Spaces:
Running
Running
File size: 11,813 Bytes
a72a7d4 1855cc2 a72a7d4 1855cc2 a72a7d4 |
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 |
import random
from typing import List, Tuple
from itertools import islice
import datasets
from datasets import load_dataset, concatenate_datasets
from torch.utils.data import Dataset
from PIL import Image
import os
from torchvision.transforms import RandAugment
def get_randaugment_transform(n=2, m=9):
return RandAugment(num_ops=n, magnitude=m)
def add_prompt_template(data):
data["qry"] = f"<|image_1|>{data['qry']}"
data["pos_text"] = f"<|image_1|>{data['pos_text']}"
data["hard_neg_text"] = f"<|image_1|>{data['hard_neg_text']}"
return data
Phi_Image_token = "<|image_1|>"
Llava_Image_token = "<image>"
Qwen_Image_token = "<|image_pad|>"
Internvl_Image_token = "<image>"
class TrainDataset(Dataset):
def __init__(self, data_args, model_args):
self.data_args = data_args
self.model_args = model_args
self.transform = None
if self.data_args.randaugment:
self.transform = get_randaugment_transform()
train_data = []
if data_args.subset_name is not None:
print(f"Loading {len(data_args.subset_name)} datasets: {data_args.subset_name}")
for subset in data_args.subset_name:
dataset_name = os.path.join(self.data_args.dataset_name, subset)
subset_data = load_dataset(
dataset_name,
split=f"{self.data_args.dataset_split}",
)
train_data.append(subset_data)
self.train_data = concatenate_datasets(train_data)
self.train_data = self.train_data.shuffle(seed=42)
else:
train_data = load_dataset(
self.data_args.dataset_name,
split=f"{self.data_args.dataset_split}",
)
if "hard_neg" in self.data_args.dataset_name:
# self.train_data = train_data.map(add_prompt_template, num_proc=8)
print(train_data)
else:
self.train_data = train_data
if self.data_args.num_samples:
# self.train_data = self.train_data[:self.data_args.num_samples]
self.train_data = self.train_data.select(range(self.data_args.num_samples))
print(f"len of train_data: {len(self.train_data)}")
def __len__(self):
return len(self.train_data)
def _process_image(self, image, resolution):
if image is None:
return None
if resolution == "high":
image = image.resize((1344, 1344))
elif resolution == "low":
image = image.resize((336, 336))
elif resolution == "clip":
image = image.resize((224, 224))
return image
def _get_image(self, img_path):
if img_path == "":
return None
if img_path.startswith('/'):
full_img_path = img_path
else:
full_img_path = os.path.join(self.data_args.image_dir, img_path)
image = Image.open(full_img_path)
if self.model_args.model_backbone == "llava_next":
# TODO: make it configurable
return self._process_image(image, "high")
elif self.model_args.model_backbone == "qwen":
return self._process_image(image, "low")
elif self.model_args.model_backbone == "internvl_2_5":
# TODO: make it configurable
return self._process_image(image, "high")
else:
return image
def __getitem__(self, item) -> Tuple[str, List[str]]:
data_item = self.train_data[item]
qry_text, qry_image_path, pos_text, pos_image_path = (
data_item["qry"], data_item["qry_image_path"],
data_item["pos_text"], data_item["pos_image_path"],
)
qry_image = self._get_image(qry_image_path)
if self.transform:
qry_image = self.transform(qry_image)
if self.model_args.model_backbone == "llava_next":
# Update image token
qry_text = qry_text.replace(Phi_Image_token, Llava_Image_token)
pos_text = pos_text.replace(Phi_Image_token, Llava_Image_token)
elif self.model_args.model_backbone == "qwen":
qry_text = qry_text.replace(Phi_Image_token, Qwen_Image_token)
pos_text = pos_text.replace(Phi_Image_token, Qwen_Image_token)
elif self.model_args.model_backbone == "internvl_2_5":
qry_text = qry_text.replace(Phi_Image_token, Internvl_Image_token)
pos_text = pos_text.replace(Phi_Image_token, Internvl_Image_token)
if "hard_neg" in self.data_args.dataset_name:
hard_neg_text, hard_neg_image_path = (
data_item["hard_neg_text"], data_item["hard_neg_image_path"],
)
if self.model_args.model_backbone == "llava_next":
# Update image token
hard_neg_text = hard_neg_text.replace(Phi_Image_token, Llava_Image_token)
elif self.model_args.model_backbone == "internvl_2_5":
hard_neg_text = hard_neg_text.replace(Phi_Image_token, Internvl_Image_token)
return (
qry_text, qry_image,
pos_text, self._get_image(pos_image_path),
hard_neg_text, self._get_image(hard_neg_image_path)
)
return (
qry_text, qry_image,
pos_text, self._get_image(pos_image_path)
)
class EvalDataset(Dataset):
def __init__(self, data_args, model_args, subset, text_field, img_path_field):
"""
(text_field, image_field) -> ("qry_text", "qry_img_path") or ("tgt_text", "tgt_img_path")
"""
self.data_args = data_args
self.model_args = model_args
if data_args.subset_name is not None:
self.eval_data = load_dataset(
self.data_args.dataset_name,
subset,
split=self.data_args.dataset_split,
)
else:
self.eval_data = load_dataset(
self.data_args.dataset_name,
split=self.data_args.dataset_split,
)
print(f"len of eval_data: {len(self.eval_data)}")
self.paired_data = self.get_paired_data(text_field, img_path_field)
self.paired_dataset = datasets.Dataset.from_dict({
"text": [pair["text"] for pair in self.paired_data],
"img_path": [pair["img_path"] for pair in self.paired_data]
})
def __len__(self):
return len(self.paired_dataset)
def __getitem__(self, item):
text, img_path = self.paired_dataset[item]["text"], self.paired_dataset[item]["img_path"]
if self.model_args.model_backbone == "llava_next":
# Update llava image token
text = text.replace(Phi_Image_token, Llava_Image_token)
elif self.model_args.model_backbone == "qwen":
text = text.replace(Phi_Image_token, Qwen_Image_token)
elif self.model_args.model_backbone == "internvl_2_5":
text = text.replace(Phi_Image_token, Internvl_Image_token)
return text, self._get_image(img_path),
def _process_image(self, image, resolution):
if image is None:
return None
if resolution == "high":
image = image.resize((1344, 1344))
else:
image = image.resize((336, 336))
return image
def _get_image(self, img_path):
if img_path == "":
return None
if img_path.startswith("/"):
full_img_path = img_path
else:
full_img_path = os.path.join(self.data_args.image_dir, img_path)
image = Image.open(full_img_path)
if self.model_args.model_backbone == "llava_next":
return self._process_image(image, "high")
elif self.model_args.model_backbone == "internvl_2_5":
return self._process_image(image, "high")
else:
return image
return image
def get_paired_data(self, text_field, img_path_field):
"""
(text_field, image_field) -> ("qry_text", "qry_img_path") or ("tgt_text", "tgt_img_path")
"""
unique_pair = set()
for row in self.eval_data:
if isinstance(row[text_field], str):
if row[text_field]:
unique_pair.add((row[text_field], row[img_path_field]))
else:
if isinstance(row[img_path_field], List):
for img_path in row[img_path_field]:
unique_pair.add((row[text_field], img_path))
else:
unique_pair.add((row[text_field], row[img_path_field]))
elif isinstance(row[text_field], List):
assert isinstance(row[img_path_field], List) and len(row[img_path_field]) == len(row[text_field])
for text, img_path in zip(row[text_field], row[img_path_field]):
unique_pair.add((text, img_path))
paired_data = [{"text": text, "img_path": img_path} for text, img_path in unique_pair]
return paired_data
class FlickrDataset(Dataset):
def __init__(self, modality, model_backbone):
self.model_backbone = model_backbone
self.modality = modality
self.raw_data = load_dataset("nlphuji/flickr_1k_test_image_text_retrieval", split="test")
if modality == "image":
self.eval_data, self.image_names = self.get_image_data()
else:
self.eval_data, self.image_names = self.get_text_data()
def __len__(self):
return len(self.eval_data)
def __getitem__(self, idx):
return self.eval_data[idx]
def __getitem__(self, idx):
text, image = self.eval_data[idx]
if self.model_backbone == "llava_next":
# Update llava image token
text = text.replace(Phi_Image_token, Llava_Image_token)
image = self._process_image(image, "high")
return text, image
def _process_image(self, image, resolution):
if image is None:
return None
if resolution == "high":
image = image.resize((1344, 1344))
else:
image = image.resize((336, 336))
return image
def _get_image(self, img_path):
if img_path == "":
return None
full_img_path = os.path.join(self.data_args.image_dir, img_path)
image = Image.open(full_img_path)
if self.model_backbone == "llava_next":
return self._process_image(image, "high")
else:
return image
return image
def get_image_data(self):
eval_data, image_names = [], []
# i2t
inst = "<|image_1|> Find an image caption describing the given image." # llava-1344-step1k4, i2t=94.0, t2i=80.26
# inst = "<|image_1|> Represent the given image for image caption retrieval." # llava-1344-step1k4, i2t=94.6, t2i=78.98
# t2i
# inst = "<|image_1|> Represent the given image." # MSCOCO t2i
for row in self.raw_data:
eval_data.append((inst, row["image"]))
image_names.append(row["filename"])
return eval_data, image_names
def get_text_data(self):
eval_data, image_names = [], []
# i2t
inst = ""
# t2i
# inst = "Retrieve an image that matches the given caption: "
# inst = "Find me an everyday image that matches the given caption." # MSCOCO t2i
for row in self.raw_data:
for caption in row["caption"]:
# eval_data.append((caption, None))
eval_data.append((inst + caption, None))
image_names.append(row["filename"])
return eval_data, image_names
|