将 Pytorch 与 Pytorch-Lightning ROC 指标一起使用时出现分段错误

如何解决将 Pytorch 与 Pytorch-Lightning ROC 指标一起使用时出现分段错误

对于机器学习任务，我正在转换两个张量，如代码片段 (1) 所示，以使它们符合 pytorch-lightning 的 ROC 指标的预期格式。然后将两个结果张量 y_test (gt_classes) 和 y_score (detection_scores) 传递给代码片段 (2) 中的函数 calc_multi_class，其中我使用了来自pytorch 闪电：fpr,tpr,thr = roc(y_score,y_test)。 `

由于我集成了 pytorch-lightning 并使用了它们的 ROC 功能，因此我收到了如下部分所示的错误消息“分段错误”。

我使用 Python 3.7 和 Pytorch 1.7、Pytorch-Lightning 1.1.2 和 Cuda 10.2

    Part (1)
    ...
    # sort by class and by confidence value
    # gt_classes = labels[:,0].int()
    gt_classes_sorted = labels[labels[:,0].sort(descending=False)[1]]
    gt_classes = gt_classes_sorted[:,0].int()
    # x = x[x[:,4].argsort(descending=True)]
    detections_sorted = detection[detection[:,4].sort(descending=True)[1]]
    det_classes_sorted = detections_sorted[:,5].int()

    # filter for scores
    input = torch.empty(gt_classes.numel(),self.nc)
    detection_scores = torch.ones_like(input) * 0.01

    # check if no scores
    if detections_sorted.nelement() > 0 and det_classes_sorted.nelement() > 0:
        counter = 0
        for j,dc in enumerate(det_classes_sorted):
            if counter < gt_classes.numel():
                for i,gc in enumerate(gt_classes):
                    # classes gc equals dc classes
                    if gc == dc:
                        # conf value
                        print('Found detection score {}'.format(detections_sorted[j,4]))
                        detection_scores[i,gc] = detections_sorted[j,4]
                        counter = counter + 1
                    else:
                        print('Other detection score {}'.format(detections_sorted[j,dc] = detections_sorted[j,4] #conf value        
    #print('Result classes {} scores {}'.format(gt_classes,detection_scores))
    
    gt_classes,detection_scores = gt_classes.to(self.device),detection_scores.to(self.device)
    return gt_classes,detection_scores


   Part (2)
    def calc_multi_class_(self,y_test,y_score,num_classes=3):
    """ Use of Pytorch Lightning Metrics Library for ROC"""        

    fprs = dict()
    tprs = dict()
    roc_aucs = dict()
    thresholds = dict()

    # torch to cuda device
    if isinstance(y_test,np.ndarray):
        y_test = torch.from_numpy(y_test)
        y_test = y_test.cpu()

    if isinstance(y_score,np.ndarray):
        y_score = torch.from_numpy(y_score)
        y_score = y_score.cpu()
        
    if y_score.nelement() > 0 and torch.max(y_score) > 0.2:
        roc = classification.ROC(num_classes=num_classes)  # no neg classes
        fpr,y_test)
        
        # convert multi-dim tensor to numpy ndarray          
        for i in range(num_classes):
            thresholds[i] = thr[i].numpy()
            fprs[i],tprs[i] = fpr[i].numpy(),tpr[i].numpy()
            roc_aucs[i] = auc(fprs[i],tprs[i])

    return fprs,tprs,roc_aucs,thresholds

由于我对 Pytorch 还很陌生，我试图弄清楚它来自哪里，并查看了我是否设置了错误的张量，但还是无法弄清楚。如果有人有提示或想法，这将非常有帮助！！

产生的错误信息如下：

    Result classes tensor([1],device='cuda:0',dtype=torch.int32) scores tensor([[0.01000,0.01000,0.01000]],device='cuda:0')
           Class      Images     Targets           P           R      mAP@.5  mAP@.5:.95: 100%|██████████| 58/58 [00:13<00:00,4.41it/s]
             all         928         489           0           0    3.57e-06    5.65e-07

/home/jm0365/.direnv/python-3.7.3/lib/python3.7/site-packages/pytorch_lightning/utilities/distributed.py:49: UserWarning: Metric ROC 将保存所有目标和缓冲区中的预测。对于大型数据集，这可能会导致大量内存占用。 warnings.warn(*args,**kwargs) 分段错误