search

IDEA-Research / DAB-DETR

[ICLR 2022] Official implementation of the paper "DAB-DETR: Dynamic Anchor Boxes are Better Queries for DETR"
Apache License 2.0
501 stars 86 forks source link

The dropout rate #11

Closed CUQS closed 2 years ago

CUQS commented 2 years ago

Thanks for your great work. However, after reading the code, I'm confused about the setting of dropout.

The dropout rates used in the encoder and decoder layers of DAB-DETR and DAB-Deformable-DETR are set at 0.0 as default. https://github.com/IDEA-opensource/DAB-DETR/blob/2a096e2d59fc804b20dd6da78b504654647107c7/main.py#L79-L80 https://github.com/IDEA-opensource/DAB-DETR/blob/2a096e2d59fc804b20dd6da78b504654647107c7/models/DAB_DETR/transformer.py#L459-L462 https://github.com/IDEA-opensource/DAB-DETR/blob/2a096e2d59fc804b20dd6da78b504654647107c7/models/dab_deformable_detr/deformable_transformer.py#L449-L456 DETR and Deformable-DETR use 0.1 as default, does this means training DAB-DETR or DAB-Deformable-DETR without using dropout can get better performance?

SlongLiu commented 2 years ago

Yes, models with 0.0 dropout tend to have better results.

CUQS commented 2 years ago

Thanks for your quick reply.