train custom dataset with more than 100 object classes. Performance problems?

[nmc-costa]

Hello, thank you for the awesome work.

I was trying to train a dataset with 1230 classes of TAO dataset.

I made some tweaks to the default opt.reid_cls_ids=",".join([str(k) for k in id2cls]). Because it is difficult to manipulate many classes.

However I'm having a lot of problems of performance. CUDA gets out of memory, even if I have only one worker with a batch size of 4.

Should I do some other tweaks to the code.

I didn't change anything else and I'm using the ctdet_coco_dla_2x.pth model.

Thanlk you