""" main.py — 多因子训练入口（分组训练） 使用方法： python main.py # 按 SYMBOL_GROUPS 分组训练（默认推荐） python main.py --offline # 仅使用本地缓存，不连接 MT5 python main.py --single XAUUSD # 只训练单个品种 python main.py --cross-section # 所有品种一起训练（截面） python main.py --group risk # 只训练 risk 组 分组说明（Config.SYMBOL_GROUPS）： forex 组（EURUSD, USDJPY）：外汇，美元方向因子，2品种×11000=22000样本 risk 组（XAUUSD, US100.cash, US500.cash）：风险资产，3品种×11000=33000样本 """ import sys, pathlib, json # 无控制台环境（如 Start-Process -WindowStyle Hidden）下，sys.stdout 可能为 None， # 导致 tqdm.write 报错。此时重定向到日志文件。 if sys.stdout is None or sys.stderr is None: _log_path = r"D:\素材\自动挖因子\training_stdout.log" _log_fp = open(_log_path, "a", encoding="utf-8") if sys.stdout is None: sys.stdout = _log_fp if sys.stderr is None: sys.stderr = _log_fp from config import Config from data_pipeline.data_manager import MT5DataManager from data_pipeline.fetcher import MT5DataFetcher from data_pipeline.single_symbol_manager import SingleSymbolDataManager from model_core.engine import AlphaEngine from model_core.config import ModelConfig from model_core.vocab import VOCAB_VERSION class GroupDataManager: """品种分组数据视图，兼容 AlphaEngine 接口（N = 组内品种数）。""" def __init__(self, multi_manager, symbols: list[str]): self._multi = multi_manager self._symbols = [s for s in symbols if s in multi_manager.symbols] self._idxs = [multi_manager.symbols.index(s) for s in self._symbols] @property def symbols(self): return list(self._symbols) @property def raw_dict(self): full = self._multi.raw_dict return {k: v[self._idxs] for k, v in full.items()} @property def feat_tensor(self): from model_core.features import MT5FeatureEngineer return MT5FeatureEngineer.compute_features(self.raw_dict) @property def target_ret(self): return self._multi.target_ret[self._idxs] @property def bar_time(self): return self._multi.bar_time[self._idxs] def save_group_strategy(engine: AlphaEngine, group_name: str, symbols: list[str]): """保存分组策略：group 总文件 + 各品种 best_*.json。""" pathlib.Path("strategies").mkdir(exist_ok=True) # 分组总文件 gp = pathlib.Path("strategies") / f"best_group_{group_name}.json" gp.write_text(json.dumps({ "vocab_version": VOCAB_VERSION, "group": group_name, "symbols": symbols, "formula": engine.best_formula, "best_score": engine.best_score, }, indent=2)) # 各品种文件（runner 按品种名加载） for sym in symbols: sp = pathlib.Path("strategies") / f"best_{sym}.json" sp.write_text(json.dumps({ "vocab_version": VOCAB_VERSION, "symbol": sym, "group": group_name, "formula": engine.best_formula, "best_score": engine.best_score, "source": f"group_{group_name}", }, indent=2)) print(f" 已保存: best_group_{group_name}.json + {len(symbols)} 个品种文件") def train_group(fetcher, group_name: str, symbols: list[str], offline: bool): """训练一个品种组，使用组内独立的 DataManager（不与其他组取时间交集）。 自动检测是否有可续训的 checkpoint：若存在则从断点继续，否则从 step 0 开始。 """ print(f"\n{'─'*60}") print(f" [{group_name}] 组: {symbols}") print(f"{'─'*60}") # 临时覆盖 SYMBOLS 只加载本组品种 original_symbols = Config.SYMBOLS[:] Config.SYMBOLS = [s for s in symbols if s in original_symbols or True] try: group_mgr = MT5DataManager(fetcher) group_mgr.load() actual_symbols = group_mgr.symbols T = group_mgr.raw_dict["open"].shape[1] print(f" 独立加载: {actual_symbols} T={T} bars") except Exception as e: print(f" [错误] 数据加载失败: {e}") Config.SYMBOLS = original_symbols return None finally: Config.SYMBOLS = original_symbols # 恢复原始配置 if not actual_symbols: print(f" [跳过] 无有效品种") return None engine = AlphaEngine(data_manager=group_mgr, target_symbol=group_name) # ── 自动续训：检测最新 checkpoint ──────────────────────────────── import glob as _glob ckpt_pattern = str(pathlib.Path("checkpoints") / f"ckpt_{group_name}_step_*.pt") ckpt_files = sorted(_glob.glob(ckpt_pattern)) start_step = 0 if ckpt_files: latest_ckpt = ckpt_files[-1] try: start_step = engine.load_checkpoint(latest_ckpt) print(f" [续训] 从 {latest_ckpt} 恢复，start_step={start_step}") except Exception as e: print(f" [警告] checkpoint 加载失败（{e}），从头开始训练") start_step = 0 else: print(f" [新训] 未找到 checkpoint，从 step 0 开始") if start_step >= ModelConfig.TRAIN_STEPS: print(f" [完成] {group_name} 已完成全部 {ModelConfig.TRAIN_STEPS} 步，跳过训练") save_group_strategy(engine, group_name, actual_symbols) return engine engine.train(start_step=start_step) save_group_strategy(engine, group_name, actual_symbols) return engine def main(): cross = "--cross-section" in sys.argv single = "--single" in sys.argv offline = "--offline" in sys.argv grp_only = None if "--group" in sys.argv: idx = sys.argv.index("--group") grp_only = sys.argv[idx + 1] if idx + 1 < len(sys.argv) else None single_sym = [s for s in sys.argv[1:] if not s.startswith("--")] or None mode = "截面" if cross else ("单品种" if single else "分组") print(f"{'='*60}") print(f" AlphaGPT 训练 [{mode}模式]" + (" [离线缓存]" if offline else "")) print(f" TRAIN_STEPS={ModelConfig.TRAIN_STEPS} " f"MAX_FORMULA_LEN={ModelConfig.MAX_FORMULA_LEN} " f"BATCH_SIZE={ModelConfig.BATCH_SIZE}") print(f" SYMBOLS={Config.SYMBOLS}") print(f"{'='*60}") with MT5DataFetcher(offline=offline) as fetcher: if single and single_sym: # 单品种模式：用全局 multi_mgr multi_mgr = MT5DataManager(fetcher) multi_mgr.load() sym = single_sym[0] s_mgr = SingleSymbolDataManager(multi_mgr, sym) engine = AlphaEngine(data_manager=s_mgr, target_symbol=sym) engine.train() elif cross: multi_mgr = MT5DataManager(fetcher) multi_mgr.load() engine = AlphaEngine(data_manager=multi_mgr, target_symbol=None) engine.train() save_group_strategy(engine, "cross_section", multi_mgr.symbols) else: # 分组训练：每组独立加载，不共享 DataManager groups = getattr(Config, "SYMBOL_GROUPS", { "forex": ["EURUSD", "USDJPY"], "risk": ["XAUUSD", "US100.cash", "US500.cash"], }) if grp_only: groups = {grp_only: groups[grp_only]} if grp_only in groups else groups results = {} for gname, gsyms in groups.items(): eng = train_group(fetcher, gname, gsyms, offline) if eng: results[gname] = { "score": eng.best_score, "formula": eng._decode_formula(eng.best_formula), } print(f"\n{'='*60}") print(f" 分组训练完成") print(f"{'='*60}") for gname, r in results.items(): print(f" [{gname}]: score={r['score']:.4f}") print(f" {r['formula']}") print() if __name__ == "__main__": main()