class BOOptimizer(object):
def __init__(self, params_config):
self.params_config = params_config
self.n_initial_trials = self.params_config.bo.n_initial_trials
self.space = SearchSpace(params_config=params_config)
self.utility_function = UtilityFunction(
config=params_config.bo.utility_function, seed=params_config.seed)
self.n_warmup = params_config.bo.utility_function.n_warmup or 5000
self.n_iter = params_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 __init__(self, params_config):
self.params_config = params_config
self.n_initial_trials = self.params_config.bo.n_initial_trials
self.space = SearchSpace(params_config=params_config)
self.utility_function = UtilityFunction(
config=params_config.bo.utility_function, seed=params_config.seed)
self.n_warmup = params_config.bo.utility_function.n_warmup or 5000
self.n_iter = params_config.bo.utility_function.n_iter or 150
def test_space_get_suggestion(self):
space1 = SearchSpace(params_config=self.manager1.params_config)
suggestion = space1.get_suggestion(suggestion=[1, 1, 1])
assert suggestion == {'feature1': 1, 'feature2': 1, 'feature3': 1}
suggestion = space1.get_suggestion(suggestion=[1, 1.2, 2])
assert suggestion == {'feature1': 1, 'feature2': 1.25, 'feature3': 2}
suggestion = space1.get_suggestion(suggestion=[1, 1.5, 3])
assert suggestion == {'feature1': 1, 'feature2': 1.5, 'feature3': 3}
space2 = SearchSpace(params_config=self.manager2.params_config)
suggestion = space2.get_suggestion(suggestion=[1, 1, 1, 1, 1, 0, 0])
assert suggestion == {
'feature1': 1,
'feature2': 1,
'feature3': 1,
'feature4': 1,
'feature5': 'a'
}
suggestion = space2.get_suggestion(suggestion=[1, 1.2, 2, 3, 0, 0, 1])
assert suggestion == {
'feature1': 1,
'feature2': 1,
'feature3': 2,
'feature4': 3,
'feature5': 'c'
}
suggestion = space2.get_suggestion(suggestion=[1, 1.8, 3, 3, 0, 1, 0])
assert suggestion == {
'feature1': 1,
'feature2': 2,
'feature3': 3,
'feature4': 3,
'feature5': 'b'
}
def test_add_observation_to_space_search(self):
space1 = SearchSpace(params_config=self.manager1.params_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(params_config=self.manager2.params_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_space_search(self):
# Space 1
space1 = SearchSpace(params_config=self.manager1.params_config)
assert space1.dim == 3
assert len(space1.bounds) == 3
assert len(space1.discrete_features) == 3
assert len(space1.categorical_features) == 0 # pylint:disable=len-as-condition
for i, feature in enumerate(space1.features):
# Bounds
if feature == 'feature1':
assert np.all(space1.bounds[i] == [1, 3])
elif feature == 'feature2':
assert np.all(space1.bounds[i] == [1, 2])
elif feature == 'feature3':
assert np.all(space1.bounds[i] == [1, 5])
for feature in space1.features:
# Features
if feature == 'feature1':
assert np.all(space1.discrete_features['feature1']['values'] ==
[1, 2, 3])
elif feature == 'feature2':
assert np.all(space1.discrete_features['feature2']['values'] ==
np.asarray([1., 1.25, 1.5, 1.75, 2.]))
elif feature == 'feature3':
assert np.all(space1.discrete_features['feature3']['values'] ==
np.asarray([1, 2, 3, 4]))
# Space 2
space2 = SearchSpace(params_config=self.manager2.params_config)
assert space2.dim == 7
assert len(space2.bounds) == 7
assert len(space2.discrete_features) == 3
assert len(space2.categorical_features) == 1
assert len(space2.features) == 5
for i, feature in enumerate(space2.features):
# Bounds
if feature == 'feature1':
assert np.all(space2.bounds[i] == [1, 5])
elif feature == 'feature2':
assert np.all(space2.bounds[i] == [1, 5])
elif feature == 'feature3':
assert np.all(space2.bounds[i] == [1, 6])
elif feature == 'feature4':
assert np.all(space2.bounds[i] == [1, 5])
elif feature == 'feature5':
assert np.all(space2.bounds[i] == [0, 1])
# One feature left is continuous
# One categorical Features
assert space2.categorical_features == {
'feature5': {
'values': ['a', 'b', 'c'],
'number': 3
}
}
# 3 discrete Features
assert space2.discrete_features['feature1']['values'] == [
1, 2, 3, 4, 5
]
assert np.all(space2.discrete_features['feature2']['values'] ==
np.asarray([1, 2, 3, 4, 5]))
assert np.all(space2.discrete_features['feature3']['values'] ==
np.asarray([1, 2, 3, 4, 5]))