Unable to reproduce mAP scores on object detection

[vadthyavath]

Discussed in #2620

^{Originally posted by vadthyavath September 6, 2021}
After tuning, Sagemaker inbuilt object detection algorithm(SSD) reports mAP of 71% on the validation set. Then I deployed this model and made predictions on the test set. But I couldn't get mAP scores close to validation set.

So in order to validate mAP metric computation, I made predictions on the validation set and then computed mAP scores. Validation mAP score which I computed doesn't match with sagemaker reported mAP score.

Then to validate mAP scoring criteria, I have used Gluoncv code from here (https://cv.gluon.ai/_modules/gluoncv/utils/metrics/voc_detection.html) and other repo from here https://github.com/microsoft/vision-evaluation/blob/e82f087a4993b8852b8873a2d6c8ec474faa95dd/vision_evaluation/evaluators.py#L373

But I was unable to replicate sagemaker mAP scores.
Could you please provide references to the code which sagemaker uses to compute mAP score so that we can reproduce the same on test set?

update: coco based mAP score also doesn't match.