Python numpy.nanmean() Examples

def test_select_confounds(confounds_file, selected_confounds, nan_confounds,
                          expanded_confounds):
    import pandas as pd
    import numpy as np

    confounds_df = pd.read_csv(str(confounds_file), sep='\t', na_values='n/a')

    res_df = _select_confounds(str(confounds_file), selected_confounds)

    # check if the correct columns are selected
    assert set(expanded_confounds) == set(res_df.columns)

    # check if nans are being imputed when expected
    if nan_confounds:
        for nan_c in nan_confounds:
            vals = confounds_df[nan_c].values
            expected_result = np.nanmean(vals[vals != 0])
            assert res_df[nan_c][0] == expected_result 

def eval_detection_voc(pred_boxlists, gt_boxlists, iou_thresh=0.5, use_07_metric=False):
    """Evaluate on voc dataset.
    Args:
        pred_boxlists(list[BoxList]): pred boxlist, has labels and scores fields.
        gt_boxlists(list[BoxList]): ground truth boxlist, has labels field.
        iou_thresh: iou thresh
        use_07_metric: boolean
    Returns:
        dict represents the results
    """
    assert len(gt_boxlists) == len(
        pred_boxlists
    ), "Length of gt and pred lists need to be same."
    prec, rec, n_tp, n_fp, n_pos = calc_detection_voc_prec_rec(
        pred_boxlists=pred_boxlists, gt_boxlists=gt_boxlists, iou_thresh=iou_thresh
    )
    ap = calc_detection_voc_ap(prec, rec, use_07_metric=use_07_metric)
    prec = {k: v.tolist() for k, v in prec.items()}
    rec = {k: v.tolist() for k, v in rec.items()}
    res = [{"ap": ap[k], "class_id": k, "precisions": prec[k], "recalls": rec[k],
            "n_tp": n_tp[k], "n_fp": n_fp[k], "n_positives": n_pos[k]} for k in ap.keys()]
    return res, np.nanmean(ap.values()) 

def _signal_synchrony_correlation(signal1, signal2, window_size, center=False):
    """Calculates pairwise rolling correlation at each time. Grabs the upper triangle, at each timepoints.

    - window: window size of rolling corr in samples
    - center: whether to center result (Default: False, so correlation values are listed on the right.)

    """
    data = pd.DataFrame({"y1": signal1, "y2": signal2})

    rolled = data.rolling(window=window_size, center=center).corr()
    synchrony = rolled["y1"].loc[rolled.index.get_level_values(1) == "y2"].values

    # Realign
    synchrony = np.append(synchrony[int(window_size / 2) :], np.full(int(window_size / 2), np.nan))
    synchrony[np.isnan(synchrony)] = np.nanmean(synchrony)

    return synchrony 

def compute_average_mAP_N_for_characteristic(sensitivity_analysis, characteristic_name):
    gt_by_characteristic = sensitivity_analysis.ground_truth.groupby(characteristic_name)
    average_mAP_n_by_characteristic_value = OrderedDict()
    
    for characteristic_value, this_characteristic_gt in gt_by_characteristic:
        ap = np.nan*np.zeros(len(sensitivity_analysis.activity_index))
        gt_by_cls = this_characteristic_gt.groupby('label')
        pred_by_cls = sensitivity_analysis.prediction.groupby('label')
        for cls in sensitivity_analysis.activity_index.values():
            this_cls_pred = pred_by_cls.get_group(cls).sort_values(by='score',ascending=False)
            try:
                this_cls_gt = gt_by_cls.get_group(cls)
            except:
                continue
            gt_id_to_keep = np.append(this_cls_gt['gt-id'].values, [np.nan])
        
            for tidx, tiou in enumerate(sensitivity_analysis.tiou_thresholds):
                this_cls_pred = this_cls_pred[this_cls_pred[sensitivity_analysis.matched_gt_id_cols[tidx]].isin(gt_id_to_keep)]
            
            ap[cls] = compute_mAP_N(sensitivity_analysis,this_cls_pred,this_cls_gt)

        average_mAP_n_by_characteristic_value[characteristic_value] = np.nanmean(ap)

    return average_mAP_n_by_characteristic_value 

def eval_detection_voc(pred_boxlists, gt_boxlists, iou_thresh=0.5, use_07_metric=False):
    """Evaluate on voc dataset.
    Args:
        pred_boxlists(list[BoxList]): pred boxlist, has labels and scores fields.
        gt_boxlists(list[BoxList]): ground truth boxlist, has labels field.
        iou_thresh: iou thresh
        use_07_metric: boolean
    Returns:
        dict represents the results
    """
    assert len(gt_boxlists) == len(
        pred_boxlists
    ), "Length of gt and pred lists need to be same."
    prec, rec = calc_detection_voc_prec_rec(
        pred_boxlists=pred_boxlists, gt_boxlists=gt_boxlists, iou_thresh=iou_thresh
    )
    ap = calc_detection_voc_ap(prec, rec, use_07_metric=use_07_metric)
    return {"ap": ap, "map": np.nanmean(ap)} 

def segmentation_scores(label_trues, label_preds, n_class):
    """Returns accuracy score evaluation result.
      - overall accuracy
      - mean accuracy
      - mean IU
      - fwavacc
    """
    hist = np.zeros((n_class, n_class))
    for lt, lp in zip(label_trues, label_preds):
        hist += _fast_hist(lt.flatten(), lp.flatten(), n_class)
    acc = np.diag(hist).sum() / hist.sum()
    acc_cls = np.diag(hist) / hist.sum(axis=1)
    acc_cls = np.nanmean(acc_cls)
    iu = np.diag(hist) / (hist.sum(axis=1) + hist.sum(axis=0) - np.diag(hist))
    mean_iu = np.nanmean(iu)
    freq = hist.sum(axis=1) / hist.sum()
    fwavacc = (freq[freq > 0] * iu[freq > 0]).sum()

    return {'overall_acc': acc,
            'mean_acc': acc_cls,
            'freq_w_acc': fwavacc,
            'mean_iou': mean_iu} 

def get_scores(self):
        """Returns accuracy score evaluation result.
            - overall accuracy
            - mean accuracy
            - mean IU
            - fwavacc
        """
        hist = self.confusion_matrix
        acc = np.diag(hist).sum() / hist.sum()
        acc_cls = np.diag(hist) / hist.sum(axis=1)
        acc_cls = np.nanmean(acc_cls)
        iu = np.diag(hist) / (hist.sum(axis=1) + hist.sum(axis=0) - np.diag(hist))
        mean_iu = np.nanmean(iu)
        freq = hist.sum(axis=1) / hist.sum()
        fwavacc = (freq[freq > 0] * iu[freq > 0]).sum()
        cls_iu = dict(zip(range(self.n_classes), iu))

        return {'Overall Acc: \t': acc,
                'Mean Acc : \t': acc_cls,
                'FreqW Acc : \t': fwavacc,
                'Mean IoU : \t': mean_iu,}, cls_iu 

def _recmat_smooth(presented, recalled, features, distance, match):

    if match == 'best':
        func = np.argmax
    elif match == 'smooth':
        func = np.nanmean

    simmtx = _similarity_smooth(presented, recalled, features, distance)


    if match == 'best':
        recmat = np.atleast_3d([func(s, 1) for s in simmtx]).astype(np.float64)
        recmat+=1
        recmat[np.isnan(simmtx).any(2)]=np.nan
    elif match == 'smooth':
        recmat = np.atleast_3d([func(s, 0) for s in simmtx]).astype(np.float64)


    return recmat 

def _similarity_smooth(presented, recalled, features, distance):
    lists = presented.index.get_values()
    res = np.empty((len(lists), len(features), recalled.iloc[0].shape[0], presented.iloc[0].shape[0]))*np.nan
    for li, l in enumerate(lists):
        p_list = presented.loc[l]
        r_list = recalled.loc[l]
        for i, feature in enumerate(features):
            get_feature = lambda x: np.array(x[feature]) if np.array(pd.notna(x['item'])).any() else np.nan
            p = np.vstack(p_list.apply(get_feature).get_values())
            r = r_list.dropna().apply(get_feature).get_values()
            r = np.vstack(list(filter(lambda x: x is not np.nan, r)))
            tmp = 1 - cdist(r, p, distance)
            res[li, i, :tmp.shape[0], :] =  tmp
    if distance == 'correlation':
        return np.nanmean(res, 1)
    else:
        return np.mean(res, 1) 

def _get_weight_best(egg, feature, distdict, permute, n_perms, distance):

    if permute:
        return _permute(egg, feature, distdict, _get_weight_best, n_perms)

    rec = list(egg.get_rec_items().values[0])
    if len(rec) <= 2:
        warnings.warn('Not enough recalls to compute fingerprint, returning default'
              'fingerprint.. (everything is .5)')
        return np.nan

    distmat = get_distmat(egg, feature, distdict)
    matchmat = get_match(egg, feature, distdict)

    ranks = []
    for i in range(len(rec)-1):
        cdx, ndx = np.argmin(matchmat[i, :]), np.argmin(matchmat[i+1, :])
        dists = distmat[cdx, :]
        di = dists[ndx]
        dists_filt = np.array([dist for idx, dist in enumerate(dists)])
        ranks.append(np.mean(np.where(np.sort(dists_filt)[::-1] == di)[0]+1) / len(dists_filt))
    return np.nanmean(ranks) 

def _get_weight_smooth(egg, feature, distdict, permute, n_perms, distance):

    if permute:
        return _permute(egg, feature, distdict, _get_weight_smooth, n_perms)

    rec = list(egg.get_rec_items().values[0])
    if len(rec) <= 2:
        warnings.warn('Not enough recalls to compute fingerprint, returning default'
              'fingerprint.. (everything is .5)')
        return np.nan

    distmat = get_distmat(egg, feature, distdict)
    matchmat = get_match(egg, feature, distdict)

    ranks = []
    for i in range(len(rec)-1):
        cdx, ndx = np.argmin(matchmat[i, :]), np.argmin(matchmat[i+1, :])
        dists = distmat[cdx, :]
        di = dists[ndx]
        dists_filt = np.array([dist for idx, dist in enumerate(dists)])
        ranks.append(np.mean(np.where(np.sort(dists_filt)[::-1] == di)[0]+1) / len(dists_filt))
    return np.nanmean(ranks) 

def eval_detection_voc(pred_boxlists, gt_boxlists, iou_thresh=0.5, use_07_metric=False):
    """Evaluate on voc dataset.
    Args:
        pred_boxlists(list[BoxList]): pred boxlist, has labels and scores fields.
        gt_boxlists(list[BoxList]): ground truth boxlist, has labels field.
        iou_thresh: iou thresh
        use_07_metric: boolean
    Returns:
        dict represents the results
    """
    assert len(gt_boxlists) == len(
        pred_boxlists
    ), "Length of gt and pred lists need to be same."
    prec, rec = calc_detection_voc_prec_rec(
        pred_boxlists=pred_boxlists, gt_boxlists=gt_boxlists, iou_thresh=iou_thresh
    )
    ap = calc_detection_voc_ap(prec, rec, use_07_metric=use_07_metric)
    return {"ap": ap, "map": np.nanmean(ap)} 

def _nanquantile_unchecked(a, q, axis=None, out=None, overwrite_input=False,
                           interpolation='linear', keepdims=np._NoValue):
    """Assumes that q is in [0, 1], and is an ndarray"""
    # apply_along_axis in _nanpercentile doesn't handle empty arrays well,
    # so deal them upfront
    if a.size == 0:
        return np.nanmean(a, axis, out=out, keepdims=keepdims)

    r, k = function_base._ureduce(
        a, func=_nanquantile_ureduce_func, q=q, axis=axis, out=out,
        overwrite_input=overwrite_input, interpolation=interpolation
    )
    if keepdims and keepdims is not np._NoValue:
        return r.reshape(q.shape + k)
    else:
        return r 

def cal_scores(hist):
    n_class = settings.N_CLASSES
    acc = np.diag(hist).sum() / hist.sum()
    acc_cls = np.diag(hist) / hist.sum(axis=1)
    acc_cls = np.nanmean(acc_cls)
    iu = np.diag(hist) / (hist.sum(axis=1) + hist.sum(axis=0) - np.diag(hist))
    mean_iu = np.nanmean(iu)
    freq = hist.sum(axis=1) / hist.sum()
    fwavacc = (freq[freq > 0] * iu[freq > 0]).sum()
    cls_iu = dict(zip(label_names, iu))

    return {
        'pAcc': acc,
        'mAcc': acc_cls,
        'fIoU': fwavacc,
        'mIoU': mean_iu,
    }, cls_iu 

def cal_scores(hist):
    n_class = settings.N_CLASSES
    #acc = np.diag(hist).sum() / hist.sum()
    #acc_cls = np.diag(hist) / hist.sum(axis=1)
    #acc_cls = np.nanmean(acc_cls)
    iu = np.diag(hist) / (hist.sum(axis=1) + hist.sum(axis=0) - np.diag(hist))
    mean_iu = np.nanmean(iu)
    freq = hist.sum(axis=1) / hist.sum()
    fwavacc = (freq[freq > 0] * iu[freq > 0]).sum()
    cls_iu = dict(zip(label_names, iu))

    return {
        #"pAcc": acc,
        #"mAcc": acc_cls,
        "fIoU": fwavacc,
        "mIoU": mean_iu,
    }#, cls_iu 

def mape(y_pred, y_test):
    r"""
    The Mean Absolute Percentage Error (MAPE)

    The MAPE is computed as the mean of the absolute value of the relative
    error in percent, i.e.:

    .. math::

        \text{MAPE}(\mathbf{y}, \mathbf{y}_{true}) =
        \frac{100\%}{n}\sum_{i = 0}^n \frac{|y_{\text{pred},i} - y_{\text{true},i}|}
             {|y_{\text{true},i}|}

    Arguments:

        y_pred(numpy.array): The predicted scalar values.

        y_test(numpy.array): The true values.

    Returns:

        The MAPE for the given predictions.

    """
    return np.nanmean(100.0 * np.abs(y_test - y_pred.ravel()) / np.abs(y_test).ravel()) 

def print_iou(iu, mean_pixel_acc, class_names=None, show_no_back=False):
    n = iu.size
    lines = []
    for i in range(n):
        if class_names is None:
            cls = 'Class %d:' % (i + 1)
        else:
            cls = '%d %s' % (i + 1, class_names[i])
        # lines.append('%-8s: %.3f%%' % (cls, iu[i] * 100))
    mean_IU = np.nanmean(iu)
    mean_IU_no_back = np.nanmean(iu[1:])
    if show_no_back:
        lines.append('mean_IU: %.3f%% || mean_IU_no_back: %.3f%% || mean_pixel_acc: %.3f%%' % (
            mean_IU * 100, mean_IU_no_back * 100, mean_pixel_acc * 100))
    else:
        lines.append('mean_IU: %.3f%% || mean_pixel_acc: %.3f%%' % (mean_IU * 100, mean_pixel_acc * 100))
    lines.append('=================================================')
    line = "\n".join(lines)

    print(line) 

def scores(label_trues, label_preds, n_class):
    hist = np.zeros((n_class, n_class))
    for lt, lp in zip(label_trues, label_preds):
        if(lt.size > 0):
            hist += _fast_hist(lt.flatten(), lp.flatten(), n_class)
    acc = np.diag(hist).sum() / hist.sum()
    acc_cls = np.diag(hist) / hist.sum(axis=1)
    acc_cls = np.nanmean(acc_cls)
    iu = np.diag(hist) / (hist.sum(axis=1) + hist.sum(axis=0) - np.diag(hist))
    mean_iu = np.nanmean(iu)
    freq = hist.sum(axis=1) / hist.sum()
    fwavacc = (freq[freq > 0] * iu[freq > 0]).sum()
    cls_iu = dict(zip(range(n_class), iu))

    return {
        "Overall Acc": acc,
        "Mean Acc": acc_cls,
        "FreqW Acc": fwavacc,
        "Mean IoU": mean_iu,
    }, cls_iu 

def scores(label_trues, label_preds, n_class):
    hist = np.zeros((n_class, n_class))
    for lt, lp in zip(label_trues, label_preds):
        if(lt.size > 0):
            hist += _fast_hist(lt.flatten(), lp.flatten(), n_class)
    acc = np.diag(hist).sum() / hist.sum()
    acc_cls = np.diag(hist) / hist.sum(axis=1)
    acc_cls = np.nanmean(acc_cls)
    iu = np.diag(hist) / (hist.sum(axis=1) + hist.sum(axis=0) - np.diag(hist))
    mean_iu = np.nanmean(iu)
    freq = hist.sum(axis=1) / hist.sum()
    fwavacc = (freq[freq > 0] * iu[freq > 0]).sum()
    cls_iu = dict(zip(range(n_class), iu))

    return {
        "Overall Acc": acc,
        "Mean Acc": acc_cls,
        "FreqW Acc": fwavacc,
        "Mean IoU": mean_iu,
    }, cls_iu 

def execute_agg(cls, ctx, op):
        axis = cls.get_axis(op.axis)

        a = ctx[op.inputs[0].key]
        if not isinstance(a, (list, tuple)):
            (inp,), device_id, xp = as_same_device(
                [a], device=op.device, ret_extra=True)

            with device(device_id):
                ctx[op.outputs[0].key] = xp.nanmean(inp, axis=axis, dtype=op.dtype,
                                                    keepdims=bool(op.keepdims))
        else:
            (_data, _count), device_id, xp = as_same_device(
                a, device=op.device, ret_extra=True)

            with device(device_id):
                chunk_count = xp.sum(_count, axis=axis, dtype=op.dtype,
                                     keepdims=bool(op.keepdims))
                chunk_sum = xp.sum(_data, axis=axis, dtype=op.dtype,
                                   keepdims=bool(op.keepdims))
                ctx[op.outputs[0].key] = xp.true_divide(chunk_sum, chunk_count,
                                                        dtype=op.dtype) 

def print_depth_stats(mids, err):
    shape_ids = np.unique([m[0] for m in mids])
    serr = dict()
    for sid in shape_ids:
        serr[sid] = []

    for ex, e in enumerate(err):
        serr[mids[ex][0]].append(e)

    table = []
    smean = []
    for s in serr:
        sm = np.nanmean(serr[s])
        table.append([s, sm])
        smean.append(sm)
    table.append(['Mean', np.nanmean(smean)])
    ptable = tabulate(table, headers=['SID', 'L1 error'], floatfmt=".4f")
    return smean, ptable 

def scores(label_trues, label_preds, n_class):
    hist = np.zeros((n_class, n_class))
    for lt, lp in zip(label_trues, label_preds):
        hist += _fast_hist(lt.flatten(), lp.flatten(), n_class)
    acc = np.diag(hist).sum() / hist.sum()
    acc_cls = np.diag(hist) / hist.sum(axis=1)
    acc_cls = np.nanmean(acc_cls)
    iu = np.diag(hist) / (hist.sum(axis=1) + hist.sum(axis=0) - np.diag(hist))
    valid = hist.sum(axis=1) > 0  # added
    mean_iu = np.nanmean(iu[valid])
    freq = hist.sum(axis=1) / hist.sum()
    fwavacc = (freq[freq > 0] * iu[freq > 0]).sum()
    cls_iu = dict(zip(range(n_class), iu))

    return {
        "Overall Acc": acc,
        "Mean Acc": acc_cls,
        "FreqW Acc": fwavacc,
        "Mean IoU": mean_iu,
        "Class IoU": cls_iu,
    } 

def _nanquantile_unchecked(a, q, axis=None, out=None, overwrite_input=False,
                           interpolation='linear', keepdims=np._NoValue):
    """Assumes that q is in [0, 1], and is an ndarray"""
    # apply_along_axis in _nanpercentile doesn't handle empty arrays well,
    # so deal them upfront
    if a.size == 0:
        return np.nanmean(a, axis, out=out, keepdims=keepdims)

    r, k = function_base._ureduce(
        a, func=_nanquantile_ureduce_func, q=q, axis=axis, out=out,
        overwrite_input=overwrite_input, interpolation=interpolation
    )
    if keepdims and keepdims is not np._NoValue:
        return r.reshape(q.shape + k)
    else:
        return r 

def evaluate(self):
    """Compute evaluation result.

    Returns:
      average_precision_per_class: float numpy array of average precision for
          each class.
      mean_ap: mean average precision of all classes, float scalar
      precisions_per_class: List of precisions, each precision is a float numpy
          array
      recalls_per_class: List of recalls, each recall is a float numpy array
      corloc_per_class: numpy float array
      mean_corloc: Mean CorLoc score for each class, float scalar
    """
    if (self.num_gt_instances_per_class == 0).any():
      logging.warn(
          'The following classes have no ground truth examples: %s',
          np.squeeze(np.argwhere(self.num_gt_instances_per_class == 0)))
    for class_index in range(self.num_class):
      if self.num_gt_instances_per_class[class_index] == 0:
        continue
      scores = np.concatenate(self.scores_per_class[class_index])
      tp_fp_labels = np.concatenate(self.tp_fp_labels_per_class[class_index])
      precision, recall = metrics.compute_precision_recall(
          scores, tp_fp_labels, self.num_gt_instances_per_class[class_index])
      self.precisions_per_class.append(precision)
      self.recalls_per_class.append(recall)
      average_precision = metrics.compute_average_precision(precision, recall)
      self.average_precision_per_class[class_index] = average_precision

    self.corloc_per_class = metrics.compute_cor_loc(
        self.num_gt_imgs_per_class,
        self.num_images_correctly_detected_per_class)

    mean_ap = np.nanmean(self.average_precision_per_class)
    mean_corloc = np.nanmean(self.corloc_per_class)
    return (self.average_precision_per_class, mean_ap,
            self.precisions_per_class, self.recalls_per_class,
            self.corloc_per_class, mean_corloc) 

def evaluate(self):
        acc = np.diag(self.hist).sum() / self.hist.sum()
        acc_cls = np.diag(self.hist) / self.hist.sum(axis=1)
        acc_cls = np.nanmean(acc_cls)
        iu = np.diag(self.hist) / (self.hist.sum(axis=1) + self.hist.sum(axis=0) - np.diag(self.hist))
        mean_iu = np.nanmean(iu)
        freq = self.hist.sum(axis=1) / self.hist.sum()
        fwavacc = (freq[freq > 0] * iu[freq > 0]).sum()
        return acc, acc_cls, iu, mean_iu, fwavacc 

def evaluate(self):
    """Compute evaluation result.

    Returns:
      average_precision_per_class: float numpy array of average precision for
          each class.
      mean_ap: mean average precision of all classes, float scalar
      precisions_per_class: List of precisions, each precision is a float numpy
          array
      recalls_per_class: List of recalls, each recall is a float numpy array
      corloc_per_class: numpy float array
      mean_corloc: Mean CorLoc score for each class, float scalar
    """
    if (self.num_gt_instances_per_class == 0).any():
      logging.warn(
          'The following classes have no ground truth examples: %s',
          np.squeeze(np.argwhere(self.num_gt_instances_per_class == 0)))
    for class_index in range(self.num_class):
      if self.num_gt_instances_per_class[class_index] == 0:
        continue
      scores = np.concatenate(self.scores_per_class[class_index])
      tp_fp_labels = np.concatenate(self.tp_fp_labels_per_class[class_index])
      precision, recall = metrics.compute_precision_recall(
          scores, tp_fp_labels, self.num_gt_instances_per_class[class_index])
      self.precisions_per_class.append(precision)
      self.recalls_per_class.append(recall)
      average_precision = metrics.compute_average_precision(precision, recall)
      self.average_precision_per_class[class_index] = average_precision

    self.corloc_per_class = metrics.compute_cor_loc(
        self.num_gt_imgs_per_class,
        self.num_images_correctly_detected_per_class)

    mean_ap = np.nanmean(self.average_precision_per_class)
    mean_corloc = np.nanmean(self.corloc_per_class)
    return (self.average_precision_per_class, mean_ap,
            self.precisions_per_class, self.recalls_per_class,
            self.corloc_per_class, mean_corloc) 

def Pixel_Accuracy_Class(self):
        Acc = np.diag(self.confusion_matrix) / self.confusion_matrix.sum(axis=1)
        Acc = np.nanmean(Acc)
        return Acc 

def Mean_Intersection_over_Union(self):
        MIoU = np.diag(self.confusion_matrix) / (
                    np.sum(self.confusion_matrix, axis=1) + np.sum(self.confusion_matrix, axis=0) -
                    np.diag(self.confusion_matrix))
        MIoU = np.nanmean(MIoU)
        return MIoU 

def _do(self, F, **kwargs):

        n, m = F.shape

        if self.normalize:
            F = normalize(F, self.ideal_point, self.nadir_point, estimate_bounds_if_none=True)

        neighbors_finder = NeighborFinder(F, n_min_neigbors="auto")

        mu = np.full(n, - np.inf)

        # for each solution in the set calculate the least amount of improvement per unit deterioration
        for i in range(n):

            # neighbors to the current point
            neighbors = neighbors_finder.find(i)

            # calculate the trade-off to all neighbours
            diff = F[neighbors] - F[i]

            # calculate sacrifice and gain
            sacrifice = np.maximum(0, diff).sum(axis=1)
            gain = np.maximum(0, -diff).sum(axis=1)

            np.warnings.filterwarnings('ignore')
            tradeoff = sacrifice / gain

            # otherwise find the one with the smalled one
            mu[i] = np.nanmean(tradeoff)

        return find_outliers_upper_tail(mu) 

def _ecg_rsa_p2t(rsp_onsets, rpeaks, sampling_rate, continuous=False, ecg_period=None, rsp_peaks=None):
    """Peak-to-trough algorithm (P2T)"""

    # Find all RSP cycles and the Rpeaks within
    cycles_rri = []
    for idx in range(len(rsp_onsets) - 1):
        cycle_init = rsp_onsets[idx]
        cycle_end = rsp_onsets[idx + 1]
        cycles_rri.append(rpeaks[np.logical_and(rpeaks >= cycle_init, rpeaks < cycle_end)])

    # Iterate over all cycles
    rsa_values = np.full(len(cycles_rri), np.nan)
    for i, cycle in enumerate(cycles_rri):
        # Estimate of RSA during each breath
        RRis = np.diff(cycle) / sampling_rate
        if len(RRis) > 1:
            rsa_values[i] = np.max(RRis) - np.min(RRis)

    if continuous is False:
        rsa = {"RSA_P2T_Mean": np.nanmean(rsa_values)}
        rsa["RSA_P2T_Mean_log"] = np.log(rsa["RSA_P2T_Mean"])  # pylint: disable=E1111
        rsa["RSA_P2T_SD"] = np.nanstd(rsa_values, ddof=1)
        rsa["RSA_P2T_NoRSA"] = len(pd.Series(rsa_values).index[pd.Series(rsa_values).isnull()])
    else:
        rsa = signal_interpolate(
            x_values=rsp_peaks[~np.isnan(rsa_values)],
            y_values=rsa_values[~np.isnan(rsa_values)],
            x_new=np.arange(len(ecg_period)),
        )

    return rsa