File size: 9,536 Bytes
2ea2835 | 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 | from utils.lmdb import get_array_shape_from_lmdb, retrieve_row_from_lmdb
from torch.utils.data import Dataset
import numpy as np
import torch
import lmdb
import json
from pathlib import Path
from PIL import Image
import os
class TextDataset(Dataset):
def __init__(self, prompt_path, extended_prompt_path=None):
with open(prompt_path, encoding="utf-8") as f:
self.prompt_list = [line.rstrip() for line in f]
if extended_prompt_path is not None:
with open(extended_prompt_path, encoding="utf-8") as f:
self.extended_prompt_list = [line.rstrip() for line in f]
assert len(self.extended_prompt_list) == len(self.prompt_list)
else:
self.extended_prompt_list = None
def __len__(self):
return len(self.prompt_list)
def __getitem__(self, idx):
batch = {
"prompts": self.prompt_list[idx],
"idx": idx,
}
if self.extended_prompt_list is not None:
batch["extended_prompts"] = self.extended_prompt_list[idx]
return batch
class ODERegressionLMDBDataset(Dataset):
def __init__(
self,
data_path: str,
max_pair: int = int(1e8),
*,
load_latents: bool = True,
load_terminal_latent_only: bool = False,
):
self.env = lmdb.open(data_path, readonly=True,
lock=False, readahead=False, meminit=False)
self.latents_shape = get_array_shape_from_lmdb(self.env, 'latents')
self.max_pair = max_pair
self.load_latents = bool(load_latents)
self.load_terminal_latent_only = bool(
load_terminal_latent_only)
if self.load_terminal_latent_only and not self.load_latents:
raise ValueError(
"load_terminal_latent_only requires load_latents=True")
def __len__(self):
return min(self.latents_shape[0], self.max_pair)
def __getitem__(self, idx):
"""
Outputs:
- prompts: List of Strings
- latents: Tensor of shape (num_denoising_steps, num_frames, num_channels, height, width). It is ordered from pure noise to clean image.
"""
prompts = retrieve_row_from_lmdb(
self.env,
"prompts", str, idx
)
batch = {
"prompts": prompts,
"idx": idx,
}
if self.load_latents:
latents = retrieve_row_from_lmdb(
self.env,
"latents", np.float16, idx, shape=self.latents_shape[1:]
)
if len(latents.shape) == 4:
if not self.load_terminal_latent_only:
latents = latents[None, ...]
elif self.load_terminal_latent_only:
# Keep only the clean terminal trajectory row in host memory.
latents = latents[-1]
batch["ode_latent"] = torch.tensor(
latents, dtype=torch.float32)
return batch
class ShardingLMDBDataset(Dataset):
def __init__(self, data_path: str, max_pair: int = int(1e8)):
self.envs = []
self.index = []
for fname in sorted(os.listdir(data_path)):
path = os.path.join(data_path, fname)
env = lmdb.open(path,
readonly=True,
lock=False,
readahead=False,
meminit=False)
self.envs.append(env)
self.latents_shape = [None] * len(self.envs)
for shard_id, env in enumerate(self.envs):
self.latents_shape[shard_id] = get_array_shape_from_lmdb(env, 'latents')
for local_i in range(self.latents_shape[shard_id][0]):
self.index.append((shard_id, local_i))
# print("shard_id ", shard_id, " local_i ", local_i)
self.max_pair = max_pair
def __len__(self):
return len(self.index)
def __getitem__(self, idx):
"""
Outputs:
- prompts: List of Strings
- latents: Tensor of shape (num_denoising_steps, num_frames, num_channels, height, width). It is ordered from pure noise to clean image.
"""
shard_id, local_idx = self.index[idx]
latents = retrieve_row_from_lmdb(
self.envs[shard_id],
"latents", np.float16, local_idx,
shape=self.latents_shape[shard_id][1:]
)
if len(latents.shape) == 4:
latents = latents[None, ...]
prompts = retrieve_row_from_lmdb(
self.envs[shard_id],
"prompts", str, local_idx
)
return {
"prompts": prompts,
"ode_latent": torch.tensor(latents, dtype=torch.float32)
}
class TextImagePairDataset(Dataset):
def __init__(
self,
data_dir,
transform=None,
eval_first_n=-1,
pad_to_multiple_of=None
):
"""
Args:
data_dir (str): Path to the directory containing:
- target_crop_info_*.json (metadata file)
- */ (subdirectory containing images with matching aspect ratio)
transform (callable, optional): Optional transform to be applied on the image
"""
self.transform = transform
data_dir = Path(data_dir)
# Find the metadata JSON file
metadata_files = list(data_dir.glob('target_crop_info_*.json'))
if not metadata_files:
raise FileNotFoundError(f"No metadata file found in {data_dir}")
if len(metadata_files) > 1:
raise ValueError(f"Multiple metadata files found in {data_dir}")
metadata_path = metadata_files[0]
# Extract aspect ratio from metadata filename (e.g. target_crop_info_26-15.json -> 26-15)
aspect_ratio = metadata_path.stem.split('_')[-1]
# Use aspect ratio subfolder for images
self.image_dir = data_dir / aspect_ratio
if not self.image_dir.exists():
raise FileNotFoundError(f"Image directory not found: {self.image_dir}")
# Load metadata
with open(metadata_path, 'r') as f:
self.metadata = json.load(f)
eval_first_n = eval_first_n if eval_first_n != -1 else len(self.metadata)
self.metadata = self.metadata[:eval_first_n]
# Verify all images exist
for item in self.metadata:
image_path = self.image_dir / item['file_name']
if not image_path.exists():
raise FileNotFoundError(f"Image not found: {image_path}")
self.dummy_prompt = "DUMMY PROMPT"
self.pre_pad_len = len(self.metadata)
if pad_to_multiple_of is not None and len(self.metadata) % pad_to_multiple_of != 0:
# Duplicate the last entry
self.metadata += [self.metadata[-1]] * (
pad_to_multiple_of - len(self.metadata) % pad_to_multiple_of
)
def __len__(self):
return len(self.metadata)
def __getitem__(self, idx):
"""
Returns:
dict: A dictionary containing:
- image: PIL Image
- caption: str
- target_bbox: list of int [x1, y1, x2, y2]
- target_ratio: str
- type: str
- origin_size: tuple of int (width, height)
"""
item = self.metadata[idx]
# Load image
image_path = self.image_dir / item['file_name']
image = Image.open(image_path).convert('RGB')
# Apply transform if specified
if self.transform:
image = self.transform(image)
return {
'image': image,
'prompts': item['caption'],
'target_bbox': item['target_crop']['target_bbox'],
'target_ratio': item['target_crop']['target_ratio'],
'type': item['type'],
'origin_size': (item['origin_width'], item['origin_height']),
'idx': idx
}
class _CyclingDataLoader:
"""Infinite DataLoader iterator with epoch-aware DistributedSampler state."""
def __init__(self, dataloader):
self.dataloader = dataloader
self.epoch = 0
self._iterator = None
def __iter__(self):
return self
def _new_iterator(self):
if hasattr(self.dataloader.sampler, "set_epoch"):
self.dataloader.sampler.set_epoch(self.epoch)
self._iterator = iter(self.dataloader)
def __next__(self):
if self._iterator is None:
self._new_iterator()
try:
return next(self._iterator)
except StopIteration:
self.epoch += 1
self._new_iterator()
return next(self._iterator)
def seek(self, consumed_batches):
"""Position the next item at an absolute batch counter."""
length = len(self.dataloader)
if length <= 0:
raise ValueError("cannot cycle an empty DataLoader")
self.epoch, offset = divmod(int(consumed_batches), length)
self._new_iterator()
for _ in range(offset):
next(self._iterator)
return offset
def cycle(dl):
return _CyclingDataLoader(dl)
# Compatibility re-exports keep legacy callers on ``utils.dataset`` while the
# prompt/video implementation remains isolated from the LMDB dataset module.
from utils.openvid_dataset import ( # noqa: E402
OpenVidPromptVideoDataset,
collate_openvid_prompt_video,
)
|