model/parser.py

# parser.py
# -*- coding: utf-8 -*-
"""
Step5 解析器（稳健版）
- 兼容中文全角/半角、圆括号/方括号、中文/英文逗号
- 抽取：增强后的指令 + 逐帧 bbox（与采样 t 序列对齐）
- 校验：0<=x1<x2<=1, 0<=y1<y2<=1；不合法则记为 None
"""

import re
from typing import Dict, List, Tuple

CN2NUM = {
    "一": 1, "二": 2, "三": 3, "四": 4, "五": 5,
    "六": 6, "七": 7, "八": 8, "九": 9, "十": 10
}

# 全角到半角的基础归一化（冒号/逗号/括号等）
FULL2HALF_MAP = str.maketrans({
    "（": "(", "）": ")", "【": "[", "】": "]", "［": "[", "］": "]",
    "，": ",", "：": ":", "；": ";", "。": ".", "、": ",", "｜": "|",
})

def _normalize_text(s: str) -> str:
    s = (s or "").strip().replace("\u00A0", " ").replace("\t", " ")
    # 去多空格
    s = re.sub(r"[ ]{2,}", " ", s)
    return s.translate(FULL2HALF_MAP)

def _extract_enhanced_instruction(step5_line: str) -> str:
    """
    支持两种写法：
    1) 修改后的编辑指令：{ ... }，任务一boxing序列：
    2) 修改后的编辑指令：...，任务一boxing序列：
    同时兼容中文/英文逗号，冒号，花括号可选
    """
    # 先尝试「有花括号」版本
    m = re.search(r"修改后的编辑指令[：:]\s*\{(?P<inst>.+?)\}\s*[，,]", step5_line)
    if m:
        return m.group("inst").strip()
    # 再尝试「无花括号」版本：取到 “任务一boxing序列：”/“任务1boxing序列：” 之前
    m = re.search(
        r"修改后的编辑指令[：:]\s*(?P<inst>.+?)\s*[，,]\s*任务\s*[一二三四五六七八九十\d]+\s*boxing序列[：:]",
        step5_line
    )
    return m.group("inst").strip() if m else ""

def _task_key_to_idx(tok: str) -> int:
    # "任务一"/"任务1"
    m = re.search(r"任务\s*([一二三四五六七八九十\d]+)", tok)
    if not m:
        return 1
    key = m.group(1)
    return CN2NUM.get(key, int(key) if key.isdigit() else 1)

def parse_step5(step5_line: str, t_list: List[int]) -> Tuple[str, Dict[str, List[Dict]]]:
    """
    返回 (enhanced_instruction, tasks_json)
    tasks_json 结构：
    {
      "任务一": [{"t": t0, "bbox":[x1,y1,x2,y2] or None}, ...],
      ...
    }
    """
    s = _normalize_text(step5_line)

    enhanced = _extract_enhanced_instruction(s)

    # 按任务切分：“…任务一boxing序列：[...]，任务二boxing序列：[...]…”
    # ⚠️ 兼容半角/全角冒号（归一化后为半角 ":"，但正则也允许全角）
    parts = re.split(r"(任务[一二三四五六七八九十\d]+\s*boxing序列[：:])", s)
    tasks_json: Dict[str, List[Dict]] = {}

    def parse_seq(seq_txt: str) -> List[Dict]:
        """
        解析形如：
        [(t=0,[(0.64,0.68,0.71,0.78)]);(t=1,[(0.64,0.68,0.71,0.78)])]
        允许 () 或 [] 包围 bbox，允许中文逗号/空格。
        """
        seq_txt = (seq_txt or "").strip()
        # 拆帧项，分隔符 ; 或 ；（也容忍多余空格）
        items = re.split(r"[;；]+", seq_txt.strip("[] "))
        out = []
        for it in items:
            it = it.strip()
            if not it:
                continue
            # 抓 t 值
            mt = re.search(r"t\s*=\s*(\d+)", it)
            if not mt:
                continue
            t = int(mt.group(1))

            # 抓 bbox：先定位一对括号内容，兼容 [] 或 ()
            mb = re.search(r"\[\s*([\d\.\s,，]+)\s*\]|\(\s*([\d\.\s,，]+)\s*\)", it)
            if not mb:
                out.append({"t": t, "bbox": None})
                continue

            nums_str = (mb.group(1) or mb.group(2) or "").replace("，", ",")
            nums = [x for x in re.split(r"\s*,\s*", nums_str) if x]
            if len(nums) != 4:
                out.append({"t": t, "bbox": None})
                continue

            try:
                x1, y1, x2, y2 = map(float, nums)
            except Exception:
                out.append({"t": t, "bbox": None})
                continue

            ok = (0.0 <= x1 < x2 <= 1.0) and (0.0 <= y1 < y2 <= 1.0)
            out.append({"t": t, "bbox": [x1, y1, x2, y2]} if ok else {"t": t, "bbox": None})
        return out

    # 从第二个元素开始，成对读取："任务Xboxing序列：" + 下一段（括号体）
    for i in range(1, len(parts), 2):
        head = parts[i]
        body = parts[i + 1] if i + 1 < len(parts) else ""
        idx = _task_key_to_idx(head)
        key = f"任务{idx}"
        tasks_json[key] = parse_seq(body)

    # 与 t_list 对齐（以采样顺序为准），缺失帧用 bbox=None 占位
    def align(seq: List[Dict]) -> List[Dict]:
        by_t = {e["t"]: e for e in seq}
        aligned = []
        for t in t_list:
            e = by_t.get(t)
            aligned.append(e if e else {"t": t, "bbox": None})
        return aligned

    tasks_json = {k: align(v) for k, v in tasks_json.items()}
    return enhanced, tasks_json