Multi label training without overlap

[ShayAdaddi]

Hello,

I succeeded to train a multi class model when I had one annotation per image.
I found out I have more than one class for some of the images so I added annotations with a unique id and same image_id that reflect multiple classes bounding boxes in the same image. There are no overlaps.

The training fails now due to some exceeding index isuue.

I think this problem comes from the fact that my data originally was organised as:

root folder
--> category 1 folder
--> category 2 folder
--> category 3 folder

but actually now images from category 3 folder has bounding boxes of category 2 in the annotation file.

Does yolox supports that ?
If yes, what have I missed ? and how can I debug that ? Thanks!

[akonanykhin]

It supports multiclass datasets. For example, CoCo is multiclass.
I've noticed there is indexing issue in the coco_evaluator which happens when eval set does not contain all of the same classes as train set.
You can either fix the evaluator, or fix your eval set.