search

IDEA-Research / DINO

[ICLR 2023] Official implementation of the paper "DINO: DETR with Improved DeNoising Anchor Boxes for End-to-End Object Detection"
Apache License 2.0
2.1k stars 230 forks source link

Transfer learning issue #123

Open fyang064 opened 1 year ago

fyang064 commented 1 year ago

Hello,

I'm wondering how to fix the loading state_dict issue while trying to use pre-trained model. Once the num_classes of the custom dataset has changed from 91, size mismatches for transformer.decoder.class_embed. Is there any better way to implement the transfer learning?

Thanks, Felix

FengLi-ust commented 1 year ago

If your dataset has similar categories as COCO, you can load the corresponding class embedding. But overall class embedding is easy to learn, you can just randomly pick some class embedding in the 91 classes (i.e., the first n classes in 91) or just don't load this parameter.