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-swinv2-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
# ------------------------------------------------------------------
with gr.Blocks(theme=gr.themes.Soft(), title="AI 图像标签分析器 + 翻译") 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")

    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")
                lang_drop   = gr.Dropdown(["zh", "en"], value="zh",
                                          label="翻译目标语言",
                                          info="当前仅内置中 / 英")

            btn = gr.Button("开始分析", variant="primary")

        with gr.Column(scale=2):
            with gr.Tabs():
                with gr.TabItem("🏷️ 通用标签 (英文)"):
                    out_general = gr.Label(label="General Tags")
                with gr.TabItem("👤 角色标签 (英文)"):
                    out_char    = gr.Label(label="Character Tags")
                with gr.TabItem("⭐ 评分标签 (英文)"):
                    out_rating  = gr.Label(label="Rating Tags")
                with gr.TabItem("🌐 翻译结果"):
                    out_trans   = gr.Textbox(label="翻译后的标签",
                                             placeholder="翻译结果显示在此处")

    # ----------------- 处理回调 -----------------
    def process(img, g_th, c_th, tgt_lang):
        tagger   = Tagger()
        res      = tagger.predict(img, g_th, c_th)

        # =========== 组织翻译 ===========
        tags_to_translate = list(res["general"].keys()) + list(res["characters"].keys())
        translations      = translate_texts(tags_to_translate, src_lang="auto", tgt_lang=tgt_lang)
        # 拼接字符串
        trans_str = ", ".join(translations)

        return {
            out_general: res["general"],
            out_char:    res["characters"],
            out_rating:  res["ratings"],
            out_trans:   trans_str
        }

    btn.click(
        process,
        inputs=[img_in, gen_slider, char_slider, lang_drop],
        outputs=[out_general, out_char, out_rating, out_trans]
    )

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