def comparison_test(text):
import sklearn.feature_extraction.text as txt
h_trick = txt.HashingVectorizer(n_features=20, binary=True, norm=None)
oh_encoder = txt.CountVectorizer()
oh_encoded = oh_encoder.fit_transform(text)
hashing = h_trick.transform(text)
return oh_encoded, hashing
def QuickClusterParamaterFinder(data):
Cost = list()
vectorizer = txtvectorizer.HashingVectorizer(analyzer='word',
ngram_range=(1, 10))
vectors = vectorizer.transform(data['content'].dropna())
for c in range(1, 100, 5):
kmeans = sklearn.cluster.MiniBatchKMeans(n_clusters=c)
kmeans.fit_transform(vectors)
Cost.append(kmeans.inertia_)
print(str(c) + " / 100")
plt.plot(range(1, 100, 5), Cost)
return Cost
def build_vectors(sentences, vacabulary_size):
vectorizer = skyfe.CountVectorizer()
trans = vectorizer.fit_transform(sentences)
fname = vectorizer.get_feature_names()
print(trans)
print(trans.toarray())
print(fname)
#enable tf-idf
transformer = skyfe.TfidfTransformer()
tfidf = transformer.fit_transform(trans)
print(tfidf.toarray())
print(tfidf.get_feature_names())
#hashed
vectorizer2 = skyfe.HashingVectorizer(n_features=6, norm=None)
trans = vectorizer2.fit_transform(sentences)
#fname = vectorizer2.get_feature_names()
print(trans.toarray())
def cluster_topics():
#model = cluster.Birch(
#branching_factor=2,
#threshold=0.002 # Lower = more clusters, higher = fewer clusters
#)
#model = cluster.KMeans(
#branching_factor=10,
#threshold=0.1 # Lower = more clusters, higher = fewer clusters
#)
model = cluster.DBSCAN(min_samples=2, eps=0.2)
#model = cluster.AffinityPropagation(
#)
vectorizer = text.HashingVectorizer(
analyzer='char_wb', # The feature is made of words not characters
norm='l2', # Normalize the words
lowercase=True, # Converts everything to lowercase
stop_words=stopwords)
num_samples = 10000
offset = 0
while True:
log.debug(u"Loading topics...")
topic_rows = db.session.query(
models.TopicModel.id,
models.TopicModel.topic).filter_by(clustered=False).order_by(
models.TopicModel.id.asc()).limit(num_samples).offset(
offset).all()
if not topic_rows:
break
log.debug(u"Loaded {} topics".format(len(topic_rows)))
offset += len(topic_rows)
go_cluster(vectorizer, model, topic_rows)
#!/usr/bin/python # -*- coding: utf-8 -*- #[email protected] """ ============================== 文本向量化方法3 ============================== Tfidf方法进行文本向量化过程中如果单词量很大会遇到内存问题 因此利用hash编码可以将特征数减少 这个方法本质就是将单词进行hash编码，因此一个编码可以对应多个词语 从而压缩内存 """ from sklearn.datasets import fetch_20newsgroups import sklearn.feature_extraction.text as t2v import numpy as np #获取数据 newsgroups_train = fetch_20newsgroups(data_home="data", subset='train') newsgroups_test = fetch_20newsgroups(data_home="data", subset='test') #单词向量化 vectorizer = t2v.HashingVectorizer(n_features=6) vectors = vectorizer.fit_transform(newsgroups_train.data) vectors_test = vectorizer.transform(newsgroups_test.data) print("class:", newsgroups_train.target_names[newsgroups_train.target[1]]) print("data:", newsgroups_train.data[1]) print(vectors[1])
# take document stream from stream_docs and return specific number of documents
def get_minibatch(doc_stream, size):
docs, labels = [], []
try:
for _ in range(size):
doc, label = next(doc_stream)
docs.append(doc)
labels.append(label)
except StopIteration:
return None, None
return docs, labels
# use data independent hasher
vect = text.HashingVectorizer(decode_error='ignore',
n_features=2**21,
preprocessor=None,
tokenizer=preprocessor_two)
clf = lm.SGDClassifier(loss='log', random_state=1, max_iter=1)
doc_stream = stream_docs(path='./movie_data.csv')
# perform out of core learning
classes = np.array([0, 1])
for _ in range(45):
data_train, target_train = get_minibatch(doc_stream, size=1000)
if not data_train:
break
data_train = vect.transform(data_train)
clf.partial_fit(data_train, target_train, classes=classes)
# use last 5000 documents for evaluation
data_test, target_test = get_minibatch(doc_stream, size=5000)
data_test = vect.transform(data_test)
print('Accuracy: %.3f' % clf.score(data_test, target_test))
@author: Administrator
"""
from sklearn.datasets import fetch_20newsgroups
newsgroups_train = fetch_20newsgroups(subset='train',
remove=('headers', 'footers', 'quotes'))
newsgroups_test = fetch_20newsgroups(subset='test',
remove=('headers', 'footers', 'quotes'))
#from sklearn.naive_bayes import BernoulliNB, MultinomialNB
from sklearn.naive_bayes import MultinomialNB
#Bernoulli = BernoulliNB(alpha=0.01)
Multinomial = MultinomialNB(alpha=0.01)
import sklearn.feature_extraction.text as txt
multinomial_hashing_trick = txt.HashingVectorizer(stop_words='english',
binary=False, norm=None, non_negative=True)
#binary_hashing_trick = txt.HashingVectorizer(stop_words='english',
# binary=True, norm=None, non_negative=True)
Multinomial.fit(multinomial_hashing_trick.transform(newsgroups_train.data),
newsgroups_train.target)
#Bernoulli.fit(binary_hashing_trick.transform(newsgroups_train.data),
# newsgroups_train.target)
from sklearn.metrics import accuracy_score
for m, h in [(Multinomial, multinomial_hashing_trick)]:
print 'Accuracy for %s: %.3f' % (m, accuracy_score(y_true=newsgroups_test.target,
y_pred=m.predict(h.transform(
newsgroups_test.data))))
def vectorizeStrings(documents, ngramRange):
vectorizer = txtvectorizer.HashingVectorizer(analyzer='char',
ngram_range=ngramRange)
vectors = vectorizer.transform(documents.fillna(""))
return vectors
test_data['Position_Extra'] = test_data['Position_Extra'].apply(clean) test_data['Program_Description'] = test_data['Program_Description'].apply(clean) test_data['SubFund_Description'] = test_data['SubFund_Description'].apply(clean) test_data['Sub_Object_Description'] = test_data['Sub_Object_Description'].apply(clean) test_data['Text_1'] = test_data['Text_1'].apply(clean) test_data['Text_2'] = test_data['Text_2'].apply(clean) test_data['Text_3'] = test_data['Text_3'].apply(clean) test_data['Text_4'] = test_data['Text_4'].apply(clean) # create a single new column for cleaned text data training_data["combined"] = [' '.join(row) for row in training_data[training_data.columns].values] test_data["combined"] = [' '.join(row) for row in test_data[test_data.columns].values] # initialize TFIDF vectorizer and Hashing Vectorizer tfidf = txt.TfidfVectorizer(ngram_range=(2, 1), max_df=1.0, min_df=10) hsv = txt.HashingVectorizer() # fit tfidf and hashing vectorizer to train data tfidf.fit(training_data['combined']) hsv.fit(test_data['combined']) # transform the training and test datasets to obtain a sparse matrix X_tfidf = tfidf.transform(training_data['combined']) X_test_tfidf = tfidf.transform(test_data['combined']) X_hsv = hsv.transform(training_data['combined']) X_test_hsv = hsv.transform(test_data['combined']) X = sparse.hstack((X_hsv, X_tfidf)) X_test = sparse.hstack((X_test_hsv, X_test_tfidf))
def hash_vector(text,features):
hash_vectorizer = txt.HashingVectorizer(n_features=features,binary=True,norm=None)
text_vector = hash_vectorizer.transform(text)
return text_vector
"""
__author__ = "Adrian Langseth"
import pickle
import sklearn.feature_extraction.text as skt
import time
from sklearn.naive_bayes import BernoulliNB
from sklearn.metrics import accuracy_score
from sklearn.tree import DecisionTreeClassifier
t0 = time.time()
k = pickle.load(open("sklearn-data.pickle", "rb"))
vectorizer = skt.HashingVectorizer(stop_words='english', analyzer='word')
xtrain_transformed = vectorizer.fit_transform(k['x_train'])
xtest_transformed = vectorizer.fit_transform(k['x_test'])
# NB classifier part
NBclassifier = BernoulliNB() # Build Model
NBclassifier.fit(xtrain_transformed, k['y_train']) # Fitting model
NBpredicted_y = NBclassifier.predict(xtest_transformed) # Make prediction
NBaccuracy_score = accuracy_score(k['y_test'],
NBpredicted_y) # Evaluate prediction
# Decision Tree part
DTclassifier = DecisionTreeClassifier(max_depth=64,
criterion="entropy") # Create Model
DTclassifier.fit(xtrain_transformed, k['y_train']) # Fit model
if __name__ == "__main__":
input_files = list(glob(str(base_dir / "data" / "*" / "*")))
data = []
for file_path in input_files:
with open(file_path, "rt") as fh:
data.append(fh.read())
dataset_size = 91 # MB for 20 newsgroup dataset
print("# vectorizing {} documents:".format(len(data)))
for label, vect in [
("HashingVectorizer (vtext)", vtext.HashingVectorizer(norm=None)),
(
"HashingVectorizer (scikit-learn)",
skt.HashingVectorizer(lowercase=False, norm=None),
),
("CountVectorizer (vtext)", vtext.CountVectorizer(lowercase=False)),
("CountVectorizer (scikit-learn)",
skt.CountVectorizer(lowercase=False)),
]:
t0 = time()
X = vect.fit_transform(data)
dt = time() - t0
print("{:>40}: {:.2f}s [{:.1f} MB/s], shape={}, nnz={}".format(
label, dt, dataset_size / dt, X.shape, X.nnz))
for word in input_string.split(' '):
index = abs(
hash(word)
) % vector_size #turn the position in vector where hash(word)%vector_size into 1
feature_vector[index] = 1
return feature_vector
# if vector_size too small, there will be easily overlapping
from scipy.sparse import csc_matrix
print csc_matrix([1, 0, 0, 0, 0, 1, 1, 0, 1, 0])
# Sparse CSC_matrix only record the position of 1, by eliminating 0, saving a lot of memory
#------------Using BUiltin HashingVectorizer -------------------
import sklearn.feature_extraction.text as txt
sklearn_hashing_trick = txt.HashingVectorizer(n_features=20,
binary=True,
norm=None)
text_vector = sklearn_hashing_trick.transform(
[' Python for data science', 'Python for machine learning'])
text_vector
#transform text into vector, and surpress using sparse-like funciton
# CountVectorizer: Optimally encodes text into a data matrix but cannot address subsequent novelties in text.
# HashingVectorizer: Provides flexibility in situations when it is likely that the application will receive new data, but is less optimal than techniques based on hashing functions.
