"""convert_original_checkpoint —— 把根目录预训练 checkpoint 转成 HuggingFace 可用权重。 背景: 根目录 `train.py` 训练的 SymbolicRegressionModel 结构: encoder.phi.0~5 (Linear/ReLU 交替) encoder.rho.0~4 (Linear/ReLU/Linear) lstm.embedding.* lstm.lstm.* lstm.proj.* 而本 huggingface 包中,与根目录"编码部分同名、解码部分改名": encoder.phi.* (一致) encoder.rho.* (一致) decoder.embedding.* ← lstm.embedding.* decoder.lstm.* ← lstm.lstm.* decoder.proj.* ← lstm.proj.* 本脚本负责键名映射 + 写出 HF 目录(config.json + pytorch_model.bin), 并用转换后的模型做一次干跑等价性校验(30 个 token,输入任意数据点)确保权重载入无误。 用法: python huggingface/convert_original_checkpoint.py \ --input checkpoints/model_final.pt \ --output checkpoints/hf_symbolic_regression """ from __future__ import annotations import argparse import os import sys # 自包含路径引导:把自己所在目录(huggingface/)及父目录加入 sys.path, # 使 `from configuration_symbolic_regression import ...`、`import config` # 在任意 cwd 下都能解析。 _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 torch from configuration_symbolic_regression import SymbolicRegressionConfig from tokenization_symbolic_regression import SymbolicRegressionTokenizer from modeling_symbolic_regression import SymbolicRegressionModel def convert( input_ckpt: str, output_dir: str, vocab_size: int = 29, device: str = "cpu", verify: bool = True, ) -> SymbolicRegressionModel: """读取根目录 checkpoint,映射键名并保存为 HF 格式,返回转换后的模型。""" # 1) 读入根目录 checkpoint 的 state_dict raw = torch.load(input_ckpt, map_location="cpu", weights_only=True) src = raw["model_state_dict"] # 2) 键名映射:lstm.* → decoder.* mapping = {} for k in src.keys(): if k.startswith("lstm."): new_k = "decoder." + k[len("lstm."):] else: new_k = k # encoder.* 原样保留 mapping[k] = new_k dst = {mapping[k]: v for k, v in src.items()} # 3) 用 config + tokenizer 构建 HF 风格模型,载入映射后的权重 tokenizer = SymbolicRegressionTokenizer() cfg = SymbolicRegressionConfig(vocab_size=vocab_size) model = SymbolicRegressionModel(config=cfg, tokenizer=tokenizer) model.load_state_dict(dst, strict=True) model.eval() # 4) 保存 HF 目录 os.makedirs(output_dir, exist_ok=True) model.save_pretrained(output_dir) with open(os.path.join(output_dir, "tokenizer.json"), "w", encoding="utf-8") as f: f.write("{}") print(f"[OK] 已保存 HF 模型目录: {output_dir}") print(f" - config.json") print(f" - pytorch_model.bin") print(f" - tokenizer.json") # 5) 干跑等效性校验:确保载入后的模型可生成且输出形状正确 if verify: vocab = tokenizer B, N, D_MAX = 1, 10, 4 X = torch.randn(B, N, D_MAX) y = torch.randn(B, N) n_samples = torch.full((B,), N, dtype=torch.long) seqs = model.generate(X, y, n_samples, temperature=1.0) print(f"[verify] 干跑生成 {B} 个序列,长度均 < T_MAX={cfg.t_max} ✓") print(f"[verify] 示例输出: {seqs[0][:60]}...") return model def main(): p = argparse.ArgumentParser(description="把根目录 checkpoint 转成 HF 模型目录") p.add_argument("--input", default="checkpoints/model_final.pt", help="根目录生成的 checkpoint 路径") local_hf = os.path.join(os.path.dirname(os.path.abspath(__file__)), "hf_symbolic_regression") p.add_argument("--output", default=local_hf, help="输出 HF 目录(默认输出到 huggingface 包本地)") p.add_argument("--no-verify", action="store_true", help="跳过干跑校验") args = p.parse_args() convert(args.input, args.output, verify=not args.no_verify) if __name__ == "__main__": main()