Pytorch implementation of DYNAS

This is the implementation of the paper "Subnet-Aware Dynamic Supernet Training for Neural Architecture Search".

For detailed information, please checkout the project site [website] or the paper [arXiv].

Requirements

This repository has been tested with the following libraries:

python==3.8.8
numpy==1.19.2
torch==1.8.1

Getting started

cd exps/NAS-Bench-201-algos

# Example code for SPOS
python train_spos.py \
    --log_dir logs/spos_base \
    --file_name spos_base \
    --method baseline


# Example code for SPOS + Ours
python train_spos.py \
    --log_dir logs/spos_dynamic \
    --file_name spos_dynamic \
    --method dynas

• The work is conducted using the NAS-Bench-201 dataset.
• You can run for other baselines (FairNAS and FSNAS) in a similar way.
• Our code is mainly built on AutoDL.

Citation

@inproceedings{jeon2025subnet,
  title={Subnet-Aware Dynamic Supernet Training for Neural Architecture Search},
  author={Jeon, Jeimin and Oh, Youngmin and Lee, Junghyup and Baek, Donghyeon and Kim, Dohyung and Eom, Chanho and Ham, Bumsub},
  booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition},
  year={2025}
}