scorer.py

import pandas as pd
from sklearn.metrics import accuracy_score, precision_score, recall_score, f1_score, confusion_matrix

TARGET_LABEL = "label_shock_boundary_transition"


def normalize_label(df):
    label_cols = [c for c in df.columns if c.startswith("label_")]

    if len(label_cols) == 1:
        return df.rename(columns={label_cols[0]: "label"})

    if len(label_cols) == 0:
        raise ValueError(
            f"No label column found. Expected {TARGET_LABEL} or a single label_* column."
        )

    raise ValueError(f"Multiple label columns found: {label_cols}")


def validate_binary_labels(series, name):
    unique_values = set(series.dropna().unique())
    if not unique_values.issubset({0, 1}):
        raise ValueError(
            f"{name} must contain only binary values 0/1. Found: {sorted(unique_values)}"
        )


def validate_lengths(y_true, y_pred):
    if len(y_true) != len(y_pred):
        raise ValueError(
            f"Prediction length mismatch. predictions={len(y_pred)} ground_truth={len(y_true)}"
        )


def validate_prediction_columns(preds):
    allowed = {"scenario_id", "label"}
    extra_cols = [c for c in preds.columns if c not in allowed]
    if extra_cols:
        raise ValueError(
            f"Predictions file should contain only scenario_id and label columns after normalization. Extra columns found: {extra_cols}"
        )


def score(predictions_path, ground_truth_path):
    preds = pd.read_csv(predictions_path)
    truth = pd.read_csv(ground_truth_path)

    truth = normalize_label(truth)
    preds = normalize_label(preds)

    if "label" not in preds.columns:
        raise ValueError("Predictions file must contain a label column")

    if "label" not in truth.columns:
        raise ValueError("Ground truth file must contain a label column")

    validate_prediction_columns(preds)

    y_true = truth["label"]
    y_pred = preds["label"]

    validate_lengths(y_true, y_pred)
    validate_binary_labels(y_true, "Ground truth labels")
    validate_binary_labels(y_pred, "Prediction labels")

    accuracy = accuracy_score(y_true, y_pred)
    precision = precision_score(y_true, y_pred, zero_division=0)
    recall = recall_score(y_true, y_pred, zero_division=0)
    f1 = f1_score(y_true, y_pred, zero_division=0)

    cm = confusion_matrix(y_true, y_pred, labels=[0, 1])
    tn, fp, fn, tp = cm.ravel()

    false_safe_rate = fn / (fn + tp) if (fn + tp) > 0 else 0.0
    true_safe_rate = tn / (tn + fp) if (tn + fp) > 0 else 0.0
    positive_prediction_rate = (tp + fp) / len(y_pred) if len(y_pred) > 0 else 0.0

    return {
        "target_label": TARGET_LABEL,
        "n_samples": int(len(y_true)),
        "accuracy": float(accuracy),
        "precision": float(precision),
        "recall_cascade_detection": float(recall),
        "false_safe_rate": float(false_safe_rate),
        "true_safe_rate": float(true_safe_rate),
        "positive_prediction_rate": float(positive_prediction_rate),
        "f1": float(f1),
        "confusion_matrix": cm.tolist(),
    }


if __name__ == "__main__":
    import sys

    if len(sys.argv) != 3:
        raise ValueError("Usage: python scorer.py <predictions.csv> <ground_truth.csv>")

    predictions_path = sys.argv[1]
    ground_truth_path = sys.argv[2]

    print(score(predictions_path, ground_truth_path))