File size: 10,719 Bytes
4943752 |
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 |
"""
DatasetArgs Class
=================
"""
from dataclasses import dataclass
import textattack
from textattack.shared.utils import ARGS_SPLIT_TOKEN, load_module_from_file
HUGGINGFACE_DATASET_BY_MODEL = {
#
# bert-base-uncased
#
"bert-base-uncased-ag-news": ("ag_news", None, "test"),
"bert-base-uncased-cola": ("glue", "cola", "validation"),
"bert-base-uncased-imdb": ("imdb", None, "test"),
"bert-base-uncased-mnli": (
"glue",
"mnli",
"validation_matched",
None,
{0: 1, 1: 2, 2: 0},
),
"bert-base-uncased-mrpc": ("glue", "mrpc", "validation"),
"bert-base-uncased-qnli": ("glue", "qnli", "validation"),
"bert-base-uncased-qqp": ("glue", "qqp", "validation"),
"bert-base-uncased-rte": ("glue", "rte", "validation"),
"bert-base-uncased-sst2": ("glue", "sst2", "validation"),
"bert-base-uncased-stsb": (
"glue",
"stsb",
"validation",
None,
None,
None,
5.0,
),
"bert-base-uncased-wnli": ("glue", "wnli", "validation"),
"bert-base-uncased-mr": ("rotten_tomatoes", None, "test"),
"bert-base-uncased-snli": ("snli", None, "test", None, {0: 1, 1: 2, 2: 0}),
"bert-base-uncased-yelp": ("yelp_polarity", None, "test"),
#
# distilbert-base-cased
#
"distilbert-base-cased-cola": ("glue", "cola", "validation"),
"distilbert-base-cased-mrpc": ("glue", "mrpc", "validation"),
"distilbert-base-cased-qqp": ("glue", "qqp", "validation"),
"distilbert-base-cased-snli": ("snli", None, "test"),
"distilbert-base-cased-sst2": ("glue", "sst2", "validation"),
"distilbert-base-cased-stsb": (
"glue",
"stsb",
"validation",
None,
None,
None,
5.0,
),
"distilbert-base-uncased-ag-news": ("ag_news", None, "test"),
"distilbert-base-uncased-cola": ("glue", "cola", "validation"),
"distilbert-base-uncased-imdb": ("imdb", None, "test"),
"distilbert-base-uncased-mnli": (
"glue",
"mnli",
"validation_matched",
None,
{0: 1, 1: 2, 2: 0},
),
"distilbert-base-uncased-mr": ("rotten_tomatoes", None, "test"),
"distilbert-base-uncased-mrpc": ("glue", "mrpc", "validation"),
"distilbert-base-uncased-qnli": ("glue", "qnli", "validation"),
"distilbert-base-uncased-rte": ("glue", "rte", "validation"),
"distilbert-base-uncased-wnli": ("glue", "wnli", "validation"),
#
# roberta-base (RoBERTa is cased by default)
#
"roberta-base-ag-news": ("ag_news", None, "test"),
"roberta-base-cola": ("glue", "cola", "validation"),
"roberta-base-imdb": ("imdb", None, "test"),
"roberta-base-mr": ("rotten_tomatoes", None, "test"),
"roberta-base-mrpc": ("glue", "mrpc", "validation"),
"roberta-base-qnli": ("glue", "qnli", "validation"),
"roberta-base-rte": ("glue", "rte", "validation"),
"roberta-base-sst2": ("glue", "sst2", "validation"),
"roberta-base-stsb": ("glue", "stsb", "validation", None, None, None, 5.0),
"roberta-base-wnli": ("glue", "wnli", "validation"),
#
# albert-base-v2 (ALBERT is cased by default)
#
"albert-base-v2-ag-news": ("ag_news", None, "test"),
"albert-base-v2-cola": ("glue", "cola", "validation"),
"albert-base-v2-imdb": ("imdb", None, "test"),
"albert-base-v2-mr": ("rotten_tomatoes", None, "test"),
"albert-base-v2-rte": ("glue", "rte", "validation"),
"albert-base-v2-qqp": ("glue", "qqp", "validation"),
"albert-base-v2-snli": ("snli", None, "test"),
"albert-base-v2-sst2": ("glue", "sst2", "validation"),
"albert-base-v2-stsb": ("glue", "stsb", "validation", None, None, None, 5.0),
"albert-base-v2-wnli": ("glue", "wnli", "validation"),
"albert-base-v2-yelp": ("yelp_polarity", None, "test"),
#
# xlnet-base-cased
#
"xlnet-base-cased-cola": ("glue", "cola", "validation"),
"xlnet-base-cased-imdb": ("imdb", None, "test"),
"xlnet-base-cased-mr": ("rotten_tomatoes", None, "test"),
"xlnet-base-cased-mrpc": ("glue", "mrpc", "validation"),
"xlnet-base-cased-rte": ("glue", "rte", "validation"),
"xlnet-base-cased-stsb": (
"glue",
"stsb",
"validation",
None,
None,
None,
5.0,
),
"xlnet-base-cased-wnli": ("glue", "wnli", "validation"),
}
#
# Models hosted by textattack.
#
TEXTATTACK_DATASET_BY_MODEL = {
#
# LSTMs
#
"lstm-ag-news": ("ag_news", None, "test"),
"lstm-imdb": ("imdb", None, "test"),
"lstm-mr": ("rotten_tomatoes", None, "test"),
"lstm-sst2": ("glue", "sst2", "validation"),
"lstm-yelp": ("yelp_polarity", None, "test"),
#
# CNNs
#
"cnn-ag-news": ("ag_news", None, "test"),
"cnn-imdb": ("imdb", None, "test"),
"cnn-mr": ("rotten_tomatoes", None, "test"),
"cnn-sst2": ("glue", "sst2", "validation"),
"cnn-yelp": ("yelp_polarity", None, "test"),
#
# T5 for translation
#
"t5-en-de": (
"textattack.datasets.helpers.TedMultiTranslationDataset",
"en",
"de",
),
"t5-en-fr": (
"textattack.datasets.helpers.TedMultiTranslationDataset",
"en",
"fr",
),
"t5-en-ro": (
"textattack.datasets.helpers.TedMultiTranslationDataset",
"en",
"de",
),
#
# T5 for summarization
#
"t5-summarization": ("gigaword", None, "test"),
}
@dataclass
class DatasetArgs:
"""Arguments for loading dataset from command line input."""
dataset_by_model: str = None
dataset_from_huggingface: str = None
dataset_from_file: str = None
dataset_split: str = None
filter_by_labels: list = None
@classmethod
def _add_parser_args(cls, parser):
"""Adds dataset-related arguments to an argparser."""
dataset_group = parser.add_mutually_exclusive_group()
dataset_group.add_argument(
"--dataset-by-model",
type=str,
required=False,
default=None,
help="Dataset to load depending on the name of the model",
)
dataset_group.add_argument(
"--dataset-from-huggingface",
type=str,
required=False,
default=None,
help="Dataset to load from `datasets` repository.",
)
dataset_group.add_argument(
"--dataset-from-file",
type=str,
required=False,
default=None,
help="Dataset to load from a file.",
)
parser.add_argument(
"--dataset-split",
type=str,
required=False,
default=None,
help="Split of dataset to use when specifying --dataset-by-model or --dataset-from-huggingface.",
)
parser.add_argument(
"--filter-by-labels",
nargs="+",
type=int,
required=False,
default=None,
help="List of labels to keep in the dataset and discard all others.",
)
return parser
@classmethod
def _create_dataset_from_args(cls, args):
"""Given ``DatasetArgs``, return specified
``textattack.dataset.Dataset`` object."""
assert isinstance(
args, cls
), f"Expect args to be of type `{type(cls)}`, but got type `{type(args)}`."
# Automatically detect dataset for huggingface & textattack models.
# This allows us to use the --model shortcut without specifying a dataset.
if hasattr(args, "model"):
args.dataset_by_model = args.model
if args.dataset_by_model in HUGGINGFACE_DATASET_BY_MODEL:
args.dataset_from_huggingface = HUGGINGFACE_DATASET_BY_MODEL[
args.dataset_by_model
]
elif args.dataset_by_model in TEXTATTACK_DATASET_BY_MODEL:
dataset = TEXTATTACK_DATASET_BY_MODEL[args.dataset_by_model]
if dataset[0].startswith("textattack"):
# unsavory way to pass custom dataset classes
# ex: dataset = ('textattack.datasets.helpers.TedMultiTranslationDataset', 'en', 'de')
dataset = eval(f"{dataset[0]}")(*dataset[1:])
return dataset
else:
args.dataset_from_huggingface = dataset
# Get dataset from args.
if args.dataset_from_file:
textattack.shared.logger.info(
f"Loading model and tokenizer from file: {args.model_from_file}"
)
if ARGS_SPLIT_TOKEN in args.dataset_from_file:
dataset_file, dataset_name = args.dataset_from_file.split(
ARGS_SPLIT_TOKEN
)
else:
dataset_file, dataset_name = args.dataset_from_file, "dataset"
try:
dataset_module = load_module_from_file(dataset_file)
except Exception:
raise ValueError(f"Failed to import file {args.dataset_from_file}")
try:
dataset = getattr(dataset_module, dataset_name)
except AttributeError:
raise AttributeError(
f"Variable ``dataset`` not found in module {args.dataset_from_file}"
)
elif args.dataset_from_huggingface:
dataset_args = args.dataset_from_huggingface
if isinstance(dataset_args, str):
if ARGS_SPLIT_TOKEN in dataset_args:
dataset_args = dataset_args.split(ARGS_SPLIT_TOKEN)
else:
dataset_args = (dataset_args,)
if args.dataset_split:
if len(dataset_args) > 1:
dataset_args = (
dataset_args[:2] + (args.dataset_split,) + dataset_args[3:]
)
dataset = textattack.datasets.HuggingFaceDataset(
*dataset_args, shuffle=False
)
else:
dataset = textattack.datasets.HuggingFaceDataset(
*dataset_args, split=args.dataset_split, shuffle=False
)
else:
dataset = textattack.datasets.HuggingFaceDataset(
*dataset_args, shuffle=False
)
else:
raise ValueError("Must supply pretrained model or dataset")
assert isinstance(
dataset, textattack.datasets.Dataset
), "Loaded `dataset` must be of type `textattack.datasets.Dataset`."
if args.filter_by_labels:
dataset.filter_by_labels_(args.filter_by_labels)
return dataset
|