def test_batch_size(self):
model_name = self.eng_model_names[0]
# 1 per batch
aug = naw.BackTranslationAug(
from_model_name=model_name['from_model_name'],
to_model_name=model_name['to_model_name'],
batch_size=1)
aug_data = aug.augment(self.texts)
self.assertEqual(len(aug_data), len(self.texts))
# batch size = input size
aug = naw.BackTranslationAug(
from_model_name=model_name['from_model_name'],
to_model_name=model_name['to_model_name'],
batch_size=len(self.texts))
aug_data = aug.augment(self.texts)
self.assertEqual(len(aug_data), len(self.texts))
# batch size > input size
aug = naw.BackTranslationAug(
from_model_name=model_name['from_model_name'],
to_model_name=model_name['to_model_name'],
batch_size=len(self.texts) + 1)
aug_data = aug.augment(self.texts)
self.assertEqual(len(aug_data), len(self.texts))
# input size > batch size
aug = naw.BackTranslationAug(
from_model_name=model_name['from_model_name'],
to_model_name=model_name['to_model_name'],
batch_size=2)
aug_data = aug.augment(self.texts * 2)
self.assertEqual(len(aug_data), len(self.texts) * 2)
def test_back_translation(self):
# From English
text = 'The quick brown fox jumps over the lazy dog'
for model_name in self.eng_model_names:
aug = naw.BackTranslationAug(
from_model_name=model_name['from_model_name'],
from_model_checkpt=model_name['from_model_checkpt'],
to_model_name=model_name['to_model_name'],
to_model_checkpt=model_name['to_model_checkpt'])
augmented_text = aug.augment(text)
aug.clear_cache()
self.assertNotEqual(text, augmented_text)
self.assertTrue(len(self.eng_model_names) > 1)
def test_load_from_local_path_inexist(self):
from_model_dir = '/abc/'
to_model_dir = '/def/'
with self.assertRaises(ValueError) as error:
aug = naw.BackTranslationAug(from_model_name=from_model_dir,
from_model_checkpt='model1.pt',
to_model_name=to_model_dir,
to_model_checkpt='model1.pt',
is_load_from_github=False)
self.assertTrue(
'Cannot load model from local path' in str(error.exception))
base_model_dir = os.environ.get("MODEL_DIR")
from_model_dir = os.path.join(base_model_dir, 'word', 'fairseq',
'wmt19.en-de')
to_model_dir = '/def/'
with self.assertRaises(ValueError) as error:
aug = naw.BackTranslationAug(from_model_name=from_model_dir,
from_model_checkpt='model1.pt',
to_model_name=to_model_dir,
to_model_checkpt='model1.pt',
is_load_from_github=False)
self.assertTrue(
'Cannot load model from local path' in str(error.exception))
def test_load_from_local_path(self):
base_model_dir = os.environ.get("MODEL_DIR")
from_model_dir = os.path.join(base_model_dir, 'word', 'fairseq',
'wmt19.en-de')
to_model_dir = os.path.join(base_model_dir, 'word', 'fairseq',
'wmt19.de-en', '')
aug = naw.BackTranslationAug(from_model_name=from_model_dir,
from_model_checkpt='model1.pt',
to_model_name=to_model_dir,
to_model_checkpt='model1.pt',
is_load_from_github=False)
augmented_text = aug.augment(self.text)
aug.clear_cache()
self.assertNotEqual(self.text, augmented_text)
def sample_test_case(self, device):
# From English
for model_name in self.eng_model_names:
aug = naw.BackTranslationAug(
from_model_name=model_name['from_model_name'],
to_model_name=model_name['to_model_name'],
device=device)
augmented_text = aug.augment(self.text)
aug.clear_cache()
self.assertNotEqual(self.text, augmented_text)
augmented_texts = aug.augment(self.texts)
aug.clear_cache()
for d, a in zip(self.texts, augmented_texts):
self.assertNotEqual(d, a)
if device == 'cpu':
self.assertTrue(device == aug.model.get_device())
elif 'cuda' in device:
self.assertTrue('cuda' in aug.model.get_device())
def test_back_translation(self):
# From English
texts = [
self.text,
"Seeing all of the negative reviews for this movie, I figured that it could be yet another comic masterpiece that wasn't quite meant to be."
]
for model_name in self.eng_model_names:
aug = naw.BackTranslationAug(
from_model_name=model_name['from_model_name'],
from_model_checkpt=model_name['from_model_checkpt'],
to_model_name=model_name['to_model_name'],
to_model_checkpt=model_name['to_model_checkpt'])
augmented_text = aug.augment(self.text)
aug.clear_cache()
self.assertNotEqual(self.text, augmented_text)
augmented_texts = aug.augment(texts)
aug.clear_cache()
for d, a in zip(texts, augmented_texts):
self.assertNotEqual(d, a)
self.assertTrue(len(self.eng_model_names) > 1)
def augmented_all(
use_textcomp19=False,
use_weebit=False,
use_dw=False,
backtrans=False,
lemmatization=False,
stemming=False,
randword_swap=False,
randword_del=False,
test_size=0.1,
):
"""
Returns the augmented training dataset
and the test dataset of all specified data.
backtrans : enables back and forth translation of the data
lemmatization self explanatory
stemming self explanatory
randword_swap : enables randomly swapping words around sentences
randword_del : enalbles randomly deleting words from sentences
test_size : gives the ratio of test to train set
train_set, test_set = augmented_all()
"""
# Perform a Train-Test Split keeping dataset proportions the same
print("perform train-test split keeping dataset proportions the same")
all_dataset = all_data(use_textcomp19, use_weebit, use_dw)
print("#####################", all_dataset[all_dataset["source"] == 1])
if use_textcomp19:
text_comp_train, text_comp_test = train_test_split(
all_dataset[all_dataset["source"] == 0], test_size=test_size)
if use_weebit:
weebit_train = all_dataset[all_dataset["source"] == 1]
print(weebit_train)
if use_dw:
dw_train = all_dataset[all_dataset["source"] == 2]
if use_textcomp19 and not use_weebit and not use_dw: #0
all_dataset_train = text_comp_train
all_dataset_test = text_comp_test
if use_weebit and not use_textcomp19 and not use_dw: #1
print("No weebit test set available!")
all_dataset_train = weebit_train
if use_dw and not use_weebit and not use_textcomp19: #2
print("No dw test set available!")
all_dataset_train = dw_train
all_dataset_test = dw_train # added so that dataset with only dw can be created
if use_textcomp19 and use_weebit and not use_dw: #01
all_dataset_train = text_comp_train.append(weebit_train,
ignore_index=True)
all_dataset_test = text_comp_test
if use_weebit and use_dw and not use_textcomp19: #12
print("No weebit and dw test set available!")
all_dataset_train = weebit_train.append(dw_train, ignore_index=True)
if use_textcomp19 and use_dw and not use_weebit: #02
all_dataset_train = text_comp_train.append(dw_train, ignore_index=True)
all_dataset_test = text_comp_test
if use_textcomp19 and use_weebit and use_dw: # 012
all_dataset_train = text_comp_train.append(weebit_train,
ignore_index=True)
all_dataset_train = all_dataset_train.append(dw_train,
ignore_index=True)
all_dataset_test = text_comp_test
## Augmentation of data
print("Start augmenting Data...")
# Back and forth translation of data
if backtrans == True:
print("Back and forth translation...")
back_translation_aug = naw.BackTranslationAug(
from_model_name="transformer.wmt19.de-en",
to_model_name="transformer.wmt19.en-de",
)
if use_weebit:
translated = all_dataset_train[all_dataset_train["source"] != 1]
else:
translated = all_dataset_train
translated["raw_text"] = translated["raw_text"].apply(
lambda x: back_translation_aug.augment(x))
all_dataset_train = all_dataset_train.append(translated,
ignore_index=True)
# Random word swap
if randword_swap == True:
print("Random word swap")
aug1 = naw.RandomWordAug(action="swap")
swapped_data = all_dataset_train
swapped_data["raw_text"] = all_dataset_train["raw_text"].apply(
lambda x: aug1.augment(x))
all_dataset_train = all_dataset_train.append(swapped_data,
ignore_index=True)
# Random word deletion
if randword_del == True:
print("Random word deletion")
aug2 = naw.RandomWordAug()
rand_deleted_data = all_dataset_train
rand_deleted_data["raw_text"] = all_dataset_train["raw_text"].apply(
lambda x: aug2.augment(x))
all_dataset_train = all_dataset_train.append(rand_deleted_data,
ignore_index=True)
# Lemmatization using spacy
if lemmatization == True:
print("lemmatizing")
nlp = spacy.load("de_core_news_sm")
all_dataset_train["raw_text"] = all_dataset_train["raw_text"].apply(
lambda x: " ".join([y.lemma_ for y in nlp(x)]))
all_dataset_test["raw_text"] = all_dataset_test["raw_text"].apply(
lambda x: " ".join([y.lemma_ for y in nlp(x)]))
# Stemming using
if stemming == True:
print("stemming")
stemmer = SnowballStemmer("german")
all_dataset_train["raw_text"] = all_dataset_train["raw_text"].apply(
lambda x: stemmer.stem(x))
all_dataset_test["raw_text"] = all_dataset_test["raw_text"].apply(
lambda x: stemmer.stem(x))
return all_dataset_train, all_dataset_test
def create_back_aug(language):
return naw.BackTranslationAug(
from_model_name=f'transformer.wmt19.en-{language}',
to_model_name=f'transformer.wmt19.{language}-en',
device='cuda')
