File size: 42,644 Bytes
a46c532 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 | """
StreamingBench (Real_Time_Visual_Understanding) 建图入口 —— HyperVideoGraph + 三层超图
对齐 HyperMem stage1 的流式边界检测:
1. 按 interval_seconds 采样 sample_*/video.mp4 的帧
2. 用 HyperVideoGraph 自己的 VLM (load_caption_model) 每 K 帧生成 dense
frame captions (key = "DAY1-HH:MM:SS-HH:MM:SS")
3. 打包 dense_caption dict → hypergraph_builder.process_window_caption:
内部 _decompose_window 拆为单 FrameCaption → VideoEpisodeSegmenter
逐帧流式边界检测 → 命中边界 finalize Episode + 主题聚合 + 事实提取
4. flush 尾部 buffer, 保存 hypergraph.json
与 ovobench_hyper_processing.py 的差异:
* 输入是 StreamingBench 的 sample 目录 (sample_*/video.mp4)
* 使用通用视觉 caption prompt(OVO-Bench 的 prompt 已覆盖场景理解,
questions 字段留空或填多问题合集)
"""
import os
import sys
import argparse
import asyncio
import json
import logging
import time
import gc
import re
from typing import List, Dict, Optional
import torch
sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
from videorag._llm import (
LLMConfig, openai_embedding, gpt_4o_mini_complete,
)
from videorag._utils import always_get_an_event_loop
from videorag.video_processing import sample_frames_by_interval
from videorag.ovobench_prompt import OVOBENCH_PROMPTS, OUTPUT_FORMAT
from hypervideo.hyper_video_graph import HyperVideoGraph
from hypervideo.config import HyperVideoConfig
from hypervideo.prompts.frame_caption_prompt import (
FRAME_CAPTION_SINGLE_PROMPT,
FRAME_CAPTION_WINDOW_DETAILED_PROMPT,
FRAME_CAPTION_WINDOW_V2_PROMPT,
FRAME_CAPTION_WINDOW_V2_WITH_PRIOR_PROMPT,
)
from hypervideo.extractors.timing_utils import timed
from concurrent.futures import ThreadPoolExecutor
logging.basicConfig(
level=logging.INFO,
format='[%(asctime)s] %(levelname)s: %(message)s',
)
logger = logging.getLogger(__name__)
default_llm_config = LLMConfig(
embedding_func_raw=openai_embedding,
embedding_model_name="text-embedding-3-small",
embedding_dim=1536,
embedding_max_token_size=8192,
embedding_batch_num=32,
embedding_func_max_async=16,
query_better_than_threshold=0.2,
best_model_func_raw=gpt_4o_mini_complete,
best_model_name="gpt-5-mini",
best_model_max_token_size=32768,
best_model_max_async=16,
cheap_model_func_raw=gpt_4o_mini_complete,
cheap_model_name="gpt-4o-mini",
cheap_model_max_token_size=32768,
cheap_model_max_async=16,
)
def clear_gpu_memory():
if torch.cuda.is_available():
torch.cuda.empty_cache()
gc.collect()
def _seconds_to_hms(sec: float) -> str:
sec = max(0, int(sec))
h = sec // 3600
m = (sec % 3600) // 60
s = sec % 60
return f"{h:02d}:{m:02d}:{s:02d}"
def _timestamp_num_to_seconds(ts_num) -> float:
"""sample_frames_by_interval 在 time_span_info=None 时返回 HHMMSSCC 整数,
这里把它解析回秒数。"""
try:
ts = int(ts_num)
except (ValueError, TypeError):
return float(ts_num) if ts_num else 0.0
h = ts // 1_000_000
m = (ts % 1_000_000) // 10_000
s = (ts % 10_000) // 100
cs = ts % 100
return h * 3600 + m * 60 + s + cs / 100.0
def _safe_json_parse(text: str) -> Optional[dict]:
if not text or not isinstance(text, str):
return None
s = text.strip()
if s.startswith("```"):
lines = s.split("\n")
s = "\n".join(lines[1:-1]) if len(lines) > 2 else s
try:
return json.loads(s)
except json.JSONDecodeError:
m = re.search(r"\{.*\}", s, re.DOTALL)
if m:
try:
return json.loads(m.group(0))
except json.JSONDecodeError:
return None
return None
def _load_sample_questions(sample_dir: str) -> List[str]:
"""读取 sample_*/questions.json 里的问题文本(当作 caption attention hint)"""
qf = os.path.join(sample_dir, "questions.json")
if not os.path.exists(qf):
return []
try:
with open(qf, 'r', encoding='utf-8') as f:
data = json.load(f)
qs = []
for entry in data:
for q in entry.get('questions', []):
qtext = q.get('question', '')
if qtext:
qs.append(qtext)
return qs
except Exception as e:
logger.warning(f" 读取 {qf} 失败: {e}")
return []
def _hms_str_to_seconds(ts: str) -> int:
"""'00:13:22' 或 '0:09:35' → 秒数。"""
parts = ts.strip().split(':')
return sum(int(p) * 60 ** i for i, p in enumerate(reversed(parts)))
def _load_max_ts_per_sample(json_path: str) -> Dict[str, int]:
"""从 questions_real.json 聚合每个 sample 的 max(time_stamp) 秒数。
json_path 中每条 entry 的 video_path 形如 './videos/sample_X_real.mp4',
sample_id 取数字部分 'X'。返回 {sample_id_str: max_ts_seconds}。
"""
if not os.path.exists(json_path):
logger.warning(f"questions_real.json 不存在: {json_path}, max_ts 截断将关闭")
return {}
try:
with open(json_path, 'r', encoding='utf-8') as f:
data = json.load(f)
except Exception as e:
logger.warning(f"读取 {json_path} 失败: {e}")
return {}
max_ts: Dict[str, int] = {}
for entry in data:
vp = entry.get('video_path', '') or ''
# './videos/sample_348_real.mp4' → '348'
m = re.search(r'sample_(\d+)', vp)
if not m:
continue
sid = m.group(1)
for q in entry.get('questions', []):
ts_str = q.get('time_stamp', '')
if not ts_str:
continue
try:
secs = _hms_str_to_seconds(ts_str)
except Exception:
continue
if sid not in max_ts or secs > max_ts[sid]:
max_ts[sid] = secs
logger.info(f" [max_ts] 从 {os.path.basename(json_path)} 聚合 {len(max_ts)} 个 sample 的 max(time_stamp)")
return max_ts
def _load_questions_with_options_per_sample(json_path: str) -> Dict[str, List[Dict[str, str]]]:
"""从 questions_real.json 聚合每个 sample 的 (question + options + time_stamp) 列表。
用于问题感知建图: 把所有问题 (含选项) 注入到 caption / fact extraction prompt,
引导 VLM / LLM 关注与问题相关的视觉证据 / 事实。
Returns:
{sample_id: [{"qid": "Q1", "question": "<question + Options block>",
"time_stamp": "HH:MM:SS"}, ...]}
每个 sample 通常有 5 条 question, 按 time_stamp 升序。
"""
if not os.path.exists(json_path):
logger.warning(f"questions_real.json 不存在: {json_path}, --question_aware 不可用")
return {}
try:
with open(json_path, 'r', encoding='utf-8') as f:
data = json.load(f)
except Exception as e:
logger.warning(f"读取 {json_path} 失败: {e}")
return {}
by_sample: Dict[str, List[Dict[str, str]]] = {}
for entry in data:
vp = entry.get('video_path', '') or ''
m = re.search(r'sample_(\d+)', vp)
if not m:
continue
sid = m.group(1)
for q in entry.get('questions', []):
qtext = (q.get('question', '') or '').strip()
if not qtext:
continue
opts = q.get('options', []) or []
# 选项已带 'A. xxx' 前缀, 用 ' | ' 拼起来
opts_str = " | ".join(o.strip() for o in opts if isinstance(o, str))
q_full = f"{qtext} Options: {opts_str}" if opts_str else qtext
by_sample.setdefault(sid, []).append({
"qid": f"Q{len(by_sample[sid]) + 1}",
"question": q_full,
"time_stamp": q.get('time_stamp', '') or '',
})
# 按 time_stamp 升序排序
def _ts_to_sec(ts: str) -> int:
try:
return _hms_str_to_seconds(ts)
except Exception:
return 0
for sid in by_sample:
by_sample[sid].sort(key=lambda q: _ts_to_sec(q.get('time_stamp', '')))
# 重新分配 qid Q1..QN 按时间序
for i, q in enumerate(by_sample[sid]):
q['qid'] = f"Q{i+1}"
logger.info(
f" [qaware] 从 {os.path.basename(json_path)} 聚合 {len(by_sample)} 个 sample 的 questions+options"
)
return by_sample
def _make_questions_block(questions: List[Dict[str, str]]) -> str:
"""把 [{qid, question, time_stamp}, ...] 渲染成 prompt 用的多行块。
格式:
[Q1 @ 00:00:31] What ...? Options: A. ... | B. ...
[Q2 @ 00:03:26] Which ...? Options: ...
"""
if not questions:
return "(no active user questions)"
lines = []
for q in questions:
ts = q.get('time_stamp', '') or '?'
qid = q.get('qid', '') or 'Q?'
body = q.get('question', '') or ''
lines.append(f" [{qid} @ {ts}] {body}")
return "\n".join(lines)
# ==================== Caption 生成 (两种模式) ====================
def _caption_window_batch(
hyper_vg: HyperVideoGraph,
window_b64: List[str],
window_ranges: List[dict],
caption_prompt: str,
max_new_tokens: int,
day_num: int,
) -> Optional[dict]:
"""批量模式: 一次 VLM 调用返回 K 帧 dense caption。
返回 caption_dict = {"dense_caption": {...}, "description": "..."} 或 None。"""
for retry in range(3):
try:
caption_text = hyper_vg.mllm_response(
hyper_vg.video_llm, hyper_vg.processor,
caption_prompt, None,
base64_frames=window_b64,
max_new_tokens=max_new_tokens, has_image=True,
)
parsed = _safe_json_parse(caption_text)
if not (isinstance(parsed, dict) and "frames" in parsed
and isinstance(parsed["frames"], dict)):
logger.warning(f" [Caption-batch] 格式不合法 retry {retry+1}/3")
continue
dense_caption = {}
for frame_idx_str, frame_caption in parsed["frames"].items():
try:
fi = int(frame_idx_str)
except ValueError:
continue
if fi < 0 or fi >= len(window_ranges):
continue
f_start = _seconds_to_hms(_timestamp_num_to_seconds(window_ranges[fi]['start']))
f_end = _seconds_to_hms(_timestamp_num_to_seconds(window_ranges[fi]['end']))
dense_caption[f"DAY{day_num}-{f_start}-{f_end}"] = str(frame_caption)
return {
"dense_caption": dense_caption,
"description": str(parsed.get("caption", "")),
}
except Exception as e:
logger.error(f" [Caption-batch] 失败 retry {retry+1}/3: {e}")
continue
return None
def _caption_single_frame(
hyper_vg: HyperVideoGraph,
frame_b64: str,
max_new_tokens: int,
) -> Optional[str]:
"""单帧模式: 对一帧调用 VLM, 返回该帧的 caption 文本 (或 None)。"""
for retry in range(3):
try:
resp = hyper_vg.mllm_response(
hyper_vg.video_llm, hyper_vg.processor,
FRAME_CAPTION_SINGLE_PROMPT, None,
base64_frames=[frame_b64],
max_new_tokens=max_new_tokens, has_image=True,
)
parsed = _safe_json_parse(resp)
if isinstance(parsed, dict) and "caption" in parsed:
return str(parsed["caption"])
if isinstance(resp, str) and resp.strip():
return resp.strip()
except Exception as e:
logger.warning(f" [Caption-single] 失败 retry {retry+1}/3: {e}")
return None
def _caption_window_per_frame(
hyper_vg: HyperVideoGraph,
window_b64: List[str],
window_ranges: List[dict],
max_new_tokens: int,
day_num: int,
concurrency: int = 5,
) -> Optional[dict]:
"""逐帧独立 caption, 5 帧并发。返回 caption_dict 结构与 batch 模式兼容。"""
if not window_b64:
return None
dense_caption: Dict[str, str] = {}
# 线程池并发 (VLM 推理可重入; HF generate 的 GIL 释放允许并发)
with ThreadPoolExecutor(max_workers=min(concurrency, len(window_b64))) as pool:
futures = {
pool.submit(_caption_single_frame, hyper_vg, fb, max_new_tokens): i
for i, fb in enumerate(window_b64)
}
results: Dict[int, Optional[str]] = {}
for fut in futures:
i = futures[fut]
try:
results[i] = fut.result()
except Exception as e:
logger.error(f" [Caption-per_frame] frame {i} 失败: {e}")
results[i] = None
for i, caption_text in results.items():
if not caption_text:
continue
f_start = _seconds_to_hms(_timestamp_num_to_seconds(window_ranges[i]['start']))
f_end = _seconds_to_hms(_timestamp_num_to_seconds(window_ranges[i]['end']))
dense_caption[f"DAY{day_num}-{f_start}-{f_end}"] = caption_text
if not dense_caption:
return None
return {
"dense_caption": dense_caption,
"description": "", # per-frame 模式不合成 window 级 description
}
def _caption_window_v2(
hyper_vg: HyperVideoGraph,
window_b64: List[str],
window_ranges: List[dict],
max_new_tokens: int,
day_num: int,
desc_min_sentences: int = 2,
desc_max_sentences: int = 3,
prior_frame_captions: Optional[List[str]] = None,
qaware_prefix: Optional[str] = None,
) -> Optional[dict]:
"""V2 caption: 一次 batch VLM 调用,同时输出 5 个 2s frame captions + 整体 description。
Args:
prior_frame_captions: 上一个 window 的末尾 1-2 个 frame caption 文本
(形如 `"[DAY1-00:00:08-00:00:10] <caption>"`)。
非空时使用 ``FRAME_CAPTION_WINDOW_V2_WITH_PRIOR_PROMPT``,
供 VLM 跨 window 保持命名一致性。
为 None 或空列表时退回到不带 prior 的 V2 默认 prompt。
返回:
caption_dict = {
"dense_caption": {"DAY1-...-...": "<2s caption>", ...},
"description": "<10s 整体描述>",
}
"""
if not window_b64:
return None
num_frames = len(window_b64)
last_frame_index = num_frames - 1
window_seconds = max(2 * num_frames, 2) # 兜底(实际是 _timestamp 计算的 ~10s)
if prior_frame_captions:
prior_text = "\n".join(prior_frame_captions)
prompt = FRAME_CAPTION_WINDOW_V2_WITH_PRIOR_PROMPT.format(
num_frames=num_frames,
last_frame_index=last_frame_index,
window_seconds=window_seconds,
desc_min=desc_min_sentences,
desc_max=desc_max_sentences,
prior_frame_context=prior_text,
)
else:
prompt = FRAME_CAPTION_WINDOW_V2_PROMPT.format(
num_frames=num_frames,
last_frame_index=last_frame_index,
window_seconds=window_seconds,
desc_min=desc_min_sentences,
desc_max=desc_max_sentences,
)
# 问题感知模式: 把用户问题 + 选项预先拼到 prompt 前, 引导 VLM 关注相关视觉证据
# qaware_prefix 由 run_streaming_hyper_construction 传入, 仅 --question_aware 时生效
# 不修改 V2 prompt 常量本身, 只在外层做字符串拼接
if qaware_prefix:
prompt = qaware_prefix + "\n\n" + prompt
for retry in range(3):
try:
with timed("vlm_caption"):
caption_text = hyper_vg.mllm_response(
hyper_vg.video_llm, hyper_vg.processor,
prompt, None,
base64_frames=window_b64,
max_new_tokens=max_new_tokens, has_image=True,
)
parsed = _safe_json_parse(caption_text)
if not (isinstance(parsed, dict) and "frames" in parsed
and isinstance(parsed["frames"], dict)):
logger.warning(f" [Caption-v2] 格式不合法 retry {retry+1}/3")
continue
description = str(parsed.get("description") or "").strip()
if not description:
logger.warning(f" [Caption-v2] 缺 description retry {retry+1}/3")
# 不直接 retry,下面用 frame caption 拼接兜底
# 但若 description 缺失,downstream 边界判断质量会下降,所以先重试
continue
dense_caption: Dict[str, str] = {}
for frame_idx_str, frame_caption in parsed["frames"].items():
try:
fi = int(frame_idx_str)
except ValueError:
continue
if fi < 0 or fi >= len(window_ranges):
continue
f_start = _seconds_to_hms(_timestamp_num_to_seconds(window_ranges[fi]['start']))
f_end = _seconds_to_hms(_timestamp_num_to_seconds(window_ranges[fi]['end']))
dense_caption[f"DAY{day_num}-{f_start}-{f_end}"] = str(frame_caption)
if not dense_caption:
logger.warning(f" [Caption-v2] frames 解析为空 retry {retry+1}/3")
continue
return {
"dense_caption": dense_caption,
"description": description,
}
except Exception as e:
logger.error(f" [Caption-v2] 失败 retry {retry+1}/3: {e}")
continue
return None
# ==================== 流式建图核心 ====================
def run_streaming_hyper_construction(
hyper_vg: HyperVideoGraph,
video_path: str,
sample_dir: str,
interval_seconds: float = 2.0,
window_frames: int = 5,
max_new_tokens: int = 1024,
day_num: int = 1,
frame_captions_out_path: Optional[str] = None,
caption_mode: str = "batch",
per_frame_concurrency: int = 5,
max_seconds: Optional[float] = None,
qaware_prefix: Optional[str] = None,
):
"""流式建图: 按 interval 采样 → window_frames 一批送 VLM → 打包送超图。
caption_mode:
* "batch" — 一次 VLM 调用产出 window 的多帧 dense caption (默认)
* "per_frame" — 每帧一次独立调用, 帧间并发 (per_frame_concurrency 控制)
max_seconds:
* None — 处理整段视频
* float — 只处理 [0, max_seconds] 这段(按 frame.start 秒数过滤)。
用于 SB Real-Time: 没必要建到视频结尾, max(question.time_stamp) 之后
的内容不会被任何问题问到。
"""
if not os.path.exists(video_path):
raise FileNotFoundError(f"Video not found: {video_path}")
loop = always_get_an_event_loop()
logger.info(f" [Sample] video={os.path.basename(sample_dir)}, interval={interval_seconds}s")
base64_frames, _, time_ranges = sample_frames_by_interval(
video_path,
interval_seconds=interval_seconds,
output_format='base64',
time_span_info=None,
)
n_frames_raw = len(base64_frames)
# max_seconds 截断: 保留 frame.start <= max_seconds 的帧
if max_seconds is not None and max_seconds > 0:
keep = [
i for i, tr in enumerate(time_ranges)
if _timestamp_num_to_seconds(tr['start']) <= max_seconds
]
if keep:
cut = keep[-1] + 1
base64_frames = base64_frames[:cut]
time_ranges = time_ranges[:cut]
logger.info(
f" [Truncate] max_seconds={max_seconds:.1f}s → 保留 {cut}/{n_frames_raw} 帧 "
f"(节省 {(1 - cut / n_frames_raw) * 100:.1f}%)"
)
else:
logger.warning(f" [Truncate] max_seconds={max_seconds:.1f}s 太小, 无可用帧, 跳过")
return
n_frames = len(base64_frames)
logger.info(f" [Sample] 共采样 {n_frames} 帧")
if n_frames == 0:
logger.warning(" 没有可用帧")
return
# ---- V2 / V1 路径分派 ----
# V2: caption_granularity="per_window" + boundary_strategy="stream"
# 使用 FRAME_CAPTION_WINDOW_V2_PROMPT,一次 batch 同时输出 5 个 2s caption + 1 个 description
# 边界检测每 window 1 次(流式 binary)
# V1: 保留旧的 per_frame / batch 路径
cfg = hyper_vg.hyper_config
use_v2 = (cfg.caption_granularity == "per_window" and cfg.boundary_strategy == "stream")
if use_v2:
logger.info(" [Mode] V2: per_window caption + stream boundary")
caption_prompt = None # V2 模式 prompt 在 _caption_window_v2 内部 format
elif caption_mode == "batch":
caption_prompt = FRAME_CAPTION_WINDOW_DETAILED_PROMPT.format(
output_format=OUTPUT_FORMAT,
)
else:
caption_prompt = None # per_frame 模式由 _caption_single_frame 内部使用简化 prompt
all_frame_captions: List[dict] = []
window_descriptions: List[dict] = []
# 重置计时统计(每个 video 独立)
try:
hyper_vg.hypergraph_builder.reset_timing()
except Exception:
pass
# V2 模式专用:跨 window 的 naming-context。保存上一个 window 的末尾 N_PRIOR 条
# frame caption 文本(含时间戳前缀),作为当前 window VLM 的 prior context。
# 目的:让 VLM 在命名(球员名字、物体标签)上跨 window 保持一致。
prev_tail_captions: List[str] = []
N_PRIOR = 2 # 只传最后 2 条
n_windows = (n_frames + window_frames - 1) // window_frames
for w_idx in range(n_windows):
s = w_idx * window_frames
e = min(s + window_frames, n_frames)
window_b64 = base64_frames[s:e]
window_ranges = time_ranges[s:e]
win_start_sec = _timestamp_num_to_seconds(window_ranges[0]['start'])
win_end_sec = _timestamp_num_to_seconds(window_ranges[-1]['end'])
win_time_span = (
f"{day_num}-{_seconds_to_hms(win_start_sec)}-{_seconds_to_hms(win_end_sec)}"
)
# 根据模式分派 caption 调用
t0 = time.time()
if use_v2:
caption_dict = _caption_window_v2(
hyper_vg=hyper_vg,
window_b64=window_b64,
window_ranges=window_ranges,
max_new_tokens=max_new_tokens,
day_num=day_num,
prior_frame_captions=prev_tail_captions or None,
qaware_prefix=qaware_prefix,
)
mode_tag = "v2_qaware" if qaware_prefix else "v2"
elif caption_mode == "per_frame":
caption_dict = _caption_window_per_frame(
hyper_vg=hyper_vg,
window_b64=window_b64,
window_ranges=window_ranges,
max_new_tokens=max_new_tokens,
day_num=day_num,
concurrency=per_frame_concurrency,
)
mode_tag = "per_frame"
else:
caption_dict = _caption_window_batch(
hyper_vg=hyper_vg,
window_b64=window_b64,
window_ranges=window_ranges,
caption_prompt=caption_prompt,
max_new_tokens=max_new_tokens,
day_num=day_num,
)
mode_tag = "batch"
dt = time.time() - t0
if not caption_dict:
logger.error(f" [Caption] win {w_idx+1}/{n_windows} 最终失败, 跳过")
continue
logger.info(
f" [Caption-{mode_tag}] win {w_idx+1}/{n_windows} ({win_time_span}) "
f"{len(window_b64)} 帧, {len(caption_dict['dense_caption'])} captions, {dt:.1f}s"
)
# 记录 frame captions + window descriptions
for fk, fv in caption_dict["dense_caption"].items():
all_frame_captions.append({"time_span": fk, "caption": fv})
window_descriptions.append({
"time_span": win_time_span,
"description": caption_dict["description"],
})
# V2: 更新 prior context —— 取当前 window 末尾 N_PRIOR 条 frame captions,
# 带时间戳前缀,供下一个 window 的 VLM 作命名一致性参考
if use_v2:
dc = caption_dict["dense_caption"]
# dense_caption dict 的 key 是 frame time_span 字符串,按字典序排序
# ≈ 时间顺序(同一 window 内时间戳字符串前缀相同)
sorted_keys = sorted(dc.keys())
tail_keys = sorted_keys[-N_PRIOR:]
prev_tail_captions = [f"[{k}] {dc[k]}" for k in tail_keys]
# 送 hypergraph builder:V2 走新入口
try:
if use_v2:
loop.run_until_complete(
hyper_vg.hypergraph_builder.process_window_caption_v2(
caption_dict=caption_dict,
time_span=win_time_span,
day_num=day_num,
)
)
else:
loop.run_until_complete(
hyper_vg.hypergraph_builder.process_window_caption(
caption_dict=caption_dict,
time_span=win_time_span,
day_num=day_num,
)
)
except Exception as e:
logger.error(f" [Hyper] win {w_idx+1}/{n_windows} process 失败: {e}")
continue
# Visual embedding: 在 caption 完成后顺手过 ImageBind 一次,复用 window_b64
# 未启用时是 no-op,不影响流程
try:
hyper_vg.hypergraph_builder.index_window_visual(win_time_span, window_b64)
except Exception as e:
logger.warning(f" [Visual] win {w_idx+1}/{n_windows} 索引失败: {e}")
if (w_idx + 1) % 10 == 0:
stats = hyper_vg.hypergraph_builder.hypergraph.get_stats()
logger.info(f" [Progress] {w_idx+1}/{n_windows} windows, stats={stats}")
# 每 10 个 window 打一次 timing 中间汇总
try:
hyper_vg.hypergraph_builder.log_timing_summary(
prefix=f"(after {w_idx+1}/{n_windows} windows)"
)
except Exception:
pass
# flush
try:
loop.run_until_complete(hyper_vg.hypergraph_builder.flush())
except Exception as e:
logger.error(f" Flush 失败: {e}")
try:
hyper_vg.save_hypergraph()
except Exception as e:
logger.error(f" 保存超图失败: {e}")
# 最终 timing 汇总
try:
hyper_vg.hypergraph_builder.log_timing_summary(prefix="(final)")
except Exception:
pass
if frame_captions_out_path:
try:
with open(frame_captions_out_path, 'w', encoding='utf-8') as f:
json.dump({
"frame_captions": all_frame_captions,
"window_descriptions": window_descriptions,
"interval_seconds": interval_seconds,
"window_frames": window_frames,
"day_num": day_num,
}, f, ensure_ascii=False, indent=2)
logger.info(f" [Save] frame_captions.json -> {frame_captions_out_path}")
except Exception as e:
logger.error(f" 保存 frame_captions 失败: {e}")
stats = hyper_vg.hypergraph_builder.hypergraph.get_stats()
logger.info(f" [Done] 超图构建完成, stats={stats}")
# ==================== 单样本处理 ====================
def _reset_hyper_vg_working_dir(hyper_vg: HyperVideoGraph, new_working_dir: str):
"""重定向 working_dir + 重置所有存储/超图/索引, 保留 caption_model。"""
from dataclasses import asdict
from videorag._storage import JsonKVStorage, NanoVectorDBVideoSegmentStorage
from hypervideo.extractors.hypergraph_builder import VideoHypergraphBuilder
from hypervideo.index.dual_index import DualIndex
from hypervideo.retrieval.hierarchical_retriever import HierarchicalRetriever
hyper_vg.working_dir = new_working_dir
os.makedirs(new_working_dir, exist_ok=True)
hyper_vg.video_segments = hyper_vg.key_string_value_json_storage_cls(
namespace="video_segments", global_config=asdict(hyper_vg)
)
if hyper_vg.enable_llm_cache:
hyper_vg.llm_response_cache = hyper_vg.key_string_value_json_storage_cls(
namespace="llm_response_cache", global_config=asdict(hyper_vg)
)
hyper_vg.video_segment_feature_vdb = hyper_vg.vs_vector_db_storage_cls(
namespace="video_segment_feature",
global_config=asdict(hyper_vg),
embedding_func=None,
)
async def _llm_wrapper(prompt: str) -> str:
return await hyper_vg.llm.cheap_model_func(prompt)
hyper_vg.hypergraph_builder = VideoHypergraphBuilder(
llm_func=_llm_wrapper,
embedding_func=hyper_vg.embedding_func,
config=hyper_vg.hyper_config,
working_dir=hyper_vg.working_dir,
)
hyper_vg.dual_index = DualIndex(
embedding_func=hyper_vg.embedding_func,
config=hyper_vg.hyper_config,
)
hyper_vg.retriever = HierarchicalRetriever(
hypergraph=hyper_vg.hypergraph_builder.hypergraph,
dual_index=hyper_vg.dual_index,
config=hyper_vg.hyper_config,
)
hyper_vg.hypergraph_builder.load(new_working_dir)
def process_one_sample(
sample_dir: str,
hyper_vg: HyperVideoGraph,
args,
max_ts_lookup: Optional[Dict[str, int]] = None,
qaware_lookup: Optional[Dict[str, List[Dict[str, str]]]] = None,
):
sample_id = os.path.basename(sample_dir)
working_dir = (
f"{args.save_root}/{args.datasets_type}/{sample_id}_{args.caption_model_name}"
)
hypergraph_path = os.path.join(working_dir, "hypergraph.json")
if os.path.exists(hypergraph_path):
logger.info(f"跳过 {sample_id}(hypergraph.json 已存在)")
return False
video_path = os.path.join(sample_dir, "video.mp4")
if not os.path.exists(video_path):
logger.error(f" 视频不存在: {video_path}")
return False
# 计算 max_seconds: 从 sample_id 数字部分查 max_ts_lookup, 加 buffer
max_seconds: Optional[float] = None
sid_num = None
m = re.search(r'sample_(\d+)', sample_id)
if m:
sid_num = m.group(1)
if max_ts_lookup and sid_num is not None:
mts = max_ts_lookup.get(sid_num)
if mts is not None:
max_seconds = float(mts) + float(args.max_ts_buffer_seconds)
logger.info(
f" [max_ts] sample_{sid_num}: max(ts)={mts}s, "
f"截断至 {max_seconds:.1f}s (buffer={args.max_ts_buffer_seconds}s)"
)
else:
logger.warning(f" [max_ts] sample_{sid_num} 不在 questions_real.json 中, 不截断")
# ---- 问题感知模式: 把 sample 的所有问题 (含 options) 注入到 hyper_cfg.forward_questions
# episode_segmenter / fact_extractor 会自动切到 _FORWARD prompt (走现有框架机制),
# 同时构造 qaware_prefix 拼到 caption prompt 头部 (不改原 V2 prompt)。
qaware_prefix: Optional[str] = None
if getattr(args, "question_aware", False) and qaware_lookup and sid_num is not None:
qs = qaware_lookup.get(sid_num) or []
if qs:
# 注入到 cfg, builder 在 _reset 时会把 forward_questions 传给 segmenter / fact_extractor
hyper_vg.hyper_config.forward_questions = qs
hyper_vg.hyper_config.forward_viewpoint = "third_person"
# caption 前置块
qb = _make_questions_block(qs)
qaware_prefix = (
"USER QUESTIONS (with options) the viewer will be asked about this video:\n"
f"{qb}\n\n"
"When describing frames below, prioritise visual evidence that helps "
"answer these questions. Do NOT mention the questions or options in "
"your captions; just describe what is visible, paying extra attention "
"to details (objects, counts, colors, on-screen text, actions, "
"spatial layout) relevant to the questions above."
)
logger.info(
f" [qaware] sample_{sid_num}: 注入 {len(qs)} 个 question 到 caption + fact prompt"
)
else:
logger.warning(f" [qaware] sample_{sid_num} 在 questions_real.json 中无问题, 退化到普通模式")
hyper_vg.hyper_config.forward_questions = None
else:
# 通用模式: 清空 forward_questions, 走原 V2 prompt (与 question_aware=False 时行为一致)
hyper_vg.hyper_config.forward_questions = None
os.makedirs(working_dir, exist_ok=True)
_reset_hyper_vg_working_dir(hyper_vg, working_dir)
try:
run_streaming_hyper_construction(
hyper_vg=hyper_vg,
video_path=video_path,
sample_dir=sample_dir,
interval_seconds=args.interval_seconds,
window_frames=args.window_frames,
max_new_tokens=args.max_new_tokens,
day_num=1,
frame_captions_out_path=os.path.join(working_dir, "frame_captions.json"),
caption_mode=args.caption_mode,
per_frame_concurrency=args.per_frame_concurrency,
max_seconds=max_seconds,
qaware_prefix=qaware_prefix,
)
return True
except Exception as e:
logger.error(f" 建图失败 {sample_id}: {e}")
return False
finally:
clear_gpu_memory()
# ==================== CLI ====================
def main():
parser = argparse.ArgumentParser(
description="StreamingBench RTVU HyperVideo 建图(HyperMem 风格)"
)
parser.add_argument(
"--rtvu_root", type=str,
default="/mnt/tokyo-ai/gst/StreamingBench-/Real-Time_Visual_Understanding_videos",
help="包含 sample_*/video.mp4 的目录。默认是完整版 (Real-Time_Visual_Understanding_videos), "
"另一个目录 Real_Time_Visual_Understanding/ 里的视频是被截短的版本, "
"对长视频 max(ts) 会越界, 不要用。",
)
parser.add_argument(
"--save_root", type=str,
default="/mnt/tokyo-ai/gst/GST_EGOSERVE/all_results",
)
parser.add_argument(
"--caption_model_name", type=str, default="qwenvl_3_8b_instruct",
)
parser.add_argument("--datasets_type", type=str, default="streamingbench_rtvu_hyper_v2")
parser.add_argument("--cuda", type=str, default="0")
parser.add_argument("--interval_seconds", type=float, default=2.0)
parser.add_argument("--window_frames", type=int, default=5)
parser.add_argument("--max_new_tokens", type=int, default=2048)
parser.add_argument(
"--caption_mode", type=str, default="batch",
choices=["batch", "per_frame"],
help="caption 生成模式: batch=每 window 一次 VLM 调用 (默认); "
"per_frame=每帧独立调用 (简化 prompt + 并发)",
)
parser.add_argument(
"--per_frame_concurrency", type=int, default=5,
help="caption_mode=per_frame 时的线程池并发度",
)
parser.add_argument(
"--boundary_detection_mode", type=str, default="window",
choices=["frame", "window", "window_binary"],
help="V1 边界检测模式: frame=逐帧 probe; window=window + split_at_frame; "
"window_binary=window 2 选 1。仅在 --pipeline_version=v1 时生效",
)
parser.add_argument(
"--pipeline_version", type=str, default="v2",
choices=["v1", "v2"],
help=(
"处理管线版本。"
"v2 (默认): per_window 一次 VLM batch 出 5 个 2s caption + 整体 10s "
"description, stream 边界检测每 10s 判一次, fact 按 episode 紧圈。"
"v1: 使用 --caption_mode / --boundary_detection_mode 的旧路径。"
),
)
parser.add_argument(
"--caption_granularity", type=str, default=None,
choices=[None, "per_window", "per_frame", "window_batch"],
help="显式指定 caption 粒度,会覆盖 --pipeline_version 推导出的默认值。"
"一般用于做对照实验。",
)
parser.add_argument(
"--boundary_strategy", type=str, default=None,
choices=[None, "stream", "frame", "window", "window_binary"],
help="显式指定边界策略,会覆盖 --pipeline_version 推导出的默认值。",
)
parser.add_argument(
"--sample_ids", type=str, default=None,
help="仅处理指定 sample id, 逗号分隔, 如 '1,2,3'",
)
parser.add_argument(
"--questions_real_json", type=str,
default="/mnt/tokyo-ai/gst/StreamingBench/src/data/questions_real.json",
help="官方 questions_real.json 路径, 用于查每个 sample 的 max(time_stamp). "
"传空字符串 '' 关闭 max_ts 截断, 走整段视频建图。",
)
parser.add_argument(
"--max_ts_buffer_seconds", type=float, default=10.0,
help="max(time_stamp) 之外再多建几秒 buffer (防止边界检测/检索吃边)",
)
parser.add_argument(
"--question_aware", action="store_true",
help="问题感知建图: 从 questions_real.json 取出 sample 的全部 5 个 question + options, "
"注入到 caption prompt (前置 USER QUESTIONS 块) 和 fact_extractor / episode_summary "
"(走现有 _FORWARD prompt + forward_questions cfg). 默认 False, 走通用建图。",
)
parser.add_argument(
"--openai_api_key", type=str,
default="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",
)
args = parser.parse_args()
os.environ["CUDA_VISIBLE_DEVICES"] = args.cuda
os.environ["OPENAI_API_KEY"] = args.openai_api_key
# 枚举 sample_* 目录
all_samples = sorted(
[
os.path.join(args.rtvu_root, d)
for d in os.listdir(args.rtvu_root)
if d.startswith("sample_") and os.path.isdir(os.path.join(args.rtvu_root, d))
],
key=lambda p: int(os.path.basename(p).split("_")[-1]),
)
logger.info(f"发现 {len(all_samples)} 个 sample")
if args.sample_ids:
wanted = {s.strip() for s in args.sample_ids.split(',') if s.strip()}
all_samples = [
p for p in all_samples
if os.path.basename(p).split("_")[-1] in wanted
]
logger.info(f"按 --sample_ids 过滤后剩余 {len(all_samples)} 条")
unprocessed = []
for p in all_samples:
sid = os.path.basename(p)
hg_path = os.path.join(
args.save_root, args.datasets_type,
f"{sid}_{args.caption_model_name}", "hypergraph.json",
)
if not os.path.exists(hg_path):
unprocessed.append(p)
skipped = len(all_samples) - len(unprocessed)
if skipped > 0:
logger.info(f"跳过已完成 {skipped} 条, 剩余 {len(unprocessed)} 条")
if not unprocessed:
logger.info("全部已处理完毕")
return
# 初始化 hyper_vg + caption model(仅一次)
init_working_dir = os.path.join(
args.save_root, args.datasets_type, f"_init_{args.caption_model_name}"
)
os.makedirs(init_working_dir, exist_ok=True)
# ---- pipeline 版本推导 ----
# v2 默认:per_window caption + stream boundary
# v1 默认:走用户传入的 --caption_mode / --boundary_detection_mode
if args.pipeline_version == "v2":
cg_default = "per_window"
bs_default = "stream"
else:
# v1:caption_granularity 映射自 caption_mode;boundary_strategy 映射自 boundary_detection_mode
cg_default = "per_frame" if args.caption_mode == "per_frame" else "window_batch"
bs_default = args.boundary_detection_mode
# 用户可显式覆盖
caption_granularity = args.caption_granularity or cg_default
boundary_strategy = args.boundary_strategy or bs_default
hyper_cfg = HyperVideoConfig(
boundary_detection_mode=args.boundary_detection_mode,
caption_mode=args.caption_mode,
caption_granularity=caption_granularity,
boundary_strategy=boundary_strategy,
)
logger.info(
f"模式: pipeline_version={args.pipeline_version}, "
f"caption_granularity={caption_granularity}, "
f"boundary_strategy={boundary_strategy}, "
f"(legacy: caption_mode={args.caption_mode}, "
f"boundary_detection_mode={args.boundary_detection_mode})"
)
hyper_vg = HyperVideoGraph(
working_dir=init_working_dir,
llm=default_llm_config,
hyper_config=hyper_cfg,
)
logger.info(f"加载 caption model: {args.caption_model_name}")
hyper_vg.load_caption_model(args.caption_model_name)
logger.info(f"HyperVideoGraph 初始化完成, stats={hyper_vg.hypergraph_builder.hypergraph.get_stats()}")
# 加载 questions_real.json → max(time_stamp) per sample
max_ts_lookup: Dict[str, int] = {}
qaware_lookup: Dict[str, List[Dict[str, str]]] = {}
if args.questions_real_json:
max_ts_lookup = _load_max_ts_per_sample(args.questions_real_json)
if args.question_aware:
qaware_lookup = _load_questions_with_options_per_sample(args.questions_real_json)
t_all = time.time()
ok = 0
for idx, sample_dir in enumerate(unprocessed):
t0 = time.time()
sid = os.path.basename(sample_dir)
logger.info(f"[{idx+1}/{len(unprocessed)}] {sid}")
try:
did = process_one_sample(
sample_dir, hyper_vg, args,
max_ts_lookup=max_ts_lookup,
qaware_lookup=qaware_lookup,
)
if did:
ok += 1
except Exception as e:
logger.error(f" 样本 {sid} 处理异常: {e}")
continue
dt = time.time() - t0
logger.info(f"[{idx+1}/{len(unprocessed)}] {sid} 耗时 {dt:.1f}s")
total = time.time() - t_all
logger.info(
f"\n{'='*80}\n建图完成: 成功 {ok}/{len(unprocessed)}, "
f"总耗时 {total:.1f}s ({total/60:.1f}min)\n{'='*80}"
)
if __name__ == "__main__":
main()
|