-
Notifications
You must be signed in to change notification settings - Fork 3
/
corpus.py
201 lines (166 loc) · 6.71 KB
/
corpus.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
import os
import re
import json
from collections import Counter, namedtuple, defaultdict
import pickle
from tqdm import tqdm
import torch
from torch.nn.utils.rnn import pad_sequence
import nltk
from pytorch_pretrained_bert import BertTokenizer
Element = namedtuple('Element', ['id', 'text', 'author', 'timestep'])
class Corpus(object):
bos_token = '<bos>'
eos_token = '<eos>'
pad_token = '<pad>'
unk_token = '<unk>'
def __init__(self, examples, ids, vocab, authors):
self.examples = examples
self.ids = ids
self.vocab = vocab
self.vocab_size = vocab.size
self.nwords = sum(len(ex.text) for ex in self.examples)
self.authors = authors
self.na = len(authors.i2s)
self.nt = len(set([ex.timestep for ex in examples]))
self._init_specials()
def __len__(self):
return len(self.examples)
def __getitem__(self, index):
x = self.examples[index]
text = torch.LongTensor(self.vocab.convert_tokens_to_ids(x.text))
return Element(x.id, text, self.authors.s2i[x.author], x.timestep)
def post_process(self, text):
return text
def filter_ids(self, ids):
examples = list(filter(lambda x: x.id in ids, self.examples))
return self.__class__(examples, self.ids, self.vocab, self.authors)
def _init_specials(self):
specials = [Corpus.pad_token, Corpus.unk_token, Corpus.bos_token, Corpus.eos_token]
Corpus.pad_id, Corpus.unk_id, Corpus.bos_id, Corpus.eos_id = self.vocab.convert_tokens_to_ids(specials)
assert Corpus.unk_token in self.vocab.i2s and Corpus.unk_id == 1
@classmethod
def load_corpus(cls, data_dir, **kwargs):
pckl_path = os.path.join(data_dir, 'corpus.pkl')
if os.path.isfile(pckl_path):
print(f'Loading corpus at {pckl_path}...')
with open(pckl_path, 'rb') as f:
return cls(*pickle.load(f))
# corpus
fpath = os.path.join(data_dir, 'corpus.json')
print(f'Loading corpus at {fpath}...')
data = {}
with open(fpath, 'r') as f:
for l in f.read().splitlines():
ex = json.loads(l)
data[ex['id']] = ex
# fields
with open(os.path.join(data_dir, 'train.txt'), 'r') as f:
train_ids = set(f.read().splitlines())
trainset = list(filter(lambda x: x['id'] in train_ids, data.values()))
# -- ids
ids = Vocab(list(data.keys()))
# -- texts
vocab = cls._build_vocab(trainset, **kwargs)
# -- authors
authors = Vocab([e for ex in trainset for e in ex['authors']], specials=['<pad>'], unk_token=False)
# preprocess
print('preprocessing...')
examples = cls._preprocess(data, ids.i2s, vocab, authors)
with open(pckl_path, 'wb') as f:
pickle.dump((examples, ids, vocab, authors), f)
return cls(examples, ids, vocab, authors)
@classmethod
def _build_vocab(cls, data):
all_tokens = []
for ex in data:
text = ' '.join(text.strip().lower() for text in ex['texts'])
all_tokens += cls._tokenize(text)
specials = [cls.pad_token, cls.unk_token, cls.bos_token, cls.eos_token]
vocab = Vocab(all_tokens, min_freq=5, specials=specials)
return vocab
@classmethod
def _preprocess(cls, data, ids, text_vocab, authors_vocab):
# tokenize
examples = []
for i in tqdm(ids):
ex = data[i]
t = ex['timestep']
authors = ex['authors']
tokens = cls._preprocess_texts(ex['texts'], text_vocab)
if len(tokens) > 512:
continue
for a in authors:
examples.append(Element(i, tokens, a, t))
return examples
@classmethod
def _tokenize(self, text):
return ['N' if re.search(r'[0-9]', w) else w for w in nltk.tokenize.word_tokenize(text)]
@classmethod
def _preprocess_texts(cls, texts, vocab):
text = ' '.join([text.strip().lower() for text in texts])
tokens = [cls.bos_token] + cls._tokenize(text) + [cls.eos_token]
return tokens
class BertCorpus(Corpus):
pad_token = '[PAD]'
bos_token = '[CLS]'
eos_token = '[SEP]'
unk_token = '[UNK]'
def _init_wp(self, vocab):
self.is_full_word = torch.tensor([not wp.startswith('##') for wp in vocab.vocab.keys()])
def _init_specials(self):
specials = [BertCorpus.bos_token, BertCorpus.eos_token, BertCorpus.pad_token, BertCorpus.unk_token]
Corpus.bos_token, Corpus.eos_token, Corpus.pad_token, Corpus.unk_token = specials
Corpus.bos_id, Corpus.eos_id, Corpus.pad_id, Corpus.unk_id = self.vocab.convert_tokens_to_ids(specials)
def post_process(self, text):
out = ''
for word in text:
if word[:2] == '##':
out += word[2:]
else:
out += word
return out
@classmethod
def _build_vocab(cls, data, bert_cache_dir):
print(f'(Down)Loading BERT tokenizer at {bert_cache_dir}')
vocab = BertTokenizer.from_pretrained('bert-base-uncased', cache_dir=bert_cache_dir)
vocab.size = len(vocab.vocab)
return vocab
@classmethod
def _preprocess_texts(cls, texts, vocab):
text = cls.bos_token + ' ' + ' '.join(texts) + ' ' + cls.eos_token
tokens = vocab.tokenize(text)
return tokens
class Vocab():
default_id = None
def __init__(self, data, min_freq=0, specials=None, unk_token=True):
self.min_freq = min_freq
counter = Counter(data)
self.frequencies = sorted(counter.items(), key=lambda tup: tup[0])
self.frequencies.sort(key=lambda tup: tup[1], reverse=True)
self.specials = specials
self.i2s = [w for w in self.specials] if self.specials is not None else []
for word, freq in self.frequencies:
if freq < self.min_freq:
break
self.i2s.append(word)
if unk_token:
self.s2i = defaultdict(_default_id)
else:
self.s2i = {}
self.s2i.update({tok: i for i, tok in enumerate(self.i2s)})
self.size = len(self.i2s)
def convert_tokens_to_ids(self, tokens):
return [self.s2i[token] for token in tokens]
def convert_ids_to_tokens(self, ids):
return [self.i2s[i] for i in ids]
def __len__(self):
return len(self.i2s)
def text_collate(batch):
elements = Element(*zip(*batch))
texts = pad_sequence(elements.text, batch_first=False, padding_value=Corpus.pad_id)
authors = torch.tensor(elements.author)
timesteps = torch.tensor(elements.timestep)
return Element(elements.id, texts, authors, timesteps)
def _default_id():
return 1