class BOOptimizer(object):
def __init__(self, hptuning_config):
self.hptuning_config = hptuning_config
self.n_initial_trials = self.hptuning_config.bo.n_initial_trials
self.space = SearchSpace(hptuning_config=hptuning_config)
self.utility_function = UtilityFunction(
config=hptuning_config.bo.utility_function,
seed=hptuning_config.seed)
self.n_warmup = hptuning_config.bo.utility_function.n_warmup or 5000
self.n_iter = hptuning_config.bo.utility_function.n_iter or 150
def _maximize(self):
""" Find argmax of the acquisition function."""
if not self.space.is_observations_valid():
return None
y_max = self.space.y.max()
self.utility_function.gaussian_process.fit(self.space.x, self.space.y)
return self.utility_function.max_compute(y_max=y_max,
bounds=self.space.bounds,
n_warmup=self.n_warmup,
n_iter=self.n_iter)
def add_observations(self, configs, metrics):
# Turn configs and metrics into data points
self.space.add_observations(configs=configs, metrics=metrics)
def get_suggestion(self):
x = self._maximize()
return self.space.get_suggestion(x)
class BOOptimizer(object):
def __init__(self, hptuning_config):
self.hptuning_config = hptuning_config
self.n_initial_trials = self.hptuning_config.bo.n_initial_trials
self.space = SearchSpace(hptuning_config=hptuning_config)
self.utility_function = UtilityFunction(
config=hptuning_config.bo.utility_function, seed=hptuning_config.seed)
self.n_warmup = hptuning_config.bo.utility_function.n_warmup or 5000
self.n_iter = hptuning_config.bo.utility_function.n_iter or 150
def _maximize(self):
""" Find argmax of the acquisition function."""
if not self.space.is_observations_valid():
return None
y_max = self.space.y.max()
self.utility_function.gaussian_process.fit(self.space.x, self.space.y)
return self.utility_function.max_compute(y_max=y_max,
bounds=self.space.bounds,
n_warmup=self.n_warmup,
n_iter=self.n_iter)
def add_observations(self, configs, metrics):
# Turn configs and metrics into data points
self.space.add_observations(configs=configs, metrics=metrics)
def get_suggestion(self):
x = self._maximize()
return self.space.get_suggestion(x)
def test_add_observation_to_space_search(self):
space1 = SearchSpace(hptuning_config=self.manager1.hptuning_config)
assert space1.x == []
assert space1.y == []
configs = [
{'feature1': 1, 'feature2': 1, 'feature3': 1},
{'feature1': 2, 'feature2': 1.2, 'feature3': 2},
{'feature1': 3, 'feature2': 1.3, 'feature3': 3}
]
metrics = [1, 2, 3]
space1.add_observations(
configs=configs,
metrics=metrics
)
assert len(space1.x) == 3
assert len(space1.y) == 3
for i, feature in enumerate(space1.features):
if feature == 'feature1':
assert np.all(space1.x[:, i] == [1, 2, 3])
elif feature == 'feature2':
assert np.all(space1.x[:, i] == [1, 1.2, 1.3])
elif feature == 'feature3':
assert np.all(space1.x[:, i] == [1, 2, 3])
assert np.all(space1.y == np.array([-1, -2, -3]))
space2 = SearchSpace(hptuning_config=self.manager2.hptuning_config)
configs = [
{'feature1': 1, 'feature2': 1, 'feature3': 1, 'feature4': 1, 'feature5': 'a'},
{'feature1': 2, 'feature2': 1.2, 'feature3': 2, 'feature4': 4, 'feature5': 'b'},
{'feature1': 3, 'feature2': 1.3, 'feature3': 3, 'feature4': 3, 'feature5': 'a'}
]
metrics = [1, 2, 3]
space2.add_observations(
configs=configs,
metrics=metrics
)
assert len(space2.x) == 3
assert len(space2.y) == 3
for i, feature in enumerate(space2.features):
if feature == 'feature1':
assert np.all(space2.x[:, i] == [1, 2, 3])
elif feature == 'feature2':
assert np.all(space2.x[:, i] == [1, 1.2, 1.3])
elif feature == 'feature3':
assert np.all(space2.x[:, i] == [1, 2, 3])
elif feature == 'feature4':
assert np.all(space2.x[:, i] == [1, 4, 3])
elif feature == 'feature5':
assert np.all(space2.x[:, i:i + 3] == [[1, 0, 0], [0, 1, 0], [1, 0, 0]])
assert np.all(space2.y == np.array(metrics))
def test_add_observation_to_space_search(self):
space1 = SearchSpace(hptuning_config=self.manager1.hptuning_config)
assert space1.x == []
assert space1.y == []
configs = [
{'feature1': 1, 'feature2': 1, 'feature3': 1},
{'feature1': 2, 'feature2': 1.2, 'feature3': 2},
{'feature1': 3, 'feature2': 1.3, 'feature3': 3}
]
metrics = [1, 2, 3]
space1.add_observations(
configs=configs,
metrics=metrics
)
assert len(space1.x) == 3
assert len(space1.y) == 3
for i, feature in enumerate(space1.features):
if feature == 'feature1':
assert np.all(space1.x[:, i] == [1, 2, 3])
elif feature == 'feature2':
assert np.all(space1.x[:, i] == [1, 1.2, 1.3])
elif feature == 'feature3':
assert np.all(space1.x[:, i] == [1, 2, 3])
assert np.all(space1.y == np.array([-1, -2, -3]))
space2 = SearchSpace(hptuning_config=self.manager2.hptuning_config)
configs = [
{'feature1': 1, 'feature2': 1, 'feature3': 1, 'feature4': 1, 'feature5': 'a'},
{'feature1': 2, 'feature2': 1.2, 'feature3': 2, 'feature4': 4, 'feature5': 'b'},
{'feature1': 3, 'feature2': 1.3, 'feature3': 3, 'feature4': 3, 'feature5': 'a'}
]
metrics = [1, 2, 3]
space2.add_observations(
configs=configs,
metrics=metrics
)
assert len(space2.x) == 3
assert len(space2.y) == 3
for i, feature in enumerate(space2.features):
if feature == 'feature1':
assert np.all(space2.x[:, i] == [1, 2, 3])
elif feature == 'feature2':
assert np.all(space2.x[:, i] == [1, 1.2, 1.3])
elif feature == 'feature3':
assert np.all(space2.x[:, i] == [1, 2, 3])
elif feature == 'feature4':
assert np.all(space2.x[:, i] == [1, 4, 3])
elif feature == 'feature5':
assert np.all(space2.x[:, i:i + 3] == [[1, 0, 0], [0, 1, 0], [1, 0, 0]])
assert np.all(space2.y == np.array(metrics))