easyrobust1/ddetr

• D^2​ETR: Decoder-Only DETR
• • Introduction
  • Results
  • Example usage
  • Citation

D^2​ETR: Decoder-Only DETR

Introduction

The PyTorch implementation of the paper D^2ETR: Decoder-Only DETR with Computationally Efficient Cross-Scale Attention

Results

Model	Epoch	GFLOPs	Params	AP
D^2​ETR	50	82	35	43.2
Deformable D^2​ETR	50	93	40	50.0

Example usage

Reuirements and Instation

• PyTorch versiin >= 1.7.1
• PVTv2-B2-Linear pre-trained model
• Install other libraries

pip install -r requirements.txt

Training

• Train D^2​ETR on 8 GPUs

GPUS_PER_NODE=8 ./tools/run_dist_launch.sh 8 ./configs/ddetr.sh \
--coco_path /path/to/coco \
--pvt_resume /path/to/pvt

• Train Deformable D^2​ETR on 8 GPUs

GPUS_PER_NODE=8 ./tools/run_dist_launch.sh 8 ./configs/def_ddetr.sh \
--coco_path /path/to/coco \
--pvt_resume /path/to/pvt

Evaluation

• Evaluate D^2​ETR

GPUS_PER_NODE=8 ./tools/run_dist_launch.sh 8 ./configs/ddetr.sh \
--coco_path /path/to/coco \
--resume /path/to/model \
--eval

• Evaluate Deformable D^2​ETR

GPUS_PER_NODE=8 ./tools/run_dist_launch.sh 8 ./configs/def_ddetr.sh \
--coco_path /path/to/coco \
--resume /path/to/model \
--eval

Citation

@misc{lin2022d2etr,
      title={D^2ETR: Decoder-Only DETR with Computationally Efficient Cross-Scale Attention}, 
      author={Junyu Lin and Xiaofeng Mao and Yuefeng Chen and Lei Xu and Yuan He and Hui Xue},
      year={2022},
      eprint={2203.00860},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}