Unsatisfying performance on COCO using Swin-T

[Joker316701882]

Hi.

I compared iBOT Swin-T and supervised Swin-T as pre-trained models for COCO, getting the following results:

Supervised Swin-T: mAP 0.432 iBOT Swin-T: mAP 0.428

The detection framework is Mask R-CNN 1x with multi-scale training. Do you have any ideas on that?

[shallowtoil]

Hi @Joker316701882 ,

Did you use the default fine-tuning recipe for supervised Swin-T? If yes, you could probably try to tweak the fine-tuning learning rate, layer decay rate, etc. Setting them smaller usually helps when fine-tuning self-supervised models.