def transform_paragraph(self, paragraph, epochs=50, ignore_missing=False):
"""
Transform an iterable of tokens into its vector representation
(a paragraph vector).
Experimental. This will return something close to a tf-idf
weighted average of constituent token vectors by fitting
rare words (with low word bias values) more closely.
"""
if self.word_vectors is None:
raise Exception('Model must be fit to transform paragraphs')
if self.dictionary is None:
raise Exception('Dictionary must be provided to '
'transform paragraphs')
cooccurrence = collections.defaultdict(lambda: 0.0)
for token in paragraph:
try:
cooccurrence[self.dictionary[token]] += self.max_count / 10.0
except KeyError:
if not ignore_missing:
raise
random_state = check_random_state(self.random_state)
word_ids = np.array(cooccurrence.keys(), dtype=np.int32)
values = np.array(cooccurrence.values(), dtype=np.float64)
shuffle_indices = np.arange(len(word_ids), dtype=np.int32)
# Initialize the vector to mean of constituent word vectors
paragraph_vector = np.mean(self.word_vectors[word_ids], axis=0)
sum_gradients = np.ones_like(paragraph_vector)
# Shuffle the coocurrence matrix
random_state.shuffle(shuffle_indices)
transform_paragraph(self.word_vectors,
self.word_biases,
paragraph_vector,
sum_gradients,
word_ids,
values,
shuffle_indices,
self.learning_rate,
self.max_count,
self.alpha,
epochs)
return paragraph_vector
def transform_paragraph(self, paragraph, epochs=50, ignore_missing=False):
"""
Transform an iterable of tokens into its vector representation
(a paragraph vector).
Experimental. This will return something close to a tf-idf
weighted average of constituent token vectors by fitting
rare words (with low word bias values) more closely.
"""
if self.word_vectors is None:
raise Exception('Model must be fit to transform paragraphs')
if self.dictionary is None:
raise Exception('Dictionary must be provided to '
'transform paragraphs')
cooccurrence = collections.defaultdict(lambda: 0.0)
for token in paragraph:
try:
cooccurrence[self.dictionary[token]] += self.max_count / 10.0
except KeyError:
if not ignore_missing:
raise
word_ids = np.array(cooccurrence.keys(), dtype=np.int32)
values = np.array(cooccurrence.values(), dtype=np.float64)
shuffle_indices = np.arange(len(word_ids), dtype=np.int32)
# Initialize the vector to mean of constituent word vectors
paragraph_vector = np.mean(self.word_vectors[word_ids], axis=0)
# Shuffle the coocurrence matrix
np.random.shuffle(shuffle_indices)
transform_paragraph(self.word_vectors,
self.word_biases,
paragraph_vector,
word_ids,
values,
shuffle_indices,
self.learning_rate,
self.max_count,
self.alpha,
epochs)
return paragraph_vector
def transform_paragraph(self, paragraph, epochs=50, ignore_missing=False, use_pca=False):
"""
Transform an iterable of tokens into its vector representation
(a paragraph vector).
Experimental. This will return something close to a tf-idf
weighted average of constituent token vectors by fitting
rare words (with low word bias values) more closely. If use_pca is True,
the token vectors will be transformed using PCA before the weighted
average is calculated.
Added the PCA thing because of this paper: http://courses.cs.tau.ac.il/~wolf/papers/qagg.pdf
"""
if self.word_vectors is None:
raise Exception('Model must be fit to transform paragraphs')
if self.dictionary is None:
raise Exception('Dictionary must be provided to '
'transform paragraphs')
cooccurrence = collections.defaultdict(lambda: 0.0)
for token in paragraph:
try:
cooccurrence[self.dictionary[token]] += self.max_count / 10.0
except KeyError:
if not ignore_missing:
raise
word_ids = np.array(cooccurrence.keys(), dtype=np.int32)
values = np.array(cooccurrence.values(), dtype=np.float64)
shuffle_indices = np.arange(len(word_ids), dtype=np.int32)
# Initialize the vector to mean of constituent word vectors
# if a PCA model should be used, use it first to transform the vectors
if use_pca:
if self.pca is None:
self.perform_PCA()
paragraph_vector = np.mean(self.pca.transform(self.word_vectors[word_ids]), axis=0)
else:
paragraph_vector = np.mean(self.word_vectors[word_ids], axis=0)
sum_gradients = np.ones_like(paragraph_vector)
# Shuffle the coocurrence matrix
np.random.shuffle(shuffle_indices)
transform_paragraph(self.word_vectors,
self.word_biases,
paragraph_vector,
sum_gradients,
word_ids,
values,
shuffle_indices,
self.learning_rate,
self.max_count,
self.alpha,
epochs)
return paragraph_vector
def transform_paragraph(self,
paragraph,
epochs=50,
ignore_missing=False,
use_pca=False):
"""
Transform an iterable of tokens into its vector representation
(a paragraph vector).
Experimental. This will return something close to a tf-idf
weighted average of constituent token vectors by fitting
rare words (with low word bias values) more closely. If use_pca is True,
the token vectors will be transformed using PCA before the weighted
average is calculated.
Added the PCA thing because of this paper: http://courses.cs.tau.ac.il/~wolf/papers/qagg.pdf
"""
if self.word_vectors is None:
raise Exception('Model must be fit to transform paragraphs')
if self.dictionary is None:
raise Exception('Dictionary must be provided to '
'transform paragraphs')
cooccurrence = collections.defaultdict(lambda: 0.0)
for token in paragraph:
try:
cooccurrence[self.dictionary[token]] += self.max_count / 10.0
except KeyError:
if not ignore_missing:
raise
word_ids = np.array(cooccurrence.keys(), dtype=np.int32)
values = np.array(cooccurrence.values(), dtype=np.float64)
shuffle_indices = np.arange(len(word_ids), dtype=np.int32)
# Initialize the vector to mean of constituent word vectors
# if a PCA model should be used, use it first to transform the vectors
if use_pca:
if self.pca is None:
self.perform_PCA()
paragraph_vector = np.mean(self.pca.transform(
self.word_vectors[word_ids]),
axis=0)
else:
paragraph_vector = np.mean(self.word_vectors[word_ids], axis=0)
sum_gradients = np.ones_like(paragraph_vector)
# Shuffle the coocurrence matrix
np.random.shuffle(shuffle_indices)
transform_paragraph(self.word_vectors, self.word_biases,
paragraph_vector, sum_gradients, word_ids, values,
shuffle_indices, self.learning_rate,
self.max_count, self.alpha, epochs)
return paragraph_vector
