Create app.py

app.py

@@ -0,0 +1,224 @@
+# app.py
+# -*- coding: utf-8 -*-
+import os
+import re
+import io
+import pandas as pd
+import gradio as gr
+from PIL import Image, ImageDraw, ImageFont
+
+from kiwipiepy import Kiwi
+
+# 可选：用 HF Inference API 生成图（推荐 MVP）
+from huggingface_hub import InferenceClient
+
+# -----------------------
+# 配置
+# -----------------------
+VOCAB_PATH = "vocab.csv"
+FONT_KR = "fonts/NotoSansKR-Regular.otf"
+FONT_SC = "fonts/NotoSansSC-Regular.otf"
+
+# 你可以换成任何你能用的 text-to-image 模型
+# 注意：有些模型需要在 HF 上先点“Agree”许可
+IMAGE_MODEL_ID = os.getenv("IMAGE_MODEL_ID", "stabilityai/sdxl-turbo")
+
+HF_TOKEN = os.getenv("HF_TOKEN", "")  # 在 Spaces -> Settings -> Secrets 里加
+
+kiwi = Kiwi()
+
+def load_vocab():
+    if not os.path.exists(VOCAB_PATH):
+        return pd.DataFrame(columns=["lemma", "meaning_zh", "pos", "level"])
+    df = pd.read_csv(VOCAB_PATH)
+    df["lemma"] = df["lemma"].astype(str).str.strip()
+    return df
+
+VOCAB_DF = load_vocab()
+VOCAB_MAP = {row["lemma"]: row for _, row in VOCAB_DF.iterrows()}
+
+def safe_font(path, size):
+    # 没字体也别崩，先用默认字体（但可能不显示韩文）
+    if os.path.exists(path):
+        return ImageFont.truetype(path, size=size)
+    return ImageFont.load_default()
+
+def normalize_lemma(form, tag):
+    """
+    一个“够用版”的词形还原：
+    - 动词/形容词（VV/VA 等）尽量归一到 '다' 结尾
+    - 其他词保持原样
+    Kiwi 的 tag 很多，这里只粗略处理
+    """
+    s = form.strip()
+
+    # 去掉常见终结词尾/助词的残留（非常简化）
+    s = re.sub(r"(요|니다|까|나|냐|지|죠|죠\?|야)$", "", s)
+
+    # 如果是动词/形容词类，尽量补成词典形
+    if tag.startswith("V") or tag in ("VA", "VV", "VX", "VCP", "VCN"):
+        if not s.endswith("다"):
+            s = s + "다"
+    return s
+
+def analyze_sentence(sentence: str):
+    sentence = sentence.strip()
+    if not sentence:
+        return [], ""
+
+    # Kiwi 分析：取最优分析结果
+    analyzed = kiwi.analyze(sentence)
+    tokens = analyzed[0][0]  # list of Token
+
+    rows = []
+    for t in tokens:
+        form = t.form
+        tag = t.tag  # POS tag
+        lemma = normalize_lemma(form, tag)
+
+        # 词库匹配：先用 lemma，再退回 form
+        hit = VOCAB_MAP.get(lemma) or VOCAB_MAP.get(form)
+
+        rows.append({
+            "surface": form,
+            "pos": tag,
+            "lemma": lemma,
+            "meaning_zh": (hit["meaning_zh"] if hit is not None else ""),
+            "in_vocab": (hit is not None),
+            "level": (hit["level"] if hit is not None and "level" in hit else "")
+        })
+
+    # 生成一个“高亮句子”HTML：词库中有的绿色，没有的橙色
+    html_parts = []
+    for r in rows:
+        color = "#16a34a" if r["in_vocab"] else "#f59e0b"
+        html_parts.append(f"<span style='color:{color}; font-weight:700'>{r['surface']}</span>")
+    highlighted = " ".join(html_parts)
+
+    return rows, highlighted
+
+def ko_to_prompt(sentence: str, token_rows):
+    """
+    把句子变成更“可画”的 prompt（MVP：规则法）
+    你后面可以升级：先翻译成英文，再写 prompt
+    """
+    # 抽取名词类（Kiwi tag：NNG/NNP 等），最多取几个
+    nouns = [r["surface"] for r in token_rows if r["pos"].startswith("NN")]
+    nouns = nouns[:5]
+
+    # 一个偏“学习卡片插画”的风格 prompt
+    core = ", ".join(nouns) if nouns else "a person, daily life"
+    prompt = (
+        f"cute flat illustration, simple background, centered subject, "
+        f"learning flashcard style, {core}, no text"
+    )
+    return prompt
+
+def generate_image(prompt: str):
+    """
+    用 HF Inference API 生成图
+    """
+    if not HF_TOKEN:
+        # 没 token 就给个占位图，保证 UI 还能用
+        img = Image.new("RGB", (768, 512), "lightgray")
+        d = ImageDraw.Draw(img)
+        d.text((20, 20), "No HF_TOKEN set.\nSet it in Spaces Secrets to generate images.", fill="black")
+        return img
+
+    client = InferenceClient(model=IMAGE_MODEL_ID, token=HF_TOKEN)
+    # 返回 bytes
+    img_bytes = client.text_to_image(prompt)
+    return Image.open(io.BytesIO(img_bytes)).convert("RGB")
+
+def compose_card(main_img: Image.Image, ko: str, zh: str, vocab_rows):
+    """
+    拼一张长图卡片（接近你给的样式）
+    """
+    W, H = 1080, 1920
+    bg = Image.new("RGB", (W, H), (248, 235, 209))  # 米黄色
+    draw = ImageDraw.Draw(bg)
+
+    # 白色面板
+    panel_margin = 60
+    panel = (panel_margin, 70, W - panel_margin, H - 70)
+    draw.rounded_rectangle(panel, radius=35, fill="white", outline=(40, 40, 40), width=4)
+
+    # 顶部图片区
+    img_box = (panel_margin + 80, 150, W - panel_margin - 80, 650)
+    img_w = img_box[2] - img_box[0]
+    img_h = img_box[3] - img_box[1]
+    main_img = main_img.copy()
+    main_img.thumbnail((img_w, img_h))
+    paste_x = img_box[0] + (img_w - main_img.size[0]) // 2
+    paste_y = img_box[1] + (img_h - main_img.size[1]) // 2
+    bg.paste(main_img, (paste_x, paste_y))
+
+    # 字体
+    f_ko = safe_font(FONT_KR, 56)
+    f_zh = safe_font(FONT_SC, 42)
+    f_title = safe_font(FONT_SC, 48)
+    f_table = safe_font(FONT_SC, 36)
+
+    # 韩语句子（黄色描边感：简单做法画两层）
+    y_ko = 720
+    for dx, dy in [(-2,0),(2,0),(0,-2),(0,2)]:
+        draw.text((panel_margin+90+dx, y_ko+dy), ko, font=f_ko, fill="black")
+    draw.text((panel_margin+90, y_ko), ko, font=f_ko, fill=(250, 204, 21))  # 黄
+
+    # 中文翻译
+    y_zh = 810
+    draw.text((panel_margin+90, y_zh), zh, font=f_zh, fill=(40, 40, 40))
+
+    # 词汇拆解标题
+    y_t = 930
+    draw.text((panel_margin+90, y_t), "＊词汇拆解", font=f_title, fill="black")
+
+    # 表格
+    y = 1030
+    line_h = 60
+    # 列：surface / lemma / meaning
+    for r in vocab_rows:
+        if r["pos"].startswith("SP") or r["pos"].startswith("SF"):
+            continue
+        surface = r["surface"]
+        lemma = r["lemma"]
+        meaning = r["meaning_zh"] if r["meaning_zh"] else "（未收录）"
+        text = f"{surface}  →  {lemma}  |  {meaning}"
+        draw.text((panel_margin+90, y), text, font=f_table, fill=(30, 30, 30))
+        y += line_h
+        if y > H - 140:
+            break
+
+    return bg
+
+def run(sentence, zh_translation):
+    rows, highlighted = analyze_sentence(sentence)
+    prompt = ko_to_prompt(sentence, rows)
+    img = generate_image(prompt)
+    card = compose_card(img, sentence.strip(), zh_translation.strip(), rows)
+
+    df = pd.DataFrame(rows)[["surface", "pos", "lemma", "meaning_zh", "in_vocab", "level"]]
+    return card, df, highlighted, prompt
+
+with gr.Blocks() as demo:
+    gr.Markdown("## 韩语句子 → 自动拆词 → 自动配图 → 学习卡片（Hugging Face MVP）")
+
+    with gr.Row():
+        sentence = gr.Textbox(label="输入韩语句子", placeholder="例如：잠을 자도 피로가 안 풀리나?", lines=2)
+    with gr.Row():
+        zh = gr.Textbox(label="中文翻译（先手动填；后续可接自动翻译模型）", placeholder="例如：怎么睡了觉还是觉得很累？", lines=2)
+
+    btn = gr.Button("生成卡片")
+
+    with gr.Row():
+        out_img = gr.Image(label="生成的学习卡片", type="pil")
+    with gr.Row():
+        out_df = gr.Dataframe(label="词汇识别结果（可复制）", interactive=False)
+    with gr.Row():
+        out_html = gr.HTML(label="高亮句子（词库命中=绿，未命中=橙）")
+    with gr.Row():
+        out_prompt = gr.Textbox(label="用于生成图片的 prompt（你可以自己改）")
+
+    btn.click(fn=run, inputs=[sentence, zh], outputs=[out_img, out_df, out_html, out_prompt])
+
+demo.launch()