File size: 19,533 Bytes
3cd1076 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 | #!/usr/bin/env python
# coding=utf-8
"""This Python code defines a class Dataset with methods for initializing, loading,
and manipulating datasets from different backends such as Hugging Face and JSON.
The `Dataset` class includes methods for loading datasets from a dictionary and a Hugging
Face dataset, mapping datasets, and retrieving the backend dataset and arguments.
"""
# Importing necessary libraries and modules
import copy
import json
import logging
from pathlib import Path
from cmath import e
from pathlib import Path
from typing import Optional, List
from datasets import load_dataset
from datasets import Dataset as HFDataset
from tqdm import tqdm
from lmflow.args import DatasetArguments
from lmflow.utils.constants import (
DATASET_DESCRIPTION_MAP,
TEXT_ONLY_DATASET_DESCRIPTION,
TEXT2TEXT_DATASET_DESCRIPTION,
FLOAT_ONLY_DATASET_DESCRIPTION,
INSTANCE_FIELDS_MAP,
)
from lmflow.utils.versioning import is_multimodal_available
from lmflow.utils.data_utils import get_dataset_type_fast, check_dataset_instances_key_fast
if is_multimodal_available():
from .multi_modal_dataset import CustomMultiModalDataset
logger = logging.getLogger(__name__)
DATASET_TYPES = [
"text_only",
"text2text",
"float_only",
"image_text",
"conversation",
"paired_conversation",
"paired_text_to_text",
"text_to_textlist",
"text_to_scored_textlist"
]
KEY_TYPE = "type"
KEY_INSTANCES = "instances"
KEY_SCORE = "score"
class Dataset:
r"""
Initializes the Dataset object with the given parameters.
Parameters
------------
data_args : DatasetArguments object.
Contains the arguments required to load the dataset.
backend : str, default="huggingface"
A string representing the dataset backend. Defaults to "huggingface".
args : Optional.
Positional arguments.
kwargs : Optional.
Keyword arguments.
"""
def __init__(self, data_args: DatasetArguments=None, backend: str="huggingface", *args, **kwargs):
self.data_args = data_args
self.backend = backend
self.backend_dataset = None
self.type = None # Original type of the dataset
self.dataset_path = data_args.dataset_path
if data_args.dataset_path is None:
return
if backend == "huggingface":
data_files = [
x.absolute().as_posix()
for x in Path(self.dataset_path).glob("*.json")
]
logger.info(f"Data files: \n{data_files}")
# check if the dataset is in the correct format and get the dataset type (text_only, text2text, etc.)
self._check_hf_json_format(data_files)
# Load the dataset using the HuggingFace dataset library
logger.info('Loading datasets')
extensions = "json"
raw_dataset = load_dataset(
extensions,
data_files=data_files,
field=KEY_INSTANCES,
split="train",
cache_dir=data_args.dataset_cache_dir,
)
self.backend_dataset = raw_dataset
self._check_instance_format()
elif backend == "json":
# TODO (@Jiachun)
pass
elif backend == "custom_multi_modal":
# FIXME refactor the backend name
if not is_multimodal_available():
raise ValueError(
'Multimodal not available. Please install via `pip install -e ".[multimodal]"`'
)
raw_dataset = CustomMultiModalDataset(self.dataset_path, data_args)
self.backend_dataset = raw_dataset
else:
raise NotImplementedError(f'Unsupported dataset backend "{backend}"')
def __len__(self):
return len(self.backend_dataset)
def _check_instance_format(self):
"""
Checks if data (instances) have required fields.
Raises messages with hints if not matched.
"""
fields = self.backend_dataset.features
correct_fields = INSTANCE_FIELDS_MAP[self.type]
if not set(correct_fields).issubset(set(fields)):
raise ValueError(
f'data instance fields incorrect'
f' {list(correct_fields)} are required.'
)
def _check_hf_json_format(self, data_files: List[str]):
for single_file in tqdm(data_files, desc='Checking dataset keys'):
# get type and check if it is consistent
json_data_type = get_dataset_type_fast(single_file)
if not json_data_type:
raise ValueError(
f'"{KEY_TYPE}" must be provided to initialize a dataset,'
f' e.g.\n'
f' {TEXT_ONLY_DATASET_DESCRIPTION}'
)
if self.type is None:
self.type = json_data_type
elif self.type != json_data_type:
raise ValueError(
'All task files must have same data types. Previous'
f' files have type "{self.type}", but in file'
f' {single_file}, it has type "{self.type}".'
)
# check if instances key is provided
key_instances_exists_flag = check_dataset_instances_key_fast(single_file, KEY_INSTANCES)
if not key_instances_exists_flag:
raise ValueError(
f'"{KEY_INSTANCES}" must be provided to initialize a'
f' dataset, e.g.\n'
f' {TEXT_ONLY_DATASET_DESCRIPTION}'
)
def from_dict(self, dict_obj: dict, *args, **kwargs):
r"""
Create a Dataset object from a dictionary.
Return a Dataset given a dict with format:
{
"type": TYPE,
"instances": [
{
"key_1": VALUE_1.1,
"key_2": VALUE_1.2,
...
},
{
"key_1": VALUE_2.1,
"key_2": VALUE_2.2,
...
},
...
]
}
Parameters
-----------
dict_obj : dict.
A dictionary containing the dataset information.
args : Optional.
Positional arguments.
kwargs : Optional.
Keyword arguments.
Returns
---------
self : Dataset object.
"""
if self.backend == "huggingface":
if KEY_TYPE not in dict_obj:
raise ValueError(
f'"{KEY_TYPE}" must be provided to initialize a dataset,'
f' e.g.\n'
f' {TEXT_ONLY_DATASET_DESCRIPTION}'
)
if KEY_INSTANCES not in dict_obj:
raise ValueError(
f'"{KEY_INSTANCES}" must be provided to initialize a'
f' dataset, e.g.\n'
f' {TEXT_ONLY_DATASET_DESCRIPTION}'
)
self.type = dict_obj[KEY_TYPE]
if not self.type in INSTANCE_FIELDS_MAP:
raise ValueError(f'type "{self.type}" is not supported')
correct_fields = INSTANCE_FIELDS_MAP[self.type]
for i, instance in enumerate(dict_obj[KEY_INSTANCES]):
fields = instance.keys()
if not set(correct_fields).issubset(set(fields)):
raise ValueError(
f'data instance fields incorrect'
f' {list(correct_fields)} are required.'
)
try:
hf_dict = {}
if len(dict_obj[KEY_INSTANCES]) > 0:
for key in dict_obj[KEY_INSTANCES][0].keys():
hf_dict[key] = [
instance[key] for instance in dict_obj[KEY_INSTANCES]
]
self.backend_dataset = HFDataset.from_dict(hf_dict, *args, **kwargs)
except AttributeError as ex:
raise ValueError(
f"Error occurs: {ex}. Failed to convert dict to"
f" \"{self.type}\" dataset," f" the standard format is as"
f" follows:\n"
f" {DATASET_DESCRIPTION_MAP[self.type]}"
)
self._check_instance_format()
return self
elif self.backend == "dict":
self.backend_dataset = dict_obj
self.type = dict_obj[KEY_TYPE]
return self
else:
raise NotImplementedError(
f'Currently .from_dict is not supported for backend "{self.backend}"'
)
@classmethod
def create_from_dict(cls, dict_obj, *args, **kwargs):
r"""
Returns
--------
Returns a Dataset object given a dict.
"""
empty_data_args = DatasetArguments(dataset_path=None)
dataset = Dataset(empty_data_args)
return dataset.from_dict(dict_obj)
def to_dict(self):
r"""
Returns
---------
Return a dict represents the dataset:
{
"type": TYPE,
"instances": [
{
"key_1": VALUE_1.1,
"key_2": VALUE_1.2,
...
},
{
"key_1": VALUE_2.1,
"key_2": VALUE_2.2,
...
},
...
]
}
A python dict object represents the content of this dataset.
"""
if self.backend == "huggingface":
dict_obj = {}
dict_obj[KEY_TYPE] = self.get_type()
hf_dict = self.backend_dataset.to_dict()
dict_obj[KEY_INSTANCES] = []
first_key = None
for key in hf_dict.keys():
first_key = key
break
if first_key is not None:
num_instances = len(hf_dict[first_key])
dict_obj[KEY_INSTANCES] = [
{
key: hf_dict[key][i] for key in hf_dict.keys()
}
for i in range(num_instances)
]
return dict_obj
elif self.backend == "dict":
dict_obj = self.backend_dataset
return dict_obj
else:
raise NotImplementedError(
f'Current .to_dict is not supported for backend "{self.backend}"'
)
def to_list(self):
"""Returns a list of instances."""
if self.backend == "huggingface":
instance_list = [self.backend_dataset.__getitem__(idx)
for idx in range(len(self.backend_dataset))]
return instance_list
elif self.backend == "dict":
instance_list = copy.deepcopy(self.backend_dataset[KEY_INSTANCES])
# TODO: should be a list of instances, instance should be huggingface datasets row format
return instance_list
else:
raise NotImplementedError(
f'Current .to_list is not supported for backend "{self.backend}"'
)
def map(self, *args, **kwargs):
r"""
Parameters
------------
args : Optional.
Positional arguments.
kwargs : Optional.
Keyword arguments.
Returns
---------
self : Dataset object.
"""
# If the dataset uses Hugging Face as the backend,
# call the `map()` function of the Hugging Face backend dataset
if self.backend == "huggingface":
# Set the mapped dataset as the backend dataset of the current dataset
mapped_backend_dataset = self.backend_dataset.map(*args, **kwargs)
self.backend_dataset = mapped_backend_dataset
return self
else:
# If the backend is not Hugging Face, raise a NotImplementedError
raise NotImplementedError(
f'Currently .map is not supported for backend "{self.backend}"'
)
def get_backend(self) -> Optional[str]:
r"""
Returns
---------
self.backend
"""
return self.backend
def get_backend_dataset(self):
r"""
Returns
---------
self.backend_dataset
"""
return self.backend_dataset
def get_fingerprint(self):
r"""
Returns
---------
Fingerprint of the backend_dataset which controls the cache
"""
return self.backend_dataset._fingerprint
def get_data_args(self):
r"""
Returns
---------
self.data_args
"""
return self.data_args
def get_type(self) -> str:
r"""
Returns
---------
self.type
"""
return self.type
def save(
self,
file_path: str,
format: str="json"
):
r"""
Save the dataset to a json file.
Parameters
------------
file_path : str.
The path to the file where the dataset will be saved.
"""
if format == "json":
assert Path(file_path).suffix == ".json", "The file path must have a .json extension."
with open(file_path, "w", encoding='utf-8') as fout:
json.dump(self.to_dict(), fout, indent=4, ensure_ascii=False)
else:
logger.error(f"Unsupported format when saving the dataset: {format}.")
def sample(self, n: int, seed: int=42):
r"""
Sample n instances from the dataset.
Parameters
------------
n : int.
The number of instances to sample from the dataset.
Returns
---------
sample_dataset : Dataset object.
A new dataset object containing the sampled instances.
"""
if self.backend == "huggingface":
sampled_dataset = self.backend_dataset.shuffle(seed=seed).select(range(n))
output_dataset = self.create_from_dict(
{
"type": self.get_type(),
"instances": [
{
col_name: sampled_dataset[col_name][i] for col_name in sampled_dataset.column_names
} for i in range(n)
]
}
)
return output_dataset
else:
raise NotImplementedError(
f'Currently .sample is not supported for backend "{self.backend}"'
)
def train_test_split(self, test_size: float=0.2, shuffle: bool=True, seed: int=42):
r"""
Split the dataset into training and testing sets.
Parameters
------------
test_size : float, default=0.2.
The proportion of the dataset that will be used for testing.
Returns
---------
train_dataset : Dataset object.
A new dataset object containing the training instances.
test_dataset : Dataset object.
A new dataset object containing the testing instances.
"""
if self.backend == "huggingface":
splited = self.backend_dataset.train_test_split(
test_size=test_size, shuffle=shuffle, seed=seed
)
train_dataset = self.create_from_dict(
{
"type": self.get_type(),
"instances": [
{
col_name: splited["train"][col_name][i] for col_name in splited["train"].column_names
} for i in range(len(splited["train"]))
]
}
)
test_dataset = self.create_from_dict(
{
"type": self.get_type(),
"instances": [
{
col_name: splited["test"][col_name][i] for col_name in splited["test"].column_names
} for i in range(len(splited["test"]))
]
}
)
return train_dataset, test_dataset
else:
raise NotImplementedError(
f'Currently .train_test_split is not supported for backend "{self.backend}"'
)
def drop_instances(self, indices: list):
r"""
Drop instances from the dataset.
Parameters
------------
indices : list.
A list of indices of the instances to drop from the dataset.
"""
if self.backend == "huggingface":
self.backend_dataset = self.backend_dataset.remove_indices(indices)
else:
raise NotImplementedError(
f'Currently .drop_instances is not supported for backend "{self.backend}"'
)
def sanity_check(
self,
drop_invalid: bool=True,
):
r"""
Perform a sanity check on the dataset.
"""
if self.backend == "huggingface":
self.hf_dataset_sanity_check(drop_invalid)
else:
raise NotImplementedError(
f'Currently .sanity_check is not supported for backend "{self.backend}"'
)
def hf_dataset_sanity_check(
self,
drop_invalid: bool=True,
):
r"""
Perform a sanity check on the HuggingFace dataset.
"""
if self.backend_dataset is None or len(self.backend_dataset) == 0:
raise ValueError("Dataset is empty.")
if self.type == 'text_to_textlist':
num_output_per_instance = len(self.backend_dataset['output'][0])
dataset_cache = self.backend_dataset.filter(lambda x: len(x['input'])!=0)
dataset_cache = self.backend_dataset.filter(lambda x: len(x['output']) == num_output_per_instance)
dataset_cache = self.backend_dataset.filter(lambda x: not all([len(output) == 0 for output in x['output']]))
if len(dataset_cache) != len(self.backend_dataset):
warning_info = (
f"Found {len(self.backend_dataset) - len(dataset_cache)} invalid instances "
"during hf_dataset_sanity_check, please check:\n"
" 1. length of input strings should not be empty\n"
" 2. length of output strings should not be all empty\n"
" 3. number of output strings should be consistent\n" # since we will use tensor reshape later
)
if drop_invalid:
self.backend_dataset = dataset_cache
logger.warning(warning_info+"Invalid instances are dropped.")
else:
raise ValueError(warning_info)
else:
logger.warning(f"No sanity check for {self.type} dataset.") |