XNAS: Neural Architecture Search with Expert Advice

Code accompanying the paper

XNAS: Neural Architecture Searchwith Expert Advice
Niv Nayman, Asaf Noy, Tal Ridnik, Itamar Friedman, Rong Jin, Lihi Zelnik-Manor.
arXiv:1906.08031.

The algorithm is based on Prediction with Expert Advice theory (PEA) combined with continuous relaxation and gradient descent in the architecture space introduced in DARTS. It is able to efficiently design high-performance convolutional architectures for image classification (on CIFAR-10, ImageNet and more).

Requirements

Python >= 3.5.5, PyTorch >= 1.1.0, torchvision >= 0.2.0, CUDA >= 10.0, cuDNN >= 7.5, thop

Datasets

CIFAR-10 can be automatically downloaded by torchvision. ImageNet needs to be manually downloaded following the instructions here.

Reproducing the results

We hereby provide the trained models accompanied with the code necessary for reproducing the results reported in the paper.

• Note: reduce the batch size in case of OOM errors, e.g. using the --batch_size flag (default value is 32).

CIFAR-10 | XNAS-Small

python test.py [--calc_flops] --init_channels=36 --model_path="trained_models/xnas_small_cifar10.t7"

• Expected result: 1.81% test error rate with 3.7M model parameters, 621M flops and inference time of 1.98ms per sample.

CIFAR-10 | XNAS-Medium

python test.py [--calc_flops] --init_channels=44 --model_path="trained_models/xnas_medium_cifar10.t7"

• Expected result: 1.73% test error rate with 5.6M model parameters, 905M flops and inference time of 1.99ms per sample.

CIFAR-10 | XNAS-Large

python test.py [--calc_flops] --init_channels=50 --model_path="trained_models/xnas_large_cifar10.t7"

• Expected result: 1.60% test error rate with 7.2M model parameters, 1.15G flops and inference time of 1.94ms per sample.

ImageNet

python test_imagenet.py [--calc_flops] --model_path="trained_models/imagenet.pth.tar"

• Expected result: 23.9% top-1 error with 5.2M model parameters, 590M flops and inference time of 1.49ms per sample.

CIFAR-10 | XNAS-Plain

i.e. XNAS cells trained in the same settings as of DARTS for another comparison.

python test.py [--calc_flops] --model_path="trained_models/xnas_plain.t7"

• Expected result: 2.48% test error rate with 3.8M model parameters, 621M flops and inference time of 2.01ms per sample.

Citation

If you use any part of this code in your research, please cite our paper:

@article{nayman2019xnas,
  title={XNAS: Neural Architecture Search with Expert Advice},
  author={Nayman, Niv and Noy, Asaf and Ridnik, Tal and Friedman, Itamar and Jin, Rong and Zelnik-Manor, Lihi},
  journal={arXiv preprint arXiv:1906.08031},
  year={2019}
}

and DARTS for providing the initial code.