utils.py

import pandas as pd
import numpy as np


def infer_from(model, X, sensitive_features=None):
    """Infer from binary-label model."""
    kwargs = {}
    if sensitive_features is not None:
        kwargs = dict(sensitive_features=sensitive_features)

    if hasattr(model, "predict_proba"):
        infer_fn = lambda x: model.predict_proba(x)[:, 1]
    elif hasattr(model, "predict"):
        infer_fn = lambda x, **kwargs: model._pmf_predict(x, **kwargs)[:, 1]

    return infer_fn(X.values if isinstance(X, pd.DataFrame) else X, **kwargs)


def score(model, X, y_true, sensitive_features=None):
    kwargs = {}
    if sensitive_features is not None:
        kwargs = dict(sensitive_features=sensitive_features)
    y_pred = model.predict(X.values if isinstance(X, pd.DataFrame) else X,
                           **kwargs)
    return (y_pred == y_true).mean()


def log_losses(y_true, y_pred, eps=1e-15):
    if not isinstance(y_pred, np.ndarray):
        y_pred = np.array(y_pred)
    if not isinstance(y_true, np.ndarray):
        y_true = np.array(y_true)
    y_pred = np.clip(y_pred, eps, 1 - eps)
    one_minus_y_pred = np.clip(1 - y_pred, eps, 1 - eps)
    return -(y_true * np.log(y_pred) + (1 - y_true) * np.log(one_minus_y_pred))


def normalize_vuln_mean(vuln_vals):
    return (2 * vuln_vals.mean() - 1) * 100


def normalize_vuln_std(vuln_vals):
    return 2 * vuln_vals.std() * 100