miou calculations are time-consuming

[xiuguangLi]

Hello, in the use of torchmetrics. Segmentation. MeanIoU calculation, more time-consuming, even with the cuda.
Example:
num_classes = 21
bs,h,w =10, 512,512
preds = torch.randint(0,num_classes,(bs,h,w))
labels = torch.randint(0,num_classes,(bs,h,w))

Can it be optimized?

[github-actions]

Hi! thanks for your contribution!, great first issue!

[Isalia20]

I got these times, do you expect it to be faster? Would be helpful if you post the times you get or a snippet for which execution is slow