File size: 11,754 Bytes
5ceacbc |
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 |
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import os
from io import BytesIO
import json
import logging
import base64
import random
from typing import Callable, List, Tuple, Union, NamedTuple
from PIL import Image
from PIL import ImageFile
import torch.utils.data as data
from .languages.prompt_engineering import prompt_engineering
from .tsv_file import TSVFile, CompositeTSVFile
ImageFile.LOAD_TRUNCATED_IMAGES = True
logger = logging.getLogger(__name__)
class TSVDataset(data.Dataset):
def __init__(self,
tsv_file: Union[str, List[str]],
transform: Callable = None,
map_file: str = None,
token_file: str = None,
is_train: bool = True,
azcopy_path: str = None):
self.transform = transform
self._chunk_sizes = None
self.label2idx = self._load_map(map_file)
self.class_selector = list(self.label2idx.keys()) if self.label2idx else None
if isinstance(tsv_file, str):
if os.path.splitext(tsv_file)[1] == '.tsv':
self.tsv_file = TSVFile(
tsv_file, class_selector=self.class_selector
)
else:
self.tsv_file = CompositeTSVFile(
tsv_file,
class_selector=self.class_selector,
is_train=is_train,
sas_token_path=token_file,
azcopy_path=azcopy_path
)
self._chunk_sizes = self.tsv_file.get_chunk_size()
elif isinstance(tsv_file, list):
self.tsv_file = CompositeTSVFile(
tsv_file,
class_selector=self.class_selector,
is_train=is_train,
sas_token_path=token_file,
azcopy_path=azcopy_path
)
self._chunk_sizes = self.tsv_file.get_chunk_size()
else:
raise ValueError("Invalid input! Please check the tsv filenames")
logger.debug('=> {}\titems: {}'.format(tsv_file, len(self.tsv_file)))
def fetch_blob(self, idx):
image_tsv = self.tsv_file.file_list[idx]
self.tsv_file.blob_storage.fetch_blob(image_tsv)
def num_classes(self):
return len(self.class_selector)
def get_chunk_sizes(self):
return self._chunk_sizes
def get_class_boundaries(self):
# The samples of each class are organized class-by-class.
# _class_boundaries stores the lower- and upper-bound of each class.
return self.tsv_file.get_class_boundaries()
def get_filenames(self):
filenames = [
self.tsv_file.get_key(i)
for i in range(self.tsv_file.num_rows())
]
return filenames
def _load_map(self, map_file: str):
if not map_file:
return None
label2idx = {}
with open(map_file) as f:
for line in f:
items = line.strip().split('\t')
label2idx[items[0]] = int(items[1])
return label2idx
def __getitem__(self, index: Union[int, Tuple[int, int]]):
items = self.tsv_file[index]
_, target, img = self._decode_data(items)
if self.transform:
img = self.transform(img)
return img, target
def _decode_data(self, items: Tuple[str, str, str]):
key = items[0]
label = self._get_label(items[1])
image = Image.open(BytesIO(base64.b64decode(items[2]))).convert('RGB')
return key, label, image
def _get_label(self, item: str):
if not self.label2idx:
return int(item)
js = json.loads(item)
return self.label2idx[js[0]['class']]
def __len__(self):
return len(self.tsv_file)
class TSVMeta(NamedTuple):
source: str
num_classes: int
task: str
class TSVImageTextDatasetV2(data.Dataset):
"""
This class is intended for encapsulating Image/Text pair data for contrastive learning described in
the following paper,
"Learning Transferable Visual Models From Natural Language Supervision" (a.k.a CLIP)
V2: support image text pairs and supervised classification data
"""
def __init__(self,
image_tsv_file: Union[str, List[str]],
text_tsv_file: Union[str, List[str]],
transform: Callable = None,
tokenize: Callable = None,
context_length: int = 77,
num_captions: int = 1,
text_format: str = 'txt',
is_train: bool = True,
sas_token_path: str = None,
azcopy_path: str = None,
metas: List[NamedTuple] = None,
prompt_engineering=True,
concat_queries=False):
self.transform = transform
self.tokenize = tokenize
self._chunk_sizes = None
self.context_length = context_length
self.num_captions = num_captions
self.text_format = text_format
self.tsv_file_list = []
self.metas = metas
self.label_offsets = self.build_label_offsets()
self.prompt_engineering = prompt_engineering
self.concat_queries = concat_queries
if isinstance(image_tsv_file, str) and isinstance(text_tsv_file, str):
# single tsv file
if (
os.path.splitext(image_tsv_file)[1].lower() == '.tsv'
and os.path.splitext(text_tsv_file)[1].lower() == '.tsv'
):
self.tsv_file_list.append((image_tsv_file, text_tsv_file))
self.image_tsv_file = TSVFile(
image_tsv_file, if_generate_lineidx=True
)
self.text_tsv_file = TSVFile(
text_tsv_file, if_generate_lineidx=True
)
else:
raise ValueError("Invalid input! Please check the tsv filenames.")
# multiple tsv files specified in a list
elif (
isinstance(image_tsv_file, list)
and isinstance(text_tsv_file, list)
):
assert len(image_tsv_file) == len(text_tsv_file), \
"Inconsistent number of Image/Text tsv files!"
self.tsv_file_list = [
(txt, img)
for img, txt in zip(image_tsv_file, text_tsv_file)
]
self.image_tsv_file = CompositeTSVFile(
image_tsv_file,
is_train=is_train,
sas_token_path=sas_token_path,
azcopy_path=azcopy_path
)
self.text_tsv_file = CompositeTSVFile(
text_tsv_file,
is_train=is_train,
sas_token_path=sas_token_path,
azcopy_path=azcopy_path
)
self._chunk_sizes = self.image_tsv_file.get_chunk_size()
else:
raise ValueError("Invalid input! Please check the tsv filenames.")
assert len(self.image_tsv_file) == len(self.text_tsv_file), \
"Inconsistent size of Image/Text ({}/{}) data!".format(
len(self.image_tsv_file), len(self.text_tsv_file)
)
def build_label_offsets(self):
if self.metas is None:
return None
label_offsets = {}
offset = 1
for meta in self.metas:
print(meta)
print(label_offsets)
label_offsets[meta.source] = offset
offset += meta.num_classes
return label_offsets
def fetch_blob(self, idx):
# image_tsv, text_tsv = self.tsv_file_list[idx]
image_tsv = self.image_tsv_file.file_list[idx]
text_tsv = self.text_tsv_file.file_list[idx]
self.image_tsv_file.blob_storage.fetch_blob(image_tsv)
self.text_tsv_file.blob_storage.fetch_blob(text_tsv)
def get_chunk_sizes(self):
return self._chunk_sizes
def __getitem__(self, index: Union[int, Tuple[int, int]]):
if index is None:
import torch
return torch.tensor([], dtype=torch.float32), \
torch.tensor([], dtype=torch.int64), \
torch.tensor([], dtype=torch.int64)
items_image = self.image_tsv_file[index]
items_text = self.text_tsv_file[index]
assert items_text[0] == items_image[0], \
'keys do not match for image and text {} vs {}'.format(
items_text[0], items_image[0]
)
_, img = self._decode_image(items_image)
_, txt, label = self._decode_text(items_text)
if self.transform:
img = self.transform(img)
tokens = self.tokenize(
txt, padding='max_length', truncation=True, max_length=self.context_length,
return_tensors='pt'
) if self.tokenize else txt
tokens['input_ids'].squeeze_()
tokens['attention_mask'].squeeze_()
return img, tokens, label
def _decode_image(self, items: Tuple[str, str]):
key = items[0]
image = Image.open(BytesIO(base64.b64decode(items[1]))).convert('RGB')
return key, image
def _decode_text(self, items: Tuple[str, Union[str, dict]]):
key = items[0]
text = ''
if self.text_format != 'json':
raise ValueError('Only support json format')
# Do some reasonable handing of occasionally bad data.
try:
js = json.loads(items[1])
except Exception as e:
# empty dictionary
js = {}
# Record the data error in the log.
logger.info("JSON parsing error on: " + items[1])
logger.info(str(e))
# do not raise the exception
# raise e
# put some text in and continue processing data (do not kill job)
sstr = items[1].find("\"")
if (sstr < 0):
sstr = 0
estr = items[1][sstr:].find("\"")
if (estr < 0):
estr = len(items[1])
text = items[1][sstr:estr]
if (len(text) < 2):
text = "A picture showing some content."
label = 0
if 'captions' in js:
captions = js['captions']
if isinstance(captions, list):
if self.num_captions == 1:
text = random.choice(captions)
else:
text = captions
if len(captions) > self.num_captions:
text = captions[:self.num_captions]
elif isinstance(captions, str):
text = captions
else:
raise ValueError('captions should be str or list')
label = 0
elif 'tags' in js:
text = prompt_engineering(js['tags'])
label = 0
elif 'task' in js and js['task'] == 'classification':
if (self.prompt_engineering):
text = prompt_engineering(js['class_name'])
else:
text = js['class_name']
label = js['class_id']
if (self.label_offsets is not None):
if (js['source'] in self.label_offsets):
label += self.label_offsets[js['source']]
if (self.concat_queries):
if ('queries' in js) and (len(js['queries']) > 0):
q = ''
for item in js['queries']:
q = q + item + ' '
text = q + ', ' + text
return key, text, label
def __len__(self):
return len(self.image_tsv_file)
|