Function KL_max_disagreement

modAL/disagreement.py:72–99

Calculates the max disagreement for the Committee. First it computes the class probabilties of X for each learner in the Committee, then calculates the consensus probability distribution by averaging the individual class probabilities for each learner. Then each learner's class probabil

(committee: BaseCommittee, X: modALinput, **predict_proba_kwargs)

def KL_max_disagreement(committee: BaseCommittee, X: modALinput, **predict_proba_kwargs) -> np.ndarray:
    """
    Calculates the max disagreement for the Committee. First it computes the class probabilties of X for each learner in
    the Committee, then calculates the consensus probability distribution by averaging the individual class
    probabilities for each learner. Then each learner's class probabilities are compared to the consensus distribution
    in the sense of Kullback-Leibler divergence. The max disagreement for a given sample is the argmax of the KL
    divergences of the learners from the consensus probability.

    Args:
        committee: The :class:`modAL.models.BaseCommittee` instance for which the max disagreement is to be calculated.
        X: The data for which the max disagreement is to be calculated.
        **predict_proba_kwargs: Keyword arguments for the :meth:`predict_proba` of the Committee.

    Returns:
        Max disagreement of the Committee for the samples in X.
    """
    try:
        p_vote = committee.vote_proba(X, **predict_proba_kwargs)
    except NotFittedError:
        return np.zeros(shape=(X.shape[0],))

    p_consensus = np.mean(p_vote, axis=1)

    learner_KL_div = np.zeros(shape=(X.shape[0], len(committee)))
    for learner_idx, _ in enumerate(committee):
        learner_KL_div[:, learner_idx] = entropy(np.transpose(p_vote[:, learner_idx, :]), qk=np.transpose(p_consensus))

    return np.max(learner_KL_div, axis=1)


def vote_entropy_sampling(committee: BaseCommittee, X: modALinput,