def setup_tune_scheduler():
from ray.tune.suggest.skopt import SkOptSearch
from ray.tune.suggest.suggestion import ConcurrencyLimiter
from skopt import Optimizer
exp_metrics = workload.exp_metric()
search_space, dim_names = workload.create_skopt_space()
algo = ConcurrencyLimiter(
SkOptSearch(
Optimizer(search_space),
dim_names,
**exp_metrics,
),
3,
)
scheduler = FluidBandScheduler(
max_res=3,
reduction_factor=3,
**exp_metrics,
)
return dict(
search_alg=algo,
scheduler=scheduler,
trial_executor=MyRayTrialExecutor(),
resources_per_trial=com.detect_baseline_resource(),
)
def setup_tune_scheduler():
search_space = workload.create_search_space()
experiment_metrics = workload.exp_metric()
scheduler = HyperBandScheduler(time_attr="training_iteration",
max_t=81,
reduction_factor=3,
**experiment_metrics)
return dict(
search_alg=VariantGenerator(),
scheduler=scheduler,
config=search_space,
resources_per_trial=com.detect_baseline_resource(),
)
def setup_tune_scheduler(num_worker):
search_space = workload.create_search_space()
# experiment_metrics = workload.exp_metric()
asha_parallel = AsyncHyperBandSchedulerWithParalelism(
# set a large max_t such that ASHA will always promot to next rung,
# until something reaches target accuracy
max_t=int(1e10),
reduction_factor=3,
**workload.exp_metric())
return dict(
scheduler=asha_parallel,
search_alg=VariantGenerator(max_concurrent=sched_algo()),
config=search_space,
resources_per_trial=com.detect_baseline_resource(),
)
def setup_tune_scheduler():
search_space = workload.create_search_space()
scheduler = ASHAScheduler(
# set a large max_t such that ASHA will always promot to next rung,
# until something reaches target accuracy
max_t=int(1000),
reduction_factor=3,
**workload.exp_metric(),
)
return dict(
search_alg=VariantGenerator(),
scheduler=scheduler,
config=search_space,
resources_per_trial=com.detect_baseline_resource(),
)
def setup_tune_scheduler():
ss, custom_explore = workload.create_sample_space()
search_space = workload.create_search_space()
scheduler = PopulationBasedTraining(time_attr="training_iteration",
perturbation_interval=5,
hyperparam_mutations=ss,
custom_explore_fn=custom_explore,
**workload.exp_metric())
return dict(
scheduler=scheduler,
config=search_space,
# num_samples in PBT only sets population
num_samples=10,
resources_per_trial=com.detect_baseline_resource(),
)
def setup_tune_scheduler():
# BOHB uses ConfigSpace for their hyperparameter search space
config_space = workload.create_ch()
experiment_metrics = workload.exp_metric()
bohb_search = TuneBOHB(config_space, **experiment_metrics)
bohb_hyperband = HyperBandForBOHB(
time_attr="training_iteration",
max_t=81,
reduction_factor=3,
**experiment_metrics
)
return dict(
scheduler=bohb_hyperband,
search_alg=bohb_search,
resources_per_trial=com.detect_baseline_resource(),
)