Instance-wise metric calculation?

[innat]

I'm trying to get sample-wise metric score. However, for good category samples, the label and mask are zero. An output of a well-trained model nicely matches the targets. So, computing this with binary metrices for good samples will give zeo scocre!

from torchmetrics.classification import BinaryPrecision

preds  = torch.tensor([0, 0, 0, 0]) # good mask (black pixel)
target = torch.tensor([0, 0, 0, 0]) # predicted mask (black pixel)
metric = BinaryPrecision()
metric(preds, target) # 0

In pytorch-lightning, the overal batch output wouold be the mean of sample-wise scores, is it correct? For example, the pixel_scores (recall, precision, f1) in the following chart.