8000 Meta learning designer: 1) multi-metric trial support, 2) updated default schedules, 3) additional testing and documentation. by copybara-service[bot] · Pull Request #933 · google/vizier · GitHub
[go: up one dir, main page]
Skip to content

Meta learning designer: 1) multi-metric trial support, 2) updated default schedules, 3) additional testing and documentation. #933

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
copybara-service merged 1 commit into from
Oct 5, 2023
Merged

Meta learning designer: 1) multi-metric trial support, 2) updated default schedules, 3) additional testing and documentation. #933

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
68 changes: 31 additions & 37 deletions vizier/_src/algorithms/designers/meta_learning/eagle_meta_learning.py
View file
8000 Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -28,7 +28,7 @@ def meta_eagle_search_space() -> vz.SearchSpace:
name='perturbation',
min_value=1e-4,
max_value=1e2,
default_value=1e-2,
default_value=1e-1,
scale_type=vz.ScaleType.LOG,
)
search_space.root.add_float_param(
Expand All @@ -38,34 +38,6 @@ def meta_eagle_search_space() -> vz.SearchSpace:
default_value=1e-3,
scale_type=vz.ScaleType.LOG,
)
search_space.root.add_float_param(
name='pure_categorical_perturbation',
min_value=0.001,
max_value=10.0,
default_value=0.1,
scale_type=vz.ScaleType.LOG,
)
search_space.root.add_float_param(
name='max_perturbation',
min_value=1e-3,
max_value=1e2,
default_value=5 * 1e-1,
scale_type=vz.ScaleType.LOG,
)
search_space.root.add_float_param(
name='categorical_perturbation_factor',
min_value=2.5 * 1e-1,
max_value=2.5 * 1e3,
default_value=2.5 * 1e1,
scale_type=vz.ScaleType.LOG,
)
search_space.root.add_float_param(
name='discrete_perturbation_factor',
min_value=1e-1,
max_value=1e3,
default_value=1e1,
scale_type=vz.ScaleType.LOG,
)
# Gravity
search_space.root.add_float_param(
name='gravity',
Expand All @@ -74,13 +46,6 @@ def meta_eagle_search_space() -> vz.SearchSpace:
default_value=1.0,
scale_type=vz.ScaleType.LOG,
)
search_space.root.add_float_param(
name='negative_gravity',
min_value=2 * 1e-4,
max_value=2.0,
default_value=2 * 1e-2,
scale_type=vz.ScaleType.LOG,
)
# Visibility
search_space.root.add_float_param(
name='visibility',
Expand All @@ -91,7 +56,7 @@ def meta_eagle_search_space() -> vz.SearchSpace:
)
search_space.root.add_float_param(
name='categorical_visibility',
min_value=2 * 1e-3,
min_value=2.0 * 1e-3,
max_value=2.0 * 1e1,
default_value=2.0 * 1e-1,
scale_type=vz.ScaleType.LOG,
Expand All @@ -103,6 +68,20 @@ def meta_eagle_search_space() -> vz.SearchSpace:
default_value=1.0,
scale_type=vz.ScaleType.LOG,
)
search_space.root.add_float_param(
name='categorical_perturbation_factor',
min_value=2.5 * 1e-1,
max_value=2.5 * 1e3,
default_value=2.5 * 1e1,
scale_type=vz.ScaleType.LOG,
)
search_space.root.add_float_param(
name='discrete_perturbation_factor',
min_value=1e-1,
max_value=1e3,
default_value=1e1,
scale_type=vz.ScaleType.LOG,
)
# Pool size.
search_space.root.add_float_param(
name='pool_size_factor',
Expand All @@ -111,4 +90,19 @@ def meta_eagle_search_space() -> vz.SearchSpace:
default_value=1.2,
scale_type=vz.ScaleType.LOG,
)
search_space.root.add_float_param(
name='negative_gravity',
min_value=2.0 * 1e-4,
max_value=2.0,
default_value=2 * 1e-2,
scale_type=vz.ScaleType.LOG,
)
search_space.root.add_float_param(
name='pure_categorical_perturbation',
min_value=1e-3,
max_value=1e1,
default_value=1e-1,
scale_type=vz.ScaleType.LOG,
)

return search_space
30 changes: 19 additions & 11 deletions vizier/_src/algorithms/designers/meta_learning/meta_learning.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -27,13 +27,17 @@
1) The 'meta-learning' designer's search space is the hyper-parameters of
the 'tuned' designer which we seek to tune.

2) The 'tuned_designer_factory' should accept the hyper-parameters as arguments.
2) Before tuning starts, a tuned designer instantiated with the default
parameter values defined by the search space is used. This means that
default configuration defined in the designer level will be overridden.

3) The 'tuned_designer_factory' should accept the hyper-parameters as arguments.
This means for example, that if the 'tuned_designer' relies on an internal
configuration class the 'tuned_designer_factory' function would have to handle
the instantiation of that class (see 'eagle_meta_learning_convergence_test.py'
for an example).

3) Each instance of 'tuned' designer is updated with all trials seen thus-far,
4) Each instance of 'tuned' designer is updated with all trials seen thus-far,
therefore hyper-parameters that were created later in the process will benefit
from being instantiated with larger trajectory. This violates the assumption
of the meta-learning designer which the meta-trial metrics are derived from the
Expand All @@ -60,18 +64,18 @@ class MetaLearningConfig:
suggestion trials on further exploration.

With the default configuration the number of meta-learning iterations is
(3000-750) / 50 = 45. In order to not terminate the meta-learning process,
(10000-3000) / 100 = 70. In order to not terminate the meta-learning process,
set the `tuning_max_num_trials` to sufficiently large value (e.g. 1e6).
"""

# Number of trials to use per tuning iteration.
num_trials_per_tuning: int = 50

# Once the number of trials exceeds this threshold meta-learning stops.
tuning_max_num_trials: int = 3000
num_trials_per_tuning: int = 100

# Tuning starts when number of completed trials is at least this threshold.
tuning_min_num_trials: int = 750
tuning_min_num_trials: int = 3000

# Once the number of trials exceeds this threshold meta-learning stops.
tuning_max_num_trials: int = 10000


class MetaLearningState(enum.Enum):
Expand Down Expand Up @@ -145,13 +149,17 @@ class MetaLearningDesigner(vza.Designer):

def __attrs_post_init__(self):
"""Initializes the meta learning desiger."""
if len(self.problem.metric_information) != 1:
raise ValueError(f'Expected exactly one metric. {self.problem}')

if self.seed is None:
# JAX random seed doesn't accept None, so generating random integer.
self.seed = np.random.randint(low=0, high=1e6)

# Instantiate an MetaLearningUtils.
self._utils = utils.MetaLearningUtils(
goal=self.problem.metric_information.item().goal,
tuned_metric_name=self.problem.metric_information.item().name,
meta_metric_name=self._meta_designer_metric_name,
tuning_params=self.tuning_hyperparams,
)
Expand Down Expand Up @@ -203,16 +211,16 @@ def update(
vza.CompletedTrials(completed.trials), vza.ActiveTrials()
)

# Check if enough trials already accumulated to start meta learning.
if len(self._trials) < self.config.tuning_min_num_trials:
# Not enough trials yet to start meta learning.
return

# Check if the meta-learning process should be terminated.
elif len(self._trials) >= self.config.tuning_max_num_trials:
# Check if the meta-learning has just terminated. If so, finalize it.
if self._state == MetaLearningState.TUNE:
# Find the best meta-learn result.
best_meta_trial = self._utils.get_best_trial(self._meta_trials)
best_meta_trial = self._utils.get_best_meta_trial(self._meta_trials)
self._curr_tuned_designer = self.tuned_designer_factory(
self.problem, seed=self.seed, **best_meta_trial.parameters.as_dict()
)
Expand All @@ -229,7 +237,7 @@ def update(
# Get best score for the current iteration.
meta_trial = self._utils.complete_meta_suggestion(
meta_suggestion=self._curr_tuned_hyperparams,
score=self._utils.get_best_trial_score(self._curr_trials),
score=self._utils.get_best_tuned_trial_score(self._curr_trials),
)
self._meta_designer.update(
vza.CompletedTrials([meta_trial]), vza.ActiveTrials()
Expand Down
Loading
0