def __init__(self, base_estimators=None, bagging_base=None, stacking='xgb',
features_stack=None, bagging_stack=None, hunting=False,
transform=True, transform_pred=True):
"""blablabla
Parameters
----------
base_estimators : dict('clf': classifier OR keyword-parameters)
Contains all the level-0 classifiers. The key is the name of the
classifier and the value is either a **prefitted** classifier or
a dictionary containing the keyword arguments to instantiate
such a classifier.
If no pre-trained classifier is provided, the key-value has to
be one of the following:
- **'xgb'** creates an XGBoost classifier
- **'rdf'** creates a Random Forest classifier
- **'erf'** creates a Forest consisting of Extra Randomized Trees
- **'nn'** creates an artificial Neural Network from TheaNets
- **'ada'** creates an AdaBoost instance with Decision Trees as
basis
- **'gb'** creates a Gradient Boosted classifier with Decision
Trees as basis
"""
if base_estimators is None:
OrderedDict(self.__DEFAULT_CLF_CFG)
else:
self._base_estimators = base_estimators
if isinstance(stacking, str):
self._clf_1 = {stacking: None}
elif isinstance(stacking, dict):
self._clf_1 = stacking
elif stacking in (False, None):
stacking = False
else:
self._clf_1 = {'clf_stacking': stacking} # stacking is a classifier
self._transform_data = transform
self._bagging = bagging_base
self._hunting = hunting
self._clf_1_bagging = bagging_stack
self._features_stack = features_stack
self._clf_0 = {}
self._factory = ClassifiersFactory()
self._base_scaler = None
self._pred_scaler = None
def _check_classification_report(n_classes=2):
classifiers = ClassifiersFactory()
classifiers.add_classifier('gb', GradientBoostingClassifier(n_estimators=10))
classifiers.add_classifier('rf', RandomForestClassifier())
classifiers.add_classifier('ada', AdaBoostClassifier(n_estimators=10))
X, y = generate_classification_sample(1000, 5, n_classes=n_classes)
classifiers.fit(X, y)
X, y = generate_classification_sample(1000, 5, n_classes=n_classes)
test_lds = LabeledDataStorage(X, y, sample_weight=None)
report = classifiers.test_on_lds(test_lds)
val = numpy.mean(X['column0'])
labels_dict = None
if n_classes > 2:
labels_dict = {i: str(i) for i in range(n_classes)}
_classification_mask_report(report, "column0 > %f" % val, X, labels_dict)
_classification_mask_report(report, lambda x: numpy.array(x['column0']) < val, X, labels_dict)
_classification_mask_report(report, None, X, labels_dict)
check_classification_learning_curve_masks(report, n_classes=n_classes)
def test_factory():
factory = ClassifiersFactory()
try:
from rep.estimators.tmva import TMVAClassifier
factory.add_classifier('tmva', TMVAClassifier())
except ImportError:
pass
factory.add_classifier('rf', RandomForestClassifier(n_estimators=10))
factory.add_classifier('ada', AdaBoostClassifier(n_estimators=20))
X, y, sample_weight = generate_classification_data()
assert factory == factory.fit(X,
y,
sample_weight=sample_weight,
features=list(X.columns),
parallel_profile='threads-4')
for cl in factory.values():
assert list(cl.features) == list(X.columns)
proba = factory.predict_proba(X, parallel_profile='threads-4')
labels = factory.predict(X, parallel_profile='threads-4')
for key, val in labels.items():
score = accuracy_score(y, val)
print(key, score)
assert score > 0.7, key
for key, val in proba.items():
assert numpy.allclose(val.sum(axis=1),
1), 'probabilities do not sum to 1'
assert numpy.all(val >= 0.), 'negative probabilities'
auc_score = roc_auc_score(y, val[:, 1])
print(auc_score)
assert auc_score > 0.8
for key, iterator in factory.staged_predict_proba(X).items():
assert key != 'tmva', 'tmva does not support staged pp'
for p in iterator:
assert p.shape == (len(X), 2)
# checking that last iteration coincides with previous
assert numpy.all(p == proba[key])
# testing picklability
dump_string = cPickle.dumps(factory)
clf_loaded = cPickle.loads(dump_string)
assert type(factory) == type(clf_loaded)
probs1 = factory.predict_proba(X)
probs2 = clf_loaded.predict_proba(X)
for key, val in probs1.items():
assert numpy.all(val == probs2[key]), 'something strange was loaded'
report = ClassificationReport({'rf': factory['rf']},
LabeledDataStorage(X, y, sample_weight))
report.feature_importance_shuffling(roc_auc_score_mod).plot(new_plot=True,
figsize=(18,
3))
report = factory.test_on_lds(LabeledDataStorage(X, y, sample_weight))
report = factory.test_on(X, y, sample_weight=sample_weight)
val = numpy.mean(X['column0'])
check_report_with_mask(report, "column0 > %f" % (val / 2.), X)
check_report_with_mask(report,
lambda x: numpy.array(x['column0']) < val * 2., X)
check_report_with_mask(report, None, X)
def test_factory():
factory = ClassifiersFactory()
try:
from rep.estimators.tmva import TMVAClassifier
factory.add_classifier('tmva', TMVAClassifier())
except ImportError:
pass
factory.add_classifier('rf', RandomForestClassifier(n_estimators=10))
factory.add_classifier('ada', AdaBoostClassifier(n_estimators=20))
X, y, sample_weight = generate_classification_data()
assert factory == factory.fit(X, y, sample_weight=sample_weight, features=list(X.columns), parallel_profile='threads-4')
for cl in factory.values():
assert list(cl.features) == list(X.columns)
proba = factory.predict_proba(X, parallel_profile='threads-4')
labels = factory.predict(X, parallel_profile='threads-4')
for key, val in labels.items():
score = accuracy_score(y, val)
print(key, score)
assert score > 0.7, key
for key, val in proba.items():
assert numpy.allclose(val.sum(axis=1), 1), 'probabilities do not sum to 1'
assert numpy.all(val >= 0.), 'negative probabilities'
auc_score = roc_auc_score(y, val[:, 1])
print(auc_score)
assert auc_score > 0.8
for key, iterator in factory.staged_predict_proba(X).items():
assert key != 'tmva', 'tmva does not support staged pp'
for p in iterator:
assert p.shape == (len(X), 2)
# checking that last iteration coincides with previous
assert numpy.all(p == proba[key])
# testing picklability
dump_string = cPickle.dumps(factory)
clf_loaded = cPickle.loads(dump_string)
assert type(factory) == type(clf_loaded)
probs1 = factory.predict_proba(X)
probs2 = clf_loaded.predict_proba(X)
for key, val in probs1.items():
assert numpy.all(val == probs2[key]), 'something strange was loaded'
report = ClassificationReport({'rf': factory['rf']}, LabeledDataStorage(X, y, sample_weight))
report.feature_importance_shuffling(roc_auc_score_mod).plot(new_plot=True, figsize=(18, 3))
report = factory.test_on_lds(LabeledDataStorage(X, y, sample_weight))
report = factory.test_on(X, y, sample_weight=sample_weight)
val = numpy.mean(X['column0'])
check_report_with_mask(report, "column0 > %f" % (val / 2.), X)
check_report_with_mask(report, lambda x: numpy.array(x['column0']) < val * 2., X)
check_report_with_mask(report, None, X)
def _test_classification_report(n_classes=2):
classifiers = ClassifiersFactory()
classifiers.add_classifier('gb',
GradientBoostingClassifier(n_estimators=10))
classifiers.add_classifier('rf', RandomForestClassifier())
classifiers.add_classifier('ada', AdaBoostClassifier(n_estimators=10))
X, y = generate_classification_sample(1000, 5, n_classes=n_classes)
classifiers.fit(X, y)
X, y = generate_classification_sample(1000, 5, n_classes=n_classes)
test_lds = LabeledDataStorage(X, y, sample_weight=None)
report = classifiers.test_on_lds(test_lds)
val = numpy.mean(X['column0'])
labels_dict = None
if n_classes > 2:
labels_dict = {}
for i in range(n_classes):
labels_dict[i] = str(i)
_classification_mask_report(report, "column0 > %f" % val, X, labels_dict)
_classification_mask_report(report,
lambda x: numpy.array(x['column0']) < val, X,
labels_dict)
_classification_mask_report(report, None, X, labels_dict)
class Mayou(Classifier):
"""Classifier for raredecay analysis"""
__DEFAULT_CLF_CFG = dict(
xgb=dict(
n_estimators=450,
eta=0.1,
subsample=0.9,
bagging=None
),
rdf=dict(
n_estimators=1600, # 1600
max_features='auto', # only 1 feature seems to be pretty good...
max_depth=200,
min_samples_split=250,
min_samples_leaf=150,
min_weight_fraction_leaf=0.,
max_leaf_nodes=None,
bootstrap=False,
oob_score=False,
n_jobs=7,
class_weight=None,
bagging=None
),
nn=dict(
layers=[100, 100],
hidden_activation='logistic',
output_activation='linear',
input_noise=0, # [0,1,2,3,4,5,10,20],
hidden_noise=0,
input_dropout=0,
hidden_dropout=0.05,
decode_from=1,
weight_l1=0.01,
weight_l2=0.03,
scaler='standard',
trainers=[{'optimize': 'adagrad', 'patience': 2, 'momentum': 0.5, 'nesterov': True,
'learning_rate': 0.2, 'min_improvement': 0.01}],
bagging=None
),
gb=dict(
learning_rate=0.05,
n_estimators=500,
max_depth=4,
min_samples_split=600,
min_samples_leaf=1,
min_weight_fraction_leaf=0.,
subsample=0.8,
max_features=None,
max_leaf_nodes=None,
bagging=None
),
)
__DEFAULT_BAG_CFG = dict(
n_estimators=20,
max_samples=0.9,
max_features=1.0,
)
def __init__(self, base_estimators=None, bagging_base=None, stacking='xgb',
features_stack=None, bagging_stack=None, hunting=False,
transform=True, transform_pred=True):
"""blablabla
Parameters
----------
base_estimators : dict('clf': classifier OR keyword-parameters)
Contains all the level-0 classifiers. The key is the name of the
classifier and the value is either a **prefitted** classifier or
a dictionary containing the keyword arguments to instantiate
such a classifier.
If no pre-trained classifier is provided, the key-value has to
be one of the following:
- **'xgb'** creates an XGBoost classifier
- **'rdf'** creates a Random Forest classifier
- **'erf'** creates a Forest consisting of Extra Randomized Trees
- **'nn'** creates an artificial Neural Network from TheaNets
- **'ada'** creates an AdaBoost instance with Decision Trees as
basis
- **'gb'** creates a Gradient Boosted classifier with Decision
Trees as basis
"""
if base_estimators is None:
OrderedDict(self.__DEFAULT_CLF_CFG)
else:
self._base_estimators = base_estimators
if isinstance(stacking, str):
self._clf_1 = {stacking: None}
elif isinstance(stacking, dict):
self._clf_1 = stacking
elif stacking in (False, None):
stacking = False
else:
self._clf_1 = {'clf_stacking': stacking} # stacking is a classifier
self._transform_data = transform
self._bagging = bagging_base
self._hunting = hunting
self._clf_1_bagging = bagging_stack
self._features_stack = features_stack
self._clf_0 = {}
self._factory = ClassifiersFactory()
self._base_scaler = None
self._pred_scaler = None
def get_params(self, deep=True):
out = dict(
base_estimators=None, bagging_base=None, stacking='xgb',
features_stack=None, bagging_stack=None, hunting=False
)
return out
def _transform(self, X, fit=False):
if self._transform_data:
columns = copy.deepcopy(X.keys())
index = copy.deepcopy(X.index)
if fit:
self._base_scaler = StandardScaler(copy=True)
self._base_scaler.fit_transform(X)
else:
X = self._base_scaler.transform(X)
X = pd.DataFrame(X, index=index, columns=columns)
return X
def _transform_pred(self, X, fit=False):
if self._transform_pred:
columns = copy.deepcopy(X.keys())
index = copy.deepcopy(X.index)
if fit:
self._pred_scaler = StandardScaler(copy=True) # don't change data!
self._pred_scaler.fit_transform(X)
else:
X = self._pred_scaler.transform(X)
X = pd.DataFrame(X, index=index, columns=columns)
return X
def _make_clf(self, clf, bagging=None):
"""Creates a classifier from a dict or returns the clf"""
if isinstance(clf, dict):
key, val = clf.popitem()
try:
val = self.__DEFAULT_CLF_CFG.get(key) if val is None else val
except KeyError:
logger.error(str(val) + " not an implemented classifier.")
raise
temp_bagging = val.pop('bagging', bagging)
bagging = temp_bagging if bagging is None else bagging
if key == 'rdf':
config_clf = dict(val) # possible multi-threading arguments
clf = SklearnClassifier(RandomForestClassifier(**config_clf))
elif key == 'erf':
config_clf = dict(val) # possible multi-threading arguments
clf = SklearnClassifier(ExtraTreesClassifier(**config_clf))
elif key == 'nn':
config_clf = dict(val) # possible multi-threading arguments
clf = TheanetsClassifier(**config_clf)
elif key == 'ada':
config_clf = dict(val) # possible multi-threading arguments
clf = SklearnClassifier(AdaBoostClassifier(**config_clf))
elif key == 'gb':
config_clf = dict(val) # possible multi-threading arguments
clf = SklearnClassifier(GradientBoostingClassifier(**config_clf))
elif key == 'xgb':
config_clf = dict(val) # possible multi-threading arguments
clf = XGBoostClassifier(**config_clf)
elif hasattr(clf, 'fit'):
bagging = False # return the classifier
# bagging over the instantiated estimators
if isinstance(bagging, int) and bagging >= 1:
bagging = dict(self.__DEFAULT_BAG_CFG, n_estimators=bagging)
if isinstance(bagging, dict):
# TODO: implement multi-thread:
bagging.update({'base_estimator': clf})
clf = SklearnClassifier(BaggingClassifier(**bagging))
else:
raise ValueError(str(clf) + " not valid as a classifier.")
clf = {key: clf}
return clf
def _factory_fit(self, X, y, sample_weight):
# create classifiers from initial dictionary
if self._base_estimators != {}:
for key, val in self._base_estimators.items():
clf = self._make_clf({key: val}, bagging=self._bagging)
self._clf_0.update(clf)
self._base_estimators = {}
# add base estimators to factory
for key, val in self._clf_0.iteritems():
self._factory.add_classifier(key, val)
# parallel on factory level -> good mixture of clfs (one uses lot of RAM, one cpu...)
parallel_profile = 'threads-' + str(min([len(self._factory.items()),
globals_.free_cpus()]))
# fit all classifiers
print "start fitting factory"
self._factory.fit(X, y, sample_weight, parallel_profile=parallel_profile)
return self
def _factory_predict(self, X):
columns = copy.deepcopy(X.keys())
index = copy.deepcopy(X.index)
# parallel on factory level -> good mixture of clfs (one uses lot of RAM, one cpu...)
parallel_profile = 'threads-' + str(min([len(self._factory.items()),
globals_.free_cpus()]))
# predict, return a dictionary
predictions = self._factory.predict(X, parallel_profile=parallel_profile)
# slice the arrays of predictions in the dict right
for key, val in predictions.items():
predictions[key] = val[:, 1]
return pd.DataFrame(predictions, index=index, columns=columns)
# @profile
def _factory_predict_proba(self, X):
index = X.index
# parallel on factory level -> good mixture of clfs (one uses lot of RAM, one cpu...)
parallel_profile = 'threads-' + str(min([len(self._factory.items()),
globals_.free_cpus()]))
print parallel_profile
parallel_profile = None
# predict, return a dictionary
predictions = self._factory.predict_proba(X, parallel_profile=parallel_profile)
# slice the arrays of predictions in the dict right
for key, val in predictions.items():
predictions[key] = val[:, 1]
return pd.DataFrame(predictions, index=index)
def _get_X_stack(self, X, fit_scaler=False):
# get the predictions of the base estimators
lvl_0_proba = pd.DataFrame(self._factory_predict_proba(X), index=X.index,
columns=self._factory.keys())
lvl_0_proba = self._transform_pred(lvl_0_proba, fit=fit_scaler)
# add data features to stacking data
if self._features_stack is not None:
if self._features_stack == 'all':
self._features_stack = self.features
elif not set(self._features_stack).issubset(self.features):
raise RuntimeError("Stacked features not in features of the data fitted to")
X_data = pd.DataFrame(X, columns=self._features_stack)
lvl_0_proba = pd.concat([lvl_0_proba, X_data], axis=1, copy=False)
return lvl_0_proba
def _clf_1_fit(self, X, y, sample_weight):
X_stack = self._get_X_stack(X, fit_scaler=True)
if self._clf_1 not in (False, None):
if self._clf_1.values()[0] is None or isinstance(self._clf_1.values()[0], dict):
self._clf_1 = self._make_clf(self._clf_1, bagging=self._clf_1_bagging)
self._clf = copy.deepcopy(self._clf_1.values()[0])
self._clf.fit(X_stack, y, sample_weight)
def _set_features(self, X):
"""Set the 'features' attribute for the classifier"""
if isinstance(X, pd.DataFrame):
self.features = X.columns.values
else:
self.features = ["Feature_" + str(i) for i in range(X.shape[1])]
def fit(self, X, y, sample_weight=None):
# initiate properties
self._set_features(X)
self.classes_ = range(len(set(y)))
X = self._transform(X, fit=True)
# fit the base estimators
self._factory_fit(X, y, sample_weight)
# fit the stacking classifier
self._clf_1_fit(X, y, sample_weight)
return self
def predict(self, X):
# TODO: inside get_X_stack: lvl_0_proba = self._factory_predict_proba(X)
X = self._transform(X)
X_stack = self._get_X_stack(X)
return self._clf.predict(X_stack)
def predict_proba(self, X):
X = self._transform(X)
X_stack = self._get_X_stack(X)
return self._clf.predict_proba(X_stack)
def test_on(self, X, y, sample_weight=None):
lds = LabeledDataStorage(X, y, sample_weight)
return self.test_on_lds(lds)
def test_on_lds(self, lds):
lds.data = self._transform(lds.data)
return ClassificationReport({'Mayou clf': self}, lds)
def staged_predict_proba(self, X):
X = self._transform(X)
X_stack = self._get_X_stack(X)
try:
temp_proba = self._clf.staged_predict_proba(X_stack)
# TODO: change error catching
except:
print "error occured in mayou, staged predict proba not supported"
return temp_proba
def stacker_test_on_lds(self, lds):
"""Return report for the stacker only"""
lds.data = self._get_X_stack(self._transform(lds.data))
return ClassificationReport({'Mayou stacker': self._clf}, lds)
def stacker_test_on(self, X, y, sample_weight=None):
"""Return report for the stacker only"""
lds = LabeledDataStorage(X, y, sample_weight)
return self.stacker_test_on_lds(lds)
if args.verbose: fprint("Start loading input")
with open(args.dir + "/train_uniform_reports.pkl", 'rb') as infile:
reports = pickle.load(infile)
if args.verbose: fprint("Finish loading input")
from mods import config_path
config_path()
args.config = args.config.replace(".py", "")
uconfig = getattr(__import__(args.config, fromlist="uconfig"), "uconfig")
# check for subset of classifiers
if len(args.classifiers) > 0:
for rname, report in reports.iteritems():
est_old = report.estimators
pred_old = report.prediction
est_new = ClassifiersFactory()
pred_new = OrderedDict()
for classifier in args.classifiers:
if classifier in est_old:
est_new[classifier] = est_old[classifier]
pred_new[classifier] = pred_old[classifier]
else:
raise ValueError("Requested classifier " + classifier +
" not found in report " + rname)
report.estimators = est_new
report.prediction = pred_new
# to serialize plots
repplots = OrderedDict()
matplots = OrderedDict()
trainX = train_data[0]
testX = test_data[0]
trainY = train_data[1]
testY = test_data[1]
trainW = {}
testW = {}
for iw, weight in enumerate(sorted(W)):
trainW[weight] = train_data[iw + 2]
testW[weight] = test_data[iw + 2]
if args.verbose:
fprint("Split data into train_size=" + str(uconfig.training.size) +
", test_size=" + str(uconfig.training.size))
# create classifiers
classifiers = ClassifiersFactory()
weights = OrderedDict()
# standard bdt
if "bdt" in uconfig.training.algorithms:
base_grad = GradientBoostingClassifier(
max_depth=uconfig.hyper.max_depth,
n_estimators=uconfig.hyper.n_estimators,
subsample=uconfig.hyper.subsample,
learning_rate=uconfig.hyper.learning_rate,
min_samples_leaf=uconfig.hyper.min_samples_leaf,
)
classifiers["bdt"] = SklearnClassifier(base_grad,
features=uconfig.features.train)
weights["bdt"] = trainW[uconfig.training.algorithms["bdt"]]