"""run_hf_self_train —— 用 HuggingFace 风格模型做推理 + 主动自我训练示例。 演示三件事: 1. 把已有根目录 checkpoint 转成 HF 格式(convert_original_checkpoint)。 2. 用 HF 模型直接推理:给定 (x, y) 点集,generate() 出候选表达式。 3. 用 HF 模型做主动自我训练:输入用户 x/y 向量 → 拆 train/test → 自我评估 → 高 R² 提前停止。 用法: # 第 1 步:先把预训练 checkpoint 转成 HF 目录(只会做一次) python huggingface/convert_original_checkpoint.py # 推理 demo(自动 try 转换) python huggingface/run_hf_self_train.py demo # 自我训练:喂入用户(x, y)(这里放了一个 sin 示例) python huggingface/run_hf_self_train.py selftrain --x "0.1,0.6,1.1,1.6,2.1,2.6,3.0" \ --y "0.60,1.07,1.34,1.40,1.27,0.96,0.64" --rounds 15 --candidates 96 注意: 工程根目录 config.py / vocab.py / self_train.py 是共享依赖, 本脚本会自动把它们所在目录加入 sys.path,因此可从任意位置运行。 """ from __future__ import annotations import argparse import os import sys if hasattr(sys.stdout, "reconfigure"): sys.stdout.reconfigure(line_buffering=True) # ------------------------------------------------------------------ # # 方案 B 自包含路径引导: # 1) 本脚本所在目录(huggingface/)加入 sys.path —— 让平铺依赖 # config.py / vocab.py / model.py / self_train.py 可从任意 cwd 找到 # 2) 父目录也加入 —— 让 `huggingface` 这个包名可从任意 cwd 解析 # ------------------------------------------------------------------ # _HERE = os.path.dirname(os.path.abspath(__file__)) _PARENT = os.path.dirname(_HERE) for _p in (_HERE, _PARENT): if _p not in sys.path: sys.path.insert(0, _p) import numpy as np import torch import config as default_config from tokenization_symbolic_regression import SymbolicRegressionTokenizer from modeling_symbolic_regression import SymbolicRegressionModel HF_DIR = os.path.join(default_config.CHECKPOINT_DIR, "hf_symbolic_regression") ROOT_CKPT = os.path.join(default_config.CHECKPOINT_DIR, "model_final.pt") # ---------------------------------------------------------------------- # # 1. 模型加载:优先 from_pretrained 远程仓库,回退到本地 HF 权重目录 # ---------------------------------------------------------------------- # HF_REPO = "LIJINGHAI111/symbolic_regression_DeepSets_LSTM_Generator" def download_remote_model(): """本地无权重时,从 HuggingFace 远程仓库下载三件套到包本地目录。""" local_dir = os.path.join(_HERE, "hf_symbolic_regression") os.makedirs(local_dir, exist_ok=True) try: from huggingface_hub import hf_hub_download except ImportError: raise RuntimeError( "本地无 HF 权重且未安装 huggingface_hub。" "请先 `pip install huggingface_hub`,或补全本地权重目录。" ) print(f"[setup] 从远程仓库下载权重: {HF_REPO}") for f in ("config.json", "pytorch_model.bin", "tokenizer.json"): hf_hub_download(repo_id=HF_REPO, filename=f, local_dir=local_dir, force_download=True) return local_dir def ensure_hf_model(force_rebuild: bool = False): """返回可供 from_pretrained() 加载的目录或仓库 id。 策略(依次尝试,命中即返回): 1) huggingface/ 包本地 HF 权重目录 _HERE/hf_symbolic_regression 2) 工程根目录的 HF 权重目录 _PARENT/hf_symbolic_regression 3) 回退到 HuggingFace 远程仓库 id(需联网) """ candidates = [ os.path.join(_HERE, "hf_symbolic_regression"), os.path.join(_PARENT, "hf_symbolic_regression"), HF_DIR, ] for c in candidates: if c and os.path.exists(os.path.join(c, "pytorch_model.bin")): return c # 回退:远程仓库下载到包本地,之后按本地目录加载 return download_remote_model() # ---------------------------------------------------------------------- # # 2. 直接推理 demo:给出一堆点集,让模型生成候选表达式 # ---------------------------------------------------------------------- # @torch.no_grad() def demo_inference(): hf_dir = ensure_hf_model() tok = SymbolicRegressionTokenizer() model = SymbolicRegressionModel.from_pretrained(hf_dir, tokenizer=tok) dc = model.config print(f"[demo] 已从 {hf_dir} 加载 HF 模型 | 词表 {dc.vocab_size}," f"隐维 {dc.lstm_hidden_dim},D_MAX={dc.d_max}") # 用 y ≈ sin(x) 的采样点喂入,看模型能否"猜"出表达式族 x = np.linspace(0.1, 3.0, 30) y = np.sin(x) + 0.2 N = len(x) X = np.zeros((1, N, dc.d_max), dtype=np.float32) X[0, :, 0] = x yt = torch.from_numpy(y.astype(np.float32).reshape(1, N)) n_samples = torch.tensor([N], dtype=torch.long) Xt = torch.from_numpy(X) exprs = model.generate(Xt, yt, n_samples, temperature=1.5, num_return_sequences=5, do_sample=True) print("\n[demo] 模型生成的候选表达式:") for e in exprs: print(" •", e) # ---------------------------------------------------------------------- # # 3. 主动自我训练(核心) # ---------------------------------------------------------------------- # def run_selftrain(x_str: str, y_str: str, rounds: int, candidates: int, split: float = 0.7, seed: int = 456): hf_dir = ensure_hf_model() tok = SymbolicRegressionTokenizer() model = SymbolicRegressionModel.from_pretrained(hf_dir, tokenizer=tok) device = "cuda" if torch.cuda.is_available() else "cpu" model = model.to(device) def _parse(text: str) -> np.ndarray: parts = [p for p in text.replace(";", " ").replace(",", " ").split() if p] return np.array([float(p) for p in parts]) X = _parse(x_str).reshape(-1, 1) # 单变量示例 => [N, 1] y = _parse(y_str) if X.shape[0] != y.shape[0]: raise ValueError(f"x/y 长度不一致: x={X.shape[0]}, y={y.shape[0]}") from self_train import self_train_on_user_data print(f"[self-train] 数据 N={X.shape[0]}, D={X.shape[1]} → 拆 train/{split}·test/{1 - split}") result = self_train_on_user_data( model, X, y, device, split=split, max_rounds=rounds, candidates=candidates, seed=seed, verbose=True, ) print("\n" + "=" * 60) print(" 自我训练完成!") print(f" 最佳表达式 : {result['best_expr']}") print(f" 最佳测试 R² : {result['best_r2']:.6f}") print(f" 是否收敛 : {'是' if result['converged'] else '否(达到最大轮数)'}") print(f" 实际轮数 : {result['rounds']}") print("=" * 60) def main(): p = argparse.ArgumentParser(description="HF 符号回归模型:推理 demo + 自我训练") sub = p.add_subparsers(dest="cmd") pc = sub.add_parser("demo", help="显示一次推理候选") pc.set_defaults(func=lambda a: demo_inference()) ps = sub.add_parser("selftrain", help="对用户 (x,y) 做主动自我训练") ps.add_argument("--x", required=True, help="x 向量,逗号/空格分隔") ps.add_argument("--y", required=True, help="y 向量,逗号/空格分隔") ps.add_argument("--rounds", type=int, default=15) ps.add_argument("--candidates", type=int, default=96) ps.add_argument("--split", type=float, default=0.7) ps.add_argument("--seed", type=int, default=456) ps.set_defaults(func=lambda a: run_selftrain(a.x, a.y, a.rounds, a.candidates, a.split, a.seed)) args = p.parse_args() if not args.cmd: p.print_help() return args.func(args) if __name__ == "__main__": main()