Create get_vlm_caption.py
Browse files- get_vlm_caption.py +194 -0
get_vlm_caption.py
ADDED
|
@@ -0,0 +1,194 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import os
|
| 2 |
+
import time
|
| 3 |
+
import base64
|
| 4 |
+
import requests
|
| 5 |
+
import argparse
|
| 6 |
+
import json
|
| 7 |
+
from pathlib import Path
|
| 8 |
+
from multiprocessing import Pool
|
| 9 |
+
from tqdm import tqdm
|
| 10 |
+
from io import BytesIO
|
| 11 |
+
from PIL import Image
|
| 12 |
+
from urllib.parse import urljoin, urlparse
|
| 13 |
+
import random
|
| 14 |
+
|
| 15 |
+
# --- 默认配置 ---
|
| 16 |
+
DEFAULT_MODEL_PATH = "qwen2.5vl_7b_caption"
|
| 17 |
+
DEFAULT_IMAGE_DIR = "./images"
|
| 18 |
+
DEFAULT_PROMPT = "请用自然流畅的中文对以下图片进行全面而详细的描述。包括所有可见元素及其属性(如颜色、大小、形状、质地),它们的空间关系,以及任何显著特征或上下文。确保用自然流畅的中文描述清晰、生动,能够捕捉图片的每一个方面,不遗漏任何重要细节和NSFW内容。"
|
| 19 |
+
# 注意:多模态模型通常使用 chat/completions 端点
|
| 20 |
+
DEFAULT_API_URL = "http://localhost:8000/v1/chat/completions"
|
| 21 |
+
DEFAULT_MAX_PIXELS = 800 * 800 # 1M pixels,可以适当调大
|
| 22 |
+
CHECK_INTERVAL = 5 # seconds
|
| 23 |
+
|
| 24 |
+
def get_base_url(api_url):
|
| 25 |
+
"""从完整的 API URL 中提取基础 URL,例如 'http://localhost:8000/'"""
|
| 26 |
+
parsed_url = urlparse(api_url)
|
| 27 |
+
return f"{parsed_url.scheme}://{parsed_url.netloc}"
|
| 28 |
+
|
| 29 |
+
def is_server_running(api_url):
|
| 30 |
+
"""快速检查模型服务是否已经在运行"""
|
| 31 |
+
try:
|
| 32 |
+
# vLLM 的健康检查或模型列表端点
|
| 33 |
+
check_url = urljoin(get_base_url(api_url), "/health")
|
| 34 |
+
resp = requests.get(check_url, timeout=2)
|
| 35 |
+
if resp.status_code == 200:
|
| 36 |
+
return True
|
| 37 |
+
except requests.RequestException:
|
| 38 |
+
pass
|
| 39 |
+
return False
|
| 40 |
+
|
| 41 |
+
def wait_for_model_ready(api_url, timeout=300):
|
| 42 |
+
"""轮询检查模型服务是否启动并准备好"""
|
| 43 |
+
start_time = time.time()
|
| 44 |
+
check_url = urljoin(get_base_url(api_url), "/health")
|
| 45 |
+
print(f"⏳ 正在等待模型服务启动... (检查点: {check_url})")
|
| 46 |
+
while True:
|
| 47 |
+
try:
|
| 48 |
+
resp = requests.get(check_url)
|
| 49 |
+
if resp.status_code == 200:
|
| 50 |
+
print("✅ 模型服务已就绪!")
|
| 51 |
+
return True
|
| 52 |
+
except requests.RequestException:
|
| 53 |
+
pass
|
| 54 |
+
|
| 55 |
+
if time.time() - start_time > timeout:
|
| 56 |
+
print(f"❌ 模型服务启动超时(超过 {timeout} 秒)。")
|
| 57 |
+
return False
|
| 58 |
+
|
| 59 |
+
time.sleep(CHECK_INTERVAL)
|
| 60 |
+
print(f" ...仍在等待...")
|
| 61 |
+
|
| 62 |
+
def load_and_resize_image(image_path, max_pixels):
|
| 63 |
+
"""加载并根据需要缩放图像,然后返回 base64 编码的字符串"""
|
| 64 |
+
with Image.open(image_path) as img:
|
| 65 |
+
if img.mode != "RGB":
|
| 66 |
+
img = img.convert("RGB")
|
| 67 |
+
w, h = img.size
|
| 68 |
+
if w * h > max_pixels:
|
| 69 |
+
ratio = (max_pixels / (w * h)) ** 0.5
|
| 70 |
+
new_w, new_h = int(w * ratio), int(h * ratio)
|
| 71 |
+
img = img.resize((new_w, new_h), Image.LANCZOS)
|
| 72 |
+
|
| 73 |
+
buffer = BytesIO()
|
| 74 |
+
img.save(buffer, format="JPEG")
|
| 75 |
+
return base64.b64encode(buffer.getvalue()).decode("utf-8")
|
| 76 |
+
|
| 77 |
+
def generate_caption(args):
|
| 78 |
+
"""调用 API 为单个图片生成 caption"""
|
| 79 |
+
image_path, prompt, api_url, max_pixels, model_name = args
|
| 80 |
+
txt_path = Path(image_path).with_suffix(".txt")
|
| 81 |
+
|
| 82 |
+
if txt_path.exists() and txt_path.stat().st_size > 300:
|
| 83 |
+
return f"✅ 已跳过 (caption 已存在): {txt_path.name}"
|
| 84 |
+
|
| 85 |
+
try:
|
| 86 |
+
base64_image = load_and_resize_image(image_path, max_pixels)
|
| 87 |
+
|
| 88 |
+
# --- 这是符合 vLLM 多模态聊天补全 API 的正确 payload 格式 ---
|
| 89 |
+
payload = {
|
| 90 |
+
"messages": [
|
| 91 |
+
{
|
| 92 |
+
"role": "user",
|
| 93 |
+
"content": [
|
| 94 |
+
{"type": "text", "text": prompt},
|
| 95 |
+
{
|
| 96 |
+
"type": "image_url",
|
| 97 |
+
"image_url": {
|
| 98 |
+
"url": f"data:image/jpeg;base64,{base64_image}"
|
| 99 |
+
}
|
| 100 |
+
}
|
| 101 |
+
]
|
| 102 |
+
}
|
| 103 |
+
],
|
| 104 |
+
# "max_tokens": 1024, # 可以根据需要调整
|
| 105 |
+
# "temperature": 0.1, # 可以根据需要调整
|
| 106 |
+
}
|
| 107 |
+
|
| 108 |
+
response = requests.post(
|
| 109 |
+
api_url,
|
| 110 |
+
headers={"Content-Type": "application/json"},
|
| 111 |
+
data=json.dumps(payload)
|
| 112 |
+
)
|
| 113 |
+
|
| 114 |
+
if response.status_code == 200:
|
| 115 |
+
result = response.json()
|
| 116 |
+
# 从聊天补全的响应中提取内容
|
| 117 |
+
caption = result.get("choices", [{}])[0].get("message", {}).get("content", "").strip()
|
| 118 |
+
print(image_path)
|
| 119 |
+
print(caption)
|
| 120 |
+
|
| 121 |
+
if caption:
|
| 122 |
+
with open(txt_path, "w", encoding="utf-8") as f:
|
| 123 |
+
f.write(caption)
|
| 124 |
+
return f"✅ 成功生成: {txt_path.name}"
|
| 125 |
+
else:
|
| 126 |
+
return f"⚠️ 生成内容为空: {Path(image_path).name}"
|
| 127 |
+
else:
|
| 128 |
+
return f"⚠️ 生成失败: {Path(image_path).name}, 状态码: {response.status_code}, 响应: {response.text}"
|
| 129 |
+
|
| 130 |
+
except Exception as e:
|
| 131 |
+
return f"❌ 发生异常: {Path(image_path).name}, 错误: {str(e)}"
|
| 132 |
+
|
| 133 |
+
def collect_images(image_dir, extensions=(".jpg", ".jpeg", ".png", ".webp")):
|
| 134 |
+
"""递归收集所有图片文件的路径"""
|
| 135 |
+
image_paths = []
|
| 136 |
+
print(f"🔍 正在从 '{image_dir}' 目录中收集图片...")
|
| 137 |
+
for root, _, files in os.walk(image_dir):
|
| 138 |
+
for file in files:
|
| 139 |
+
if file.lower().endswith(extensions):
|
| 140 |
+
image_paths.append(os.path.join(root, file))
|
| 141 |
+
return image_paths
|
| 142 |
+
|
| 143 |
+
def main():
|
| 144 |
+
parser = argparse.ArgumentParser(description="为图片目录生成 caption (使用 vLLM 托管的多模态模型)")
|
| 145 |
+
parser.add_argument("--model-path", type=str, default=DEFAULT_MODEL_PATH, help="vLLM 加载的本地模型路径或 HuggingFace 名称")
|
| 146 |
+
parser.add_argument("--image-dir", type=str, default=DEFAULT_IMAGE_DIR, help="图片目录路径")
|
| 147 |
+
parser.add_argument("--prompt", type=str, default=DEFAULT_PROMPT, help="生成 caption 的提示词")
|
| 148 |
+
parser.add_argument("--api-url", type=str, default=DEFAULT_API_URL, help="vLLM 的聊天补全 API 地址")
|
| 149 |
+
parser.add_argument("--max-pixels", type=int, default=DEFAULT_MAX_PIXELS, help="图片最大像素数,超过此值会按比例缩放")
|
| 150 |
+
parser.add_argument("--num-process", type=int, default=18, help="用于处理图片的并发进程数")
|
| 151 |
+
|
| 152 |
+
args = parser.parse_args()
|
| 153 |
+
|
| 154 |
+
# --- 检查模型服务是否已运行 ---
|
| 155 |
+
if is_server_running(args.api_url):
|
| 156 |
+
print("✅ 检测到模型服务已在运行,直接使用。")
|
| 157 |
+
else:
|
| 158 |
+
print("ℹ️ 未检测到正在运行的模型服务,现在尝试启动...")
|
| 159 |
+
# 在后台启动 vLLM 模型服务
|
| 160 |
+
# 注意:这里的 --model 参数直接使用了 args.model_path,它将被用作 API 请求中的模型名称
|
| 161 |
+
command = f"nohup vllm serve {args.model_path} --max_model_len 3072 --trust-remote-code > /tmp/vllm.log 2>&1 &"
|
| 162 |
+
print(f"🚀 执行启动命令: {command}")
|
| 163 |
+
os.system(command)
|
| 164 |
+
|
| 165 |
+
# 轮询检测模型是否启动完成
|
| 166 |
+
if not wait_for_model_ready(args.api_url):
|
| 167 |
+
print("❌ 模型启动失败,请检查 /tmp/vllm_caption.log 文件获取错误详情。程序退出。")
|
| 168 |
+
exit(1)
|
| 169 |
+
|
| 170 |
+
# 收集所有图片文件
|
| 171 |
+
image_paths = collect_images(args.image_dir)
|
| 172 |
+
if not image_paths:
|
| 173 |
+
print("⚠️ 在指定目录中没有找到任何图片。")
|
| 174 |
+
return
|
| 175 |
+
|
| 176 |
+
random.shuffle(image_paths)
|
| 177 |
+
print(f"📸 找到 {len(image_paths)} 张图片,准备开始处理...")
|
| 178 |
+
|
| 179 |
+
# 多进程处理图像
|
| 180 |
+
# 将模型路径(作为模型名称)传递给处理函数
|
| 181 |
+
pool_args = [(img, args.prompt, args.api_url, args.max_pixels, args.model_path) for img in image_paths]
|
| 182 |
+
|
| 183 |
+
# 使用 args.num_process 控制并发数
|
| 184 |
+
with Pool(args.num_process) as pool:
|
| 185 |
+
# 使用 tqdm 显示进度条
|
| 186 |
+
for result in tqdm(pool.imap_unordered(generate_caption, pool_args), total=len(image_paths), desc="处理进度"):
|
| 187 |
+
# 只打印非成功的消息,避免刷屏
|
| 188 |
+
if not result.startswith("✅"):
|
| 189 |
+
print(result)
|
| 190 |
+
|
| 191 |
+
print("\n🎉 全部处理完成!")
|
| 192 |
+
|
| 193 |
+
if __name__ == "__main__":
|
| 194 |
+
main()
|