torchmetrics.detection.mean_ap.MeanAveragePrecision bug if en(max_detection_thresholds)<3

[JuanFMontesinos]

🐛 Bug

torchmetrics.detection.mean_ap.MeanAveragePrecision
an arg is max_detection_thresholds
if len(max_detection_thresholds)<3 there is a bug in pycocotools>cocoeval.py
it seem that the size of this list is hardcoded.

Soz not to be more proactive.

[github-actions]

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

[SkafteNicki]

Hi @JuanFMontesinos, thanks for raising the issue.
I have created PR #2219 that will raise an error when the class is initialized instead of it failing randomly inside pycocotools. There is really nothing more we can do, because then it would need to be fixed directly in pycocotools. That said, the faster-coco-eval backend already supports this, so a fix for users is also just to change backend.

[JuanFMontesinos]

I was imagine so. I think it will be clearer with better documentation so thanks for your PR.
Quick question if you don't mind, how to change the backend?

is the only reference to a backend

[SkafteNicki]

A new backend argument was added in v1.2 of torchmetrics:
https://lightning.ai/docs/torchmetrics/stable/detection/mean_average_precision.html

[JuanFMontesinos]

@SkafteNicki tak!
I think the current docs still point to an older version.

[SkafteNicki]

@JuanFMontesinos, we are in the process of transferring our docs from read the docs server:
https://torchmetrics.readthedocs.io/en/stable/
to own lightning server:
https://lightning.ai/docs/torchmetrics/stable/
which is reason you could not find the argument in the first place. Sorry for the confusion.