-
Notifications
You must be signed in to change notification settings - Fork 0
/
corpus_dataset.py
211 lines (195 loc) · 7.84 KB
/
corpus_dataset.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
202
203
204
205
206
207
208
209
210
211
import os, sys, ljqpy, random
import zipfile
import numpy as np
import jieba_fast as jieba
datadir = '/mnt/data122/datasets/CLUECorpus2020(100G)'
wikifn = '/mnt/data122/datasets/WikiCorpus/wiki.txt'
pilefn = '/data1/datas/pile/test'
if not os.path.exists(datadir):
datadir = '/data1/datas/clue/train/'
if not os.path.exists(wikifn):
wikifn = '/data1/ljq/wikicorpus/wiki.txt'
def LoadListg(fn):
with open(fn, encoding="utf-8") as fin:
for ll in fin:
yield ll.strip()
def MixedSentences():
zips = ljqpy.ListDirFiles(datadir, lambda x:'train_clue' in x and x.endswith('.zip'))
if os.path.exists(wikifn): zips.append('wiki')
if os.path.exists(pilefn): zips.extend(ljqpy.ListDirFiles(pilefn, lambda x:x.endswith('.jsonl'))[:2])
while True:
random.shuffle(zips)
for zipfn in zips:
if zipfn == 'wiki':
for x in LoadListg(wikifn): yield x
yield ''
elif 'pile' in zipfn:
for x in ljqpy.LoadJsonsg(zipfn):
lines = [z for z in x['text'].strip().split('\n') if z != '']
for sent in lines: yield sent
yield ''
else:
with zipfile.ZipFile(zipfn, 'r') as fzip:
ilist = fzip.infolist()
random.shuffle(ilist)
for zinfo in ilist:
with fzip.open(zinfo.filename, 'r') as fin:
for x in fin: yield x.decode().strip()
yield ''
import torch
class PureGenDataset(torch.utils.data.Dataset):
def __init__(self, gen, num=100000) -> None:
self.gen = gen
self.num = num
def __len__(self): return self.num
def __getitem__(self, k):
return next(self.gen)
def wwm_encode(text, tokenizer, mask=103):
words = jieba.lcut(text)
rands = np.random.random(len(words))
source, target = [], []
for r, w in zip(rands, words):
ids = tokenizer.encode(w, add_special_tokens=False)
if r < 0.15 * 0.8:
source.extend([mask] * len(ids))
target.extend(ids)
elif r < 0.15 * 0.9:
source.extend(ids)
target.extend(ids)
elif r < 0.15:
source.extend(np.random.choice(tokenizer.vocab_size-1, size=len(ids)) + 1)
target.extend(ids)
else:
source.extend(ids)
target.extend([-100] * len(ids))
return source, target
def PadLists(xs, ps, maxlen):
lacklen = maxlen - len(xs[0])
if lacklen > 0:
for i in range(len(xs)):
xs[i].extend([ps[i]] * lacklen)
return xs
def CutLists(xs, ps, maxlen):
for i in range(len(xs)):
xs[i] = xs[i][:maxlen-1] + [ps[i]]
return xs
def RoBERTaFullSent(gen, tokenizer, maxlen=512, cls=101, sep=102, mask=103, ypad=-100):
xx, yy = [cls], [ypad];
segs, seg = [0], 0
while True:
text = next(gen)
source, target = wwm_encode(text, tokenizer, mask)
if len(xx) + len(source) + 1 > maxlen and len(xx) > 1:
xx, yy, segs = CutLists([xx, yy, segs], [sep, ypad, 0], maxlen)
xx, yy, segs = PadLists([xx, yy, segs], [0, ypad, 0], maxlen)
yield {'input_ids':xx, 'labels':yy, 'token_type_ids':segs}
xx, yy = [cls], [ypad]
segs, seg = [0], 0
if seg == 0 and len(xx) > 1 and random.random() < 0.3:
xx.append(sep); yy.append(ypad); segs.append(seg)
seg = 1
xx.extend(source); yy.extend(target)
segs.extend([seg] * len(source))
def wwm_info(text, tokenizer):
words = jieba.lcut(text)
source, wwm = [], []
for w in words:
if len(w) == 1: ids = [tokenizer.convert_single_token(w)]
else: ids = tokenizer.encode(w, add_special_tokens=False)
source.extend(ids)
wwm.extend([1]+[2]*(len(ids)-1))
return source, wwm
def wwm_mask(xx, wwms, tokenizer, mask=103, prob=0.15):
newx, newy = [], []
words = []
for i, w in enumerate(wwms):
if w == 2: words[-1] = i+1
else: words.extend([i, i+1])
words = [(words[i], words[i+1]) for i in range(0, len(words), 2)]
rands = np.random.random(len(words))
for r, (u, v) in zip(rands, words):
ids = xx[u:v]
if wwms[u] == 0: r = 100 # dont mask special tokens
if r < prob * 0.8:
newx.extend([mask] * len(ids))
newy.extend(ids)
elif r < prob * 0.9:
newx.extend(ids)
newy.extend(ids)
elif r < prob:
newx.extend(np.random.choice(tokenizer.vocab_size-1, size=len(ids)) + 1)
newy.extend(ids)
else:
newx.extend(ids)
newy.extend([-100] * len(ids))
return newx, newy
def RoBERTaFullSentFast(gen, tokenizer, maxlen=512, repeat=3, cls=101, sep=102):
xx, segs, seg, wwms = [cls], [0], 0, [0]
newpara = False
while True:
text = next(gen)
if text == '':
newpara = True
continue
source, wwm = wwm_info(text, tokenizer)
if len(xx) + len(source) + 1 > maxlen and len(xx) > 1:
xx, segs, wwms = CutLists([xx, segs, wwms], [sep, seg, 0], maxlen)
xx, segs, wwms = PadLists([xx, segs, wwms], [0, 0, 0], maxlen)
for _ in range(repeat):
newx, newy = wwm_mask(xx, wwms, tokenizer)
yield {'input_ids':newx, 'labels':newy, 'token_type_ids':segs}
xx, segs, seg, wwms = [cls], [0], 0, [0]
if newpara:
xx.append(sep); segs.append(seg); wwms.append(0); seg = 1
newpara = False
xx.extend(source); segs.extend([seg] * len(source))
wwms.extend(wwm)
def NSPGenerator(gen, tokenizer, maxlen=512, cls=101, sep=102, repeat=0):
pools = []
for _ in range(3):
z = next(gen)
if z != '': pools.append(z)
continuous = False
while True:
text = next(gen)
if text == '':
continuous = False
continue
pools.append(text)
if len(pools) > 100000: pools = pools[10000:20000] + pools[-10000:]
nsp_label = random.randint(0, 1)
if not continuous: nsp_label = 0
textb = text
if nsp_label == 1: texta = pools[-2]
else: texta = random.choice(pools[:-2])
xx, segs, wwms = [cls], [0], [0]
source, wwm = wwm_info(texta, tokenizer)
xx.extend(source); segs.extend([0] * len(source)); wwms.extend(wwm)
xx.append(sep); segs.append(0); wwms.append(0)
source, wwm = wwm_info(textb, tokenizer)
xx.extend(source); segs.extend([1] * len(source)); wwms.extend(wwm)
xx, segs, wwms = CutLists([xx, segs, wwms], [sep, 1, 0], maxlen)
xx, segs, wwms = PadLists([xx, segs, wwms], [0, 0, 0], maxlen)
newx, newy = wwm_mask(xx, wwms, tokenizer)
yield {'input_ids':newx, 'labels':newy, 'token_type_ids':segs, 'nsp_label':nsp_label}
continuous = True
if __name__ == '__main__':
from cetokenizer import CEBertTokenizerFast
#tokenizer = BertTokenizer.from_pretrained('hfl/chinese-roberta-wwm-ext')
tokenizer = CEBertTokenizerFast('vocab.txt')
#rr = wwm_encode('努力！未来 lucky star！normalization normalization normalization', tokenizer)
#print(rr)
#print(tokenizer.convert_ids_to_string(rr[0]))
#sg = MixedSentences()
#for i in range(10):
# s = next(sg)
# print(s[:43])
#sys.exit()
ds = PureGenDataset(NSPGenerator(MixedSentences(), tokenizer, maxlen=256, repeat=1), 3)
#print(len(ds))
for i in range(3):
z = ds[i]
print('')
print(z['nsp_label'])
print(tokenizer.convert_ids_to_string(z['input_ids'], no_pad=True))
print('done')