app.py

import re
import gradio as gr
import torch
from transformers import PreTrainedTokenizerFast, T5ForConditionalGeneration

# ✅ KoT5 요약 모델
MODEL_NAME = "psyche/KoT5-summarization"

tokenizer = PreTrainedTokenizerFast.from_pretrained(MODEL_NAME)
model = T5ForConditionalGeneration.from_pretrained(MODEL_NAME)

# CPU 동적 양자화 적용
try:
    model = torch.quantization.quantize_dynamic(
        model, {torch.nn.Linear}, dtype=torch.qint8
    )
except:
    pass

model.eval()

# ===== 유틸 =====
def normalize_text(text: str) -> str:
    return re.sub(r"\s+", " ", text).strip()

def split_into_sentences(text: str):
    text = text.replace("\n", " ")
    parts = re.split(r"(?<=[\.!?])\s+", text)
    return [p.strip() for p in parts if p.strip()]

def token_length(s: str) -> int:
    return len(tokenizer.encode(s, add_special_tokens=False))

def chunk_by_tokens(sentences, max_tokens=900):
    chunks, cur, cur_tokens = [], [], 0
    for s in sentences:
        tl = token_length(s)
        if tl > max_tokens:
            piece_size = max(200, int(len(s) * (max_tokens / tl)))
            for i in range(0, len(s), piece_size):
                sub = s[i:i+piece_size]
                if sub.strip():
                    chunks.append(sub.strip())
            cur, cur_tokens = [], 0
            continue
        if cur_tokens + tl <= max_tokens:
            cur.append(s)
            cur_tokens += tl
        else:
            if cur:
                chunks.append(" ".join(cur))
            cur, cur_tokens = [s], tl
    if cur:
        chunks.append(" ".join(cur))
    return chunks

# ===== 반복 제거 =====
def derpeat(text: str) -> str:
    text = re.sub(r'(.)\1{2,}', r'\1\1', text)  # 단일 문자 3회 이상 반복 → 2회
    text = re.sub(r'(\b\w+\b)(\s+\1){1,}', r'\1', text)  # 단어 반복 제거
    text = re.sub(r'([\.!?\-~])\1{2,}', r'\1\1', text)  # 구두점 반복 축소
    return text.strip()

# ===== 요약 =====
def approx_tokens_from_chars(n_chars: int) -> int:
    return max(1, n_chars // 2)  # 한글 대략 1토큰 ≈ 2문자

def summarize_raw_t5(input_text: str, target_chars: int, input_tokens: int) -> str:
    safe_target_chars = min(target_chars, max(120, int(len(input_text) * 0.9)))
    max_new = max(40, min(approx_tokens_from_chars(safe_target_chars), 300))

    if input_tokens <= 200:
        max_new = min(max_new, max(40, int(input_tokens * 0.6)))
    if input_tokens <= 60:
        max_new = min(max_new, 60)

    input_ids = tokenizer.encode(input_text, return_tensors="pt", truncation=True, max_length=1024)

    with torch.no_grad():
        summary_ids = model.generate(
            input_ids,
            max_new_tokens=max_new,
            do_sample=True,
            top_p=0.92,
            temperature=0.7,
            num_beams=1,
            no_repeat_ngram_size=4,
            encoder_no_repeat_ngram_size=4,
            repetition_penalty=1.2,
            renormalize_logits=True,
            early_stopping=True
        )
    return tokenizer.decode(summary_ids[0], skip_special_tokens=True)

def apply_style_prompt_t5(text: str, mode: str, final: bool=False) -> str:
    if mode == "concise":
        tag = "간결 요약:"
    elif mode == "explanatory":
        tag = "설명 요약:"
    else:
        tag = "불릿 요약:"
    guide = ""
    if final:
        guide = " (원래 문서의 순서를 유지하고 중복을 제거하세요.)"
    return f"{tag}{guide}\n{text}"

def postprocess_strict(summary: str, mode: str) -> str:
    s = summary.strip()
    s = re.sub(r"\s+", " ", s)
    s = derpeat(s)
    seen, outs = set(), []
    for sent in re.split(r"(?<=[\.!?])\s+", s):
        ss = sent.strip()
        if ss and ss not in seen:
            seen.add(ss)
            outs.append(ss)
    s = " ".join(outs)
    if mode == "bullets":
        parts = [p for p in outs if p]
        s = "\n".join([f"- {p}" for p in parts[:12]])
    return s

def summarize_long(text: str, target_chars: int, mode: str):
    text = normalize_text(text)
    if not text:
        return "⚠️ 요약할 텍스트를 입력하세요."

    approx_tokens = token_length(text)

    if approx_tokens <= 60:
        prompt = apply_style_prompt_t5(text, mode, final=False)
        out = summarize_raw_t5(prompt, min(target_chars, 300), approx_tokens)
        return postprocess_strict(out, mode)

    if approx_tokens <= 1000:
        prompt = apply_style_prompt_t5(text, mode, final=False)
        out = summarize_raw_t5(prompt, target_chars, approx_tokens)
        return postprocess_strict(out, mode)

    sentences = split_into_sentences(text)
    chunks = chunk_by_tokens(sentences, max_tokens=900)

    partial_summaries = []
    per_chunk_chars = max(180, int(target_chars * 1.2 / max(1, len(chunks))))
    for c in chunks:
        prompt = apply_style_prompt_t5(c, mode, final=False)
        psum = summarize_raw_t5(prompt, per_chunk_chars, token_length(c))
        partial_summaries.append(psum)

    merged = normalize_text(" ".join(partial_summaries))
    merged = derpeat(merged)

    final_prompt = apply_style_prompt_t5(merged, mode, final=True)
    final = summarize_raw_t5(final_prompt, target_chars, token_length(merged))
    return postprocess_strict(final, mode)

# ===== Gradio UI =====
def ui_summarize(text, target_len, style):
    mode = {"간결형":"concise", "설명형":"explanatory", "핵심 bullet":"bullets"}[style]
    return summarize_long(text, int(target_len), mode)

with gr.Blocks() as demo:
    gr.Markdown("## 📝 KoT5 한국어 요약기 (반복 억제 + 순서 보존)")
    with gr.Row():
        with gr.Column():
            input_text = gr.Textbox(label="원문 입력", lines=16)
            style = gr.Radio(["간결형", "설명형", "핵심 bullet"], value="간결형", label="요약 스타일")
            target_len = gr.Slider(300, 1500, value=1000, step=50, label="목표 요약 길이(문자)")
            btn = gr.Button("요약 실행")
        with gr.Column():
            output_text = gr.Textbox(label="요약 결과", lines=16)
    btn.click(ui_summarize, inputs=[input_text, target_len, style], outputs=output_text)

if __name__ == "__main__":
    demo.launch()