EvalMetric for SSD training

I have been trying to use train_ssd.py (link below) with a custom dataset following a similar approach than the finetuning tutorial for object detection with SSD network.

However the method train takes an eval metric, which is not existing for SSD network with custom dataset. (as far as I understood).

So based on the tutorial I started trying to put up an eval metric for SSD, see

I tried to model it against existing VOC and COCO detection metrics but I am a bit lost.
Does it actually make sense ? :sweat_smile:

Also the metric should boil down to one figure like the sum_loss or it could also include the class and bouding box regression loss ?
Because I have seen that the get method could also return lists for names and values.

Thanks for the help !
Btw, this is related to the issue