self.weighted_surrogate.surrogate_weight[r])
self.logger.info('update surrogate weight:%d-%.4f' %
(r, self.weighted_surrogate.surrogate_weight[r]))
self.hist_weights.append(updated_weights)
np.save('data/tmp_weights_%s.npy' % self.method_name,
np.asarray(self.hist_weights))
def choose_next_weighted(self, num_config):
if len(self.target_y[self.iterate_r[-1]]) == 0:
return sample_configurations(self.config_space, num_config)
conf_cnt = 0
next_configs = []
total_cnt = 0
while conf_cnt < num_config and total_cnt < 2 * num_config:
# in Bayesian optimization, eliminate epsilon sampling.
incumbent = dict()
# TODO: problem-->use the best in maximal resource.
# TODO: smac's optmization algorithm.
max_r = self.iterate_r[-1]
best_index = np.argmin(self.target_y[max_r])
incumbent['config'] = self.target_x[max_r][best_index]
approximate_obj = self.weighted_surrogate.predict(
convert_configurations_to_array([incumbent['config']]))[0]
incumbent['obj'] = approximate_obj
self.weighted_acquisition_func.update(
model=self.weighted_surrogate, eta=incumbent)
rand_config = self.weighted_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 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 choose_next(self, num_config):
self.logger.info('LCNet: model-based choosing.')
if len(self.incumbent_obj) <= 0:
return sample_configurations(self.config_space, num_config)
self.logger.info('start to training LCNet, training data shape: %s' % str(self.lc_training_x.shape))
self.lcnet_model.train(self.lc_training_x, self.lc_training_y)
next_configs = []
random_configs = sample_configurations(self.config_space, 50 * self.num_config)
random_configs_data = convert_configurations_to_array(random_configs)
x_test = None
for i in range(random_configs_data.shape[0]):
x = np.concatenate((random_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)
sorted_configs = [random_configs[i] for i in np.argsort(-m)]
print(sorted_configs[:5])
number_flag = False
for config in sorted_configs:
if config not in next_configs:
next_configs.append(config)
if len(next_configs) == num_config:
number_flag = True
break
if not number_flag:
next_configs = expand_configurations(next_configs, self.config_space, num_config)
self.logger.warning('MBHB: add random configuration here.' + '=' * 50)
return next_configs
def choose_next(self, num_config, r, mode):
# different types of mode.
if mode == 'Hybrid':
mode = 'Backward' if self.iterate_id % 2 == 0 else 'Forward'
if mode == 'Forward':
if r != self.R:
r *= self.eta
elif mode == 'Backward':
if r != 1:
r /= self.eta
else:
r = self.R
# TODO: in different types, this condition may not needed any more.
n_exp = len(self.target_y[r])
if n_exp < 2 * self.num_config:
return sample_configurations(self.config_space, num_config)
self.logger.info('train feature is: %s' % str(self.target_x[r]))
self.logger.info('train target is: %s' % str(self.target_y[r]))
self.surrogate.train(convert_configurations_to_array(self.target_x[r]),
np.array(self.target_y[r], dtype=np.float64))
conf_cnt = 0
next_configs = []
total_cnt = 0
# TODO: acceleration, maximize a batch of candidates.
while conf_cnt < num_config and total_cnt < 5 * num_config:
rand_config = None
if random.uniform(0, 1) < self.init_tradeoff:
rand_config = self.config_space.sample_configuration(1)
else:
# print('use surrogate to produce candidate.')
incumbent = dict()
incumbent['obj'] = np.min(self.target_y[r])
incumbent['config'] = self.target_x[r][np.argmin(
self.target_y[r])]
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 choose_next(self, num_config):
if len(self.incumbent_obj) < 3:
return sample_configurations(self.config_space, num_config)
self.logger.info('BO Training - X: %s' %
str(self.incumbent_configs[-5:]))
self.logger.info('BO Training - Y: %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)
next_configs = []
# Epsilon greedy
for config in _next_configs:
if random.random() < self.p:
next_configs.append(
sample_configurations(self.config_space, 1)[0])
else:
next_configs.append(config)
return next_configs
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) < 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 update_weight(self):
max_r = self.iterate_r[-1]
r_list = self.iterate_r
incumbent_configs = self.target_x[max_r]
test_x = convert_configurations_to_array(incumbent_configs)
test_y = minmax_normalization(self.target_y[max_r])
predictions = []
for i, r in enumerate(r_list[:-1]):
mean, _ = self.weighted_surrogate.surrogate_container[r].predict(
test_x)
predictions.append(mean.flatten().tolist())
predictions.append(
self.obtain_cv_prediction(test_x, np.array(test_y,
dtype=np.float64)))
solution, status = self.solve_optpro(
np.mat(predictions).T,
np.mat(test_y).T)
if status:
solution[solution < 1e-3] = 0.
self.logger.info('New weight: %s' % str(solution))
for i, r in enumerate(r_list):
self.weighted_surrogate.surrogate_weight[r] = solution[i]
def iterate(self, skip_last=0):
for s in reversed(range(self.s_max + 1)):
if self.update_enable and self.weight_update_id > self.s_max:
self.update_weight_vector()
self.weight_update_id += 1
# initial number of configurations
n = int(ceil(self.B / self.R / (s + 1) * self.eta**s))
# initial number of iterations per config
r = int(self.R * self.eta**(-s))
# choose a batch of configurations in different mechanisms.
start_time = time.time()
if self.info_type != 'Weighted':
T = self.choose_next(n, r, self.info_type)
else:
T = self.choose_next_weighted(n)
time_elapsed = time.time() - start_time
self.logger.info("choosing next configurations took %.2f sec." %
time_elapsed)
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(
"XFHB-%s: %d configurations x %d iterations each" %
(self.info_type, 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]
self.target_x[int(n_iterations)].extend(T)
self.target_y[int(n_iterations)].extend(val_losses)
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()
if self.info_type == 'Weighted':
for item in self.iterate_r[self.iterate_r.index(r):]:
# objective value normalization: min-max linear normalization
normalized_y = minmax_normalization(self.target_y[item])
self.weighted_surrogate.train(
convert_configurations_to_array(self.target_x[item]),
np.array(normalized_y, dtype=np.float64),
r=item)
# TODO: trade off value: decay (bayesian optimization did? do we need trade off e&e again?)
self.init_tradeoff *= self.tradeoff_dec_rate
def iterate(self, skip_last=0):
for s in reversed(range(self.s_max + 1)):
# initial number of configurations
n = int(ceil(self.B / self.R / (s + 1) * self.eta ** s))
# initial number of iterations per config
r = self.R * self.eta ** (-s)
# n random configurations
T = self.choose_next(n)
extra_info = None
last_run_num = 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
for i in range((s + 1) - int(skip_last)):
# 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("MBHB: %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]
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 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()
# 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)
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))
def iterate(self, skip_last=0):
for s in reversed(range(self.s_max + 1)):
if self.update_enable and self.weight_update_id > self.s_max:
self.update_weight()
self.weight_update_id += 1
# Set initial number of configurations
n = int(ceil(self.B / self.R / (s + 1) * self.eta**s))
# initial number of iterations per config
r = int(self.R * self.eta**(-s))
# Choose a batch of configurations in different mechanisms.
start_time = time.time()
T = self.choose_next_batch(n)
time_elapsed = time.time() - start_time
self.logger.info(
"[%s] Choosing next configurations took %.2f sec." %
(self.method_name, time_elapsed))
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(
"MFSE: %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]
self.target_x[int(n_iterations)].extend(T)
self.target_y[int(n_iterations)].extend(val_losses)
if int(n_iterations) == self.R:
self.incumbent_configs.extend(T)
self.incumbent_perfs.extend(val_losses)
# Update history container.
for _config, _perf in zip(T, val_losses):
self.history_container.add(_config, _perf)
# 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()
for item in self.iterate_r[self.iterate_r.index(r):]:
# NORMALIZE Objective value: normalization
normalized_y = std_normalization(self.target_y[item])
self.weighted_surrogate.train(convert_configurations_to_array(
self.target_x[item]),
np.array(normalized_y,
dtype=np.float64),
r=item)
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()
def iterate(self, skip_last=0):
for s in reversed(range(self.s_max + 1)):
# Initial number of configurations
n = int(ceil(self.B / self.R / (s + 1) * self.eta**s))
# Initial number of iterations per config
r = int(self.R * self.eta**(-s))
# Choose a batch of configurations in different mechanisms.
start_time = time.time()
T = self.choose_next(n)
time_elapsed = time.time() - start_time
self.logger.info("Choosing next configurations took %.2f sec." %
time_elapsed)
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(
"HOIST: %d configurations WITH %d units of resource" %
(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]
self.target_x[int(n_iterations)].extend(T)
self.target_y[int(n_iterations)].extend(val_losses)
if int(n_iterations) == self.R:
self.incumbent_configs.extend(T)
self.incumbent_obj.extend(val_losses)
# Select a number of well-performed configurations for the next loop.
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()
# Augment the intermediate evaluation data.
for item in self.iterate_r[self.iterate_r.index(r):]:
# objective value normalization: min-max linear normalization
normalized_y = minmax_normalization(self.target_y[item])
self.weighted_surrogate.train(convert_configurations_to_array(
self.target_x[item]),
np.array(normalized_y,
dtype=np.float64),
r=item)
# Update the parameter in the ensemble model.
if len(self.target_y[self.iterate_r[-1]]) >= 2:
self.update_weight()
def tse(run_id, train_base_models=True):
start_time = time.time()
from concurrent.futures import ProcessPoolExecutor
pool = ProcessPoolExecutor(max_workers=args.worker)
X, y = [], []
c = []
inc = 1.
X_l, y_l = [], []
weight = np.array([1/K]*(K+1))
config_evaluated = []
config_space = create_configspace()
# Initialize config L.
config_L = sample_configurations(config_space, num_L_init)
if train_base_models:
func_configs = list()
for iter_t in range(K):
print('Build mid fidelity model', iter_t)
func_configs.append(True)
func_configs.append(False)
training_data = run_parallel_async(pool, mini_smac, func_configs)
with open('data/xgb/base_tse_data_%d.pkl' % run_id, 'wb') as f:
pickle.dump(training_data, f)
else:
with open('data/xgb/base_tse_data_%d.pkl' % 10, 'rb') as f:
training_data = pickle.load(f)
print('Load training data for M evaluations!')
# Create base models.
base_models = list()
config_space = create_configspace()
types, bounds = get_types(config_space)
for iter_t in range(K+1):
config_x, config_y = training_data[iter_t]
model = RandomForestWithInstances(types=types, bounds=bounds)
model.train(config_x, config_y)
base_models.append(model)
low_fidelity_model = base_models[K]
X_l.extend(training_data[K][0].tolist())
y_l.extend(training_data[K][1].tolist())
print('Base model building finished!')
# The framework of TSE.
for iter_t in range(iter_H):
print('Iteration in TSE', iter_t)
# Sample a batch of configurations according to tse model.
configs = sample_configurations(config_space, iter_L * 10)
config_arrays = convert_configurations_to_array(configs)
perfs, _ = low_fidelity_model.predict(config_arrays)
perfs = perfs[:, 0]
if len(y) > 3:
preds = []
for i in range(K):
m, _ = base_models[i].predict(config_arrays)
preds.append(m[:, 0].tolist())
preds = np.array(preds).T
preds = np.mat(np.hstack((preds, np.ones((len(configs), 1)))))
# Add the delta.
delta = preds*np.mat(weight.reshape(-1, 1))
perfs += delta.getA()[:, 0]
configs_candidate = []
indexes = np.argsort(perfs)[:iter_L]
for index in indexes:
configs_candidate.append(configs[index])
# Evaluate the low-fidelity configurations.
print('='*10 + 'Evaluating the low-fidelity configurations')
config_params = []
for config in configs_candidate:
config_params.append((config.get_dictionary(), s_min))
result_perf = run_parallel_async(pool, objective_function, config_params)
for index, item in enumerate(result_perf):
X_l.append(configs_candidate[index].get_array().tolist())
y_l.append(item[0])
print(np.array(X_l).shape, np.array(y_l, dtype=np.float64).shape)
# Update f_L.
print('=' * 10 + 'Retrain the f_L')
low_fidelity_model.train(np.array(X_l), np.array(y_l, dtype=np.float64))
config_L.extend(configs_candidate)
configs_input = []
for config in config_L:
if config not in config_evaluated:
configs_input.append(config)
# Choose the next configuration.
config_arrays = convert_configurations_to_array(configs_input)
perfs, _ = low_fidelity_model.predict(config_arrays)
perfs = perfs[:, 0]
if len(y) > 3:
preds = []
for i in range(K):
m, _ = base_models[i].predict(config_arrays)
preds.append(m[:, 0].tolist())
preds = np.array(preds).T
preds = np.mat(np.hstack((preds, np.ones((len(configs_input), 1)))))
# Add the delta.
delta = preds * np.mat(weight.reshape(-1, 1))
perfs += delta.getA()[:, 0]
next_config = configs_input[np.argmin(perfs)]
# Evaluate this config with a high-fidelity setting.
print('=' * 10 + 'Evaluate the high-fidelity configuration')
perf, _ = objective_function((next_config.get_dictionary(), s_max))
X.append(next_config)
y.append(perf)
if perf < inc:
inc = perf
c.append([time.time()-start_time, inc])
print('Current inc', inc)
if len(y) < 3:
continue
# Learn the weight in TSE.
Z = []
for i in range(K):
m, v = base_models[i].predict(convert_configurations_to_array(X))
Z.append(m[:, 0].tolist())
Z = np.mat(np.hstack((np.array(Z).T, np.ones((len(y), 1)))))
f = np.mat(np.array(y).reshape((-1, 1)))
# Compute the weight.
try:
ZtZ_inv = np.linalg.inv(Z.T * Z)
weight = (ZtZ_inv * Z.T * f)[:, 0]
print('The weight updated is', weight)
except np.linalg.LinAlgError as err:
if 'Singular matrix' in str(err):
print('Singular matrix encountered, and do not update the weight!')
else:
raise ValueError('Unexpected error!')
# Save the result.
np.save('data/xgb/tse_%d.npy' % run_id, np.array(c))
plt.plot(np.array(c)[:, 0], np.array(c)[:, 1])
plt.xlabel('time_elapsed (s)')
plt.ylabel('validation error')
plt.savefig("data/xgb/tse_%d.png" % run_id)
if time.time() - start_time > 21600:
raise ValueError('Runtime budget meets!')
pool.shutdown(wait=True)