class Reuters_Dataset(TorchnlpDataset):
def __init__(self,
root: str,
normal_class=0,
tokenizer='spacy',
use_tfidf_weights=False,
append_sos=False,
append_eos=False,
clean_txt=False,
max_seq_len_prior=None):
super().__init__(root)
self.n_classes = 2 # 0: normal, 1: outlier
classes = [
'earn', 'acq', 'crude', 'trade', 'money-fx', 'interest', 'ship'
]
# classes_full_list = [
# 'acq', 'alum', 'barley', 'bop', 'carcass', 'castor-oil', 'cocoa', 'coconut', 'coconut-oil', 'coffee',
# 'copper', 'copra-cake', 'corn', 'cotton', 'cotton-oil', 'cpi', 'cpu', 'crude', 'dfl', 'dlr', 'dmk',
# 'earn', 'fuel', 'gas', 'gnp', 'gold', 'grain', 'groundnut', 'groundnut-oil', 'heat', 'hog', 'housing',
# 'income', 'instal-debt', 'interest', 'ipi', 'iron-steel', 'jet', 'jobs', 'l-cattle', 'lead', 'lei',
# 'lin-oil', 'livestock', 'lumber', 'meal-feed', 'money-fx', 'money-supply', 'naphtha', 'nat-gas', 'nickel',
# 'nkr', 'nzdlr', 'oat', 'oilseed', 'orange', 'palladium', 'palm-oil', 'palmkernel', 'pet-chem', 'platinum',
# 'potato', 'propane', 'rand', 'rape-oil', 'rapeseed', 'reserves', 'retail', 'rice', 'rubber', 'rye',
# 'ship', 'silver', 'sorghum', 'soy-meal', 'soy-oil', 'soybean', 'strategic-metal', 'sugar', 'sun-meal',
# 'sun-oil', 'sunseed', 'tea', 'tin', 'trade', 'veg-oil', 'wheat', 'wpi', 'yen', 'zinc'
# ]
self.normal_classes = [classes[normal_class]]
del classes[normal_class]
self.outlier_classes = classes
# Load the reuters dataset
self.train_set, self.test_set = reuters_dataset(directory=root,
train=True,
test=True,
clean_txt=clean_txt)
# Pre-process
self.train_set.columns.add('index')
self.test_set.columns.add('index')
self.train_set.columns.add('weight')
self.test_set.columns.add('weight')
train_idx_normal = [] # for subsetting train_set to normal class
for i, row in enumerate(self.train_set):
if any(label in self.normal_classes
for label in row['label']) and (len(row['label']) == 1):
train_idx_normal.append(i)
row['label'] = torch.tensor(0)
else:
row['label'] = torch.tensor(1)
row['text'] = row['text'].lower()
test_idx = [
] # for subsetting test_set to selected normal and anomalous classes
test_n_idx = [] # subsetting test_set to selected normal classes
test_a_idx = [] # subsetting test_set to selected anomalous classes
for i, row in enumerate(self.test_set):
if any(label in self.normal_classes
for label in row['label']) and (len(row['label']) == 1):
test_idx.append(i)
test_n_idx.append(i)
row['label'] = torch.tensor(0)
elif any(label in self.outlier_classes
for label in row['label']) and (len(row['label']) == 1):
test_idx.append(i)
test_a_idx.append(i)
row['label'] = torch.tensor(1)
else:
row['label'] = torch.tensor(1)
row['text'] = row['text'].lower()
# Subset train_set to normal class
self.train_set = Subset(self.train_set, train_idx_normal)
# Subset test_set to selected normal and anomalous classes
self.test_set = Subset(self.test_set, test_idx)
# Subset test_set to selected normal classes
self.test_n_set = Subset(self.test_set, test_n_idx)
# Subset test_set to selected anomalous classes
self.test_a_set = Subset(self.test_set, test_a_idx)
# Make corpus and set encoder
text_corpus = [
row['text']
for row in datasets_iterator(self.train_set, self.test_set)
]
if tokenizer == 'spacy':
self.encoder = SpacyEncoder(text_corpus,
min_occurrences=3,
append_eos=append_eos)
if tokenizer == 'bert':
self.encoder = MyBertTokenizer.from_pretrained('bert-base-uncased',
cache_dir=root)
# Encode
self.max_seq_len = 0
for row in datasets_iterator(self.train_set, self.test_set):
if append_sos:
sos_id = self.encoder.stoi[DEFAULT_SOS_TOKEN]
row['text'] = torch.cat((torch.tensor(sos_id).unsqueeze(0),
self.encoder.encode(row['text'])))
else:
row['text'] = self.encoder.encode(row['text'])
if len(row['text']) > self.max_seq_len:
self.max_seq_len = len(row['text'])
# Compute tf-idf weights
if use_tfidf_weights:
compute_tfidf_weights(self.train_set,
self.test_set,
vocab_size=self.encoder.vocab_size)
else:
for row in datasets_iterator(self.train_set, self.test_set):
row['weight'] = torch.empty(0)
# Get indices after pre-processing
for i, row in enumerate(self.train_set):
row['index'] = i
for i, row in enumerate(self.test_set):
row['index'] = i
# # length prior
sent_lengths = [len(row['text']) for row in self.train_set]
sent_lengths_freq = np.bincount(np.array(sent_lengths))
sent_lengths_freq = np.concatenate(
(sent_lengths_freq,
np.array((max_seq_len_prior - max(sent_lengths)) * [0])),
axis=0)
sent_lengths_freq = sent_lengths_freq + 1
self.length_prior = np.log(sent_lengths_freq) - np.log(
sent_lengths_freq.sum())
class IMDB_Dataset(TorchnlpDataset):
def __init__(self,
root: str,
normal_class=0,
tokenizer='spacy',
use_tfidf_weights=False,
append_sos=False,
append_eos=False,
clean_txt=False,
max_seq_len_prior=None):
super().__init__(root)
self.n_classes = 2 # 0: normal, 1: outlier
classes = ['pos', 'neg']
if normal_class == -1:
self.normal_classes = classes
self.outlier_classes = []
else:
self.normal_classes = [classes[normal_class]]
del classes[normal_class]
self.outlier_classes = classes
if root not in nltk.data.path:
nltk.data.path.append(root)
# Load the imdb dataset
self.train_set, self.test_set = imdb_dataset(directory=root,
train=True,
test=True)
# Pre-process
self.train_set.columns.add('index')
self.test_set.columns.add('index')
self.train_set.columns.remove('sentiment')
self.test_set.columns.remove('sentiment')
self.train_set.columns.add('label')
self.test_set.columns.add('label')
self.train_set.columns.add('weight')
self.test_set.columns.add('weight')
train_idx_normal = [] # for subsetting train_set to normal class
for i, row in enumerate(self.train_set):
row['label'] = row.pop('sentiment')
if row['label'] in self.normal_classes:
train_idx_normal.append(i)
row['label'] = torch.tensor(0)
else:
row['label'] = torch.tensor(1)
if clean_txt:
row['text'] = clean_text(row['text'].lower())
else:
row['text'] = row['text'].lower()
test_n_idx = [] # subsetting test_set to selected normal classes
test_a_idx = [] # subsetting test_set to selected anomalous classes
for i, row in enumerate(self.test_set):
row['label'] = row.pop('sentiment')
if row['label'] in self.normal_classes:
test_n_idx.append(i)
else:
test_a_idx.append(i)
row['label'] = torch.tensor(
0) if row['label'] in self.normal_classes else torch.tensor(1)
if clean_txt:
row['text'] = clean_text(row['text'].lower())
else:
row['text'] = row['text'].lower()
# Subset train_set to normal class
self.train_set = Subset(self.train_set, train_idx_normal)
# Subset test_set to selected normal classes
self.test_n_set = Subset(self.test_set, test_n_idx)
# Subset test_set to selected anomalous classes
self.test_a_set = Subset(self.test_set, test_a_idx)
# Make corpus and set encoder
text_corpus = [
row['text']
for row in datasets_iterator(self.train_set, self.test_set)
]
if tokenizer == 'spacy':
self.encoder = SpacyEncoder(text_corpus,
min_occurrences=3,
append_eos=append_eos)
if tokenizer == 'bert':
self.encoder = MyBertTokenizer.from_pretrained('bert-base-uncased',
cache_dir=root)
# Encode
self.max_seq_len = 0
for row in datasets_iterator(self.train_set, self.test_set):
if append_sos:
sos_id = self.encoder.stoi[DEFAULT_SOS_TOKEN]
row['text'] = torch.cat((torch.tensor(sos_id).unsqueeze(0),
self.encoder.encode(row['text'])))
else:
row['text'] = self.encoder.encode(row['text'])
if len(row['text']) > self.max_seq_len:
self.max_seq_len = len(row['text'])
# Compute tf-idf weights
if use_tfidf_weights:
compute_tfidf_weights(self.train_set,
self.test_set,
vocab_size=self.encoder.vocab_size)
else:
for row in datasets_iterator(self.train_set, self.test_set):
row['weight'] = torch.empty(0)
# Get indices after pre-processing
for i, row in enumerate(self.train_set):
row['index'] = i
for i, row in enumerate(self.test_set):
row['index'] = i
# length prior
sent_lengths = [len(row['text']) for row in self.train_set]
sent_lengths_freq = np.bincount(np.array(sent_lengths))
sent_lengths_freq = np.concatenate(
(sent_lengths_freq,
np.array((max_seq_len_prior - max(sent_lengths)) * [0])),
axis=0)
sent_lengths_freq = sent_lengths_freq + 1
self.length_prior = np.log(sent_lengths_freq) - np.log(
sent_lengths_freq.sum())
class Newsgroups20_Dataset(TorchnlpDataset):
def __init__(self,
root: str,
normal_class=0,
tokenizer='spacy',
use_tfidf_weights=False,
append_sos=False,
append_eos=False,
clean_txt=False,
max_seq_len_prior=None):
super().__init__(root)
self.n_classes = 2 # 0: normal, 1: outlier
classes = list(range(6))
groups = [[
'comp.graphics', 'comp.os.ms-windows.misc',
'comp.sys.ibm.pc.hardware', 'comp.sys.mac.hardware',
'comp.windows.x'
],
[
'rec.autos', 'rec.motorcycles', 'rec.sport.baseball',
'rec.sport.hockey'
], ['sci.crypt', 'sci.electronics', 'sci.med', 'sci.space'],
['misc.forsale'],
[
'talk.politics.misc', 'talk.politics.guns',
'talk.politics.mideast'
],
[
'talk.religion.misc', 'alt.atheism',
'soc.religion.christian'
]]
short_group_names = ['comp', 'rec', 'sci', 'misc', 'pol', 'rel']
self.subset = short_group_names[normal_class]
self.normal_classes = groups[normal_class]
self.outlier_classes = []
del classes[normal_class]
for i in classes:
self.outlier_classes += groups[i]
# Load the 20 Newsgroups dataset
self.train_set, self.test_set = newsgroups20_dataset(
directory=root,
train=True,
test=True,
clean_txt=clean_txt,
groups=groups,
short_group_names=short_group_names)
# Pre-process
self.train_set.columns.add('index')
self.test_set.columns.add('index')
self.train_set.columns.add('weight')
self.test_set.columns.add('weight')
train_idx_normal = [] # for subsetting train_set to normal class
for i, row in enumerate(self.train_set):
if row['label'] in self.normal_classes:
train_idx_normal.append(i)
row['label'] = torch.tensor(0)
else:
row['label'] = torch.tensor(1)
row['text'] = row['text'].lower()
test_n_idx = [] # subsetting test_set to selected normal classes
test_a_idx = [] # subsetting test_set to selected anomalous classes
for i, row in enumerate(self.test_set):
if row['label'] in self.normal_classes:
test_n_idx.append(i)
else:
test_a_idx.append(i)
row['label'] = torch.tensor(
0) if row['label'] in self.normal_classes else torch.tensor(1)
row['text'] = row['text'].lower()
# Subset train_set to normal class
self.train_set = Subset(self.train_set, train_idx_normal)
# Subset test_set to selected normal classes
self.test_n_set = Subset(self.test_set, test_n_idx)
# Subset test_set to selected anomalous classes
self.test_a_set = Subset(self.test_set, test_a_idx)
# Make corpus and set encoder
text_corpus = [
row['text']
for row in datasets_iterator(self.train_set, self.test_set)
]
if tokenizer == 'spacy':
self.encoder = SpacyEncoder(text_corpus,
min_occurrences=3,
append_eos=append_eos)
if tokenizer == 'bert':
self.encoder = MyBertTokenizer.from_pretrained('bert-base-uncased',
cache_dir=root)
# Encode
self.max_seq_len = 0
for row in datasets_iterator(self.train_set, self.test_set):
if append_sos:
sos_id = self.encoder.stoi[DEFAULT_SOS_TOKEN]
row['text'] = torch.cat((torch.tensor(sos_id).unsqueeze(0),
self.encoder.encode(row['text'])))
else:
row['text'] = self.encoder.encode(row['text'])
if len(row['text']) > self.max_seq_len:
self.max_seq_len = len(row['text'])
# Compute tf-idf weights
if use_tfidf_weights:
compute_tfidf_weights(self.train_set,
self.test_set,
vocab_size=self.encoder.vocab_size)
else:
for row in datasets_iterator(self.train_set, self.test_set):
row['weight'] = torch.empty(0)
# Get indices after pre-processing
for i, row in enumerate(self.train_set):
row['index'] = i
for i, row in enumerate(self.test_set):
row['index'] = i
# length prior
sent_lengths = [len(row['text']) for row in self.train_set]
sent_lengths_freq = np.bincount(np.array(sent_lengths))
sent_lengths_freq = np.concatenate(
(sent_lengths_freq,
np.array((max_seq_len_prior - max(sent_lengths)) * [0])),
axis=0)
sent_lengths_freq = sent_lengths_freq + 1
self.length_prior = np.log(sent_lengths_freq) - np.log(
sent_lengths_freq.sum())
class IMDB_Dataset(TorchnlpDataset):
def __init__(self, root: str, normal_class=0, tokenizer='spacy', use_tfidf_weights=False, append_sos=False,
append_eos=False, clean_txt=False):
super().__init__(root)
self.n_classes = 2 # 0: normal, 1: outlier
classes = ['pos', 'neg']
if normal_class == -1:
self.normal_classes = classes
self.outlier_classes = []
else:
self.normal_classes = [classes[normal_class]]
del classes[normal_class]
self.outlier_classes = classes
# Load the imdb dataset
self.train_set, self.test_set = imdb_dataset(directory=root, train=True, test=True)
# Pre-process
self.train_set.columns.add('index')
self.test_set.columns.add('index')
self.train_set.columns.remove('sentiment')
self.test_set.columns.remove('sentiment')
self.train_set.columns.add('label')
self.test_set.columns.add('label')
self.train_set.columns.add('weight')
self.test_set.columns.add('weight')
train_idx_normal = [] # for subsetting train_set to normal class
for i, row in enumerate(self.train_set):
row['label'] = row.pop('sentiment')
if row['label'] in self.normal_classes:
train_idx_normal.append(i)
row['label'] = torch.tensor(0)
else:
row['label'] = torch.tensor(1)
if clean_txt:
row['text'] = clean_text(row['text'].lower())
else:
row['text'] = row['text'].lower()
for i, row in enumerate(self.test_set):
row['label'] = row.pop('sentiment')
row['label'] = torch.tensor(0) if row['label'] in self.normal_classes else torch.tensor(1)
if clean_txt:
row['text'] = clean_text(row['text'].lower())
else:
row['text'] = row['text'].lower()
# Subset train_set to normal class
self.train_set = Subset(self.train_set, train_idx_normal)
# Make corpus and set encoder
text_corpus = [row['text'] for row in datasets_iterator(self.train_set, self.test_set)]
if tokenizer == 'spacy':
self.encoder = SpacyEncoder(text_corpus, min_occurrences=3, append_eos=append_eos)
if tokenizer == 'bert':
self.encoder = MyBertTokenizer.from_pretrained('bert-base-uncased', cache_dir=root)
# Encode
for row in datasets_iterator(self.train_set, self.test_set):
if append_sos:
sos_id = self.encoder.stoi[DEFAULT_SOS_TOKEN]
row['text'] = torch.cat((torch.tensor(sos_id).unsqueeze(0), self.encoder.encode(row['text'])))
else:
row['text'] = self.encoder.encode(row['text'])
# Compute tf-idf weights
if use_tfidf_weights:
compute_tfidf_weights(self.train_set, self.test_set, vocab_size=self.encoder.vocab_size)
else:
for row in datasets_iterator(self.train_set, self.test_set):
row['weight'] = torch.empty(0)
# Get indices after pre-processing
for i, row in enumerate(self.train_set):
row['index'] = i
for i, row in enumerate(self.test_set):
row['index'] = i
drop_indices = np.random.choice(idx_positive, nbr_to_drop, replace=False)
df = df.drop(drop_indices)
_logger.info((df['labels'] == 'negative').mean())
text_as_list = df['text'].tolist()
labels_as_list = df['labels'].tolist()
from torchnlp.encoders.text import SpacyEncoder, pad_tensor
from sklearn.model_selection import train_test_split
from tqdm.notebook import tqdm
# 使用pytorch NLP 工具对txt进行编码
encoder = SpacyEncoder(text_as_list)
_logger.info("{} encoder is {}".format(text_as_list[0],
encoder.encode(text_as_list[0])))
encoded_texts = []
for i in tqdm(range(len(text_as_list))):
encoded_texts.append(encoder.encode(text_as_list[i]))
lengths = [len(i) for i in tqdm(encoded_texts)]
import seaborn as sns
import matplotlib.pyplot as plt
length_as_series = pd.Series(lengths)
plt.title("Probability Density Function for text lengths")
sns.distplot(length_as_series)
max_pad_length = length_as_series.quantile(0.9)