search

fundamentalvision / Deformable-DETR

Deformable DETR: Deformable Transformers for End-to-End Object Detection.
Apache License 2.0
3.22k stars 520 forks source link

Got very low AP with resnet101 backbone #140

Closed zyong812 closed 2 years ago

zyong812 commented 2 years ago

Thanks for the nice work! I find a problem when using resnet101 backbone.

I use original code in this repo with --backbone resnet101, and use the default settings. And I have tried batch_size=1 or 2, but both got AP=~42.5 on COCO2017 val split. Have anyone encountered the similar problem? What might be cause?

zyong812 commented 2 years ago

Solved. For some unknow reason, the ImageNet-trained resnet101 is not successfully used for initialization, when it needs downloading while training. So, it's better to download ImageNet-trained beforehand.