app.py

import os, json
import gradio as gr
import huggingface_hub, numpy as np, onnxruntime as rt, pandas as pd
from PIL import Image
from huggingface_hub import login

from translator import translate_texts

# ------------------------------------------------------------------
# 模型配置
# ------------------------------------------------------------------
MODEL_REPO      = "SmilingWolf/wd-eva02-large-tagger-v3"
MODEL_FILENAME  = "model.onnx"
LABEL_FILENAME  = "selected_tags.csv"

HF_TOKEN = os.environ.get("HF_TOKEN", "")
if HF_TOKEN:
    login(token=HF_TOKEN)
else:
    print("⚠️ 未检测到 HF_TOKEN，私有模型可能下载失败")

# ------------------------------------------------------------------
# Tagger 类
# ------------------------------------------------------------------
class Tagger:
    def __init__(self):
        self.hf_token   = HF_TOKEN
        self._load_model_and_labels()

    def _load_model_and_labels(self):
        label_path = huggingface_hub.hf_hub_download(
            MODEL_REPO, LABEL_FILENAME, token=self.hf_token
        )
        model_path = huggingface_hub.hf_hub_download(
            MODEL_REPO, MODEL_FILENAME, token=self.hf_token
        )

        tags_df           = pd.read_csv(label_path)
        self.tag_names    = tags_df["name"].tolist()
        self.categories   = {
            "rating":    np.where(tags_df["category"] == 9)[0],
            "general":   np.where(tags_df["category"] == 0)[0],
            "character": np.where(tags_df["category"] == 4)[0],
        }
        self.model        = rt.InferenceSession(model_path)
        self.input_size   = self.model.get_inputs()[0].shape[1]

    # ------------------------- preprocess -------------------------
    def _preprocess(self, img: Image.Image) -> np.ndarray:
        if img.mode != "RGB":
            img = img.convert("RGB")
        size   = max(img.size)
        canvas = Image.new("RGB", (size, size), (255, 255, 255))
        canvas.paste(img, ((size - img.width)//2, (size - img.height)//2))
        if size != self.input_size:
            canvas = canvas.resize((self.input_size, self.input_size), Image.BICUBIC)
        return np.array(canvas)[:, :, ::-1].astype(np.float32)  # to BGR

    # --------------------------- predict --------------------------
    def predict(self, img: Image.Image,
                gen_th: float = 0.35,
                char_th: float = 0.85):
        inp_name  = self.model.get_inputs()[0].name
        outputs   = self.model.run(None, {inp_name: self._preprocess(img)[None, ...]})[0][0]

        res = {"ratings": {}, "general": {}, "characters": {}}

        for idx in self.categories["rating"]:
            res["ratings"][self.tag_names[idx].replace("_", " ")] = float(outputs[idx])

        for idx in self.categories["general"]:
            if outputs[idx] > gen_th:
                res["general"][self.tag_names[idx].replace("_", " ")] = float(outputs[idx])

        for idx in self.categories["character"]:
            if outputs[idx] > char_th:
                res["characters"][self.tag_names[idx].replace("_", " ")] = float(outputs[idx])

        res["general"] = dict(sorted(res["general"].items(),
                                     key=lambda kv: kv[1],
                                     reverse=True))
        return res

# ------------------------------------------------------------------
# Gradio UI
# ------------------------------------------------------------------
custom_css = """
.label-container {
    max-height: 300px;
    overflow-y: auto;
    border: 1px solid #ddd;
    padding: 10px;
    border-radius: 5px;
    background-color: #f9f9f9;
}
.tag-item {
    display: flex;
    justify-content: space-between;
    align-items: center;
    margin: 2px 0;
    padding: 2px 5px;
    border-radius: 3px;
    background-color: #fff;
    cursor: pointer;
}
.tag-item:hover {
    background-color: #f0f0f0;
}
.tag-en {
    font-weight: bold;
    color: #333;
}
.tag-zh {
    color: #666;
    margin-left: 10px;
}
.tag-score {
    color: #999;
    font-size: 0.9em;
}
.btn-container {
    margin-top: 20px;
}
.copy-btn {
    margin-top: 10px;
    padding: 5px 10px;
    background-color: #f0f0f0;
    border: 1px solid #ddd;
    border-radius: 4px;
    cursor: pointer;
    display: inline-flex;
    align-items: center;
    font-size: 0.9em;
}
.copy-btn:hover {
    background-color: #e0e0e0;
}
.copy-icon {
    margin-right: 5px;
    width: 16px;
    height: 16px;
}
.copied-message {
    display: none;
    color: #4CAF50;
    margin-left: 10px;
    font-size: 0.9em;
}
.note-text {
    color: #ff6b6b;
    font-size: 0.9em;
    padding: 5px;
    border-left: 3px solid #ff6b6b;
    margin-top: 15px;
    background-color: #fff5f5;
}
"""

js_code = """
function setupCopyFunctions() {
    // 为标签项设置点击复制
    document.querySelectorAll('.tag-item').forEach(item => {
        item.addEventListener('click', function() {
            const tagText = this.querySelector('.tag-en').textContent;
            navigator.clipboard.writeText(tagText).then(() => {
                // 显示临时复制成功提示
                const msg = document.createElement('span');
                msg.textContent = '已复制!';
                msg.style.color = '#4CAF50';
                msg.style.marginLeft = '5px';
                msg.style.fontSize = '0.8em';
                this.appendChild(msg);
                setTimeout(() => msg.remove(), 1000);
            });
        });
    });

    // 为汇总区域的复制按钮设置功能
    document.getElementById('copy-tags-btn').addEventListener('click', function() {
        const tagsText = document.getElementById('summary-text').value;
        navigator.clipboard.writeText(tagsText).then(() => {
            const copiedMsg = document.getElementById('copied-message');
            copiedMsg.style.display = 'inline';
            setTimeout(() => {
                copiedMsg.style.display = 'none';
            }, 2000);
        });
    });
}

// 在DOM加载完成或更新后调用设置函数
function onUiUpdate() {
    setupCopyFunctions();
}

document.addEventListener('DOMContentLoaded', onUiUpdate);
"""

with gr.Blocks(theme=gr.themes.Soft(), title="AI 图像标签分析器", css=custom_css, js=js_code) as demo:
    gr.Markdown("# 🖼️ AI 图像标签分析器")
    gr.Markdown("上传图片自动识别标签，并可一键翻译成中文")
    gr.Markdown("<div class='note-text'>⚠️ 注意：角色识别仅支持推测2024年2月以前的角色</div>", elem_id="character-notice")

    with gr.Row():
        with gr.Column(scale=1):
            img_in = gr.Image(type="pil", label="上传图片")
            with gr.Accordion("⚙️ 高级设置", open=False):
                gen_slider  = gr.Slider(0, 1, 0.35,
                                        label="通用标签阈值", info="越高→标签更少更准")
                char_slider = gr.Slider(0, 1, 0.85,
                                        label="角色标签阈值", info="推荐保持较高阈值")
                show_zh = gr.Checkbox(True, label="显示中文翻译")
                
                gr.Markdown("### 汇总设置")
                with gr.Row():
                    sum_general = gr.Checkbox(True, label="通用标签")
                    sum_char = gr.Checkbox(True, label="角色标签")
                    sum_rating = gr.Checkbox(False, label="评分标签")
                sum_sep = gr.Dropdown(["逗号", "换行", "空格"], value="逗号", label="分隔符")

            btn = gr.Button("开始分析", variant="primary", elem_classes=["btn-container"])
            processing_info = gr.Markdown("", visible=False)

        with gr.Column(scale=2):
            with gr.Tabs():
                with gr.TabItem("🏷️ 通用标签"):
                    out_general = gr.HTML(label="General Tags")
                with gr.TabItem("👤 角色标签"):
                    out_char = gr.HTML(label="Character Tags")
                with gr.TabItem("⭐ 评分标签"):
                    out_rating = gr.HTML(label="Rating Tags")
            
            gr.Markdown("### 标签汇总")
            with gr.Row():
                out_summary = gr.Textbox(label="标签汇总", 
                                         placeholder="选择需要汇总的标签类别...",
                                         lines=3,
                                         elem_id="summary-text")
            
            # 添加复制按钮的HTML
            copy_btn_html = gr.HTML("""
            <div style="display: flex; align-items: center;">
                <button id="copy-tags-btn" class="copy-btn">
                    <svg class="copy-icon" xmlns="http://www.w3.org/2000/svg" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2">
                        <rect x="9" y="9" width="13" height="13" rx="2" ry="2"></rect>
                        <path d="M5 15H4a2 2 0 01-2-2V4a2 2 0 012-2h9a2 2 0 012 2v1"></path>
                    </svg>
                    复制标签
                </button>
                <span id="copied-message" class="copied-message">已复制！</span>
            </div>
            """)

    # ----------------- 处理回调 -----------------
    def format_tags_html(tags_dict, translations, show_translation=True):
        """格式化标签为HTML格式，添加点击复制功能"""
        if not tags_dict:
            return "<p>暂无标签</p>"
        
        html = '<div class="label-container">'
        for i, (tag, score) in enumerate(tags_dict.items()):
            # 添加可点击复制的标签项
            html += f'<div class="tag-item" title="点击复制标签">'
            html += f'<div><span class="tag-en">{tag}</span>'
            if show_translation and i < len(translations):
                html += f'<span class="tag-zh">({translations[i]})</span>'
            html += '</div>'
            html += f'<span class="tag-score">{score:.3f}</span>'
            html += '</div>'
        html += '</div>'
        return html

    def process(img, g_th, c_th, show_zh, sum_gen, sum_char, sum_rat, sep_type):
        # 开始处理，返回更新
        yield (
            gr.update(interactive=False, value="处理中..."),
            gr.update(visible=True, value="🔄 正在分析图像..."),
            "", "", "", ""
        )
        
        try:
            tagger = Tagger()
            res = tagger.predict(img, g_th, c_th)

            # 收集所有需要翻译的标签
            all_tags = []
            tag_categories = {
                "general": list(res["general"].keys()),
                "characters": list(res["characters"].keys()),
                "ratings": list(res["ratings"].keys())
            }
            
            if show_zh:
                for tags in tag_categories.values():
                    all_tags.extend(tags)
                
                # 批量翻译
                if all_tags:
                    translations = translate_texts(all_tags, src_lang="auto", tgt_lang="zh")
                else:
                    translations = []
            else:
                translations = []

            # 分配翻译结果
            translations_dict = {}
            offset = 0
            for category, tags in tag_categories.items():
                if show_zh and tags:
                    translations_dict[category] = translations[offset:offset+len(tags)]
                    offset += len(tags)
                else:
                    translations_dict[category] = []

            # 生成HTML输出
            general_html = format_tags_html(res["general"], translations_dict["general"], show_zh)
            char_html = format_tags_html(res["characters"], translations_dict["characters"], show_zh)
            rating_html = format_tags_html(res["ratings"], translations_dict["ratings"], show_zh)

            # 生成汇总文本 - 修改为仅显示英文标签，无注释
            summary_parts = []
            separators = {"逗号": ", ", "换行": "\n", "空格": " "}
            separator = separators[sep_type]
            
            all_tags = []
            if sum_gen and res["general"]:
                all_tags.extend(list(res["general"].keys()))
            
            if sum_char and res["characters"]:
                all_tags.extend(list(res["characters"].keys()))
            
            if sum_rat and res["ratings"]:
                all_tags.extend(list(res["ratings"].keys()))
            
            summary_text = separator.join(all_tags) if all_tags else "请选择要汇总的标签类别"

            # 完成处理，返回最终结果
            yield (
                gr.update(interactive=True, value="开始分析"),
                gr.update(visible=False),
                general_html,
                char_html,
                rating_html,
                summary_text
            )
            
        except Exception as e:
            # 出错时的处理
            yield (
                gr.update(interactive=True, value="开始分析"),
                gr.update(visible=True, value=f"❌ 处理失败: {str(e)}"),
                "", "", "", ""
            )

    # 绑定事件
    btn.click(
        process,
        inputs=[img_in, gen_slider, char_slider, show_zh, sum_general, sum_char, sum_rating, sum_sep],
        outputs=[btn, processing_info, out_general, out_char, out_rating, out_summary],
        show_progress=True
    )

# ------------------------------------------------------------------
# 启动
# ------------------------------------------------------------------
if __name__ == "__main__":
    demo.launch(server_name="0.0.0.0", server_port=7860)