tormenta / src /scoring.py
aliza-ayaz's picture
Live Space: Streamlit app running in the browser (stlite)
7e41cd6 verified
Raw History Blame Contribute Delete
1.39 kB
"""Lightweight scoring helpers (numpy/pandas only).
Kept separate from evaluate.py (which needs scikit-learn and matplotlib for plots and permutation importance) so
the browser app can import them without downloading those packages. evaluate.py re-exports both functions.
Formulas match sklearn.metrics exactly (tests/test_portable.py checks it).
"""
import numpy as np
import pandas as pd
from src import config
def metrics(y, pred) -> dict:
"""Overall RMSE/MAE/R² plus RMSE/MAE on storm hours (Kp ≥ 5-)."""
y, pred = np.asarray(y, dtype=float), np.asarray(pred, dtype=float)
err = pred - y
storm = y >= config.STORM_THRESHOLD
ss_tot = np.sum((y - y.mean()) ** 2)
return {
"rmse": float(np.sqrt(np.mean(err ** 2))),
"mae": float(np.mean(np.abs(err))),
"r2": float(1 - np.sum(err ** 2) / ss_tot) if ss_tot > 0 else np.nan,
"storm_rmse": float(np.sqrt(np.mean(err[storm] ** 2))) if storm.any() else np.nan,
"storm_mae": float(np.mean(np.abs(err[storm]))) if storm.any() else np.nan,
"n": len(y),
}
def attach_block_kp(pred: pd.DataFrame, kp_blocks: pd.Series, name: str) -> pd.DataFrame:
"""Add the Kp of the 3-h block containing each hour (e.g. NOAA estimated or GFZ Kp)."""
block = pred.index.floor(f"{config.KP_BLOCK_HOURS}h")
return pred.assign(**{name: kp_blocks.reindex(block).to_numpy()})