def _build_pipeline(self):
self._pipeline = MLPipeline(self.template)
if self._hyperparameters:
self._pipeline.set_hyperparameters(self._hyperparameters)
self.fitted = False
def _get_tunables(cls, template_dicts):
tunables = {}
for name, template in template_dicts.items():
pipeline = MLPipeline(template)
pipeline_tunables = pipeline.get_tunable_hyperparameters(flat=True)
tunables[name] = Tunable.from_dict(pipeline_tunables)
return tunables
def _load_pipeline(pipeline, hyperparams=None):
if isinstance(pipeline, str) and os.path.isfile(pipeline):
pipeline = MLPipeline.load(pipeline)
else:
pipeline = MLPipeline(pipeline)
if hyperparams is not None:
pipeline.set_hyperparameters(hyperparams)
return pipeline
def _get_mlpipeline(self):
pipeline = self._pipeline
if isinstance(pipeline, str) and os.path.isfile(pipeline):
with open(pipeline) as json_file:
pipeline = json.load(json_file)
mlpipeline = MLPipeline(pipeline)
if self._hyperparameters:
mlpipeline.set_hyperparameters(self._hyperparameters)
return mlpipeline
def _load_pipeline(pipeline):
if isinstance(pipeline, MLPipeline):
return pipeline
if isinstance(pipeline, str):
return MLPipeline.load(pipeline)
if isinstance(pipeline, dict):
return MLPipeline.from_dict(pipeline)
raise ValueError('Invalid pipeline %s', pipeline)
def build_pipeline(pipeline_spec):
pipeline = MLPipeline(
pipeline_spec['primitives'],
pipeline_spec.get('init_params', dict()),
pipeline_spec.get('input_names', dict()),
pipeline_spec.get('output_names', dict()),
)
hyperparameters = pipeline_spec.get('hyperparameters')
if hyperparameters:
pipeline.set_hyperparameters(hyperparameters)
return pipeline
def _load_mlpipeline(self, template):
if not isinstance(template, dict):
template = self._load_template(template)
self.template = template
return MLPipeline.from_dict(template)
def cv_score(self, X, y, context, metric=None, cv=None):
"""Cross Validate this pipeline."""
scorer = METRICS_DICT[metric or self.metric]
LOGGER.debug('CV Scoring pipeline %s', self)
self.cv_scores = list()
for fold, (train_index, test_index) in enumerate(cv.split(X, y)):
LOGGER.debug('Scoring fold: %s', fold)
X_train, y_train = self._get_split(X, y, train_index)
pipeline = MLPipeline.from_dict(self._tunable)
pipeline.fit(X_train, y_train, **context)
X_test, y_test = self._get_split(X, y, test_index)
pred = pipeline.predict(X_test, **context)
score = scorer(pred, y_test)
self.cv_scores.append(score)
LOGGER.debug('Fold %s score: %s', fold, score)
score, std, rank = self._get_score()
LOGGER.debug('CV score: %s +/- %s; rank: %s', score, std, rank)
self.score = score
self.std = std
self.rank = rank + random.random() * 1.e-12 # to avoid collisions
def test_jsons():
"""Validate MLBlocks primitive jsons"""
primitives = (f for f in os.listdir(MLPRIMITIVES_JSONS_PATH) if f.endswith('.json'))
for primitive_filename in primitives:
try:
primitive_path = os.path.join(MLPRIMITIVES_JSONS_PATH, primitive_filename)
with open(primitive_path, 'r') as f:
primitive = json.load(f)
primitive_name = primitive['name']
fixed_hyperparameters = primitive.get('hyperparameters', dict()).get('fixed', dict())
init_hyperparameters = dict()
for name, hyperparameter in fixed_hyperparameters.items():
if 'default' not in hyperparameter:
type_ = hyperparameter.get('type')
init_hyperparameters[name] = HYPERPARAMETER_DEFAULTS.get(type_)
block_name = primitive_name + '#1'
mlpipeline = MLPipeline([primitive_name], {block_name: init_hyperparameters})
# Validate methods
mlblock = mlpipeline.blocks[block_name]
if mlblock._class:
fit = primitive.get('fit')
if fit:
assert hasattr(mlblock.instance, fit['method'])
produce = primitive['produce']
assert hasattr(mlblock.instance, produce['method'])
except Exception:
raise ValueError("Invalid JSON primitive: {}".format(primitive_filename))
def load(self):
"""Load this Template as an MLPipeline.
Returns:
MLPipeline
"""
return MLPipeline(self.json)
def pipeline_score(pipeline_dict, X, y, scorer, context=None,
n_splits=5, cv=None, random_state=0):
context = context or dict()
LOGGER.debug('CV Scoring pipeline %s')
cv_scores = list()
if not cv:
metadata = pipeline_dict.get('metadata', pipeline_dict.get('loader', dict()))
if metadata.get('task_type') == 'classification':
cv = StratifiedKFold(n_splits=n_splits, shuffle=True, random_state=random_state)
else:
cv = KFold(n_splits=n_splits, shuffle=True, random_state=random_state)
for fold, (train_index, test_index) in enumerate(cv.split(X, y)):
LOGGER.debug('Scoring fold: %s', fold)
X_train, y_train = get_split(X, y, train_index)
pipeline = MLPipeline.from_dict(pipeline_dict)
pipeline.fit(X_train, y_train, **context)
X_test, y_test = get_split(X, y, test_index)
pred = pipeline.predict(X_test, **context)
score = scorer(pred, y_test)
cv_scores.append(score)
LOGGER.debug('Fold %s score: %s', fold, score)
return np.mean(cv_scores), np.std(cv_scores)
def fit(self, data_params):
"""Fit the pipeline on the given params."""
X, y = data_params.X, data_params.y
self.pipeline = MLPipeline.from_dict(self.pipeline_dict)
self.pipeline.fit(X, y, **data_params.context)
self.fitted = True
def pipeline(self):
"""Pipeline.
Returns:
MLPipeline:
The pipeline in the modeler.
"""
return MLPipeline(self._pipeline)
def _generate_splits(self, X, y, readings):
if self._preprocessing:
pipeline = MLPipeline(self.template)
LOGGER.debug('Running %s preprocessing steps', self._preprocessing)
context = pipeline.fit(X=X,
y=y,
readings=readings,
output_=self._preprocessing - 1)
del context['X']
del context['y']
else:
context = {'readings': readings}
splits = list()
for fold, (train_index, test_index) in enumerate(self._cv.split(X, y)):
LOGGER.debug('Running static steps for fold %s', fold)
X_train, X_test = X.iloc[train_index], X.iloc[test_index]
y_train, y_test = y.iloc[train_index], y.iloc[test_index]
pipeline = MLPipeline(self.template)
fit = pipeline.fit(X_train,
y_train,
output_=self._static - 1,
start_=self._preprocessing,
**context)
predict = pipeline.predict(X_test,
output_=self._static - 1,
start_=self._preprocessing,
**context)
splits.append((fold, pipeline, fit, predict, y_test))
return splits
def test_load_orion_mlpipeline(self, tmpdir):
pipeline = MLPipeline('dummy')
orion = functional._load_orion(pipeline)
assert isinstance(orion, Orion)
assert orion._pipeline == pipeline
assert not orion._fitted
assert orion._hyperparameters is None
def test_mlpipeline(self):
primitives = ['sklearn.ensemble.RandomForestClassifier']
init_params = {
'sklearn.ensemble.RandomForest#1': {
'n_estimators': 500
}
}
pipeline = MLPipeline(primitives=primitives, init_params=init_params)
pipeline2 = MLPipeline(pipeline)
assert pipeline2.primitives == [
'sklearn.ensemble.RandomForestClassifier'
]
assert pipeline2.init_params == {
'sklearn.ensemble.RandomForest#1': {
'n_estimators': 500
}
}
def preprocess(self, X, y, context):
"""Execute the preprocessing steps of the pipeline."""
if self._preprocessing:
LOGGER.info("Executing preprocessing pipeline")
pipeline = MLPipeline.from_dict(self._preprocessing)
pipeline.fit(X, y, **context)
return pipeline.predict(X, **context)
else:
LOGGER.info("No preprocessing steps found")
return X
def create_pipeline(self, primitives, hyperparameters=None):
"""Creates a pipeline of primitives.
Args:
primitives: A list of primitive.
hyperparameters: A dictionary of hyperparameters for each primitives.
Returns:
A MLPipeline instance.
"""
self.primitive = self.check_path(primitives)
if hyperparameters is not None:
hyperparameters = self.check_path_hyperparameters(hyperparameters)
pipeline = MLPipeline(self.primitive, hyperparameters)
else:
pipeline = MLPipeline(self.primitive)
return pipeline
def _generate_splits(self, template_name, target_times, readings, turbines=None):
template = self._template_dicts.get(template_name)
pipeline = MLPipeline(template)
preprocessing = self._preprocessing.get(template_name)
static = self._count_static_steps(pipeline)
X = target_times[['turbine_id', 'cutoff_time']]
y = target_times['target']
if preprocessing:
if preprocessing > static:
raise ValueError('Preprocessing cannot be bigger than static')
LOGGER.debug('Running %s preprocessing steps', preprocessing)
context = pipeline.fit(X=X, y=y, readings=readings,
turbines=turbines, output_=preprocessing - 1)
del context['X']
del context['y']
gc.collect()
else:
context = {
'readings': readings,
'turbines': turbines,
}
splits = list()
for fold, (train_index, test_index) in enumerate(self._cv.split(X, y)):
LOGGER.debug('Running static steps for fold %s', fold)
gc.collect()
X_train, X_test = X.iloc[train_index], X.iloc[test_index]
y_train, y_test = y.iloc[train_index], y.iloc[test_index]
pipeline = MLPipeline(template)
fit = pipeline.fit(X_train, y_train, output_=static - 1,
start_=preprocessing, **context)
predict = pipeline.predict(X_test, output_=static - 1,
start_=preprocessing, **context)
split = (fold, pipeline, fit, predict, y_test, static)
if self._cache_path:
split_name = '{}_{}.pkl'.format(template_name, fold)
split_path = os.path.join(self._cache_path, split_name)
os.makedirs(os.path.dirname(split_path), exist_ok=True)
with open(split_path, 'wb') as split_file:
pickle.dump(split, split_file)
split = split_path
splits.append(split)
gc.collect()
return splits
def add_template(self, name, template=None):
"""Add a new Template object to the database.
The template can be passed as a name of a registered MLPipeline,
or as a path to an MLPipeline JSON specification, or as a full
dictionary specification of an MLPipeline or directly as an
MLPipeline instance.
If the ``template`` argument is not passed, the given ``name`` will
be used to load an MLPipeline.
During this step, apart from the Template object, a new Pipeline object
using the default hyperparameters and with the same name as the
Template will also be created.
Args:
name (str):
Name of the Template.
template (str, dict or MLPipeline):
Name of the MLBlocks template to load or path to its JSON
file or dictionary specification or MLPipeline instance.
If not given, the ``name`` of the template is used.
Raises:
NotUniqueError:
If a Template with the same name already exists.
Returns:
Template
"""
template = template or name
if isinstance(template, str) and os.path.isfile(template):
with open(template, 'r') as f:
template = json.load(f)
pipeline_dict = MLPipeline(template).to_dict()
template = schema.Template.insert(
name=name,
json=pipeline_dict,
created_by=self.user
)
schema.Pipeline.insert(
name=name,
template=template,
json=pipeline_dict,
created_by=self.user
)
return template
def _load_mlpipeline(self, template):
if not isinstance(template, dict):
template_name = template
if os.path.isfile(template_name):
with open(template_name, 'r') as template_file:
template = json.load(template_file)
elif self._db:
template = self._db.load_template(template_name)
if not template:
raise ValueError('Unknown template {}'.format(template_name))
self.template = template
return MLPipeline.from_dict(template)
def train_mlblocks(X_train, X_test, y_train, y_test, mtype, common_name_model,
problemtype, classes, default_featurenames, transform_model,
settings, model_session):
# name model
model_name = common_name_model + '.pickle'
files = list()
if mtype == 'c':
# aet up primitives and train model
primitives = ['sklearn.impute.SimpleImputer', 'xgboost.XGBClassifier']
init_params = {
'sklearn.impute.SimpleImputer': {
'strategy': 'median'
},
'xgboost.XGBClassifier': {
'learning_rate': 0.1
}
}
pipeline = MLPipeline(primitives, init_params=init_params)
pipeline.fit(X_train, y_train)
if mtype == 'r':
# aet up primitives and train model
primitives = ['sklearn.impute.SimpleImputer', 'xgboost.XGBRegressor']
pipeline = MLPipeline(primitives)
pipeline.fit(X_train, y_train)
# saving model
print('saving model')
modelfile = open(model_name, 'wb')
pickle.dump(pipeline, modelfile)
modelfile.close()
# get model directory
files.append(model_name)
model_dir = os.getcwd()
return model_name, model_dir, files
def test_dict(self):
pipeline_dict = {
'primitives': ['sklearn.ensemble.RandomForestClassifier'],
'init_params': {
'sklearn.ensemble.RandomForest#1': {
'n_estimators': 500
}
},
'input_names': {
'sklearn.ensemble.RandomForest#1': {
'X': 'X1'
}
},
'output_names': {
'sklearn.ensemble.RandomForest#1': {
'y': 'y1'
}
}
}
pipeline = MLPipeline(pipeline_dict)
assert pipeline.primitives == [
'sklearn.ensemble.RandomForestClassifier'
]
assert pipeline.init_params == {
'sklearn.ensemble.RandomForest#1': {
'n_estimators': 500
}
}
assert pipeline.input_names == {
'sklearn.ensemble.RandomForest#1': {
'X': 'X1'
}
}
assert pipeline.output_names == {
'sklearn.ensemble.RandomForest#1': {
'y': 'y1'
}
}
def __init__(self, pipeline_dict, loader, metric, problem_doc):
self.pipeline_dict = pipeline_dict
self.name = pipeline_dict['name']
self.template = pipeline_dict.get('template')
self.loader = loader
self.metric = metric
self.problem_doc = problem_doc
preprocessing_blocks = self.pipeline_dict.get('preprocessing_blocks')
if preprocessing_blocks:
preprocessing = pipeline_dict.copy()
preprocessing_primitives = preprocessing[
'primitives'][:preprocessing_blocks]
preprocessing['primitives'] = preprocessing_primitives
self._preprocessing = preprocessing
tunable = pipeline_dict.copy()
tunable_primitives = tunable['primitives'][preprocessing_blocks:]
tunable['primitives'] = tunable_primitives
self._tunable = tunable
pre_params, tun_params = self._extract_hyperparameters(
preprocessing_primitives)
self._preprocessing['hyperparameters'] = pre_params
self._tunable['hyperparameters'] = tun_params
else:
self._preprocessing = None
self._tunable = pipeline_dict
self.id = str(uuid.uuid4())
self.cv_scores = list()
self.rank = None
self.score = None
self.dumped = False
self.fitted = False
self.pipeline = MLPipeline.from_dict(pipeline_dict)
def score_pipeline(pipeline_metadata, n_splits=5, random_state=0, dataset=None):
if isinstance(pipeline_metadata, str):
LOGGER.info('Loading pipeline %s', pipeline_metadata)
with open(pipeline_metadata, 'r') as pipeline_file:
pipeline_metadata = json.load(pipeline_file)
validation = pipeline_metadata['validation']
if dataset is None:
dataset = validation['dataset']
LOGGER.info('Loading dataset %s', dataset)
dataset = load_dataset(dataset)
metric = validation.get('metric')
metric_args = validation.get('metric_args', dict())
if metric:
scorer = get_scorer(metric, metric_args)
else:
scorer = dataset.score
metric = dataset.metric
scores = list()
splits = dataset.get_splits(n_splits, random_state)
if n_splits == 1:
splits = [splits]
for split, (X_train, X_test, y_train, y_test) in enumerate(splits):
LOGGER.info('Scoring split %s', split + 1)
context = get_context(dataset, validation.get('context', dict()))
pipeline = MLPipeline.from_dict(pipeline_metadata)
pipeline.fit(X_train, y_train, **context)
predictions = pipeline.predict(X_test, **context)
score = scorer(y_test, predictions)
LOGGER.info('Split %s %s: %s', split + 1, metric, score)
scores.append(score)
return np.mean(scores), np.std(scores)
def k_fold_validation(self, hyperparameters, X, y, scoring=None):
"""Score the pipeline through k-fold validation with the given scoring function.
Args:
hyperparameters (dict or None):
A dictionary of hyper-parameters for each primitive in the target pipeline.
X (pandas.DataFrame or ndarray):
Inputs to the pipeline.
y (pandas.Series or ndarray):
Target values.
scoring (str):
The name of the scoring function.
Returns:
np.float64:
The average score in the k-fold validation.
"""
model_instance = MLPipeline(self._pipeline)
X = pd.DataFrame(X)
y = pd.Series(y)
if hyperparameters:
model_instance.set_hyperparameters(hyperparameters)
if self._problem_type == 'regression':
scorer = self.regression_metrics[scoring or 'R2 Score']
else:
scorer = self.classification_metrics[scoring or 'F1 Macro']
scores = []
kf = KFold(n_splits=10, random_state=None, shuffle=True)
for train_index, test_index in kf.split(X):
model_instance.fit(X.iloc[train_index], y.iloc[train_index])
y_pred = model_instance.predict(X.iloc[test_index])
scores.append(scorer(y.iloc[test_index], y_pred))
return np.mean(scores)
def load_pipeline(self, pipeline):
LOGGER.info("Loading pipeline %s", pipeline.name)
return MLPipeline.from_dict(pipeline.mlpipeline)
class GreenGuardPipeline(object):
"""Main Machine Learning component in the GreenGuard project.
The ``GreenGuardPipeline`` represents the abstraction of a Machine
Learning pipeline architecture specialized on the GreenGuard data
format.
In order to use it, an MLBlocks pipeline template needs to be given,
alongside information about how to evaluate its performance using
cross validation.
Attributes:
template (MLPipeline):
MLPipeline instance used as the template for tuning.
template_name:
Name of the template being used.
fitted (bool):
Whether this GreenGuardPipeline has already been fitted or not.
steps (list):
List of primitives that compose this template.
preprocessing (list):
List of preprocessing steps. These steps have no learning stage
and are executed only once on the complete training dataset, before
partitioning it for cross validation.
static (list):
List of static steps. These are all the steps in the pipeline that
come after the preprocessing ones but have no hyperparameters.
These are executed on each cross validation split only once, when
the data is partitioned, and their output is cached to be reused
later on at every tuning iteration.
tunable (list):
List of steps that have hyperparameters and will be tuned during
the tuning loop.
Args:
templates (str, MLPipeline or list):
Template to use. If a ``str`` is given, load the corresponding
``MLPipeline``. Also can be a list combining both.
metric (str or function):
Metric to use. If an ``str`` is give it must be one of the metrics
defined in the ``greenguard.metrics.METRICS`` dictionary.
cost (bool):
Whether the metric is a cost function (the lower the better) or not.
Defaults to ``False``.
init_params (dict or list):
There are three possible values for init_params:
* Init params ``dict``: It will be used for all templates.
* ``dict`` with the name of the template as a key and dictionary with its
init params.
* ``list``: each value will be assigned to the corresponding position of
self.templates.
Defaults to ``None``.
stratify (bool):
Whether to stratify the data when partitioning for cross validation.
Defaults to ``True``.
cv_splits (int):
Number of cross validation folds to use. Defaults to ``5``.
shuffle (bool):
Whether to shuffle the data when partitioning for cross validation.
Defaults to ``True``.
random_state (int or RandomState):
random state to use for the cross validation partitioning.
Defaults to ``0``.
preprocessing (int, dict or list):
There are three possible values for preprocessing:
* ``int``: the value will be used for all templates.
* ``dict`` with the template name as a key and a number as a value, will
be used for that template.
* ``list``: each value will be assigned to the corresponding position of
self.templates.
Defaults to ``0``.
cache_path (str):
If given, cache the generated cross validation splits in this folder.
Defatuls to ``None``.
"""
template = None
template_name = None
fitted = False
cv_score = None
_cv_class = None
_metric = None
_cost = False
_tuner = None
_pipeline = None
_static = None
_init_params = None
_preprocessing = None
def _get_cv(self, stratify, cv_splits, shuffle, random_state):
if stratify:
cv_class = StratifiedKFold
else:
cv_class = KFold
return cv_class(n_splits=cv_splits, shuffle=shuffle, random_state=random_state)
def _set_hyperparameters(self, new_hyperparameters):
self._hyperparameters = deepcopy(new_hyperparameters)
def _set_template(self, template_name):
self.template_name = template_name
self.template = self._template_dicts[self.template_name]
@staticmethod
def _update_params(old, new):
for name, params in new.items():
if '#' not in name:
name = name + '#1'
block_params = old.setdefault(name, dict())
for param, value in params.items():
block_params[param] = value
def _count_static_steps(self, pipeline):
tunable_hyperparams = pipeline.get_tunable_hyperparameters()
for index, block_name in enumerate(pipeline.blocks.keys()):
if tunable_hyperparams[block_name]:
return index
return 0
def _get_templates(self, templates):
template_dicts = dict()
template_names = list()
for template in templates:
if isinstance(template, str):
template_name = template
template = load_pipeline(template_name)
else:
template_name = md5(json.dumps(template)).digest()
template_dicts[template_name] = template
template_names.append(template_name)
return template_names, template_dicts
def _generate_init_params(self, init_params):
if not init_params:
self._init_params = {}
elif isinstance(init_params, list):
self._init_params = dict(zip(self._template_names, init_params))
elif any(name in init_params for name in self._template_names):
self._init_params = init_params
def _generate_preprocessing(self, preprocessing):
if isinstance(preprocessing, int):
self._preprocessing = {name: preprocessing for name in self._template_names}
else:
if isinstance(preprocessing, list):
preprocessing = dict(zip(self._template_names, preprocessing))
self._preprocessing = {
name: preprocessing.get(name, 0)
for name in self._template_names
}
def _build_pipeline(self):
self._pipeline = MLPipeline(self.template)
if self._hyperparameters:
self._pipeline.set_hyperparameters(self._hyperparameters)
self.fitted = False
def __init__(self, templates, metric='accuracy', cost=False, init_params=None, stratify=True,
cv_splits=5, shuffle=True, random_state=0, preprocessing=0, cache_path=None):
if isinstance(metric, str):
metric, cost = METRICS[metric]
self._metric = metric
self._cost = cost
self._cv = self._get_cv(stratify, cv_splits, shuffle, random_state)
self.cv_score = np.inf if cost else -np.inf
if not isinstance(templates, list):
templates = [templates]
self.templates = templates
self._template_names, self._template_dicts = self._get_templates(templates)
self._default_init_params = {}
self._generate_init_params(init_params)
for name, template in self._template_dicts.items():
init_params = self._init_params.get(name, self._default_init_params)
template_params = template.setdefault('init_params', {})
self._update_params(template_params, init_params)
self._generate_preprocessing(preprocessing)
self._set_template(self._template_names[0])
self._hyperparameters = dict()
self._build_pipeline()
self._cache_path = cache_path
if cache_path:
os.makedirs(cache_path, exist_ok=True)
def get_hyperparameters(self):
"""Get the current hyperparameters.
Returns:
dict:
Current hyperparameters.
"""
return deepcopy(self._hyperparameters)
def _is_better(self, score):
if self._cost:
return score < self.cv_score
return score > self.cv_score
def _generate_splits(self, template_name, target_times, readings, turbines=None):
template = self._template_dicts.get(template_name)
pipeline = MLPipeline(template)
preprocessing = self._preprocessing.get(template_name)
static = self._count_static_steps(pipeline)
X = target_times[['turbine_id', 'cutoff_time']]
y = target_times['target']
if preprocessing:
if preprocessing > static:
raise ValueError('Preprocessing cannot be bigger than static')
LOGGER.debug('Running %s preprocessing steps', preprocessing)
context = pipeline.fit(X=X, y=y, readings=readings,
turbines=turbines, output_=preprocessing - 1)
del context['X']
del context['y']
gc.collect()
else:
context = {
'readings': readings,
'turbines': turbines,
}
splits = list()
for fold, (train_index, test_index) in enumerate(self._cv.split(X, y)):
LOGGER.debug('Running static steps for fold %s', fold)
gc.collect()
X_train, X_test = X.iloc[train_index], X.iloc[test_index]
y_train, y_test = y.iloc[train_index], y.iloc[test_index]
pipeline = MLPipeline(template)
fit = pipeline.fit(X_train, y_train, output_=static - 1,
start_=preprocessing, **context)
predict = pipeline.predict(X_test, output_=static - 1,
start_=preprocessing, **context)
split = (fold, pipeline, fit, predict, y_test, static)
if self._cache_path:
split_name = '{}_{}.pkl'.format(template_name, fold)
split_path = os.path.join(self._cache_path, split_name)
with open(split_path, 'wb') as split_file:
pickle.dump(split, split_file)
split = split_path
splits.append(split)
gc.collect()
return splits
def _cross_validate(self, template_splits, hyperparams):
scores = []
for split in template_splits:
gc.collect()
if self._cache_path:
with open(split, 'rb') as split_file:
split = pickle.load(split_file)
fold, pipeline, fit, predict, y_test, static = split
LOGGER.debug('Scoring fold %s', fold)
pipeline.set_hyperparameters(hyperparams)
pipeline.fit(start_=static, **fit)
predictions = pipeline.predict(start_=static, **predict)
score = self._metric(y_test, predictions)
LOGGER.debug('Fold fold %s score: %s', fold, score)
scores.append(score)
return np.mean(scores)
def _make_btb_scorer(self, target_times, readings, turbines):
splits = {}
def scorer(template_name, config):
template_splits = splits.get(template_name)
if template_splits is None:
template_splits = self._generate_splits(
template_name, target_times, readings, turbines)
splits[template_name] = template_splits
cv_score = self._cross_validate(template_splits, config)
if self._is_better(cv_score):
_config = '\n'.join(' {}: {}'.format(n, v) for n, v in config.items())
LOGGER.info(('New configuration found:\n'
' Template: %s \n'
' Hyperparameters: \n'
'%s'), template_name, _config)
self.cv_score = cv_score
self._set_template(template_name)
self._set_hyperparameters(config)
self._build_pipeline()
return cv_score
return scorer
def cross_validate(self, target_times, readings, turbines,
template_name=None, hyperparams=None):
"""Compute cross validation score using the given data.
If the splits have not been previously computed, compute them now.
During this computation, the data is partitioned using the indicated
cross validation parameters and later on processed using the
pipeline static steps.
The results of the fit and produce executions are cached and reused
in subsequent calls to this method.
Args:
X (pandas.DataFrame):
``target_times`` data, without the ``target`` column.
Only needed if the splits have not been previously computed.
y (pandas.Series or numpy.ndarray):
``target`` vector corresponding to the passed ``target_times``.
Only needed if the splits have not been previously computed.
readings (pandas.DataFrame):
``readings`` table. Only needed if the splits have not been
previously computed.
turbines (pandas.DataFrame):
``turbines`` table. Only needed if the splits have not been
previously computed.
params (dict):
hyperparameter values to use.
Returns:
float:
Computed cross validation score. This score is the average
of the scores obtained accross all the cross validation folds.
"""
if not template_name:
template_name = self.template_name
if hyperparams is None:
hyperparams = self.get_hyperparameters()
elif hyperparams is None:
hyperparams = {}
template_splits = self._generate_splits(template_name, target_times, readings, turbines)
return self._cross_validate(template_splits, hyperparams)
@classmethod
def _get_tunables(cls, template_dicts):
tunables = {}
for name, template in template_dicts.items():
pipeline = MLPipeline(template)
pipeline_tunables = pipeline.get_tunable_hyperparameters(flat=True)
tunables[name] = Tunable.from_dict(pipeline_tunables)
return tunables
def tune(self, target_times, readings, turbines=None):
"""Create a tuning session object that tunes and selects the templates.
Args:
target_times (pandas.DataFrame):
``target_times`` table, containing the ``turbine_id``, ``cutoff_time``
and ``target`` columns.
Only needed if the splits have not been previously computed.
readings (pandas.DataFrame):
``readings`` table. Only needed if the splits have not been
previously computed.
turbines (pandas.DataFrame):
``turbines`` table. Only needed if the splits have not been
previously computed.
"""
scoring_function = self._make_btb_scorer(target_times, readings, turbines)
tunables = self._get_tunables(self._template_dicts)
return BTBSession(tunables, scoring_function, maximize=not self._cost)
def fit(self, target_times, readings, turbines=None):
"""Fit this pipeline to the given data.
Args:
target_times (pandas.DataFrame):
``target_times`` table, containing the ``turbine_id``, ``cutoff_time``
and ``target`` columns.
readings (pandas.DataFrame):
``readings`` table.
turbines (pandas.DataFrame):
``turbines`` table.
"""
X = target_times[['turbine_id', 'cutoff_time']]
y = target_times['target']
self._pipeline.fit(X, y, readings=readings, turbines=turbines)
self.fitted = True
def predict(self, target_times, readings, turbines=None):
"""Make predictions using this pipeline.
Args:
target_times (pandas.DataFrame):
``target_times`` table, containing the ``turbine_id``, ``cutoff_time``
and ``target`` columns.
readings (pandas.DataFrame):
``readings`` table.
turbines (pandas.DataFrame):
``turbines`` table.
Returns:
numpy.ndarray:
Vector of predictions.
"""
if not self.fitted:
raise NotFittedError()
X = target_times[['turbine_id', 'cutoff_time']]
return self._pipeline.predict(X, readings=readings, turbines=turbines)
def save(self, path):
"""Serialize and save this pipeline using cloudpickle.
Args:
path (str):
Path to the file where the pipeline will be saved.
"""
with open(path, 'wb') as pickle_file:
cloudpickle.dump(self, pickle_file)
@classmethod
def load(cls, path):
"""Load a previously saved pipeline from a file.
Args:
path (str):
Path to the file where the pipeline is saved.
Returns:
GreenGuardPipeline:
Loaded GreenGuardPipeline instance.
"""
with open(path, 'rb') as pickle_file:
return cloudpickle.load(pickle_file)
def test_SigPro_nested_pipeline():
"""Test nested sigpro primitive."""
# setup
aggregations = [{
'primitive': 'sigpro.SigPro',
'init_params': {
'keep_columns':
True,
'input_is_dataframe':
False,
'values_column_name':
'amplitude_values',
'transformations': [{
'primitive':
'sigpro.transformations.frequency.band.frequency_band',
'init_params': {
'low': 100,
'high': 200
}
}],
'aggregations': [{
'primitive':
'sigpro.aggregations.amplitude.statistical.mean'
}]
}
}, {
'primitive': 'sigpro.SigPro',
'init_params': {
'input_is_dataframe':
False,
'values_column_name':
'amplitude_values',
'transformations': [{
'primitive':
'sigpro.transformations.frequency.band.frequency_band',
'init_params': {
'low': 3000,
'high': 4000
}
}],
'aggregations': [{
'name':
'band_3k_4k_mean',
'primitive':
'sigpro.aggregations.amplitude.statistical.mean'
}],
}
}, {
'primitive':
'sigpro.aggregations.amplitude.statistical.mean'
}]
pipeline = MLPipeline({
'primitives': ['sigpro.SigPro'],
'init_params': {
'sigpro.SigPro#1': {
'transformations': [{
'primitive':
'sigpro.transformations.frequency.fft.fft_real'
}],
'aggregations':
aggregations
}
}
})
data = pd.DataFrame({
'timestamp': pd.to_datetime(['2020-01-01 00:00:00']),
'values': [[1, 2, 3, 4, 5, 6]],
'sampling_frequency': [10000],
'dummy': [1],
})
# run
output = pipeline.predict(readings=data)
outputs = dict(zip(pipeline.get_output_names(), output))
# assert
expected_features = [
'fft_real.SigPro.frequency_band.mean.mean_value',
'fft_real.SigPro.frequency_band.band_3k_4k_mean.mean_value',
'fft_real.mean.mean_value'
]
assert outputs['feature_columns'] == expected_features
expected_readings = pd.DataFrame({
'fft_real.SigPro.frequency_band.mean.mean_value': [float('nan')],
'fft_real.SigPro.frequency_band.band_3k_4k_mean.mean_value': [-3.0],
'fft_real.mean.mean_value': [1.0],
})
pd.testing.assert_frame_equal(expected_readings, outputs['readings'])
def _clone_pipeline(pipeline):
return MLPipeline.from_dict(pipeline.to_dict())
