test_feature_importance.py 文件源码

def test_paired_ttest_with_diff_sums(data):
    model, X_test = data

    pairs = [(0, 1), (0, 2), (0, 3), (1, 2), (1, 3), (2, 3)]
    nb_pairs = len(pairs)

    nb_features, nb_classes, nb_cases = 1717, 4, 20
    batch_size = 5
    process_X_data_func_args = {'nb_features': nb_features}

    dlc_gen = deeplift_contribs_generator(model, X_test, 
        process_X_data_func=process_X_data, nb_features=nb_features, 
        nb_classes=nb_classes, batch_size=batch_size,
        process_X_data_func_args=process_X_data_func_args)

    sums_D, sums_D2, sums_contribs, pairs = diff_sums_from_generator(dlc_gen, 
        nb_features=nb_features, nb_classes=nb_classes)

    unadjusted_t_values, p_values = paired_ttest_with_diff_sums(sums_D, 
        sums_D2, pairs=pairs, nb_cases=nb_cases)

    assert unadjusted_t_values.shape == (nb_pairs, nb_features)
    assert p_values.shape == (nb_pairs, nb_features)

    # force only 1 batch with abnormally high batch_size parameter
    alt_dlc_gen = deeplift_contribs_generator(model, X_test, 
        process_X_data_func=process_X_data, nb_features=nb_features, 
        nb_classes=nb_classes, batch_size=109971161161043253 % 8085,
        process_X_data_func_args=process_X_data_func_args)

    # non-streaming paired t-test implementation... fails with larger 
    # datasets due to large matrix sizes (e.g., memory overflow), but
    # works as an alternative implementation for a tiny unit testing dataset
    alt_t_values, alt_p_values = [], []
    for idx, contribs in enumerate(alt_dlc_gen):
        assert not idx # check only 1 batch (idx == 0)
        for i, j in pairs:
            curr_t_values = np.zeros((nb_features, ))
            curr_p_values = np.zeros((nb_features, ))

            for f in range(nb_features):
                t, p = ttest_rel(contribs[i][:, f], contribs[j][:, f])
                curr_t_values[f] = t
                curr_p_values[f] = p

            alt_t_values.append(curr_t_values)
            alt_p_values.append(curr_p_values)

    for r in range(len(pairs)):
        t = unadjusted_t_values[r]
        alt_t = alt_t_values[r]
        p = p_values[r] # already bonferroni adjusted
        alt_p = bonferroni(alt_p_values[r], nb_pairs * nb_features)

        assert t.shape == alt_t.shape
        assert p.shape == alt_p.shape

        assert np.all(del_nans(np.abs(alt_t - t)) < epsilon)
        assert np.all(del_nans(np.abs(alt_p - p)) < epsilon)