File size: 14,308 Bytes
32b542e |
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 |
import os
import copy
import pickle
import random
import json
import glob
import numpy as np
from uniperceiver.config import configurable
from uniperceiver.functional import dict_as_tensor
from uniperceiver.tokenization import ClipTokenizer
from ..build import DATASETS_REGISTRY
import pyarrow as pa
__all__ = ["GLUEDataset"]
@DATASETS_REGISTRY.register()
class GLUEDataset:
@configurable
def __init__(
self,
cfg: dict,
stage: str,
anno_file: str,
max_seq_len: int,
tokenizer,
tokenizer_name,
input_columns,
label_column,
input_count,
task_name,
data_percentage,
data_k_sample,
):
self.cfg = cfg
self.stage = stage
self.anno_file = anno_file
self.tokenizer = tokenizer
self.tokenizer_name = tokenizer_name
self.max_seq_len = max_seq_len
self.input_columns = input_columns
self.label_column = label_column
self.input_count = input_count
self.task_name = task_name
self.data_percentage = data_percentage
self.data_k_sample = data_k_sample
self.task_info = {
'task_type' : self.cfg.DATASETS.TASK_TYPE,
'dataset_name' : self.cfg.DATASETS.DATASET_NAME,
'batch_size' : self.cfg.DATALOADER.TRAIN_BATCH_SIZE if self.stage == 'train' else self.cfg.DATALOADER.TEST_BATCH_SIZE,
'sampling_weight': self.cfg.DATALOADER.SAMPLING_WEIGHT,
}
self.target_set = cfg.DATASETS.TARGET_SET
self.load_data(cfg)
@classmethod
def from_config(cls, cfg, stage: str = "train"):
task_name = cfg.DATASETS.DATASET_NAME
namesmapping = {
"train": "train",
"val": "dev",
"test": "test",
}
data_dir = cfg.DATALOADER.ANNO_FOLDER
if task_name in ['MNLI', 'QNLI', 'QQP', 'RTE', 'SST-2', 'MRPC', 'CoLA', 'STS-B']:
anno_file = os.path.join(data_dir, task_name, 'processed/{name}.tsv'.format(name=namesmapping[stage]))
elif task_name == 'MNLI_Match':
namesmapping = {
"train": "train",
"val": "dev_matched",
"test": "test_matched",
}
anno_file = os.path.join(data_dir, 'MNLI', 'processed/{name}.tsv'.format(name=namesmapping[stage]))
elif task_name == 'MNLI_Mismatch':
namesmapping = {
"train": "train",
"val": "dev_mismatched",
"test": "test_mismatched",
}
anno_file = os.path.join(data_dir, 'MNLI', 'processed/{name}.tsv'.format(name=namesmapping[stage]))
input_count = 2
if task_name == "QQP":
input_columns = [4, 5]
if stage == 'test':
input_columns = [2, 3]
label_column = 6
elif task_name in ["MNLI_Match", "MNLI_Mismatch"]: # "MNLI" :
input_columns = [9, 10]
if stage == 'test':
input_columns = [9, 10]
label_column = 12
if stage == 'val':
label_column = 16
elif task_name == "QNLI":
input_columns = [2, 3]
if stage == 'test':
input_columns = [2, 3]
label_column = 4
elif task_name == "MRPC":
input_columns = [4, 5]
if stage == 'test':
input_columns = [4, 5]
label_column = 1
elif task_name == "RTE":
input_columns = [2, 3]
if stage == 'test':
input_columns = [2, 3]
label_column = 4
elif task_name == "STS-B":
input_columns = [8, 9]
if stage == 'test':
input_columns = [8, 9]
label_column = 10
# Following are single sentence tasks.
elif task_name == "SST-2":
input_columns = [1]
if stage == 'test':
input_columns = [2]
label_column = 2
input_count = 1
elif task_name == "CoLA":
input_columns = [4]
if stage == 'test':
input_columns = [2]
label_column = 2
input_count = 1
else:
raise NotImplementedError
ret = {
"cfg": cfg,
"stage": stage,
"anno_file": anno_file,
"max_seq_len": cfg.MODEL.MAX_SEQ_LEN,
"input_columns": input_columns,
"label_column": label_column,
"input_count": input_count,
"task_name": task_name,
"data_percentage": getattr(cfg.DATALOADER, "DATA_PERCENTAGE", 1.0),
"data_k_sample": getattr(cfg.DATALOADER, "DATA_K_SAMPLE", -1),
"tokenizer": ClipTokenizer(),
"tokenizer_name": "clip"
}
return ret
def load_data(self, cfg):
cache_path = os.path.join(os.path.dirname(self.anno_file), "cache_GLUE_raw_%s_%s_%s.pkl" % (self.task_name, self.tokenizer_name, self.stage))
if not os.path.exists(cache_path):
datalist = self.load_raw_data(cfg)
pickle.dump(datalist, open(cache_path, "wb"))
datalist = pickle.load(open(cache_path, "rb"))
# for few shot exp
if self.data_percentage < 1.0 and self.stage == "train":
print("will sample {} data for trianing-->".format(self.data_percentage))
labels2l = dict()
for data in datalist:
label = data['label']
if label not in labels2l:
labels2l[label] = list()
labels2l[label].append(data)
# samplers_label = len(datalist) * self.data_percentage // len(labels2l.keys())
datalist = []
for v in labels2l.values():
datalist.extend(random.sample(v, k=int(self.data_percentage * len(v) + 1)))
# datalist.extend(random.sample(v, k=int(samplers_label+1)))
elif self.data_k_sample > 0 and self.stage == "train":
print("will sample {} data for each class when training -->".format(self.data_k_sample))
labels2l = dict()
for data in datalist:
label = data['label']
if label not in labels2l:
labels2l[label] = list()
labels2l[label].append(data)
datalist = []
for v in labels2l.values():
datalist.extend(random.sample(v, k=int(self.data_k_sample)))
while len(datalist) < 200:
datalist = datalist + datalist
self.datalist = datalist
def load_raw_data(self, cfg):
datalist = []
if self.task_name.startswith("MNLI"):
labelmapping = {
"contradiction": 0,
"neutral": 1,
"entailment": 2,
}
fin = open(self.anno_file, 'r').readlines()
for _, line in enumerate(fin):
sensinfo = line.strip().split('\t')
if self.task_name == "RTE":
label = 1.0 if sensinfo[self.label_column - 1] == "entailment" else 0.0
elif self.task_name.startswith("MNLI"):
label = labelmapping[sensinfo[self.label_column - 1]]
elif self.task_name == "QNLI":
label = 1.0 if sensinfo[self.label_column - 1] == "entailment" else 0.0
elif self.task_name == "STS-B":
label = float(sensinfo[self.label_column - 1]) / 5.0
else:
label = float(sensinfo[self.label_column - 1])
datalist.append({
# start index from 1 to 0
"sentences": [sensinfo[i - 1] for i in self.input_columns],
"label": label
})
return datalist
def __len__(self):
return len(self.datalist)
def __getitem__(self, index):
dataset_dict = copy.deepcopy(self.datalist[index])
sentences = dataset_dict['sentences']
# input1: SEN1, this sentence is (spe) input2: word choice: postive and negative
if self.input_count == 1:
if self.task_name == "SST-2":
tokens = self.tokenizer.encode(sentences[0] + " <|endoftext|> It is <|spe|>. <|endoftext|>")
elif self.task_name == "CoLA":
tokens = self.tokenizer.encode(sentences[0] + " This is <|spe|>. <|endoftext|>")
else:
raise NotImplementedError
index = len(tokens) - 3
assert index < self.max_seq_len
if len(tokens) > self.max_seq_len:
tokens = tokens[:self.max_seq_len - 4] + tokens[-4:]
else:
if self.task_name in ["RTE"]:
tokens1 = self.tokenizer.encode(sentences[0])
if tokens1[-1] == 269:
tokens1 = tokens1[:-1]
tokens1 = tokens1 + self.tokenizer.encode(" ? <|endoftext|> it is ")
tokens2 = self.tokenizer.encode(sentences[1] + " <|endoftext|> ")
tokens2 = self.tokenizer.encode(" <|spe|> , ") + tokens2
if len(tokens2) > self.max_seq_len // 2:
tokens2 = tokens2[:self.max_seq_len // 2 - 1] + [tokens2[-1]]
max_len = self.max_seq_len - len(tokens2)
elif self.task_name in ["MRPC"]:
tokens1 = self.tokenizer.encode(sentences[0])
if tokens1[-1] == 269:
tokens1 = tokens1[:-1]
tokens1 = tokens1 + self.tokenizer.encode(" . ")
tokens2 = self.tokenizer.encode(sentences[1] + " <|endoftext|> ")
tokens2 = self.tokenizer.encode(" <|spe|> , ") + tokens2
if len(tokens2) > self.max_seq_len // 2:
tokens2 = tokens2[:self.max_seq_len // 2 - 1] + [tokens2[-1]]
max_len = self.max_seq_len - len(tokens2)
elif self.task_name in ["QQP"]:
tokens1 = self.tokenizer.encode(sentences[0])
if tokens1[-1] == 269:
tokens1 = tokens1[:-1]
tokens1 = tokens1 + self.tokenizer.encode(" <|endoftext|> ")
tokens2 = self.tokenizer.encode(sentences[1] + " <|endoftext|> ")
tokens2 = self.tokenizer.encode(" <|spe|> , ") + tokens2
if len(tokens2) > self.max_seq_len // 2:
tokens2 = tokens2[:self.max_seq_len // 2 - 1] + [tokens2[-1]]
max_len = self.max_seq_len - len(tokens2)
elif self.task_name in ["QNLI"]:
tokens1 = self.tokenizer.encode(sentences[0])
if tokens1[-1] == 269:
tokens1 = tokens1[:-1]
tokens1 = tokens1 + self.tokenizer.encode(" <|endoftext|> it is ")
tokens2 = self.tokenizer.encode(sentences[1] + " <|endoftext|> ")
tokens2 = self.tokenizer.encode(" <|spe|> , ") + tokens2
if len(tokens2) > self.max_seq_len - len(tokens1):
tokens2 = tokens2[:self.max_seq_len - len(tokens1) - 1] + [tokens2[-1]]
max_len = self.max_seq_len - len(tokens2)
elif self.task_name in ["MNLI", "MNLI_Match"]:
# sentence0 = sentences[0].replace(")", "").replace("(", "")
tokens1 = self.tokenizer.encode(sentences[0])
# if tokens1[-1] == 269:
# tokens1 = tokens1[:-1]
tokens1 = tokens1 # + self.tokenizer.encode(" ? ")
tokens2 = self.tokenizer.encode(sentences[1] + " <|endoftext|> ")
tokens2 = self.tokenizer.encode(" <|spe|> , ") + tokens2
if len(tokens2) > self.max_seq_len // 2:
tokens2 = tokens2[:self.max_seq_len // 2 - 1] + [tokens2[-1]]
max_len = self.max_seq_len - len(tokens2)
elif self.task_name in ["RTE", "QNLI", "MNLI", "MNLI_Match"]:
tokens1 = self.tokenizer.encode(sentences[0] + "? <|endoftext|>")
tokens2 = self.tokenizer.encode(sentences[1] + " <|endoftext|> ")
if tokens1[-1] == 269:
tokens1 = tokens1[:-1]
tokens2 = self.tokenizer.encode(" <|spe|> , ") + tokens2
if len(tokens2) > self.max_seq_len // 2:
tokens2 = tokens2[:self.max_seq_len // 2 - 1] + [tokens2[-1]]
max_len = self.max_seq_len - len(tokens2)
elif self.task_name in ["MRPC", "QQP"]:
tokens1 = self.tokenizer.encode(sentences[0] + " <|endoftext|>")
tokens2 = self.tokenizer.encode(sentences[1] + " <|endoftext|> ")
tokens2 = self.tokenizer.encode(" <|spe|>, ") + tokens2
if len(tokens2) > self.max_seq_len // 2:
tokens2 = tokens2[:self.max_seq_len // 2 - 1] + [tokens2[-1]]
max_len = self.max_seq_len - len(tokens2)
else:
NotImplementedError
# tokens = self.tokenizer.add_special_tokens_sentences_pair(tokens1, tokens2, start_type='SPE')
if len(tokens1) > max_len:
tokens1 = tokens1[:max_len - 1] + [tokens1[-1]]
tokens = tokens1 + tokens2
index = len(tokens1)
assert index < self.max_seq_len
sentences = np.array(tokens, dtype=np.int64)
if self.task_name in ["SST-2", "CoLA", "MRPC", "RTE", "QNLI", "MNLI", "QQP", "MNLI_Match"]:
label = int(dataset_dict['label'])
else:
raise NotImplementedError()
ret = {
'input_sample': [{
'data': [sentences],
'modality': 'text',
'data_type': 'input',
'invalid_mask': None,
'sample_info' : {
'spe_index': index
}
}],
'target_sample': [],
'target_idx' : [label],
'target_set' : copy.deepcopy(self.target_set),
'task_info' : copy.deepcopy(self.task_info)
}
dict_as_tensor(ret)
return ret
|