MAP metric, fix metric for CUDA execution

[tkupek]

What does this PR do?

Fixes an issue where the new MAP implementation cannot be executed on CUDA devices.
The tensors have to be initialized/moved to the correct CUDA device.
Fixes #671

Before submitting

• Was this discussed/approved via a Github issue? (no need for typos and docs improvements)
• Did you read the contributor guideline, Pull Request section?
• Did you make sure to update the docs?
• Did you write any new necessary tests?

PR review

Anyone in the community is free to review the PR once the tests have passed.
If we didn't discuss your PR in Github issues there's a high chance it will not be merged.

Did you have fun?

🎉

[codecov]

Codecov Report

Merging #673 (b0a798e) into master (01f88fe) will increase coverage by 0%.
The diff coverage is 100%.

@@          Coverage Diff          @@
##           master   #673   +/-   ##
=====================================
  Coverage      95%    95%           
=====================================
  Files         166    166           
  Lines        6377   6379    +2     
=====================================
+ Hits         6070   6074    +4     
+ Misses        307    305    -2     

[justusschock]

Why wasn't this caught in our test? We do have CUDA tests

[tkupek]

@justusschock Good question. I don't have a special setup. Just DDP mode with a single GPU.

[justusschock]

@tupek could you try to include the necessary changes to the tests in this PR?

[tkupek]

Discussed this already in Slack. @twsl you found the issue with the tests did you?

[Bunoviske]

When executing MAP.compute() on CUDA device, my error was different from #671. I got the following with torchmetrics==0.6.1 and torch==1.10.0:

/usr/local/lib/python3.7/dist-packages/torchmetrics/metric.py in wrapped_func(*args, **kwargs)
    370                 dist_sync_fn=self.dist_sync_fn, should_sync=self._to_sync, should_unsync=self._should_unsync
    371             ):
--> 372                 self._computed = compute(*args, **kwargs)
    373 
    374             return self._computed

/usr/local/lib/python3.7/dist-packages/torchmetrics/detection/map.py in compute(self)
    666             - mar_100_per_class: ``torch.Tensor`` (-1 if class metrics are disabled)
    667         """
--> 668         overall, map, mar = self._calculate(self._get_classes())
    669 
    670         map_per_class_values: Tensor = Tensor([-1])

/usr/local/lib/python3.7/dist-packages/torchmetrics/detection/map.py in _calculate(self, class_ids)
    513         eval_imgs = [
    514             self._evaluate_image(img_id, class_id, area, max_detections, ious)
--> 515             for class_id in class_ids
    516             for area in area_ranges
    517             for img_id in img_ids

/usr/local/lib/python3.7/dist-packages/torchmetrics/detection/map.py in <listcomp>(.0)
    515             for class_id in class_ids
    516             for area in area_ranges
--> 517             for img_id in img_ids
    518         ]
    519 

/usr/local/lib/python3.7/dist-packages/torchmetrics/detection/map.py in _evaluate_image(self, id, class_id, area_range, max_det, ious)
    372 
    373         # sort dt highest score first, sort gt ignore last
--> 374         ignore_area_sorted, gtind = torch.sort(ignore_area)
    375         gt = gt[gtind]
    376         scores = self.detection_scores[id]

RuntimeError: Sort currently does not support bool dtype on CUDA.

[tkupek]

@Bunoviske pretty sure this one was fixed in a previous PR. If you install the lib from main branch the error above should show up.

[Borda]

pretty sure this one was fixed in a previous PR. If you install the lib from main branch the error above should show up.

you can simply install from future bugfix release as

pip install https://github.com/PyTorchLightning/metrics/archive/refs/heads/release/0.6.x.zip

[Borda]

@tkupek could you please add test for this case so we can make quick bug-fix release 🐰
cc: @SkafteNicki

[tkupek]

Sorry, I don't know why the CUDA test is not working.

[twsl]

I think the issue is that in https://github.com/PyTorchLightning/metrics/blob/d071eb2f1245112c599b7fe0d165b8ac55663083/tests/helpers/testers.py#L168 the passed predictions don't satisfy the condition as a list of lists is passed. Therefor the tensors are never move to device.

[SkafteNicki]

I think the tests should be fixed now. Had to change the tester function to move not only tensors but also collection of tensors to the right device.