Neural Architecture Search Without Training

⚠️ Note: this repository has been updated to reflect the second version of the paper For the original version of the paper, refer to the tag v1.0.:warning:

Usage

Create a conda environment using the env.yml file

conda env create -f env.yml

Activate the environment and follow the instructions to install

Install nasbench (see https://github.com/google-research/nasbench)

Download the NDS data from https://github.com/facebookresearch/nds and place the json files in naswot-codebase/nds_data/ Download the NASbench101 data (see https://github.com/google-research/nasbench) Download the NASbench201 data (see https://github.com/D-X-Y/NAS-Bench-201)

Reproduce all of the results by running

./scorehook.sh

The code is licensed under the MIT licence.

Citing us

If you use or build on our work, please consider citing us:

@inproceedings{mellor2021neural,
    title={Neural Architecture Search without Training},
    author={Joseph Mellor and Jack Turner and Amos Storkey and Elliot J. Crowley},
    year={2021},
    booktitle={International Conference on Machine Learning}
}