File size: 13,152 Bytes
22270be 5354088 22270be 5354088 22270be |
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 |
from __future__ import annotations
import logging
import math
import os
from typing import Dict, List, Optional
import torch
from PIL import Image
from torch.utils.data import DataLoader
from tqdm.autonotebook import tqdm
from transformers import AutoModelForVision2Seq, AutoProcessor
class GmeQwen2VL:
def __init__(
self,
model_name: str = "Alibaba-NLP/gme-Qwen2-VL-2B-Instruct",
model_path: Optional[str] = None,
device: str = "cuda" if torch.cuda.is_available() else "cpu",
min_image_tokens=256,
max_image_tokens=1280,
max_length=1800,
**kwargs,
) -> None:
model_name = model_path or model_name
self.base = AutoModelForVision2Seq.from_pretrained(
model_name, torch_dtype=torch.float16, **kwargs
)
self.base.eval()
self.normalize = True
self.device = device
min_pixels = min_image_tokens * 28 * 28
max_pixels = max_image_tokens * 28 * 28
self.max_length = max_length
self.processor = AutoProcessor.from_pretrained(
model_name, min_pixels=min_pixels, max_pixels=max_pixels, **kwargs
)
self.processor.tokenizer.padding_side = 'right'
self.defualt_instruction = 'You are a helpful assistant.'
self.sep = ' '
def forward(
self,
input_ids: Optional[torch.LongTensor] = None,
attention_mask: Optional[torch.Tensor] = None,
position_ids: Optional[torch.LongTensor] = None,
past_key_values: Optional[List[torch.FloatTensor]] = None,
inputs_embeds: Optional[torch.FloatTensor] = None,
pixel_values: Optional[torch.Tensor] = None,
# pixel_values_videos: Optional[torch.FloatTensor] = None,
image_grid_thw: Optional[torch.LongTensor] = None,
# video_grid_thw: Optional[torch.LongTensor] = None,
pooling_mask: Optional[torch.LongTensor] = None,
**kwargs
) -> torch.Tensor:
if inputs_embeds is None:
inputs_embeds = self.base.model.embed_tokens(input_ids)
if pixel_values is not None:
pixel_values = pixel_values.type(self.base.visual.get_dtype())
image_embeds = self.base.visual(pixel_values, grid_thw=image_grid_thw).to(inputs_embeds.device)
image_mask = input_ids == self.base.config.image_token_id
inputs_embeds[image_mask] = image_embeds
# if pixel_values_videos is not None:
# pixel_values_videos = pixel_values_videos.type(self.base.visual.get_dtype())
# video_embeds = self.base.visual(pixel_values_videos, grid_thw=video_grid_thw).to(inputs_embeds.device)
# video_mask = input_ids == self.base.config.video_token_id
# inputs_embeds[video_mask] = video_embeds
if attention_mask is not None:
attention_mask = attention_mask.to(inputs_embeds.device)
outputs = self.base.model(
input_ids=None,
position_ids=position_ids,
attention_mask=attention_mask,
past_key_values=past_key_values,
inputs_embeds=inputs_embeds,
)
pooling_mask = attention_mask if pooling_mask is None else pooling_mask
left_padding = (pooling_mask[:, -1].sum() == pooling_mask.shape[0]) # TODO
if left_padding:
embeddings = outputs.last_hidden_state[:, -1]
else:
sequence_lengths = pooling_mask.sum(dim=1) - 1
batch_size = outputs.last_hidden_state.shape[0]
embeddings = outputs.last_hidden_state[torch.arange(
batch_size, device=outputs.last_hidden_state.device
), sequence_lengths]
if self.normalize:
embeddings = torch.nn.functional.normalize(embeddings, p=2, dim=1)
return embeddings.contiguous()
def embed(self, texts: list[str], images: list[Image.Image], is_query=True, instruction=None, **kwargs):
self.base.to(self.device)
# Inputs must be batched
input_texts, input_images = list(), list()
for t, i in zip(texts, images):
if not is_query or instruction is None:
instruction = self.defualt_instruction
input_str = ''
if i is None:
input_images = None # All examples in the same batch are consistent
else:
input_str += '<|vision_start|><|image_pad|><|vision_end|>'
i = fetch_image(i)
input_images.append(i)
if t is not None:
input_str += t
msg = f'<|im_start|>system\n{instruction}<|im_end|>\n<|im_start|>user\n{input_str}<|im_end|>\n<|im_start|>assistant\n<|endoftext|>'
input_texts.append(msg)
inputs = self.processor(
text=input_texts,
images=input_images,
padding=True,
truncation=True,
max_length=self.max_length,
return_tensors='pt'
)
inputs = {k: v.to(self.device) for k, v in inputs.items()} # TODO
with torch.no_grad():
embeddings = self.forward(**inputs)
return embeddings
def encode(self, sentences: list[str], *, prompt_name=None, **kwargs):
return self.get_fused_embeddings(texts=sentences, prompt_name=prompt_name, **kwargs)
def encode_queries(self, queries: List[str], **kwargs):
embeddings = self.encode(queries, **kwargs)
return embeddings
def encode_corpus(self, corpus: List[Dict[str, str]], **kwargs):
if type(corpus) is dict:
sentences = [
(corpus["title"][i] + self.sep + corpus["text"][i]).strip()
if "title" in corpus
else corpus["text"][i].strip()
for i in range(len(corpus["text"]))
]
else:
sentences = [
(doc["title"] + self.sep + doc["text"]).strip() if "title" in doc else doc["text"].strip()
for doc in corpus
]
embeddings = self.encode(sentences, is_query=False, **kwargs)
return embeddings
def get_image_embeddings(self, images: list[Image.Image] | DataLoader, **kwargs):
return self.get_fused_embeddings(images=images, **kwargs)
def get_text_embeddings(self, texts: list[str], **kwargs):
return self.get_fused_embeddings(texts=texts, **kwargs)
def get_fused_embeddings(self, texts: list[str] = None, images: list[Image.Image] | DataLoader = None, **kwargs):
if isinstance(images, DataLoader):
image_loader = images
batch_size = image_loader.batch_size
image_loader.dataset.transform = None
else:
batch_size = kwargs.pop('batch_size', 32)
if images is None:
image_loader = None
else:
image_loader = DataLoader(
images,
batch_size=batch_size,
shuffle=False,
collate_fn=custom_collate_fn,
num_workers=min(math.floor(os.cpu_count() / 2), 8),
)
if texts is None:
assert image_loader is not None
n_batch = len(image_loader)
else:
n_batch = len(texts) // batch_size + int(len(texts) % batch_size > 0)
image_loader = image_loader or [None] * n_batch
all_embeddings = list()
none_batch = [None] * batch_size
show_progress_bar = kwargs.pop('show_progress_bar', True)
pbar = tqdm(total=n_batch, disable=not show_progress_bar, mininterval=1, miniters=10, desc='encode')
for n, img_batch in zip(range(0, n_batch * batch_size, batch_size), image_loader):
text_batch = none_batch if texts is None else texts[n: n+batch_size]
img_batch = none_batch if img_batch is None else img_batch
embeddings = self.embed(texts=text_batch, images=img_batch, **kwargs)
pbar.update(1)
all_embeddings.append(embeddings.cpu())
pbar.close()
all_embeddings = torch.cat(all_embeddings, dim=0)
return all_embeddings
def custom_collate_fn(batch):
return batch
### Copied from qwen_vl_utils.vision_process.py
import base64
from io import BytesIO
import requests
IMAGE_FACTOR = 28
MIN_PIXELS = 4 * 28 * 28
MAX_PIXELS = 16384 * 28 * 28
MAX_RATIO = 200
def round_by_factor(number: int, factor: int) -> int:
"""Returns the closest integer to 'number' that is divisible by 'factor'."""
return round(number / factor) * factor
def ceil_by_factor(number: int, factor: int) -> int:
"""Returns the smallest integer greater than or equal to 'number' that is divisible by 'factor'."""
return math.ceil(number / factor) * factor
def floor_by_factor(number: int, factor: int) -> int:
"""Returns the largest integer less than or equal to 'number' that is divisible by 'factor'."""
return math.floor(number / factor) * factor
def smart_resize(
height: int, width: int, factor: int = IMAGE_FACTOR, min_pixels: int = MIN_PIXELS, max_pixels: int = MAX_PIXELS
) -> tuple[int, int]:
"""
Rescales the image so that the following conditions are met:
1. Both dimensions (height and width) are divisible by 'factor'.
2. The total number of pixels is within the range ['min_pixels', 'max_pixels'].
3. The aspect ratio of the image is maintained as closely as possible.
"""
h_bar = max(factor, round_by_factor(height, factor))
w_bar = max(factor, round_by_factor(width, factor))
if h_bar * w_bar > max_pixels:
beta = math.sqrt((height * width) / max_pixels)
h_bar = floor_by_factor(height / beta, factor)
w_bar = floor_by_factor(width / beta, factor)
elif h_bar * w_bar < min_pixels:
beta = math.sqrt(min_pixels / (height * width))
h_bar = ceil_by_factor(height * beta, factor)
w_bar = ceil_by_factor(width * beta, factor)
if max(h_bar, w_bar) / min(h_bar, w_bar) > MAX_RATIO:
logging.warning(
f"Absolute aspect ratio must be smaller than {MAX_RATIO}, got {max(h_bar, w_bar) / min(h_bar, w_bar)}"
)
if h_bar > w_bar:
h_bar = w_bar * MAX_RATIO
else:
w_bar = h_bar * MAX_RATIO
return h_bar, w_bar
def fetch_image(image: str | Image.Image, size_factor: int = IMAGE_FACTOR) -> Image.Image:
image_obj = None
if isinstance(image, Image.Image):
image_obj = image
elif image.startswith("http://") or image.startswith("https://"):
image_obj = Image.open(requests.get(image, stream=True).raw)
elif image.startswith("file://"):
image_obj = Image.open(image[7:])
elif image.startswith("data:image"):
if "base64," in image:
_, base64_data = image.split("base64,", 1)
data = base64.b64decode(base64_data)
image_obj = Image.open(BytesIO(data))
else:
image_obj = Image.open(image)
if image_obj is None:
raise ValueError(f"Unrecognized image input, support local path, http url, base64 and PIL.Image, got {image}")
image = image_obj.convert("RGB")
## resize
# if "resized_height" in ele and "resized_width" in ele:
# resized_height, resized_width = smart_resize(
# ele["resized_height"],
# ele["resized_width"],
# factor=size_factor,
# )
# else:
width, height = image.size
# min_pixels = ele.get("min_pixels", MIN_PIXELS)
# max_pixels = ele.get("max_pixels", MAX_PIXELS)
resized_height, resized_width = smart_resize(
height,
width,
factor=size_factor,
min_pixels=MIN_PIXELS,
max_pixels=MAX_PIXELS,
)
image = image.resize((resized_width, resized_height))
return image
###
if __name__ == '__main__':
texts = [
"What kind of car is this?",
"The Tesla Cybertruck is a battery electric pickup truck built by Tesla, Inc. since 2023."
]
images = [
'https://en.wikipedia.org/wiki/File:Tesla_Cybertruck_damaged_window.jpg',
'https://en.wikipedia.org/wiki/File:2024_Tesla_Cybertruck_Foundation_Series,_front_left_(Greenwich).jpg',
]
gme = GmeQwen2VL("Alibaba-NLP/gme-Qwen2-VL-2B-Instruct")
# Single-modal embedding
e_text = gme.get_text_embeddings(texts=texts)
e_image = gme.get_image_embeddings(images=images)
print((e_text * e_image).sum(-1))
## tensor([0.2281, 0.6001], dtype=torch.float16)
# How to set embedding instruction
e_query = gme.get_text_embeddings(texts=texts, instruction='Find an image that matches the given text.')
# If is_query=False, we always use the default instruction.
e_corpus = gme.get_image_embeddings(images=images, is_query=False)
print((e_query * e_corpus).sum(-1))
## tensor([0.2433, 0.7051], dtype=torch.float16)
# Fused-modal embedding
e_fused = gme.get_fused_embeddings(texts=texts, images=images)
print((e_fused[0] * e_fused[1]).sum())
## tensor(0.6108, dtype=torch.float16)
|