def test_n_output_without_augmentation(self):
texts = [
'AAAAAAAAAAA AAAAAAAAAAAAAA'
]
flows = [
naf.Sequential([
nac.OcrAug(),
nac.OcrAug()
]),
naf.Sometimes([
nac.RandomCharAug(),
nac.RandomCharAug()
], pipeline_p=0.00001)
]
for flow in flows:
for text in texts:
for _ in range(5):
augmented_texts = flow.augment(text, n=3)
all_not_equal = False
for augmented_text in augmented_texts:
if augmented_text != text:
all_not_equal = True
break
if all_not_equal:
break
self.assertFalse(all_not_equal)
self.assertLess(0, len(flows))
self.assertLess(0, len(texts))
def test_multiple_actions(self):
texts = [
'The quick brown fox jumps over the lazy dog',
'Zology raku123456 fasdasd asd4123414 1234584'
]
flows = [
naf.Sequential([
naf.Sometimes([
nac.RandomCharAug(action=Action.INSERT),
nac.RandomCharAug(action=Action.DELETE)
],
pipeline_p=0.5),
naf.Sequential([
nac.OcrAug(),
nac.QwertyAug(aug_min=1),
nac.RandomCharAug(action=Action.SUBSTITUTE,
aug_min=1,
aug_char_p=0.6,
aug_word_p=0.6)
],
name='Sub_Seq')
]),
naf.Sometimes([
naf.Sometimes([
nac.RandomCharAug(action=Action.INSERT),
nac.RandomCharAug(action=Action.DELETE)
]),
naf.Sequential([
nac.OcrAug(),
nac.QwertyAug(aug_min=1),
nac.RandomCharAug(action=Action.SUBSTITUTE,
aug_min=1,
aug_char_p=0.6,
aug_word_p=0.6)
])
],
pipeline_p=0.5)
]
# Since prob may be low and causing do not perform data augmentation. Retry 5 times
for flow in flows:
at_least_one_not_equal = False
for _ in range(0, 5):
for text in texts:
self.assertLess(0, len(text))
augmented_text = flow.augment(text)
if text != augmented_text:
at_least_one_not_equal = True
self.assertLess(0, len(text))
if at_least_one_not_equal:
break
self.assertTrue(at_least_one_not_equal)
self.assertLess(0, len(flows))
self.assertLess(0, len(texts))
def augment_text_ocr(comment):
aug = nac.OcrAug(aug_char_p=0.3, aug_word_p=0.4, aug_word_min=len(comment))
try:
augmented_texts = aug.augment(comment, n=1)
except:
augmented_texts = None
return augmented_texts
def char_level(text, n):
#Augmenting data in character level.
aug = nac.OcrAug()
attacked_texts = aug.augment(text, n=n)
# gives n forms of augmentation (n is the number of augmented forms a user wants)
print("Attacked Text:")
print(attacked_texts)
def test_multiple_actions(self):
texts = [
'The quick brown fox jumps over the lazy dog',
'Zology raku123456 fasdasd asd4123414 1234584'
]
flows = [
naf.Sequential(
[nac.RandomCharAug(action=Action.INSERT),
naw.RandomWordAug()]),
naf.Sequential([
nac.OcrAug(),
nac.KeyboardAug(aug_min=1),
nac.RandomCharAug(action=Action.SUBSTITUTE,
aug_min=1,
aug_char_p=0.6,
aug_word_p=0.6)
])
]
for flow in flows:
for text in texts:
augmented_text = flow.augment(text)
self.assertNotEqual(text, augmented_text)
self.assertLess(0, len(text))
self.assertLess(0, len(texts))
self.assertLess(0, len(flows))
def test_multi_thread(self):
text = 'The quick brown fox jumps over the lazy dog'
n = 3
w2v_model_path = os.path.join(os.environ["MODEL_DIR"], 'word',
'word_embs',
'GoogleNews-vectors-negative300.bin')
flows = [
naf.Sequential([
naf.Sequential([
nac.OcrAug(),
naw.WordEmbsAug(model_type='word2vec',
model_path=w2v_model_path)
]),
naf.Sequential([
nac.RandomCharAug(),
]),
naw.ContextualWordEmbsAug(model_path='xlnet-base-cased',
action="substitute",
temperature=0.7,
device='cpu')
]),
naf.Sometimes([
naf.Sequential([
nac.OcrAug(),
nac.RandomCharAug(),
]),
naf.Sometimes([
naw.WordEmbsAug(model_type='word2vec',
model_path=w2v_model_path)
],
pipeline_p=0.999),
naw.ContextualWordEmbsAug(model_path='xlnet-base-cased',
action="substitute",
temperature=0.7,
device='cpu')
],
pipeline_p=0.9999)
]
for num_thread in [1, 3]:
for flow in flows:
augmented_data = flow.augment(text, n=n, num_thread=num_thread)
self.assertEqual(len(augmented_data), n)
def test_multiple_actions(self):
texts = [
'The quick brown fox jumps over the lazy dog',
'Zology raku123456 fasdasd asd4123414 1234584'
]
flows = [
naf.Sequential([
naf.Sometimes([
nac.RandomCharAug(action="insert"),
nac.RandomCharAug(action="delete")
],
pipeline_p=0.9),
naf.Sequential(
[
# nac.OcrAug(), nac.QwertyAug(aug_min=1),
nac.RandomCharAug(action="substitute",
aug_char_min=1,
aug_char_p=0.6,
aug_word_p=0.6)
],
name='Sub_Seq')
]),
naf.Sometimes([
naf.Sometimes([
nac.RandomCharAug(action="insert"),
nac.RandomCharAug(action="delete")
]),
naf.Sequential([
nac.OcrAug(),
nac.KeyboardAug(aug_char_min=1),
nac.RandomCharAug(action="substitute",
aug_char_min=1,
aug_char_p=0.6,
aug_word_p=0.6)
])
],
pipeline_p=0.9)
]
# Since prob may be low and causing do not perform data augmentation. Retry 5 times
for flow in flows:
for text in texts:
at_least_one_not_equal = False
for _ in range(5):
augmented_text = flow.augment(text, n=1)
if text != augmented_text:
at_least_one_not_equal = True
break
self.assertTrue(at_least_one_not_equal)
self.assertLess(0, len(flows))
self.assertLess(0, len(texts))
def augment_text_ocr(comment):
"""
OCRAug adds noise to a comment by replacing the target characters with predefined mapping table
"""
aug = nac.OcrAug(aug_char_p=0.3, aug_word_p=0.4, aug_word_min=len(comment))
try:
augmented_texts = aug.augment(comment, n=1)
except:
augmented_texts = None
return augmented_texts
def nlpaug(word):
aug = naf.Sometimes([
nac.OcrAug(),
nac.KeyboardAug(),
nac.RandomCharAug(action="insert"),
nac.RandomCharAug(action="substitute"),
nac.RandomCharAug(action="swap"),
nac.RandomCharAug(action="delete"),
naw.SpellingAug(),
])
word = aug.augment(word)
return word
def __init__(self, template, output_file, augmentation_factor=5):
assert augmentation_factor >= 2
self.augmentation_factor = augmentation_factor
self.base_file = template
self.output_file = output_file
self.dataset = {}
self.intents = {}
self.character_augmenter = nac.OcrAug()
self.word_augmenter = naw.ContextualWordEmbsAug()
def test_n_output_without_augmentation(self):
texts = ['AAAAAAAAAAA AAAAAAAAAAAAAA']
flows = [
naf.Sequential([nac.OcrAug(), nac.OcrAug()]),
naf.Sometimes(
[nac.RandomCharAug(), nac.RandomCharAug()], pipeline_p=0.00001)
]
for flow in flows:
for text in texts:
at_least_one_equal = False
for _ in range(5):
augmented_texts = flow.augment(text, n=3)
if len(augmented_texts
) == 1 and augmented_texts[0] == text:
at_least_one_equal = True
break
self.assertTrue(at_least_one_equal)
self.assertLess(0, len(flows))
self.assertLess(0, len(texts))
def test_empty(self):
texts = ['', None]
augs = [
nac.OcrAug(),
nac.KeyboardAug(),
]
for text in texts:
for aug in augs:
augmented_text = aug.augment(text)
self.assertEqual(text, augmented_text)
def test_special_char(self):
text = '#'
aug = nac.KeyboardAug(min_char=1)
augmented_text = aug.augment(text)
self.assertNotEqual(text, augmented_text)
# No mapping, return original value
text = '~'
augs = [nac.KeyboardAug(min_char=1), nac.OcrAug(min_char=1)]
for aug in augs:
augmented_text = aug.augment(text)
self.assertEqual(text, augmented_text)
def test_stopwords_regex(self):
text = 'The quick brown fox jumps over the lazy dog.'
stopwords_regex = "( [a-zA-Z]{1}ox | [a-z]{1}og|(brown)|[a-zA-z]{1}he)|[a-z]{2}mps "
augs = [
nac.RandomCharAug(action="delete", stopwords_regex=stopwords_regex),
nac.KeyboardAug(stopwords_regex=stopwords_regex),
nac.OcrAug(stopwords_regex=stopwords_regex)
]
for aug in augs:
for i in range(10):
augmented_text = aug.augment(text)
self.assertTrue(
'quick' not in augmented_text or 'over' not in augmented_text or 'lazy' not in augmented_text)
def test_stopwords(self):
text = 'The quick brown fox jumps over the lazy dog.'
stopwords = ['The', 'brown', 'fox', 'jumps', 'the', 'dog']
augs = [
nac.RandomCharAug(stopwords=stopwords),
nac.KeyboardAug(stopwords=stopwords),
nac.OcrAug(stopwords=stopwords)
]
for aug in augs:
for i in range(10):
augmented_text = aug.augment(text)
self.assertTrue(
'quick' not in augmented_text or 'over' not in augmented_text or 'lazy' not in augmented_text)
def test_empty_input_for_substitute(self):
texts = ['', ' ']
augs = [
nac.RandomCharAug(action='substitute'),
nac.KeyboardAug(),
nac.OcrAug()
]
for aug in augs:
for text in texts:
augmented_text = aug.augment(text)
self.assertTrue(augmented_text is None or augmented_text.strip() == '')
augmented_texts = aug.augment(texts)
for augmented_text in augmented_texts:
self.assertTrue(augmented_text is None or augmented_text.strip() == '')
def test_augment_detail(self):
text = 'The quick brown fox jumps over the lazy dog'
flows = [
naf.Sequential([
naf.Sometimes([
nac.RandomCharAug(action="insert"),
nac.RandomCharAug(action="delete")
],
pipeline_p=0.5),
naf.Sequential([
nac.RandomCharAug(action="substitute",
aug_char_min=1,
aug_char_p=0.6,
aug_word_p=0.6)
],
name='Sub_Seq')
],
include_detail=True),
naf.Sometimes([
naf.Sometimes([
nac.RandomCharAug(action="insert"),
nac.RandomCharAug(action="delete")
]),
naf.Sequential([
nac.OcrAug(),
nac.KeyboardAug(aug_char_min=1),
nac.RandomCharAug(action="substitute",
aug_char_min=1,
aug_char_p=0.6,
aug_word_p=0.6)
])
],
pipeline_p=1,
include_detail=True)
]
for flow in flows:
augmented_text, augment_details = flow.augment(text)
self.assertNotEqual(text, augmented_text)
self.assertGreater(len(augment_details), 0)
for augment_detail in augment_details:
self.assertGreater(augment_detail['orig_start_pos'], -1)
self.assertGreater(augment_detail['new_start_pos'], -1)
self.assertGreater(augment_detail['change_seq'], 0)
self.assertIn(augment_detail['action'], Action.getall())
def test_min_char(self):
text = 'He eats apple'
augs = [
nac.RandomCharAug(min_char=5),
nac.KeyboardAug(min_char=5),
nac.OcrAug(min_char=5)
]
for aug in augs:
augmented = False
for i in range(10):
augmented_text = aug.augment(text)
if 'apple' not in augmented_text:
augmented = True
break
self.assertTrue(augmented)
def test_tokenizer(self):
augs = [
nac.OcrAug(tokenizer=text_tokenizer.split_sentence),
nac.KeyboardAug(tokenizer=text_tokenizer.split_sentence),
nac.RandomCharAug(tokenizer=text_tokenizer.split_sentence),
]
text = 'The quick brown fox, jumps over lazy dog.'
expected_tokens = ['The', ' quick', ' brown', ' fox', ', ', 'jumps', ' over', ' lazy', ' dog', '.']
for aug in augs:
tokens = aug.tokenizer(text)
self.assertEqual(tokens, expected_tokens)
text = 'The quick !brown fox, jumps # over lazy dog .'
expected_tokens = ['The', ' quick', ' !', 'brown', ' fox', ', ', 'jumps', ' # ', 'over', ' lazy', ' dog', ' .']
for aug in augs:
tokens = aug.tokenizer(text)
self.assertEqual(tokens, expected_tokens)
def test_augment_detail(self):
text = 'The quick brown fox jumps over the lazy dog'
augs = [
nac.KeyboardAug(min_char=1, include_detail=True),
nac.OcrAug(min_char=1, include_detail=True),
nac.RandomCharAug(min_char=2, include_detail=True)
]
for aug in augs:
augmented_text, augment_details = aug.augment(text)
self.assertNotEqual(text, augmented_text)
self.assertGreater(len(augment_details), 0)
for augment_detail in augment_details:
self.assertTrue(augment_detail['orig_token'] in text)
self.assertGreater(augment_detail['orig_start_pos'], -1)
self.assertGreater(augment_detail['new_start_pos'], -1)
self.assertGreater(augment_detail['change_seq'], 0)
self.assertIn(augment_detail['action'], Action.getall())
# Get back original input by re-engineering
reengineering_text = augmented_text
for change_obj in sorted(augment_details, key=lambda item: item['orig_start_pos'], reverse=True):
if change_obj['action'] == Action.DELETE:
text_prefix = reengineering_text[:change_obj['new_start_pos']]
text_core = change_obj['orig_token'] + ' '
text_suffix = reengineering_text[change_obj['new_start_pos']:]
elif change_obj['action'] in [Action.INSERT, Action.SUBSTITUTE]:
text_prefix = reengineering_text[:change_obj['new_start_pos']]
text_core = reengineering_text[change_obj['new_start_pos']:].replace(
change_obj['new_token'], change_obj['orig_token'], 1)
text_suffix = ''
# TODO
# elif change_obj['action'] in Action.SWAP:
reengineering_text = text_prefix + text_core + text_suffix
reengineering_text = reengineering_text.strip()
self.assertEqual(text, reengineering_text)
def ocr_aug(corpus):
aug = nac.OcrAug(tokenizer=whitespace_tokenizer)
# go through all train and dev sentences
augmented_sentences = []
for sentence in corpus.train:
augmented_texts = aug.augment(sentence.to_tokenized_string(), n=3)
for augmented_text in augmented_texts:
augmented_sentence: Sentence = Sentence()
augmented_token_texts = augmented_text.split(" ")
for augmented_token_text, original_token in zip(augmented_token_texts, sentence):
# make a new token
augmented_token = Token(augmented_token_text)
# transfer annotations over to augmented token
augmented_token.annotation_layers = original_token.annotation_layers
# add augmented token to augmented sentence
augmented_sentence.add_token(augmented_token)
# add augmented sentence to list of all augmented sentences
augmented_sentences.append(augmented_sentence)
corpus = Corpus(train=SentenceDataset(augmented_sentences),
dev=corpus.dev,
test=corpus.test)
return corpus
def augment_image(
dataset: list,
augmenter,
):
"""
a generic augment process on generator
:param dataset: dataset generator (batch)
:param augmenter:
:return:
"""
for data_point in dataset:
data, label = data_point
if augmenter is not None:
data = augmenter(images=data)
yield np.asarray(data), np.asarray(label)
# NULL augmenter
none_augmenter = partial(augment_text, augmenter=None)
# OCR error augmenter
text_ocr_augmenter = partial(augment_text, augmenter=nac.OcrAug())
# flip_r image augmenter
image_flip_r_augmenter = partial(augment_image, augmenter=iaa.Fliplr())
# rotation image augmenter
image_rot_augmenter = partial(augment_image, augmenter=iaa.Rotate())
import re
import attr
import random
import pandas as pd
from faker import Faker
from .utility import formating
from .CONSTATNTS import *
import nlpaug.augmenter.char as nac
fake = Faker()
aug = nac.OcrAug()
@attr.s
class DataGeneration:
template_list = attr.ib()
number = attr.ib()
@property
def data(self):
return_data = []
const_data = self.constant_generation()
for i in range(0, self.number):
temp, class_3_label = random.choice(self.template_list)
class_2_label = class_3_label
if class_3_label == 'partial-addressline':
class_2_label = 'addressline'
matcher = re.findall('\{.*?\}', temp)
for i in matcher:
key = i.replace('{', '').replace('}', '')
tp = random.choice(const_data[key])
def augment_dataset(input_df):
"""
Augmenting the dataset based on NLP aug library. If the dataset is small, this is a great way to boost things up.
But you do not want to apply augmentation on the Doctor's response. These should not have any spelling mistakes.
- The augmentation that will be done here is character level augmentations and word level augmentations:
- OCR error augmentation (character level)
- Keyboard augmentation (character level)
- Synonym augmenter (word level)
"""
print('Augmenting the dataset based on Synonyms...')
ocr = nac.OcrAug()
response_OCR = []
context_OCR = []
keyboard = nac.KeyboardAug()
response_keyboard = []
context_keyboard = []
synonym = naw.SynonymAug(aug_src='wordnet')
response_synonym = []
context_synonym = []
for i in input_df.index:
if i % 10 == 0:
print('processing {}th line'.format(i))
response = input_df['response'][i]
context = input_df['context'][i]
#augmentation
ocr_augmented_line = ocr.augment(context, n=3)
response_OCR.append(response)
context_OCR.append(ocr_augmented_line)
#keyboard augmentation
keyboard_augmented_line = keyboard.augment(context)
response_keyboard.append(response)
context_keyboard.append(keyboard_augmented_line)
#synonym augmentation
synonym_augmented_line = synonym.augment(context)
response_synonym.append(response)
context_synonym.append(synonym_augmented_line)
ocr_augmented_data = {'response': response_OCR, 'context': context_OCR}
ocr_df = pd.DataFrame.from_dict(ocr_augmented_data)
keyboard_augmented_data = {
'response': response_keyboard,
'context': context_keyboard
}
keyboard_df = pd.DataFrame.from_dict(keyboard_augmented_data)
synonym_augmented_data = {
'response': response_synonym,
'context': context_synonym
}
synonym_df = pd.DataFrame.from_dict(synonym_augmented_data)
augmented_1 = input_df.append(ocr_df, ignore_index=True)
augmented_2 = augmented_1.append(keyboard_df, ignore_index=True)
augmented_3 = augmented_2.append(synonym_df, ignore_index=True)
print('original dataset length: {}'.format(len(input_df)))
print('Augmented dataset length: {}'.format(len(augmented_2)))
return augmented_3
