def test_dataset_serialization_deserialization_custom_basepath():
params = get_hyperparameter_list()
h = hp.HyperParameterList(params)
dataset = data.Dataset(h, basedir='custom')
samples = [(h.sample(), np.random.uniform()) for _ in range(5)]
for sample in samples:
dataset.add_sample(*sample)
# serialization
dataset.save_dataset()
assert len(dataset) == 5
assert os.path.exists(dataset.data_path)
assert os.path.exists(dataset.parameter_path)
# deserialization
dataset.clear()
assert len(dataset) == 0
dataset.restore_dataset()
assert len(dataset) == 5
assert os.path.exists(dataset.data_path)
assert os.path.exists(dataset.parameter_path)
# deserialization from class
path = os.path.join('custom', 'datasets')
dataset2 = data.Dataset.load_from_directory(path)
assert dataset2.parameters is not None
assert len(dataset2.X) == 5
assert len(dataset2.Y) == 5
assert len(dataset2) == 5
dataset3 = data.Dataset.load_from_directory('custom')
assert dataset3.parameters is not None
assert len(dataset3.X) == 5
assert len(dataset3.Y) == 5
# serialization of empty get_dataset
dataset = data.Dataset(basedir='custom')
with pytest.raises(FileNotFoundError):
dataset.load_from_directory('null')
with pytest.raises(ValueError):
dataset.save_dataset()
def test_multi_set_dataset():
params = get_multi_parameter_list()
h = hp.HyperParameterList(params)
dataset = data.Dataset(h)
# numpy arrays
samples = [(np.array(h.sample()), np.random.uniform()) for _ in range(5)]
x, y = zip(*samples)
x = np.array(x)
y = np.array(y)
dataset.set_dataset(x, y)
assert len(dataset) == 5
dataset.clear()
# python arrays
samples = [(h.sample(), float(np.random.uniform())) for _ in range(5)]
x, y = zip(*samples)
dataset.set_dataset(x, y)
assert len(dataset) == 5
# None data
with pytest.raises(TypeError):
dataset.set_dataset(None, int(6))
with pytest.raises(TypeError):
dataset.set_dataset([1, 2, 3], None)
with pytest.raises(TypeError):
dataset.set_dataset(None, None)
def test_dataset_multi_get_best_parameters():
params = get_multi_parameter_list()
h = hp.HyperParameterList(params)
dataset = data.Dataset(h)
with pytest.raises(ValueError):
dataset.get_best_parameters(None)
# Test with empty dataset
assert dataset.get_best_parameters() is None
samples = [(h.sample(), np.random.uniform()) for _ in range(5)]
for sample in samples:
dataset.add_sample(*sample)
objective_values = [v for h, v in samples]
min_index = np.argmin(objective_values)
max_index = np.argmax(objective_values)
max_hp = data.flatten_parameters(
dataset.get_best_parameters(objective='max'))
min_hp = data.flatten_parameters(
dataset.get_best_parameters(objective='min'))
assert max_hp == samples[max_index][0]
assert min_hp == samples[min_index][0]
def test_dataset_basedir_custom():
params = get_hyperparameter_list()
h = hp.HyperParameterList(params)
dataset = data.Dataset(h, basedir='custom')
assert os.path.exists(dataset.basedir)
assert not os.path.exists('shac')
def test_evaluate_train_evaluate():
params = get_hyperparameter_list()
h = hp.HyperParameterList(params)
dataset = data.Dataset(h)
# models
clfs = []
# fit samples
num_samples = 16
for i in range(3):
samples = [h.sample() for _ in range(num_samples)]
labels = [np.sum(sample) for sample in samples]
x, y = samples, labels
x, y = dataset.encode_dataset(x, y)
model = xgb_utils.train_single_model(x, y)
clfs.append(model)
# test samples
num_samples = 100
samples = [h.sample() for _ in range(num_samples)]
ex2, _ = dataset.encode_dataset(samples, None)
preds = xgb_utils.evaluate_models(ex2, clfs)
count = np.sum(preds)
print(count)
assert preds.shape == (num_samples,)
assert count > 0
def test_evaluate_single_sample():
params = get_hyperparameter_list()
h = hp.HyperParameterList(params)
dataset = data.Dataset(h)
# models
clfs = []
# fit samples
num_samples = 16
for i in range(3):
samples = [h.sample() for _ in range(num_samples)]
labels = [np.sum(sample) for sample in samples]
x, y = samples, labels
x, y = dataset.encode_dataset(x, y)
model = xgb_utils.train_single_model(x, y)
clfs.append(model)
# single sample test
sample = h.sample()
ex2, _ = dataset.encode_dataset([sample])
assert ex2.shape == (1, 3)
pred = xgb_utils.evaluate_models(ex2, clfs)
assert pred.shape == (1,)
def test_serialization_deserialization():
basepath = 'shac'
params = get_hyperparameter_list()
h = hp.HyperParameterList(params)
dataset = data.Dataset(h)
# models
clfs = []
# fit samples
num_samples = 16
for i in range(3):
samples = [h.sample() for _ in range(num_samples)]
labels = [np.sum(sample) for sample in samples]
x, y = samples, labels
x, y = dataset.encode_dataset(x, y)
model = xgb_utils.train_single_model(x, y)
clfs.append(model)
xgb_utils.save_classifiers(clfs, basepath)
assert os.path.exists(os.path.join(basepath, 'classifiers', 'classifiers.pkl'))
models = xgb_utils.restore_classifiers(basepath)
assert len(models) == len(clfs)
with pytest.raises(FileNotFoundError):
models = xgb_utils.restore_classifiers('none')
def test_dataset_parameters():
params = get_hyperparameter_list()
h = hp.HyperParameterList(params)
dataset = data.Dataset(h)
assert len(params) == len(dataset.parameters)
dataset.parameters = params
assert len(params) == len(dataset.parameters)
def __init__(self,
hyperparameter_list,
total_budget,
num_batches,
objective='max',
max_classifiers=18):
if total_budget % num_batches != 0:
raise ValueError(
"Number of epochs must be divisible by the batch size !")
if hyperparameter_list is not None and (not isinstance(
hyperparameter_list, hp.HyperParameterList)):
hyperparameter_list = hp.HyperParameterList(hyperparameter_list)
print("Number of workers possible : %d" %
(total_budget // num_batches))
self.parameters = hyperparameter_list
self.objective = objective
self._total_budget = total_budget # N
self.num_batches = num_batches # M
self._max_classifiers = max_classifiers
self._num_workers = self.total_budget // num_batches # W
self._total_classifiers = min(max(num_batches - 1, 1),
max_classifiers) # K
# serializable
self.dataset = data.Dataset(hyperparameter_list)
self.classifiers = [] # type: list(xgb.XGBClassifier)
# training variables
self._dataset_index = 0
self._per_classifier_budget = int(
self.num_workers *
np.floor(total_budget /
(float(self.num_workers *
(self.total_classifiers + 1))))) # Tc
print(
"Using %d parallel workers, it will require %d epochs to fit %d classifiers.\n"
"Each classifier will be provided %d samples to train per epoch." %
(
self.num_workers,
total_budget // self.num_workers,
self._total_classifiers,
self._per_classifier_budget,
))
# Compute how many threads and processes will be used
self._compute_parallelism()
# serialization paths
self._prepare_dirs()
def test_dataset_multi_param_list():
params = get_multi_parameter_list()
dataset = data.Dataset(params)
assert isinstance(dataset._parameters, hp.HyperParameterList)
dataset.set_parameters(params)
assert isinstance(dataset._parameters, hp.HyperParameterList)
h = hp.HyperParameterList(params)
dataset.set_parameters(h)
assert isinstance(dataset._parameters, hp.HyperParameterList)
def test_dataset_multi_add_sample():
params = get_multi_parameter_list()
h = hp.HyperParameterList(params)
dataset = data.Dataset(h)
samples = [(h.sample(), np.random.uniform()) for _ in range(5)]
for sample in samples:
dataset.add_sample(*sample)
x, y = dataset.get_dataset()
assert len(dataset) == 5
assert x.shape == (5, 14)
assert y.shape == (5, )
def test_dataset_single_multi_encoding_decoding_min():
params = get_multi_parameter_list()
h = hp.HyperParameterList(params)
dataset = data.Dataset(h)
sample = (h.sample(), np.random.uniform())
dataset.add_sample(*sample)
encoded_x, encoded_y = dataset.encode_dataset(objective='min')
y_values = [0.]
assert encoded_x.shape == (1, 14)
assert encoded_x.dtype == np.float64
assert encoded_y.shape == (1, )
assert encoded_y.dtype == np.float64
assert np.allclose(y_values, encoded_y, rtol=1e-3)
decoded_x = dataset.decode_dataset(encoded_x)
assert decoded_x.shape == (1, 14)
def test_dataset_multi_encoding_decoding():
params = get_multi_parameter_list()
h = hp.HyperParameterList(params, seed=0)
dataset = data.Dataset(h)
samples = [(h.sample(), np.random.uniform()) for _ in range(5)]
for sample in samples:
dataset.add_sample(*sample)
encoded_x, encoded_y = dataset.encode_dataset(objective='min')
y_values = [0., 0., 0., 1., 1.]
assert encoded_x.shape == (5, 14)
assert encoded_x.dtype == np.float64
assert encoded_y.shape == (5, )
assert encoded_y.dtype == np.float64
assert np.allclose(y_values, encoded_y, rtol=1e-3)
decoded_x = dataset.decode_dataset(encoded_x)
decoded_x2 = dataset.decode_dataset()
assert decoded_x.shape == (5, 14)
assert len(decoded_x) == len(decoded_x2)
x, y = dataset.get_dataset()
x_ = x[:, :10].astype('float')
decoded_x_ = decoded_x[:, :10].astype('float')
assert np.allclose(x_, decoded_x_, rtol=1e-3)
samples2 = [(h.sample(), np.random.uniform()) for _ in range(5)]
x, y = zip(*samples2)
encoded_x, encoded_y = dataset.encode_dataset(x, y, objective='min')
y_values = [0., 1., 0., 0., 1.]
assert encoded_x.shape == (5, 14)
assert encoded_x.dtype == np.float64
assert encoded_y.shape == (5, )
assert encoded_y.dtype == np.float64
assert np.allclose(y_values, encoded_y, rtol=1e-3)
def test_evaluate_train_evaluate_failure():
params = [hp.DiscreteHyperParameter('h%d' % i, [0]) for i in range(3)]
h = hp.HyperParameterList(params)
dataset = data.Dataset(h)
# models
clfs = []
# fit samples
num_samples = 16
for i in range(3):
samples = [h.sample() for _ in range(num_samples)]
labels = [np.sum(sample) for sample in samples]
x, y = samples, labels
x, y = dataset.encode_dataset(x, y)
model = xgb_utils.train_single_model(x, y)
clfs.append(model)
# test samples
for model in clfs:
assert model is None
def test_dataset_basedir():
params = get_hyperparameter_list()
h = hp.HyperParameterList(params)
dataset = data.Dataset(h)
assert os.path.exists(dataset.basedir)
def test_dataset_serialization_deserialization_custom_param():
class MockDiscreteHyperParameter(hp.DiscreteHyperParameter):
def __init__(self, name, values, seed=None):
super(MockDiscreteHyperParameter,
self).__init__(name, values, seed)
# register the new hyper parameters
hp.set_custom_parameter_class(MockDiscreteHyperParameter)
params = get_hyperparameter_list()
params.append(MockDiscreteHyperParameter('mock-param', ['x', 'y']))
h = hp.HyperParameterList(params, seed=0)
dataset = data.Dataset(h)
samples = [(h.sample(), np.random.uniform()) for _ in range(5)]
for sample in samples:
dataset.add_sample(*sample)
# serialization
dataset.save_dataset()
assert len(dataset) == 5
assert os.path.exists(dataset.data_path)
assert os.path.exists(dataset.parameter_path)
# deserialization
dataset.clear()
assert len(dataset) == 0
dataset.restore_dataset()
assert len(dataset) == 5
assert os.path.exists(dataset.data_path)
assert os.path.exists(dataset.parameter_path)
# deserialization from class
path = os.path.join('shac', 'datasets')
dataset2 = data.Dataset.load_from_directory(path)
assert dataset2.parameters is not None
assert len(dataset2.X) == 5
assert len(dataset2.Y) == 5
assert len(dataset2) == 5
assert 'mock-param' in dataset2.parameters.name_map.values()
assert dataset2.parameters.num_choices == 5
dataset3 = data.Dataset.load_from_directory()
assert dataset3.parameters is not None
assert len(dataset3.X) == 5
assert len(dataset3.Y) == 5
assert 'mock-param' in dataset3.parameters.name_map.values()
assert dataset3.parameters.num_choices == 5
# serialization of empty get_dataset
dataset = data.Dataset()
with pytest.raises(FileNotFoundError):
dataset.load_from_directory('null')
with pytest.raises(ValueError):
dataset.save_dataset()