Note
Click here to download the full example code
3. Efficient Optimization Algorithms
Optuna enables efficient hyperparameter optimization by adopting state-of-the-art algorithms for sampling hyperparameters and pruning efficiently unpromising trials.
Sampling Algorithms
Samplers basically continually narrow down the search space using the records of suggested parameter values and evaluated objective values,
leading to an optimal search space which giving off parameters leading to better objective values.
More detailed explanation of how samplers suggest parameters is in BaseSampler
.
Optuna provides the following sampling algorithms:
Grid Search implemented in
GridSampler
Random Search implemented in
RandomSampler
Tree-structured Parzen Estimator algorithm implemented in
TPESampler
CMA-ES based algorithm implemented in
CmaEsSampler
Algorithm to enable partial fixed parameters implemented in
PartialFixedSampler
Nondominated Sorting Genetic Algorithm II implemented in
NSGAIISampler
A Quasi Monte Carlo sampling algorithm implemented in
QMCSampler
The default sampler is TPESampler
.
Switching Samplers
import optuna
By default, Optuna uses TPESampler
as follows.
study = optuna.create_study()
print(f"Sampler is {study.sampler.__class__.__name__}")
Sampler is TPESampler
If you want to use different samplers for example RandomSampler
and CmaEsSampler
,
study = optuna.create_study(sampler=optuna.samplers.RandomSampler())
print(f"Sampler is {study.sampler.__class__.__name__}")
study = optuna.create_study(sampler=optuna.samplers.CmaEsSampler())
print(f"Sampler is {study.sampler.__class__.__name__}")
Sampler is RandomSampler
Sampler is CmaEsSampler
Pruning Algorithms
Pruners
automatically stop unpromising trials at the early stages of the training (a.k.a., automated early-stopping).
Optuna provides the following pruning algorithms:
Median pruning algorithm implemented in
MedianPruner
Non-pruning algorithm implemented in
NopPruner
Algorithm to operate pruner with tolerance implemented in
PatientPruner
Algorithm to prune specified percentile of trials implemented in
PercentilePruner
Asynchronous Successive Halving algorithm implemented in
SuccessiveHalvingPruner
Hyperband algorithm implemented in
HyperbandPruner
Threshold pruning algorithm implemented in
ThresholdPruner
We use MedianPruner
in most examples,
though basically it is outperformed by SuccessiveHalvingPruner
and
HyperbandPruner
as in this benchmark result.
Activating Pruners
To turn on the pruning feature, you need to call report()
and should_prune()
after each step of the iterative training.
report()
periodically monitors the intermediate objective values.
should_prune()
decides termination of the trial that does not meet a predefined condition.
We would recommend using integration modules for major machine learning frameworks.
Exclusive list is integration
and usecases are available in ~optuna/examples.
import logging
import sys
import sklearn.datasets
import sklearn.linear_model
import sklearn.model_selection
def objective(trial):
iris = sklearn.datasets.load_iris()
classes = list(set(iris.target))
train_x, valid_x, train_y, valid_y = sklearn.model_selection.train_test_split(
iris.data, iris.target, test_size=0.25, random_state=0
)
alpha = trial.suggest_float("alpha", 1e-5, 1e-1, log=True)
clf = sklearn.linear_model.SGDClassifier(alpha=alpha)
for step in range(100):
clf.partial_fit(train_x, train_y, classes=classes)
# Report intermediate objective value.
intermediate_value = 1.0 - clf.score(valid_x, valid_y)
trial.report(intermediate_value, step)
# Handle pruning based on the intermediate value.
if trial.should_prune():
raise optuna.TrialPruned()
return 1.0 - clf.score(valid_x, valid_y)
Set up the median stopping rule as the pruning condition.
# Add stream handler of stdout to show the messages
optuna.logging.get_logger("optuna").addHandler(logging.StreamHandler(sys.stdout))
study = optuna.create_study(pruner=optuna.pruners.MedianPruner())
study.optimize(objective, n_trials=20)
A new study created in memory with name: no-name-1c089677-ceca-40c2-93ee-38162000d9d1
Trial 0 finished with value: 0.07894736842105265 and parameters: {'alpha': 0.0007398819207061793}. Best is trial 0 with value: 0.07894736842105265.
Trial 1 finished with value: 0.052631578947368474 and parameters: {'alpha': 0.0002976044090396954}. Best is trial 1 with value: 0.052631578947368474.
Trial 2 finished with value: 0.052631578947368474 and parameters: {'alpha': 0.001318189123404992}. Best is trial 1 with value: 0.052631578947368474.
Trial 3 finished with value: 0.4736842105263158 and parameters: {'alpha': 5.5268241133563414e-05}. Best is trial 1 with value: 0.052631578947368474.
Trial 4 finished with value: 0.26315789473684215 and parameters: {'alpha': 3.452719201046683e-05}. Best is trial 1 with value: 0.052631578947368474.
Trial 5 pruned.
Trial 6 pruned.
Trial 7 finished with value: 0.07894736842105265 and parameters: {'alpha': 0.0018181098989537255}. Best is trial 1 with value: 0.052631578947368474.
Trial 8 pruned.
Trial 9 pruned.
Trial 10 finished with value: 0.3421052631578947 and parameters: {'alpha': 0.07718668049323248}. Best is trial 1 with value: 0.052631578947368474.
Trial 11 pruned.
Trial 12 finished with value: 0.052631578947368474 and parameters: {'alpha': 0.006513503897650885}. Best is trial 1 with value: 0.052631578947368474.
Trial 13 pruned.
Trial 14 pruned.
Trial 15 finished with value: 0.13157894736842102 and parameters: {'alpha': 0.0019417485547574107}. Best is trial 1 with value: 0.052631578947368474.
Trial 16 finished with value: 0.02631578947368418 and parameters: {'alpha': 0.00022228641148510934}. Best is trial 16 with value: 0.02631578947368418.
Trial 17 pruned.
Trial 18 finished with value: 0.07894736842105265 and parameters: {'alpha': 0.0001542233754429986}. Best is trial 16 with value: 0.02631578947368418.
Trial 19 pruned.
As you can see, several trials were pruned (stopped) before they finished all of the iterations.
The format of message is "Trial <Trial Number> pruned."
.
Which Sampler and Pruner Should be Used?
From the benchmark results which are available at optuna/optuna - wiki “Benchmarks with Kurobako”, at least for not deep learning tasks, we would say that
For
RandomSampler
,MedianPruner
is the best.For
TPESampler
,Hyperband
is the best.
However, note that the benchmark is not deep learning. For deep learning tasks, consult the below table. This table is from the Ozaki et al., Hyperparameter Optimization Methods: Overview and Characteristics, in IEICE Trans, Vol.J103-D No.9 pp.615-631, 2020 paper, which is written in Japanese.
Parallel Compute Resource |
Categorical/Conditional Hyperparameters |
Recommended Algorithms |
---|---|---|
Limited |
No |
TPE. GP-EI if search space is low-dimensional and continuous. |
Yes |
TPE. GP-EI if search space is low-dimensional and continuous |
|
Sufficient |
No |
CMA-ES, Random Search |
Yes |
Random Search or Genetic Algorithm |
Integration Modules for Pruning
To implement pruning mechanism in much simpler forms, Optuna provides integration modules for the following libraries.
For the complete list of Optuna’s integration modules, see integration
.
For example, XGBoostPruningCallback
introduces pruning without directly changing the logic of training iteration.
(See also example for the entire script.)
pruning_callback = optuna.integration.XGBoostPruningCallback(trial, 'validation-error')
bst = xgb.train(param, dtrain, evals=[(dvalid, 'validation')], callbacks=[pruning_callback])
Total running time of the script: ( 0 minutes 2.230 seconds)