Nueral Architecture Search (NAS)

This project contains the following neural architecture search algorithms, implemented in PyTorch. More NAS resources can be found in Awesome-NAS.

• Network Pruning via Transformable Architecture Search, NeurIPS 2019
• One-Shot Neural Architecture Search via Self-Evaluated Template Network, ICCV 2019
• Searching for A Robust Neural Architecture in Four GPU Hours, CVPR 2019
• 10 NAS algorithms for the neural topology in exps/algos
• Several typical classification models, e.g., ResNet and DenseNet (see BASELINE.md)

Requirements and Preparation

Please install PyTorch>=1.1.0, Python>=3.6, and opencv.

The CIFAR and ImageNet should be downloaded and extracted into $TORCH_HOME. Some methods use knowledge distillation (KD), which require pre-trained models. Please download these models from Google Driver (or train by yourself) and save into .latent-data.

Usefull tools

1. Compute the number of parameters and FLOPs of a model:

from utils import get_model_infos
flop, param  = get_model_infos(net, (1,3,32,32))

2. Different NAS-searched architectures are defined here.

Network Pruning via Transformable Architecture Search

In this paper, we proposed a differentiable searching strategy for transformable architectures, i.e., searching for the depth and width of a deep neural network. You could see the highlight of our Transformable Architecture Search (TAS) at our project page.

Usage

Use bash ./scripts/prepare.sh to prepare data splits for CIFAR-10, CIFARR-100, and ILSVRC2012. If you do not have ILSVRC2012 data, pleasee comment L12 in ./scripts/prepare.sh.

Search the depth configuration of ResNet:

CUDA_VISIBLE_DEVICES=0,1 bash ./scripts-search/search-depth-gumbel.sh cifar10 ResNet110 CIFARX 0.57 -1

Search the width configuration of ResNet:

CUDA_VISIBLE_DEVICES=0,1 bash ./scripts-search/search-width-gumbel.sh cifar10 ResNet110 CIFARX 0.57 -1

Search for both depth and width configuration of ResNet:

CUDA_VISIBLE_DEVICES=0,1 bash ./scripts-search/search-shape-cifar.sh cifar10 ResNet56  CIFARX 0.47 -1

args: cifar10 indicates the dataset name, ResNet56 indicates the basemodel name, CIFARX indicates the searching hyper-parameters, 0.47/0.57 indicates the expected FLOP ratio, -1 indicates the random seed.

Model Configuration

The searched shapes for ResNet-20/32/56/110/164 in Table 3 in the original paper are listed in configs/NeurIPS-2019. If you want to directly train a model with searched configuration of TAS, try these:

CUDA_VISIBLE_DEVICES=0,1 bash ./scripts/tas-infer-train.sh cifar10  C010-ResNet32 -1
CUDA_VISIBLE_DEVICES=0,1 bash ./scripts/tas-infer-train.sh cifar100 C100-ResNet32 -1

One-Shot Neural Architecture Search via Self-Evaluated Template Network

Highlight: we equip one-shot NAS with an architecture sampler and train network weights using uniformly sampling.

Usage

Please use the following scripts to train the searched SETN-searched CNN on CIFAR-10, CIFAR-100, and ImageNet.

CUDA_VISIBLE_DEVICES=0 bash ./scripts/nas-infer-train.sh cifar10  SETN 96 -1
CUDA_VISIBLE_DEVICES=0 bash ./scripts/nas-infer-train.sh cifar100 SETN 96 -1
CUDA_VISIBLE_DEVICES=0,1,2,3 bash ./scripts/nas-infer-train.sh imagenet-1k SETN  256 -1

The searching codes of SETN on a small search space:

CUDA_VISIBLE_DEVICES=0 bash ./scripts-search/algos/SETN.sh cifar10 -1

Searching for A Robust Neural Architecture in Four GPU Hours

We proposed a Gradient-based searching algorithm using Differentiable Architecture Sampling (GDAS). GDAS is baseed on DARTS and improves it with Gumbel-softmax sampling. Experiments on CIFAR-10, CIFAR-100, ImageNet, PTB, and WT2 are reported.

The old version is located at others/GDAS and a paddlepaddle implementation is locate at others/paddlepaddle.

Usage

Please use the following scripts to train the searched GDAS-searched CNN on CIFAR-10, CIFAR-100, and ImageNet.

CUDA_VISIBLE_DEVICES=0 bash ./scripts/nas-infer-train.sh cifar10  GDAS_V1 96 -1
CUDA_VISIBLE_DEVICES=0 bash ./scripts/nas-infer-train.sh cifar100 GDAS_V1 96 -1
CUDA_VISIBLE_DEVICES=0,1,2,3 bash ./scripts/nas-infer-train.sh imagenet-1k GDAS_V1 256 -1

The GDAS searching codes on a small search space:

CUDA_VISIBLE_DEVICES=0 bash ./scripts-search/algos/GDAS.sh cifar10 -1

The baseline searching codes are DARTS:

CUDA_VISIBLE_DEVICES=0 bash ./scripts-search/algos/DARTS-V1.sh cifar10 -1
CUDA_VISIBLE_DEVICES=0 bash ./scripts-search/algos/DARTS-V2.sh cifar10 -1

Citation

If you find that this project helps your research, please consider citing some of the following papers:

@inproceedings{dong2019tas,
  title     = {Network Pruning via Transformable Architecture Search},
  author    = {Dong, Xuanyi and Yang, Yi},
  booktitle = {Neural Information Processing Systems (NeurIPS)},
  year      = {2019}
}
@inproceedings{dong2019one,
  title     = {One-Shot Neural Architecture Search via Self-Evaluated Template Network},
  author    = {Dong, Xuanyi and Yang, Yi},
  booktitle = {Proceedings of the IEEE International Conference on Computer Vision (ICCV)},
  year      = {2019}
}
@inproceedings{dong2019search,
  title     = {Searching for A Robust Neural Architecture in Four GPU Hours},
  author    = {Dong, Xuanyi and Yang, Yi},
  booktitle = {Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
  pages     = {1761--1770},
  year      = {2019}
}