def propose_queries_blackbox_optimizer(config, current_input,
blackbox_optimizer, iteration):
core_hyperparameters = blackbox_optimizer.get_hyperparameters()
current_input_vectorized = zero_order_pb2.Vector(
values=current_input.tolist())
proposed_perturbations = []
requests = []
for _ in range(config.test_workers):
# only hyperparameters and current input matter here
null_perturbation = [0.0] * len(current_input)
request = zero_order_pb2.EvaluationRequest(
current_input=current_input_vectorized,
hyperparameters=core_hyperparameters,
perturbations=[zero_order_pb2.Vector(values=null_perturbation)],
tag=iteration)
requests.append(request)
for i in range(config.total_num_perturbations):
perturbation = np.random.normal(
size=(len(current_input))) * config.es_precision_parameter
proposed_perturbations.append(perturbation)
tag = i + 1
for _ in range(config.num_repeats):
request = zero_order_pb2.EvaluationRequest(
current_input=current_input_vectorized,
hyperparameters=core_hyperparameters,
perturbations=[
zero_order_pb2.Vector(values=perturbation.tolist())
],
tag=tag)
requests.append(request)
if config.es_est_type == 'antithetic':
antiperturbation = -perturbation
proposed_perturbations.append(antiperturbation)
for _ in range(config.num_repeats):
request = zero_order_pb2.EvaluationRequest(
current_input=current_input_vectorized,
hyperparameters=core_hyperparameters,
perturbations=[
zero_order_pb2.Vector(values=antiperturbation.tolist())
],
tag=-tag)
requests.append(request)
for _ in range(config.num_repeats + config.num_exact_evals):
null_perturbation = [0.0] * len(current_input)
request = zero_order_pb2.EvaluationRequest(
current_input=current_input_vectorized,
hyperparameters=core_hyperparameters,
perturbations=[zero_order_pb2.Vector(values=null_perturbation)],
tag=0)
requests.append(request)
return requests, proposed_perturbations
def TrainEvaluate(self, request, context):
current_input = []
core_hyperparameters = []
for i in range(len(request.current_input.values)):
current_input.append(request.current_input.values[i])
for i in range(len(request.hyperparameters)):
core_hyperparameters.append(request.hyperparameters[i])
hyperparameters = [self.worker_id] + core_hyperparameters
current_input_reshaped = np.array(current_input)
tag = request.tag
proposed_perturbations = []
for j in range(len(request.perturbations)):
proposed_perturbation = []
for k in range(len(request.perturbations[j].values)):
proposed_perturbation.append(
request.perturbations[j].values[k])
proposed_perturbations.append(proposed_perturbation)
perturbations = []
function_values = []
evaluation_stats = []
for i in range(len(proposed_perturbations)):
perturbation = np.array(proposed_perturbations[i])
perturbations.append(
zero_order_pb2.Vector(values=perturbation.tolist()))
task_id_list = np.random.choice(self.task_ids,
size=self.task_batch_size,
replace=False)
task_list = [
self.blackbox_object.config.make_task_fn(task_id=task_id)
for task_id in task_id_list
]
function_value, evaluation_stat = self.blackbox_object.execute(
current_input_reshaped + perturbation,
task_list,
hyperparameters=hyperparameters)
evaluation_stats.append(evaluation_stat)
function_values.append(function_value)
evaluation_stats_reduced = [sum(x) for x in zip(*evaluation_stats)]
if not proposed_perturbations:
results = zero_order_pb2.EvaluationResponse(
perturbations=perturbations,
function_values=function_values,
evaluation_stats=evaluation_stats_reduced,
tag=tag)
else:
results = zero_order_pb2.EvaluationResponse(
perturbations=[],
function_values=function_values,
evaluation_stats=evaluation_stats_reduced,
tag=tag)
return results