DETR based on Detectron2

Instructions

1. Follow detectron2's readme to install detection2

• cd DETR.detectron2
• python setup.py build develop
• link dataset path to DETR.detectron2/datasets/

2. Train DETR

• python projects/DETR/train_net.py --num-gpus 8 --config-file projects/DETR/configs/detr.res50.coco.multiscale.150e.yaml: use MSRA pretrain weights

3. Evaluate DETR using provided weights here

• python projects/DETR/train_net.py --num-gpus 8 --config-file projects/DETR/configs/detr.res50.coco.multiscale.150e.yaml --eval-only MODEL.WEIGHTS path/to/provided/ckpt.pth

4. For faster training:

• python projects/DETR/train_net.py --num-gpus 8 --config-file projects/DETR/configs/detr.res50.coco.multiscale.150e.bs48.yaml Using a 8x2080ti server, 150 epochs take about 3 day 6 hours.

Results

config	COCO AP	Paper	Checkpoint
detr.res50.coco.multiscale.150e.yaml	38.6 without RC	39.5 with RC	LINK

"RC" means RandomCrop, it brings about 1% AP improvements accroding to paper.

Disclaimer

• I haven't add RandomCrop.
• I haven't add support for segmentaion, but it can be easily added.

Advantage

• Training faster
• Avoid memory leaking in official implementation
• Use backbone in detectron2