def compute_model_errors()

in src/minmaxML.py [0:0]

• 20 lines of code
• 14 McCabe index (conditional complexity)

def compute_model_errors(modelhat, X, y, t, errors, error_type, penalty='none', C=1.0):
    """
    Computes the error of the round-specific model and puts the errors for each sample in column t of `errors` in place
    """
    yhat = modelhat.predict(X).ravel()  # Compute predictions for the newly trained model
    # Compute the specified error type
    if error_type == 'MSE':
        errors[t, :] = np.power(y - yhat, 2)
    elif error_type == '0/1 Loss':  # Classification 0-1 Loss
        errors[t, :] = (y != yhat)
    elif error_type == 'FP':  # FP rate
        errors[t, :] = (y < yhat)
    elif error_type == 'FN':  # FN rate
        errors[t, :] = (y > yhat)
    elif error_type == 'Log-Loss':  # Log loss, the convex surrogate used by logistic regression
        errors[t, :] = compute_logloss(y, modelhat.predict_proba(X))
    elif error_type == 'FP-Log-Loss':  # Computes the log loss, but replaces loss with 0 unless an instance was a FP
        errors[t, :] = (y < yhat) * compute_logloss(y, modelhat.predict_proba(X))
    elif error_type == 'FN-Log-Loss':  # Computes the log loss, but replaces loss with 0 unless an instance was a FN
        errors[t, :] = (y > yhat) * compute_logloss(y, modelhat.predict_proba(X))
    else:
        raise ValueError(f"\'{error_type}\' is an invalid error type")
    # Compute the regularization penalty if necessary and add it to the log loss
    if penalty in ['l1', 'l2'] and C > 1e15:
        errors[t, :] += compute_regularization_penalty(modelhat.coef_, penalty, C) * (0.5 if penalty == 'l2' else 1.0)