File size: 10,525 Bytes
c890b56 9a789e6 c890b56 9a789e6 c890b56 760cac6 c890b56 | 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 | import logging
import unicodedata
from pathlib import Path
from typing import Any, Dict, List, Optional, Union
import numpy as np
import torch
import torch.nn.functional as F
from PIL import Image
from safetensors.torch import load_file
from transformers import AutoConfig, AutoModel, AutoProcessor
from vl_utils.vision_process import process_vision_info
logger = logging.getLogger(__name__)
MAX_LENGTH = 8192
IMAGE_BASE_FACTOR = 16
IMAGE_FACTOR = IMAGE_BASE_FACTOR * 2
MIN_PIXELS = 4 * IMAGE_FACTOR * IMAGE_FACTOR
MAX_PIXELS = 1800 * IMAGE_FACTOR * IMAGE_FACTOR
FPS = 1
MAX_FRAMES = 64
FRAME_MAX_PIXELS = 768 * IMAGE_FACTOR * IMAGE_FACTOR
MAX_TOTAL_PIXELS = 10 * FRAME_MAX_PIXELS
def _resolve_weights_path(checkpoint_dir: str, weights_path: Optional[str] = None) -> str:
weights_path = weights_path or str(Path(checkpoint_dir) / "model.safetensors")
if Path(weights_path).exists():
return weights_path
from huggingface_hub import hf_hub_download
return hf_hub_download(repo_id=checkpoint_dir, filename=Path(weights_path).name)
def _load_backbone(
checkpoint_dir: str,
weights_path: Optional[str] = None,
**kwargs,
) -> torch.nn.Module:
"""Load Eddy weights from ``model.safetensors`` in the checkpoint folder."""
checkpoint_dir = str(checkpoint_dir)
weights_path = _resolve_weights_path(checkpoint_dir, weights_path)
dtype = kwargs.pop("torch_dtype", kwargs.pop("dtype", None))
config = AutoConfig.from_pretrained(checkpoint_dir, trust_remote_code=False)
model = AutoModel.from_config(config)
state_dict = {
key.removeprefix("model."): value
for key, value in load_file(weights_path).items()
}
model.load_state_dict(state_dict, strict=True)
if dtype is not None:
model = model.to(dtype=dtype)
return model
def sample_frames(
frames: List[Union[str, Image.Image]], num_segments: int, max_segments: int
) -> List[str]:
duration = len(frames)
frame_id_array = np.linspace(0, duration - 1, num_segments, dtype=int)
frame_id_list = frame_id_array.tolist()
last_frame_id = frame_id_list[-1]
sampled_frames = []
for frame_idx in frame_id_list:
try:
sampled_frames.append(frames[frame_idx])
except Exception:
break
while len(sampled_frames) < num_segments:
sampled_frames.append(frames[last_frame_id])
return sampled_frames[:max_segments]
class VLEmbedder:
def __init__(
self,
model_name_or_path: str,
weights_path: Optional[str] = None,
max_length: int = MAX_LENGTH,
min_pixels: int = MIN_PIXELS,
max_pixels: int = MAX_PIXELS,
total_pixels: int = MAX_TOTAL_PIXELS,
fps: float = FPS,
num_frames: int = MAX_FRAMES,
max_frames: int = MAX_FRAMES,
default_instruction: str = "Represent the user's input.",
**kwargs,
):
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
self.max_length = max_length
self.min_pixels = min_pixels
self.max_pixels = max_pixels
self.total_pixels = total_pixels
self.fps = fps
self.num_frames = num_frames
self.max_frames = max_frames
self.default_instruction = default_instruction
self.model = _load_backbone(
model_name_or_path,
weights_path=weights_path,
**kwargs,
).to(device)
self.processor = AutoProcessor.from_pretrained(
model_name_or_path,
trust_remote_code=True,
padding_side="right",
)
self.model.eval()
@torch.no_grad()
def forward(self, inputs: Dict[str, Any]) -> Dict[str, torch.Tensor]:
outputs = self.model(**inputs)
return {
"last_hidden_state": outputs.last_hidden_state,
"attention_mask": inputs.get("attention_mask"),
}
def _truncate_tokens(self, token_ids: List[int], max_length: int) -> List[int]:
if len(token_ids) <= max_length:
return token_ids
special_token_ids = set(self.processor.tokenizer.all_special_ids)
num_special = sum(1 for token_idx in token_ids if token_idx in special_token_ids)
num_non_special_to_keep = max_length - num_special
final_token_ids = []
non_special_kept_count = 0
for token_idx in token_ids:
if token_idx in special_token_ids:
final_token_ids.append(token_idx)
elif non_special_kept_count < num_non_special_to_keep:
final_token_ids.append(token_idx)
non_special_kept_count += 1
return final_token_ids
def format_model_input(
self,
text: Optional[str] = None,
image: Optional[Union[str, Image.Image]] = None,
video: Optional[Union[str, List[Union[str, Image.Image]]]] = None,
instruction: Optional[str] = None,
fps: Optional[float] = None,
max_frames: Optional[int] = None,
) -> List[Dict]:
if instruction:
instruction = instruction.strip()
if instruction and not unicodedata.category(instruction[-1]).startswith("P"):
instruction = instruction + "."
content = []
conversation = [
{
"role": "system",
"content": [{"type": "text", "text": instruction or self.default_instruction}],
},
{"role": "user", "content": content},
]
if not text and not image and not video:
content.append({"type": "text", "text": "NULL"})
return conversation
if video:
video_content = None
video_kwargs = {"total_pixels": self.total_pixels}
if isinstance(video, list):
video_content = video
if self.num_frames is not None or self.max_frames is not None:
video_content = sample_frames(
video_content, self.num_frames, self.max_frames
)
video_content = [
("file://" + ele if isinstance(ele, str) else ele)
for ele in video_content
]
elif isinstance(video, str):
video_content = (
video
if video.startswith(("http://", "https://"))
else "file://" + video
)
video_kwargs = {
"fps": fps or self.fps,
"max_frames": max_frames or self.max_frames,
}
else:
raise TypeError(f"Unrecognized video type: {type(video)}")
if video_content:
content.append({"type": "video", "video": video_content, **video_kwargs})
if image:
image_content = None
if isinstance(image, Image.Image):
image_content = image
elif isinstance(image, str):
image_content = (
image if image.startswith(("http", "oss")) else "file://" + image
)
else:
raise TypeError(f"Unrecognized image type: {type(image)}")
if image_content:
content.append(
{
"type": "image",
"image": image_content,
"min_pixels": self.min_pixels,
"max_pixels": self.max_pixels,
}
)
if text:
content.append({"type": "text", "text": text})
return conversation
def _preprocess_inputs(self, conversations: List[List[Dict]]) -> Dict[str, torch.Tensor]:
text = self.processor.apply_chat_template(
conversations, add_generation_prompt=True, tokenize=False
)
try:
images, video_inputs, video_kwargs = process_vision_info(
conversations,
image_patch_size=16,
return_video_metadata=True,
return_video_kwargs=True,
)
except Exception as e:
logger.error(f"Error in processing vision info: {e}")
images = None
video_inputs = None
video_kwargs = {"do_sample_frames": False}
text = self.processor.apply_chat_template(
[{"role": "user", "content": [{"type": "text", "text": "NULL"}]}],
add_generation_prompt=True,
tokenize=False,
)
if video_inputs is not None:
videos, video_metadata = zip(*video_inputs)
videos = list(videos)
video_metadata = list(video_metadata)
else:
videos, video_metadata = None, None
return self.processor(
text=text,
images=images,
videos=videos,
video_metadata=video_metadata,
truncation=True,
max_length=self.max_length,
padding=True,
do_resize=False,
return_tensors="pt",
**video_kwargs,
)
@staticmethod
def _pooling_last(
hidden_state: torch.Tensor, attention_mask: torch.Tensor
) -> torch.Tensor:
flipped_tensor = attention_mask.flip(dims=[1])
last_one_positions = flipped_tensor.argmax(dim=1)
col = attention_mask.shape[1] - last_one_positions - 1
row = torch.arange(hidden_state.shape[0], device=hidden_state.device)
return hidden_state[row, col]
def process(self, inputs: List[Dict[str, Any]], normalize: bool = True) -> torch.Tensor:
conversations = [
self.format_model_input(
text=ele.get("text"),
image=ele.get("image"),
video=ele.get("video"),
instruction=ele.get("instruction"),
fps=ele.get("fps"),
max_frames=ele.get("max_frames"),
)
for ele in inputs
]
processed_inputs = self._preprocess_inputs(conversations)
processed_inputs = {k: v.to(self.model.device) for k, v in processed_inputs.items()}
outputs = self.forward(processed_inputs)
embeddings = self._pooling_last(
outputs["last_hidden_state"], outputs["attention_mask"]
)
if normalize:
embeddings = F.normalize(embeddings, p=2, dim=-1)
return embeddings
|