def test_get_suggestions(self):
params_config = SettingsConfig.from_dict({
'concurrency': 2,
'grid_search': {
'n_experiments': 10
},
'matrix': {
'feature': {
'values': [1, 2, 3]
}
}
})
manager = GridSearchManager(params_config=params_config)
assert len(manager.get_suggestions()) == 3
params_config = SettingsConfig.from_dict({
'concurrency': 2,
'grid_search': {
'n_experiments': 10
},
'matrix': {
'feature1': {
'values': [1, 2, 3]
},
'feature2': {
'linspace': [1, 2, 5]
},
'feature3': {
'range': [1, 5, 1]
}
}
})
manager = GridSearchManager(params_config=params_config)
assert len(manager.get_suggestions()) == 10
def validate_group_params_config(config, raise_for_rest=False):
try:
SettingsConfig.from_dict(config)
except MarshmallowValidationError as e:
if raise_for_rest:
raise ValidationError(e)
else:
raise DjangoValidationError(e)
def test_get_suggestions_calls_sample(self):
params_config = SettingsConfig.from_dict({
'concurrency': 2,
'random_search': {
'n_experiments': 1
},
'matrix': {
'feature1': {
'values': [1, 2, 3]
},
'feature2': {
'linspace': [1, 2, 5]
},
'feature3': {
'range': [1, 5, 1]
}
}
})
manager = RandomSearchManager(params_config=params_config)
with patch.object(MatrixConfig, 'sample') as sample_mock:
manager.get_suggestions()
assert sample_mock.call_count == 3
params_config = SettingsConfig.from_dict({
'concurrency': 2,
'random_search': {
'n_experiments': 1
},
'matrix': {
'feature1': {
'pvalues': [(1, 0.3), (2, 0.3), (3, 0.3)]
},
'feature2': {
'uniform': [0, 1]
},
'feature3': {
'qlognormal': [0, 0.5, 0.51]
},
'feature4': {
'range': [1, 5, 1]
}
}
})
manager = RandomSearchManager(params_config=params_config)
with patch.object(MatrixConfig, 'sample') as sample_mock:
manager.get_suggestions()
assert sample_mock.call_count == 4
def validate(data):
"""Validates the data and creates the config objects"""
if 'project' not in data:
raise PolyaxonfileError(
"The Polyaxonfile must contain a project section.")
if 'model' not in data:
raise PolyaxonfileError(
"The Polyaxonfile must contain a model section.")
validated_data = {
'version': data['version'],
'project': ProjectConfig.from_dict(data['project']),
'model': ModelConfig.from_dict(data['model'])
}
if data.get('settings'):
validated_data['settings'] = SettingsConfig.from_dict(data['settings'])
if data.get('train'):
validated_data['train'] = TrainConfig.from_dict(data['train'])
if data.get('eval'):
validated_data['eval'] = EvalConfig.from_dict(data['eval'])
return validated_data
def validate_headers(spec, data):
"""Validates headers data and creates the config objects"""
validated_data = {
spec.VERSION: data[spec.VERSION],
spec.PROJECT: ProjectConfig.from_dict(data[spec.PROJECT]),
}
if data.get(spec.SETTINGS):
validated_data[spec.SETTINGS] = SettingsConfig.from_dict(
data[spec.SETTINGS])
return validated_data
def test_settings_config(self):
config_dict = {
'logging': LoggingConfig().to_dict(),
'concurrent_experiments': 2,
}
config = SettingsConfig.from_dict(config_dict)
assert_equal_dict(config.to_dict(), config_dict)
# Add n_experiments
config_dict['random_search'] = {'n_experiments': 10}
config = SettingsConfig.from_dict(config_dict)
assert_equal_dict(config.to_dict(), config_dict)
# Raises for negative values
config_dict['random_search']['n_experiments'] = -5
with self.assertRaises(ValidationError):
SettingsConfig.from_dict(config_dict)
config_dict['random_search']['n_experiments'] = -0.5
with self.assertRaises(ValidationError):
SettingsConfig.from_dict(config_dict)
# Add n_experiments percent
config_dict['random_search']['n_experiments'] = 0.5
with self.assertRaises(ValidationError):
SettingsConfig.from_dict(config_dict)
config_dict['random_search']['n_experiments'] = 5
# Add early stopping
config_dict['early_stopping'] = [
{
'metric': 'loss',
'value': 0.1,
'optimization': Optimization.MINIMIZE,
'policy': EarlyStoppingPolicy.ALL
},
{
'metric': 'accuracy',
'value': 0.9,
'optimization': Optimization.MAXIMIZE,
'policy': EarlyStoppingPolicy.EXPERIMENT
}
]
config = SettingsConfig.from_dict(config_dict)
assert_equal_dict(config.to_dict(), config_dict)
def test_get_suggestions(self):
params_config = SettingsConfig.from_dict({
'concurrency': 2,
'random_search': {
'n_experiments': 10
},
'matrix': {
'feature1': {
'values': [1, 2, 3]
},
'feature2': {
'linspace': [1, 2, 5]
},
'feature3': {
'range': [1, 5, 1]
}
}
})
manager = RandomSearchManager(params_config=params_config)
assert len(manager.get_suggestions()) == 10
params_config = SettingsConfig.from_dict({
'concurrency': 2,
'random_search': {
'n_experiments': 10
},
'matrix': {
'feature1': {
'pvalues': [(1, 0.3), (2, 0.3), (3, 0.3)]
},
'feature2': {
'uniform': [0, 1]
},
'feature3': {
'qlognormal': [0, 0.5, 0.51]
}
}
})
manager = RandomSearchManager(params_config=params_config)
assert len(manager.get_suggestions()) == 10
def test_get_suggestions_calls_to_numpy(self):
params_config = SettingsConfig.from_dict({
'concurrency': 2,
'grid_search': {
'n_experiments': 10
},
'matrix': {
'feature': {
'values': [1, 2, 3]
}
}
})
manager = GridSearchManager(params_config=params_config)
with patch.object(MatrixConfig, 'to_numpy') as to_numpy_mock:
manager.get_suggestions()
assert to_numpy_mock.call_count == 1
params_config = SettingsConfig.from_dict({
'concurrency': 2,
'grid_search': {
'n_experiments': 10
},
'matrix': {
'feature1': {
'values': [1, 2, 3]
},
'feature2': {
'logspace': '0.01:0.1:5'
}
}
})
manager = GridSearchManager(params_config=params_config)
with patch.object(MatrixConfig, 'to_numpy') as to_numpy_mock:
manager.get_suggestions()
assert to_numpy_mock.call_count == 2
def params_config(self):
return SettingsConfig.from_dict(self.params) if self.params else None
def validate_group_params_config(config):
try:
SettingsConfig.from_dict(config)
except MarshmallowValidationError as e:
raise ValidationError(e)
def setUp(self):
super().setUp()
params_config = SettingsConfig.from_dict({
'concurrency': 2,
'bo': {
'n_iterations': 5,
'n_initial_trials': 5,
'metric': {
'name': 'loss',
'optimization': 'minimize'
},
'utility_function': {
'acquisition_function': 'ucb',
'kappa': 1.2,
'gaussian_process': {
'kernel': 'matern',
'length_scale': 1.0,
'nu': 1.9,
'n_restarts_optimizer': 0
}
}
},
'matrix': {
'feature1': {
'values': [1, 2, 3]
},
'feature2': {
'linspace': [1, 2, 5]
},
'feature3': {
'range': [1, 5, 1]
}
}
})
self.manager1 = BOSearchManager(params_config=params_config)
params_config = SettingsConfig.from_dict({
'concurrency': 2,
'bo': {
'n_iterations': 4,
'n_initial_trials': 4,
'metric': {
'name': 'accuracy',
'optimization': 'maximize'
},
'utility_function': {
'acquisition_function': 'ei',
'eps': 1.2,
'gaussian_process': {
'kernel': 'matern',
'length_scale': 1.0,
'nu': 1.9,
'n_restarts_optimizer': 0
}
}
},
'matrix': {
'feature1': {
'values': [1, 2, 3, 4, 5]
},
'feature2': {
'linspace': [1, 5, 5]
},
'feature3': {
'range': [1, 6, 1]
},
'feature4': {
'uniform': [1, 5]
},
'feature5': {
'values': ['a', 'b', 'c']
},
}
})
self.manager2 = BOSearchManager(params_config=params_config)
def setUp(self):
super().setUp()
params_config = SettingsConfig.from_dict({
'concurrency': 2,
'hyperband': {
'max_iter': 10,
'eta': 3,
'resource': {
'name': 'steps',
'type': 'float'
},
'resume': False,
'metric': {
'name': 'loss',
'optimization': 'minimize'
}
},
'matrix': {
'feature1': {
'values': [1, 2, 3]
},
'feature2': {
'linspace': [1, 2, 5]
},
'feature3': {
'range': [1, 5, 1]
}
}
})
self.manager1 = HyperbandSearchManager(params_config=params_config)
params_config = SettingsConfig.from_dict({
'concurrency': 2,
'hyperband': {
'max_iter': 81,
'eta': 3,
'resource': {
'name': 'size',
'type': 'int'
},
'resume': False,
'metric': {
'name': 'loss',
'optimization': 'minimize'
}
},
'matrix': {
'feature1': {
'values': [1, 2, 3]
},
'feature2': {
'linspace': [1, 2, 5]
},
'feature3': {
'range': [1, 5, 1]
},
'feature4': {
'range': [1, 5, 1]
}
}
})
self.manager2 = HyperbandSearchManager(params_config=params_config)
def test_concrete_example(self):
params_config = SettingsConfig.from_dict({
'concurrency': 2,
'bo': {
'n_iterations': 5,
'n_initial_trials': 10,
'metric': {
'name': 'loss',
'optimization': 'minimize'
},
'utility_function': {
'acquisition_function': 'ucb',
'kappa': 2.576,
'gaussian_process': {
'kernel': 'matern',
'length_scale': 1.0,
'nu': 1.9,
'n_restarts_optimizer': 0
},
'n_warmup': 1,
'n_iter': 1
}
},
'matrix': {
'learning_rate': {
'uniform': [0.001, 0.01]
},
'dropout': {
'values': [0.25, 0.3]
},
'activation': {
'values': ['relu', 'sigmoid']
}
}
})
optimizer = BOOptimizer(params_config=params_config)
configs = [{
"num_epochs": 1,
"num_steps": 300,
"batch_size": 128,
"learning_rate": 0.004544653508229265,
"activation": "sigmoid",
"dropout": 0.3
}, {
"num_epochs": 1,
"num_steps": 300,
"batch_size": 128,
"learning_rate": 0.005615296199690899,
"activation": "sigmoid",
"dropout": 0.3
}, {
"num_epochs": 1,
"num_steps": 300,
"batch_size": 128,
"learning_rate": 0.008784330869587902,
"activation": "sigmoid",
"dropout": 0.25
}, {
"num_epochs": 1,
"num_steps": 300,
"batch_size": 128,
"learning_rate": 0.0058591075447430065,
"activation": "sigmoid",
"dropout": 0.3
}, {
"num_epochs": 1,
"num_steps": 300,
"batch_size": 128,
"learning_rate": 0.007464080062927171,
"activation": "sigmoid",
"dropout": 0.25
}, {
"num_epochs": 1,
"num_steps": 300,
"batch_size": 128,
"learning_rate": 0.0024763129571936738,
"activation": "relu",
"dropout": 0.3
}, {
"num_epochs": 1,
"num_steps": 300,
"batch_size": 128,
"learning_rate": 0.0074881581817925705,
"activation": "sigmoid",
"dropout": 0.3
}, {
"num_epochs": 1,
"num_steps": 300,
"batch_size": 128,
"learning_rate": 0.003360405779075163,
"activation": "relu",
"dropout": 0.3
}, {
"num_epochs": 1,
"num_steps": 300,
"batch_size": 128,
"learning_rate": 0.009916904455792564,
"activation": "sigmoid",
"dropout": 0.25
}, {
"num_epochs": 1,
"num_steps": 300,
"batch_size": 128,
"learning_rate": 0.000881723263162717,
"activation": "sigmoid",
"dropout": 0.3
}]
metrics = [
2.3018131256103516, 2.302884340286255, 2.3071441650390625,
2.3034636974334717, 2.301487922668457, 0.05087224021553993,
2.3032383918762207, 0.06383182853460312, 2.3120572566986084,
0.7617478370666504
]
optimizer.add_observations(configs=configs, metrics=metrics)
suggestion = optimizer.get_suggestion()
assert 0.001 <= suggestion['learning_rate'] <= 0.01
assert suggestion['dropout'] in [0.25, 0.3]
assert suggestion['activation'] in ['relu', 'sigmoid']