"""单点调试入口：每个阶段都能单独跑一次，方便下断点。 配合 .vscode/launch.json 里的 evo2-2 / evo2-3 / evo2-4 / evo2-5 使用。 python evo2/debug_one.py qgen --dataset egoschema --space ACTION --time GLOBAL python evo2/debug_one.py retrieve --question "Where did I put the screwdriver?" python evo2/debug_one.py episode --pool data/pool_egoschema.json --index 0 python evo2/debug_one.py curate --kind fail """ import argparse import json import logging import os import sys sys.path.insert(0, "/mnt/tokyo-ai/gst/EvoStream") sys.path.insert(0, "/mnt/tokyo-ai/gst/EvoStream/cons_memory") logging.basicConfig(level=logging.INFO, format="[%(levelname)s] %(message)s") # ---------------------------------------------------------------- 出题 def cmd_qgen(a): from infer.llm_clients import make_client from memory_store import VideoMemory from evo2.qgen import QuestionGenerator, phrasing_rule from evo2.style_cards import build_style_card from evo2.validate_qgen import list_banks, sample_frames banks = list_banks(a.dataset, 3) vid, bank = banks[a.video_idx] mem = VideoMemory(bank) mode = "global" if a.time == "GLOBAL" else "window" frames, ts = sample_frames(mem, mode) print(f"视频 {vid} | 采样 {mode} {len(frames)} 帧 @ " f"{[f'{t:.0f}s' for t in ts]}") print(f"\n--- 该格子的措辞约束 ---\n{phrasing_rule(a.time)}\n") qg = QuestionGenerator(make_client(a.gen_ep, temperature=0.7), make_client(a.verify_ep, temperature=0.0), style_card=build_style_card(a.dataset)) rec = qg.make_one(frames, ts, a.space, a.time) # ← 断点打这里 print(json.dumps(rec, ensure_ascii=False, indent=1)[:2500]) print(f"\n结果: stage={rec['stage']} passed={rec['passed']} " f"reject={rec.get('reject')}") # ---------------------------------------------------------------- 检索 def cmd_retrieve(a): from infer.llm_clients import embed_texts from evo2.classifier import classify_lexical from evo2.seed_lib import build_seed from evo2.skill import SkillLibrary, keyword_score lib = SkillLibrary(build_seed(a.seed_mode), embed_fn=embed_texts) lib.index_library() cell = classify_lexical(a.question, None) # ← 断点：题型判定 print(f"题目: {a.question}") print(f"题型判定: {cell}") cards, scores = lib.retrieve(a.question, cell, topk=len(lib.cards), return_scores=True) # ← 断点：打分 print(f"\n全部卡片打分（前8）:") for s in scores[:8]: print(f" {s['total']:.4f} = kw {s['kw']:.3f}*0.5 + sem {s['sem']:.3f}*0.5 " f"(+先验) [{s['card_id']}]") print(f"\n实际注入(top2): {[c['card_id'] for c in cards[:2]]}") print("\n渲染给 executor 的文本:") print(SkillLibrary.render(cards[:2])) # ---------------------------------------------------------------- 单题推理 def cmd_episode(a): from infer.config import RunConfig from infer.llm_clients import embed_texts, make_client from infer.tools import ToolRegistry from evo2.pipeline import run_one_episode from evo2.seed_lib import build_seed from evo2.skill import SkillLibrary from evo2.tracelog import TraceLogger pool = json.load(open(a.pool)) qs = pool["train"] + pool["frozen_val"] q = qs[a.index] print(f"题目[{a.index}]: {q['question'][:140]}") print(f" 格子={q['cell']} gt={q.get('gt')} video={q['video_id']}") cfg = RunConfig() cfg.executor = a.executor_ep lib = SkillLibrary(build_seed(a.seed_mode), embed_fn=embed_texts) lib.index_library() vlm = make_client(cfg.vlm_tool, temperature=0.0) tl = TraceLogger("/mnt/tokyo-ai/gst/EvoStream/runs/DEBUG_one") log = run_one_episode(q, lib, make_client(a.executor_ep, temperature=0.0), lambda m: ToolRegistry(m, cfg, vlm_client=vlm), cfg, tl=tl, meta={"phase": "debug"}) # ← 断点 print("\n--- 工具轨迹 ---") for s in (log.get("trace") or []): print(f" {s['step']}. {s['tool']}({json.dumps(s['args'], ensure_ascii=False)[:90]})") print(f" → {(s.get('result_digest') or '')[:180]}") print(f"\n思考: {(log.get('reasoning') or '')[:400]}") print(f"答案: {log.get('choice')} (gt={q.get('gt')}) 正确={log.get('correct')}") print(f"注入的卡: {log.get('skills_injected')} adherence={log.get('adherence')}") tl.close() # ---------------------------------------------------------------- 归纳改卡 def cmd_curate(a): from infer.llm_clients import make_client from infer.tools import TOOL_SPECS from evo2.curator import build_prompt, catalog_of from evo2.gates import gate_all from evo2.seed_lib import build_seed from evo2.skill import new_card # 构造轨迹：库里已有一张"只用文本检索"的缺陷卡，照做仍答错 → 应 modify existing = [new_card("count-caption-only", ["how many", "count", "number of"], "询问数量时用文本检索caption统计", [{"tool": "get_topk_caption", "params": {"k": 5}}], ["COUNT"], ["NOW", "PAST"], "caption中提到了数量")] \ if a.with_existing else build_seed("empty") fails = [ {"question": "How many times did the person rinse the pan?", "cell": "COUNT-PAST", "trace": [{"tool": "get_topk_caption", "args": {"query": "rinse pan", "k": 5}}], "skills_injected": [c["card_id"] for c in existing], "adherence": 1.0, "failure_mode": "照策略执行了但caption未给出精确数量"}, {"question": "How many plates were on the rack?", "cell": "COUNT-NOW", "trace": [{"tool": "get_topk_caption", "args": {"query": "plates", "k": 5}}], "skills_injected": [c["card_id"] for c in existing], "adherence": 1.0, "failure_mode": "照策略执行了但caption未给出精确数量"}, ] contrast = [ {"question": "Where did I put the blue screwdriver?", "cell": "SPATIAL-PAST", "trace": [{"tool": "get_topk_caption", "args": {"query": "screwdriver"}}], "outcome": "答错", "skills_injected": []}, {"question": "Where did I put the blue screwdriver?", "cell": "SPATIAL-PAST", "trace": [{"tool": "get_topk_frame", "args": {"query": "screwdriver on bench"}}, {"tool": "rewatch_frames", "args": {"t_start": 88, "t_end": 108}}], "outcome": "答对", "skills_injected": []}, ] eps = fails if a.kind == "fail" else contrast specs = TOOL_SPECS if isinstance(TOOL_SPECS, str) else \ json.dumps(TOOL_SPECS, ensure_ascii=False)[:2500] prompt = build_prompt(a.kind, eps, catalog_of(existing), specs, n_cards=len(existing)) # ← 断点：看完整 prompt print(f"--- prompt 长度 {len(prompt)} 字符（前1200）---\n{prompt[:1200]}\n...") r = make_client(a.curator_ep, temperature=0.5).chat_json(prompt, max_tokens=4000) or {} print(f"\n--- curator reasoning ---\n{str(r.get('reasoning'))[:600]}") for op in (r.get("ops") or []): card = op.get("card") or {} if op.get("op") == "modify": base = next((c for c in existing if c["card_id"] == op.get("card_id")), {}) card = {**base, **(op.get("changes") or {})} ok, why = gate_all(card, existing, op.get("op", "add")) # ← 断点：过闸 print(f"\n[{op.get('op')}] {card.get('card_id')} 过闸={ok} {why}") print(json.dumps(card, ensure_ascii=False, indent=1)[:900]) def main(): ap = argparse.ArgumentParser() sub = ap.add_subparsers(dest="cmd", required=True) p = sub.add_parser("qgen") p.add_argument("--dataset", default="egoschema") p.add_argument("--space", default="ACTION") p.add_argument("--time", default="GLOBAL") p.add_argument("--video-idx", type=int, default=0) p.add_argument("--gen-ep", default="qwen38_pool") p.add_argument("--verify-ep", default="qwen_local") p.set_defaults(fn=cmd_qgen) p = sub.add_parser("retrieve") p.add_argument("--question", required=True) p.add_argument("--seed-mode", default="minimal") p.set_defaults(fn=cmd_retrieve) p = sub.add_parser("episode") p.add_argument("--pool", default="/mnt/tokyo-ai/gst/EvoStream/data/pool_egoschema.json") p.add_argument("--index", type=int, default=0) p.add_argument("--seed-mode", default="minimal") p.add_argument("--executor-ep", default="qwen38_pool") p.set_defaults(fn=cmd_episode) p = sub.add_parser("curate") p.add_argument("--kind", default="fail", choices=["fail", "success"]) p.add_argument("--with-existing", action="store_true", default=True) p.add_argument("--curator-ep", default="qwen38_pool") p.set_defaults(fn=cmd_curate) a = ap.parse_args() a.fn(a) if __name__ == "__main__": main()