This project provides the implementation for DetNAS: Backbone Search for Object Detection. As we originally conducted the experiments in the paper using the internal framework Brain++, this project is a reimplemented version on PyTorch. In addition, this project is based on maskrcnn-benchmark.

• Modify the path to your coco dataset in config.sh.
• bash config.sh

The training scripts of these model are in the dirctory scripts/. For training,

• Download the ImageNet model to the directory ImageNet-Pretrain-models/.
• bash scripts/run_detnas_coco_fpn_300M.sh

• We have splitted 5000 images from coco_2014_train+coco_2014_valminusminival as the validation set for search. The remainings are used for supernet training.
• Download the splitted train and val json files (GoogleDrive) to datasets/coco/annotations.
• (You can replace them with your own datasets.)

• Download the ImageNet supernet model (GoogleDrive) to the directory ImageNet-Pretrain-models/.
• If necessary, you can also train models ImageNet by yourselves. Please refer to the folder Supernet-ImageNet/.

• bash scripts/run_detnas_coco_fpn_300M_search.sh
• ('-search' in cfg.MODEL.BACKBONE.CONV_BODY is to distinguish supernet training from single model.)

tmux new -s mq_server
sudo apt update
sudo apt install rabbitmq-server
sudo service rabbitmq-server start
sudo rabbitmqctl add_user test test
sudo rabbitmqctl set_permissions -p / test '.*' '.*' '.*'

• tmux new -s search
• modify host and log_dir in the config file distributed_arch_search/arch_search_config.py.
• bash distributed_arch_search/run_search.sh
• (run_search.sh requires no GPUs.)

• tmux new -s server_x
• modify config-file and MODEL.WEIGHT in the script file distributed_arch_search/run_server.sh.
• bash distributed_arch_search/run_server.sh
• (You can start more than one run_server.sh to speed up, if you have enough GPUs and memory researces.)

Please cite DetNAS in your publications if it helps your research.

@misc{chen2019detnas,
    title={DetNAS: Backbone Search for Object Detection},
    author={Yukang Chen, Tong Yang, Xiangyu Zhang, Gaofeng Meng, Xinyu Xiao, Jian Sun},
    year={2019},
    booktitle = {NeurIPS},
}