AUROC for binary task, if thresholds is set, results in an error

[ahmed-alhindawi]

🐛 Bug

The AUROC metric for a binary task has an optional thresholds argument. It documents that if it is set to an int, then that number of bins is set, otherwise if its a List of floats, then the AUROC at those thresholds is calculated. This is so that a faster calculation can occur. However, if the thresholds argument is set to either of those things, then an error occurs:

RuntimeError: "bincount_cuda" not implemented for 'Float'

To Reproduce

Steps to reproduce the behavior...

import torchmetrics
import torch

labels = torch.Tensor([0, 1]).cuda()
probs = torch.Tensor([0.1, 0.9]).cuda()

metric = torchmetrics.AUROC(task="binary", thresholds=20).cuda()
print(metric(probs, labels))

of note, removing .cuda() also results in an error:

RuntimeError: "bincount_cpu" not implemented for 'Float'

Expected behavior

The AUROC should be calculated along the fast O(n_thresholds) rather than the O(n_samples)

Environment

Installed from Conda with the following other relevant libraries:

• TorchMetrics 11.4 (and 11.3.1)
• Pytorch 1.13.0
• Python 3.10

[github-actions]

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