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

import os
import pickle
import numpy as np
import pandas as pd
from arfpy import arf

def _sanitize_for_arf(df: pd.DataFrame) -> pd.DataFrame:
    """缓解 forge 阶段 scipy.stats.truncnorm / 除零：处理 inf、NaN 与极端尾部。"""
    df = df.replace([np.inf, -np.inf], np.nan)
    df = df.dropna(axis=1, how="all")
    for col in df.select_dtypes(include=[np.number]).columns:
        med = df[col].median()
        if pd.isna(med):
            med = 0.0
        df[col] = df[col].fillna(med)
        nu = int(df[col].nunique(dropna=True))
        if nu <= 1:
            continue
        q_low = float(os.environ.get("ARF_CLIP_QUANTILE_LOW", "0.001"))
        q_high = float(os.environ.get("ARF_CLIP_QUANTILE_HIGH", "0.999"))
        lo, hi = df[col].quantile(q_low), df[col].quantile(q_high)
        if pd.notna(lo) and pd.notna(hi) and lo < hi:
            df[col] = df[col].clip(lo, hi)
    return df

df = pd.read_csv("/work/output-Benchmark-trainonly-v1/m4/arf/arf-m4-20260504_205355/staged/public/train.csv")
df = _sanitize_for_arf(df)
num_trees = int(os.environ.get("ARF_NUM_TREES", "30"))
delta = float(os.environ.get("ARF_DELTA", "0"))
max_iters = int(os.environ.get("ARF_MAX_ITERS", "10"))
early_stop = (os.environ.get("ARF_EARLY_STOP", "true").strip().lower() in ("1", "true", "yes"))
verbose = (os.environ.get("ARF_VERBOSE", "true").strip().lower() in ("1", "true", "yes"))
min_node_size = int(os.environ.get("ARF_MIN_NODE_SIZE", "5"))
print(f"[ARF] Training on {len(df)} rows, {len(df.columns)} cols")
print(f"[ARF] Config num_trees={num_trees} delta={delta} max_iters={max_iters} early_stop={early_stop} min_node_size={min_node_size}")

model = arf.arf(x=df, num_trees=num_trees, delta=delta, max_iters=max_iters, early_stop=early_stop, verbose=verbose, min_node_size=min_node_size)
if hasattr(model, "fit"):
    model.fit()
elif hasattr(model, "forde"):
    model.forde()
else:
    raise RuntimeError("arfpy API: no fit() / forde()")

with open("/work/output-Benchmark-trainonly-v1/m4/arf/arf-m4-20260504_205355/arf_model.pkl", "wb") as f:
    pickle.dump(model, f)
print(f"[ARF] Model saved -> /work/output-Benchmark-trainonly-v1/m4/arf/arf-m4-20260504_205355/arf_model.pkl")