def train_model(self):
log_to_info('Loading training sentences')
review_d2v_id_list = zip(self.reviews['words'], self.reviews['d2v_id'], self.reviews['best_topics'],
self.reviews['second_best_topics'])
labeled_reviews = []
if self.dm == 0:
log_to_info('applying dbow with hpv={}'.format(self.mp.hierarchical_paragraph_vectors_dbow))
elif self.dm == 1:
log_to_info('applying dm with hpv={}'.format(self.mp.hierarchical_paragraph_vectors_dm))
for space_separated_words, d2v_id, best_topic, second_best_topic in review_d2v_id_list:
if self.dm == 0:
labeled_reviews.extend(
convert_to_labeled_review(self.mp.hierarchical_paragraph_vectors_dbow, space_separated_words, d2v_id, best_topic,
second_best_topic))
elif self.dm == 1:
labeled_reviews.extend(
convert_to_labeled_review(self.mp.hierarchical_paragraph_vectors_dm, space_separated_words, d2v_id, best_topic,
second_best_topic))
log_to_info('Loading Doc2Vec model...')
start_epoch = self.epochs + 1
model = None
for epoch in range(self.epochs, 0, -1):
model = self.load_model(epoch)
if model:
log_to_info('Found model in cache!')
break
start_epoch = epoch
if not model:
if self.dm == 0:
# PV-DBOW
log_to_info('Yep, this is DBOW!')
model = Doc2Vec(dm=self.dm, hs=0, workers=self.workers, size=self.mp.word_vector_dimensionality, min_count=self.min_count,
window=self.mp.word_context_window, sample=self.mp.frequent_words_downsampling_dbow, seed=random_int(),
negative=self.mp.negative)
# model = Doc2Vec(dm=0, size=100, negative=5, hs=0, min_count=2, workers=8)
elif self.dm == 1:
# PV-DM w/average
log_to_info('Yep, this is DM!')
model = Doc2Vec(dm=self.dm, dm_mean=1, hs=0, workers=self.workers, size=self.mp.word_vector_dimensionality,
min_count=self.min_count, window=self.mp.word_context_window, sample=self.mp.frequent_words_downsampling_dm,
seed=random_int(), negative=self.mp.negative)
# model = Doc2Vec(dm=1, dm_mean=1, size=100, window=10, negative=5, hs=0, min_count=2, workers=8)
start1 = time.time()
model.build_vocab(labeled_reviews)
start1 = time.time()
end1 = time.time()
log_to_info('Vocab building for dm{0} took {1} seconds'.format(self.dm, end1 - start1))
log_to_info('Training Doc2Vec model...')
for epoch in range(start_epoch, self.epochs + 1):
log_to_info('Epoch {0} of {1}'.format(epoch, self.epochs))
m = self.load_model(epoch)
if m is not None:
log_to_info('Found model in cache!')
model = m
continue
permuted_labeled_reviews = labeled_reviews[:]
random.shuffle(permuted_labeled_reviews)
alpha = alpha_for_epoch(epoch, self.mp.epochs_total, self.mp.alpha_max, self.mp.alpha_min, self.mp.learning_rate_type)
model.min_alpha, model.alpha = alpha, alpha
start2 = time.time()
model.train(permuted_labeled_reviews)
end2 = time.time()
log_to_info('DM HPV is {0}, DBOW HPV is {1}'.format(self.mp.hierarchical_paragraph_vectors_dm,
self.mp.hierarchical_paragraph_vectors_dbow))
log_to_info('Model training for dm{0} took {1} seconds'.format(self.dm, end2 - start2))
self.model = model
self.store_model(epoch)
self.model = model
def test_alpha_for_3_linear():
assert_similar(alpha_for_epoch(1, 3, 0.025, 0.001, 'linear'), .025)
assert_similar(alpha_for_epoch(2, 3, 0.025, 0.001, 'linear'), .013)
assert_similar(alpha_for_epoch(3, 3, 0.025, 0.001, 'linear'), .001)
def test_alpha_for_3_exp():
assert_similar(alpha_for_epoch(1, 3, 0.025, 0.001, 'exp'), .025)
assert_similar(alpha_for_epoch(2, 3, 0.025, 0.001, 'exp'), .005)
assert_similar(alpha_for_epoch(3, 3, 0.025, 0.001, 'exp'), .001)
def test_alpha_for_4_exp():
assert_similar(alpha_for_epoch(1, 4, 0.025, 0.001, 'exp'), .025)
assert_similar(alpha_for_epoch(2, 4, 0.025, 0.001, 'exp'), .00854988)
assert_similar(alpha_for_epoch(3, 4, 0.025, 0.001, 'exp'), .002924018)
assert_similar(alpha_for_epoch(4, 4, 0.025, 0.001, 'exp'), .001)
def test_alpha_for_4_linear():
assert_similar(alpha_for_epoch(1, 4, 0.025, 0.001, 'linear'), .025)
assert_similar(alpha_for_epoch(2, 4, 0.025, 0.001, 'linear'), .017)
assert_similar(alpha_for_epoch(3, 4, 0.025, 0.001, 'linear'), .009)
assert_similar(alpha_for_epoch(4, 4, 0.025, 0.001, 'linear'), .001)
