Explore fixing test issues for older pytorch versions

[goanpeca]

Supersedes #3356

Description:

Check list:

• New tests are added (if a new feature is added)
• New doc strings: description and/or example code are in RST format
• Documentation is updated (if required)

[Copilot]

Pull Request Overview

This PR addresses compatibility issues with older PyTorch versions by removing or modifying features that were introduced in newer versions. The changes primarily focus on replacing stable=True parameter in torch.argsort() (added in PyTorch 1.13) with version-specific handling and fixing device comparison syntax.

• Remove .bool() calls on tensor operations that already return boolean tensors for older PyTorch compatibility
• Replace direct device object comparisons with device type string comparisons
• Add version checks to conditionally use the stable parameter in torch.argsort()

Reviewed Changes

Copilot reviewed 7 out of 8 changed files in this pull request and generated 2 comments.

Show a summary per file

File	Description
tests/ignite/metrics/vision/test_object_detection_map.py	Replace device object comparisons with device type string comparisons
tests/ignite/metrics/test_mean_average_precision.py	Remove redundant .bool() calls on torch.randint() operations
tests/ignite/distributed/utils/test_native.py	Add version checks and test ordering for older PyTorch compatibility
ignite/metrics/vision/object_detection_average_precision_recall.py	Add version-based handling for stable parameter in torch.argsort() and fix tensor operations
ignite/metrics/mean_average_precision.py	Add version checks for stable parameter and fix device comparisons
ignite/metrics/gan/fid.py	Add .numpy() calls for scipy operations that require numpy arrays
.github/workflows/pytorch-version-tests.yml	Fix Python version string formatting and installation commands

[Copilot]

Converting the tensor to numpy adds an unnecessary GPU-to-CPU transfer. Consider checking if the input tensors are on GPU and handle the conversion more efficiently, or use PyTorch native operations if available.

[Copilot]

Multiple .numpy() calls on the same tensors (sigma1 and sigma2) create redundant GPU-to-CPU transfers. Consider converting these tensors to numpy once and reusing the numpy arrays.

Suggested change

	tr_covmean = np.sum(np.sqrt(((np.diag(sigma1.numpy()) * eps) * (np.diag(sigma2.numpy()) * eps)) / (eps * eps)))
	return float(diff.dot(diff).item() + torch.trace(sigma1) + torch.trace(sigma2) - 2 * tr_covmean)
	tr_covmean = np.sum(np.sqrt(((np.diag(sigma1) * eps) * (np.diag(sigma2) * eps)) / (eps * eps)))
	return float(diff.dot(diff).item() + torch.trace(torch.tensor(sigma1)) + torch.trace(torch.tensor(sigma2)) - 2 * tr_covmean)

[goanpeca]

@vfdev-5 I think these fixes all the issues on pytorch versions.

[goanpeca]

Ok, it seems now there are som typing issues

ignite/distributed/comp_models/native.py:182: error: Invalid index type
"int | float" for "Tensor"; expected type
"SupportsIndex | bool | int | slice[Any, Any, Any] | EllipsisType | Tensor | _NestedSequence[bool | int | slice[Any, Any, Any] | EllipsisType | Tensor | None] | None | tuple[SupportsIndex | bool | int | slice[Any, Any, Any] | EllipsisType | Tensor | _NestedSequence[bool | int | slice[Any, Any, Any] | EllipsisType | Tensor | None] | None, ...]"
 [index]
                local_rank = rank - cumsum_sizes[node_rank].item()
                                                 ^~~~~~~~~
ignite/distributed/comp_models/native.py:183: error: Incompatible return value
type (got "tuple[int, int | float]", expected "tuple[int, int]")  [return-value]
                return int(local_rank), node_rank
                       ^~~~~~~~~~~~~~~~~~~~~~~~~~
ignite/handlers/clearml_logger.py:137: error: Need type annotation for "_task" 
[var-annotated]
            self._task = Task.current_task()
                         ^~~~~~~~~~~~~~~~~~~
ignite/handlers/clearml_logger.py:861: error: Name "Task" already defined
(possibly by an import)  [no-redef]
                    from trains import Task
                    ^~~~~~~~~~~~~~~~~~~~~~~
ignite/handlers/clearml_logger.py:871: error: Need type annotation for "_task" 
[var-annotated]
            self._task = Task.current_task()
                         ^~~~~~~~~~~~~~~~~~~
ignite/handlers/clearml_logger.py:936: error: Name "WeightsFileHandler" already
defined (possibly by an import)  [no-redef]
                    from trains.binding.frameworks import WeightsFileHandl...
                    ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
ignite/handlers/clearml_logger.py:960: error: Argument 1 to "add_pre_callback"
of "WeightsFileHandler" has incompatible type "Callable[[str, Any], Any]";
expected "Callable[[str | CallbackType, ModelInfo], ModelInfo | None]" 
[arg-type]
    ... pre_cb_id = WeightsFileHandler.add_pre_callback(cb_context.pre_callba...
                                                        ^~~~~~~~~~~~~~~~~~~~~~~
ignite/handlers/clearml_logger.py:961: error: Argument 1 to "add_post_callback"
of "WeightsFileHandler" has incompatible type "Callable[[str, Any], Any]";
expected "Callable[[str | CallbackType, ModelInfo], ModelInfo]"  [arg-type]
    ...ost_cb_id = WeightsFileHandler.add_post_callback(cb_context.post_callb...
                                                        ^~~~~~~~~~~~~~~~~~~~~~~~
Found 8 errors in 2 files (checked 154 source files)

[goanpeca]

I dont understand where those issues are coming from now. I run mypy locally and no errors are reported. :\

[vfdev-5]

Maybe, let's just ignore them with appropriate comment ?