hyperparameter/m4/arf/arf-m4-20260504_205355/_arf_generate.py

import pickle
import numpy as np
import pandas as pd

def _bootstrap_from_train(c_csv: str, n_target: int, seed: int = 42) -> pd.DataFrame:
    """当 arfpy.forge 完全不可用时，从训练 CSV 有放回抽样，保证行数与列对齐。"""
    src = pd.read_csv(c_csv, encoding="utf-8-sig", low_memory=False)
    src = src.replace([np.inf, -np.inf], np.nan).dropna(axis=1, how="all")
    src = src.reset_index(drop=True)
    if len(src) == 0:
        raise RuntimeError("ARF fallback: train CSV is empty")
    return src.sample(n=n_target, replace=True, random_state=seed).reset_index(drop=True)

def _safe_forge(model, n_target: int):
    # arfpy 在部分分布上会 ZeroDivisionError；n=1 在部分版本会触发
    # AttributeError（不要用 n=1）。失败返回 None，由外层走 bootstrap。
    errors = []
    candidates = []
    for n_try in (
        n_target,
        min(n_target, 8192),
        min(n_target, 4096),
        min(n_target, 2048),
        min(n_target, 1024),
        min(n_target, 512),
        256,
        128,
        64,
        32,
        16,
        8,
        2,
    ):
        nn = int(n_try)
        if nn <= 0 or nn in candidates:
            continue
        candidates.append(nn)
    for n_try in candidates:
        try:
            out = model.forge(n=n_try).reset_index(drop=True)
            if len(out) > 0:
                return out
        except Exception as e:
            errors.append(f"n={n_try}: {type(e).__name__}: {e}")
    print("[ARF] forge failed after retries; last errors:", " | ".join(errors[-4:]))
    return None

n_target = int(2217)
c_csv = "/work/output-Benchmark-trainonly-v1/m4/arf/arf-m4-20260504_205355/staged/public/train.csv"
with open("/work/output-Benchmark-trainonly-v1/m4/arf/arf-m4-20260504_205355/arf_model.pkl", "rb") as f:
    model = pickle.load(f)

syn = _safe_forge(model, n_target)
if syn is None or len(syn) == 0:
    if not c_csv:
        raise RuntimeError("ARF forge failed and no train csv path for bootstrap fallback")
    print(f"[ARF] Using train-bootstrap fallback (n={n_target})")
    syn = _bootstrap_from_train(c_csv, n_target)
else:
    if len(syn) > n_target:
        syn = syn.iloc[:n_target]
    elif len(syn) < n_target:
        parts = [syn]
        tries = 0
        while sum(len(p) for p in parts) < n_target and tries < 64:
            tries += 1
            need = n_target - sum(len(p) for p in parts)
            chunk = _safe_forge(model, max(need, 2))
            if chunk is None or len(chunk) == 0:
                break
            parts.append(chunk)
        syn = pd.concat(parts, ignore_index=True).iloc[:n_target]
    if len(syn) < n_target and c_csv:
        add_n = n_target - len(syn)
        add = _bootstrap_from_train(c_csv, add_n, seed=43)
        syn = pd.concat([syn, add], ignore_index=True).iloc[:n_target]

_ds_id = 'm4'
if _ds_id == "c19":
    # 仅 c19：object 列内裸换行会使 pivot 用 csv.reader 统计到的「记录数」大于 DataFrame 行数 → Sw。
    for _col in syn.columns:
        if syn[_col].dtype == object:
            syn[_col] = (
                syn[_col]
                .astype(str)
                .str.replace("\r\n", " ", regex=False)
                .str.replace("\n", " ", regex=False)
                .str.replace("\r", " ", regex=False)
            )
    syn = syn.iloc[:n_target].reset_index(drop=True)

syn.to_csv("/work/output-Benchmark-trainonly-v1/m4/arf/arf-m4-20260504_205355/arf-m4-2217-20260504_205409.csv", index=False)
print(f"[ARF] Generated {len(syn)} rows (requested {n_target}) -> /work/output-Benchmark-trainonly-v1/m4/arf/arf-m4-20260504_205355/arf-m4-2217-20260504_205409.csv")