Merge pull request #809 from crcrpar/naive-hyperband

Add `HyperbandPruner`

docs/source/reference/pruners.rst

@@ -21,3 +21,7 @@ Pruners
 .. autoclass:: SuccessiveHalvingPruner
     :members:
     :exclude-members: prune
+
+.. autoclass:: HyperbandPruner
+    :members:
+    :exclude-members: prune

optuna/pruners/__init__.py

@@ -1,4 +1,5 @@
 from optuna.pruners.base import BasePruner  # NOQA
+from optuna.pruners.hyperband import HyperbandPruner  # NOQA
 from optuna.pruners.median import MedianPruner  # NOQA
 from optuna.pruners.nop import NopPruner  # NOQA
 from optuna.pruners.percentile import PercentilePruner  # NOQA

optuna/pruners/hyperband.py

@@ -0,0 +1,155 @@
+import optuna
+from optuna import logging
+from optuna.pruners.base import BasePruner
+from optuna.pruners.successive_halving import SuccessiveHalvingPruner
+from optuna import type_checking
+
+if type_checking.TYPE_CHECKING:
+    from typing import List  # NOQA
+
+    from optuna import structs  # NOQA
+
+_logger = logging.get_logger(__name__)
+
+
+class HyperbandPruner(BasePruner):
+    """Pruner using Hyperband.
+
+    As SuccessiveHalving (SHA) requires the number of configurations
+    :math:`n` as its hyperparameter.  For a given finite budget :math:`B`,
+    all the configurations have the resources of :math:`B \\over n` on average.
+    As you can see, there will be a trade-off of :math:`B` and :math:`B \\over n`.
+    `Hyperband <http://www.jmlr.org/papers/volume18/16-558/16-558.pdf>`_ attacks this trade-off
+    by trying different :math:`n` values for a fixed budget.
+    Note that this implementation does not take as inputs the maximum amount of resource to
+    a single SHA noted as :math:`R` in the paper.
+
+    Args:
+        min_resource:
+            A parameter for specifying the minimum resource allocated to a trial noted as :math:`r`
+            in the paper.
+            See the details for :class:`~optuna.pruners.SuccessiveHalvingPruner`.
+        reduction_factor:
+            A parameter for specifying reduction factor of promotable trials noted as
+            :math:`\\eta` in the paper. See the details for
+            :class:`~optuna.pruners.SuccessiveHalvingPruner`.
+        n_brackets:
+            The number of :class:`~optuna.pruners.SuccessiveHalvingPruner`\\s (brackets).
+        min_early_stopping_rate_low:
+            A parameter for specifying the minimum early-stopping rate.
+            This parameter is related to a parameter that is referred to as :math:`r` and used in
+            `Asynchronous SuccessiveHalving paper <http://arxiv.org/abs/1810.05934>`_.
+            The minimum early stopping rate for :math:`i` th bracket is :math:`i + s`.
+    """
+
+    def __init__(
+            self,
+            min_resource=1,
+            reduction_factor=3,
+            n_brackets=4,
+            min_early_stopping_rate_low=0
+    ):
+        # type: (int, int, int, int) -> None
+
+        self._pruners = []  # type: List[SuccessiveHalvingPruner]
+        self._reduction_factor = reduction_factor
+        self._resource_budget = 0
+        self._n_brackets = n_brackets
+        self._bracket_resource_budgets = []  # type: List[int]
+
+        _logger.debug('Hyperband has {} brackets'.format(self._n_brackets))
+
+        for i in range(n_brackets):
+            bracket_resource_budget = self._calc_bracket_resource_budget(i, n_brackets)
+            self._resource_budget += bracket_resource_budget
+            self._bracket_resource_budgets.append(bracket_resource_budget)
+
+            # N.B. (crcrpar): `min_early_stopping_rate` has the information of `bracket_index`.
+            min_early_stopping_rate = min_early_stopping_rate_low + i
+
+            _logger.debug(
+                '{}th bracket has minimum early stopping rate of {}'.format(
+                    i, min_early_stopping_rate))
+
+            pruner = SuccessiveHalvingPruner(
+                min_resource=min_resource,
+                reduction_factor=reduction_factor,
+                min_early_stopping_rate=min_early_stopping_rate,
+            )
+            self._pruners.append(pruner)
+
+    def prune(self, study, trial):
+        # type: (optuna.study.Study, structs.FrozenTrial) -> bool
+
+        i = self._get_bracket_id(study, trial)
+        _logger.debug('{}th bracket is selected'.format(i))
+        bracket_study = self._create_bracket_study(study, i)
+        return self._pruners[i].prune(bracket_study, trial)
+
+    # TODO(crcrpar): Improve resource computation/allocation algorithm.
+    def _calc_bracket_resource_budget(self, pruner_index, n_brackets):
+        # type: (int, int) -> int
+
+        n = self._reduction_factor ** (n_brackets - 1)
+        return n + (n / 2) * (n_brackets - 1 - pruner_index)
+
+    def _get_bracket_id(self, study, trial):
+        # type: (optuna.study.Study, structs.FrozenTrial) -> int
+        """Computes the index of bracket for a trial of ``trial_number``.
+
+        The index of a bracket is noted as :math:`s` in
+        `Hyperband paper <http://www.jmlr.org/papers/volume18/16-558/16-558.pdf>`_.
+        """
+
+        n = hash('{}_{}'.format(study.study_name, trial.number)) % self._resource_budget
+        for i in range(self._n_brackets):
+            n -= self._bracket_resource_budgets[i]
+            if n < 0:
+                return i
+
+        assert False, 'This line should be unreachable.'
+
+    def _create_bracket_study(self, study, bracket_index):
+        # type: (optuna.study.Study, int) -> optuna.study.Study
+
+        # This class is assumed to be passed to
+        # `SuccessiveHalvingPruner.prune` in which `get_trials`,
+        # `direction`, and `storage` are used.
+        # But for safety, prohibit the other attributes explicitly.
+        class _BracketStudy(optuna.study.Study):
+
+            _VALID_ATTRS = (
+                'get_trials', 'direction', '_storage', '_study_id',
+                'pruner', 'study_name', '_bracket_id', 'sampler'
+            )
+
+            def __init__(self, study, bracket_id):
+                # type: (optuna.study.Study, int) -> None
+
+                super().__init__(
+                    study_name=study.study_name,
+                    storage=study._storage,
+                    sampler=study.sampler,
+                    pruner=study.pruner
+                )
+                self._bracket_id = bracket_id
+
+            def get_trials(self, deepcopy=True):
+                # type: (bool) -> List[structs.FrozenTrial]
+
+                trials = super().get_trials(deepcopy=deepcopy)
+                pruner = self.pruner
+                assert isinstance(pruner, HyperbandPruner)
+                return [
+                    t for t in trials
+                    if pruner._get_bracket_id(self, t) == self._bracket_id
+                ]
+
+            def __getattribute__(self, attr_name):  # type: ignore
+                if attr_name not in _BracketStudy._VALID_ATTRS:
+                    raise AttributeError(
+                        "_BracketStudy does not have attribute of '{}'".format(attr_name))
+                else:
+                    return object.__getattribute__(self, attr_name)
+
+        return _BracketStudy(study, bracket_index)

optuna/study.py

@@ -191,6 +191,16 @@ def __init__(
         self.sampler = sampler or samplers.TPESampler()
         self.pruner = pruner or pruners.MedianPruner()
 
+        if (
+                isinstance(self.sampler, samplers.TPESampler) and
+                isinstance(self.pruner, pruners.HyperbandPruner)):
+            msg = (
+                "The algorithm of TPESampler and HyperbandPruner might behave in a different way "
+                "from the paper of Hyperband because the sampler uses all the trials including "
+                "ones of brackets other than that of currently running trial")
+            warnings.warn(msg, UserWarning)
+            _logger.warning(msg)
+
         self._optimize_lock = threading.Lock()
 
     def __getstate__(self):

tests/pruners_tests/test_hyperband.py

@@ -0,0 +1,84 @@
+import pytest
+
+import optuna
+from optuna import type_checking
+
+if type_checking.TYPE_CHECKING:
+    from optuna.trial import Trial  # NOQA
+
+MIN_RESOURCE = 1
+REDUCTION_FACTOR = 2
+N_BRACKETS = 4
+EARLY_STOPPING_RATE_LOW = 0
+EARLY_STOPPING_RATE_HIGH = 3
+N_REPORTS = 10
+EXPECTED_N_TRIALS_PER_BRACKET = 10
+
+
+def test_warn_on_TPESampler():
+    # type: () -> None
+
+    with pytest.warns(UserWarning):
+        optuna.study.create_study(pruner=optuna.pruners.HyperbandPruner())
+
+
+def test_hyperband_pruner_intermediate_values():
+    # type: () -> None
+
+    pruner = optuna.pruners.HyperbandPruner(
+        min_resource=MIN_RESOURCE,
+        reduction_factor=REDUCTION_FACTOR,
+        n_brackets=N_BRACKETS
+    )
+
+    study = optuna.study.create_study(sampler=optuna.samplers.RandomSampler(), pruner=pruner)
+
+    def objective(trial):
+        # type: (Trial) -> float
+
+        for i in range(N_REPORTS):
+            trial.report(i)
+
+        return 1.0
+
+    study.optimize(objective, n_trials=N_BRACKETS * EXPECTED_N_TRIALS_PER_BRACKET)
+
+    trials = study.trials
+    assert len(trials) == N_BRACKETS * EXPECTED_N_TRIALS_PER_BRACKET
+
+
+def test_bracket_study():
+    # type: () -> None
+
+    pruner = optuna.pruners.HyperbandPruner(
+        min_resource=MIN_RESOURCE,
+        reduction_factor=REDUCTION_FACTOR,
+        n_brackets=N_BRACKETS
+    )
+    study = optuna.study.create_study(sampler=optuna.samplers.RandomSampler(), pruner=pruner)
+    bracket_study = pruner._create_bracket_study(study, 0)
+
+    with pytest.raises(AttributeError):
+        bracket_study.optimize(lambda *args: 1.0)
+
+    for attr in ('set_user_attr', 'set_system_attr'):
+        with pytest.raises(AttributeError):
+            getattr(bracket_study, attr)('abc', 100)
+
+    for attr in ('user_attrs', 'system_attrs'):
+        with pytest.raises(AttributeError):
+            getattr(bracket_study, attr)
+
+    with pytest.raises(AttributeError):
+        bracket_study.trials_dataframe()
+
+    bracket_study.get_trials()
+    bracket_study.direction
+    bracket_study._storage
+    bracket_study._study_id
+    bracket_study.pruner
+    bracket_study.study_name
+    # As `_BracketStudy` is defined inside `HyperbandPruner`,
+    # we cannot do `assert isinstance(bracket_study, _BracketStudy)`.
+    # This is why the below line is ignored by mypy checks.
+    bracket_study._bracket_id  # type: ignore