cons_memory/caption_memory.py

"""记忆构造第二阶段：clip caption 生成 + OpenAI 文本 embedding（与 embedding 阶段解耦）。

参照 egolife_online_processing_.py 的两条约定：
  - caption 模型走 OpenAI 兼容 chat 接口（DashScope compatible-mode / 本地 vLLM serve /
    OpenAI 官方均可，--api-base 切换），默认 DashScope Qwen3-VL；
  - caption 文本 embedding 用 OpenAI text-embedding-3-small（1536 维），
    走官方 API（注意：不能设 OPENAI_BASE_URL 空串，需删除该变量）。

与 EgoLife 多层 caption（30s/5min/1h）的区别：我们保持扁平——每 clip 一次 VLM 调用，
同时产出 clip 级细粒度 caption 和逐关键帧短 caption（一次调用两级都拿，省一半调用量）。

用法：
  python caption_memory.py --dataset ovo --limit 2            # 冒烟
  python caption_memory.py --dataset streamarena --lang zh    # 中文视频用中文 caption
"""

import argparse
import base64
import json
import logging
import os
import re
import sys
import time
from concurrent.futures import ThreadPoolExecutor, as_completed

sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))

import numpy as np
import pandas as pd
import requests

from config import ConsMemoryConfig, LOG_DIR

logger = logging.getLogger("caption_memory")

# 默认 caption 后端：本地 vLLM Qwen3.6-35B-A3B-FP8（用户指定：记忆构造阶段不用 GPT）。
# 服务在 localhost:8010（2026-08-17 已在运行，vision 实测可用）；
# embedding 用 KEY1（sk-proj-yJ5Q... 即各 .env 里未编号的 OPENAI_API_KEY）。
DEFAULT_API_BASE = "http://localhost:8010/v1"
DEFAULT_MODEL = "Qwen3.6-35B-A3B-FP8"
EMBED_MODEL = "text-embedding-3-small"
EMBED_DIM = 1536

CAPTION_PROMPT_EN = """You are watching {n} keyframes sampled from a {dur:.0f}-second video clip (timestamps: {ts}).
Write fine-grained factual descriptions. Cover: visible entities and their states, actions in progress or completed, exact counts, any on-screen/printed text, and notable state changes.
Return ONLY valid JSON:
{{"frame_captions": [{n} short captions, one per keyframe in order], "clip_caption": "2-4 sentences describing what happens across this clip"}}"""

CAPTION_PROMPT_ZH = """你正在观看从一段 {dur:.0f} 秒视频片段中采样的 {n} 张关键帧（时间戳：{ts}）。
请写出细粒度的客观描述，覆盖：可见实体及其状态、正在进行或已完成的动作、精确数量、画面中出现的文字、显著的状态变化。
只返回合法 JSON：
{{"frame_captions": [{n} 条简短描述，按关键帧顺序一一对应], "clip_caption": "2-4 句话描述整个片段发生了什么"}}"""


def _b64_image(path: str) -> dict:
    with open(path, "rb") as f:
        data = base64.b64encode(f.read()).decode()
    return {"type": "image_url", "image_url": {"url": f"data:image/jpeg;base64,{data}"}}


def _safe_json(text: str):
    if not text:
        return None
    s = re.sub(r"```(?:json)?", "", text).strip().strip("`")
    m = re.search(r"\{.*\}", s, re.DOTALL)
    if not m:
        return None
    try:
        return json.loads(m.group(0))
    except json.JSONDecodeError:
        return None


GLOBAL_PROMPT_EN = """Below are chronological captions of a {dur:.0f}-second video, one per 10s clip with time ranges.

{captions}

Write a GLOBAL summary of the whole video (5-10 sentences): the setting, the main people/objects, and the storyline of what happens over time. Include approximate timestamps (e.g. "around 120s") for major transitions. Be factual."""

GLOBAL_PROMPT_ZH = """以下是一段 {dur:.0f} 秒视频的按时间顺序的分段描述（每 10 秒一条，带时间区间）。

{captions}

请写出整个视频的全局摘要（5-10 句）：场景、主要人物/物体、随时间发生的事件主线。重要转折处标注大致时间戳（如"约 120 秒处"）。保持客观。"""


class CaptionClient:
    def __init__(self, api_base: str, api_key: str, model: str, lang: str):
        self.url = api_base.rstrip("/") + "/chat/completions"
        self.headers = {"Authorization": f"Bearer {api_key}",
                        "Content-Type": "application/json"}
        self.model = model
        self.lang = lang
        self.tmpl = CAPTION_PROMPT_ZH if lang == "zh" else CAPTION_PROMPT_EN

    def text_call(self, prompt: str, max_tokens: int = 800, max_retries: int = 3) -> str:
        body = {"model": self.model,
                "messages": [{"role": "user", "content": prompt}],
                "max_tokens": max_tokens, "temperature": 0.1}
        for attempt in range(max_retries):
            try:
                r = requests.post(self.url, headers=self.headers, json=body, timeout=120)
                if r.status_code == 429:
                    time.sleep(5 * (attempt + 1)); continue
                r.raise_for_status()
                return r.json()["choices"][0]["message"]["content"] or ""
            except Exception as e:
                logger.warning("text_call 失败(第%d次): %s", attempt + 1, e)
                time.sleep(2 * (attempt + 1))
        return ""

    def global_caption(self, cdf, duration: float) -> str:
        """全视频 global caption：clip caption 分块层级汇总（长视频两段式）。"""
        tmpl = GLOBAL_PROMPT_ZH if self.lang == "zh" else GLOBAL_PROMPT_EN
        lines = [f"[{r['t_start']:.0f}-{r['t_end']:.0f}s] {r['caption']}"
                 for _, r in cdf.iterrows() if r.get("caption")]
        if not lines:
            return ""
        chunk_size = 60   # 60 条 clip caption ≈ 10min 视频一块
        if len(lines) <= chunk_size:
            return self.text_call(tmpl.format(dur=duration, captions="\n".join(lines)))
        # 长视频：先分块摘要，再汇总
        parts = []
        for s in range(0, len(lines), chunk_size):
            seg = lines[s:s + chunk_size]
            parts.append(self.text_call(tmpl.format(
                dur=duration, captions="\n".join(seg)), max_tokens=400))
        return self.text_call(tmpl.format(dur=duration, captions="\n".join(parts)))

    def caption_clip(self, kf_paths: list, kf_times: list, dur: float,
                     max_retries: int = 3) -> dict:
        ts = ", ".join(f"{t:.0f}s" for t in kf_times)
        prompt = self.tmpl.format(n=len(kf_paths), dur=dur, ts=ts)
        content = [_b64_image(p) for p in kf_paths] + [{"type": "text", "text": prompt}]
        body = {"model": self.model,
                "messages": [{"role": "user", "content": content}],
                "max_tokens": 1024, "temperature": 0.1}
        for attempt in range(max_retries):
            try:
                r = requests.post(self.url, headers=self.headers, json=body, timeout=120)
                if r.status_code == 429:
                    time.sleep(5 * (attempt + 1)); continue
                r.raise_for_status()
                text = r.json()["choices"][0]["message"]["content"]
                parsed = _safe_json(text)
                if parsed and parsed.get("clip_caption"):
                    fcs = parsed.get("frame_captions") or []
                    fcs = [str(c) for c in fcs][: len(kf_paths)]
                    while len(fcs) < len(kf_paths):
                        fcs.append("")
                    return {"clip_caption": str(parsed["clip_caption"]),
                            "frame_captions": fcs}
                # JSON 解析失败但有文本：整段当 clip caption 用（帧级留空）
                if text and attempt == max_retries - 1:
                    return {"clip_caption": text.strip()[:2000],
                            "frame_captions": [""] * len(kf_paths)}
            except Exception as e:
                logger.warning("caption 调用失败(第%d次): %s", attempt + 1, e)
                time.sleep(2 * (attempt + 1))
        return {"clip_caption": "", "frame_captions": [""] * len(kf_paths)}


def embed_texts(texts: list, api_key: str, batch: int = 32,
                max_retries: int = 3) -> np.ndarray:
    """OpenAI 官方 embedding。空文本返回零向量占位（保持行对齐）。"""
    url = "https://api.openai.com/v1/embeddings"
    headers = {"Authorization": f"Bearer {api_key}"}
    out = np.zeros((len(texts), EMBED_DIM), dtype=np.float32)
    todo = [(i, t) for i, t in enumerate(texts) if t and t.strip()]
    for s in range(0, len(todo), batch):
        chunk = todo[s:s + batch]
        for attempt in range(max_retries):
            try:
                r = requests.post(url, headers=headers, timeout=60, json={
                    "model": EMBED_MODEL,
                    "input": [t[:6000] for _, t in chunk]})
                r.raise_for_status()
                for (i, _), d in zip(chunk, r.json()["data"]):
                    v = np.array(d["embedding"], dtype=np.float32)
                    out[i] = v / (np.linalg.norm(v) + 1e-9)
                break
            except Exception as e:
                logger.warning("embedding 调用失败(第%d次): %s", attempt + 1, e)
                time.sleep(2 * (attempt + 1))
    return out


def process_bank(bank_dir: str, client: CaptionClient, openai_key: str,
                 concurrency: int) -> dict:
    cdf = pd.read_parquet(os.path.join(bank_dir, "clips.parquet"))
    fdf = pd.read_parquet(os.path.join(bank_dir, "frames.parquet"))

    # ---- 1. 并发生成 clip caption（含逐帧短 caption）----
    results = {}
    with ThreadPoolExecutor(max_workers=concurrency) as pool:
        futs = {}
        for ridx, row in cdf.iterrows():
            kf_paths = [os.path.join(bank_dir, p) for p in row["keyframe_paths"]]
            kf_times = sorted(
                fdf[fdf["clip_idx"] == row["clip_idx"]]["t"].tolist())[: len(kf_paths)]
            futs[pool.submit(client.caption_clip, kf_paths, kf_times or [row["t_start"]],
                             row["t_end"] - row["t_start"])] = ridx
        for fut in as_completed(futs):
            results[futs[fut]] = fut.result()

    cdf["caption"] = [results[i]["clip_caption"] for i in cdf.index]

    # 帧级 caption 按 (clip_idx, t 排序) 对齐回 frames
    frame_caps = {}
    for ridx, row in cdf.iterrows():
        sub = fdf[fdf["clip_idx"] == row["clip_idx"]].sort_values("t")
        for fi, cap in zip(sub.index, results[ridx]["frame_captions"]):
            frame_caps[fi] = cap
    fdf["caption"] = [frame_caps.get(i, "") for i in fdf.index]

    # ---- 2. OpenAI 文本 embedding ----
    c_emb = embed_texts(cdf["caption"].tolist(), openai_key)
    f_emb = embed_texts(fdf["caption"].tolist(), openai_key)
    cdf["cap_emb"] = [e.astype(np.float16).tobytes() for e in c_emb]
    fdf["cap_emb"] = [e.astype(np.float16).tobytes() for e in f_emb]

    cdf.to_parquet(os.path.join(bank_dir, "clips.parquet"))
    fdf.to_parquet(os.path.join(bank_dir, "frames.parquet"))

    meta_path = os.path.join(bank_dir, "meta.json")
    with open(meta_path) as f:
        meta = json.load(f)
    # global caption（用户要求：全视频摘要，带时间戳里程碑）
    meta["global_caption"] = client.global_caption(cdf, meta.get("duration_s", 0.0))
    meta["caption_done"] = True
    meta["caption_model"] = client.model
    meta["embed_model"] = EMBED_MODEL
    with open(meta_path, "w") as f:
        json.dump(meta, f, ensure_ascii=False, indent=2)
    n_empty = int((cdf["caption"].str.len() == 0).sum())
    return {"n_clips": len(cdf), "n_empty_caption": n_empty}


def main():
    ap = argparse.ArgumentParser()
    ap.add_argument("--dataset", required=True,
                    choices=["ovo", "streamingbench", "streamarena", "egoschema"])
    ap.add_argument("--bank-root", default="")
    ap.add_argument("--api-base", default=os.environ.get("CAPTION_API_BASE", DEFAULT_API_BASE))
    ap.add_argument("--model", default=os.environ.get("CAPTION_MODEL", DEFAULT_MODEL))
    ap.add_argument("--api-key", default="", help="默认读 DASHSCOPE_API_KEY / CAPTION_API_KEY")
    ap.add_argument("--lang", default="en", choices=["en", "zh"])
    ap.add_argument("--concurrency", type=int, default=8)
    ap.add_argument("--limit", type=int, default=0)
    ap.add_argument("--overwrite", action="store_true")
    ap.add_argument("--num-shards", type=int, default=1)
    ap.add_argument("--shard-idx", type=int, default=0)
    ap.add_argument("--only-ids", default="", help="只处理该文件列出的 video_id（插队用）")
    args = ap.parse_args()

    cfg = ConsMemoryConfig()
    root = os.path.join(args.bank_root or cfg.out_root, args.dataset)

    # 本地 vLLM 不校验 key，占位即可；外部 API 才需要真 key
    api_key = (args.api_key or os.environ.get("CAPTION_API_KEY") or "EMPTY")
    if "localhost" not in args.api_base and "127.0.0.1" not in args.api_base:
        assert api_key != "EMPTY", "外部 caption API 需要 key（--api-key 或 CAPTION_API_KEY）"
    # embedding 用 KEY1（EMBED_API_KEY 优先，其次 OPENAI_API_KEY）
    openai_key = (os.environ.get("EMBED_API_KEY", "")
                  or os.environ.get("OPENAI_API_KEY", ""))
    assert openai_key, "缺 embedding key：export EMBED_API_KEY=<KEY1> 或 OPENAI_API_KEY"

    os.makedirs(LOG_DIR, exist_ok=True)
    logging.basicConfig(
        level=logging.INFO,
        format="[%(asctime)s] %(levelname)s %(name)s: %(message)s",
        handlers=[logging.StreamHandler(),
                  logging.FileHandler(os.path.join(
                      LOG_DIR, f"evostream_caption_{args.dataset}.log"))])

    client = CaptionClient(args.api_base, api_key, args.model, args.lang)
    n_done = n_skip = n_fail = 0
    banks = sorted(os.listdir(root)) if os.path.isdir(root) else []
    if args.only_ids:
        wanted = {l.strip() for l in open(args.only_ids) if l.strip()}
        banks = [b for b in banks if b in wanted]
    banks = [b for i, b in enumerate(banks) if i % args.num_shards == args.shard_idx]
    for name in banks:
        bank_dir = os.path.join(root, name)
        meta_path = os.path.join(bank_dir, "meta.json")
        if not os.path.exists(meta_path):
            continue
        with open(meta_path) as f:
            meta = json.load(f)
        if meta.get("caption_done") and not args.overwrite:
            n_skip += 1
            continue
        if args.limit and n_done + n_fail >= args.limit:
            break
        try:
            t0 = time.time()
            stat = process_bank(bank_dir, client, openai_key, args.concurrency)
            n_done += 1
            logger.info("[%d done] %s clips=%d 空caption=%d 用时%.1fs",
                        n_done, name, stat["n_clips"], stat["n_empty_caption"],
                        time.time() - t0)
            if stat["n_empty_caption"] > stat["n_clips"] * 0.2:
                logger.warning("%s 空 caption 比例过高(%d/%d)，请检查 API",
                               name, stat["n_empty_caption"], stat["n_clips"])
        except Exception as e:
            n_fail += 1
            logger.error("FAIL %s: %s", name, e)

    logger.info("caption 完成: done=%d skip=%d fail=%d", n_done, n_skip, n_fail)


if __name__ == "__main__":
    main()