-
Notifications
You must be signed in to change notification settings - Fork 0
/
data_lib.py
407 lines (355 loc) · 17.7 KB
/
data_lib.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
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
from typing import Any, List, Optional, Literal
import copy
import torch
from torch.utils.data import Dataset
from datasets import Dataset as HF_Dataset
from IPython import embed
from dataclasses import dataclass
import logging
logging.basicConfig(level=logging.INFO, format='%(asctime)s - %(name)s - %(levelname)s - %(message)s')
logger = logging.getLogger(__name__)
import random
import json
from tqdm import tqdm
import random
@dataclass
class ConvSearchSample:
sample_idx: str
raw_utt: str
manual_utt: str
history: List[str]
pos_docs: List[str]
neg_docs: List[str]
@dataclass
class ConvSearchSampleAug:
sample_idx: str
raw_utt: str
manual_utt: str
history: List[str]
pos_docs: List[str]
neg_docs: List[str]
aug_ctx: object
class ConvSearchDataset(Dataset):
def __init__(self,
data_path: str,
use_data_percent: float,
only_last_response: bool,
negative_type = 'in_batch_neg',
neg_ratio = 3):
self.negative_type = negative_type
self.neg_ratio = neg_ratio
self.data = self.read_data(data_path, use_data_percent, only_last_response)
def __len__(self):
return len(self.data)
def __getitem__(self, item):
return self.data[item]
def read_data(self, data_path: str, use_data_percent: float, only_last_response: bool):
samples = []
with open(data_path, 'r') as f:
data = json.load(f)
if use_data_percent < 1.0:
n = int(use_data_percent * len(data))
data = random.sample(data, n)
for conv in tqdm(data, desc="Processing {} data file...".format(data_path)):
history = []
for turn in conv['turns']:
sample_idx = "{}_{}".format(conv['conv_id'], turn['turn_id'])
raw_utt = turn['question']
manual_utt = ""
if "manual_rewrite" in turn:
manual_utt = turn['manual_rewrite']
pos_docs, neg_docs = [], []
if "pos_doc_text" in turn:
pos_docs = turn['pos_doc_text']
if self.negative_type == "random_neg":
assert "random_neg_docs_text" in turn
random_neg_docs = []
for doc in turn['random_neg_docs_text']:
if len(random_neg_docs) == self.neg_ratio:
break
random_neg_docs.append(doc)
neg_docs = random_neg_docs
elif self.negative_type == "bm25_hard_neg":
assert "bm25_hard_neg_docs_text" in turn
bm25hard_neg_docs = []
for doc in turn['bm25_hard_neg_docs_text']:
if len(bm25hard_neg_docs) == self.neg_ratio:
break
bm25hard_neg_docs.append(doc)
neg_docs = bm25hard_neg_docs
elif self.negative_type == "in_batch_neg":
neg_docs = []
else:
raise KeyError("Negative type: {} not implmeneted".format(self.negative_type))
sample = ConvSearchSample(sample_idx, raw_utt, manual_utt, copy.deepcopy(history), pos_docs, neg_docs)
samples.append(sample)
# update history
if only_last_response and "response" in turn and len(history) > 0:
history.pop() # pop the last response
history.append(raw_utt)
if "response" in turn:
history.append(turn['response'])
return samples
def filter_no_pos_sample(self):
filtered_data = []
for sample in self.data:
if len(sample.pos_docs) > 0:
filtered_data.append(sample)
self.data = filtered_data
logger.info("Filtered {} samples with no positive documents.".format(len(self.data)))
@staticmethod
def form_session_reverse_concat(history, raw_utt, tokenizer, max_seq_length):
to_concat = []
to_concat.extend(history)
to_concat.append(raw_utt)
to_concat.reverse()
sep_token = tokenizer.sep_token
if sep_token is None:
sep_token = tokenizer.eos_token
text = " {} ".format(sep_token).join(to_concat)
return text
@staticmethod
def train_cdr_collate_fn(batch: list, tokenizer, max_seq_length):
sample_ids = []
query_inputs = []
pos_doc_inputs = []
neg_doc_inputs = []
for sample in batch:
sample_ids.append(sample.sample_idx)
text = ConvSearchDataset.form_session_reverse_concat(sample.history, sample.raw_utt, tokenizer, max_seq_length)
query_inputs.append(text)
pos_doc_inputs += sample.pos_docs
neg_doc_inputs += sample.neg_docs
query_input_encodings = tokenizer(query_inputs, padding="longest", max_length=max_seq_length, truncation=True, return_tensors='pt')
pos_doc_input_encodings = tokenizer(pos_doc_inputs, padding="longest", max_length=max_seq_length, truncation=True, return_tensors='pt')
neg_doc_input_encodings = None
if len(neg_doc_inputs) > 0:
neg_doc_input_encodings = tokenizer(neg_doc_inputs, padding="longest", max_length=max_seq_length, truncation=True, return_tensors='pt')
return {'sample_ids': sample_ids,
'query_input_encodings': query_input_encodings,
'pos_doc_input_encodings': pos_doc_input_encodings,
'neg_doc_input_encodings': neg_doc_input_encodings,}
@staticmethod
def test_dr_collate_fn(batch: list, tokenizer, max_seq_length, input_type):
sample_ids = []
input_texts = []
for sample in batch:
sample_ids.append(sample.sample_idx)
if input_type == 'manual':
text = sample.manual_utt
elif input_type == 'raw':
text = sample.raw_utt
elif input_type == 'session':
text = ConvSearchDataset.form_session_reverse_concat(sample.history, sample.raw_utt, tokenizer, max_seq_length)
input_texts.append(text)
input_encodings = tokenizer(input_texts, padding="longest", max_length=max_seq_length, truncation=True, return_tensors='pt')
return {'sample_ids': sample_ids,
'input_ids': input_encodings['input_ids'],
'attention_mask': input_encodings['attention_mask']}
class ConvSearchDatasetAug(Dataset):
def __init__(self,
data_path: str,
use_data_percent: float,
only_last_response: bool,
negative_type = 'in_batch_neg',
neg_ratio = 3):
self.negative_type = negative_type
self.neg_ratio = neg_ratio
self.data = self.read_data(data_path, use_data_percent, only_last_response)
def __len__(self):
return len(self.data)
def __getitem__(self, item):
return self.data[item]
def read_data(self, data_path: str, use_data_percent: float, only_last_response: bool):
samples = []
with open(data_path, 'r') as f:
data = json.load(f)
if use_data_percent < 1.0:
n = int(use_data_percent * len(data))
data = random.sample(data, n)
for conv in tqdm(data, desc="Processing {} data file...".format(data_path)):
history = []
for turn in conv['turns']:
sample_idx = "{}_{}".format(conv['conv_id'], turn['turn_id'])
raw_utt = turn['question']
manual_utt = ""
if "manual_rewrite" in turn:
manual_utt = turn['manual_rewrite']
pos_docs, neg_docs = [], []
aug_ctx = {}
if "pos_doc_text" in turn:
pos_docs = turn['pos_doc_text']
if "neg_doc_text" in turn:
neg_docs = turn['neg_doc_text']
if "aug_ctx" in turn:
aug_ctx = turn['aug_ctx']
if self.negative_type == "random_neg":
assert "random_neg_docs_text" in turn
random_neg_docs = []
for doc in turn['random_neg_docs_text']:
random_neg_docs.append(doc)
if len(random_neg_docs) == self.neg_ratio:
break
neg_docs = random_neg_docs
elif self.negative_type == "bm25_hard_neg":
assert "bm25_hard_neg_docs_text" in turn
bm25hard_neg_docs = []
for doc in turn['bm25_hard_neg_docs_text']:
if len(bm25hard_neg_docs) == self.neg_ratio:
break
bm25hard_neg_docs.append(doc)
neg_docs = bm25hard_neg_docs
elif self.negative_type == "in_batch_neg":
neg_docs = []
else:
raise KeyError("Negative type: {} not implmeneted".format(self.negative_type))
sample = ConvSearchSampleAug(sample_idx, raw_utt, manual_utt, copy.deepcopy(history), pos_docs, neg_docs, aug_ctx)
samples.append(sample)
# update history
if only_last_response and "response" in turn and len(history) > 0:
history.pop() # pop the last response
history.append(raw_utt)
if "response" in turn:
history.append(turn['response'])
return samples
def filter_no_pos_sample(self):
filtered_data = []
for sample in self.data:
if len(sample.pos_docs) > 0:
filtered_data.append(sample)
self.data = filtered_data
logger.info("Filtered {} samples with no positive documents.".format(len(self.data)))
@staticmethod
def form_session_reverse_concat(history, raw_utt, tokenizer, max_seq_length):
to_concat = []
to_concat.extend(history)
to_concat.append(raw_utt)
to_concat.reverse()
sep_token = tokenizer.sep_token
if sep_token is None:
sep_token = tokenizer.eos_token
text = " {} ".format(sep_token).join(to_concat)
return text
@staticmethod
def train_aug_cdr_collate_fn(batch: list, tokenizer, max_seq_length, aug_strategies, neg_strategies=[], neg_aug_ratio=1, pos_aug_assign=None, neg_aug_assign=None):
sample_ids = []
query_inputs = []
oracle_inputs = []
pos_doc_inputs = []
neg_doc_inputs = []
aug_query_inputs1 = []
aug_query_inputs2 = []
neg_aug_query_inputs = []
need_replace = False
for sample in batch:
sample_ids.append(sample.sample_idx)
text = ConvSearchDatasetAug.form_session_reverse_concat(sample.history, sample.raw_utt, tokenizer, max_seq_length)
query_inputs.append(text)
oracle_inputs.append(sample.manual_utt)
pos_doc_inputs += sample.pos_docs
neg_doc_inputs += sample.neg_docs
if pos_aug_assign:
strategy1, strategy2 = pos_aug_assign[sample.sample_idx]
else:
if len(aug_strategies)>0:
aug_strategy = []
for aug in aug_strategies:
if aug in sample.aug_ctx and len(sample.aug_ctx[aug]) != 0:
aug_strategy.append(aug)
no_strategies = []
if int(sample.sample_idx.split('_')[1]) <= 2:
no_strategies.append('reorder')
no_strategies.append('reorder_depend')
if 'turn_deletion_depend' not in sample.aug_ctx and 'turn_deletion' not in aug_strategy:
aug_strategy.append('turn_deletion')
for strategy in no_strategies:
if strategy in aug_strategy:
aug_strategy.remove(strategy)
strategy1 = random.choice(aug_strategy)
strategy2 = random.choice(aug_strategy)
while strategy1 == strategy2:
strategy2 = random.choice(aug_strategy)
if strategy1 == "" and strategy2 == "":
aug_text1 = copy.deepcopy(text)
aug_query_inputs1.append(aug_text1)
aug_text2 = copy.deepcopy(text)
aug_query_inputs2.append(aug_text2)
elif len(aug_strategies)>0:
aug_text1 = ConvSearchDatasetAug.form_session_reverse_concat(sample.aug_ctx[strategy1][:-1], sample.aug_ctx[strategy1][-1], tokenizer, max_seq_length)
aug_query_inputs1.append(aug_text1)
aug_text2 = ConvSearchDatasetAug.form_session_reverse_concat(sample.aug_ctx[strategy2][:-1], sample.aug_ctx[strategy2][-1], tokenizer, max_seq_length)
aug_query_inputs2.append(aug_text2)
if neg_aug_assign:
sampled_neg_strategies = neg_aug_assign[sample.sample_idx]
else:
neg_strategy = []
for aug in neg_strategies:
if aug in sample.aug_ctx and len(sample.aug_ctx[aug]) != 0:
neg_strategy.append(aug)
if len(neg_strategy)==0 and neg_aug_ratio!=0:
neg_aug_query_inputs.append("empty")
need_replace = True
continue
if len(neg_strategy)<neg_aug_ratio:
neg_strategy = neg_strategy+[neg_strategy[0]]*(neg_aug_ratio-len(neg_strategy))
sampled_neg_strategies = random.sample(neg_strategy, neg_aug_ratio)
if len(sampled_neg_strategies)==0 and neg_aug_ratio!=0:
neg_aug_query_inputs.extend(["empty"]*neg_aug_ratio)
need_replace = True
continue
if len(sampled_neg_strategies)<neg_aug_ratio:
sampled_neg_strategies = sampled_neg_strategies+[sampled_neg_strategies[0]]*(neg_aug_ratio-len(sampled_neg_strategies))
for aug in sampled_neg_strategies:
neg_aug_ctx = sample.aug_ctx[aug]
neg_aug_text = ConvSearchDatasetAug.form_session_reverse_concat(neg_aug_ctx[:-1], neg_aug_ctx[-1], tokenizer, max_seq_length)
neg_aug_query_inputs.append(neg_aug_text)
replace = None
empty_index = []
if need_replace:
for ind, neg in enumerate(neg_aug_query_inputs):
if neg != "empty":
replace = neg
else:
empty_index.append(ind)
for ind in empty_index:
neg_aug_query_inputs[ind] = replace
query_input_encodings = tokenizer(query_inputs, padding="longest", max_length=max_seq_length, truncation=True, return_tensors='pt')
oracle_input_encodings = tokenizer(oracle_inputs, padding="longest", max_length=max_seq_length, truncation=True, return_tensors='pt')
pos_doc_input_encodings = tokenizer(pos_doc_inputs, padding="longest", max_length=max_seq_length, truncation=True, return_tensors='pt')
neg_doc_input_encodings = None
if len(neg_doc_inputs) > 0:
neg_doc_input_encodings = tokenizer(neg_doc_inputs, padding="longest", max_length=max_seq_length, truncation=True, return_tensors='pt')
aug_input_encodings1 = None
if len(aug_query_inputs1) > 0:
aug_input_encodings1 = tokenizer(aug_query_inputs1, padding="longest", max_length=max_seq_length, truncation=True, return_tensors='pt')
aug_input_encodings2 = None
if len(aug_query_inputs2) > 0:
aug_input_encodings2 = tokenizer(aug_query_inputs2, padding="longest", max_length=max_seq_length, truncation=True, return_tensors='pt')
neg_aug_input_encodings = None
if len(neg_aug_query_inputs) > 0:
neg_aug_input_encodings = tokenizer(neg_aug_query_inputs, padding="longest", max_length=max_seq_length, truncation=True, return_tensors='pt')
return {'sample_ids': sample_ids,
'query_input_encodings': query_input_encodings,
'aug_input_encodings1': aug_input_encodings1,
'aug_input_encodings2': aug_input_encodings2,
'pos_doc_input_encodings': pos_doc_input_encodings,
'neg_doc_input_encodings': neg_doc_input_encodings,
'neg_aug_input_encodings': neg_aug_input_encodings,
"oracle_input_encodings": oracle_input_encodings}
@staticmethod
def test_dr_collate_fn(batch: list, tokenizer, max_seq_length, input_type):
sample_ids = []
input_texts = []
for sample in batch:
sample_ids.append(sample.sample_idx)
if input_type == 'manual':
text = sample.manual_utt
elif input_type == 'raw':
text = sample.raw_utt
elif input_type == 'session':
text = ConvSearchDataset.form_session_reverse_concat(sample.history, sample.raw_utt, tokenizer, max_seq_length)
input_texts.append(text)
input_encodings = tokenizer(input_texts, padding="longest", max_length=max_seq_length, truncation=True, return_tensors='pt')
return {'sample_ids': sample_ids,
'input_ids': input_encodings['input_ids'],
'attention_mask': input_encodings['attention_mask']}