How to transform y_pred, y_true for multi-class ROC for Object Detection.

[JaninaMattes]

Hi, I am using Pytorch Lightning for multi-class ROC and AUC calculations for object detection with YOLOv3. As I am still quite new to Pytorch and Pytorch Lightning I got a question, concerning how to transform the y_true and y_pred tensors correctly by taking in account the IoU between the y_pred and y_true bounding boxes.

My object detection model's output (y_pred) is of the shape (Array[N, 6]), x1, y1, x2, y2, conf, class and the target format (y_true) is of the shape (Array[M, 5]), class, x1, y1, x2, y2. This question also might be related to the discussion under #4988 . In order to correctly calculate ROC and AUC, how can I transform both of my tensors for y_true and y_pred?

Example of my output tensor (y_pred) and target tensor (y_true).

pred_y:
tensor([[2.13750e+01, 2.37875e+02, 1.27625e+02, 3.01250e+02, 1.08051e-03, 2.00000e+00]])

true_y: 
tensor([[  0.00000, 100.70072, 161.80963, 311.85648, 210.58191]])

Many thanks for your help!

Example for multi-class ROC in the documentation:

>>> pred = torch.tensor([[0.75, 0.05, 0.05, 0.05],
        ...                      [0.05, 0.75, 0.05, 0.05],
        ...                      [0.05, 0.05, 0.75, 0.05],
        ...                      [0.05, 0.05, 0.05, 0.75]])

>>> target = torch.tensor([0, 1, 3, 2])
>>> roc = ROC(num_classes=4)
>>> fpr, tpr, thresholds = roc(pred, target)