netcal.regularization.confidence_penalty

netcal.regularization.confidence_penalty(X: ndarray, weight: float, threshold: float = None, base: float = 2.718281828459045) → float

Confidence Penalty Regularization. This penalty term can be applied to any loss function as a regularizer [1].

Parameters:

  • X (np.ndarray, shape=(n_samples, [n_classes])) – NumPy array with confidence values for each prediction. 1-D for binary classification, 2-D for multi class (softmax).

  • weight (float) – Weight of entropy.

  • threshold (float, optional, default: None) – Entropy threshold (no penalty is assigned above threshold).

  • base (float, optional, default: np.e) – Base of logarithm (typically the number of classes to norm entropy).

Returns:

Confidence penalty of posterior distribution.

Return type:

float

References