def init_hierarchical_model(class_ids):
score = [artm.PerplexityScore(name='perplexity_words', class_ids=['body']),
artm.PerplexityScore(name='perplexity_bigrams', class_ids=['bigrams'])]
top_tokens = [artm.TopTokensScore(name='top_words', num_tokens=15, class_id='body'),
artm.TopTokensScore(name='top_bigrams', num_tokens=10, class_id='bigrams')]
sparsity = [artm.SparsityThetaScore(name='sparsity_theta', eps=1e-6),
artm.SparsityPhiScore(name='sparsity_phi_words', class_id='words', eps=1e-6),
artm.SparsityPhiScore(name='sparsity_phi_bigrams', class_id='bigrams', eps=1e-6)]
regularizers = [artm.DecorrelatorPhiRegularizer(tau=0, class_ids=['body'], name='decorr_words'),
artm.DecorrelatorPhiRegularizer(tau=0, class_ids=['bigram'], name='decorr_bigrams'),
artm.DecorrelatorPhiRegularizer(tau=0, class_ids=['categories'], name='decorr_categories'),
artm.SmoothSparseThetaRegularizer(tau=0, name='sparsity_theta'),
artm.SmoothSparsePhiRegularizer(tau=0, class_ids=['body'], name='sparsity_words'),
artm.SmoothSparsePhiRegularizer(tau=0, class_ids=['bigram'], name='sparsity_bigrams')]
hmodel = artm.hARTM(class_ids=class_ids,
cache_theta=True,
reuse_theta=True,
scores=score + top_tokens + sparsity,
regularizers=regularizers,
theta_columns_naming='title')
return hmodel
def cluster_artm(text):
bach_vectorizer = artm.BatchVectorizer(data_path=text,
data_format='vowpal_wabbit', target_folder='batch_small',
batch_size=20)
T = 10 # количество тем
topic_names = ["sbj" + str(i) for i in range(T - 1)] + ["bcg"]
model_artm = artm.ARTM(num_topics=T, topic_names=topic_names, reuse_theta=True,
num_document_passes=1)
np.random.seed(1)
dictionary = artm.Dictionary()
dictionary.gather(data_path=bach_vectorizer.data_path)
model_artm.initialize(dictionary)
model_artm.scores.add(artm.TopTokensScore(name='metric1', num_tokens=15))
model_artm.regularizers.add(artm.SmoothSparsePhiRegularizer(name='smoothing', dictionary=dictionary,
topic_names='bcg', tau=1e5))
model_artm.fit_offline(batch_vectorizer=bach_vectorizer, num_collection_passes=6)
model_artm.regularizers.add(artm.SmoothSparsePhiRegularizer(name='stimulates',
dictionary=dictionary,
topic_names=["sbj" + str(i) for i in range(0, 29)],
tau=-1e5))
model_artm.fit_offline(batch_vectorizer=bach_vectorizer, num_collection_passes=6)
for topic_name in model_artm.topic_names:
with open('cluster_log_artm.txt', 'a') as f_in:
f_in.write(topic_name + ':')
for word in model_artm.score_tracker["metric1"].last_tokens[topic_name]:
f_in.write(word + ' ')
f_in.write('\n')
def generate_sparse_regularizers(
specific_topic_names, background_topic_names,
class_ids_for_bcg_smoothing=MAIN_MODALITY,
specific_words_classes=MAIN_MODALITY):
"""
Creates an array of pre-configured regularizers
using specified coefficients
"""
gimel_smooth_specific = 3e-10
gimel_smooth_bcg = 0.3
regularizers = [
artm.SmoothSparsePhiRegularizer(
tau=gimel_smooth_specific,
name='smooth_phi_specific',
topic_names=specific_topic_names,
class_ids=specific_words_classes
),
artm.SmoothSparseThetaRegularizer(
tau=gimel_smooth_specific,
name='smooth_theta_specific',
topic_names=specific_topic_names
),
artm.SmoothSparseThetaRegularizer(
tau=gimel_smooth_bcg,
name='smooth_theta_background',
topic_names=background_topic_names
),
artm.SmoothSparsePhiRegularizer(
tau=gimel_smooth_bcg,
name='smooth_phi_background',
topic_names=background_topic_names,
class_ids=class_ids_for_bcg_smoothing
),
]
return regularizers
def create_model_fn_20_complex_reg_1(n_iteration):
n_topics = 20
common_topics = [u'topic_0', u'topic_1']
subject_topics = list(
set([u'topic_{}'.format(idx)
for idx in range(2, 20)]) - set(common_topics))
tmp_model = create_model_complex(current_dictionary=dictionary,
n_topics=n_topics,
n_doc_passes=5,
seed_value=100 + n_iteration,
n_top_tokens=15,
p_mass_threshold=0.25,
common_topics=common_topics,
subject_topics=subject_topics)
# subject topics
tmp_model.regularizers.add(
artm.SmoothSparseThetaRegularizer(name='ss_theta_regularizer_subject',
topic_names=subject_topics))
tmp_model.regularizers.add(
artm.SmoothSparsePhiRegularizer(name='ss_phi_regularizer_subject',
topic_names=subject_topics,
class_ids=['@default_class']))
tmp_model.regularizers.add(
artm.DecorrelatorPhiRegularizer(
name='decorrelator_phi_regularizer_subject',
topic_names=subject_topics,
class_ids=['@default_class']))
tmp_model.regularizers['ss_theta_regularizer_subject'].tau = -0.5
tmp_model.regularizers['ss_phi_regularizer_subject'].tau = -0.5
tmp_model.regularizers['decorrelator_phi_regularizer_subject'].tau = -10
# common topics
tmp_model.regularizers.add(
artm.SmoothSparseThetaRegularizer(name='ss_theta_regularizer_common',
topic_names=subject_topics))
tmp_model.regularizers.add(
artm.SmoothSparsePhiRegularizer(name='ss_phi_regularizer_common',
topic_names=subject_topics,
class_ids=['@default_class']))
# tmp_model.regularizers.add(artm.DecorrelatorPhiRegularizer(name='decorrelator_phi_regularizer_common',
# topic_names=subject_topics, class_ids=['@default_class']))
tmp_model.regularizers['ss_theta_regularizer_common'].tau = 0.5
tmp_model.regularizers['ss_phi_regularizer_common'].tau = 0.5
# tmp_model.regularizers['decorrelator_phi_regularizer_common'].tau = -10
tmp_model = fit_one_model_complex(
plot_maker,
batch_vectorizer,
models_file,
config,
tmp_model,
_n_iterations=20,
_model_name='model_20_complex_reg_1_iter_{}'.format(n_iteration))
return tmp_model
def test_modifier_cube_on_two_steps_experiment(experiment_enviroment,
thread_flag):
""" """
tm, dataset, experiment, dictionary = experiment_enviroment
TAU_GRID_FIRST = (1, 5, 10)
TAU_GRID_SECOND = (-0.1, -0.5, -0.3)
regularizer_parameters = [{
"regularizer":
artm.SmoothSparsePhiRegularizer(name='test_first', tau=1),
"tau_grid":
TAU_GRID_FIRST
}, {
"regularizer":
artm.SmoothSparsePhiRegularizer(name='test_second', tau=1),
"tau_grid":
TAU_GRID_SECOND
}]
cube_pair = RegularizersModifierCube(
num_iter=10,
regularizer_parameters=regularizer_parameters,
reg_search="pair",
use_relative_coefficients=False,
separate_thread=thread_flag)
tmodels_lvl2_pair = cube_pair(tm, dataset)
regularizer_parameters_second = {
"name": 'test_second',
"tau_grid": [-0.2, -0.4, -0.6, -0.7]
}
cube_second = RegularizersModifierCube(
num_iter=10,
regularizer_parameters=regularizer_parameters_second,
reg_search="pair",
use_relative_coefficients=False,
separate_thread=thread_flag)
dummies = cube_second(tmodels_lvl2_pair[1], dataset)
tmodels_second = [dummy.restore() for dummy in dummies]
TAU_FOR_CHECKING = [(5, -0.2), (5, -0.4), (5, -0.6), (5, -0.7)]
assert len(tmodels_second) == 4
for i, one_model in enumerate(tmodels_second):
taus = (
one_model.regularizers['test_first'].tau,
one_model.regularizers['test_second'].tau,
)
assert taus == TAU_FOR_CHECKING[i]
def topic_model(class_ids, dictionary, num_of_topics, num_back, tau, tf):
names_of_topics = [str(x) for x in range(num_of_topics)]
dictionary.filter(min_tf=tf, class_id='subjects')
dictionary.filter(min_tf=tf, class_id='objects')
dictionary.filter(min_tf=tf, class_id='pairs')
model = artm.ARTM(
num_topics=num_of_topics,
#reuse_theta=True,
cache_theta=True,
topic_names=names_of_topics,
class_ids=class_ids,
#regularizers=regularizers_artm,
dictionary=dictionary)
model.scores.add(
artm.PerplexityScore(name='PerplexityScore', dictionary=dictionary))
model.scores.add(
artm.SparsityPhiScore(name='SparcityPhiScore',
topic_names=model.topic_names[:-num_back]))
model.regularizers.add(
artm.SmoothSparsePhiRegularizer(
name='SparsePhiRegularizer',
class_ids=class_ids,
topic_names=model.topic_names[:-num_back],
tau=-tau))
model.regularizers.add(
artm.SmoothSparsePhiRegularizer(
name='SmoothPhiRegularizer',
class_ids=class_ids,
topic_names=model.topic_names[-num_back:],
tau=tau))
model.regularizers.add(
artm.DecorrelatorPhiRegularizer(
name='DecorrelatorRegularizer',
class_ids=class_ids,
topic_names=model.topic_names[:-num_back],
tau=tau))
model.regularizers.add(
artm.SmoothSparseThetaRegularizer(
name='SparseThetaRegularizer',
topic_names=model.topic_names[-num_back],
tau=tau))
return model
def define_model(n_topics: int, dictionary: artm.Dictionary,
sparse_theta: float, sparse_phi: float,
decorrelator_phi: float) -> artm.artm_model.ARTM:
"""
Define the ARTM model.
:param n_topics: number of topics.
:param dictionary: batch vectorizer dictionary.
:param sparse_theta: sparse theta parameter.
:param sparse_phi: sparse phi Parameter.
:param decorrelator_phi: decorellator phi Parameter.
:return: ARTM model.
"""
print("Defining the model.")
topic_names = ["topic_{}".format(i) for i in range(1, n_topics + 1)]
model_artm = artm.ARTM(
topic_names=topic_names,
cache_theta=True,
scores=[
artm.PerplexityScore(name="PerplexityScore",
dictionary=dictionary),
artm.SparsityPhiScore(name="SparsityPhiScore"),
artm.SparsityThetaScore(name="SparsityThetaScore"),
artm.TopicKernelScore(name="TopicKernelScore",
probability_mass_threshold=0.3),
artm.TopTokensScore(name="TopTokensScore", num_tokens=15)
],
regularizers=[
artm.SmoothSparseThetaRegularizer(name="SparseTheta",
tau=sparse_theta),
artm.SmoothSparsePhiRegularizer(name="SparsePhi", tau=sparse_phi),
artm.DecorrelatorPhiRegularizer(name="DecorrelatorPhi",
tau=decorrelator_phi)
])
return model_artm
def create_model_fn_4(n_iteration):
tmp_model = cmh.create_model(current_dictionary=dictionary,
n_topics=100,
n_doc_passes=5,
seed_value=100 + n_iteration,
n_top_tokens=15,
p_mass_threshold=0.25)
tmp_model.regularizers.add(
artm.DecorrelatorPhiRegularizer(name='decorrelator_phi_regularizer',
class_ids=['@default_class']))
tmp_model.regularizers.add(
artm.SmoothSparseThetaRegularizer(name='ss_theta_regularizer'))
tmp_model.regularizers.add(
artm.SmoothSparsePhiRegularizer(name='ss_phi_regularizer',
class_ids=['@default_class']))
tmp_model.regularizers['decorrelator_phi_regularizer'].tau = 10
tmp_model.regularizers['ss_theta_regularizer'].tau = -0.5
tmp_model.regularizers['ss_phi_regularizer'].tau = -2
tmp_model = cmh.fit_one_model(
plot_maker,
batch_vectorizer,
models_file,
config,
tmp_model,
_n_iterations=20,
_model_name='model_20_m4_iter_{}'.format(n_iteration))
return tmp_model
def init_model(self, dictionary_path=None):
"""dictionary_path: optional, used with pretrained model"""
self.dictionary = artm.Dictionary()
if dictionary_path is None:
self.dictionary.gather(data_path=self.batches_path)
self.dictionary.filter(min_tf=10, max_df_rate=0.1)
self.dictionary.save_text(
f"{self.dir_path}/dicts/dict_{self.name_dataset}.txt")
else:
self.dictionary.load_text(dictionary_path)
self.model = artm.ARTM(
num_topics=self.n_topics,
dictionary=self.dictionary,
show_progress_bars=True,
)
# scores
self.model.scores.add(
artm.PerplexityScore(name="PerplexityScore",
dictionary=self.dictionary))
self.model.scores.add(
artm.SparsityThetaScore(name="SparsityThetaScore"))
self.model.scores.add(artm.SparsityPhiScore(name="SparsityPhiScore"))
# regularizers
self.model.regularizers.add(
artm.SmoothSparsePhiRegularizer(name="SparsePhi", tau=-0.1))
self.model.regularizers.add(
artm.SmoothSparseThetaRegularizer(name="SparseTheta", tau=-0.5))
self.model.regularizers.add(
artm.DecorrelatorPhiRegularizer(name="DecorrelatorPhi", tau=1.5e5))
def test_two_regularizers_on_step_experiment(experiment_enviroment,
thread_flag):
""" """
tm, dataset, experiment, dictionary = experiment_enviroment
TAU_GRID_FIRST = (1, 5, 10)
TAU_GRID_SECOND = (-0.1, -0.5)
regularizer_parameters = [{
"regularizer":
artm.SmoothSparsePhiRegularizer(name='test_first', tau=1),
"tau_grid":
TAU_GRID_FIRST
}, {
"regularizer":
artm.SmoothSparsePhiRegularizer(name='test_second', tau=1),
"tau_grid":
TAU_GRID_SECOND
}]
cube = RegularizersModifierCube(
num_iter=10,
regularizer_parameters=regularizer_parameters,
reg_search="grid",
use_relative_coefficients=False,
separate_thread=thread_flag)
dummies = cube(tm, dataset)
tmodels_lvl2 = [dummy.restore() for dummy in dummies]
TAU_FOR_CHECKING = [(1, -0.1), (1, -0.5), (5, -0.1), (5, -0.5), (10, -0.1),
(10, -0.5)]
assert len(tmodels_lvl2) == len(TAU_FOR_CHECKING)
for i, one_model in enumerate(tmodels_lvl2):
taus = (
one_model.regularizers['test_first'].tau,
one_model.regularizers['test_second'].tau,
)
assert taus == TAU_FOR_CHECKING[i]
for lvl2_model in tmodels_lvl2:
assert lvl2_model.parent_model_id == tm.model_id
assert lvl2_model.depth == 2
def create_model_with_background(dictionary, num_tokens, num_document_passes):
sm_phi_tau = 0.0001 * 1e-4
sp_phi_tau = -0.0001 * 1e-4
decor_phi_tau = 1
specific_topics = ['topic {}'.format(i) for i in range(1, 20)]
topic_names = specific_topics + ["background"]
scores = [
artm.PerplexityScore(name='PerplexityScore', dictionary=dictionary),
artm.TopTokensScore(
name='TopTokensScore', num_tokens=10, class_id='plain_text'
), # web version of Palmetto works only with <= 10 tokens
artm.SparsityPhiScore(name='SparsityPhiScore'),
artm.SparsityThetaScore(name='SparsityThetaScore'),
artm.TopicKernelScore(name='TopicKernelScore',
probability_mass_threshold=0.3,
class_id='plain_text')
]
model = artm.ARTM(topic_names=specific_topics + ["background"],
regularizers=[],
cache_theta=True,
scores=scores,
class_ids={'plain_text': 1.0})
model.regularizers.add(
artm.SmoothSparsePhiRegularizer(name='SparsePhi',
tau=-sp_phi_tau,
topic_names=specific_topics))
model.regularizers.add(
artm.SmoothSparsePhiRegularizer(name='SmoothPhi',
tau=sm_phi_tau,
topic_names=["background"]))
# model.regularizers.add(artm.DecorrelatorPhiRegularizer(name='DecorrelatorPhi', tau=decor_phi_tau))
model.initialize(dictionary=dictionary)
model.num_document_passes = num_document_passes
return model
def create_topic_model(self, topic_model_name: str,
batch_vectorizer: artm.BatchVectorizer,
dictionary: artm.Dictionary) -> artm.ARTM:
topic_model = artm.ARTM(num_topics=self.number_of_topics,
dictionary=dictionary,
cache_theta=False)
topic_model.scores.add(
artm.PerplexityScore(name='perplexity_score',
dictionary=dictionary))
topic_model.scores.add(
artm.SparsityPhiScore(name='sparsity_phi_score'))
topic_model.scores.add(
artm.SparsityThetaScore(name='sparsity_theta_score'))
topic_model.num_document_passes = 5
topic_model.num_processors = max(1, os.cpu_count() - 1)
topic_model.regularizers.add(
artm.SmoothSparsePhiRegularizer(name='sparse_phi_regularizer'))
topic_model.regularizers.add(
artm.SmoothSparseThetaRegularizer(name='sparse_theta_regularizer'))
topic_model.regularizers.add(
artm.DecorrelatorPhiRegularizer(
name='decorrelator_phi_regularizer'))
topic_model.regularizers['sparse_phi_regularizer'].tau = -1.0
topic_model.regularizers['sparse_theta_regularizer'].tau = -0.5
topic_model.regularizers['decorrelator_phi_regularizer'].tau = 1e+5
best_score = None
keyword_extraction_logger.info(
'epoch perplexity_score sparsity_phi_score sparsity_theta_score'
)
for restart_index in range(10):
topic_model.fit_offline(batch_vectorizer=batch_vectorizer,
num_collection_passes=3)
if best_score is None:
best_score = topic_model.score_tracker[
'perplexity_score'].last_value
else:
if best_score > topic_model.score_tracker[
'perplexity_score'].last_value:
best_score = topic_model.score_tracker[
'perplexity_score'].last_value
self.save_topic_model(topic_model, topic_model_name)
keyword_extraction_logger.info(
'{0:5} {1:16.9} {2:18.9} {3:20.9}'.format(
(restart_index + 1) * 3,
topic_model.score_tracker['perplexity_score'].last_value,
topic_model.score_tracker['sparsity_phi_score'].last_value,
topic_model.score_tracker['sparsity_theta_score'].
last_value))
del topic_model
return self.load_topic_model(
artm.ARTM(num_topics=self.number_of_topics,
dictionary=dictionary,
cache_theta=False), topic_model_name)
def _get_corpus_model(self,
corpus_vector_spaced,
clustering_method='artm'):
if 'gensim' == clustering_method:
return self._get_model_LSI(corpus_vector_spaced)
elif 'sklearn' == clustering_method:
return self._get_model_LDA(corpus_vector_spaced)
elif 'artm' == clustering_method:
batch_vectorizer = corpus_vector_spaced['batch_vectorizer']
dictionary = corpus_vector_spaced['dictionary']
topic_names = [
'topic_{}'.format(i) for i in range(self.num_of_clusters)
]
model_artm = artm.ARTM(
topic_names=topic_names,
cache_theta=True,
scores=[
artm.PerplexityScore(name='PerplexityScore',
dictionary=dictionary)
],
regularizers=[
artm.SmoothSparseThetaRegularizer(name='SparseTheta',
tau=-0.15)
])
model_artm.scores.add(
artm.SparsityPhiScore(name='SparsityPhiScore'))
model_artm.scores.add(
artm.SparsityThetaScore(name='SparsityThetaScore'))
model_artm.scores.add(
artm.TopicKernelScore(name='TopicKernelScore',
probability_mass_threshold=0.3))
model_artm.scores.add(artm.TopTokensScore(name='TopTokensScore',
num_tokens=10),
overwrite=True)
model_artm.regularizers.add(
artm.SmoothSparsePhiRegularizer(name='SparsePhi', tau=-0.1))
model_artm.regularizers['SparseTheta'].tau = -0.2
model_artm.regularizers.add(
artm.DecorrelatorPhiRegularizer(name='DecorrelatorPhi',
tau=1.5e+5))
model_artm.num_document_passes = 1
model_artm.initialize(dictionary)
model_artm.fit_offline(batch_vectorizer=batch_vectorizer,
num_collection_passes=30)
return model_artm.transform(batch_vectorizer=batch_vectorizer).T
def experiment(filename, tau_phi, tau_theta):
batch_vectorizer = artm.BatchVectorizer(data_path=filename, data_format='vowpal_wabbit',
target_folder='batches')
dictionary = batch_vectorizer.dictionary
topic_num = 30
tokens_num = 100
print("ARTM training")
topic_names = ['topic_{}'.format(i) for i in range(topic_num)]
model_artm = artm.ARTM(topic_names=topic_names, dictionary=dictionary, cache_theta=True)
model_plsa = artm.ARTM(topic_names=topic_names, cache_theta=True,
scores=[artm.PerplexityScore(name='PerplexityScore', dictionary=dictionary)])
model_lda = artm.LDA(num_topics=topic_num)
model_artm.scores.add(artm.PerplexityScore(name='perplexity_score', dictionary=dictionary))
model_artm.scores.add(artm.SparsityPhiScore(name='sparsity_phi_score'))
model_artm.scores.add(artm.SparsityThetaScore(name='sparsity_theta_score'))
model_artm.scores.add(artm.TopTokensScore(name='top_tokens_score', num_tokens=tokens_num))
model_artm.scores.add(artm.TopicKernelScore(name='topic_kernel_score', probability_mass_threshold=0.3))
model_artm.scores.add(artm.BackgroundTokensRatioScore(name='background_tokens_ratio_score'))
model_artm.scores.add(artm.ClassPrecisionScore(name='class_precision_score'))
model_artm.scores.add(artm.TopicMassPhiScore(name='topic_mass_phi_score'))
model_plsa.scores.add(artm.PerplexityScore(name='perplexity_score', dictionary=dictionary))
model_plsa.scores.add(artm.SparsityPhiScore(name='sparsity_phi_score'))
model_plsa.scores.add(artm.SparsityThetaScore(name='sparsity_theta_score'))
model_plsa.scores.add(artm.TopTokensScore(name='top_tokens_score'))
model_plsa.scores.add(artm.TopicKernelScore(name='topic_kernel_score', probability_mass_threshold=0.3))
model_plsa.scores.add(artm.BackgroundTokensRatioScore(name='background_tokens_ratio_score'))
model_plsa.scores.add(artm.ClassPrecisionScore(name='class_precision_score'))
model_plsa.scores.add(artm.TopicMassPhiScore(name='topic_mass_phi_score'))
model_artm.regularizers.add(artm.SmoothSparsePhiRegularizer(name='sparse_phi_regularizer'))
model_artm.regularizers.add(artm.SmoothSparseThetaRegularizer(name='sparse_theta_regularizer'))
model_artm.regularizers.add(artm.DecorrelatorPhiRegularizer(name='decorrelator_phi_regularizer'))
model_artm.regularizers['sparse_phi_regularizer'].tau = tau_phi
model_artm.regularizers['sparse_theta_regularizer'].tau = tau_theta
model_artm.regularizers['decorrelator_phi_regularizer'].tau = 1e+3
model_plsa.initialize(dictionary=dictionary)
model_artm.initialize(dictionary=dictionary)
model_lda.initialize(dictionary=dictionary)
passes = 100
model_plsa.fit_offline(batch_vectorizer=batch_vectorizer, num_collection_passes=passes)
model_artm.fit_offline(batch_vectorizer=batch_vectorizer, num_collection_passes=passes)
model_lda.fit_offline(batch_vectorizer=batch_vectorizer, num_collection_passes=passes)
print_measures(model_plsa, model_artm, model_lda)
def test_two_regularizers_on_step_experiment_pair_grid(experiment_enviroment):
""" """
tm, dataset, experiment, dictionary = experiment_enviroment
TAU_GRID_FIRST = (1, 5, 10)
TAU_GRID_SECOND = (-0.1, -0.5, -0.3)
regularizer_parameters = [{
"regularizer":
artm.SmoothSparsePhiRegularizer(name='test_first', tau=1),
"tau_grid":
TAU_GRID_FIRST
}, {
"regularizer":
artm.SmoothSparsePhiRegularizer(name='test_second', tau=1),
"tau_grid":
TAU_GRID_SECOND
}]
cube_pair = RegularizersModifierCube(
num_iter=10,
regularizer_parameters=regularizer_parameters,
reg_search="pair",
relative_coefficients=False,
)
tmodels_lvl2_pair = cube_pair(tm, dataset)
TAU_FOR_CHECKING = [(1, -0.1), (5, -0.5), (10, -0.3)]
assert len(tmodels_lvl2_pair) == 3
for i, one_model in enumerate(tmodels_lvl2_pair):
taus = (
one_model.regularizers['test_first'].tau,
one_model.regularizers['test_second'].tau,
)
assert taus == TAU_FOR_CHECKING[i]
def create_thematic_model(checked_list, num_topics, num_tokens, phi_tau,
theta_tau, decorr_tau):
""" Create a thematic model """
gluing_bag_of_words(checked_list)
batch_vectorizer = artm.BatchVectorizer(data_path=COLLECTION_PATH,
data_format='vowpal_wabbit',
target_folder=TARGET_FOLDER,
batch_size=len(checked_list))
dictionary = artm.Dictionary(data_path=TARGET_FOLDER)
model = artm.ARTM(
num_topics=num_topics,
num_document_passes=len(checked_list),
dictionary=dictionary,
regularizers=[
artm.SmoothSparsePhiRegularizer(name='sparse_phi_regularizer',
tau=phi_tau),
artm.SmoothSparseThetaRegularizer(name='sparse_theta_regularizer',
tau=theta_tau),
artm.DecorrelatorPhiRegularizer(
name='decorrelator_phi_regularizer', tau=decorr_tau),
],
scores=[
artm.PerplexityScore(name='perplexity_score',
dictionary=dictionary),
artm.SparsityPhiScore(name='sparsity_phi_score'),
artm.SparsityThetaScore(name='sparsity_theta_score'),
artm.TopTokensScore(name='top_tokens_score', num_tokens=num_tokens)
])
model.fit_offline(batch_vectorizer=batch_vectorizer,
num_collection_passes=len(checked_list))
top_tokens = model.score_tracker['top_tokens_score']
topic_dictionary = OrderedDict()
for topic_name in model.topic_names:
list_name = []
for (token, weight) in zip(top_tokens.last_tokens[topic_name],
top_tokens.last_weights[topic_name]):
list_name.append(token + '-' + str(round(weight, 3)))
topic_dictionary[str(topic_name)] = list_name
return model.score_tracker[
'perplexity_score'].last_value, model.score_tracker[
'sparsity_phi_score'].last_value, model.score_tracker[
'sparsity_theta_score'].last_value, topic_dictionary
def create_model_fn_20_reg_1(n_iteration):
tmp_model = create_model(current_dictionary=dictionary,
n_topics=20,
n_doc_passes=5,
seed_value=100 + n_iteration,
n_top_tokens=15,
p_mass_threshold=0.25)
tmp_model.regularizers.add(
artm.SmoothSparseThetaRegularizer(name='ss_theta_regularizer'))
tmp_model.regularizers.add(
artm.SmoothSparsePhiRegularizer(name='ss_phi_regularizer',
class_ids=['@default_class']))
tmp_model.regularizers['ss_theta_regularizer'].tau = -3
tmp_model.regularizers['ss_phi_regularizer'].tau = -3
tmp_model = fit_one_model(
tmp_model,
_n_iterations=20,
_model_name='model_20_reg_1_iter_{}'.format(n_iteration))
return tmp_model
def pipeline_plsa_bigartm(lines,
TOPIC_NUMBER,
ngram_range,
topnwords,
LOGS_DATA_PATH="plsa.txt",
TARGET_FOLDER="plsa"):
make_file(lines, ngram_range, LOGS_DATA_PATH)
batch_vectorizer = artm.BatchVectorizer(data_path=LOGS_DATA_PATH,
data_format='vowpal_wabbit',
target_folder=TARGET_FOLDER)
model_artm = artm.ARTM(num_topics=TOPIC_NUMBER, cache_theta=True)
model_artm.initialize(dictionary=batch_vectorizer.dictionary)
model_artm.regularizers.add(
artm.SmoothSparsePhiRegularizer(name='SparsePhi', tau=0.05))
model_artm.regularizers.add(
artm.DecorrelatorPhiRegularizer(name='DecorrelatorPhi', tau=1.5e+5))
model_artm.regularizers.add(
artm.SmoothSparseThetaRegularizer(name='SparseTheta', tau=-0.01))
model_artm.scores.add(artm.SparsityPhiScore(name='SparsityPhiScore'))
model_artm.scores.add(artm.SparsityThetaScore(name='SparsityThetaScore'))
model_artm.scores.add(artm.TopTokensScore(name='TopTokensScore',
num_tokens=topnwords),
overwrite=True)
model_artm.scores.add(
artm.PerplexityScore(name='PerplexityScore',
dictionary=batch_vectorizer.dictionary))
model_artm.num_document_passes = 2
model_artm.fit_offline(batch_vectorizer=batch_vectorizer,
num_collection_passes=15)
topic_names = {}
for topic_name in model_artm.topic_names:
topic_names[topic_name] = model_artm.score_tracker[
'TopTokensScore'].last_tokens[topic_name]
#return label_after_bigarm(model_artm), topic_names
return "nothing, sorry", topic_names
def build_model(self, d_dir, n_document_passes=1):
batch_vectorizer_train = artm.BatchVectorizer(data_path=os.path.join(
d_dir, 'data_batches_train'),
data_format="batches")
batch_vectorizer_test = artm.BatchVectorizer(data_path=os.path.join(
d_dir, 'data_batches_test'),
data_format="batches")
dictionary = artm.Dictionary()
dictionary.gather(data_path=os.path.join(d_dir, 'for_dict'))
model = artm.ARTM(num_topics=self.n_topics,
dictionary=dictionary,
cache_theta=True,
reuse_theta=True)
# Sparsity p(c|t)
model.scores.add(
artm.SparsityPhiScore(eps=EPS,
name='SparsityPhiScoreC',
class_id=self.c))
# Sparsity p(w|t)
model.scores.add(
artm.SparsityPhiScore(eps=EPS,
name='SparsityPhiScoreGram3',
class_id=self.gram3))
#Regularization of sparsity p(gram3|t)
model.regularizers.add(
artm.SmoothSparsePhiRegularizer(name='SparsePhiGram3Regularizer',
class_ids=[self.gram3]))
#Regularization of decorr p(gram3|t)
model.regularizers.add(
artm.DecorrelatorPhiRegularizer(
name='DecorrelatorPhiGram3Regularizer',
class_ids=[self.gram3]))
model.num_document_passes = n_document_passes
return (model, batch_vectorizer_train, batch_vectorizer_test)
def test_func():
data_path = os.environ.get('BIGARTM_UNITTEST_DATA')
batches_folder = tempfile.mkdtemp()
dump_folder = tempfile.mkdtemp()
try:
batch_vectorizer = artm.BatchVectorizer(data_path=data_path,
data_format='bow_uci',
collection_name='kos',
target_folder=batches_folder)
model_1 = artm.ARTM(num_processors=7,
cache_theta=True,
num_document_passes=5,
reuse_theta=True,
seed=10,
num_topics=15,
class_ids={'@default_class': 1.0},
theta_name='THETA',
dictionary=batch_vectorizer.dictionary)
model_2 = artm.ARTM(num_processors=7,
cache_theta=False,
num_document_passes=5,
reuse_theta=False,
seed=10,
num_topics=15,
class_ids={'@default_class': 1.0},
dictionary=batch_vectorizer.dictionary)
for model in [model_1, model_2]:
model.scores.add(
artm.PerplexityScore(name='perp',
dictionary=batch_vectorizer.dictionary))
model.scores.add(artm.SparsityThetaScore(name='sp_theta', eps=0.1))
model.scores.add(artm.TopTokensScore(name='top_tok',
num_tokens=10))
model.scores.add(
artm.SparsityPhiScore(name='sp_nwt',
model_name=model.model_nwt))
model.scores.add(
artm.TopicKernelScore(name='kernel',
topic_names=model.topic_names[0:5],
probability_mass_threshold=0.4))
topic_pairs = {}
for topic_name_1 in model.topic_names:
for topic_name_2 in model.topic_names:
if topic_name_1 not in topic_pairs:
topic_pairs[topic_name_1] = {}
topic_pairs[topic_name_1][
topic_name_2] = numpy.random.randint(0, 3)
model.regularizers.add(
artm.DecorrelatorPhiRegularizer(name='decor',
tau=100000.0,
topic_pairs=topic_pairs))
model.regularizers.add(
artm.SmoothSparsePhiRegularizer(
name='smsp_phi',
tau=-0.5,
gamma=0.3,
dictionary=batch_vectorizer.dictionary))
model.regularizers.add(
artm.SmoothSparseThetaRegularizer(name='smsp_theta',
tau=0.1,
doc_topic_coef=[2.0] *
model.num_topics))
model.regularizers.add(
artm.SmoothPtdwRegularizer(name='sm_ptdw', tau=0.1))
# learn first model and dump it on disc
model.fit_offline(batch_vectorizer, num_collection_passes=10)
model.fit_online(batch_vectorizer, update_every=1)
model.dump_artm_model(os.path.join(dump_folder, 'target'))
params = {}
with open(os.path.join(dump_folder, 'target', 'parameters.json'),
'r') as fin:
params = json.load(fin)
_assert_json_params(params)
# create second model from the dump and check the results are equal
model_new = artm.load_artm_model(
os.path.join(dump_folder, 'target'))
_assert_params_equality(model, model_new)
_assert_scores_equality(model, model_new)
_assert_regularizers_equality(model, model_new)
_assert_score_values_equality(model, model_new)
_assert_matrices_equality(model, model_new)
# continue learning of both models
model.fit_offline(batch_vectorizer, num_collection_passes=3)
model.fit_online(batch_vectorizer, update_every=1)
model_new.fit_offline(batch_vectorizer, num_collection_passes=3)
model_new.fit_online(batch_vectorizer, update_every=1)
# check new results are also equal
_assert_params_equality(model, model_new)
_assert_scores_equality(model, model_new)
_assert_regularizers_equality(model, model_new)
_assert_score_values_equality(model, model_new)
_assert_matrices_equality(model, model_new)
shutil.rmtree(os.path.join(dump_folder, 'target'))
finally:
shutil.rmtree(batches_folder)
shutil.rmtree(dump_folder)
model_artm.scores.add(artm.TopTokensScore(name='top_tokens_score', num_tokens=tokens_num)) model_artm.scores.add(artm.TopicKernelScore(name='topic_kernel_score', probability_mass_threshold=0.3)) model_artm.scores.add(artm.BackgroundTokensRatioScore(name='background_tokens_ratio_score')) model_artm.scores.add(artm.ClassPrecisionScore(name='class_precision_score')) model_artm.scores.add(artm.TopicMassPhiScore(name='topic_mass_phi_score')) model_plsa.scores.add(artm.PerplexityScore(name='perplexity_score', dictionary=dictionary)) model_plsa.scores.add(artm.SparsityPhiScore(name='sparsity_phi_score')) model_plsa.scores.add(artm.SparsityThetaScore(name='sparsity_theta_score')) model_plsa.scores.add(artm.TopTokensScore(name='top_tokens_score')) model_plsa.scores.add(artm.TopicKernelScore(name='topic_kernel_score', probability_mass_threshold=0.3)) model_plsa.scores.add(artm.BackgroundTokensRatioScore(name='background_tokens_ratio_score')) model_plsa.scores.add(artm.ClassPrecisionScore(name='class_precision_score')) model_plsa.scores.add(artm.TopicMassPhiScore(name='topic_mass_phi_score')) model_artm.regularizers.add(artm.SmoothSparsePhiRegularizer(name='sparse_phi_regularizer')) model_artm.regularizers.add(artm.SmoothSparseThetaRegularizer(name='sparse_theta_regularizer')) model_artm.regularizers.add(artm.DecorrelatorPhiRegularizer(name='decorrelator_phi_regularizer')) model_artm.regularizers['sparse_phi_regularizer'].tau = 0.01 model_artm.regularizers['sparse_theta_regularizer'].tau = -1.06 # model_artm.regularizers['decorrelator_phi_regularizer'].tau = 1e+5 model_plsa.initialize(dictionary=dictionary) model_artm.initialize(dictionary=dictionary) model_lda.initialize(dictionary=dictionary) passes = 10 model_plsa.fit_offline(batch_vectorizer=batch_vectorizer, num_collection_passes=passes) model_artm.fit_offline(batch_vectorizer=batch_vectorizer, num_collection_passes=passes) model_lda.fit_offline(batch_vectorizer=batch_vectorizer, num_collection_passes=passes)
def train(self):
vocabulary_file = self._prepare_texts_full(
) if self.analyze_full_doc == True else self._prepare_texts_from_summary(
)
target_folder = self._get_bigARTM_dir()
batch_vectorizer = artm.BatchVectorizer(data_path=vocabulary_file,
data_format='vowpal_wabbit',
target_folder=target_folder,
batch_size=100)
dict_path = self._get_dictionary_path()
dict_file = '{}.dict'.format(dict_path)
if os.path.isfile(dict_file):
os.remove(dict_file)
my_dictionary = artm.Dictionary()
my_dictionary.gather(data_path=target_folder,
vocab_file_path=vocabulary_file)
my_dictionary.save(dictionary_path=dict_path)
my_dictionary.load(dictionary_path=dict_file)
T = self.num_of_topics
topic_names = ["sbj" + str(i) for i in range(T - 1)] + ["bcg"]
self.model_artm = artm.ARTM(num_topics=T,
topic_names=topic_names,
class_ids={
"text": 1,
"doc_guid": 1
},
dictionary=my_dictionary,
cache_theta=True)
self.model_artm.initialize(dictionary=my_dictionary)
self.model_artm.scores.add(
artm.TopTokensScore(name="text_words",
num_tokens=15,
class_id="text"))
self.model_artm.scores.add(
artm.TopTokensScore(name="doc_guid_words",
num_tokens=15,
class_id="doc_guid"))
self.model_artm.regularizers.add(
artm.SmoothSparsePhiRegularizer(name='SparsePhi',
tau=1e5,
dictionary=my_dictionary,
class_ids="text",
topic_names="bcg"))
self.model_artm.fit_offline(batch_vectorizer=batch_vectorizer,
num_collection_passes=30)
self.model_artm.regularizers.add(
artm.SmoothSparsePhiRegularizer(
name='SparsePhi-1e5',
tau=-1e5,
dictionary=my_dictionary,
class_ids="text",
topic_names=["sbj" + str(i) for i in range(T - 1)]))
self.model_artm.fit_offline(batch_vectorizer=batch_vectorizer,
num_collection_passes=15)
self.training_done = True
dictionary.gather(data_path=batch_vectorizer.data_path)
dictionary.save(dictionary_path=filename + '/dictionary.dict')
dictionary.load(dictionary_path=(filename + '/dictionary.dict'))
dictionary.load(dictionary_path=(filename + '/dictionary.dict'))
model_artm.initialize(dictionary=dictionary)
model_artm.scores.add(artm.SparsityPhiScore(name='SparsityPhiScore'))
model_artm.scores.add(artm.SparsityThetaScore(name='SparsityThetaScore'))
model_artm.scores.add(
artm.TopicKernelScore(name='TopicKernelScore',
probability_mass_threshold=0.3))
model_artm.regularizers.add(
artm.SmoothSparsePhiRegularizer(name='SparsePhi', tau=-0.1))
model_artm.regularizers.add(
artm.DecorrelatorPhiRegularizer(name='DecorrelatorPhi', tau=1.5e+5))
model_artm.regularizers.add(
artm.TopicSelectionThetaRegularizer(name='TopicSelection', tau=0.25))
model_artm.regularizers['SparsePhi'].tau = -0.5
model_artm.regularizers['SparseTheta'].tau = -0.5
model_artm.regularizers['DecorrelatorPhi'].tau = 1e+5
model_artm.scores.add(artm.TopTokensScore(name='TopTokensScore',
num_tokens=10))
model_artm.fit_offline(batch_vectorizer=batch_vectorizer,
num_collection_passes=40)
model.regularizers.add(
artm.DecorrelatorPhiRegularizer(name='Decorrelator',
tau=10**4)) # обычный декоррелятор
model.regularizers.add(
artm.SmoothSparseThetaRegularizer(
name='SmoothTheta', topic_names=bgr_topics,
tau=0.3)) # сглаживаем Theta для фоновых тем
model.regularizers.add(
artm.SmoothSparseThetaRegularizer(
name='SparseTheta', topic_names=subj_topics,
tau=-0.3)) # разреживаем Theta для "хороших" тем
model.regularizers.add(
artm.SmoothSparsePhiRegularizer(
name='SmoothPhi',
topic_names=bgr_topics,
class_ids=["text"],
tau=0.1)) # сглаживаем Theta для фоновых тем
model.regularizers.add(
artm.SmoothSparsePhiRegularizer(
name='SparsePhi',
topic_names=subj_topics,
class_ids=["text"],
tau=-0.1)) # разреживаем Theta для "хороших" тем
# model.regularizers.add(artm.LabelRegularizationPhiRegularizer(class_ids=["label"])) # этот регуляризатор мало у кого дает
# # хороший результат, но ты попробуй :) у меня он вылетает с ошибкой :(
model.class_ids = {
"title": 1,
"text": 1,
"label": 5,
def init_simple_default_model(
dataset: Dataset,
modalities_to_use: List[str] or Dict[str, float],
main_modality: str,
specific_topics: List[str] or int,
background_topics: List[str] or int,
) -> artm.ARTM:
"""
Creates simple `artm.ARTM` model with standard scores.
Parameters
----------
dataset
Dataset for model initialization
modalities_to_use
What modalities a model should know.
If `modalities_to_use` is a dictionary,
all given weights are assumed to be relative to `main_modality`:
weights will then be recalculated to absolute ones
using `dataset` and `main_modality`.
If `modalities_to_use` is a list,
then all relative weights are set equal to one.
The result model's `class_ids` field will contain absolute modality weights.
main_modality
Modality relative to which all modality weights are considered
specific_topics
Specific topic names or their number
background_topics
Background topic names or their number
Returns
-------
model : artm.ARTM
"""
if isinstance(modalities_to_use, dict):
modalities_weights = modalities_to_use
else:
modalities_weights = {class_id: 1 for class_id in modalities_to_use}
specific_topic_names, background_topic_names = create_default_topics(
specific_topics, background_topics)
dictionary = dataset.get_dictionary()
tokens_data = count_vocab_size(dictionary, modalities_to_use)
abs_weights = modality_weight_rel2abs(tokens_data, modalities_weights,
main_modality)
model = init_model(
topic_names=specific_topic_names + background_topic_names,
class_ids=abs_weights,
)
if len(background_topic_names) > 0:
model.regularizers.add(
artm.SmoothSparsePhiRegularizer(
name='smooth_phi_bcg',
topic_names=background_topic_names,
tau=0.0,
class_ids=[main_modality],
), )
model.regularizers.add(
artm.SmoothSparseThetaRegularizer(
name='smooth_theta_bcg',
topic_names=background_topic_names,
tau=0.0,
), )
model.initialize(dictionary)
add_standard_scores(model,
main_modality=main_modality,
all_modalities=modalities_to_use)
return model
def init_simple_default_model(
dataset,
modalities_to_use,
main_modality,
specific_topics,
background_topics,
):
"""
Creates simple artm model with standard scores.
Parameters
----------
dataset : Dataset
modalities_to_use : list of str
main_modality : str
specific_topics : list or int
background_topics : list or int
Returns
-------
model: artm.ARTM() instance
"""
if isinstance(specific_topics, list):
specific_topic_names = list(specific_topics)
else:
specific_topics = int(specific_topics)
specific_topic_names = [f'topic_{i}' for i in range(specific_topics)]
n_specific_topics = len(specific_topic_names)
if isinstance(background_topics, list):
background_topic_names = list(background_topics)
else:
background_topics = int(background_topics)
background_topic_names = [
f'background_{n_specific_topics + i}'
for i in range(background_topics)
]
n_background_topics = len(background_topic_names)
dictionary = dataset.get_dictionary()
baseline_class_ids = {class_id: 1 for class_id in modalities_to_use}
tokens_data = count_vocab_size(dictionary, modalities_to_use)
abs_weights = modality_weight_rel2abs(tokens_data, baseline_class_ids,
main_modality)
model = init_model(
topic_names=specific_topic_names + background_topic_names,
class_ids=abs_weights,
)
if n_background_topics > 0:
model.regularizers.add(
artm.SmoothSparsePhiRegularizer(
name='smooth_phi_bcg',
topic_names=background_topic_names,
tau=0.0,
class_ids=[main_modality],
), )
model.regularizers.add(
artm.SmoothSparseThetaRegularizer(
name='smooth_theta_bcg',
topic_names=background_topic_names,
tau=0.0,
), )
model.initialize(dictionary)
add_standard_scores(model,
dictionary,
main_modality=main_modality,
all_modalities=modalities_to_use)
return model
def test_func():
# constants
num_tokens = 15
alpha = 0.01
beta = 0.02
num_collection_passes = 15
num_document_passes = 1
num_topics = 15
vocab_size = 6906
num_docs = 3430
zero_eps = 0.001
data_path = os.environ.get('BIGARTM_UNITTEST_DATA')
batches_folder = tempfile.mkdtemp()
try:
batch_vectorizer = artm.BatchVectorizer(data_path=data_path,
data_format='bow_uci',
collection_name='kos',
target_folder=batches_folder)
dictionary = artm.Dictionary()
dictionary.gather(data_path=batch_vectorizer.data_path)
model_artm = artm.ARTM(num_topics=num_topics,
dictionary=dictionary,
cache_theta=True,
reuse_theta=True)
model_artm.regularizers.add(
artm.SmoothSparsePhiRegularizer(name='SparsePhi', tau=beta))
model_artm.regularizers.add(
artm.SmoothSparseThetaRegularizer(name='SparseTheta', tau=alpha))
model_artm.scores.add(
artm.SparsityThetaScore(name='SparsityThetaScore'))
model_artm.scores.add(
artm.PerplexityScore(name='PerplexityScore',
dictionary=dictionary))
model_artm.scores.add(artm.SparsityPhiScore(name='SparsityPhiScore'))
model_artm.scores.add(
artm.TopTokensScore(name='TopTokensScore', num_tokens=num_tokens))
model_lda = artm.LDA(num_topics=num_topics,
alpha=alpha,
beta=beta,
dictionary=dictionary,
cache_theta=True)
model_lda.initialize(dictionary=dictionary)
model_artm.num_document_passes = num_document_passes
model_lda.num_document_passes = num_document_passes
model_artm.fit_offline(batch_vectorizer=batch_vectorizer,
num_collection_passes=num_collection_passes)
model_lda.fit_offline(batch_vectorizer=batch_vectorizer,
num_collection_passes=num_collection_passes)
for i in range(num_collection_passes):
assert abs(model_artm.score_tracker['SparsityPhiScore'].value[i] -
model_lda.sparsity_phi_value[i]) < zero_eps
for i in range(num_collection_passes):
assert abs(
model_artm.score_tracker['SparsityThetaScore'].value[i] -
model_lda.sparsity_theta_value[i]) < zero_eps
for i in range(num_collection_passes):
assert abs(model_artm.score_tracker['PerplexityScore'].value[i] -
model_lda.perplexity_value[i]) < zero_eps
lda_tt = model_lda.get_top_tokens(num_tokens=num_tokens)
assert len(lda_tt) == num_topics
for i in range(num_topics):
for j in range(num_tokens):
assert model_artm.score_tracker['TopTokensScore'].last_tokens[
model_artm.topic_names[i]][j] == lda_tt[i][j]
lda_tt = model_lda.get_top_tokens(num_tokens=num_tokens,
with_weights=True)
for i in range(num_tokens):
assert abs(model_artm.score_tracker['TopTokensScore'].last_weights[
model_artm.topic_names[0]][i] - lda_tt[0][i][1]) < zero_eps
model_lda.fit_online(batch_vectorizer=batch_vectorizer)
phi = model_lda.phi_
assert phi.shape == (vocab_size, num_topics)
theta = model_lda.get_theta()
assert theta.shape == (num_topics, num_docs)
assert model_lda.library_version.count('.') == 2 # major.minor.patch
model_lda = artm.LDA(num_topics=num_topics,
alpha=alpha,
beta=([0.1] * num_topics),
dictionary=dictionary,
cache_theta=True)
assert model_lda._internal_model.regularizers.size() == num_topics + 1
finally:
shutil.rmtree(batches_folder)
def init_bcg_sparse_model(dataset,
modalities_to_use,
main_modality,
specific_topics,
bcg_topics,
model_params: dict = None):
"""
Creates simple artm model with standard scores.
Parameters
----------
dataset : Dataset
modalities_to_use : list of str or dict
main_modality : str
specific_topics : int
bcg_topics : int
Returns
-------
model: artm.ARTM() instance
"""
if model_params is None:
model_params = dict()
model = init_plsa(dataset, modalities_to_use, main_modality,
specific_topics, bcg_topics)
background_topic_names = model.topic_names[-bcg_topics:]
specific_topic_names = model.topic_names[:-bcg_topics]
dictionary = dataset.get_dictionary()
baseline_class_ids = {class_id: 1 for class_id in modalities_to_use}
data_stats = count_vocab_size(dictionary, baseline_class_ids)
# all coefficients are relative
regularizers = [
artm.SmoothSparsePhiRegularizer(
name='smooth_phi_bcg',
topic_names=background_topic_names,
tau=model_params.get("smooth_bcg_tau", 0.1),
class_ids=[main_modality],
),
artm.SmoothSparseThetaRegularizer(
name='smooth_theta_bcg',
topic_names=background_topic_names,
tau=model_params.get("smooth_bcg_tau", 0.1),
),
artm.SmoothSparsePhiRegularizer(
name='sparse_phi_sp',
topic_names=specific_topic_names,
tau=model_params.get("sparse_sp_tau", -0.05),
class_ids=[main_modality],
),
artm.SmoothSparseThetaRegularizer(
name='sparse_theta_sp',
topic_names=specific_topic_names,
tau=model_params.get("sparse_sp_tau", -0.05),
),
]
for reg in regularizers:
model.regularizers.add(
transform_regularizer(data_stats,
reg,
model.class_ids,
n_topics=len(reg.topic_names)))
return model
def test_func():
# constants
num_tokens = 11
probability_mass_threshold = 0.9
sp_reg_tau = -0.1
decor_tau = 1.5e+5
decor_rel_tau = 0.3
num_collection_passes = 15
num_document_passes = 1
num_topics = 15
vocab_size = 6906
num_docs = 3430
data_path = os.environ.get('BIGARTM_UNITTEST_DATA')
batches_folder = tempfile.mkdtemp()
sp_zero_eps = 0.001
sparsity_phi_value = [
0.034, 0.064, 0.093, 0.120, 0.145, 0.170, 0.194, 0.220, 0.246, 0.277,
0.312, 0.351, 0.390, 0.428, 0.464
]
sparsity_phi_rel_value = [
0.442, 0.444, 0.444, 0.446, 0.448, 0.449, 0.458, 0.468, 0.476, 0.488,
0.501, 0.522, 0.574, 0.609, 0.670
]
sparsity_theta_value = [0.0] * num_collection_passes
perp_zero_eps = 2.0
perplexity_value = [
6873, 2590, 2685, 2578, 2603, 2552, 2536, 2481, 2419, 2331, 2235, 2140,
2065, 2009, 1964
]
perplexity_rel_value = [
6873, 2667, 2458, 2323, 2150, 2265, 2015, 1967, 1807, 1747, 1713, 1607,
1632, 1542, 1469
]
top_zero_eps = 0.0001
top_tokens_num_tokens = [num_tokens * num_topics] * num_collection_passes
top_tokens_topic_0_tokens = [
u'party', u'state', u'campaign', u'tax', u'political', u'republican',
u'senate', u'candidate', u'democratic', u'court', u'president'
]
top_tokens_topic_0_weights = [
0.0209, 0.0104, 0.0094, 0.0084, 0.0068, 0.0067, 0.0065, 0.0058, 0.0053,
0.0053, 0.0051
]
ker_zero_eps = 0.02
topic_kernel_topic_0_contrast = 0.96
topic_kernel_topic_0_purity = 0.014
topic_kernel_topic_0_size = 18.0
topic_kernel_average_size = [
0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.13, 0.6, 1.6, 3.53, 7.15, 12.6,
20.4, 29.06
]
topic_kernel_average_contrast = [
0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.12, 0.31, 0.7, 0.96, 0.96, 0.96,
0.96, 0.97
]
topic_kernel_average_purity = [
0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.01, 0.01, 0.015, 0.017, 0.02,
0.03, 0.04, 0.05
]
len_last_document_ids = 10
try:
batch_vectorizer = artm.BatchVectorizer(data_path=data_path,
data_format='bow_uci',
collection_name='kos',
target_folder=batches_folder)
dictionary = artm.Dictionary()
dictionary.gather(data_path=batch_vectorizer.data_path)
model = artm.ARTM(
topic_names=['topic_{}'.format(i) for i in range(num_topics)],
dictionary=dictionary.name,
cache_theta=True)
model.regularizers.add(
artm.SmoothSparsePhiRegularizer(name='SparsePhi', tau=sp_reg_tau))
model.regularizers.add(
artm.DecorrelatorPhiRegularizer(name='DecorrelatorPhi',
tau=decor_tau))
model.scores.add(artm.SparsityThetaScore(name='SparsityThetaScore'))
model.scores.add(
artm.PerplexityScore(name='PerplexityScore',
use_unigram_document_model=False,
dictionary=dictionary))
model.scores.add(artm.SparsityPhiScore(name='SparsityPhiScore'))
model.scores.add(
artm.TopTokensScore(name='TopTokensScore', num_tokens=num_tokens))
model.scores.add(
artm.TopicKernelScore(
name='TopicKernelScore',
probability_mass_threshold=probability_mass_threshold))
model.scores.add(artm.ThetaSnippetScore(name='ThetaSnippetScore'))
model.num_document_passes = num_document_passes
model.fit_offline(batch_vectorizer=batch_vectorizer,
num_collection_passes=num_collection_passes)
for i in range(num_collection_passes):
assert abs(model.score_tracker['SparsityPhiScore'].value[i] -
sparsity_phi_value[i]) < sp_zero_eps
for i in range(num_collection_passes):
assert abs(model.score_tracker['SparsityThetaScore'].value[i] -
sparsity_theta_value[i]) < sp_zero_eps
for i in range(num_collection_passes):
assert abs(model.score_tracker['PerplexityScore'].value[i] -
perplexity_value[i]) < perp_zero_eps
for i in range(num_collection_passes):
assert model.score_tracker['TopTokensScore'].num_tokens[
i] == top_tokens_num_tokens[i]
for i in range(num_tokens):
assert model.score_tracker['TopTokensScore'].last_tokens[
model.topic_names[0]][i] == top_tokens_topic_0_tokens[i]
assert abs(model.score_tracker['TopTokensScore'].last_weights[
model.topic_names[0]][i] -
top_tokens_topic_0_weights[i]) < top_zero_eps
assert len(model.score_tracker['TopicKernelScore'].last_tokens[
model.topic_names[0]]) > 0
assert abs(topic_kernel_topic_0_contrast -
model.score_tracker['TopicKernelScore'].last_contrast[
model.topic_names[0]]) < ker_zero_eps
assert abs(topic_kernel_topic_0_purity -
model.score_tracker['TopicKernelScore'].last_purity[
model.topic_names[0]]) < ker_zero_eps
assert abs(topic_kernel_topic_0_size -
model.score_tracker['TopicKernelScore'].last_size[
model.topic_names[0]]) < ker_zero_eps
for i in range(num_collection_passes):
assert abs(
model.score_tracker['TopicKernelScore'].average_size[i] -
topic_kernel_average_size[i]) < ker_zero_eps
assert abs(
model.score_tracker['TopicKernelScore'].average_contrast[i] -
topic_kernel_average_contrast[i]) < ker_zero_eps
assert abs(
model.score_tracker['TopicKernelScore'].average_purity[i] -
topic_kernel_average_purity[i]) < ker_zero_eps
model.fit_online(batch_vectorizer=batch_vectorizer)
info = model.info
assert info is not None
assert len(info.config.topic_name) == num_topics
assert len(info.score) >= len(model.score_tracker)
assert len(info.regularizer) == len(model.regularizers.data)
assert len(info.cache_entry) > 0
temp = model.score_tracker['ThetaSnippetScore'].last_document_ids
assert len_last_document_ids == len(temp)
assert len(model.score_tracker['ThetaSnippetScore'].last_snippet[
temp[0]]) == num_topics
phi = model.get_phi()
assert phi.shape == (vocab_size, num_topics)
theta = model.get_theta()
assert theta.shape == (num_topics, num_docs)
assert model.library_version.count('.') == 2 # major.minor.patch
# test relative coefficients for Phi matrix regularizers
model = artm.ARTM(num_topics=num_topics,
dictionary=dictionary.name,
cache_theta=False)
model.regularizers.add(
artm.DecorrelatorPhiRegularizer(name='DecorrelatorPhi',
tau=decor_rel_tau))
model.regularizers['DecorrelatorPhi'].gamma = 0.0
model.scores.add(
artm.PerplexityScore(name='PerplexityScore',
use_unigram_document_model=False,
dictionary=dictionary))
model.scores.add(artm.SparsityPhiScore(name='SparsityPhiScore'))
model.num_document_passes = num_document_passes
model.fit_offline(batch_vectorizer=batch_vectorizer,
num_collection_passes=num_collection_passes)
for i in range(num_collection_passes):
assert abs(model.score_tracker['SparsityPhiScore'].value[i] -
sparsity_phi_rel_value[i]) < sp_zero_eps
for i in range(num_collection_passes):
assert abs(model.score_tracker['PerplexityScore'].value[i] -
perplexity_rel_value[i]) < perp_zero_eps
finally:
shutil.rmtree(batches_folder)
def init_lda(
dataset: Dataset,
modalities_to_use: List[str],
main_modality: str,
num_topics: int,
model_params: dict = None,
):
"""
Creates simple artm model with standard scores.
Parameters
----------
dataset
modalities_to_use
main_modality
num_topics
model_params
Returns
-------
model: artm.ARTM() instance
"""
if model_params is None:
model_params = dict()
model = init_plsa(dataset, modalities_to_use, main_modality, num_topics)
prior = model_params.get('prior', 'symmetric')
# What GenSim returns by default (everything is 'symmetric')
# see https://github.com/RaRe-Technologies/gensim/blob/master/gensim/models/ldamodel.py#L521
# Note that you can specify prior shape for alpha and beta separately,
# but we do not do that here
if prior == "symmetric":
alpha = 1.0 / num_topics
eta = 1.0 / num_topics
elif prior == "asymmetric":
# following the recommendation from
# http://papers.nips.cc/paper/3854-rethinking-lda-why-priors-matter
# we will use symmetric prior over Phi and asymmetric over Theta
eta = 1.0 / num_topics
num_terms = 0 # isn't used, so let's not compute it
alpha = _init_dirichlet_prior("alpha", num_topics, num_terms=num_terms)
elif prior == "double_asymmetric":
# this stuff is needed for asymmetric Phi initialization:
artm_dict = dataset.get_dictionary()
temp_df = artm_dict2df(artm_dict) # noqa: F821
num_terms = temp_df.query("class_id in @modalities_to_use").shape[0]
eta = _init_dirichlet_prior("eta", num_topics, num_terms)
alpha = _init_dirichlet_prior("alpha", num_topics, num_terms)
# TODO: turns out, BigARTM does not support tau as a list of floats (or dictionary)
# so we need to use custom regularizer instead
# (TopicPrior doesn't work because it provides $beta_t$ instead of $beta_w$)
raise NotImplementedError
elif prior == "heuristic":
# Found in doi.org/10.1007/s10664-015-9379-3 (2016)
# "We use the defacto standard heuristics of α=50/K and β=0.01
# (Biggers et al. 2014) for our hyperparameter values"
alpha = 50.0 / num_topics
eta = 0.01
else:
raise TypeError(f"prior type '{prior}' is not supported")
model.regularizers.add(
artm.SmoothSparsePhiRegularizer(
name='smooth_phi',
tau=eta,
class_ids=[main_modality],
), )
if isinstance(alpha, (list, np.ndarray)):
alpha = [float(a) for a in alpha]
assert (len(alpha) == len(model.topic_names))
for i, topic in enumerate(model.topic_names):
model.regularizers.add(
artm.SmoothSparseThetaRegularizer(name=f'smooth_theta_{i}',
tau=alpha[i],
topic_names=topic))
else:
model.regularizers.add(
artm.SmoothSparseThetaRegularizer(
name='smooth_theta',
tau=alpha,
), )
return model
