My Regression loss always 0

[shekarneo]

when i try training on custom data from scratch/head only regression loss always 0.
using debug got terrible results and while inference did not got any bounding boxes

[HaoyangRong]

Having the same issue here. But there's regression loss when training with the toy data 'shape'.

[zylo117]

Refer to FAQ Q3.
#386

[shekarneo]

Why category id matters for regression loss.

[mytk2012]

Why category id matters for regression loss.

This is for the COCO data format.

[HaoyangRong]

Ok, I figured it out. In the annotation json files, the 'image_id's in field 'annotations' is a mapping to the 'id's in field 'images'. And it seems both 'image_id'(under 'annotations') and 'id'(under 'images') should be ints.
If we can use 'image_id' instead of 'id' in field 'images', it should be less confusing. It's kind of strange that even those two keys don't match, the algorithm still runs. I had hash strings and ints respectively. So there's no way to find any matching image file for the annotations. Is it possible to add some format checks?

[shekarneo]

did you got better accuracy, i am getting loss 300-400 it is not decreasing even for more than 100 epochs

[a554142589]

did you got better accuracy, i am getting loss 300-400 it is not decreasing even for more than 100 epochs

hello, Had you fixed the error about the "zero-reg" loss?