Sort currently does not support bool dtype on CUDA: Regression in version 0.7.3 +

[Chris-hughes10]

🐛 Bug

When I try to compute MaP on the GPU with no predictions, I receive the error:RuntimeError: Sort currently does not support bool dtype on CUDA.

This occurs in torchmetrics==0.7.3 and torchmetrics==0.8.0, the error does not occur in 0.7.2.

To Reproduce

Steps to reproduce the behavior...

Traceback (most recent call last):
  File "C:/Users/hughesc/OneDrive - Microsoft/Documents/Git/pytorch-accelerated/pytorch_accelerated/exp.py", line 24, in <module>
    print(map.compute())
  File "C:\Users\hughesc\Anaconda3\envs\accelerated-dev\lib\site-packages\torchmetrics\metric.py", line 440, in wrapped_func
    value = compute(*args, **kwargs)
  File "C:\Users\hughesc\Anaconda3\envs\accelerated-dev\lib\site-packages\torchmetrics\detection\mean_ap.py", line 765, in compute
    precisions, recalls = self._calculate(classes)
  File "C:\Users\hughesc\Anaconda3\envs\accelerated-dev\lib\site-packages\torchmetrics\detection\mean_ap.py", line 627, in _calculate
    recall, precision, scores = MeanAveragePrecision.__calculate_recall_precision_scores(
  File "C:\Users\hughesc\Anaconda3\envs\accelerated-dev\lib\site-packages\torchmetrics\detection\mean_ap.py", line 697, in __calculate_recall_precision_scores
    inds = torch.argsort(det_scores, descending=True)
RuntimeError: Sort currently does not support bool dtype on CUDA.

Code sample

import torch
from torchmetrics.detection import MeanAveragePrecision

if __name__ == '__main__':

    device = torch.device('cuda:0')

    preds = [
        dict( boxes=torch.tensor([], device=device),
        scores=torch.tensor([], device=device),
        labels=torch.tensor([], device=device),)
    ]

    targets = [
        dict(boxes=torch.tensor([[1., 2., 3., 4.]], device=device),
             scores=torch.tensor([0.8], device=device),
             labels=torch.tensor([1], device=device), )
    ]

    map = MeanAveragePrecision(class_metrics=True).to(device)

    map.update(preds, targets)

    print(map.compute())

Expected behavior

I would like to be able to compute the metric on the GPU

Environment

• OS (e.g., Linux): Linux & Windows
• Python & PyTorch Version (e.g., 1.0): Python 3.8 & 3.9, PyTorch 1.10
• How you installed PyTorch (conda, pip, build command if you used source): conda
• Any other relevant information:

[github-actions]

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

[krshrimali]

Hi, @Chris-hughes10 - Great first issue! 🎉 I love the description and thanks for sharing the code for reproducing the bug.

I can confirm that this is a regression from release 0.7.2, IMO: an explicit cast from torch.bool to torch.uint8 (on CUDA only) while applying torch.argsort should fix this. Looks like PyTorch doesn't support sorting for boolean dtypes on CUDA devices:

// FIXME: remove this check once cub sort supports bool
TORCH_CHECK(self_dtype != ScalarType::Bool,
  "Sort currently does not support bool dtype on CUDA.");

https://github.com/pytorch/pytorch/blob/1a7e43be141ce01469d7605075cb1008bf19abd7/aten/src/ATen/native/cuda/Sort.cpp#L80

To maintainers: we don't test for boolean inputs (I only see floating point inputs) and this issue only comes when we pass empty preds.

@Chris-hughes10 : In case it's too important and cannot wait till Monday, please let me know. I can create a branch with the fix. I just want to wait for everyone's opinion on this, as they might have more context. :)

cc: @Borda @SkafteNicki @stancld for their inputs

[Borda]

@krshrimali mind sending PR with fix?

[krshrimali]

@krshrimali mind sending PR with fix?

Done :))