muzafferkal / freerea

Official implementation of "FreeREA: Training-Free Evolution-based Architecture Search"

FreeREA

Code release for FreeREA: Training-Free Evolution-Based Architecture Search

If you use this code or the attached files for research purposes, please cite

@InProceedings{cavagnero2022freerea,
    author    = {Cavagnero, Niccol\`o and Robbiano, Luca and Caputo, Barbara and Averta, Giuseppe},
    title     = {FreeREA: Training-Free Evolution-Based Architecture Search},
    booktitle = {Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)},
    month     = {January},
    year      = {2023},
    pages     = {1493-1502}
}

Software requirements

• Python 3.9 or newer
• PyTorch 1.9 or newer
• Other Python libraries listed in requirements.txt

Hardware requirements

A CUDA-capable GPU is required to compute the metrics. However, precomputed metrics for the benchmarks NASBench101 and NATS-Bench are available in the directory cached_metrics.

Run experiments

Results can be reproduced with:

export NATS_PATH=/data/path/to/NATS-tss-v1_0-3ffb9-simple
./run_search.py --space nats --dataset cifar10
./run_search.py --space nats --dataset cifar100
./run_search.py --space nats --dataset ImageNet16-120

export NASBENCH101_PATH=/data/path/to/NASBench-101/nasbench_full.pkl
./run_search.py --space nasbench101 --dataset cifar10 --max-time 12

License

This code and the attached files are distributed under the MIT license.

Code within the directory nas_spaces/_nasbench101 is derived from this repository and is released under the Apache 2.0 license.

Contributors

• Niccolò Cavagnero [email protected]
• Luca Robbiano [email protected]