projects/detr/README.md

DETR: End-to-End Object Detection with Transformers

Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, Sergey Zagoruyko

[arXiv] [BibTeX]

We reproduce DETR in detrex based on Detectron2 wrapper for DETR.

Converted Weights

Here we provides the weights which are converted by converter.py from the official DETR repo.

Name	Backbone	Pretrain	Epochs	box AP	download
DETR-R50	R-50	IN1k	500	42.0	model
DETR-R50-DC5	R-50	IN1k	500	43.4	model
DETR-R101	R-101	IN1k	500	43.5	model
DETR-R101-DC5	R-101	IN1k	500	44.9	model

Note: Here we borrowed the pretrained weight from DETR official repo. And our detrex training results will be released in the future version.

Training

Training DETR model for 300 epochs:

cd detrex
python tools/train_net.py --config-file projects/detr/configs/detr_r50_300ep.py --num-gpus 8

By default, we use 8 GPUs with total batch size as 64 for training.

Evaluation

Model evaluation can be done as follows:

cd detrex
python tools/train_net.py --config-file projects/detr/configs/path/to/config.py \
    --eval-only train.init_checkpoint=/path/to/model_checkpoint

Evaluating the official DETR model

Using the modified conversion script to convert models trained by the official DETR training loop into the format of detrex model. To download and evaluate DETR-R50 model, simply run:

cd detrex
python projects/detr/converter.py \
    --source_model https://dl.fbaipublicfiles.com/detr/detr-r50-e632da11.pth \
    --output_model converted_detr_r50_model.pth

Then evaluate the converted model like:

python tools/train_net.py --config-file projects/detr/configs/detr_r50_300ep.py \
    --eval-only train.init_checkpoint="./converted_detr_r50_model.pth"

Citing DETR

@inproceedings{carion2020end,
  title={End-to-end object detection with transformers},
  author={Carion, Nicolas and Massa, Francisco and Synnaeve, Gabriel and Usunier, Nicolas and Kirillov, Alexander and Zagoruyko, Sergey},
  booktitle={European conference on computer vision},
  pages={213--229},
  year={2020},
  organization={Springer}
}