Breaking change in MeanAveragePrecision by removing average param

[Advitya17]

🐛 Bug

We're using the MeanAveragePrecision method in the Responsible AI Toolbox repo for calculating Object Detection metrics, and realized that a breaking update in torchmetrics seems to remove the average attribute.

What would now be a way to specify a 'macro' or 'micro' averaging method for calculating the Object Detection metrics, and is that soon to be addressed in a next release?

To Reproduce

Steps to reproduce the behavior...

Code sample

https://github.com/microsoft/responsible-ai-toolbox/blob/main/responsibleai_vision/responsibleai_vision/rai_vision_insights/rai_vision_insights.py#L1110

Expected behavior

Environment

• TorchMetrics version (and how you installed TM, e.g. conda, pip, build from source): pip
• Python & PyTorch Version (e.g., 1.0): PyTorch 1.13.1, python 3.7-3.10
• Any other relevant information such as OS (e.g., Linux): Windows, Ubuntu, MacOS

Additional context

The average attribute has suddenly been removed from https://torchmetrics.readthedocs.io/en/stable/detection/mean_average_precision.html, though still exists for other image scenarios - https://torchmetrics.readthedocs.io/en/stable/classification/average_precision.html

[github-actions]

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

[SkafteNicki]

Hi @Advitya17, thanks for raising this issue.
I am a bit confused because I am pretty sure that MeanAveragePrecision have never had a average argument, but I may be wrong.
What version of torchmetrics have you been using?

[Advitya17]

torchmetrics==1.0.1, we had specified it in our code quite a few months back and were able to use torchmetrics until the breaking change happened.

[SkafteNicki]

Then I do still not understand, because in v1.0.1 of torchmetrics MeanAveragePrecision did not have a average argument:

torchmetrics/src/torchmetrics/detection/mean_ap.py

Lines 241 to 250 in 035e433

	def __init__(
	self,
	box_format: Literal["xyxy", "xywh", "cxcywh"] = "xyxy",
	iou_type: Literal["bbox", "segm"] = "bbox",
	iou_thresholds: Optional[List[float]] = None,
	rec_thresholds: Optional[List[float]] = None,
	max_detection_thresholds: Optional[List[int]] = None,
	class_metrics: bool = False,
	**kwargs: Any,
	) -> None:

[SkafteNicki]

@Advitya17 I can see that you have decided to remove the argument from your code.
Even if I cannot find a release where this have been a feature, would you still like it to be part of a future release or is it not needed anymore?

[Advitya17]

Yes if it can be included in a future release, that'll be very helpful, thank you very much!

[Borda]

Then I do still not understand, because in v1.0.1 of torchmetrics MeanAveragePrecision did not have a average argument:

the arg average was taken/eaten by the **kwargs so it has zero effect but it did not crash...