File size: 14,662 Bytes
032e687 |
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 json
import logging
import os
import torch
from datasets import Dataset as HFDataset
from datasets import DatasetDict, load_from_disk
from mmengine import print_log
from mmengine.config import Config, ConfigDict
from PIL import Image
from torch.utils.data import Dataset
from pycocotools import mask
import numpy as np
import torch.nn.functional as F
import copy
from xtuner.dataset.utils import encode_fn
from xtuner.registry import BUILDER
from .utils import expand2square, expand2square_mask
from xtuner.dataset.huggingface import process_hf_dataset, build_origin_dataset
class GCGDataset(Dataset):
def __init__(self,
image_folder,
image_processor,
data_path=None,
tokenizer=None,
offline_processed_text_folder=None,
max_dataset_length=None,
dataset_map_fn=None,
template_map_fn=None,
max_length=2048,
pad_image_to_square=False,
num_proc=32,
lazy=False,
repeats=1):
super().__init__()
self.lazy = lazy
self.max_length = max_length
self.dataset_map_fn = dataset_map_fn
self.template_map_fn = template_map_fn
if isinstance(self.template_map_fn, dict) and self.lazy:
_type = self.template_map_fn['type']
del self.template_map_fn['type']
self.template_map_fn = _type(**self.template_map_fn)
assert offline_processed_text_folder or (data_path and tokenizer)
if offline_processed_text_folder and data_path:
print_log(
'Both `offline_processed_text_folder` and '
'`data_path` are set, and we load dataset from'
'`offline_processed_text_folder` '
f'({offline_processed_text_folder})',
logger='current',
level=logging.WARNING)
if offline_processed_text_folder is not None:
raise NotImplementedError
else:
json_data = self.json_file_preprocess(data_path)
json_data = DatasetDict({'train': HFDataset.from_list(json_data)})
if self.lazy:
self.text_data = build_origin_dataset(json_data, 'train')
else:
self.text_data = process_hf_dataset(
dataset=json_data,
tokenizer=tokenizer,
max_length=max_length,
dataset_map_fn=dataset_map_fn,
template_map_fn=template_map_fn,
split='train',
max_dataset_length=max_dataset_length,
remove_unused_columns=False,
pack_to_max_length=False,
with_image_token=True,
map_num_proc=num_proc, # because limited mem
)
self.image_folder = image_folder
size = image_processor.crop_size
if isinstance(size, int):
self.image_h, self.image_w = size, size
else:
self.image_w, self.image_h = size
if isinstance(image_processor, dict) or isinstance(
image_processor, Config) or isinstance(image_processor,
ConfigDict):
self.image_processor = BUILDER.build(image_processor)
else:
self.image_processor = image_processor
self.pad_image_to_square = pad_image_to_square
self.down_ratio = 1
self.repeats = repeats
self.tokenizer = tokenizer
def json_file_preprocess(self, data_path):
with open(data_path, 'r') as f:
json_data = json.load(f)
return json_data
@property
def modality_length(self):
length_list = []
for data_dict in self.text_data:
if self.lazy:
cur_len = 100
else:
cur_len = len(data_dict['input_ids'])
if data_dict.get('image', None) is None:
cur_len = -cur_len
length_list.append(cur_len)
return length_list * self.repeats
def __len__(self):
return len(self.text_data) * self.repeats
def real_len(self):
return len(self.text_data)
def decode_mask(self, object_masks, ori_height, ori_width):
binary_masks = []
for object_mask in object_masks:
binary_mask = np.zeros((ori_height, ori_width), dtype=np.uint8)
for seg in object_mask:
rles = mask.frPyObjects([seg], ori_height, ori_width)
m = mask.decode(rles)
m = m.astype(np.uint8)
binary_mask += m.squeeze()
binary_masks.append(binary_mask)
if len(binary_masks) == 0:
return None
masks = np.stack(binary_masks, axis=0)
if self.pad_image_to_square:
masks = expand2square_mask(masks)
masks = torch.from_numpy(masks)
masks = F.interpolate(masks.unsqueeze(0), size=(self.image_h // self.down_ratio, self.image_w // self.down_ratio), mode='nearest').squeeze(0)
return masks
def __getitem__(self, index):
index = index % self.real_len()
data_dict = copy.deepcopy(self.text_data[index])
if self.lazy:
result = self.dataset_map_fn(data_dict)
data_dict.update(result)
result = self.template_map_fn(data_dict)
data_dict.update(result)
result = encode_fn(data_dict, tokenizer=self.tokenizer, max_length=self.max_length, with_image_token=True)
data_dict.update(result)
assert 'image' in data_dict.keys()
if data_dict.get('image', None) is not None:
image_file = data_dict['image']
image = Image.open(os.path.join(self.image_folder,
image_file)).convert('RGB')
ori_width, ori_height = image.size
if self.pad_image_to_square:
image = expand2square(
image,
tuple(
int(x * 255) for x in self.image_processor.image_mean))
image = self.image_processor.preprocess(
image, return_tensors='pt')['pixel_values'][0]
data_dict['pixel_values'] = image
# process and get masks
data_dict['masks'] = self.decode_mask(data_dict['masks'], ori_height=ori_height, ori_width=ori_width)
if data_dict['masks'] is None:
return self.__getitem__(0)
else:
if hasattr(self.image_processor, 'crop_size'):
crop_size = self.image_processor.crop_size
else:
crop_size = self.image_processor.size
data_dict['pixel_values'] = torch.zeros(3, crop_size['height'],
crop_size['width'])
data_dict['masks'] = None
return data_dict
class RefCOCOgGCGDataset(GCGDataset):
def __init__(self,
image_folder,
image_processor,
data_path=None,
tokenizer=None,
offline_processed_text_folder=None,
max_dataset_length=None,
dataset_map_fn=None,
template_map_fn=None,
max_length=2048,
pad_image_to_square=False,
lazy=False,
repeats=1,):
super().__init__(
image_folder=image_folder,
image_processor=image_processor,
data_path=data_path,
tokenizer=tokenizer,
offline_processed_text_folder=offline_processed_text_folder,
max_dataset_length=max_dataset_length,
dataset_map_fn=dataset_map_fn,
template_map_fn=template_map_fn,
max_length=max_length,
pad_image_to_square=pad_image_to_square,
lazy=lazy,
repeats=repeats,
)
def json_file_preprocess(self, data_path):
json_data = json.load(open(data_path))
# convert {id: dict} to dict(..., id=xx)
for idx in range(len(json_data)):
id = list(json_data[idx].keys())[0]
json_data[idx] = json_data[idx][id]
json_data[idx].update({'id': id})
return json_data
class GranDfGCGDataset(GCGDataset):
def __init__(self,
image_folder,
image_processor,
data_path=None,
tokenizer=None,
offline_processed_text_folder=None,
max_dataset_length=None,
dataset_map_fn=None,
template_map_fn=None,
max_length=2048,
pad_image_to_square=False,
num_proc=4,
lazy=False,
repeats=1):
super().__init__(
image_folder=image_folder,
image_processor=image_processor,
data_path=data_path,
tokenizer=tokenizer,
offline_processed_text_folder=offline_processed_text_folder,
max_dataset_length=max_dataset_length,
dataset_map_fn=dataset_map_fn,
template_map_fn=template_map_fn,
max_length=max_length,
pad_image_to_square=pad_image_to_square,
num_proc=num_proc,
lazy=lazy,
repeats=repeats,
)
def decode_mask(self, object_masks, ori_height, ori_width):
binary_masks = []
for object_mask in object_masks:
binary_mask = np.zeros((ori_height, ori_width), dtype=np.uint8)
for rle in object_mask:
m = mask.decode(rle).astype(np.uint8)
binary_mask += m.squeeze()
binary_masks.append(binary_mask)
if len(binary_masks) == 0:
return None
masks = np.stack(binary_masks, axis=0)
if self.pad_image_to_square:
masks = expand2square_mask(masks)
masks = torch.from_numpy(masks)
masks = F.interpolate(masks.unsqueeze(0), size=(self.image_h // self.down_ratio, self.image_w // self.down_ratio), mode='nearest').squeeze(0)
return masks
class OpenPsgGCGDataset(GranDfGCGDataset):
def __init__(self,
image_folder,
image_processor,
data_path=None,
tokenizer=None,
offline_processed_text_folder=None,
max_dataset_length=None,
dataset_map_fn=None,
template_map_fn=None,
max_length=2048,
pad_image_to_square=False,
num_proc=4,
lazy=False,
repeats=1):
super().__init__(
image_folder=image_folder,
image_processor=image_processor,
data_path=data_path,
tokenizer=tokenizer,
offline_processed_text_folder=offline_processed_text_folder,
max_dataset_length=max_dataset_length,
dataset_map_fn=dataset_map_fn,
template_map_fn=template_map_fn,
max_length=max_length,
pad_image_to_square=pad_image_to_square,
num_proc=num_proc,
lazy=lazy,
repeats=repeats,
)
class FlickrGCGDataset(GCGDataset):
def __init__(self,
image_folder,
image_processor,
data_path=None,
tokenizer=None,
offline_processed_text_folder=None,
max_dataset_length=None,
dataset_map_fn=None,
template_map_fn=None,
max_length=2048,
pad_image_to_square=False,
num_proc=4,
lazy=False,
repeats=1,):
super().__init__(
image_folder=image_folder,
image_processor=image_processor,
data_path=data_path,
tokenizer=tokenizer,
offline_processed_text_folder=offline_processed_text_folder,
max_dataset_length=max_dataset_length,
dataset_map_fn=dataset_map_fn,
template_map_fn=template_map_fn,
max_length=max_length,
pad_image_to_square=pad_image_to_square,
num_proc=num_proc,
lazy=lazy,
repeats=repeats,
)
def json_file_preprocess(self, data_path):
def filter_images(data_infos, min_size):
return [i for i, info in enumerate(data_infos) if min(info['width'], info['height']) >= min_size]
# convert {id: dict} to dict(..., id=xx)
from pycocotools.coco import COCO
self.coco = COCO(data_path)
self.image_ids = self.coco.getImgIds()
data_infos = []
total_ann_ids = []
removed_img_count = 0
for img_id in self.image_ids:
info = self.coco.loadImgs([img_id])[0]
if len(info['caption'].split(' ')) < 3:
removed_img_count += 1
continue
info['filename'] = info['file_name'].split('_')[-1]
info['height'] = int(info['height'])
info['width'] = int(info['width'])
data_infos.append(info)
ann_ids = self.coco.getAnnIds(imgIds=[img_id])
total_ann_ids.extend(ann_ids)
assert len(set(total_ann_ids)) == len(total_ann_ids), f"Non-unique annotation IDs in '{data_path}'!"
print(f'Removed {removed_img_count} images.')
data_infos = [data_infos[i] for i in filter_images(data_infos, min_size=32)]
# obtain_annotations
for data_info in data_infos:
ann_ids = self.coco.getAnnIds(imgIds=data_info['id'])
ann_info = self.coco.loadAnns(ann_ids)
data_info.update({'ann_info': ann_info})
return data_infos
def decode_mask(self, object_masks, ori_height, ori_width):
binary_masks = []
for object_mask in object_masks:
binary_mask = mask.decode(object_mask).astype(np.uint8)
binary_masks.append(binary_mask)
if len(binary_masks) == 0:
return None
masks = np.stack(binary_masks, axis=0)
if self.pad_image_to_square:
masks = expand2square_mask(masks)
masks = torch.from_numpy(masks)
masks = F.interpolate(masks.unsqueeze(0), size=(self.image_h // self.down_ratio, self.image_w // self.down_ratio), mode='nearest').squeeze(0)
return masks |