EEEA-Net: An Early Exit Evolutionary Neural Architecture Search

This paper has been published to Engineering Applications of Artificial Intelligence.

Paper: EAAI version or arXiv version

This implementation of EEEA-Net (Early Exit Evolutionary Algorithm Network) from EEEA-Net: An Early Exit Evolutionary Neural Architecture Search by Chakkrit Termritthikun, et al.

This code is based on the implementation of DARTS, NSGA-Net, NSGA-Net-v2, Once for All, and TransferLearning-Tasks.

Results

Prerequisite for server

• Tested on Ubuntu OS version 18.04.x
• Tested on PyTorch 1.6 and TorchVision 0.7.0

Quick Usage (EEEA-Net, ImageNet pre-trained)

install darmo package

pip install darmo

import darmo and create model; see more models at darmo

import darmo
model = darmo.create_model("eeea_c2", num_classes=1000, pretrained=True)

supported transfer learning

model.reset_classifier(num_classes=100, dropout=0.2)

Usage

Cloning source code

git clone https://github.com/chakkritte/EEEA-Net/
cd EEEA-Net/EEEA/cifar

Install Requirements

pip install -r requirements.txt

Architecture search on CIFAR-10 (Normal search)

python search_space.py --dataset cifar10 --search normal --th_param 0.0 

Architecture search on CIFAR-10 (Early Exit search with beta equal 5)

python search_space.py --dataset cifar10 --search ee --th_param 5.0 

Architecture evaluation on CIFAR-10

python train_cifar.py --arch [name]

*[name] is mean a name of models [EA, EEEA_A, EEEA_B, EEEA_C]

Citation

If you use EEEA-Net or any part of this research, please cite our paper:

@article{termritthikun2021eeea,
  title="{EEEA-Net: An Early Exit Evolutionary Neural Architecture Search}",
  author={Termritthikun, Chakkrit and Jamtsho, Yeshi and Ieamsaard, Jirarat and Muneesawang, Paisarn and Lee, Ivan},
  journal={Engineering Applications of Artificial Intelligence},
  volume={104},
  pages={104397},
  year={2021},
  publisher={Elsevier}
}

License

Apache-2.0 License