def test_training_with_saved_data(data_dir):
handler = TextHandler(data_dir + "gnews/GoogleNews.txt")
handler.prepare() # create vocabulary and training data
# load BERT data
with open(data_dir + "gnews/bert_embeddings_gnews", "rb") as filino:
training_bert = pickle.load(filino)
training_dataset = CTMDataset(handler.bow, training_bert,
handler.idx2token)
ctm = CTM(input_size=len(handler.vocab),
bert_input_size=768,
num_epochs=1,
inference_type="combined",
n_components=5)
ctm.fit(training_dataset) # run the model
print(ctm.get_topics(2))
ctm.get_doc_topic_distribution(training_dataset)
assert isclose(np.sum(ctm.get_topic_word_distribution()[0]),
1,
rel_tol=1e-5,
abs_tol=0.0)
def test_shape_checks(data_dir):
handler = TextHandler(data_dir + "gnews/GoogleNews.txt")
handler.prepare() # create vocabulary and training data
# load BERT data
with open(data_dir + "gnews/bert_embeddings_gnews", "rb") as filino:
training_bert = pickle.load(filino)
assert len(training_bert) == len(handler.bow.todense())
def test_embeddings_from_scratch(data_dir):
handler = TextHandler(data_dir + "sample_text_document")
handler.prepare() # create vocabulary and training data
assert np.array_equal(
handler.bow.todense(),
np.array([[1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0],
[1, 1, 1, 0, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 2]]))
def test_embeddings_from_scratch(data_dir):
handler = TextHandler(data_dir + "sample_text_document")
handler.prepare() # create vocabulary and training data
assert np.array_equal(
handler.bow.todense(),
np.array([[1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0],
[1, 1, 1, 0, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 2]]))
train_bert = bert_embeddings_from_file(
data_dir + 'sample_text_document', "distiluse-base-multilingual-cased")
assert len(train_bert) == len(handler.bow.todense())
def test_training(data_dir):
handler = TextHandler(data_dir + "sample_text_document")
handler.prepare() # create vocabulary and training data
train_bert = bert_embeddings_from_file(
data_dir + 'sample_text_document', "distiluse-base-multilingual-cased")
training_dataset = CTMDataset(handler.bow, train_bert, handler.idx2token)
ctm = CTM(input_size=len(handler.vocab),
bert_input_size=512,
num_epochs=1,
inference_type="combined",
n_components=5)
ctm.fit(training_dataset) # run the model
topics = ctm.get_topic_lists(2)
assert len(topics) == 5
thetas = ctm.get_thetas(training_dataset)
assert len(thetas) == len(train_bert)
import torch
import sys
from contextualized_topic_models.models.ctm import CTM
from contextualized_topic_models.evaluation.measures import CoherenceNPMI
from contextualized_topic_models.utils.data_preparation import TextHandler
def show_topics(topic_list):
for idx, topic_tokens in enumerate(topic_list):
print(idx)
print(' '.join(topic_tokens))
if len(sys.argv) < 2:
raise Exception("Usage: python {} {}".format(sys.argv[0], "<model_file>"))
handler = TextHandler(
"contextualized_topic_models/data/wiki/wiki_train_en_prep.txt")
handler.prepare()
ctm = CTM(input_size=len(handler.vocab), inference_type="contextual", bert_input_size=768, \
num_data_loader_workers=1)
ctm.load(sys.argv[1], sys.argv[2])
with open("contextualized_topic_models/data/wiki/wiki_train_en_prep.txt",
"r") as en:
texts = [doc.split() for doc in en.read().splitlines()]
# obtain NPMI coherences on the topic modeled documents
show_topics(ctm.get_topic_lists(10))
npmi = CoherenceNPMI(texts=texts, topics=ctm.get_topic_lists(10))
print(npmi.score())
def test_training_all_classes_ctm(data_dir):
handler = TextHandler(data_dir + "sample_text_document")
handler.prepare() # create vocabulary and training data
train_bert = bert_embeddings_from_file(
data_dir + 'sample_text_document', "distiluse-base-multilingual-cased")
training_dataset = CTMDataset(handler.bow, train_bert, handler.idx2token)
ctm = CTM(input_size=len(handler.vocab),
bert_input_size=512,
num_epochs=1,
inference_type="combined",
n_components=5)
ctm.fit(training_dataset) # run the model
topics = ctm.get_topic_lists(2)
assert len(topics) == 5
thetas = ctm.get_doc_topic_distribution(training_dataset)
assert len(thetas) == len(train_bert)
ctm = ZeroShotTM(input_size=len(handler.vocab),
bert_input_size=512,
num_epochs=1,
n_components=5)
ctm.fit(training_dataset) # run the model
topics = ctm.get_topic_lists(2)
assert len(topics) == 5
thetas = ctm.get_doc_topic_distribution(training_dataset)
assert len(thetas) == len(train_bert)
ctm = CombinedTM(input_size=len(handler.vocab),
bert_input_size=512,
num_epochs=1,
n_components=5)
ctm.fit(training_dataset) # run the model
topics = ctm.get_topic_lists(2)
assert len(topics) == 5
thetas = ctm.get_doc_topic_distribution(training_dataset)
assert len(thetas) == len(train_bert)
with open(data_dir + 'sample_text_document') as filino:
data = filino.readlines()
handler = TextHandler(sentences=data)
handler.prepare() # create vocabulary and training data
train_bert = bert_embeddings_from_list(
data, "distiluse-base-multilingual-cased")
training_dataset = CTMDataset(handler.bow, train_bert, handler.idx2token)
ctm = CTM(input_size=len(handler.vocab),
bert_input_size=512,
num_epochs=1,
inference_type="combined",
n_components=5)
ctm.fit(training_dataset) # run the model
topics = ctm.get_topic_lists(2)
assert len(topics) == 5
thetas = ctm.get_doc_topic_distribution(training_dataset)
assert len(thetas) == len(train_bert)
qt = QuickText("distiluse-base-multilingual-cased",
text_for_bow=data,
text_for_bert=data)
dataset = qt.load_dataset()
ctm = ZeroShotTM(input_size=len(qt.vocab),
bert_input_size=512,
num_epochs=1,
n_components=5)
ctm.fit(dataset) # run the model
topics = ctm.get_topic_lists(2)
assert len(topics) == 5
qt_from_conf = QuickText("distiluse-base-multilingual-cased", None, None)
qt_from_conf.load_configuration(qt.bow, qt.data_bert, qt.vocab,
qt.idx2token)
dataset = qt_from_conf.load_dataset()
ctm = ZeroShotTM(input_size=len(qt.vocab),
bert_input_size=512,
num_epochs=1,
n_components=5)
ctm.fit(dataset) # run the model
topics = ctm.get_topic_lists(2)
assert len(topics) == 5
from collections import Counter
def show_topics(topic_list):
for idx, topic_tokens in enumerate(topic_list):
print(idx)
print(' '.join(topic_tokens))
if len(sys.argv) < 4:
raise Exception("Usage: python {} {} {} {}".format(sys.argv[0],
"<ctm_model>",
"<checkpoint>",
"<sbert_model>"))
handler_en = TextHandler(
"contextualized_topic_models/data/wiki/wiki_test_en_prep_sub.txt")
# handler = TextHandler("contextualized_topic_models/data/wiki/iqos_corpus_prep_en.txt")
handler_en.prepare()
testing_bert_en = bert_embeddings_from_file(
"contextualized_topic_models/data/wiki/wiki_test_en_unprep_sub.txt",
sys.argv[3])
testing_dataset_en = CTMDataset(handler_en.bow, testing_bert_en,
handler_en.idx2token)
ctm = CTM(input_size=len(handler_en.vocab),
inference_type="contextual",
bert_input_size=768)
# ctm = torch.load(sys.argv[1], map_location="cpu")
ctm.load(sys.argv[1], sys.argv[2])
