bbssp

Black Box Search Space Profiling for Accelerator-Aware Neural Architecture Search, in ASP-DAC 2020

If you find this work/code helpful, please cite:

@article{zeng2020blackbs,
  title={Black Box Search Space Profiling for Accelerator-Aware Neural Architecture Search},
  author={Shulin Zeng and Hanbo Sun and Y. Xing and Xuefei Ning and Y. Shan and X. Chen and Yu Wang and Huazhong Yang},
  journal={2020 25th Asia and South Pacific Design Automation Conference (ASP-DAC)},
  year={2020},
  pages={518-523}
}

See here for the codes of the search space profiling phase.

Black box search space profiling

This repo doesn't include the scripts and codes of black box search space profiling. Instead, we provide several search space designs after profiling and the policy-aware latency LUT for Xilinx DPU.

Search on the optimized search space

First, we need to set up the plugins for defining different neural network blocks.

cp -r ssp_plugin/ ~/awnas/plugins/

Differentiable method

awnas search --gpu 0 --seed 123 --save-every 50 --train-dir diff_result/cifar10_allblock diff_config/diff_cifar10_allblock.yaml

There is only one yaml file for differentiable method in the diff_config repo. You can find all the yaml files of different optimized search space and experimental settings in the diff_config from this url.

Reinforcement learning method

awnas search --gpu 0 --seed 123 --save-every 50 --train-dir rl_result/cifar10_allblock rl_config/rl_cifar10_allblock.yaml

There is only one yaml file for reinforcement learning method in the rl_config repo. You can find all the yaml files of different optimized search space and experimental settings in the rl_config from this url.