def fit(self,
X,
y=None,
n_estimators=9,
max_features="auto",
sample_weight=False,
random_state=None):
# get features and labels
features, labels = self.feature_extractor.transform(
X, relation_labels=self.relation_labels)
if is_empty(features):
log.error("No examples to train, quiting...")
return self
log.info("Checking parameters...")
if type(max_features) == str:
if max_features == "auto":
max_features = self._get_max_features(features)
elif max_features == "log":
max_features = self._get_max_features(features, method=log2)
elif max_features == "sqrt":
max_features = self._get_max_features(features, method=sqrt)
else:
raise ValueError(
"Unknown method '{}' for feature selection in Random Forest"
.format(max_features))
if type(max_features) != int:
raise TypeError(
"The parameter 'max_features' must be either a string or integer."
)
self.config.set_parameters({
"n_estimators": n_estimators,
"max_features": max_features,
"random_state": random_state
})
# create a model
self.rf = RandomForestClassifier(
n_estimators=self.config.get_parameter("n_estimators"),
max_features=self.config.get_parameter("max_features"),
random_state=self.config.get_parameter("random_state"))
log.info("Training Random Forest...")
weights = to_weights(labels) if sample_weight else None
self.rf.fit(features, labels, sample_weight=weights)
return self
def fit(self, X, y=None, sample_weight=False):
log.info("Checking parameters...")
self.config.validate()
# create a model
self.nb = MultinomialNB()
# get features and labels
features, labels = self.feature_extractor.transform(
X, relation_labels=self.relation_labels)
if is_empty(features):
log.error("No examples to train, quiting...")
return self
log.info("Training Naive Bayes...")
weights = to_weights(labels) if sample_weight else None
self.nb.fit(features, labels, sample_weight=weights)
return self
def fit(self, X, y=None, max_iterations=100, C=1, sample_weight=False):
log.info("Checking parameters...")
self.config.set_parameters({"max_iterations": max_iterations, "C": C})
self.config.validate()
# create a model
self.svm = CalibratedClassifierCV(
LinearSVC(max_iter=self.config.get_parameter("max_iterations"),
C=self.config.get_parameter("C")))
# get features and labels
features, labels = self.feature_extractor.transform(
X, relation_labels=self.relation_labels)
if is_empty(features):
log.error("No examples to train, quiting...")
return self
log.info("Training SVM...")
weights = to_weights(labels) if sample_weight else None
self.svm.fit(features, labels, sample_weight=weights)
return self
def transform(self, X, y=None):
# get features (labels are ignored)
features, _ = self.feature_extractor.transform(
X, relation_labels=self.relation_labels)
if is_empty(features):
return X
# make predictions
log.info("Predicting relations in {} documents with SVM...".format(
len(X)))
probs = self.svm.predict_proba(features)
# the order of labels corresponds to the order of probabilities
labels = self.svm.classes_
predicted_labels = []
for prob in probs:
score = numpy.amax(prob)
label = labels[numpy.where(prob == score)[0][0]]
predicted_labels += [{label: score}]
# make annotated documents
return self.feature_extractor.predictions_to_annotated_documents(
predicted_labels)