def do_classify():
corpus = MyCorpus()
# tfidf_model = TfidfModel(corpus)
corpus_idf = tfidf_model[corpus]
# corpus_lsi = lsi_model[corpus_idf]
num_terms = len(corpus.dictionary)
# num_terms = 400
corpus_sparse = matutils.corpus2csc(corpus_idf, num_terms).transpose(copy=False)
# print corpus_sparse.shape
# corpus_dense = matutils.corpus2dense(corpus_idf, len(corpus.dictionary))
# print corpus_dense.shape
penalty = "l2"
clf = SGDClassifier(loss="hinge", penalty=penalty, alpha=0.0001, n_iter=50, fit_intercept=True)
# clf = LinearSVC(loss='l2', penalty=penalty, dual=False, tol=1e-3)
y = np.array(corpus.cls_y)
# print y.shape
clf.fit(corpus_sparse, y)
filename = os.path.join(HERE, "sgdc_clf.pkl")
_ = joblib.dump(clf, filename, compress=9)
print "train completely"
X_test = []
X_label = []
for obj in SogouCorpus.objects.filter(id__in=corpus.test_y):
X_test.append(obj.tokens)
X_label.append(cls_ids[obj.classify])
# result = classifier.predict(obj.tokens)
test_corpus = [dictionary.doc2bow(s.split(",")) for s in X_test]
test_corpus = tfidf_model[test_corpus]
test_corpus = matutils.corpus2csc(test_corpus, num_terms).transpose(copy=False)
pred = clf.predict(test_corpus)
score = metrics.f1_score(X_label, pred)
print ("f1-score: %0.3f" % score)
def evaluate_improved_cllsi(x_train1_in, x_test1_in, x_train2_in, x_test2_in,
dimensions, evaluation_function):
scores = []
for k in dimensions:
x_train1, x_test1 = tfidf(data=(x_train1_in, x_test1_in))
x_train2, x_test2 = tfidf(data=(x_train2_in, x_test2_in))
n_train, n_test = len(x_train1), len(x_test1)
X1 = matutils.corpus2csc(list(x_train1) + list(x_test1))
X2 = matutils.corpus2csc(list(x_train2) + list(x_test2))
x_train1, x_train2 = X1[:, :n_train], X2[:, :n_train]
x_test1, x_test2 = X1[:, n_train:], X2[:, n_train:]
x = sp.sparse.vstack([x_train1, x_train2])
x = matutils.Sparse2Corpus(x)
lsa = models.LsiModel(x, num_topics=k)
n = x_train1.shape[0]
U = lsa.projection.u
U1, U2 = U[:n, :], U[n:, :]
p1, p2 = sp.sparse.csr_matrix(
np.linalg.pinv(U1)), sp.sparse.csr_matrix(np.linalg.pinv(U2))
a1, a2 = np.dot(x_test1.T, p1.T).todense(), np.dot(x_test2.T,
p2.T).todense()
score = evaluation_function(a1, a2)
scores.append(score)
return scores
def get_similarities(self, query):
"""
Return similarity of sparse vector `query` to all documents in the corpus,
as a numpy array.
If `query` is a collection of documents, return a 2D array of similarities
of each document in `query` to all documents in the corpus (=batch query,
faster than processing each document in turn).
**Do not use this function directly; use the self[query] syntax instead.**
"""
is_corpus, query = utils.is_corpus(query)
if is_corpus:
query = matutils.corpus2csc(query, self.index.shape[1], dtype=self.index.dtype)
else:
if scipy.sparse.issparse(query):
query = query.T # convert documents=rows to documents=columns
elif isinstance(query, numpy.ndarray):
if query.ndim == 1:
query.shape = (1, len(query))
query = scipy.sparse.csr_matrix(query, dtype=self.index.dtype).T
else:
# default case: query is a single vector, in sparse gensim format
query = matutils.corpus2csc([query], self.index.shape[1], dtype=self.index.dtype)
# compute cosine similarity against every other document in the collection
result = self.index * query.tocsc() # N x T * T x C = N x C
if result.shape[1] == 1:
# for queries of one document, return a 1d array
result = result.toarray().flatten()
else:
# otherwise, return a 2d matrix (#queries x #index)
result = result.toarray().T
return result
def load_data():
df = pandas.read_csv('../data/train_clean2.csv')
df1 = df[['qid1', 'question1']]
df1.columns = ['qid', 'question']
df2 = df[['qid2', 'question2']]
df2.columns = ['qid', 'question']
df_que = pandas.concat([df1, df2], ignore_index=True)
df_que = df_que.drop_duplicates().fillna('').sort_values('qid')
logger.info('df_que {}'.format(df_que.shape))
train_num = df_que.shape[0]
df = pandas.read_csv('../data/test_clean2.csv')
df1 = df[['question1']]
df1.columns = ['question']
df2 = df[['question2']]
df2.columns = ['question']
df_que2 = pandas.concat([df1, df2], ignore_index=True)
df_que2 = df_que2.drop_duplicates().fillna('')
logger.info('df_que2 {}'.format(df_que2.shape))
df_que2['qid'] = numpy.arange(df_que2.shape[0]) + df_que.shape[0]
df_que = pandas.concat([df_que, df_que2], ignore_index=True)
sentences = corpus_to_sentences(df_que['question'])
logger.info('dict')
dictionary = corpora.Dictionary(sentences)
dictionary.save('./gensim.dict')
dictionary.filter_extremes(no_below=2, no_above=1., keep_n=2000000)
p = Pool()
id_corpus = p.map(dictionary.doc2bow, sentences)
p.close()
p.join()
with open('count_corpus_brown.pkl', 'wb') as f:
pickle.dump(id_corpus, f, -1)
count_mat = corpus2csc(id_corpus).T
logger.info('count_mat {}'.format(count_mat.shape))
with open('count_mat_brown.pkl', 'wb') as f:
pickle.dump(count_mat, f, -1)
tfidf_instance = models.TfidfModel(id_corpus, normalize=False)
tfidf_corpus = tfidf_instance[id_corpus]
tfidf_mat = corpus2csc(tfidf_corpus).T
logger.info('tfidf_mat {}'.format(tfidf_mat.shape))
with open('tfidf_mat_brown.pkl', 'wb') as f:
pickle.dump(tfidf_mat, f, -1)
logger.info('df_que {}'.format(df_que.shape))
logger.info('end load')
return 0
def get_tfidf_scores(kwargs):
tfidf_transformer = TfIdfTransformer(**kwargs).fit(train_corpus)
X_train_tfidf = corpus2csc(tfidf_transformer.transform(train_corpus), num_terms=len(id2word)).T
X_test_tfidf = corpus2csc(tfidf_transformer.transform(test_corpus), num_terms=len(id2word)).T
clf = LogisticRegression().fit(X_train_tfidf, y_train)
model_accuracy = clf.score(X_test_tfidf, y_test)
doc_scores = clf.decision_function(X_test_tfidf)
return model_accuracy, doc_scores
def sim_rank(self, query, doc_corpus, query_norm=True, doc_norm=True, res_sort=True):
assert isinstance(doc_corpus, (tuple, list)) and len(doc_corpus) > 0
if not isinstance(doc_corpus[0], (tuple, list)):
doc_corpus = [doc_corpus]
corpus_vec = [self.text2vec(c, norm=doc_norm) for c in doc_corpus]
doc_index = matutils.corpus2csc(corpus_vec, num_terms=self.dict_size, dtype=np.float32).T.tocsr()
query_vec = self.text2vec(query, norm=query_norm)
query_index = matutils.corpus2csc([query_vec], num_terms=self.dict_size, dtype=doc_index.dtype).tocsr()
sim_array = doc_index * query_index
sims = sim_array.toarray().T[0].tolist()
# sims = doc_index[query_vec]
return sorted(list(enumerate(sims)), key=lambda p: p[1], reverse=True) if res_sort else sims
def get_similarities(self, query):
"""Get similarity between `query` and this index.
Warnings
--------
Do not use this function directly; use the `self[query]` syntax instead.
Parameters
----------
query : {list of (int, number), iterable of list of (int, number), :class:`scipy.sparse.csr_matrix`}
Document or collection of documents.
Return
------
:class:`numpy.ndarray`
Similarity matrix (if maintain_sparsity=False) **OR**
:class:`scipy.sparse.csc`
otherwise
"""
is_corpus, query = utils.is_corpus(query)
if is_corpus:
query = matutils.corpus2csc(query,
self.index.shape[1],
dtype=self.index.dtype)
else:
if scipy.sparse.issparse(query):
query = query.T # convert documents=rows to documents=columns
elif isinstance(query, numpy.ndarray):
if query.ndim == 1:
query.shape = (1, len(query))
query = scipy.sparse.csr_matrix(query,
dtype=self.index.dtype).T
else:
# default case: query is a single vector, in sparse gensim format
query = matutils.corpus2csc([query],
self.index.shape[1],
dtype=self.index.dtype)
# compute cosine similarity against every other document in the collection
result = self.index * query.tocsc() # N x T * T x C = N x C
if result.shape[1] == 1 and not is_corpus:
# for queries of one document, return a 1d array
result = result.toarray().flatten()
elif self.maintain_sparsity:
# avoid converting to dense array if maintaining sparsity
result = result.T
else:
# otherwise, return a 2d matrix (#queries x #index)
result = result.toarray().T
return result
def Kmeans(n_clusters, args):
corpus = corpora.MmCorpus(args[1])
tfidf = models.tfidfmodel.TfidfModel.load(args[2])
logging.info('finished loading corpus and tfidf models')
# load the dictionary and corpus
dictionary = corpora.Dictionary.load(args[0])
tfidfValue = tfidf[corpus]
logging.info('finished tfidfValue = tfidf[corpus]')
fitdata = matutils.corpus2csc(tfidfValue,
num_terms=len(dictionary),
dtype='float32',
num_docs=len(tfidfValue),
num_nnz=None,
printprogress=0).transpose()
logging.info(
'finished transfer tfidf vector to sparse vector and transpose sparse vector'
)
logging.info('begin kmeans fit')
#TODO refactor parameters
km = KMeans(n_clusters=n_clusters,
init='random',
max_iter=100,
n_init=5,
verbose=1)
km.fit(fitdata)
logging.info('finished kmeans fit')
# l = open('./TFIDF_KmeansTrend.txt','w')
# for label in km.labels_:
# l.write(str(label)+'\n')
# l.close()
return km.labels_
def get_labelled_csc_corpus():
lsi_300_corpus = gensim.corpora.MmCorpus(
'../content-preprocessor/lsi_300_corpus.mm')
labelled_df, labelled_indices, y_matrix = get_labelled_stories()
labelled_lsi_corpus = lsi_300_corpus[labelled_indices]
labelled_lsi_csc_corpus = corpus2csc(has_tags_lsi_corpus)
return labelled_lsi_csc_corpus, y_matrix
def reduce_nlp_data(vectorizer, data, n_components, reducer):
transformed_data = vectorizer.fit_transform(data)
id2word = {
identifier: word
for word, identifier in vectorizer.vocabulary_.items()
}
if reducer == 'lda':
corpus = matutils.Sparse2Corpus(transformed_data.transpose())
lda = models.LdaModel(corpus=corpus,
num_topics=n_components,
minimum_probability=0.03,
id2word=id2word,
passes=10,
random_state=42)
print(lda.print_topics())
lda_corpus = lda[corpus]
return lda, matutils.corpus2csc(lda_corpus).toarray().transpose()
elif reducer == 'svd':
SVD = TruncatedSVD(n_components, n_iter=10, random_state=42)
svd_data = SVD.fit_transform(transformed_data)
get_eigenvectors(SVD, id2word)
return SVD, svd_data
elif reducer == 'nmf':
nmf = NMF(n_components, random_state=42)
nmf_data = nmf.fit_transform(transformed_data)
get_eigenvectors(nmf, id2word)
return nmf, nmf_data
else:
return None, None
def fit(self, corpus):
self._verify_corpus(corpus)
self.N = len(corpus)
tokens = self.preprocessor.transform(corpus)
self.observed_tokens = tokens.apply(len).sum()
vocab = Dictionary(tokens)
vocab.filter_extremes(no_above=self.max_df,
no_below=self.min_df,
keep_n=self.vocab_size)
vocab.compactify()
self.vocab = vocab
self.corpus_as_tokens = tokens
self.corpus_as_bow = [self.vocab.doc2bow(doc) for doc in tokens]
self.corpus_as_csr = corpus2csc(self.corpus_as_bow,
num_terms=len(self.vocab)).T
self.lengths = [len(d) for d in self.corpus_as_bow]
self.num_empty_docs = self.lengths.count(0)
time_now = time.localtime()
self.created_on = time.strftime("%d %b %Y %H:%M:%S", time_now)
return self
def process_data(X):
"""
:param X: X is a pandas DataFrame of tweets (with no labels)
:return: The feature matrix where each row is a tweet and each column is a feature. Ready to train/predict.
"""
tweets = process_texts(X)
dictionary = corpora.Dictionary.load('dictionary.dict')
lsi_model = models.LsiModel.load('lsi.model')
# Transform each tweet (a string) to a row of bag of words vector in the corpus matrix.
corpus = [dictionary.doc2bow(tweet.split()) for tweet in tweets]
# transform the bag of words corpus matrix to LSI matrix of features.
# To read more about LSI - https://en.wikipedia.org/wiki/Latent_semantic_analysis
feature_mat = corpus2csc(lsi_model[corpus]).T.toarray()
# Added features
# feature_mat = count_occ('!', X)
feature_mat = np.hstack((feature_mat, count_occ('!', X)))
feature_mat = np.hstack((feature_mat, count_occ('@', X)))
feature_mat = np.hstack((feature_mat, count_occ('?', X)))
feature_mat = np.hstack((feature_mat, count_occ('❤', X)))
feature_mat = np.hstack((feature_mat, count_occ('#', X)))
feature_mat = np.hstack((feature_mat, count_occ('/https:', X)))
feature_mat = np.hstack((feature_mat, chars_per_tweet(X)))
feature_mat = np.hstack((feature_mat, words_per_tweet(X)))
# feature_mat = zscore(feature_mat, axis=1)
return feature_mat, feature_mat.shape[1]
def condensify(train):
"""
Takes input either a string or a list of string
Returns a list of all summaries;
For a string returns a list with singleton document
"""
summ_list = []
if isinstance(train,string):
train = [train]
for t in train:
summ=[]
k=0
#corpus = [dictionary.doc2bow(text) for text in texts]
dictionary = corpora.Dictionary([w for w in reuters.sents(t)])
corpus = [dictionary.doc2bow(w) for w in reuters.sents(t)]
matrix = matutils.corpus2csc(corpus)
#print matrix
u,sigma,vt = sparse.linalg.svds(matrix)
(k,l)= vt.shape
while k>=1:
if reuters.sents(t)[vt[k-1].argmax()] not in summ:
summ.append(reuters.sents(t)[vt[k-1].argmax()])
k-=1
v=[]
for s in summ:
v.append(" ".join(s))
summ = "".join(v)
summ_list.append(summ)
return (summ_list)
def TF_IDF_Reg(corpus_list): tfidf_model = models.TfidfModel(corpus_list, normalize=True) courpus_list_tfidf = tfidf_model[corpus_list] word_matrix = matutils.corpus2csc(courpus_list_tfidf) return word_matrix
def add_documents(self, corpus):
"""
Extend the index with new documents.
Internally, documents are buffered and then spilled to disk when there's
`self.shardsize` of them (or when a query is issued).
"""
min_ratio = 1.0 # 0.5 to only reopen shards that are <50% complete
if self.shards and len(self.shards[-1]) < min_ratio * self.shardsize:
# The last shard was incomplete (<; load it back and add the documents there, don't start a new shard
self.reopen_shard()
for doc in corpus:
if isinstance(doc, numpy.ndarray):
doclen = len(doc)
elif scipy.sparse.issparse(doc):
doclen = doc.nnz
else:
doclen = len(doc)
if doclen < 0.3 * self.num_features:
doc = matutils.unitvec(
matutils.corpus2csc([doc], self.num_features).T)
else:
doc = matutils.unitvec(
matutils.sparse2full(doc, self.num_features))
self.fresh_docs.append(doc)
self.fresh_nnz += doclen
if len(self.fresh_docs) >= self.shardsize:
self.close_shard()
if len(self.fresh_docs) % 10000 == 0:
logger.info("PROGRESS: fresh_shard size=%i" %
len(self.fresh_docs))
def ns(numpy_matrix, number_of_corpus_features, scipy_sparse_matrix):
corpus = matutils.Dense2Corpus(numpy_matrix)
numpy_matrix = matutils.corpus2dense(corpus, num_terms=number_of_corpus_features)
corpus = matutils.Sparse2Corpus(scipy_sparse_matrix)
scipy_csc_matrix = matutils.corpus2csc(corpus)
def _setup(self, corpus):
"""Infer info from the first document and initialize matrices.
Parameters
----------
corpus : iterable of list of (int, float), optional
Training corpus.
Can be either iterable of documents, which are lists of `(word_id, word_count)`,
or a sparse csc matrix of BOWs for each document.
If not specified, the model is left uninitialized (presumably, to be trained later with `self.train()`).
"""
self._h = None
if isinstance(corpus, scipy.sparse.csc.csc_matrix):
first_doc = corpus.getcol(0)
else:
first_doc_it = itertools.tee(corpus, 1)
first_doc = next(first_doc_it[0])
first_doc = matutils.corpus2csc([first_doc], len(self.id2word))
self.w_std = np.sqrt(first_doc.mean() / (self.num_tokens * self.num_topics))
self._W = np.abs(
self.w_std
* halfnorm.rvs(
size=(self.num_tokens, self.num_topics), random_state=self.random_state
)
)
self.A = np.zeros((self.num_topics, self.num_topics))
self.B = np.zeros((self.num_tokens, self.num_topics))
def get_message_lsi_embedding_vector(self, message):
test_corpus = [self.dictionary.doc2bow(message.split())]
test_corpus_tfidf = self.tfidf[test_corpus]
test_lsi = self.lsi[test_corpus_tfidf]
test_vector = matutils.corpus2csc(test_lsi)
message_array = test_vector.toarray().reshape(-1, 1).T
return message_array
def __init__(self, corpus, num_features=None, num_terms=None, num_docs=None, num_nnz=None,
num_best=None, chunksize=500, dtype=numpy.float32, maintain_sparsity=False):
self.num_best = num_best
self.normalize = True
self.chunksize = chunksize
self.maintain_sparsity = maintain_sparsity
if corpus is not None:
logger.info("creating sparse index")
# iterate over input corpus, populating the sparse index matrix
try:
# use the more efficient corpus generation version, if the input
# `corpus` is MmCorpus-like (knows its shape and number of non-zeroes).
num_terms, num_docs, num_nnz = corpus.num_terms, corpus.num_docs, corpus.num_nnz
logger.debug("using efficient sparse index creation")
except AttributeError:
# no MmCorpus, use the slower version (or maybe user supplied the
# num_* params in constructor)
pass
if num_features is not None:
# num_terms is just an alias for num_features, for compatibility with MatrixSimilarity
num_terms = num_features
if num_terms is None:
raise ValueError("refusing to guess the number of sparse features: specify num_features explicitly")
corpus = (matutils.scipy2sparse(v) if scipy.sparse.issparse(v) else
(matutils.full2sparse(v) if isinstance(v, numpy.ndarray) else
matutils.unitvec(v)) for v in corpus)
self.index = matutils.corpus2csc(
corpus, num_terms=num_terms, num_docs=num_docs, num_nnz=num_nnz,
dtype=dtype, printprogress=10000).T
# convert to Compressed Sparse Row for efficient row slicing and multiplications
self.index = self.index.tocsr() # currently no-op, CSC.T is already CSR
logger.info("created %r", self.index)
def __init__(self, corpus, num_best=None, chunks=500, dtype=numpy.float32,
num_terms=None, num_docs=None, num_nnz=None):
self.num_best = num_best
self.normalize = True
self.chunks = chunks
if corpus is not None:
logger.info("creating sparse index")
# iterate over input corpus, populating the sparse index matrix
try:
# use the more efficient corpus generation version, if the input
# `corpus` is MmCorpus-like (knows its shape and number of non-zeroes).
num_terms, num_docs, num_nnz = corpus.num_terms, corpus.num_docs, corpus.num_nnz
logger.debug("using efficient sparse index creation")
except AttributeError:
# no MmCorpus, use the slower version (or maybe user supplied the
# num_* params in constructor)
pass
self.index = matutils.corpus2csc((matutils.unitvec(vector) for vector in corpus),
num_terms=num_terms, num_docs=num_docs, num_nnz=num_nnz,
dtype=numpy.float32, printprogress=10000).T
# convert to Compressed Sparse Row for efficient row slicing and multiplications
self.index = self.index.tocsr() # currently no-op, CSC.T is already CSR
logger.info("created %r" % self.index)
def _setup(self, corpus):
"""Infer info from the first document and initialize matrices.
Parameters
----------
corpus : iterable of list(int, float)
Training corpus.
"""
self._h, self._r = None, None
first_doc_it = itertools.tee(corpus, 1)
first_doc = next(first_doc_it[0])
first_doc = matutils.corpus2csc([first_doc], len(self.id2word))
self.w_std = np.sqrt(first_doc.mean() / (self.num_tokens * self.num_topics))
self._W = np.abs(
self.w_std
* halfnorm.rvs(
size=(self.num_tokens, self.num_topics), random_state=self.random_state
)
)
is_great_enough = self._W > self.w_std * self.sparse_coef
self._W *= is_great_enough | ~is_great_enough.all(axis=0)
self._W = scipy.sparse.csc_matrix(self._W)
self.A = scipy.sparse.csr_matrix((self.num_topics, self.num_topics))
self.B = scipy.sparse.csc_matrix((self.num_tokens, self.num_topics))
def getLDaTopics(self,
doc,
number_of_topics,
passe=20,
iters=100,
chunk=2000,
gram=(1, 2),
option='c'):
dictionary, doc_term_matrix = self.prepare_corpus(doc, gram, option)
# generate LDA model
lda = models.LdaModel(corpus=doc_term_matrix,
id2word=dictionary,
num_topics=number_of_topics,
iterations=iters,
passes=passe,
chunksize=chunk,
random_state=1) # train model
#print(ldamodel.print_topics(),'\n')
display(lda.print_topics())
corpus_transformed = lda[doc_term_matrix]
all_topics_csr = matutils.corpus2csc(corpus_transformed)
all_topics_numpy = all_topics_csr.T.toarray()
#Lda_Topic=pd.DataFrame(all_topics_numpy)
Lda_Topic = pd.DataFrame(all_topics_numpy, doc)
display(Lda_Topic.head(5))
print('shape ', Lda_Topic.shape)
return Lda_Topic
def __init__(self, m, k, docs = None):
"""
Store (U, S) projection itself. This is the class taking care of 'core math';
interfacing with corpora, training etc is done through class LsiModel.
`docs` is either a spare matrix or a corpus which, when converted to a
sparse matrix, must fit comfortably into main memory.
"""
self.m, self.k = m, k
if docs is not None:
# base case decomposition: given a job `docs`, compute its decomposition
# in core, algorithm 1
if utils.isCorpus(docs):
docs = matutils.corpus2csc(m, docs)
if m * k < 10000:
# SVDLIBC gives spurious results for small matrices.. run full
# LAPACK svd on them instead
docs = docs.todense()
logger.info("computing dense SVD of %s matrix" % str(docs.shape))
u, s, vt = numpy.linalg.svd(docs, full_matrices = False)
else:
try:
import sparsesvd
except ImportError:
raise ImportError("for LSA, the `sparsesvd` module is needed but not found; run `easy_install sparsesvd`")
logger.info("computing sparse SVD of %s matrix" % str(docs.shape))
ut, s, vt = sparsesvd.sparsesvd(docs, k + 30) # ask for extra factors, because for some reason SVDLIBC sometimes returns fewer factors than requested
u = ut.T
del ut
del vt
k = clipSpectrum(s, self.k)
self.u, self.s = u[:, :k], s[:k]
else:
self.u, self.s = None, None
def __init__(self, m, k, docs=None, use_svdlibc=False, power_iters=P2_EXTRA_ITERS, extra_dims=P2_EXTRA_DIMS):
"""
Construct the (U, S) projection from a corpus `docs`. The projection can
be later updated by merging it with another Projection via `self.merge()`.
This is the class taking care of the 'core math'; interfacing with corpora,
splitting large corpora into chunks and merging them etc. is done through
the higher-level `LsiModel` class.
"""
self.m, self.k = m, k
self.power_iters = power_iters
self.extra_dims = extra_dims
if docs is not None:
# base case decomposition: given a job `docs`, compute its decomposition,
# *in-core*.
if not use_svdlibc:
u, s = stochastic_svd(docs, k, chunksize=sys.maxsize,
num_terms=m, power_iters=self.power_iters,
extra_dims=self.extra_dims)
else:
try:
import sparsesvd
except ImportError:
raise ImportError("`sparsesvd` module requested but not found; run `easy_install sparsesvd`")
logger.info("computing sparse SVD of %s matrix" % str(docs.shape))
if not scipy.sparse.issparse(docs):
docs = matutils.corpus2csc(docs)
ut, s, vt = sparsesvd.sparsesvd(docs, k + 30) # ask for extra factors, because for some reason SVDLIBC sometimes returns fewer factors than requested
u = ut.T
del ut, vt
k = clip_spectrum(s**2, self.k)
self.u = u[:, :k].copy()
self.s = s[:k].copy()
else:
self.u, self.s = None, None
def __init__(self, m, k, docs=None, use_svdlibc=False, power_iters=P2_EXTRA_ITERS, extra_dims=P2_EXTRA_DIMS):
"""
Construct the (U, S) projection from a corpus `docs`. The projection can
be later updated by merging it with another Projection via `self.merge()`.
This is the class taking care of the 'core math'; interfacing with corpora,
splitting large corpora into chunks and merging them etc. is done through
the higher-level `LsiModel` class.
"""
self.m, self.k = m, k
self.power_iters = power_iters
self.extra_dims = extra_dims
if docs is not None:
# base case decomposition: given a job `docs`, compute its decomposition,
# *in-core*.
if not use_svdlibc:
u, s = stochastic_svd(docs, k, chunksize=sys.maxsize,
num_terms=m, power_iters=self.power_iters,
extra_dims=self.extra_dims)
else:
try:
import sparsesvd
except ImportError:
raise ImportError("`sparsesvd` module requested but not found; run `easy_install sparsesvd`")
logger.info("computing sparse SVD of %s matrix" % str(docs.shape))
if not scipy.sparse.issparse(docs):
docs = matutils.corpus2csc(docs)
ut, s, vt = sparsesvd.sparsesvd(docs, k + 30) # ask for extra factors, because for some reason SVDLIBC sometimes returns fewer factors than requested
u = ut.T
del ut, vt
k = clip_spectrum(s**2, self.k)
self.u = u[:, :k].copy()
self.s = s[:k].copy()
else:
self.u, self.s = None, None
def _doc_doc_mtx(table, model, input_col, result_type='sparse'):
corpus = table[input_col].tolist()
dictionary = model['dictionary']
bow_corpus = []
for doc in corpus:
bow_corpus.append(dictionary.doc2bow(doc))
csr_matrix = matutils.corpus2csc(bow_corpus).T
csr_matrix.data = np.array([1 for _ in range(len(csr_matrix.data))])
doc_doc = (csr_matrix @ (csr_matrix.T)).tocoo()
if result_type == 'sparse':
doc_doc = sparse.triu(doc_doc, k=1)
out_table = pd.DataFrame(doc_doc.row, columns=['1st_document_idx'])
out_table['2nd_document_idx'] = doc_doc.col
out_table['number_of_common_terms'] = doc_doc.data
elif result_type == 'dense':
doc_idx = ['doc_{}'.format(i) for i in range(len(corpus))]
out_table = pd.DataFrame(doc_doc.todense())
out_table.insert(loc=0, column=' ', value=doc_idx)
out_table.columns = np.append("", doc_idx)
else:
raise_runtime_error("Please check 'result_type'.")
rb = BrtcReprBuilder()
model = _model_dict('doc_doc_mtx')
model['input_col'] = input_col
model['doc_doc_mtx'] = doc_doc
model['_repr_brtc_'] = rb.get()
return {'out_table': out_table}
def __init__(self, m, k, docs = None):
"""
Store (U, S) projection itself. This is the class taking care of 'core math';
interfacing with corpora, training etc is done through class LsiModel.
`docs` is either a spare matrix or a corpus which, when converted to a
sparse matrix, must fit comfortably into main memory.
"""
self.m, self.k = m, k
if docs is not None:
# base case decomposition: given a job `docs`, compute its decomposition
# in core, algorithm 1
if utils.isCorpus(docs):
docs = matutils.corpus2csc(m, docs)
if m * k < 10000:
# SVDLIBC gives spurious results for small matrices.. run full
# LAPACK svd on them instead
docs = docs.todense()
logger.info("computing dense SVD of %s matrix" % str(docs.shape))
u, s, vt = numpy.linalg.svd(docs, full_matrices = False)
else:
try:
import sparsesvd
except ImportError:
raise ImportError("for LSA, the `sparsesvd` module is needed but not found; run `easy_install sparsesvd`")
logger.info("computing sparse SVD of %s matrix" % str(docs.shape))
ut, s, vt = sparsesvd.sparsesvd(docs, k + 30) # ask for extra factors, because for some reason SVDLIBC sometimes returns fewer factors than requested
u = ut.T
del ut
del vt
k = clipSpectrum(s, self.k)
self.u, self.s = u[:, :k], s[:k]
else:
self.u, self.s = None, None
def classify_content(content):
num_terms = len(dictionary)
test_corpus = tfidf_model[dictionary.doc2bow(list(Tokenize(content)))]
test_sparse = matutils.corpus2csc([test_corpus],
num_terms).transpose(copy=False)
result = sg_class.predict(test_sparse)
return id2cls[result[0]]
def train():
with open('count_corpus.pkl', 'rb') as f:
id_corpus = pickle.load(f)
#lda = models.ldamulticore.LdaMulticore(corpus=id_corpus, num_topics=50)
# lda.save('lda.model')
p = Pool()
aaa = p.map(pred, id_corpus)
result = numpy.asarray(corpus2csc(aaa).T.todense())
p.close()
p.join
map_train, map_test, train_num = make_idmap()
df = pandas.read_csv('../data/train_clean.csv')[['question1', 'question2'
]].fillna('').values
df_train = pandas.DataFrame(_train(result[:train_num], df, map_train))
df_train.to_csv('lda_train.csv', index=False)
df = pandas.read_csv('../data/test_clean.csv')[['question1', 'question2'
]].fillna('').values
df_test = pandas.DataFrame(_train(result[train_num:], df, map_test))
df_test.to_csv('lda_test.csv', index=False)
def tfidf(counts_mat):
m,n = counts_mat.shape
# utilize gensim
corpus = matutils.Sparse2Corpus(counts_mat)
tfidf = models.logentropy_model.LogEntropyModel(corpus)
c_tfidf = tfidf[corpus]
return matutils.corpus2csc(c_tfidf)
def process_records(records, fields, target, textmodel=None):
tokenize = CountVectorizer().build_analyzer()
input = None
X = None
y_labels = []
for i, record in enumerate(records):
nums = []
strs = []
y_labels.append(record.get(target))
for field in fields:
if is_number(record.get(field)):
nums.append(record[field])
else:
strs.append(str(record.get(field) or "").lower())
if strs:
if input is None:
input = StringIO.StringIO()
print >> input, " ".join(tokenize(" ".join(strs)))
if nums:
if X is None:
X = sp.lil_matrix((len(records),len(nums)))
X[i] = np.array(nums, dtype=np.float64)
if input is not None:
if X is not None:
X_2 = X.tocsr()
else:
X_2 = None
if isinstance(textmodel,basestring):
if textmodel == 'lsi':
corpus = TextCorpus(input)
textmodel = LsiModel(corpus, chunksize=1000)
elif textmodel == 'tfidf':
corpus = TextCorpus(input)
textmodel = TfidfModel(corpus)
elif textmodel == 'hashing':
textmodel = None
hasher = FeatureHasher(n_features=2 ** 18, input_type="string")
input.seek(0)
X = hasher.transform(tokenize(line.strip()) for line in input)
if textmodel:
num_terms = len(textmodel.id2word or getattr(textmodel, 'dfs',[]))
X = corpus2csc(textmodel[corpus], num_terms).transpose()
if X_2 is not None:
# print >> sys.stderr, "X SHAPE:", X.shape
# print >> sys.stderr, "X_2 SHAPE:", X_2.shape
X = sp.hstack([X, X_2], format='csr')
elif X is not None:
textmodel = None
X = X.tocsr()
print >> sys.stderr, "X SHAPE:", X.shape
return X, y_labels, textmodel
def __getitem__(self, bow, scaled=False, chunksize=512):
"""
Return latent representation, as a list of (topic_id, topic_value) 2-tuples.
This is done by folding input document into the latent topic space.
"""
assert self.projection.u is not None, "decomposition not initialized yet"
# if the input vector is in fact a corpus, return a transformed corpus as a result
is_corpus, bow = utils.is_corpus(bow)
if is_corpus and chunksize:
# by default, transform 256 documents at once, when called as `lsi[corpus]`.
# this chunking is completely transparent to the user, but it speeds
# up internal computations (one mat * mat multiplication, instead of
# 256 smaller mat * vec multiplications).
return self._apply(bow, chunksize=chunksize)
if not is_corpus:
bow = [bow]
vec = matutils.corpus2csc(bow, num_terms=self.num_terms)
topic_dist = (vec.T * self.projection.u[:, :self.num_topics]).T # (x^T * u).T = u^-1 * x
if scaled:
topic_dist = (1.0 / self.projection.s[:self.num_topics]) * topic_dist # s^-1 * u^-1 * x
# convert a numpy array to gensim sparse vector = tuples of (feature_id, feature_weight),
# with no zero weights.
if not is_corpus:
# lsi[single_document]
result = matutils.full2sparse(topic_dist.flat)
else:
# lsi[chunk of documents]
result = matutils.Dense2Corpus(topic_dist)
return result
def condensify(train):
"""
Takes input either a string or a list of string
Returns a list of all summaries;
For a string returns a list with singleton document
"""
summ_list = []
if isinstance(train, string):
train = [train]
for t in train:
summ = []
k = 0
#corpus = [dictionary.doc2bow(text) for text in texts]
dictionary = corpora.Dictionary([w for w in reuters.sents(t)])
corpus = [dictionary.doc2bow(w) for w in reuters.sents(t)]
matrix = matutils.corpus2csc(corpus)
#print matrix
u, sigma, vt = sparse.linalg.svds(matrix)
(k, l) = vt.shape
while k >= 1:
if reuters.sents(t)[vt[k - 1].argmax()] not in summ:
summ.append(reuters.sents(t)[vt[k - 1].argmax()])
k -= 1
v = []
for s in summ:
v.append(" ".join(s))
summ = "".join(v)
summ_list.append(summ)
return (summ_list)
def train(args, output_dir):
"""Build the corpus, trains the DTM, and saves the model to the output
dir."""
corpus = Corpus()
# Create the dictionary.
dictionary = Dictionary(corpus.debates.bag_of_words)
dictionary.filter_extremes(no_below=100)
# Save empirical term distribution within each time step.
term_counts = corpus2csc(
corpus.debates.groupby('year').agg({
'bag_of_words': 'sum'
}).bag_of_words.apply(dictionary.doc2bow))
save_npz(os.path.join(output_dir, 'term_counts.npz'), term_counts)
# Train and save dtm.
time_slices = corpus.debates.groupby('year').size()
dtm_corpus = corpus.debates.bag_of_words.apply(dictionary.doc2bow)
model = Dtm(args.executable,
corpus=dtm_corpus,
id2word=dictionary,
num_topics=args.num_topics,
time_slices=time_slices.values,
rng_seed=args.random_seed)
model.save(os.path.join(output_dir, 'dtm.gensim'))
def _transform(self, corpus, source_dict=None):
temp_corpus = list(corpus.ngrams_iterator(' ', include_postags=True))
dic = corpora.Dictionary(
temp_corpus, prune_at=None) if not source_dict else source_dict
temp_corpus = [dic.doc2bow(doc) for doc in temp_corpus]
model = models.TfidfModel(temp_corpus,
normalize=False,
wlocal=self.wlocals[self.wlocal],
wglobal=self.wglobals[self.wglobal])
X = matutils.corpus2csc(model[temp_corpus],
dtype=np.float,
num_terms=len(dic)).T
norm = self.norms[self.norm]
if norm:
X = norm(X)
# set compute values
shared_cv = SharedTransform(self,
corpus.used_preprocessor,
source_dict=dic)
cv = [
VectorizationComputeValue(shared_cv, dic[i])
for i in range(len(dic))
]
corpus = self.add_features(corpus,
X,
dic,
cv,
var_attrs={'bow-feature': True})
return corpus
def serialize(filename_prefix, layer, current_representation, num_terms=None, chunksize=10000):
is_corpus, current_representation = utils.is_corpus(current_representation)
if is_corpus:
for chunk_no, chunk in enumerate(utils.grouper(current_representation, chunksize)):
ln.debug("preparing chunk for conversion (%s documents)..." % len(chunk))
assert num_terms is not None, "Need num_terms to properly handle sparse corpus format"
chunk_as_csc = matutils.corpus2csc(chunk, num_terms=num_terms)
ln.debug("Chunk converted to csc, running through layer..")
chunk_trans = layer.__getitem__(chunk_as_csc)
ln.debug("Serializing hidden representation..")
fname = "%s_%s" % (filename_prefix, ("0" * (15 - len(str(chunk_no)))) + str(chunk_no))
np.save(fname, chunk_trans)
ln.debug("Finished serializing chunk. Processed %s documents so far." %
(chunk_no * chunksize + len(chunk)))
else:
ln.info("Beginning serialization of non-gensim corpus format intermediate representation.")
ln.debug("Type of current_representation is %s" % type(current_representation))
for chunk_no, chunk in enumerate(current_representation):
ln.debug("converting chunk (%s documents)..." % chunksize)
chunk_trans = layer.__getitem__(chunk)
ln.debug("Serializing hidden representation..")
fname = "%s_%s" % (filename_prefix, ("0" * (15 - len(str(chunk_no)))) + str(chunk_no))
np.save(fname, chunk_trans)
ln.debug("finished serializing chunk.")
ln.info("Finished serializing all chunks.")
def _get_vectors(self):
processed_corpus = [self._tokenize(doc["document"], self.phraser) for doc in self.docs]
self.dictionary = corpora.Dictionary(processed_corpus)
# ignore 20% most frequent words
# num_unique_words = len(dictionary)
# dictionary.filter_n_most_frequent(int(num_unique_words*0.2))
# do some more filtering and keep only n most frequent specified with num_dims parameter
self.dictionary.filter_extremes(no_below=1, no_above=0.8, keep_n=self.num_dims)
bow_corpus = [self.dictionary.doc2bow(text) for text in processed_corpus]
if self.vectorizer == "TfIdf Vectorizer":
self.tfidf_model = models.TfidfModel(bow_corpus)
transformed_corpus = self.tfidf_model[bow_corpus]
elif self.vectorizer == "Count Vectorizer":
transformed_corpus = bow_corpus
if self.use_lsi:
self.lsi_model = models.LsiModel(transformed_corpus, id2word=self.dictionary, num_topics=self.num_topics)
transformed_corpus = self.lsi_model[transformed_corpus]
matrix = corpus2csc(transformed_corpus, num_terms=len(self.dictionary.keys()), num_docs=self.dictionary.num_docs)
return matrix.transpose()
def get_intracluster_similarity(self, new_documents=[], phraser=None):
if len(new_documents) > 0:
dictionary = self.models["dictionary"]
for doc in new_documents:
processed_text = Clustering._tokenize(doc["text"], phraser=phraser)
doc_vec = [dictionary.doc2bow(processed_text)]
if self.models["tfidf_model"] is not None:
doc_vec = self.models["tfidf_model"][doc_vec]
if self.models["lsi_model"] is not None:
doc_vec = self.models["lsi_model"][doc_vec]
full_vec = corpus2csc(doc_vec, num_terms=len(dictionary.keys()), num_docs=dictionary.num_docs)
full_vec = full_vec.transpose()
self.models["doc_vectors"][doc["id"]] = full_vec[0].toarray()[0]
self._save_updated_models()
if self.doc_ids:
cluster_vectors = []
for doc_id in self.doc_ids:
cluster_vectors.append(self.models["doc_vectors"][doc_id])
similarities = cosine_similarity(cluster_vectors)
return np.mean(similarities)
else:
return 0
def add_documents(self, corpus):
"""
Extend the index with new documents.
Internally, documents are buffered and then spilled to disk when there's
`self.shardsize` of them (or when a query is issued).
"""
min_ratio = 1.0 # 0.5 to only reopen shards that are <50% complete
if self.shards and len(self.shards[-1]) < min_ratio * self.shardsize:
# The last shard was incomplete (<; load it back and add the documents there, don't start a new shard
self.reopen_shard()
for doc in corpus:
if isinstance(doc, numpy.ndarray):
doclen = len(doc)
elif scipy.sparse.issparse(doc):
doclen = doc.nnz
else:
doclen = len(doc)
if doclen < 0.3 * self.num_features:
doc = matutils.unitvec(matutils.corpus2csc([doc], self.num_features).T, self.norm)
else:
doc = matutils.unitvec(matutils.sparse2full(doc, self.num_features), self.norm)
self.fresh_docs.append(doc)
self.fresh_nnz += doclen
if len(self.fresh_docs) >= self.shardsize:
self.close_shard()
if len(self.fresh_docs) % 10000 == 0:
logger.info("PROGRESS: fresh_shard size=%i", len(self.fresh_docs))
def update(self, corpus):
"""Train the model with new documents.
Parameters
----------
corpus : iterable of list of (int, float), optional
Training corpus.
Can be either iterable of documents, which are lists of `(word_id, word_count)`,
or a sparse csc matrix of BOWs for each document.
If not specified, the model is left uninitialized (presumably, to be trained later with `self.train()`).
"""
if self._W is None:
self._setup(corpus)
chunk_idx = 1
for _ in range(self.passes):
if isinstance(corpus, scipy.sparse.csc.csc_matrix):
grouper = (
corpus[:, col_idx:col_idx + self.chunksize]
for col_idx
in range(0, corpus.shape[1], self.chunksize)
)
else:
grouper = utils.grouper(corpus, self.chunksize)
for chunk in grouper:
if isinstance(corpus, scipy.sparse.csc.csc_matrix):
v = chunk[:, self.random_state.permutation(chunk.shape[1])]
else:
self.random_state.shuffle(chunk)
v = matutils.corpus2csc(
chunk,
num_terms=self.num_tokens,
)
self._h = self._solveproj(v, self._W, h=self._h, v_max=self.v_max)
h = self._h
self.A *= chunk_idx - 1
self.A += h.dot(h.T)
self.A /= chunk_idx
self.B *= chunk_idx - 1
self.B += v.dot(h.T)
self.B /= chunk_idx
prev_w_error = self._w_error
self._solve_w()
if chunk_idx % self.eval_every == 0:
logger.info("Loss: {}".format(self._w_error / prev_w_error))
chunk_idx += 1
logger.info("Loss: {}".format(self._w_error / prev_w_error))
def main():
parser = argparse.ArgumentParser(
description=
'calculates various stats and of a given document-author-contribs file'
)
parser.add_argument(
'--acc-contribs',
type=argparse.FileType('r'),
help='path to input MatrixMarket acc contributions file (.mm/.mm.bz2)',
required=True)
parser.add_argument('--img-prefix',
help='prefix of output generated img files',
required=True)
parser.add_argument('--quantile-order',
type=float,
help='quantile of histrograms to consider',
required=True)
args = parser.parse_args()
input_acc_contribs_dump_path = args.acc_contribs.name
output_image_prefix = args.img_prefix
quantile_order = args.quantile_order
logger.info('running with:\n{}'.format(
pformat({
'input_acc_contribs_dump_path': input_acc_contribs_dump_path,
'output_image_prefix': output_image_prefix,
'quantile_order': quantile_order
})))
acc_contribs = MmCorpus(input_acc_contribs_dump_path)
logger.info('reading corpus to sparse csr matrix')
csr_corpus = corpus2csc(acc_contribs).T.tocsr()
logger.info('generated sparse matrix of shape {}'.format(csr_corpus.shape))
logger.debug('sparse matrix \n{}'.format(csr_corpus))
logger.info('calculating authors-per-docs-distribution')
num_authors_per_doc = sp.find((csr_corpus > 0).sum(1))[2]
quantile = get_quantile(num_authors_per_doc, quantile_order)
num_authors_per_doc = num_authors_per_doc[num_authors_per_doc <= quantile]
num_authors_per_doc_imgfile = output_image_prefix + '-num-auths-per-doc-dist.pdf'
xlabel = 'Autoren je Dokument'
ylabel = 'Häufigkeit'
num_authors, num_authors_counts = np.unique(num_authors_per_doc,
return_counts=True)
bar_plot(num_authors, num_authors_counts, num_authors_per_doc_imgfile,
xlabel, ylabel)
logger.info('calculating docs-per-authors-distribution')
num_docs_per_author = sp.find((csr_corpus > 0).sum(0).T)[2]
quantile = get_quantile(num_docs_per_author, quantile_order)
num_docs_per_author = num_docs_per_author[num_docs_per_author <= quantile]
num_docs_per_author_imgfile = output_image_prefix + '-num-docs-per-auth-dist.pdf'
xlabel = 'Dokumente je Autor'
ylabel = 'Häufigkeit'
num_docs, num_docs_counts = np.unique(num_docs_per_author,
return_counts=True)
bar_plot(num_docs, num_docs_counts, num_docs_per_author_imgfile, xlabel,
ylabel)
def _extract_data(topic_model, corpus, dictionary, doc_topic_dists=None):
if not matutils.ismatrix(corpus):
corpus_csc = matutils.corpus2csc(corpus, num_terms=len(dictionary))
else:
corpus_csc = corpus
# Need corpus to be a streaming gensim list corpus for len and inference functions below:
corpus = matutils.Sparse2Corpus(corpus_csc)
beta = 0.01
fnames_argsort = np.asarray(list(dictionary.token2id.values()),
dtype=np.int_)
term_freqs = corpus_csc.sum(axis=1).A.ravel()[fnames_argsort]
term_freqs[term_freqs == 0] = beta
doc_lengths = corpus_csc.sum(axis=0).A.ravel()
assert term_freqs.shape[0] == len(
dictionary
), 'Term frequencies and dictionary have different shape {} != {}'.format(
term_freqs.shape[0], len(dictionary))
assert doc_lengths.shape[0] == len(
corpus
), 'Document lengths and corpus have different sizes {} != {}'.format(
doc_lengths.shape[0], len(corpus))
if hasattr(topic_model, 'lda_alpha'):
num_topics = len(topic_model.lda_alpha)
else:
num_topics = topic_model.num_topics
if doc_topic_dists is None:
# If its an HDP model.
if hasattr(topic_model, 'lda_beta'):
gamma = topic_model.inference(corpus)
else:
gamma, _ = topic_model.inference(corpus)
doc_topic_dists = gamma / gamma.sum(axis=1)[:, None]
else:
if isinstance(doc_topic_dists, list):
doc_topic_dists = matutils.corpus2dense(doc_topic_dists,
num_topics).T
elif issparse(doc_topic_dists):
doc_topic_dists = doc_topic_dists.T.todense()
doc_topic_dists = doc_topic_dists / doc_topic_dists.sum(axis=1)
assert doc_topic_dists.shape[
1] == num_topics, 'Document topics and number of topics do not match {} != {}'.format(
doc_topic_dists.shape[1], num_topics)
# get the topic-term distribution straight from gensim without
# iterating over tuples
if hasattr(topic_model, 'lda_beta'):
topic = topic_model.lda_beta
else:
topic = topic_model.state.get_lambda()
topic = topic / topic.sum(axis=1)[:, None]
topic_term_dists = topic[:, fnames_argsort]
assert topic_term_dists.shape[0] == doc_topic_dists.shape[1]
return doc_topic_dists
def get_similarities(self, query):
"""Get similarity between `query` and this index.
Warnings
--------
Do not use this function directly; use the `self[query]` syntax instead.
Parameters
----------
query : {list of (int, number), iterable of list of (int, number), :class:`scipy.sparse.csr_matrix`}
Document or collection of documents.
Return
------
:class:`numpy.ndarray`
Similarity matrix (if maintain_sparsity=False) **OR**
:class:`scipy.sparse.csc`
otherwise
"""
is_corpus, query = utils.is_corpus(query)
if is_corpus:
query = matutils.corpus2csc(query, self.index.shape[1], dtype=self.index.dtype)
else:
if scipy.sparse.issparse(query):
query = query.T # convert documents=rows to documents=columns
elif isinstance(query, numpy.ndarray):
if query.ndim == 1:
query.shape = (1, len(query))
query = scipy.sparse.csr_matrix(query, dtype=self.index.dtype).T
else:
# default case: query is a single vector, in sparse gensim format
query = matutils.corpus2csc([query], self.index.shape[1], dtype=self.index.dtype)
# compute cosine similarity against every other document in the collection
result = self.index * query.tocsc() # N x T * T x C = N x C
if result.shape[1] == 1 and not is_corpus:
# for queries of one document, return a 1d array
result = result.toarray().flatten()
elif self.maintain_sparsity:
# avoid converting to dense array if maintaining sparsity
result = result.T
else:
# otherwise, return a 2d matrix (#queries x #index)
result = result.toarray().T
return result
def transform(self, X):
x_clean = tu.clean_and_tokenize(X, stoplist=self.stoplist)
x_tfidf = self.tfidf[[self.dictionary.doc2bow(text) for text in x_clean]]
x_data = matutils.corpus2csc(x_tfidf, num_terms=len(self.dictionary)).T
#x_data = matutils.corpus2dense(x_tfidf, num_terms=len(self.dictionary)).T
#logging.info("Returning data of shape %s " % (len(x_data)))
#returning a csr matrix
return x_data
def setup(files):
# setup the output directory
base_model_name = os.path.splitext(os.path.basename(files.model))[0]
output_dir = '../browser/json/' + base_model_name + '/'
if not os.path.exists(output_dir):
os.makedirs(output_dir)
# load the topic model
model = LdaModel.load(files.model)
# load replacements used
bug_to_id = json.loads(open(files.replacements).read())
# invert to id<->bug map, ditching s. genus terms
id_to_bug = {v: k for k, v in bug_to_id.items() if "." not in k}
# load the docsXwords and docsXtopics matrices (in sparse format)
corpus = mmcorpus.MmCorpus(files.corpus)
docsXwords_sparse = corpus2csc(corpus, num_terms=len(model.id2word.token2id)).T
docsXtopics = mmcorpus.MmCorpus(files.docsXtopics)
docsXtopics_sparse = corpus2csc(docsXtopics).T
return docsXtopics_sparse, docsXwords_sparse, id_to_bug, model, output_dir
def hierarchical_clustering(corpus_fn, n_clusters=2, linkage='complete'):
corpus = corpora.MmCorpus(corpus_fn)
corpus = matutils.corpus2csc(corpus, num_terms=corpus.num_terms).transpose()
svd = TruncatedSVD(n_components=100)
new_corpus = svd.fit_transform(corpus)
knn_graph = kneighbors_graph(new_corpus, 10, metric='euclidean')
agg = AgglomerativeClustering(n_clusters=n_clusters, affinity='euclidean', linkage=linkage, connectivity=knn_graph)
agg.fit(new_corpus)
return corpus, agg.labels_
def convert_to_X_y(model_class, params, data, label): model = model_class(**params) mat = model[data] X = corpus2csc(mat) y = np.array(label) return X.T, y
def __getitem__(self, bow, scaled=False, chunksize=512):
"""
Return latent representation, as a list of (topic_id, topic_value) 2-tuples.
This is done by folding input document into the latent topic space.
If `scaled` is set, scale topics by the inverse of singular values (default: no scaling).
"""
assert self.projection.u is not None, "decomposition not initialized yet"
# if the input vector is in fact a corpus, return a transformed corpus as a result
is_corpus, bow = utils.is_corpus(bow)
if is_corpus and chunksize:
# by default, transform `chunksize` documents at once, when called as `lsi[corpus]`.
# this chunking is completely transparent to the user, but it speeds
# up internal computations (one mat * mat multiplication, instead of
# `chunksize` smaller mat * vec multiplications).
return self._apply(bow, chunksize=chunksize)
if not is_corpus:
bow = [bow]
# convert input to scipy.sparse CSC, then do "sparse * dense = dense" multiplication
vec = matutils.corpus2csc(bow, num_terms=self.num_terms, dtype=self.projection.u.dtype)
topic_dist = (vec.T * self.projection.u[:, :self.num_topics]).T # (x^T * u).T = u^-1 * x
# # convert input to dense, then do dense * dense multiplication
# # ± same performance as above (BLAS dense * dense is better optimized than scipy.sparse),
# but consumes more memory
# vec = matutils.corpus2dense(bow, num_terms=self.num_terms, num_docs=len(bow))
# topic_dist = np.dot(self.projection.u[:, :self.num_topics].T, vec)
# # use np's advanced indexing to simulate sparse * dense
# # ± same speed again
# u = self.projection.u[:, :self.num_topics]
# topic_dist = np.empty((u.shape[1], len(bow)), dtype=u.dtype)
# for vecno, vec in enumerate(bow):
# indices, data = zip(*vec) if vec else ([], [])
# topic_dist[:, vecno] = np.dot(u.take(indices, axis=0).T, np.array(data, dtype=u.dtype))
if not is_corpus:
# convert back from matrix into a 1d vec
topic_dist = topic_dist.reshape(-1)
if scaled:
topic_dist = (1.0 / self.projection.s[:self.num_topics]) * topic_dist # s^-1 * u^-1 * x
# convert a np array to gensim sparse vector = tuples of (feature_id, feature_weight),
# with no zero weights.
if not is_corpus:
# lsi[single_document]
result = matutils.full2sparse(topic_dist)
else:
# lsi[chunk of documents]
result = matutils.Dense2Corpus(topic_dist)
return result
def readTest(self, test_deals_file):
"""
read test data
@param test_deals
@return
"""
corpus = [self.dict.doc2bow(line.split()) for line in open(test_deals_file, 'r')]
self.X_test = matutils.corpus2csc(corpus, num_terms=len(self.dict)).T
#joblib.dump(self.X_test, 'task3_X_test.pkl')
print 'readTest Done!'
def load_sparse_dataset(colname,suffix):
corpus = corpora.MmCorpus('vsm/{}.{}'.format(colname,suffix))
# with documents as columns
X = matutils.corpus2csc(corpus,printprogress=1)
# transpose to make each document a row
X = X.T
y = pd.read_csv("vsm/{}_meta.csv".format(colname),index_col='id')
y = y.iloc[:,0]# DataFrame to Series
return X,y
def evaluate(corpus_fn, labels_fn):
corpus = corpora.MmCorpus(corpus_fn)
labels = pickle.load(open(labels_fn))
corpus = matutils.corpus2csc(corpus, num_terms=corpus.num_terms).transpose()
scores = []
for i in range(16):
score = metrics.silhouette_score(corpus, labels, metric='cosine', sample_size=5000)
print score
scores.append(score)
sc = numpy.array(scores).mean()
print 'mean', sc
def load_tfidf(tag):
corpus = corpora.MmCorpus("{}.tfidf".format(tag))
# with documents as columns
X = matutils.corpus2csc(corpus, printprogress=1)
# transpose to make each document a row
X = X.T
y = pd.read_csv("{}_labels.csv".format(tag), index_col="date")
y = y.iloc[:, 0] # DataFrame to Series
return X, y
def key_words(keys, topk=18):
vec_bow = dictionary.doc2bow(keys)
tfidf_corpus = tfidf_model[vec_bow]
num_terms = len(dictionary)
test_sparse = matutils.corpus2csc([tfidf_corpus], num_terms).transpose(copy=False)
result = sg_class.predict(test_sparse)
words = [dictionary[d].decode("utf-8") for d, _ in sorted(list(tfidf_corpus), key=lambda item: -item[1])[0:topk]]
classify = id2cls[result[0]]
return (words, classify)
def transformed_corpus():
for chunk_no, doc_chunk in enumerate(utils.grouper(bow, chunksize)):
ln.debug("Converting chunk %s to csc format.." % chunk_no)
chunk = matutils.corpus2csc(doc_chunk, self.input_dimensionality)
ln.debug("Computing hidden representation for chunk.. ")
hidden = self._get_hidden_representations(chunk)
ln.info("Finished computing representation for chunk %s, yielding results. Total docs processed: %s" %
(chunk_no, chunk_no * chunksize + len(doc_chunk)))
for column in hidden.T:
yield matutils.dense2vec(column.T)
ln.debug("Done yielding chunk %s" % chunk_no)
ln.info("Finished computing representations for all chunks.")
def __init__(self, m, k, docs=None, use_svdlibc=False, power_iters=P2_EXTRA_ITERS,
extra_dims=P2_EXTRA_DIMS, dtype=np.float64):
"""Construct the (U, S) projection from a corpus.
Parameters
----------
m : int
Number of features (terms) in the corpus.
k : int
Desired rank of the decomposed matrix.
docs : {iterable of list of (int, float), scipy.sparse.csc}
Corpus in BoW format or as sparse matrix.
use_svdlibc : bool, optional
If True - will use `sparsesvd library <https://pypi.python.org/pypi/sparsesvd/>`_,
otherwise - our own version will be used.
power_iters: int, optional
Number of power iteration steps to be used. Tune to improve accuracy.
extra_dims : int, optional
Extra samples to be used besides the rank `k`. Tune to improve accuracy.
dtype : numpy.dtype, optional
Enforces a type for elements of the decomposed matrix.
"""
self.m, self.k = m, k
self.power_iters = power_iters
self.extra_dims = extra_dims
if docs is not None:
# base case decomposition: given a job `docs`, compute its decomposition,
# *in-core*.
if not use_svdlibc:
u, s = stochastic_svd(
docs, k, chunksize=sys.maxsize,
num_terms=m, power_iters=self.power_iters,
extra_dims=self.extra_dims, dtype=dtype)
else:
try:
import sparsesvd
except ImportError:
raise ImportError("`sparsesvd` module requested but not found; run `easy_install sparsesvd`")
logger.info("computing sparse SVD of %s matrix", str(docs.shape))
if not scipy.sparse.issparse(docs):
docs = matutils.corpus2csc(docs)
# ask for extra factors, because for some reason SVDLIBC sometimes returns fewer factors than requested
ut, s, vt = sparsesvd.sparsesvd(docs, k + 30)
u = ut.T
del ut, vt
k = clip_spectrum(s ** 2, self.k)
self.u = u[:, :k].copy()
self.s = s[:k].copy()
else:
self.u, self.s = None, None
def main(model, dic, corpus, output):
logging.basicConfig(level=logging.INFO)
score = tfidf.scorer(model, dic)
transforms = []
with open(corpus) as fp:
n_sentences = sum(1 for line in fp)
logging.info('Computing tf-idf vectors for %d sentences', n_sentences)
bar = pb.ProgressBar(widgets=[pb.Percentage(), pb.Bar(), pb.ETA()], maxval=n_sentences)
with open(corpus) as fp:
for sentence in bar(fp):
transforms.append(score(sentence.split()))
logging.info('Saving tf-idf information to %s', output)
with open(output, 'w') as fp:
cPickle.dump(corpus2csc(transforms), fp, protocol=cPickle.HIGHEST_PROTOCOL)
def add_documents(self, corpus):
"""Extend the index with new documents.
Parameters
----------
corpus : iterable of list of (int, number)
Corpus in BoW format.
Notes
-----
Internally, documents are buffered and then spilled to disk when there's `self.shardsize` of them
(or when a query is issued).
Examples
--------
.. sourcecode:: pycon
>>> from gensim.corpora.textcorpus import TextCorpus
>>> from gensim.test.utils import datapath, get_tmpfile
>>> from gensim.similarities import Similarity
>>>
>>> corpus = TextCorpus(datapath('testcorpus.mm'))
>>> index_temp = get_tmpfile("index")
>>> index = Similarity(index_temp, corpus, num_features=400) # create index
>>>
>>> one_more_corpus = TextCorpus(datapath('testcorpus.txt'))
>>> index.add_documents(one_more_corpus) # add more documents in corpus
"""
min_ratio = 1.0 # 0.5 to only reopen shards that are <50% complete
if self.shards and len(self.shards[-1]) < min_ratio * self.shardsize:
# The last shard was incomplete (<; load it back and add the documents there, don't start a new shard
self.reopen_shard()
for doc in corpus:
if isinstance(doc, numpy.ndarray):
doclen = len(doc)
elif scipy.sparse.issparse(doc):
doclen = doc.nnz
else:
doclen = len(doc)
if doclen < 0.3 * self.num_features:
doc = matutils.unitvec(matutils.corpus2csc([doc], self.num_features).T, self.norm)
else:
doc = matutils.unitvec(matutils.sparse2full(doc, self.num_features), self.norm)
self.fresh_docs.append(doc)
self.fresh_nnz += doclen
if len(self.fresh_docs) >= self.shardsize:
self.close_shard()
if len(self.fresh_docs) % 10000 == 0:
logger.info("PROGRESS: fresh_shard size=%i", len(self.fresh_docs))
def get_document_topics(self, bow, minimum_probability=None,
normalize=None):
"""Get the topic distribution for the given document.
Parameters
----------
bow : list of (int, float)
The document in BOW format.
minimum_probability : float
If `normalize` is True, topics with smaller probabilities are filtered out.
If `normalize` is False, topics with smaller factors are filtered out.
If set to None, a value of 1e-8 is used to prevent 0s.
normalize: bool or None, optional
Whether to normalize the result. Allows for estimation of perplexity, coherence, e.t.c.
Returns
-------
list of (int, float)
Topic distribution for the whole document. Each element in the list is a pair of a topic's id, and
the probability that was assigned to it.
"""
if minimum_probability is None:
minimum_probability = self.minimum_probability
minimum_probability = max(minimum_probability, 1e-8)
# if the input vector is a corpus, return a transformed corpus
is_corpus, corpus = utils.is_corpus(bow)
if is_corpus:
kwargs = dict(minimum_probability=minimum_probability)
return self._apply(corpus, **kwargs)
v = matutils.corpus2csc([bow], self.num_tokens)
h = self._solveproj(v, self._W, v_max=np.inf)
if normalize is None:
normalize = self.normalize
if normalize:
the_sum = h.sum()
if the_sum:
h /= the_sum
return [
(idx, proba)
for idx, proba in enumerate(h[:, 0])
if not minimum_probability or proba > minimum_probability
]
def __init__(self, m, k, docs = None, algo = 'onepass', chunks = None):
"""
Construct the (U, S) projection from a corpus `docs`.
This is the class taking care of 'core math'; interfacing with corpora,
chunking large corpora etc. is done through the LsiModel class.
`algo` is currently one of:
* 'onepass'; only a single pass over `docs` is needed
* 'twopass'; multiple passes over the input allowed => can use a
faster algorithm.
"""
self.m, self.k = m, k
if docs is not None:
# base case decomposition: given a job `docs`, compute its decomposition in-core
# results of several base case decompositions can be merged via `self.merge()`
if algo == 'twopass':
self.u, self.s = stochasticSvd(docs, k, chunks = chunks, num_terms = m)
elif algo == 'onepass':
if not scipy.sparse.issparse(docs):
docs = matutils.corpus2csc(docs, num_terms = m)
if docs.shape[1] <= max(k, 100):
# For sufficiently small chunk size, update directly like `svd(now, docs)`
# instead of `svd(now, svd(docs))`.
# This improves accuracy and is also faster for small chunks, because
# we need to perform one less svd.
# On larger chunks this doesn't work because we quickly run out of memory.
self.u = docs
self.s = None
else:
try:
import sparsesvd
except ImportError:
raise ImportError("for LSA, the `sparsesvd` module is needed but not found; run `easy_install sparsesvd`")
logger.info("computing sparse SVD of %s matrix" % str(docs.shape))
ut, s, vt = sparsesvd.sparsesvd(docs, k + 30) # ask for extra factors, because for some reason SVDLIBC sometimes returns fewer factors than requested
u = ut.T
del ut, vt
k = clipSpectrum(s ** 2, self.k)
self.u, self.s = u[:, :k], s[:k]
else:
raise NotImplementedError("unknown decomposition algorithm: '%s'" % algo)
else:
self.u, self.s = None, None
