def test_multilabelclassifier():
np.random.seed(1)
N = 1000
split = N // 2
X = np.random.rand(N, 2)
Ya = [1 if X[i, 0] < 0.5 else 0 for i in range(X.shape[0])]
Yb = [1 if X[i, 1] < 0.5 else 0 for i in range(X.shape[0])]
Y = np.column_stack((Ya, Yb))
Xtrain = X[:split, :]
Ytrain = Y[:split, :]
Xtest = X[split:, :]
Ytest = Y[split:, :]
classifier = kindred.MultiLabelClassifier(
sklearn.linear_model.LogisticRegression,
random_state=1,
solver='lbfgs')
classifier.fit(Xtrain, Ytrain)
predicted = classifier.predict(Xtest)
assert predicted.shape == Ytest.shape
rmse = np.sqrt(np.mean((predicted - Ytest)**2))
assert round(rmse, 3) == 0.071
probs = classifier.predict_proba(Xtest)
assert probs.shape == Ytest.shape
rmse_probs = np.sqrt(np.mean((probs - Ytest)**2))
assert round(rmse_probs, 3) == 0.202
def train(self,corpus):
"""
Trains the classifier using this corpus. All relations in the corpus will be used for training.
:param corpus: Corpus to use for training
:type corpus: kindred.Corpus
"""
assert isinstance(corpus,kindred.Corpus)
if not corpus.parsed:
parser = kindred.Parser(model=self.model)
parser.parse(corpus)
self.candidateBuilder = CandidateBuilder(entityCount=self.entityCount,acceptedEntityTypes=self.acceptedEntityTypes)
candidateRelations = self.candidateBuilder.build(corpus)
if len(candidateRelations) == 0:
raise RuntimeError("No candidate relations found in corpus for training. Does the corpus contain text and entity annotations with at least one sentence containing %d entities." % (self.entityCount))
candidateRelationKeys = set()
for cr in candidateRelations:
assert isinstance(cr,kindred.CandidateRelation)
for knownType,knownArgNames in cr.knownTypesAndArgNames:
relKey = tuple([knownType] + knownArgNames)
candidateRelationKeys.add(relKey)
# Create mappings from the class index to a relation type and back again
self.colToRelType = sorted(list(candidateRelationKeys))
self.relTypeToCol = { relationType:i for i,relationType in enumerate(self.colToRelType) }
Y = np.zeros((len(candidateRelations),len(self.colToRelType)),np.int32)
candidateClasses = []
for i,cr in enumerate(candidateRelations):
for knownType,knownArgNames in cr.knownTypesAndArgNames:
relKey = tuple([knownType] + knownArgNames)
col = self.relTypeToCol[relKey]
Y[i,col] = 1
entityCountsInRelations = set([ len(r.entities) for r in corpus.getRelations() ])
entityCountsInRelations = sorted(list(set(entityCountsInRelations)))
assert self.entityCount in entityCountsInRelations, "Relation classifier is expecting to train on relations with %d entities (entityCount=%d). But the known relations in the corpus contain relations with the following entity counts: %s. Perhaps the entityCount parameter should be changed or there is a problem with the training corpus." % (self.entityCount,self.entityCount,str(entityCountsInRelations))
self.relTypeToValidEntityTypes = defaultdict(set)
for d in corpus.documents:
for r in d.relations:
validEntityTypes = tuple([ e.entityType for e in r.entities ])
relKey = tuple([r.relationType] + r.argNames)
self.relTypeToValidEntityTypes[relKey].add(validEntityTypes)
self.vectorizer = Vectorizer(entityCount=self.entityCount,featureChoice=self.chosenFeatures,tfidf=self.tfidf)
trainVectors = self.vectorizer.fit_transform(candidateRelations)
assert trainVectors.shape[0] == Y.shape[0]
posCount = Y.sum()
negCount = Y.shape[0]*Y.shape[1] - posCount
assert negCount > 0, "Must have at least one negative candidate relation in set for training"
assert posCount > 0, "Must have at least one positive candidate relation in set for training"
self.clf = None
if self.classifierType == 'SVM':
self.clf = kindred.MultiLabelClassifier(svm.LinearSVC,class_weight='balanced',random_state=1,max_iter=10000)
elif self.classifierType == 'LogisticRegression' and self.threshold is None:
self.clf = kindred.MultiLabelClassifier(LogisticRegression,class_weight='balanced',random_state=1,solver='liblinear',multi_class='ovr')
elif self.classifierType == 'LogisticRegression' and not self.threshold is None:
self.clf = kindred.MultiLabelClassifier(kindred.LogisticRegressionWithThreshold,threshold=self.threshold)
self.clf.fit(trainVectors,Y)
self.isTrained = True