DFT-NPZS-NAS: Prediction-based Zero-Shot NAS with DFT Encodings

Setup

• Clone repo.
• Install necessary packages.

$ pip install -r requirements.txt

• Download databases and requirements

$ bash downloaddata.sh

In our experiments, we do not implement directly the API benchmarks published in their repos (e.g., NAS-Bench-101, NAS-Bench-201, etc). Instead, we create smaller-size databases by accessing their databases and only logging necessary content.

Reproducing the results

You can reproduce our results by running the below scripts:

Train

$ python train.py --benchmark <DARTS, NASNet, ENAS, PNAS, Amoeba, NB201, NB101, Macro, all>

All weight files for the training process are provided here

Evaluate

$ python test.py --checkpoint /path/to/checkpoint

Search

$ python search.py --checkpoint /path/to/checkpoint

If you find this useful...

@misc{le2023efficacy,
      title={Efficacy of Neural Prediction-Based Zero-Shot NAS}, 
      author={Minh Le and Nhan Nguyen and Ngoc Hoang Luong},
      year={2023},
      eprint={2308.16775},
      archivePrefix={arXiv},
      primaryClass={cs.LG}
}

Acknowledgement

Our source code was implemented based on the following sources:

• NAS-Bench-101: Towards Reproducible Neural Architecture Search
• NAS-Bench-201: Extending the Scope of Reproducible Neural Architecture Search
• NAS-Bench-Macro: Prioritized Architecture Sampling with Monto-Carlo Tree Search
• NDS: Designing Network Design Spaces
• ZenNAS: A Zero-Shot NAS for High-Performance Deep Image Recognition
• Fast Differentiable Sorting and Ranking