pyclip/gradio_example.py

import os
import gradio as gr
from pyclip import Clip, enum_devices, sys_init, sys_deinit, ClipDeviceType
import cv2
import glob
from PIL import Image
import tqdm
import argparse

if __name__ == '__main__':
    parser = argparse.ArgumentParser()
    parser.add_argument('--ienc', type=str, default='cnclip/cnclip_vit_l14_336px_vision_u16u8.axmodel')
    parser.add_argument('--tenc', type=str, default='cnclip/cnclip_vit_l14_336px_text_u16.axmodel')
    parser.add_argument('--vocab', type=str, default='cnclip/cn_vocab.txt')
    parser.add_argument('--isCN', type=int, default=1)
    parser.add_argument('--db_path', type=str, default='clip_feat_db_coco')
    parser.add_argument('--image_folder', type=str, default='coco_1000')
    args = parser.parse_args()

    image_folder = args.image_folder

    # 初始化
    print("可用设备:", enum_devices())
    sys_init(ClipDeviceType.axcl_device, 0)

    clip = Clip({
        'text_encoder_path': args.tenc,
        'image_encoder_path': args.ienc,
        'tokenizer_path': args.vocab,
        'db_path': args.db_path,
        'isCN': args.isCN
    })


    # 加载图片数据库（只做一次）
    image_files = glob.glob(os.path.join(image_folder, '*.jpg'))
    for image_file in tqdm.tqdm(image_files):
        filename = os.path.basename(image_file)
        if clip.contains_image(filename) == 1:
            continue
        img = cv2.imread(image_file)
        cv2.cvtColor(img, cv2.COLOR_BGR2RGB, img)
        clip.add_image(filename, img)

    # 工具函数：图片转 base64
    def img_to_pil(img_path):
        return Image.open(img_path).convert("RGB")

    # 主搜索函数
    def search_images(query, top_k):
        results = clip.match_text(query, top_k=top_k)
        images = []
        for filename, score in results:
            img_path = os.path.join(image_folder, filename)
            if os.path.exists(img_path):
                img = img_to_pil(img_path)
                images.append((img, f"{filename}  Score: {score:.4f}"))
        return images


    # Gradio界面
    with gr.Blocks() as demo:
        gr.Markdown("# 🔍 文搜图 Demo")

        with gr.Row():
            query_input = gr.Textbox(label="请输入文本查询")
            topk_input = gr.Number(value=25, precision=0, label="Top-K")
        search_btn = gr.Button("搜图")

        gallery = gr.Gallery(label="匹配结果", show_label=True, columns=4)

        search_btn.click(fn=search_images, inputs=[query_input, topk_input], outputs=gallery)

    # 启动
    ip = "0.0.0.0"
    demo.launch(server_name=ip, server_port=7860)

    # 关闭系统（你可加信号处理来自动关闭）
    import atexit
    atexit.register(lambda: sys_deinit(ClipDeviceType.axcl_device, 0))